# adikumar.co — full corpus
# Generated 2026-08-23 UTC
# Independent analysis of AI infrastructure, data centre power architecture, and industrial technology by Adi Kumar.
# 
# Every published essay below in reading order (series → publication order).
# See https://adikumar.co/llms.txt for the compact index with 1-line summaries and the scheduled pipeline.

# Contents (45 essays)
#   - The Capacitor Stack: How Energy Storage Became Core Compute Infrastructure
#   - The Wide-Bandgap Stack: Boom, Bust, and the Rebuild of AI's Power Semiconductor Chain
#   - The Thermal Stack: How Heat Became the Binding Constraint on AI Compute
#   - The Interconnect Stack: Transformers, Switchgear and the Grid Equipment That Gates the AI Buildout
#   - The On-Package Delivery Stack: How 48V Becomes 0.8V at 2,250 Amperes
#   - The Modular Datacenter Stack: How the Building Gets Built (When There Aren't Enough Electricians)
#   - The AI Power Chain: Vendor Screen
#   - Pricing Under Scarcity
#   - The Services Inversion
#   - The Wide-Bandgap Stack: Technical Companion
#   - The Thermal Stack: Technical Companion
#   - The Capacitor Stack: Technical Companion
#   - The Interconnect Stack: Technical Companion
#   - The On-Package Delivery Stack: Technical Companion
#   - The Modular Datacenter Stack Technical Companion
#   - How the AI buildout is financed
#   - The AI infrastructure M&A map
#   - The build-to-lease flip
#   - The coverage asymmetry
#   - The PE playbook, layer by layer
#   - Underwriting AI infrastructure debt
#   - The three downside cases
#   - Sovereign capital and industrial policy
#   - Arc behaviour, protection, and how insurance is catching up (DC-DC Transition IV)
#   - Grounding, bonding, and DC ground-fault protection at 800 VDC (DC-DC Transition V)
#   - Battery integration at 800 VDC: UPS restructured, not replaced (DC-DC Transition VI)
#   - Retrofit versus greenfield: the DC-DC decision framework (DC-DC Transition VII)
#   - 800 VDC and liquid cooling: co-emerging, not independent (DC-DC Transition VIII)
#   - Power quality, harmonics, and grid interaction at 800 VDC (DC-DC Transition IX)
#   - Standards evolution: OCP, IEC, NEC, IEEE, and the 2028-2029 convergence (DC-DC Transition X)
#   - Vendor economics: who wins the 800 VDC transition (DC-DC Transition XI)
#   - Commissioning, skills, and operational readiness at 800 VDC (DC-DC Transition XII)
#   - The ten-year view: where 800 VDC data centres land by 2028, 2030, and 2035 (DC-DC Transition XIII, series closing)
#   - The real reason data centres are going DC (and it's not capacity alone) (1/13)
#   - Two architectures wearing the same name (and why hyperscalers built their own) (2/13)
#   - The architecture map (where every stage lives, who builds it, what's standardized) (3/13)
#   - The Transition Tax: where the 800V DC capex premium lives (DC-DC Supplement A)
#   - Hybrid Forever: most facilities will run mixed AC/DC through 2035+ (DC-DC Supplement B)
#   - The Voltage Ceiling: 1500V DC and the MVDC question (DC-DC Supplement C)
#   - The Solid-State Transformer: Layer 2 supplier dynamics (DC-DC Supplement D)
#   - The Sidecar Power Rack: Layer 4 and the OCP-spec ecosystem (DC-DC Supplement E)
#   - The Connector Wars: Layer 4/5 interface engineering (DC-DC Supplement F, series close)
#   - Equinix + CPP / atNorth $4B: what DD would flag on a Nordic data centre take-out
#   - NVIDIA / OpenAI PORTS-Pike: what DD would flag on a $105B guarantee for a phased 4.25 GW Ohio campus
#   - July 2026: the month the grid became the ceiling: What Changed Issue 01

==============================================================================
# The Capacitor Stack: How Energy Storage Became Core Compute Infrastructure
URL: https://adikumar.co/the-capacitor-stack-800vdc/
Published: 2026-07-22
Summary: The capacitor stack for 800V DC data centres: HSCs, EDLCs, MLCCs, silicon caps. Four capacitor categories are the binding AI supply constraint.
==============================================================================

The AI Power Chain series · Part 7 of 15








Glossary of terms used



GaN
Gallium Nitride. Wide-bandgap semiconductor material used in RF and power applications.
HSC
Hybrid Supercapacitor. Energy storage technology bridging batteries and traditional capacitors.
OCP
Open Compute Project. Hyperscaler-led standards body developing open reference designs for data centre hardware.
SiC
Silicon Carbide. Wide-bandgap semiconductor material used in high-voltage power electronics.
TAM
Total Addressable Market. The maximum revenue opportunity available if a product served every potential customer segment.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.








Market deep dive · AI data center power


The 800VDC transition has turned four component categories (hybrid supercapacitors, EDLCs, MLCCs, and silicon capacitors) into some of the most supply-constrained, strategically contested layers of the AI buildout. A layer-by-layer analysis of the technology, the market structure, the component supply chains beneath them, and the M&A chessboard forming above them.



Executive summary, eight findings.

 1. Energy storage is now a compute platform component, not a reliability appendix: ~20x storage step-up at Vera Rubin; capacitive shelves native to the 2027 Kyber/800VDC reference architecture.
 2. HSCs own the contested 10ms. 10s window; the binding constraint is cell capacity (2026 demand ~2-3x supply), not demand or technology. EDLC carries 2026 deployments as the bridge.
 3. A bottom-up TAM build puts datacenter capacitive storage systems at roughly $0.8-1.4B in 2027, scaling toward $3-6B by 2030 in the base case, with assumptions stated and stress-testable (§TAM).
 4. Profit pools migrate: scarcity rent sits with cells and materials through 2027; systems/integration margin compresses fastest once OCP standardizes the shelf.
 5. The materials layer is tighter than the cell layer: Kuraray (cathode carbon), Nippon Kodoshi (separators), Sumitomo Metal Mining (MLCC nickel paste) are the unpriced chokepoints.
 6. China is the supply-side wildcard: 30-50% cost deltas vs. Japanese cells, held out of hyperscale sockets mainly by qualification friction and provenance rules, a moat, not a wall.
 7. The M&A window is open and closing: Musashi ES inside an auto-parts parent is the classic carve-out arbitrage. Tesla's $218M Maxwell deal (2019) is the haunting precedent benchmark.
 8. Batteries cannot retake the smoothing socket (a training workload is ~15M micro-cycles/year; no battery chemistry survives it), they defend only the minutes-scale autonomy socket.



33 kVStage 0 · The thesis

The AI Power Chain · six essays, one physical arc


The series walks a single physical path. It begins at the medium-voltage utility bus at the site fence, steps down through the substation and switchgear, arrives at the datacenter rack where 800V DC is stabilised by the capacitor stack, is converted by silicon-carbide switches to 48V, is distributed across the rack by copper busbars and whips, is stepped down again by multi-phase controllers on the accelerator board to 0.8V, and finally routed through the on-package power delivery network to a transistor gate drawing over 2,000 amperesWaste heat from every conversion stage is removed by the thermal stack. The whole thing is packaged inside a factory-modular building because there aren't enough electricians to build it stick-frame. Six essays. One 800V → 0.8V staircase.


 1. Part I. The Capacitor Stack 800VDC at the rack (you are here)
 2. Part II. The Wide-Bandgap Stack SiC and GaN conversion
 3. Part III. The Thermal Stack Removing the waste heat
 4. Part IV. The Interconnect Stack Busbars and whips
 5. Part V. The On-Package Delivery Stack 48V to 0.8V
 6. Part VI. The Modular Datacenter Stack How the building gets built



↓

800 V DC
Rack inlet · energy shelf~750 A busbar at the rack; the capacitive shelf absorbs the 10 ms. 10 s transient before it hits the grid
The contested socketHSC / LIC cells & modules; EDLC bridge; Al electrolytic (incumbent, being displaced)

↓

800 → 48 V
Power shelf DC-DCIsolated step-down inside the rack; feeds board voltage regulators
Bulk boardPolymer bulk; hybrid Al-polymer; Al electrolytic

↓

48 → 0.75 V
Board POL & VRMPoint-of-load conversion on the accelerator PCB; µs. Ms decoupling
Decoupling arrayMLCC arrays (thousands per accelerator); polymer tantalum

↓

0.75 V · 4 kA+
In-package · under the dieSub-microsecond decoupling at the silicon boundary; the zone MLCCs cannot physically reach
Structural design-inSilicon capacitors (deep-trench); embedded MLCC


Indicative topology. Six distinct capacitor markets sit on one power path; each owns a different transient window and answers to different economics.


~20xIncrease in integrated energy storage, Vera Rubin vs. prior NVIDIA generation
600 kWKyber rack power, 2027, 800VDC at the rack inlet, ~750A busbar
2-3x2026 AI-server supercapacitor demand vs. available manufacturing capacity
+15. 35%Murata MLCC price increases effective April 2026; peers followed


The forcing function: rack power density
kW per rack, NVIDIA flagship platforms, 2023-2029E

Source: NVIDIA roadmap disclosures (GTC 2025/2026), OCP; 2028-29 indicative. Author's compilation.



Rack power has grown more than 4x in two years, and the industry is openly designing for 1MW racks. Below ~200kW, conventional AC distribution and software smoothing suffice. Between 200-350kW, capacitive peak absorption becomes economically compelling. Above ~350kW, the entire 2027+ roadmap, the 800VDC architecture with native energy shelves becomes the default, and with it a new, large, and structurally growing demand pool for every layer of the capacitor stack.


STACKThe framework · Four markets, one problem



One pulsed load, four capacitor markets


The clean way to hold this space in your head is as a timescale ladder. Each rung is a distinct component market with distinct suppliers, cost structures, and competitive dynamics, and AI has tightened supply on every rung simultaneously, for the same root cause.



LayerTimescaleTechnology2026 market (est.)LeadersAI-era dynamic

In-packagens. µsSilicon capacitors, embedded MLCC~$1-2B broad; AI-substrate niche inflectingMurata/IPDiA, SEMCO, TSMCDesign-in via BSPDN
Board / POLµs. MsMLCC arrays, polymer~$15-23BMurata, SEMCO, Taiyo Yuden, TDKShortage; +15-35% pricing
Power shelf bulkms. 100msAl electrolytic / polymer → HSC migration~$8-10B Al-electrolytic categoryNippon Chemi-Con, Rubycon, Nichicon, PanasonicIncumbent being displaced up-shelf; polymer grows on-board
Rack / sidecar10ms. 10sHSC (sweet spot), EDLC bridge~$1-3B supercap categoryMusashi, Panasonic, SkeletonDemand 2-3x supply
Bridge / BBU10s. MinLi-ion BBUAdjacent (battery market)CATL, LGES, Samsung SDICoexists; fire-code capped
Campus BESSmin. HrsLi-ion LFPAdjacent (BESS market)Tesla, Fluence, SungrowInterconnect compliance





The timescale ladder: one pulsed load, six windows
Which device class owns each transient decade. The red rung is the socket AI created and where the supply gap sits.


Silicon caps · MLCCin-package / on-die

ns. µs


MLCC arrays · polymerboard / POL

µs. Ms


Al electrolytic → HSCpower shelf bulk

ms. 100 ms


HSC (sweet spot) · EDLC bridgerack & sidecar

10 ms. 10 s


Li-ion BBUbridge to generator

10 s. Min


Li-ion LFPcampus BESS

min. Hrs



nsµsmssminhrs



Log-time layout, decades roughly evenly spaced. Each rung is a distinct component market, same problem at the load, different physics at each device class.

MLCCs and HSCs don't compete, they're the two ends of the same pulsed-load problem, and AI made both scarce at the same time.The stack thesis in one line
800 VLayer 1 · The energy shelf





In brief


Four capacitor categories bind AI data centre supply: hybrid supercapacitors (HSCs), electric double-layer capacitors (EDLCs), multi-layer ceramic capacitors (MLCCs), and silicon capacitors. Each addresses a different frequency + energy-density regime in 800V DC systems. Concentration among top vendors is high; substitution across categories is limited by electrical + qualification specifics.








Hybrid supercapacitors: the contested crown jewel




The technology, briefly and precisely


A hybrid supercapacitor (HSC), commercially dominated by the lithium-ion capacitor (LIC) architecture, is an asymmetric device: a battery-type anode (graphite or hard carbon, pre-doped with lithium ions) paired with an EDLC-type cathode (high-surface-area activated carbon). The asymmetry is the entire point. The pre-doped anode drags the cell voltage up to 3.8-4.0V (vs. 2.7-3.0V for symmetric EDLCs) and contributes faradaic energy storage; the capacitive cathode preserves the power delivery and cycle life of a capacitor. The result sits precisely in the gap the AI load profile created:


 * Energy density ~20-25 Wh/kg, roughly 3-4x large-format EDLCs (~5-7 Wh/kg), up to 10x smaller ones. Enough for seconds of ride-through at hundreds of kW in a shelf-sized envelope.
 * Cycle life 50,000 to 1M+ cycles under continuous micro-cycling, the duty that destroys Li-ion in months.
 * Voltage economics: ~200-210 series cells for an 800V string vs. ~300 for EDLC, fewer cells, fewer balancing circuits, less interconnect.
 * Safety: no Li-ion thermal-runaway profile; UL-certified cells exist; dramatically lighter fire-code treatment (NFPA 855) inside white space.




Value proposition and market drivers


The HSC's job in the 800VDC architecture is the 10ms. 10s window: load smoothing, ride-through, and peak shaving at the power shelf, rack, and sidecar. Three demand drivers compound:


 1. Capex arbitrage. Buffering lets operators provision the upstream chain to average rather than peak load, the economic engine behind NVIDIA's ~30% peak-reduction claim. At gigawatt campus scale, 30% of grid allocation is worth billions.
 2. Grid compliance. Utilities increasingly cap ramp rates in interconnect agreements. Rack-level capacitive buffering converts a compliance obligation into a component purchase order.
 3. Architecture design-in. The OCP/NVIDIA reference architecture includes capacitive energy shelves natively. Neoclouds (CoreWeave, Lambda, Nebius, Together, OCI) inherit the reference design wholesale, an attach-rate story, not a persuasion story.




Market size: read the scope before the number


Supercapacitor market estimates diverge wildly, from ~$0.6B to ~$17B for 2026, and the divergence is informative rather than embarrassing: it is almost entirely a scope question (bare cells vs. modules vs. integrated systems; whether adjacent hybrid storage systems are counted). For working purposes I anchor on a device-level category of roughly $1-3B in 2026, growing at ~15-26% CAGR depending on scope, with hybrid/LIC devices the fastest-growing configuration (~17-26% CAGR; roughly a quarter to 40%+ of new installations by various counts). The datacenter segment is the swing factor: effectively near-zero in 2024, it is the single largest source of forecast variance for 2027-2030.




One market, five answers: supercapacitor market sizing by research house
Base year (2025/26) vs. each house's own outer-year forecast (2030-2036: horizons differ), USD billions, log scale. Scope differences explain the spread

Sources: Mordor Intelligence (2026→2031), MarketsandMarkets (2025→2030), IDTechEx (→2036), Fortune Business Insights (2026→2034), Astute Analytica (2025→2035). Scopes differ: cells vs. modules vs. systems.


Analyst's caveat: when a category's size estimates span an order of magnitude, the honest position is a range plus a growth vector, not false precision. What every house agrees on: double-digit CAGR, hybridization as the growth engine, and datacenter as the new demand pool. For diligence, size bottom-up from racks: energy-shelf content per MW of IT load × deployment forecasts beats top-down every time.




The defining fact of 2026: the supply gap


Demand estimates circulating in H1 2026 put AI-server supercapacitor requirements at ~15-18 million units against roughly 6.5 million units of annual capacity at Musashi, the category leader. That is a structural 2-3x shortfall. Panasonic's first datacenter-dedicated shipments (its FY2027, beginning April 2026) were effectively pre-sold before leaving the factory. Hyperscaler qualification cycles run 12-18 months, so even capacity now under construction converts to qualified supply slowly. The bottleneck runs through cell supply rather than through demand or through technology.




The gap that sets pricing power
2026E AI-server supercapacitor demand vs. leading-supplier capacity, million units

Source: trade press / industry estimates (Jianshi, May 2026); directional. Capacity additions ramping 2026-27 at Musashi, Panasonic, Skeleton, Chinese LIC entrants.





Key players and approximate positioning



PlayerPosition (Jul 2026)Est. category standingVerdict

Musashi Energy Solutions (JP)LIC category leader; cells, modules, ESS400 systems; Flex CESS partnership; UL cells; new plant rampingDominant in LIC, plausibly >50% of merchant LIC cell supplyThe anchor asset of the space; supply-gated
Panasonic (JP)EDLC scale incumbent pivoting hard to datacenter; FY2027 output pre-sold; plays both EDLC and hybrid chemistriesTop-3 supercap franchise globallyThe credible Tier-1 counterweight
Skeleton Technologies (EE/DE)Curved-graphene EDLC + "SuperBattery" hybrid; European sovereignty angle; grid/industrial tractionLeading EU independentDesign-win watch for EU datacenter builds
UCAP Power / Maxwell (US)Maxwell ultracap lineage post-Tesla divestiture; industrial/transport nichesNiche; sub-scale for hyperscaleCautionary tale; consolidation candidate
Eaton (US/IE)Supercap modules within backup/power-quality portfolio; full 800VDC chain participantChannel-strong module playerIntegration route into critical infra
LICAP, Nichicon, KYOCERA-AVX, CAP-XX, SPEL, Abracon, Long Sing + Chinese LIC entrantsDry-electrode IP (LICAP); regional/second-tier positions; Chinese entrants scaling on costFragmented tailQualification friction is the moat against the tail






Underneath the cell: the component supply chain


For PE diligence, the cell makers are the visible layer; the leverage often sits one level down. An LIC cell decomposes into six critical material systems, several with chokepoint structures tighter than the cell market itself:




ComponentWhat it isKey suppliersStructure

Cathode: activated carbon1,500-2,200 m²/g coconut-shell or resin-derived carbon; the capacitive electrodeKuraray (YP-50F, the de facto industry standard), Power Carbon Technology (KR), Haycarb (LK), Cabot/Norit, Fujian Yuanli, Beihai Sence (CN)~$126M market (2024) → ~$240M by 2032; top-3 hold ~67%; APAC ~64% of demand. Kuraray is the tightest identified chokepoint in the stack.
Anode: hard carbon / graphiteBattery-type lithium-intercalation electrodeKureha (hard carbon), Resonac (ex-Hitachi Chemical), POSCO Future M, BTR (CN)Shared with the Li-ion supply chain, deep, but quality tiers matter for cycle life
Lithium pre-doping sourceUltra-thin lithium foil or stabilized Li powder consumed in anode pre-lithiationHonjo Metal (JP), Livent/Arcadium lineage, China Energy LithiumNarrow, specialized; pre-doping process IP is a core cell-maker differentiator
SeparatorCellulose (TF40-class) or PE/PP membraneNippon Kodoshi (NKK), dominant in cellulose capacitor separators; Asahi Kasei, Toray (polyolefin)NKK is a quiet quasi-monopoly in capacitor-grade cellulose
ElectrolyteLiPF6-based (LIC) or TEABF4/acetonitrile (EDLC)Mitsubishi Chemical, Central Glass, Capchem (CN), Guotai Huarong (CN)Shared with Li-ion; acetonitrile purity is the EDLC-side constraint
Current collectors & bindersEtched/perforated Al foil (cathode), Cu foil (anode); PVDF/SBR bindersToyo Aluminium, JCC, SDX; Kureha & Solvay (PVDF), Zeon (SBR)Perforated foil for pre-doping ion transport is a specialty niche





The Kuraray point: when one Japanese chemical company's coconut-shell carbon (YP series) is the qualified cathode material in the majority of the world's supercapacitors, the materials layer, not the cell layer, is where a determined acquirer or a disruption (synthetic carbons, curved graphene, MOF-derived carbons) changes the game. Watch capacity announcements in activated carbon as a leading indicator of cell-maker conviction.




The China question: the supply-side wildcard


One line about "Chinese entrants" undersells the single largest uncertainty on the supply side. China already dominates supercapacitors in its home applications, Shanghai Aowei's ultracap buses, CRRC's rail and grid programs, Jianghai's capacitor franchise spanning electrolytics into EDLC/LIC, and Chinese LIC cell lines are scaling now with reported cost positions 30-50% below Japanese incumbents, riding the Li-ion ecosystem's shared materials base (anodes, electrolyte, lithium) at Chinese scale.


What keeps this capacity out of hyperscale sockets is gated by three stacked frictions rather than by technology: (1) qualification, 12-18 month hyperscaler cycles with reliability data requirements that favor incumbents with decades of field hours; (2) provenance, US and increasingly EU supply-chain rules, tariff exposure, and hyperscaler procurement policies that treat Chinese energy-storage content as a compliance risk in critical infrastructure; (3) design lock-in, once the OCP shelf spec and NVIDIA reference qualify specific cells, the socket is sticky for a platform generation. The strategic read: this is a moat, not a wall. The base case is a bifurcated market, Japanese/Korean/EU cells in Western hyperscale, Chinese cells in Chinese datacenters (a very large market in its own right) and the global industrial tail. The bear case for incumbent pricing is a 2028 scenario where the Western supply gap persists, a Tier-2 cloud breaks ranks to qualify a Chinese cell, and the price umbrella collapses 30%+ in a year. Watch for exactly one datapoint: the first non-Chinese cloud operator publicly deploying Chinese LIC cells. That is the umbrella's first tear.




Case study: Flex × Musashi CESS: the reference commercial construct
Teardown of the first-to-market capacitive energy storage system for AI racks


What it is: a rack-integrated Capacitor-based Energy Storage System pairing Musashi's hybrid supercapacitor cells/modules (ESS400 lineage, UL-certified) with Flex's power-shelf engineering and manufacturing scale, sold into the NVIDIA 800VDC ecosystem for transient smoothing and grid protection. Who supplies what: Musashi, cells, modules, charge/discharge management IP; Flex, shelf electronics, DC/DC interface, rack integration, hyperscaler channel (with its Anord Mardix switchgear franchise adjacent). Why it matters: it is the template for how HSC capacity reaches hyperscalers, cell maker + integrator, not cell maker direct, and the template acquirers will study. Indicative economics: if a CESS shelf lets an operator provision upstream infrastructure to average rather than peak (the ~30% headroom claim), the value released per 600kW rack is measured in tens of thousands of dollars of avoided upstream capex against a shelf costing perhaps $8-15k, a payback measured in months, which is why demand is supply-constrained rather than price-constrained.




800 VLayer 1b · The incumbent bridge



EDLCs: squeezed, but indispensable through 2027


The symmetric electric double-layer capacitor is the incumbent being compressed from both directions, HSCs above it on energy density, improving electrolytics plus software smoothing below it on cost. Yet the EDLC is what is carrying the 2026 datacenter deployments: proven, safe, extreme cycle life, mature multi-source supply, and available now while HSC capacity is oversubscribed. Panasonic's current rack/zone/distribution-level peak-shaving line is EDLC-led. The category retains durable non-datacenter strongholds, wind pitch control, rail regeneration, industrial ride-through, automotive backup, and grows overall even as classic EDLC's share of the supercapacitor category declines (EDLC held ~55% of category revenue in 2026 by Mordor's count, with hybrids growing ~3-4 points faster annually).


The Maxwell parable is worth a paragraph, because it explains today's market structure. Maxwell Technologies effectively invented the commercial ultracapacitor category, was acquired by Tesla in 2019 primarily for its dry-electrode battery IP, and its ultracap business was divested to UCAP Power in 2021. The years of ownership churn coincided exactly with the formation of the AI demand wave, leaving the category's original leader sub-scale precisely when the largest demand pool in its history arrived. The AI datacenter cycle is being won by Musashi, Panasonic, and the integrators, while the dry-electrode processing IP that Tesla kept is, ironically, now a key cost-down vector for the whole industry (LICAP being the notable independent practitioner).




The displaced incumbent: aluminium electrolytics and polymer


Completeness demands a section most analyses skip: the technology the HSC is displacing. Today's GB300 power shelves do their smoothing with aluminium electrolytic capacitors, the incumbent bulk-storage technology of every PSU ever built, an ~$8-10B global category led by Nippon Chemi-Con, Rubycon, Nichicon, and Panasonic. These are the current holders of the power-shelf socket, and they are responding: Chemi-Con has launched immersion-cooling-compatible capacitors specifically for AI servers, and shelf-grade electrolytic content per rack has been rising with the GB300 storage step-up. The displacement dynamic is therefore subtle: the electrolytic makers enjoy a strong 2025-27 cycle as storage content grows, then face structural substitution at the shelf level as the Rubin/Kyber generation migrates bulk storage to supercapacitor-class devices, while conductive polymer capacitors (Panasonic's SP-Cap franchise, polymer tantalum from Kyocera-AVX and Yageo/Kemet) keep growing on-board where their low ESR earns the socket regardless. Portfolio read: the pure-play electrolytic names are enjoying a cyclical peak inside a secular handover, the classic value trap setup if the handover is mispriced, and a pivot story for whichever of them buys or builds a hybrid-capacitor position (Nichicon, notably, already sells small LICs).



54 VLayer 2 · The board



MLCCs: the shortage economy arrives



Why AI broke the MLCC market


Multilayer ceramic capacitors operate six-plus orders of magnitude below HSCs in energy, microfarads, not farads, handling the microsecond-to-millisecond decoupling layer. They are not a substitute for HSCs and never will be; they are the same problem's other end. What AI did to this market is arithmetic: an AI server carries 10-15x the MLCC content of a general-purpose server. A GB300-class system carries ~30,000 MLCCs; a VR200 NVL72 rack approaches ~600,000The specification mix shifted simultaneously toward the hardest parts to make: ultra-high-capacitance (Murata's 47µF in an 0402 case, ~800 dielectric layers, a world first), higher voltage classes for 800V ecosystems, low-ESL geometries, and embedded MLCCs for in-substrate placement (Taiyo Yuden's 22µF 0402).


The result, visible across 2026: high-end line utilization at 90-95%, lead times stretched from ~4 to ~20 weeks, reported AI-server orders at ~2x Murata's available capacity, and the first broad, coordinated price increases in years. Murata raised prices 15-35% effective April 1, 2026 (AI-server high-cap, automotive, RF); Samsung Electro-Mechanics followed at +15-20%, Taiyo Yuden +15-25%, Yageo/Kemet +10-15%, with a further +20-30% projected for H2 2026 on high-cap grades. Murata is deploying ¥250B of FY2026 capex (¥80B for server-grade MLCC) yet guides only ~20% capacity growth over two years, deliberately short of demand growth. This is a supplier's market by choice as much as by constraint: the top three learned the lessons of the 2018 MLCC cycle and are prioritizing price over share.




2026 MLCC price actions: the whole oligopoly moved
Announced/reported price increase ranges by vendor, % (bars show midpoint; labels show range)

Sources: trade press compilations (Apr. Jun 2026); ranges vary by product grade. H2 2026 projections: +20-30% high-cap, +30-50% ultra-high-cap.





Market size and structure


The MLCC market is measured more consistently than supercapacitors: 2026 estimates cluster around $15-23B (with outliers to $32B on broader scopes), growing 6-15% CAGR depending on the house, with the AI/high-cap mix growing far faster than the blended rate. Structure is a stable oligopoly: Murata (~31-32%) leads, followed by Samsung Electro-Mechanics (~19-20%), Taiyo Yuden (~13%), TDK (~10%), Yageo/Kemet (~8-9%), with Kyocera-AVX, Walsin, Samwha, and Chinese entrants (Fenghua, Three-Circle) in the tail, the top five controlling roughly 70% of revenue. Premium AI mix concentrates economics further toward the top three, who alone can yield ultra-high-cap small-case parts at scale.




MLCC market share, approximate (2025/26 revenue basis)
Global MLCC revenue share, %: top five ≈ 70%

Sources: industry estimates (Dataintelo, TrendForce-adjacent trade press, company disclosures); approximate.





The MLCC component chain



ComponentRoleKey suppliersNote

Barium titanate powderThe dielectric, sub-100nm grades for thin layersSakai Chemical, Nippon Chemical Industrial, Toho Titanium, Ferro (Vibrantz), KCM/Sinocera (CN)Ultra-fine grades for 800-layer parts are the yield gate; Japanese suppliers dominate the top tier
Nickel inner-electrode pasteConductive layers between dielectric sheetsSumitomo Metal Mining (dominant), Shoei Chemical, NoritakeSMM's nickel powder position is an underappreciated chokepoint
Release film (PET)Carrier film for sheet casting, consumed in volumeToyobo, Mitsubishi Chemical, Toray, SKCWas the binding constraint in the 2018 cycle; watch again
Termination pastes / platingCu/Ni/Sn external electrodesShoei, Sumitomo, plating chemistry majorsCommodity-adjacent





0.75 VLayer 3 · Under the die



Silicon capacitors: the structural design-in


Deep-trench silicon capacitors compete with MLCCs at the point-of-load, not with HSCs, and their story is about geometry, not economics: as accelerator packages pull 4,000A+ at sub-volt rails, decoupling must sit inside the package, where discrete MLCCs physically cannot go. Proven for a decade in RF/mobile (Murata's IPDiA lineage), the category crossed its commercial Rubicon in May 2026 when Samsung Electro-Mechanics announced a silicon capacitor supply contract worth ~$1B with a major global customer for 2027-28 delivery, the first billion-dollar commitment in the category's history, built on SEMCO's unique full-stack position (silicon caps + FC-BGA substrate + MLCC as one bundled power-integrity offering). TSMC brings a logic-derived trench structure natural to CoWoS-class integration; Murata/IPDiA defends the incumbent trench franchise; Empower and substrate players (Ibiden, Shinko ecosystem) build embedding routes.


Hurdles are real: cost per µF remains ~5-10x MLCC (so adoption pencils only where MLCCs can't go), capacitance density trails, and the qualified supplier base is a handful of names against a vast MLCC ecosystem. The trigger that changes the slope is backside power delivery on 2nm-class accelerators (Intel PowerVia-class, TSMC A16-era): once power enters through the wafer backside, silicon capacitance is designed in structurally, a 2026-28 design-in window for 2027-29 silicon, compounded by co-packaged optics in the Feynman era. Broad-scope market estimates (~$1.9B in 2026) mostly measure the legacy RF/medical base; the AI-substrate niche is small today and is precisely where the growth concentrates. Expect silicon caps and MLCCs to split the power delivery network by physics, in-package vs. on-board, rather than fight for the same sockets.



COUNTERThe battery-side response, steelmanned



Can Li-ion take the socket back?


The battery industry's counterattack is coming, and it is well-funded, worth steelmanning. The candidates: lithium-titanate (LTO), Toshiba's SCiB heritage, 20,000+ deep cycles, wide temperature range, intrinsically safer chemistry; high-C-rate LFP, riding the largest manufacturing base in energy storage history; and sodium-ion power cells (CATL's Naxtra generation, BYD), which remove lithium exposure and improve safety further. All three are cheaper per Wh than any supercapacitor by roughly an order of magnitude, and all three are pushing power density upward each generation.


They still lose the smoothing socket on arithmetic. A training workload's power oscillation cycles roughly every couple of seconds, continuously. That is on the order of 15 million micro-cycles per yearAn exceptional LTO cell survives perhaps 40,000 deep cycles. Call it a few hundred thousand shallow ones with generous derating. The best battery on the market is consumed in weeks against a duty an HSC shrugs at for a decade. Depth-of-discharge management can stretch this (shallow cycling extends battery life materially), but shallow cycling means oversizing, and oversized LTO at datacenter reliability grades erases the cost advantage while retaining the NFPA 855 fire-code burden inside white space. The batteries' defensible ground is real but bounded: the seconds-to-minutes autonomy socket (BBU shelves, bridge-to-generator), where energy density genuinely dominates and cycling is rare. The likely 2028+ equilibrium is explicitly hybrid. HSC absorbs the churn; a smaller and longer-lived battery bank holds the autonomy. This increases total storage content per rack rather than picking one winner. The battery industry's best outcome in this niche is coexistence; its realistic worst case is watching silicon-anode HSCs (40-50+ Wh/kg on the lab-to-fab path) eat the bottom of the BBU segment from below.



OUTLOOKForecasts · Disruption · Capital



Where this goes: growth, technology, and the M&A chessboard




A bottom-up TAM: sizing from racks, not reports


Earlier I argued that when analyst estimates span an order of magnitude, the useful move is to build the number yourself. So: the model, with every assumption visible so you can break it.


 * Content per rack. A Kyber-class 600kW rack needs a usable capacitive buffer on the order of 250-500Wh (seconds-scale smoothing of a several-hundred-kW swing). At volume LIC cell economics of roughly $8-15/Wh, plus shelf electronics, controls, and integration at a 1.8-2.2x system multiple, installed capacitive-storage content lands around $8,000-15,000 per 600kW rack (proportionately less for 200-350kW racks on EDLC-class solutions, ~$2-5k).
 * Rack volumes. AI rack-scale system shipments across NVIDIA and the AMD/custom-silicon ecosystem plausibly run 60-90k units in 2027, growing 30-50% annually in the base case, with the high-density (HSC-relevant) share rising from perhaps a third in 2027 toward the large majority by 2030.
 * Attach rate. Near-100% for 800VDC-native racks (the shelf is in the reference design); partial for AC-architecture racks retrofitting peak absorption; plus a retrofit stream into the 2025-26 installed base and sidecar/facility-level systems on top.


Multiply through and you get the scenario band below: roughly $0.8-1.4B in 2027, $3-6B by 2030 in the base case, with a bear case (AI capex air-pocket, densities plateau at 350kW, software smoothing over-delivers) still clearing ~$1.5B by 2030 and a bull case (1MW racks on schedule, retrofit wave, grid-compliance mandates spreading) approaching $8-9B. For calibration: the base case implies datacenter alone becomes comparable to the entire pre-AI supercapacitor category within four years, which is exactly what the 20x storage step-up and the supply gap are already telling you.




Bottom-up TAM: datacenter capacitive energy storage systems
Installed system value, USD billions, 2026-2030E: author's model, assumptions in text

Author's model: content/rack × rack shipments × attach rate + retrofit & sidecar systems. Illustrative; every input stated above is adjustable. Excludes campus BESS and Li-ion BBU.


TAM growth decomposition, 2026 → 2030 base case
Waterfall: what turns a $0.6B market into a $4.5B market, base case, USD billions

Starting from 2026 shipping value, the drivers are: 2027 rack shipment ramp (60-90k units), attach rate rising from ~33% toward 80% as 800VDC becomes default, content per rack rising as densities move to 600kW+ and 1MW, and cost/ASP compression as suppliers scale (subtracted). Net = 2030 base case of ~$4.5B. Full workings in the paragraph above.






Profit pools: where the margin sits, and where it migrates


The scarcity rent of 2026-27 is not distributed evenly along the chain, and its distribution will invert as the market normalizes. The current stack, in indicative gross-margin terms: advanced materials (qualified activated carbon, separators, pre-doping lithium) at 30-45% and rising, chokepoint suppliers price against scarcity, not cost; cells at 35-50% in shortage conditions, the widest pool today, and the reason cell capacity is the M&A prize; modules at 20-30%; and shelves/integrated systems at 15-25%, respectable but structurally the most exposed layer. The mechanism of migration is standardization: once the OCP energy-shelf specification lands (targeted from YE2026) and second sources qualify, the shelf becomes a spec-built commodity the way ORv3 power shelves did. Integration margin compresses toward low teens by 2029, while qualified cell and materials margin holds far better because qualification, process IP (dry electrode, pre-doping), and materials chokepoints don't commoditize on the same clock. The investable conclusion: own the layers protected by qualification and process IP; rent the layers protected only by being early.




Margin migration along the HSC value chain
Indicative gross margin bands by layer, 2026 (shortage) vs. 2029E (post-standardization)

Author's estimates from component-industry analogues (Li-ion materials, ORv3 power shelf commoditization precedent). Indicative midpoints of stated ranges.





The ramp timeline


 * CY2026 (now): qualification year. Vera Rubin ramps with 20x storage; Vertiv and peers release 800VDC portfolios in H2; HSC supply gap sets pricing; EDLC carries deployments as the bridge.
 * CY2027: the inflection. Kyber/Rubin Ultra at 600kW on 800VDC; energy shelves at volume; Panasonic FY2027 shipments land; first large HSC fleet deployments at neoclouds and one or two hyperscalers. Supply still gates.
 * CY2028-29: the steep part. 800VDC default above ~300kW; multi-vendor qualified HSC base; retrofit of 2025-26 builds; solid-state transformers begin hollowing out the sidecar's conversion role, while making the capacitor shelf more necessary (an SST has none of a line-frequency transformer's forgiving thermal mass; its stiffness must be synthesized, largely by capacitive storage).




Technology disruption watch list



VectorWhat it isWhoImpact if it lands

Silicon-anode HSCReplacing graphite/hard-carbon LIC anodes with Si-composite, pushing hybrid cells toward 40-50+ Wh/kgCell-maker R&D labs; Li-ion Si-anode ecosystem (Group14, Sila) spilloverWidens HSC's lead over EDLC; starts eating the low end of the Li-ion BBU niche
Dry-electrode processingSolvent-free electrode manufacturing, lower cost, higher throughput, better thick electrodesLICAP (independent IP), Tesla (ex-Maxwell), Musashi process roadmapThe main cost-down lever for closing the supply gap profitably
Curved graphene / advanced carbonsSynthetic high-surface-area carbons replacing coconut-shell activated carbonSkeleton Technologies; MOF-derived carbon researchAttacks the Kuraray chokepoint; EU supply-sovereignty angle
Sodium-ion hybrid capacitorsNa pre-doping instead of Li, cheaper, no lithium exposureChinese cell makers, academic-to-commercial pipelineCost disruptor at the value end; qualification lag protects incumbents near-term
Ultra-high-cap MLCC & embedded47µF/0402-class parts; MLCCs embedded in substratesMurata, Taiyo Yuden, SEMCOConverts electrolytic/tantalum sockets; deepens top-3 moat
BSPDN + silicon capsBackside power delivery structurally designing capacitance into the packageTSMC, Intel, SEMCO, Murata/IPDiAThe one genuine structural design-in event in the stack; 2027-29
OCP energy-shelf standardizationOpen specs for capacitive shelvesOCP / ODCA working groupsDouble-edged: accelerates adoption, then commoditizes the shelf, value migrates to cells and materials






The M&A chessboard


Scarce, qualified, strategically located capacity in a 2-3x undersupplied category is a textbook acquisition setup. The board as I read it:



AssetSituationLogical acquirersLogic & friction

Musashi Energy SolutionsThe crown jewel, LIC leadership inside Musashi Seimitsu, a mid-cap Japanese auto-parts group (Honda-affiliated heritage) whose core business is transmissions and gears. A strategic mismatch between the parent's valuation and the subsidiary's franchise.Panasonic (consolidate Japanese leadership); Eaton / Vertiv / Schneider / Delta (vertical integration into the 800V power chain); Flex (deepen the CESS position); TDK or Murata (passives adjacency); PE carve-out with a hyperscaler anchor offtakeHighest strategic value in the space. Friction: Japanese governance, Honda-ecosystem sensitivities, and the parent's ability to fund the ramp itself. A minority-stake-plus-offtake structure is the likeliest first move, watch for exactly that.
Skeleton TechnologiesEU champion, curved-graphene differentiation, Siemens/Marubeni among backers; capital-hungry for expansionSiemens (deepen), a European power major (ABB, Schneider, Legrand), or IPO; EU sovereignty funds as anchorsThe European answer to supply-chain localization mandates; valuation will price the sovereignty premium
LICAP TechnologiesDry-electrode process IP, US-based, sub-scaleAny cell maker needing cost-down IP; US onshoring buyers (IRA-adjacent logic)Classic technology bolt-on; small check, outsized process value
UCAP PowerMaxwell lineage, industrial niches, sub-scale for the AI waveConsolidation into a larger industrial (Eaton-style) portfolio or PE roll-up of EDLC tailBrand and installed base worth more inside a bigger channel than standalone
Kuraray (activated carbon franchise)Not for sale, but the materials chokepoint of the entire categoryN/A as target; relevant as the partner every cell maker must secureLong-term supply agreements with Kuraray are diligence item #1 on any cell-maker deal
Silicon capacitor startups (Empower et al.)Package-level power delivery IP in the BSPDN design-in windowFoundry ecosystem, OSATs, SEMCO/Murata competitors needing a positionThe 2027 version of today's HSC scramble, buy before the design-ins lock





Valuation framing: what would you pay?


Three anchors. Precedent: Tesla paid ~$218M for all of Maxwell Technologies in 2019, the category's then-leader, acquired for its battery IP with the ultracap business thrown in and divested two years later for an undisclosed (small) sum. That deal priced the ultracapacitor franchise at close to zero, five years before the largest demand shock in its history. It is the benchmark every corporate development team currently "monitoring the space" should study. Scarcity heuristic: in a 2-3x undersupplied market, qualified capacity is worth a multiple of its revenue-generating potential, if a datacenter-grade LIC cell carries $50-100 of revenue and the gap persists through 2028, an EV of 2-4x forward capacity-revenue for a qualified line is defensible before any terminal-value argument, implying the leading franchise is a multi-billion-dollar asset on its ramp trajectory. Carve-out arbitrage: auto-parts parents trade at 5-8x EBITDA; scarce energy-storage pure-plays in a structural shortage command mid-teens to 20x+. The same asset is worth 2-3x more outside the parent than inside it, which is precisely the pressure that produces minority-stake-plus-offtake deals first, and control transactions when the parent needs ramp capital it cannot fund. On the private side, Skeleton's cumulative funding (north of €300M across rounds, with Siemens and Marubeni among backers) sets the reference for what EU-sovereignty-positioned capacity raises at; an IPO or strategic sale prices the sovereignty premium explicitly. None of this is a price target. It is the frame a buyer's IC will actually use.



Scarce, qualified capacity in a 2-3x undersupplied category, held inside a mid-cap auto-parts parent, Musashi is the most obvious strategic asset in the entire AI power chain.The M&A thesis in one line




Trade-offs and who wins


The trade-off map:


 * HSC vs. EDLC: energy density and voltage vs. availability, multi-sourcing, and forty years of field data. Through 2027, EDLC wins deployments it "shouldn't" purely on supply. The substitute for an HSC isn't a better device, it's an uglier system architecture (EDLC + Li-ion hybrids recreating the HSC across two supply chains), which is why HSC wins the 2028+ steady state.
 * Capacitors vs. software: NVIDIA's power-cap/ramp-control/GPU-burn stack shaves the transient envelope and could keep sub-300kW racks on cheaper storage. Above that, physics and grid compliance need hardware. The bear case for the whole stack is a software-plus-electrolytics world at lower-than-expected rack densities, worth stress-testing in any model.
 * Pricing power vs. demand destruction: MLCC and supercap price increases are rational oligopoly behavior, but H2 2026 projections of +30-50% on ultra-high-cap grades will accelerate design-outs (silicon caps, embedded solutions, electrolytic conversion in reverse) at the margin.
 * Winners: Musashi (if it can fund the ramp), Panasonic, the MLCC top three, SEMCO's full-stack silicon-cap position, Flex and the sidecar OEMs, and the quiet materials chokepoints (Kuraray, Nippon Kodoshi, Sumitomo Metal Mining). Pressured: the EDLC tail, UCAP-class sub-scale players, Li-ion BBU in white space (fire-code plus cycle-life headwinds), and any capacitor franchise still selling on 2024 price lists.
 * The demand-side risk that dominates all of it: an AI capex air-pocket shifts every curve right. The attach-rate thesis (storage per MW rises with density) survives a correction; the 2027 volume estimates may not.




Where each technology lives: the energy. Power map
Indicative specific power vs. specific energy by technology family (log. Log)

Indicative Ragone positioning, author's compilation from vendor datasheets and literature. HSC/LIC occupies the gap AI created demand for.





Standards and regulation: certification timing is adoption timing


The unglamorous layer that actually sets the ramp slope. The instruments that matter: UL 810A (electrochemical capacitors, the certification that lets an HSC shelf into a North American data hall; Musashi's UL-certified cells are a genuine moat); the IEC 62391/62576/62813 family (EDLC and LIC test methods, the common language of qualification); NFPA 855 (stationary energy storage, the fire-code regime whose Li-ion burdens are the supercapacitor's quiet sales argument, and whose treatment of large capacitive installations is still being clarified, a watch item, since an adverse interpretation is a tail risk to the white-space deployment model); the OCP energy-shelf and 800VDC standards (initial specifications targeted from YE2026, the commoditization clock of the profit-pool section); and DC-side electrical codes (IEC 60364 DC provisions, UL DC-rated protection), where standards genuinely lag the hardware and where an incident during the 2027 ramp would be the sector's biggest self-inflicted risk. Two supply-side regulatory vectors: the EU Critical Raw Materials Act lists natural graphite and lithium, both LIC anode-side inputs, creating both compliance burden and a subsidy case for EU cell capacity (Skeleton's tailwind); and US/Japanese onshoring incentives (IRA-adjacent programs, METI support) that shape where the next tranche of capacity lands.




Sustainability: the ESG case writes itself, mostly


Increasingly a gating item in both hyperscaler procurement and PE investment committees, and the capacitor stack scores unusually well. HSCs and EDLCs contain no cobalt and no nickel cathode, deliver 15+ year service lives against 3-7 for cycled batteries (less replacement, less embodied carbon per service-year), present minimal fire risk (a facility-insurance and water-usage argument as well as a safety one), and decompose at end-of-life into aluminium, copper, and carbon, a far simpler recycling problem than Li-ion's black mass. The honest caveats: LICs do consume lithium (pre-doping) and graphite (anode), modest quantities per cell, but the CRMA exposure above is real; EDLC electrolytes use acetonitrile, a toxic solvent demanding proper handling and end-of-life control; and the cathode's coconut-shell feedstock concentrates in Sri Lanka, the Philippines, and Indonesia, a supply chain whose traceability and agricultural ethics will attract scrutiny as volumes scale (and whose geographic concentration is itself a minor resilience flag). Net: a genuine ESG differentiator versus batteries, provided the industry gets ahead of the solvent and feedstock stories rather than waiting to be asked.




Where the cells are actually made



RegionCapacity & playersDynamic

JapanThe center of gravity: Musashi (LIC, new plant ramping), Panasonic (FY2027 datacenter line), Nichicon, plus the materials complex (Kuraray, Kureha, Nippon Kodoshi, Honjo, Toyo Aluminium)METI-supported; the qualification incumbency; the M&A hunting ground
KoreaLS Materials (ultracaps), Power Carbon Technology (cathode carbon), SEMCO (silicon caps/MLCC), Samsung SDI (BBU side)Fast follower with hyperscaler-adjacent channels via the chaebol electronics complex
EuropeSkeleton (Estonia; Leipzig/Markranstädt expansion), SPEL-EU distribution, Eaton modulesCRMA-subsidized sovereignty play; capacity small but strategically priced
United StatesLICAP (dry electrode), UCAP Power, KYOCERA-AVX linesSub-scale in cells; strong in integration (Flex, Vertiv, Eaton), onshoring incentives could change the cell map by 2028
ChinaJianghai, Aowei, CRRC ecosystem, Long Sing, scaling LIC entrantsLargest capacity trajectory; fenced out of Western hyperscale by qualification and provenance (§China)






The public-market exposure map


For analyst readers: the listed vehicles that touch each layer. Mapping only, nothing here is a recommendation, and purity varies enormously (most of these are diversified groups where the capacitor-stack exposure is one division or less).



LayerListed names (ticker)Exposure purity

HSC / supercap cellsMusashi Seimitsu (7220.T, parent of Musashi ES), Nichicon (6996.T), LS Materials (417200.KQ)Low-to-mid: buried in auto parts / diversified capacitors; the purity gap is the carve-out thesis
HSC materialsKuraray (3405.T), Kureha (4023.T), Nippon Kodoshi (3891.T), Resonac (4004.T), Toyobo (3101.T)Low purity, high chokepoint quality, Nippon Kodoshi the most concentrated of the set
MLCCMurata (6981.T), Samsung Electro-Mechanics (009150.KS), Taiyo Yuden (6976.T), TDK (6762.T), Yageo (2327.TW)Mid-to-high: MLCC is core P&L; the cleanest listed expression of the shortage economy
MLCC materialsSumitomo Metal Mining (5713.T), Sakai Chemical (4078.T)Low purity, real leverage to the high-cap ramp
Silicon capsSEMCO (009150.KS), Murata (6981.T), TSMC (2330.TW/TSM)Small today; SEMCO's $1B contract is the first line item that moves a model
Shelves / systems / 800VDC chainFlex (FLEX), Vertiv (VRT), Eaton (ETN), Delta (2308.TW), Schneider (SU.PA), nVent (NVT)Mid: the integration layer, strongest near-term revenue visibility, weakest long-term margin defense per the profit-pool argument
Displaced incumbentsNippon Chemi-Con (6997.T), Rubycon (private), Nichicon (6996.T)The cyclical-peak-inside-secular-handover situation flagged above






The dashboard: ten leading indicators to watch


 1.  Kuraray capacitor-carbon capacity announcements, the materials layer moves first; carbon capex signals cell-maker conviction one step upstream.
 2.  Musashi plant ramp signals (utilization, hiring, parent capex disclosures in 7220.T filings), the supply gap's closing speed, readable quarterly.
 3.  OCP energy-shelf specification publication, starts the commoditization clock on integration margin.
 4.  First hyperscaler dual-source HSC qualification, ends single-supplier pricing power; watch Panasonic's FY2027 customer disclosures.
 5.  H2 2026 MLCC price realization, whether the projected +20-30% on high-cap grades sticks tells you if oligopoly discipline holds into 2027.
 6.  First non-Chinese cloud deploying Chinese LIC cells, the price-umbrella tear (§China).
 7.  BSPDN accelerator tape-outs with embedded silicon caps, converts the silicon-cap thesis from optionality to backlog.
 8.  NFPA/AHJ treatment of large capacitive installations, the regulatory tail risk to white-space deployment.
 9.  A Musashi minority-stake or offtake announcement, the predicted first move on the M&A board; its counterparty tells you which acquirer archetype is winning.
 10. Rack-density reality vs. roadmap, actual 2027 Kyber deployment counts against plan; the whole stack's volume assumption in one number.



V→VThe conversion ladder



From 800V bus to GPU pin: the parasitic budget


An 800VDC data centre delivers power to compute silicon via a specific voltage cascade, and every step of that cascade requires capacitance sized to a particular set of parasitic constraints. Each layer of the capacitor stack maps to a specific stage; each has its own ESR (series resistance), ESL (series inductance), and bandwidth requirement.


 1. 800V bus at rack inlet. DC-link capacitance provides bulk buffering between grid rectification transients and the downstream converters. Bulk Al electrolytic (moving toward polymer as ripple frequencies rise) or film for HV rating. Required ESR: low-milliohm range at 100Hz. 10kHz. Required ESL: not critical at these frequencies. Cable and busbar inductance dominates.
 2. 800V → 48V (or 12V) intermediate bus converter. An LLC resonant converter or a DAB stage steps down. Bulk capacitance either side (polymer Al or hybrid polymer at the tens-of-µF scale) plus MLCC arrays at input and output for switching-frequency noise. ESR: sub-milliohm on the primary output. ESL: below 1nH for the MLCC bank feeding a ~500kHz switching stage.
 3. 48V → 12V board-level DC-DC. Multi-phase buck converters on the accelerator PCB. Input MLCC arrays (hundreds of parts) handle switching noise; output caps (more MLCC arrays plus polymer bulk) handle load transients. ESR: hundreds of microohms per MLCC in parallel. ESL: dominated by PCB layout inductance, 100pH per capacitor and tens of pH aggregate.
 4. 12V → 0.75V VRM at point-of-load. Multi-phase buck at 1MHz+ delivering thousands of amps to the accelerator package. Input caps decouple the VRM. Output caps have to absorb load transients on the microsecond scale at hundreds of amps: the fastest and hardest stage of the whole cascade. Requirements: sub-microohm aggregate ESR, sub-picohenry aggregate ESL. Achieved only by paralleling hundreds of MLCCs across multiple case sizes, and increasingly by supplementing with silicon capacitors embedded in the package substrate.
 5. Under-die decoupling. The final sub-nanosecond transient at the transistor. Handled by silicon capacitors (deep-trench structures in the package substrate) and, on the die itself, on-die metal-oxide-metal (MoM) or metal-insulator-metal (MIM) capacitors. Loop inductance is the ceiling. Anything with a physical wire in the loop, even a µm-scale bond wire, has more inductance than the transient can tolerate.




Loop inductance is the ceiling on speed
Total loop ESL from decoupling capacitor to transistor, by mounting location, log scale

Order-of-magnitude decreases in loop inductance as decoupling moves from board (10s of nH) to package substrate (100s of pH) to under-die embedded (sub-pH). Each step lets the transistor tolerate a faster load transient. This is the mechanism that makes silicon capacitors and BSPDN structurally necessary at the next accelerator generation, rather than optional.


Why MLCC arrays do not scale to sub-nanosecond decoupling. A single MLCC has 300-800pH of package ESL. Paralleling 1,000 parts in perfect PCB layout drops aggregate ESL by roughly √N (mutual coupling reduces the ideal 1/N scaling) to a few pH. That is adequate for hundreds-of-picoseconds transients but falls behind at the tens-of-picoseconds transients next-generation accelerators produce. Silicon capacitors, built with sub-100pH package ESL and mounted directly on the substrate, close the last order of magnitude.

HEATThe lifetime constraint · Arrhenius



Why the energy shelf lives in the sidecar


Every capacitor category on the AI power chain ages, and every one of them ages faster in heat. Aluminium electrolytics dry out. EDLC electrolytes decompose. HSC anodes lose lithium. MLCC dielectrics recrystallise. The rate follows the Arrhenius equation, which for a typical HSC or EDLC translates to a rule-of-thumb every operator recognises: lifetime halves for every 10°C rise in operating temperatureA 100,000-hour EDLC rated at 25°C degrades to 50,000 hours at 35°C, 25,000 hours at 45°C, and 12,500 hours (about 14 months) at 55°C.


That arithmetic decides where the energy shelf physically sits. A hot-aisle-adjacent placement, running at 40-50°C ambient in a densely-packed rack, cuts predicted service life by 4-8× against a datasheet-rated 25°C service life. Hyperscaler procurement teams underwriting 10-15 year deployments do not accept the arithmetic. The universal answer, visible in every 800VDC reference architecture published from 2025 onward, is sidecar placement: the energy shelf sits in a separate cabinet with dedicated liquid or air cooling, physically decoupled from the compute rack's thermal envelope. HSCs and EDLCs get to run at 25-35°C ambient; predicted lifetime lands where the datasheet claims it does.




Arrhenius lifetime scaling · why placement matters
Predicted service life vs operating temperature, log axis, 100,000-hour rated part at 25°C

L = L₀ · 2^((T₀−T)/10), the standard capacitor Arrhenius rule of thumb. A part rated 100,000 hours at 25°C runs 50,000 hours at 35°C, 25,000 hours at 45°C, 12,500 hours at 55°C. Real behaviour deviates at extremes but the 10-degree halving is the design heuristic every operator uses. Different chemistries have different absolute rated lives but the halving slope is broadly consistent.


The 15-year service claim. Any HSC or EDLC vendor pitching a 15-year service life is implicitly pitching a 25-30°C ambient operating temperature. Underwrite that claim under sidecar-cooled conditions only. In-rack placement adjacent to the hot aisle, or in retrofit deployments without dedicated cooling, cuts service life to a fraction of the datasheet number, and refresh cycles become a substantial fraction of the total cost of ownership.

$/JAdoption economics



Cost per joule, density per cm³, and the buyer's fight to delay


A hyperscale architect specifying a capacitor solution runs two numbers simultaneously: cost per joule of usable energy stored and joules per cubic centimetre of white space consumedThe trade between them decides what technology wins each socket, and the trade looks different for every layer of the stack.




Cost per joule vs energy density · the buyer's plane
USD per stored joule vs joules per cm³ of installed volume. Log axes. Volume includes cell packaging.

Author's estimates from vendor pricing (2025 spot), datasheet volumetric energy density, and hyperscale-scale procurement discussions through mid-2026. MLCC and silicon capacitor $/J are extreme, but the sockets they serve do not demand energy: they demand response speed. HSC and Li-ion dominate $/J; MLCC and silicon capacitors dominate response time. No single technology wins the whole plane.




Two structural implications follow. First, adoption is a fight to delay, not a race to adoptEvery capacitor architecture more advanced than a standard MLCC or Al electrolytic carries a cost premium the buyer pays only when board space, thermal budget or yield loss force the hand. Silicon capacitors under a GPU exist because thousands of paralleled MLCCs still cannot get die impedance low enough. HSC energy shelves exist because Li-ion cycle life collapses under training-workload duty. The premium is the price of removing a physical constraint the cheaper alternative cannot remove.


Second, the two axes of the buyer's plane trade against each other differently by socketIn-package decoupling has no meaningful cost sensitivity (a $50 silicon capacitor sits next to a $40,000 GPU) and extreme response-time sensitivity. Rack-level bulk storage is exactly opposite: the socket is cost-sensitive at the $/J level (a 500Wh energy shelf at $15/Wh is $7,500 per rack, a real BOM line) and response-time-insensitive above about 100ms. The technology that wins each socket is determined by which axis the buyer is willing to accept as dominant.



Hyperscalers do not adopt premium capacitor technologies because architects prefer them. They adopt because a cheaper alternative has hit a physical constraint the premium technology sidesteps. Every category on this stack has a "why couldn't we just use MLCCs, Al electrolytics or Li-ion" answer, and that answer is where the value sits.The adoption thesis in one line

REFComparison matrix



The stack, side-by-side


Every capacitor category discussed above, on the six variables an architect actually cares about, in one table.




CategoryEnergy
(Wh/kg)Power
(W/kg)ResponseT_max
(°C)$ / JSupply-chain risk

Silicon capacitor0.0001-0.00110⁵+sub-ns150very high per JVery concentrated; qualification-gated
MLCC (ceramic)0.001-0.0510⁵. 10⁶ns125very high per JTop 5 ~85%; AI-driven shortage
Film0.05-0.155,000-15,000ns. µs105mediumModerate; specialty for HV DC-link
Al electrolytic / polymer0.05-0.310⁴. 10⁵µs. Ms105low per JFragmented; polymer tight, being displaced up-shelf
EDLC4-82,000-15,000ms. S65medium~5 major cell makers; the bridge into HSC
HSC / LIC15-302,000-10,00010ms. 10s65medium-highVery concentrated (Musashi dominant); 2026 shortage
Li-ion (energy)180-300100-1,000s. Min60low per JScaled global supply; thermal / regulatory ceiling in white space






Where this lands


The 800VDC transition converted a reliability accessory into a platform component, and in doing so created four simultaneous supply crunches at four different layers of the same power chain. The capacitor stack is now where the memory market was in the early cloud era: structurally undersupplied, consolidating around a few names with real process moats, with materials chokepoints one level down that most of the market hasn't priced, and with an M&A window that will not stay open past the point where hyperscaler qualification locks the supplier base. For operators: design the buffering architecture now, and secure qualified supply early. For investors: qualified capacity is where the pricing power sits, and the qualification cycle is measured in years, not quarters.



ANNEXSources · Calculations · Assumptions



Annex: how every derived number was built


Every figure in this piece is either published (traceable to a named source), derived (calculated from published inputs, with the arithmetic shown), or estimated (my judgement, flagged as such). This annex covers the latter two, and matches the sourcing discipline of Parts II, III and IV.




§A1. Content per rack and the 20× storage step-up


Status: published + derived. Published: NVIDIA's Vera Rubin platform disclosed a roughly 20× step-up in integrated energy storage vs the preceding GB300 generation; capacitor-based power smoothing has been demonstrated as native to the 2027 Kyber/800VDC reference architecture. Derived: capacitive buffer size per 600kW rack ≈ 250-500Wh of usable energy, based on target smoothing of a several-hundred-kW swing over seconds. At volume LIC cell economics of $8-15/Wh, plus system multiple 1.8-2.2x for shelf electronics, controls and integration, gives ~$8-15k installed content per 600kW rack (as used in the TAM build below). Sources: NVIDIA GTC 2025/2026 disclosures, OCP Q1 2026 power distribution whitepaper.




§A2. Bottom-up TAM model


Status: derived. Full workings.


Inputs. Content per rack: $8-15k per 600kW rack (§A1). AI rack shipments: 60-90k units in 2027 across the NVIDIA and AMD/custom-silicon ecosystem, growing 30-50% annually in the base case. High-density (HSC-relevant) share of rack shipments: ~33% in 2027, rising toward majority by 2030 as 800VDC becomes default at higher power classes. Attach rate: near 100% for 800VDC-native racks; partial for retrofit; plus a facility-side stream.


Base case. ~$0.6B shipping in 2026, rising to $0.8-1.4B in 2027 as the Kyber generation ramps, then $3-6B by 2030 as 1MW racks and retrofit demand compound.


Bear case. AI capex air-pocket cuts rack shipments 40%, densities plateau at 350kW so HSC-attach rate stalls, software smoothing over-delivers on transient absorption. Result: ~$1.5B by 2030.


Bull case. 1MW racks on schedule, retrofit wave into the 2025-26 installed base, grid-compliance mandates spreading. Result: ~$8-9B by 2030.


Sensitivities in order. (1) AI rack shipment volumes and density mix; (2) attach rate on retrofit vs new-build; (3) $/Wh cell economics as the supply gap closes.




§A3. Market share and concentration estimates


Status: estimated where quoted. HSC/LIC cell market: Musashi ES estimated dominant share (widely reported ~50%+ of qualified capacity, exact number not published). MLCC market: Murata ~31%, Samsung EM ~20%, Taiyo Yuden ~13%, TDK ~10%, Yageo/Kemet ~9%, others ~17% (TrendForce/Yole compilations, 2024-25 basis, positions moved somewhat since). Al electrolytic market: Nippon Chemi-Con, Rubycon, Nichicon, Panasonic remain the top four, exact shares vary by source. Silicon capacitor market: Murata (via IPDiA), SEMCO, TSMC advanced packaging teams with a very small revenue base, structural shares not yet meaningful.




§A4. Profit pool bands


Status: estimated. No source publishes gross margin by capacitor value-chain layer. Bands used in this piece (advanced materials 30-45%, cells 35-50%, modules 20-30%, shelves/systems 15-25% today, migrating toward services / cells / materials as OCP standardises) are the author's judgement, built from company-reported segment margins where disclosed, the observed supply/demand imbalance across the layers, and industrial analogues from the Li-ion materials chain where scarce upstream commoditised faster than qualified midstream. Directional, not audited.




§A5. Arrhenius lifetime scaling


Status: published rule of thumb. The 10°C halving heuristic (L = L₀ · 2^((T₀−T)/10)) is the standard capacitor lifetime model published across manufacturer application notes for Al electrolytic, EDLC and HSC categories (Nichicon, Panasonic, Musashi, Skeleton, Nippon Chemi-Con, Kemet). Absolute rated life at reference temperature varies (5,000h for a mid-tier Al electrolytic; 50,000-100,000h for a supercap at rated V; higher for a de-rated part), but the slope is broadly consistent. Real behaviour deviates at extremes: below 0°C ion mobility collapses, above 85°C the electrolyte itself degrades non-Arrhenius. Use the halving inside the middle of the operating band.




§A6. Cost per joule estimates


Status: estimated from vendor list pricing and procurement observations. The $/J values plotted are order-of-magnitude illustrations, not procurement quotes. Silicon capacitor: ~$5-50 per joule (device pricing $10-50, energy per device sub-mJ). MLCC: ~$1-20 per joule (device pricing $0.001-1, energy per device µJ range). Film: ~$0.05-0.5 per joule. Al electrolytic / polymer: ~$0.01-0.1 per joule. EDLC: ~$0.005-0.02 per joule. HSC/LIC: ~$0.003-0.008 per joule (cell-level at $8-15/Wh translates to ~$2-4/kJ). Li-ion: ~$0.0005-0.002 per joule at $100-200/kWh cell pricing. Every band spans an order of magnitude; treat the chart as positional.




§A7. Kuraray and materials chokepoint


Status: published position, estimated share. Kuraray publicly discloses activated carbon leadership for supercapacitor cathodes in its capacitor-carbon business unit. Precise global share of AI-qualified cathode carbon is not published; industry estimates place Kuraray at leading share, with Nippon Kodoshi holding a comparable position in separator paper and Sumitomo Metal Mining in MLCC nickel paste. These are the "unpriced chokepoints" argument: narrower than the cell manufacturer concentration, structurally similar to the AMB Si₃N₄ layer in the wide-bandgap essay.




§A8. Other published figures used


 * NVIDIA Vera Rubin storage step-up: ~20× vs GB300; NVIDIA disclosures (GTC 2025).
 * GB300 power smoothing: capacitor-based smoothing built into the power shelf, cutting peak grid demand up to ~30%; NVIDIA disclosures + trade press.
 * Kyber rack: ~600kW, 800VDC at rack inlet, ~750A busbar; NVIDIA + OCP Q1 2026 whitepaper.
 * Rack power trajectory: GB200 ~120kW → GB300 ~145kW → VR NVL144 ~200kW → Kyber ~600kW → 1MW class 2028-29E; NVIDIA roadmap disclosures.
 * 2026 AI-server supercapacitor demand vs supply: ~2-3x demand/supply, ~15-18M units of AI demand vs ~6.5M unit Musashi capacity; industry compilations, 2025-26.
 * MLCC price increases: +15-35% at Murata effective April 2026, peers followed at various magnitudes; Murata, Taiyo Yuden, SEMCO, Yageo/Kemet company communications.
 * Musashi ES ownership structure: LIC business inside Musashi Seimitsu (Honda-affiliated mid-cap auto-parts group); public filings.
 * Maxwell / Tesla 2019: Tesla acquired Maxwell Technologies for ~$218M in 2019; Tesla SEC filings.



A general caution on this sector's data. Research houses disagree by factors of two to four on capacitive-storage TAM (see §A2 spread), and unit-level pricing is opaque. Where I found a spread I have shown it; where a figure rests on one source I have said so; where I filled a gap with judgement I have labelled it an estimate. This annex is a starting map for diligence, not a substitute for primary work.




Related in the AI Power Chain


For readers who want the physics, cross-sections and failure modes behind every capacitor category discussed here, see the technical companion: The Capacitor Stack: Technical Companion.


Later parts of the series: The Wide-Bandgap Stack (Part II) covers SiC and GaN power semiconductors; its technical companion is The Wide-Bandgap Stack: Technical CompanionThe Thermal Stack (Part III) covers liquid cooling, with The Thermal Stack: Technical Companion for the physics. The Interconnect Stack (Part IV) covers transformers, switchgear, grain-oriented electrical steel and the grid equipment lead times that now gate the AI buildout.




Methodology. This analysis is independent and based entirely on public information. Market sizes and shares are approximate, compiled from public research (Mordor Intelligence, IDTechEx, MarketsandMarkets, Fortune Business Insights, Astute Analytica, Dataintelo, Verified Market Reports, Intel Market Research), company disclosures, OCP publications, and trade press (Jan. Jul 2026); where estimates diverge, ranges are shown and scope differences noted. Figures marked directional or indicative should be treated as such. Nothing here is investment advice.


Comments and disagreements welcome, that's what the piece is for.












The AI Power Chain: series companions

 * The AI Power Chain: Vendor Screen. The vendor map across every layer
 * Pricing Under Scarcity. Where premium capture is compounding
 * The Services Inversion. Why services now command product-like multiples







Frequently asked


What is a hybrid supercapacitor?

A hybrid supercapacitor (HSC) combines electric double-layer capacitance with battery-like faradaic reactions in a single cell. Energy density is higher than EDLCs (up to 20 Wh/kg) with faster response than lithium-ion. In AI data centres, HSCs handle the millisecond-to-second energy buffer for rack-level ride-through where lithium batteries respond too slowly.


Why do MLCCs matter for AI data centre power?

Multi-layer ceramic capacitors (MLCCs) sit directly on GPU packages for high-frequency decoupling. Modern AI accelerators require thousands of MLCCs per package to maintain voltage stability under fast switching loads. MLCC supply concentration (top 5 vendors control >70% of high-cap-density market) creates rack-level supply risk.


==============================================================================
# The Wide-Bandgap Stack: Boom, Bust, and the Rebuild of AI's Power Semiconductor Chain
URL: https://adikumar.co/the-wide-bandgap-stack/
Published: 2026-08-01
Summary: Wide-bandgap semiconductors — SiC and GaN — enable 800V DC data centres. Boom, ~60% reported price collapse, Wolfspeed Ch.11, and the AI-era rebuild.
==============================================================================

The AI Power Chain series · Part 5 of 15








Glossary of terms used



AMB
Active Metal Brazed. A ceramic substrate technology used in high-power semiconductor packaging.
GaN
Gallium Nitride. Wide-bandgap semiconductor material used in RF and power applications.
SiC
Silicon Carbide. Wide-bandgap semiconductor material used in high-voltage power electronics.
TAM
Total Addressable Market. The maximum revenue opportunity available if a product served every potential customer segment.
WBG
Wide-Bandgap. Semiconductor material class (SiC and GaN) used in high-efficiency power electronics.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.









ADI KUMAR · POWER & DIGITAL INFRASTRUCTUREJULY 2026 · V2 · ~50 MIN READ

The AI Power Chain · Part II of IV
Market deep dive · Wide-bandgap semiconductors



The Wide-Bandgap Stack: Boom, Bust, and the Rebuild of AI's Power Semiconductor Chain


SiC and GaN are the enabling materials of the 800VDC data center. Their industry has just come through a capacity glut, reported price declines exceeding 60% across various SiC substrate + epi-wafer categories (specific series vary by source and vintage; not a clean like-for-like index), and the Chapter 11 filing of Wolfspeed, its founding US pure-play. This piece works down the chain from crystal boule to gate driver: who makes what, where the chokepoints sit, why the demand boom and the bust arrived at the same time, and what the rebuild means for capital.



By Adi Kumar
Independent analysis
Views are my own



Executive summary. Nine findings.

 1. The bust-boom overlap. Wolfspeed emerged from Chapter 11 in September 2025 having cut ~$4.6B of debt (~70%). Over the same period its AI data center revenue doubled in three quarters. The bust was automotive; the boom is data center, and they overlapped.
 2. Prices collapsed, not demand. 6-inch SiC epi wafers fell from $600+ (early 2023) to ~$400 by 2025. Chinese 6-inch spot was reported near $200 in mid-2026 on a single trade-press source, not corroborated in Western procurement quotes at time of writing. 8-inch declines exceeded 60% across multiple sources. Chinese substrate capacity went from 460k units (2022) toward a projected 7.7M by 2027.
 3. AI is re-rating the whole chain. Infineon guides AI-related revenue from €1.5B (FY26) to €2.5B (FY27) and sizes the AI power opportunity at ~€12B by 2030. onsemi says content per 1MW AI rack doubled from ~$50k to ~$100k.
 4. Content escalation is the core investable fact. Power component content per rack rises roughly 4x from GB200 to Vera Rubin and ~11.5x to Rubin Ultra. That is a BOM restructuring, not a cyclical uptick.
 5. SiC and GaN divide by physics. SiC owns high-voltage grid-to-rack conversion. GaN owns high-frequency intermediate and point-of-load stages. Navitas' 800V-to-6V board that eliminates the 48V intermediate bus is the clearest signal of where GaN is heading.
 6. Substrates are commoditizing; packaging is not. The scarcity has migrated downstream to AMB silicon-nitride substrates, sinter pastes, and thermal materials, where Japanese and German suppliers together are estimated to hold ~80-90% share of the AMB substrate + sinter-paste + die-attach layer, with Si₃N₄ >95% of the AMB mix. The specific % is a directional estimate from industry-analyst reporting; scope + year of the underlying survey vary.
 7. Gallium is a Chinese export instrument. China's export controls on gallium put GaN's raw-material base under policy risk that SiC's does not carry. The asymmetry between the two technologies is underpriced.
 8. The M&A window is wide open. A post-restructuring Wolfspeed, sub-scale GaN specialists, and distressed Chinese substrate capacity meet buyers (Infineon's "30-30" ambition, onsemi, Renesas, Samsung) who need capacity and design-ins fast.
 9. Bottom-up TAM. WBG devices in AI data centers at roughly $1.5-2.5B in 2026 scaling to $9-14B by 2030 in the base case. The bull case is gated by qualification, not by silicon.



BOULEStage 0 · Wolfspeed

The AI Power Chain · six essays, one physical arc


The series walks a single physical path. It begins at the medium-voltage utility bus at the site fence, steps down through the substation and switchgear, arrives at the datacenter rack where 800V DC is stabilised by the capacitor stack, is converted by silicon-carbide switches to 48V, is distributed across the rack by copper busbars and whips, is stepped down again by multi-phase controllers on the accelerator board to 0.8V, and finally routed through the on-package power delivery network to a transistor gate drawing over 2,000 amperesWaste heat from every conversion stage is removed by the thermal stack. The whole thing is packaged inside a factory-modular building because there aren't enough electricians to build it stick-frame. Six essays. One 800V → 0.8V staircase.


 1. Part I. The Capacitor Stack 800VDC at the rack
 2. Part II. The Wide-Bandgap Stack SiC and GaN conversion (you are here)
 3. Part III. The Thermal Stack Removing the waste heat
 4. Part IV. The Interconnect Stack Busbars and whips
 5. Part V. The On-Package Delivery Stack 48V to 0.8V
 6. Part VI. The Modular Datacenter Stack How the building gets built



~11.5xPower component content per rack, GB200 to Rubin Ultra
€2.5BInfineon's FY2027 AI-related revenue target, up from €1.5B in FY2026


The price collapse that defined the cycle
Discrete published price observations, USD per 6-inch wafer, 2023-2026. Note: the series mixes epi-wafer and bare-substrate quotes from different sources. It is a set of observations, not a like-for-like index.

[A1] Woodside Capital Partners (Jun 2025) for 2023 epi-wafer points; eeNews Europe (Nov 2025) for 2024-25 substrate points; DigiTimes via industry press (Jun 2026) for the Chinese spot figure (single-source, unconfirmed). See Annex §A1 for the full reconciliation.

MAPThe framework · Eight layers, one chain



The wide-bandgap value chain, layer by layer


The capacitor stack divides by timescaleThe wide-bandgap chain divides by process stepEach step has a different competitive structure, capital intensity, and margin profile, and the AI transition is redistributing power between them. The map below names the players at every layer. Two things to notice before reading it: the extraordinary Japanese depth in materials and packaging, and how thin the qualified supplier base becomes at the ceramic-substrate and sinter-paste layers. Those are the chokepoints most investors miss because they are looking at wafers.



The wide-bandgap power semiconductor market map
Named players by layer · leaders highlighted · July 2026

L1Crystal growth & substratesPVT boule growth, slicing, polishing · 150/200mm, first 300mm demo

Wolfspeed US · ~34%TanKeBlue CN · ~17%SICC CN · ~17%Coherent US · ~14%
SiCrystal (ROHM) DE/JPResonac JPSK Siltron CSS KRSumitomo Electric JP
Sanan Optoelectronics CNSynlight Crystal CNJSG / Jingsheng CNGlobalWafers TW
Soitec (SmartSiC) FRonsemi (ex-GTAT) US · captiveST (ex-Norstel) EU · captive



L2EpitaxyDoped SiC/GaN layers · the yield-critical step · mostly captive, partly merchant

Wolfspeed US · merchantResonac JP · merchant leaderEpiworld CNHantian / Enkris CN
IQE UK · GaN/RFEpisil / EpiCrystal TW/CNShowa Denko lineage JPNexWafe-adjacent EU



L3aSiC device makersMOSFETs, diodes, JFETs · 650V to 3300V · top five ≈ 90% of revenue

STMicroelectronics EU · #1onsemi USInfineon DEWolfspeed USROHM JP
Mitsubishi Electric JPFuji Electric JPToshiba JPBosch DENexperia NL
Microchip USLittelfuse (IXYS) USVishay USNavitas (GeneSiC) US · 2.3/3.3kVQorvo JFET → onsemi US
BYD Semiconductor CNUnited Nova (UNT) CNBASiC Semiconductor CNAccoPower CNSemiQ US



L3bGaN device makers100V. 650V HEMTs and power ICs · top five ≈ 88% · the AI-rack growth engine

Innoscience CN · ~30%Infineon (GaN Systems) DEPower Integrations USNavitas USEPC US
Renesas (Transphorm) JPSTMicroelectronics EUTexas Instruments USonsemi US · enteringSamsung KR · 2026 entry
Cambridge GaN Devices UKVisIC ILROHM · Toshiba · Sanken JPNexperia NLMACOM · Qorvo US · RF GaN



L3cFoundriesMerchant WBG wafer capacity · the fabless enabler

GlobalFoundries US · Navitas partnerPSMC · Powerchip TWVanguard (VIS) TWX-FAB EU
Polar Semiconductor USTSMC TW · exiting GaNSamsung Foundry KRCR Micro · Silan CN



L4Packaging materialsAMB/DBC ceramic substrates, sinter pastes, die attach, bond wire · top 8 ≈ 80-90% (est)

Rogers (curamik) US/DEHeraeus Electronics DE · ~15%Kyocera JPDenka JPMitsubishi Materials JP
NGK / Niterra JPToshiba Materials JPDOWA Metaltech JPProterial JPFerrotec JP
Xinzhou Electronic CNZhejiang TC Ceramic CNBYD CN · captiveKCC · Amogreentech KR
Indium Corporation US · sinterMacDermid Alpha US · sinterHenkel DE · TIM/attachTanaka · Nippon Micrometal JP · bond wire



L5Power modules & convertersHalf-bridges, IPMs, DC-DC bricks, PSUs, rack power boards

Infineon DE · modules+POLSemikron Danfoss DE/DKMitsubishi Electric JPFuji Electric JPVincotech EU
Monolithic Power (MPS) US · VRMVicor USAnalog Devices USTexas Instruments USRenesas JP
Delta Electronics TWLite-On TWFlex · Celestica US/CAVertiv USSchneider Electric FR
ABB CH/SENavitas PDB US · 800V→6VLittelfuse US · protection



L6Capital equipmentThe picks-and-shovels of the picks-and-shovels

Axcelis (Purion) US · SiC implantAixtron DE · epiLPE IT · epiVeeco US
DISCO JP · dicing/grindApplied Materials USKLA · Onto · Bruker US · metrologyTokyo Electron · Kokusai JP
Revasum US · CMPSiltectra/Infineon cold split DE · captiveAMEC · NAURA CN



JapanUnited StatesEuropeChinaKoreaTaiwanShaded tiles = category leaders


How to read this map. The shares quoted are approximate and drawn from public research of differing vintages and scopes. Treat them as positional, not precise. Company names are set as text, not logos. Several firms appear at multiple layers, which matters structurally: vertical integration is the dominant competitive strategy, and the companies that survived the glut are disproportionately the ones that owned their own substrate supply.
L1Crystal growth & substrates





In brief


Silicon carbide (SiC) and gallium nitride (GaN) are the enabling semiconductors of 800V DC data centres. Five vendors, STMicroelectronics, onsemi, Infineon, Wolfspeed, ROHM, hold over 90% of SiC merchant revenue. Infineon's $830M GaN Systems acquisition (closed October 2023 per company press release) is the largest transaction in the sector. Post-Wolfspeed Ch.11 restructuring, the AI DC positioning question remains open.








The wafer layer: chokepoint to commodity


Silicon carbide sublimes rather than melts, and that is the constraint the whole industry works around. You cannot pull a SiC boule the way you pull silicon from a Czochralski melt; you grow it by physical vapour transport, sublimating powder at ~2,300°C and re-condensing it on a seed crystal at millimetres per hour. A silicon boule grows in a couple of days and is two metres long. A SiC boule takes a week or more and yields perhaps 25-35mm of usable crystal. Then you cut it, with a material harder than almost anything except diamond, which is why wire-saw kerf loss historically destroyed a third of the crystal and why laser-based "cold split" techniques (Infineon's Siltectra acquisition) are strategically significant.


Substrates were the industry's chokepoint for a decade. They carried 40-50%+ of device cost, and every serious device maker moved to secure captive supply: onsemi bought GT Advanced Technologies, ST bought Norstel, ROHM has SiCrystal, Coherent built II-VI's franchise. The reversal has been just as violent. Once China decided SiC was strategic, capacity arrived at a pace the physics no longer constrained. Chinese substrate capacity ran from roughly 460k units in 2022 to 1.17M in 2023, 2.22M in 2024, an estimated 3.9M in 2025, with projections toward 7.7M by 2027, backed by a reported CNY 50bn (~$7bn) national fund commitment across 2024-25.




China's substrate capacity ramp: the supply shock in one chart
Chinese SiC substrate annual capacity, million units, 2022-2027E

Source: industry compilations of Chinese capacity announcements (2025). Nameplate capacity, not effective qualified output. The gap between the two is where the incumbents still live.





Market structure and share


Wolfspeed still led the substrate market with roughly 34% share in 2024, but the composition beneath it tells the story. TanKeBlue and SICC hold ~17% each, having arrived from effectively nowhere, with Coherent at ~14%. The top five (adding SiCrystal/ROHM, Resonac, SK Siltron) control about half of global capacity. SICC leads Chinese 8-inch. TanKeBlue is the volume domestic supplier. Meanwhile Japanese and Korean suppliers (Resonac, Sumitomo Electric, SK Siltron CSS) position explicitly as the non-Chinese alternative for Western automakers and, increasingly, for data center supply chains where provenance rules bite.




SiC substrate market share, 2024 basis, pre-restructuring
Approximate revenue share on 2024 revenue of ~$1.04B. This is the most recent broadly-cited breakdown, and it is two years stale

[A2] TrendForce (May 2025), 2024 data. Read with caution. Wolfspeed's share has almost certainly fallen since. The company itself disclosed customer inventory-building and second-sourcing during the Chapter 11 process, and Chinese suppliers have continued to add capacity. Treat the ranking as directional and the 2024 percentages as historic.





The 200mm (and now 300mm) transition


The economics are unambiguous. Moving from 150mm to 200mm yields roughly 2.2x more die per substrate and up to ~40% lower cost per die, with wafer-level cost roughly 35% lower. Everyone is going there: Wolfspeed's Mohawk Valley (the first 200mm SiC fab, backed by CHIPS Act funding), Infineon's Kulim Module 3 in Malaysia now ramping, ST's Catania campus, and onsemi's Czech expansion supported by a ~€530M EU state aid package. The catch is that 200mm yields have lagged. Industry commentary through 2025 put average yields below 50%. That compressed margins for everyone at the same time, because you carry the depreciation of the new fab while still selling the old wafer's economics.


Then, in January 2026, Wolfspeed reportedly demonstrated a 300mm SiC bouleThis claim rests on a single research-house reference and has not been corroborated by Wolfspeed primary disclosure at time of writing; treat it as reported rather than confirmed. If real, that positions Wolfspeed for standard-setting rather than near-term volume. Whoever defines the format that customers validate gets to shape a decade of tool roadmaps, and it is a reminder of what the restructuring actually preserved: the balance sheet failed, the materials science did not. One further caveat: a boule demonstration is many years from qualified 300mm production.



Diligence note on "capacity." Chinese nameplate capacity and qualified automotive/data-center-grade output are not the same number, and the gap is where the incumbents still live. Qualification runs 12-24 months for automotive (AEC-Q101) and is tightening for hyperscale infrastructure. When you see a capacity figure, ask three questions: what defect density, what qualification status, and whose provenance rules apply to the end customer.

L2Epitaxy



Epitaxy: where substrate quality becomes device economics


Between the polished substrate and the device sits the epitaxial layer, the doped, defect-controlled crystal film where the device actually lives. Epi determines blocking voltage and, more importantly, determines yieldA basal-plane dislocation propagating from substrate into epi will kill a MOSFET's reliability under bipolar stress. Epi is the step where substrate quality either converts into device economics or doesn't. It is also why the substrate price collapse has not fully passed through to device prices: good epi on cheap substrate is still not cheap.


Most volume epi is captive (Wolfspeed, Infineon, ST, onsemi, ROHM all run their own), which keeps the merchant market small but strategically visible. Resonac is the leading merchant epi-wafer supplier, Wolfspeed offers 200mm epi commercially, and Chinese players (Epiworld, Enkris for GaN) have scaled fast.


Three things make this layer more interesting than its revenue suggests. First, it is where the substrate glut gets absorbed or wastedA cheap substrate with high basal-plane dislocation density still produces expensive devices, because the yield loss shows up downstream. This is the mechanism by which the price war has hurt Chinese device makers less than expected and Western ones more: the incumbents were already buying good material and had less to gain from cheap wafers. Second, throughput is the hidden cost driverSiC epi grows at a few microns per minute, and a 1200V device needs perhaps 10-12µm of drift layer, so reactor time per wafer is measured in tens of minutes against seconds for many silicon steps. Reactor productivity (wafers per hour per tool) is a first-order determinant of fab economics, which is why the tool choice matters as much as the recipe. Third, the equipment layer is a genuine oligopolyAixtron and Italy's LPE dominate SiC epi reactors, with Veeco and Japanese suppliers in adjacent positions. This is one of the cleanest ways to hold exposure to WBG volume without picking a device winner. The tools get bought whether the wafers end up in a Chinese fab or a German one.


For GaN the picture inverts. GaN-on-silicon epi runs on MOCVD tools (Aixtron and Veeco again, plus Chinese suppliers) on 200mm silicon substrates, and the epi is the hard part because you are growing a lattice-mismatched film on a foreign substrate. GaN's cost curve therefore behaves like silicon's while its supply chain behaves like a compound semiconductor's. That is why foundry partnerships (Navitas / PowerChip, Navitas / GlobalFoundries, Innoscience's integrated epi) determine competitive position more than in SiC.




Anatomy of a restructuring: what Wolfspeed's Chapter 11 actually did
The sector's defining corporate event, in numbers


The problem. Wolfspeed built the first 200mm SiC fab (Mohawk Valley) and the Siler City materials facility on a debt-funded capex programme sized for an EV ramp that decelerated, arriving into a price collapse that halved wafer economics. The mechanism. A pre-packaged Chapter 11 filed 30 June 2025 with support from holders of >97% of senior secured notes, Renesas' US subsidiary (a creditor through its prepayment agreement), and >67% of convertible holders. The outcome (emerged 29 September 2025). Total debt reduced ~70%, roughly $4.6B; maturities extended to 2030; annual cash interest down ~60%; convertible holders took ~95% of new equity, existing shareholders diluted to 3-5%; five new directors appointed. The trading reality afterwards. Fiscal Q1 2026 revenue of ~$197M (roughly flat sequentially, +1% YoY), net loss cut to $85.2M from $119.8M largely by halving capex from $211.6M to $103.9M, cash of ~$926M, and guidance for revenue to decline to $150-190M the following quarter because customers had built inventory and second-sourced during the bankruptcy. The lesson for diligence. A balance sheet can be fixed in ninety days. A qualification position lost to a second source takes years to win back. That is the real cost of distress in this industry, and it does not appear in the restructuring arithmetic.




L3aSiC devices



Who are the five vendors that dominate SiC power devices?


Roughly five companies (STMicroelectronics, onsemi, Infineon, Wolfspeed, ROHM) have controlled 90%+ of SiC device revenue. ST has led on the strength of automotive MOSFETs (the Tesla relationship being the origin story of the modern SiC industry) and is completing its vertically-integrated Catania campus plus the Sanan joint venture in China. onsemi rose fast on EliteSiC with >50% substrate self-sufficiency post-GTAT, added Qorvo's SiC JFET business for $115M in January 2025, and secured EU state aid for Czech capacity. Infineon has the broadest portfolio across Si, SiC and GaN, launched trench-based SiC superjunction devices claiming ~40% lower RDS(on)×A, and has stated a "30-30" goal of 30% global SiC share by 2030.


What AI changes is the voltage and mission profileAutomotive SiC is a 750-1200V, cost-obsessed, AEC-qualified game. Data center SiC spans a wider range: 1200V for the sidecar/rectification stages, and increasingly the 2300V and 3300V classes for medium-voltage grid interface and solid-state transformer work, where Navitas' GeneSiC trench-assisted planar devices have carved out a genuine niche against much larger firms. Japanese module houses (Mitsubishi Electric, Fuji Electric, Toshiba) hold the traction and industrial-drive strongholds and are natural beneficiaries of any MV-DC build-out. Bosch is the automotive-captive wildcard.




SiC device market: a five-firm oligopoly, 2023/24 basis
Approximate revenue share. The top five have held ~90%+. The ordering below reflects 2023/24 rankings and predates both the Wolfspeed restructuring and the AI mix shift.

[A4] TrendForce 2023 ranking (ST 32.6%, top five 91.9%), cross-read with Mordor and Yole 2024/25 commentary. Individual percentages beyond ST are the author's interpolation within published constraints. See Annex §A4. Positions have moved materially since; treat as structural, not current.





Chinese domestic substitution is moving up-stack, not just at the wafer


The market essay's story about Chinese wafer capacity is well-worn; what is less discussed is that Chinese players are now moving up the value chain into qualified device production for domestic hyperscale. BYD Semiconductor, United Nova (UNT), BASiC Semiconductor and AccoPower are all shipping SiC MOSFETs into Chinese EV programmes and, increasingly, into Chinese cloud infrastructure (Alibaba, Baidu, Tencent, ByteDance). Chinese hyperscalers have their own strong provenance preferences that cut the opposite direction to Western ones. The practical implication for the TAM discussed later in this piece is that the $9-14B 2030 base case is a global figure; the Western-hyperscale addressable slice is materially smaller, because Chinese cloud demand will largely be served by Chinese devices as their qualification cycles complete through 2027-29. The annex includes a Western vs Chinese TAM split as a stress test.



L3bGaN devices



Where does GaN fit in AI data centre power?


Gallium nitride is not SiC's competitor so much as its complement, and the division is physical. SiC handles high voltage and high power with superior thermal conductivity: grid-to-rack conversion, traction inverters, MV interfaces. GaN switches faster at moderate voltages, which means smaller magnetics, higher power density, and better efficiency in the intermediate and point-of-load stages inside the rack. Gartner's framing of Infineon's advantage captures the consensus precisely: SiC for high-efficiency, high-voltage grid-to-rack conversion; GaN for ultra-dense, high-frequency intermediate stages; silicon at the processor level.


The competitive structure is younger and stranger than SiC's. Innoscience (China) leads with roughly 30% share, vertically integrated down to epi wafers, with an 8-inch GaN-on-Si platform and a partnership with ST that gives it Western reach. Infineon bought GaN Systems for $830M per company press release (announced March 2023, closed October 2023), still the largest transaction in the category, and is developing 12-inch GaN-on-Si pilot lines. Renesas acquired Transphorm for $339M and is on track to surpass $100M in GaN revenue. Power Integrations has PowiGaN plus the Odyssey vertical-GaN asset. EPC remains the low-voltage e-mode pioneer. Navitas has executed the sharpest strategic pivot in the sector, moving out of consumer chargers into "Navitas 2.0", with high-power markets contributing a majority of revenue for the first time in Q4 2025, a GlobalFoundries partnership for US-based GaN manufacturing, and 200mm GaN-on-Si at PowerChip for its 100V FETs.


Innoscience sits inside the gallium supply chain rather than downstream of it, which turns the raw-material risk flagged elsewhere in this piece into a company-specific advantage. It listed on the HKEX in December 2024, raising roughly $180M and giving the GaN category its first Chinese-domiciled listed vehicle. The 2025 co-development agreement with STMicroelectronics gives it a Western-qualified path into designs that provenance rules would otherwise close. The 8-inch GaN-on-Si platform (competitors sit at 6-inch or pilot 12-inch), plus vertical integration from epi through devices, adds up to a per-die cost structure the rest of the field is chasing on the 100-650V parts that dominate rack-power and consumer sockets. Not an M&A target in any realistic scenario. National champion status makes its capacity trajectory a major exogenous variable in every non-Chinese GaN player's model.


Navitas' 800V-to-6V power delivery board, shown in the NVIDIA MGX ecosystem at Computex 2026, eliminates the traditional 48V intermediate bus converter stage inside the compute tray altogether. Removing a conversion stage from the architecture redistributes BOM dollars permanently, which is why this product matters more than a typical efficiency step.



GaN's substrate economics: three routes, three cost curves


GaN commercial architectures split three ways on substrate, and the cost curves diverge sharply. GaN-on-Si is the volume play: buffer layers manage the ~17% lattice mismatch to a silicon substrate, and the epi runs on depreciated 200mm (increasingly 300mm) silicon foundry lines. Substrate cost per wafer is essentially the silicon commodity price. Yield is the constraint, not the substrate. This is the architecture Innoscience, Infineon (GaN Systems), EPC, Navitas and Power Integrations all use for <900V products. GaN-on-SiC uses a silicon carbide substrate for its superior thermal conductivity, and is the workhorse for RF GaN (Qorvo, Wolfspeed's RF business, MACOM). Substrate cost is roughly 5-10x GaN-on-Si, but the thermal-conductivity advantage matters where power density is extreme. Bulk GaN substrates (GaN-on-GaN) eliminate lattice mismatch entirely and enable vertical GaN devices, but production-scale GaN substrates remain rare and expensive; the market is measured in tens of thousands of wafers per year, mostly from Sumitomo Electric and Mitsubishi Chemical. Bulk GaN is the enabling substrate for the vertical GaN devices that could break the 650-900V ceiling of the lateral architecture, and its scale-up curve is a critical variable in the market essay's vertical-GaN thesis.


The near-term implication: GaN-on-Si dominates the <900V rack-power sockets because its cost curve tracks silicon; GaN-on-SiC holds specialty niches; GaN-on-GaN is the technology to watch for the 1200V+ vertical transition. Foundry yield curves matter for the first two more than substrate cost, because substrates are commoditised (Si) or affordable-if-specialised (SiC).



Two structural notes. First, foundries matter more in GaN than SiC because GaN-on-Si runs on depreciated silicon lines. TSMC's announced exit was a shock, but GlobalFoundries, PSMC, Vanguard, X-FAB and Polar have filled in, and Samsung is preparing a 2026 entry. Second, onsemi's absence is conspicuousYole's assessment that its entry is inevitable is the sort of thing that gets resolved by acquisition rather than organic development.




Power GaN: ranking is sourced, the split is estimated
Only two figures here are published: Innoscience at ~30% (2024) and a top-five total of ~88%. The individual shares of positions 2-5 are the author's estimates fitted to those two anchors.

[A3] Anchors: Yole Group, "The power GaN race" (Sep 2025) for Innoscience ~30%; Intel Market Research for top-five ≈88%. Positions 2-5 are estimates, not published data, shown to convey structure, not precision. Ordering follows Yole's qualitative ranking. See Annex §A3.


Materials risk: two-sided, not one-sided. Gallium is the visible risk in the GaN thesis (Chinese export instrument since 2023). SiC is often described as "geopolitically boring" by contrast, but that framing understates SiC's own exposure to input constraints. The honest picture is a matrix, not an asymmetry.




InputSiC exposureGaN exposure

Gallium metalNone~95% of primary supply from China, formal export controls since 2023 and tightened since. The most concentrated input in either chain.
Silicon powder / silicon substrateLow. PVT charge material, commodity supply.Silicon substrate for GaN-on-Si (the dominant commercial architecture). Commodity supply, no restriction risk.
High-purity graphiteMedium. PVT crucibles, susceptors and thermal shields all consume graphite. China holds an estimated ~65% of high-purity graphite production. Some grades are subject to Chinese export-licence scrutiny under the same regime as gallium.Medium. MOCVD reactor components draw from the same supply base.
Crystal growth & MOCVD equipmentAixtron and LPE (EU) subject to US-driven equipment export controls into China, which restricts Chinese fab tooling but does not affect Western build.Aixtron and Veeco MOCVD tools sit under the same US-driven regime.
Rare-earth alloys (magnets, susceptors)Low but non-zero. Some PVT chamber components use rare-earth-containing alloys.Low but non-zero.





The takeaway. Gallium exposure is genuinely more concentrated and more politically deployed than any single SiC input, and every bullish GaN thesis carries an implicit assumption about gallium availability. Watch Western gallium recovery projects (Rio Tinto, Nyrstar) as the mitigating vector. But SiC is not risk-free: high-purity graphite consumables and PVT/MOCVD equipment controls run in both directions across the SiC/GaN divide, and any Chinese retaliation to further US semiconductor controls could pull them in. Any thesis that positions SiC as immune to input risk is over-selling the asymmetry.
L4Packaging materials



Where did the SiC scarcity migrate to?


This is the layer most WBG analysis skips, and the one a materials-first reader should spend time on. A SiC die switching 1200V at high frequency generates heat in a smaller area than a silicon die and runs hotter. The package, not the die, is usually what limits the module. The critical component is the ceramic substrate that carries the die, insulates it, and conducts its heat out: DBC (direct bonded copper) on alumina or aluminium nitride for conventional parts, and AMB (active metal brazed) silicon nitride for the demanding ones. Si₃N₄ now represents over 95% of the AMB mix because its fracture toughness survives the thermal cycling that cracks alumina.


This market is small in revenue but structurally tight. Estimates put it somewhere between ~$250M and ~$1.2B depending on scope, growing 9-30% depending on whose definition you take. The top eight suppliers estimated to hold ~80-90%Rogers (curamik) and Heraeus (~15% share) anchor the Western supply. Kyocera, Denka, Mitsubishi Materials, NGK/Niterra, Toshiba Materials, DOWA and Proterial give Japan extraordinary depth. Production splits roughly Japan 27%, Germany 25%, China 43%, with China projected toward 57% by 2030. Most Chinese lines remain in qualification, which is the same story as substrates one cycle earlier, and should be read the same way.


Two live dynamics stand out. First, silverSintered-silver die attach is what makes high-temperature SiC modules reliable, and it ties the industry's cost base to a volatile precious metal. Heraeus now markets a silver-free AMB substrate explicitly as a hedge against precious-metal exposure, and low-silver sinter pastes are an active development race between Indium Corporation, MacDermid Alpha, Henkel, Kyocera and Heraeus. Second, bundlingThe strategic move in this layer is selling the substrate, the sinter paste and the thermal interface as one qualified system. That raises switching costs enormously and is the reason these quiet Japanese and German franchises defend margin better than the wafer makers ever did.




The AMB and sinter share of a high-power SiC module BOM
Approximate cost breakdown for a 1200V/300A SiC half-bridge module at scale, 2026 basis

Author's estimate from bill-of-materials analyses of published SiC module designs (Infineon EasyPACK, ROHM PMDUL series, onsemi ELiteSiC F1 series, Mitsubishi J-Series). AMB Si₃N₄ substrate and sintered-silver die attach together are roughly 15-22% of module BOM but ~80% of the reliability risk under thermal cycling. This is the arithmetic behind the packaging-margin thesis.


The wafer was the chokepoint of the last cycle. The ceramic substrate and the sinter paste are the chokepoints of this one, and few models capture them.The materials-layer thesis

L5Modules, converters & the rack



Where WBG meets the 800VDC architecture


In the 800VDC data center architecture, wide-bandgap devices appear at every conversion stage. SiC sits in the sidecar rectification and any solid-state transformer front end (and in the 2300/3300V class for medium-voltage interfaces). SiC also sits on the primary side of the 800V→54V/12V board-level converter, with GaN on the secondary. GaN or advanced silicon handles the point-of-load VRMs delivering ~0.7V at thousands of amps into the GPU package. An 800VDC bus carrying 140kW needs only ~175A, reported to cut copper requirements by ~45% against an equivalent 54V architecture. Cumulative chain efficiency can exceed 90% versus 70-85% for the legacy AC path. At gigawatt scale that difference is tens of megawatts released back to compute.


The content numbers are what make this investable, and management teams have started to disclose them. Infineon guides $12-15k of power semiconductor content per current 130kW AI rack, roughly $104 per kW, or ~$104k per MW. onsemi estimates content per next-generation 1MW rack has doubled from ~$50k to ~$100k, on ~$250M of 2025 revenue from these applications with high-teens sequential growth. Independent analysis puts power component content per rack rising ~4x from GB200 to Vera Rubin and ~11.5x to Rubin Ultra. Littelfuse reported 2025 data center design wins more than doubling year-over-year, with next-generation content opportunity "significantly more than double" current levels.




Content escalation: the investable fact
Indexed power component content per rack across three NVIDIA generations (GB200 = 1.0).

Source: The Diligence Stack (May 2026) estimates of ~4x GB200 to Vera Rubin and ~11.5x GB200 to Rubin Ultra. Corroborated by Infineon and onsemi disclosed content-per-rack figures. Indexed for comparability.





Who sits where in the rack



StageVoltageTechnologyPrincipal suppliers

<MV interface / SSTMV AC to 800VSiC 1700-3300VInfineon, Mitsubishi Electric, Fuji, Navitas (GeneSiC), Semikron Danfoss, Wolfspeed
Sidecar rectification / PSU480VAC to 800VDCSiC 1200V + SiInfineon (CoolSiC), ST, onsemi, Wolfspeed, ROHM; integrated by Delta, Flex, Lite-On, Vertiv
Board-level HV DC-DC800V to 54/12VSiC primary + GaN secondaryInfineon, Navitas, Power Integrations, TI, MPS, Vicor, Renesas
Direct-conversion PDB800V to 6VGaN (eliminates 48V IBC)Navitas (NVIDIA MGX ecosystem), Infineon, MPS
Point-of-load VRM54/12/6V to 0.7VGaN + advanced Si; thousands of ampsMPS, Infineon, TI, ADI, Renesas, Alpha & Omega, Vicor
Protection800VDCDC breakers, fuses, SSCBLittelfuse, ABB, Schneider, Siemens, Mersen





SCALEThe dog and the tail



Scale check: automotive is still the demand base


Everything above has been a data center argument, and a reader would be right to push back. AI data centers are not yet the principal demand driver for wide-bandgap semiconductors. Automotive is, by a wide margin. Published estimates put automotive at roughly 61% of SiC power semiconductor revenue in 2025 on one methodology and as high as ~81% of combined SiC+GaN device revenue on a broader one. Either way, the direction of EV production volumes still moves this industry's revenue more than anything happening in Santa Clara.


The case for focusing on data centers anyway rests on three points.


 1. Growth rate, not level. Automotive SiC is growing but decelerating against expectations. That is the demand miss that caused the glut. Data center WBG is compounding from a small base at rates no automotive programme can match: Infineon guiding AI-related revenue from €1.5B to €2.5B in one year, Wolfspeed's AI data-centre revenue more than doubling year-over-year in FY2026, onsemi's content per rack doubling. Marginal revenue growth is what re-rates equities and justifies capex, and it is now disproportionately coming from data centers.
 2. Margin quality. Automotive SiC is a cost-down machine. Multi-year LTAs with annual price-down clauses, brutal qualification, single-digit-cent negotiations. Data center customers buy density, efficiency and time-to-deployment. They are less price-elastic because the power semiconductor is a rounding error against the GPU. Same die, better economics.
 3. Architecture leverage. Automotive is a mature 400V/800V design space. Data center power architecture is being redesigned right now (800VDC, solid-state transformers, the elimination of the 48V intermediate bus) and design decisions taken in 2026-27 will lock supplier positions for a decade.


The framing for a model: automotive sets the industry's revenue base and its cycle; data center sets its growth rate, its mix, and its multiple. A reader building a forecast should size automotive first and treat the data center as the delta. A reader looking for where competitive position is being decided should watch the data center exclusively. The risk that follows is straightforward: this sector has swapped a demand concentration in one cyclical end-market (autos) for a growing concentration in another, more correlated one (hyperscaler capex).




Proportion, so the data center story stays in scale
Approximate end-market split of SiC power semiconductor revenue, 2025. Two published methodologies, shown side by side because they disagree.

[A5] Mordor Intelligence (automotive ~61.45% of SiC power semiconductor market, 2025) vs. Intel Market Research (automotive ~81% of combined SiC+GaN device market, 2024). Scope differences (SiC-only vs SiC+GaN, device vs system) explain most of the gap. See Annex §A5.


SiThe displaced incumbent



The silicon incumbent


Any technology-substitution story needs the incumbent's case put properly, and in power electronics the incumbent is enormous. IGBTs and superjunction MOSFETs represent roughly $11.6-18.7B depending on scope, several times the size of the entire SiC device market, and are growing 10-12% annually in their own right. Within the broader power semiconductor market of roughly $60B in 2026, silicon still does the overwhelming majority of the world's power conversion. WBG is the fast-growing minority, not the default.


Silicon also keeps improving specifically to defend itself. Toshiba's dual-sided multi-gate IGBT architecture claims ~34% lower turn-off loss than previous planar generations. Superjunction MOSFETs continue to improve the breakdown-voltage / on-resistance trade-off that was supposed to be their fundamental ceiling. Packaging innovation (double-sided cooling, embedded die) lifts silicon module performance without touching the die. Every generation of silicon improvement raises the bar that SiC's 3-5x lower switching losses have to clear economically, not just technically.


The substitution boundary is therefore a moving price line, not a fixed technical one. SiC wins where switching frequency, thermal headroom or system-level size reduction pay for the die premium: traction inverters, fast chargers, and now the 800VDC rectification stages where efficiency compounds across a gigawatt. Silicon holds everything else, and the wafer price collapse cuts both ways. It narrows SiC's cost gap, which accelerates substitution, but it also compresses the returns available to whoever wins that substitution. The bear case for WBG is SiC succeeding at prices that never justify the capital deployed to build it, rather than SiC failing technically. That is what happened between 2022 and 2025, and nothing structural prevents a repeat.



TAMSizing · Profit pools



What is AI actually worth to the wide-bandgap semiconductor industry?


Top-down forecasts for SiC and GaN are abundant and contradictory. SiC power semiconductors at ~$3.4B in 2026 growing ~25% CAGR to ~$10B by 2031 on one methodology; combined GaN+SiC at ~$4.1B in 2025 reaching ~$39B by 2035 on another. I build the number from scratch below, and I build it twiceThe two available methods disagree by a factor of three, and the disagreement is the most informative output.



Method A. Bottom-up from rack content


Infineon discloses $12-15k of power semiconductor content per current 130kW AI rack. Take the $13.5k midpoint and you get ~$104k per MW of IT loadonsemi independently puts content at ~$100k per next-generation 1MW rack. The two disclosures corroborate each other closely. Global data center power demand is running ~132GW in 2026 (up from ~104GW), with AI-optimised servers at ~31% of consumption. Roughly 18-20GW of new capacity is being added annually. Netting facility overhead to IT load (~70%) and taking the AI-relevant share of new build, I model ~8-10GW of AI IT-load additions in 2026At $104k/MW that gives ~$0.9-1.0B of rack-level power semiconductor content, of which WBG is perhaps 45-55%. Call it $0.4-0.6B.



Method B. Top-down from disclosed guidance


Infineon guides €1.5B of AI-related revenue in FY2026 and is described by Gartner as the company to beat in this segment. If its share of AI power semiconductors is 30-40%, the implied total market is €4-5B (~$4.5-5.5B) in 2026. Infineon separately sizes the AI power infrastructure opportunity at ~€12B by 2030. Applying a 45-55% WBG slice gives ~$2.0-2.9B of WBG device revenue in 2026.



Reconciliation, and why the gap matters


Method B is roughly 4x Method A, and the difference is definitional rather than arithmetic. Method A counts only what sits inside the compute rack. Method B counts the entire power tree: grid-side and facility conversion, UPS and switchgear electronics, PSUs, busbar systems, board-level converters and point-of-load regulation, plus service and replacement into the installed base, not just new build. Infineon's own description of its AI portfolio spans "utility-scale power conversion to rack-level power supplies", which confirms the wider scope. There is also a timing effect: revenue recognised in FY2026 includes design-in and pre-build ahead of 2027 deployments.


I adopt the full-power-tree definition, which is the economically meaningful boundary for anyone investing in this chain. That puts WBG device revenue from AI data centers at roughly $2.0-2.5B in 2026, with a central path to about $11B by 2030 (range $9-14B depending on the WBG share assumption and the pace of 800VDC adoption). Readers who prefer the narrow rack-only definition should divide by roughly four throughout. Full workings, including the scenario drivers, are in Annex §A6.




Bottom-up TAM: WBG devices in AI data centers
USD billions, full-power-tree definition (grid to GPU), 2026-2030E. Author's model.

[A6] Author's model. Inputs: content per MW (Infineon, onsemi disclosures); AI IT-load additions derived from Gartner capacity data and JLL/McKinsey build forecasts; WBG share of power BOM 45-55%; scenario drivers in Annex §A6. Cross-checked against Infineon's ~€12B-by-2030 AI power framing. Illustrative. Every input is stated and adjustable.





Profit pools: the inversion


The margin map of this industry has inverted over three years, and the inversion is where allocation decisions get made. In 2022, substrates were the scarce layer and carried 40-60% gross margins. Today, after the glut, merchant substrate margins have compressed brutally (Chinese suppliers reportedly selling near cost) and the layer is structurally commoditizing. Devices hold better (35-50% for qualified automotive and infrastructure parts) because qualification and reliability data are the moat, not the wafer. Packaging materials (AMB substrates, sinter pastes) have become the best-defended layer at 30-45%, protected by qualification bundling and Japanese/German process depth. Modules and rack-level converters run 20-35%, with the AI-specific parts currently enjoying scarcity pricing that will normalize as OCP standardizes. Equipment (Axcelis, Aixtron, DISCO) retains 40%+ through the cycle, because both the boom and the capacity race require tools regardless of who wins.




Margin migration along the WBG chain
Indicative gross margin bands by layer: 2022 (substrate scarcity) vs. 2026 (post-glut, AI-led).

Author's estimates from company disclosures and industry analogues. Indicative midpoints of ranges: directional, not audited.


The wafer makers built the industry and captured the least of it. The tool makers and the ceramic houses never had a bad quarter.The profit-pool lesson of the SiC cycle
CAPITALM&A · Disruption · Outlook



The M&A chessboard


A consolidating industry emerging from a price war, with a demand wave arriving and several sub-scale specialists holding critical IP, is the most target-rich environment in semiconductors right now. The precedents are already substantial: Infineon / GaN Systems ($830M), Renesas / Transphorm ($339M), onsemi / Qorvo SiC JFET ($115M), onsemi / GTAT, ST / Norstel, SK Siltron / DuPont SiC, Power Integrations / Odyssey, Innoscience's $180M IPO. On the minority-stake path, Denso and Mitsubishi Electric's investment into Coherent's SiC business demonstrated that customers will fund capacity directly when supply security matters more than ownership.



AssetSituationLogical acquirersLogic & friction

WolfspeedPost-emergence: debt cut ~70%, maturities to 2030, convertible holders own ~95%, 200mm vertically integrated US footprint, 300mm demonstrated, AI DC revenue more than doubled YoY in FY26Strategic: Infineon (fastest route to "30-30"), onsemi, Renesas (already a creditor-turned-stakeholder), a Japanese consortium. Financial: Apollo-led lenders already own itThe most interesting asset in the sector. Friction: US national-security review of any non-US acquirer of CHIPS-funded capacity is a near-certainty; new owners may prefer to run it and re-IPO. Watch for asset-level deals (a fab, an epi line) rather than whole-co.
NavitasFabless, GaN + high-voltage SiC, deep NVIDIA MGX design-in, GlobalFoundries partnership, "Navitas 2.0" pivot to high power. Small relative to design-win footprintonsemi (fills the conspicuous GaN gap), Renesas, TI, Microchip, or a power-systems strategic (Vertiv, Delta) wanting silicon-level controlDesign-ins at NVIDIA are worth more than revenue run-rate. That is precisely what an acquirer pays up for. A takeout premium here is a design-win premium.
EPC, Cambridge GaN Devices, VisICSub-scale GaN specialists with differentiated device architecturesIDMs needing a GaN position quickly: onsemi, Microchip, Nexperia, Samsung, Chinese acquirers (subject to controls)Small cheques, big capability transfer. Classic technology bolt-ons of a consolidating category.
Chinese substrate capacityMassive overbuild, price war, many sub-scale players below costDomestic consolidation under state-linked buyers. Western acquisition effectively blocked.The consolidation will happen inside China. The output will keep pricing the global merchant market for years. A structural fact, not a cycle.
AMB / ceramic substrate housesFragmenting at the tail. Top eight hold ~85%. Chinese entrants scaling.Materials majors (Heraeus, Rogers, Kyocera) buying tail capacity. Module makers integrating backward.Where I would look hardest for a mispriced asset. Small, unglamorous, structurally short, defended by qualification.
Equipment specialistsAxcelis (SiC implant), LPE (epi), Revasum (CMP)Larger equipment groups (AMAT, TEL, ASM). LPE-style tuck-ins into Aixtron-scale platforms.Picks-and-shovels consolidation. Buyers have the currency and the cycle is favourable.





Valuation framing


Three anchors for an IC discussion. Precedent multiples. Infineon paid ~$830M for GaN Systems at a reported high multiple of then-revenue: a capability price, not an earnings price. Renesas paid $339M for Transphorm on similar logic. In a market where design-in position determines a decade of share, capability multiples are rational. Distressed asset pricing. Wolfspeed's restructuring effectively valued a vertically integrated 200mm SiC franchise (Mohawk Valley, Siler City materials, CHIPS support, 300mm capability) at the enterprise value implied by ~$1.5-2B of post-emergence debt plus a diluted equity stub. Against ~$10B+ of cumulative capex sunk into that footprint, the replacement-cost argument is stark. Against continued 6-inch price erosion, so is the caution. Strategic-scarcity premium. Anyone who wants qualified non-Chinese WBG capacity in the 2027-29 window has approximately five phone calls to make, and every one of those counterparties knows it.




Technology disruption watch list



VectorWhat it isWhoImpact if it lands

300mm SiCWolfspeed's Jan 2026 boule demo. Die-per-wafer step change.Wolfspeed. Tool ecosystem must follow.Resets cost curve and standards. A decade of tool roadmaps at stake.
Trench & superjunction SiC~40% lower RDS(on)×A vs planarInfineon, ROHM, onsemi, Navitas (trench-assisted planar)Die shrink = more devices per wafer = deflation without margin loss for the leaders
Engineered substratesSmartSiC-style bonded thin-film substrates using less prime crystalSoitec and partnersBreaks the boule-growth bottleneck economically. Threatens pure substrate players.
Laser cold-split slicingEliminates most kerf loss vs wire sawingInfineon (Siltectra), Halo and othersEffective capacity increase of 20-30% from existing crystal. Deflationary.
Vertical GaN & GaN-on-SiCPushes GaN into higher voltage/power territoryPower Integrations (Odyssey), Navitas, research ecosystemErodes the clean SiC/GaN voltage boundary, which is the key competitive assumption.
Bidirectional switches (BDS)Single-device AC switching. Fewer components.Infineon CoolGaN BDS, Nexperia, researchSimplifies rectification stages. Changes converter topology BOM.
Ultra-wide bandgap: Ga₂O₃, diamond, AlNBandgaps beyond SiC/GaN. Theoretical efficiency leaps.Japanese research consortia (NICT, Flosfia), US labs, startups2030s technology. Thermal conductivity is Ga₂O₃'s unsolved problem. Watch, don't underwrite.
Silver-free AMB & low-silver sinterRemoves precious-metal exposure from module cost baseHeraeus (Condura.ultra), Indium, MacDermid AlphaStructural cost-down and supply-risk reduction at the packaging layer
Double-sided cooling & embedded dieDies embedded in PCB/substrate, cooled both facesInfineon, Delta, module houses, substrate makersDoubles power density. Shifts value further into packaging.
SiC as a thermal material~500 W/m·K heat spreaders/carriers for AI acceleratorsSubstrate makers seeking non-power demandA genuinely new demand pool for a glutted material. The glut's best escape valve.






Standards, controls and the policy layer


Four instruments shape this industry's cash flows more than any technology roadmap. Qualification standards (AEC-Q101 for automotive, JEDEC reliability suites, and the emerging and still immature expectations for hyperscale infrastructure) are the real moat. They are why nameplate capacity and sellable capacity diverge. Export controls run both ways: US restrictions on advanced equipment into China, and China's gallium controls running the other direction, with the added wrinkle that Wolfspeed's CHIPS-funded assets carry US strategic strings that constrain who could ever buy them. Industrial policy is writing cheques at scale: CHIPS Act grants and DOE loans in the US, the EU Chips Act behind ST's Catania and Bosch's Dresden, the ~€530M Czech package for onsemi, and China's national IC fund committing reported multi-billion sums into SiC. Materials regulation (the EU Critical Raw Materials Act, REACH restrictions reshaping baseplate metallization, ECHA scrutiny of die-attach chemistries) determines which processes survive in Europe. Policy operates as a primary determinant of who has capacity in 2029 rather than as background context.




Trade-offs, winners and losers


 * SiC vs GaN. Complementary today, competitive at the boundary tomorrow. Vertical GaN and GaN-on-SiC push upward in voltage. SiC trench devices push downward in cost. The 650-1200V band is where they will genuinely fight, and it happens to be where a lot of data center content sits.
 * Vertical integration vs fabless. The glut rewarded integration (owning cheap internal wafers when merchant prices collapsed) but punished the capex. Post-glut, with substrates cheap and abundant, the fabless model looks structurally better than it has in a decade. That is a large part of why Navitas' asset-light pivot is working.
 * Deflation vs volume. Every device-level improvement (trench, 200/300mm, cold split, engineered substrates) is deflationary per die. The industry grows in revenue only if volume outruns ASP erosion. AI helps enormously because data center content per rack is rising faster than device prices are falling, for now.
 * Winners. Infineon (breadth, scale, AI design-ins, "30-30" ambition), Navitas (design-in position vs size), the equipment layer (Axcelis, Aixtron, DISCO), the packaging-materials houses (Rogers, Heraeus, Kyocera, Denka, Mitsubishi Materials), the Japanese module incumbents if MV-DC and SST arrive, and a recapitalized Wolfspeed if AI mix outruns the 6-inch price war.
 * Pressured. Merchant substrate pure-plays without captive demand, sub-scale Chinese substrate makers heading into domestic consolidation, GaN specialists without a foundry partner or a design-in, and any business model that assumed 2022 substrate pricing would persist.
 * The dominant risk. An AI capex correction. Automotive demand is no longer the swing factor it was. This industry has re-levered itself to hyperscaler capital expenditure, which is a stronger driver in 2026 but also a more correlated one.




Ten leading indicators to watch


 1.  Wolfspeed's AI revenue mix and 200mm utilization. The clearest single read on whether the restructuring thesis works.
 2.  Infineon's progress against "30-30" and its AI revenue trajectory past €2.5B. The sector's benchmark disclosure.
 3.  Chinese 6-inch spot pricing. If it stabilizes above ~$250, the worst is over. If it keeps sliding, more consolidation.
 4.  200mm yield disclosures. The gap between nameplate and effective capacity closes here first.
 5.  onsemi's GaN entry. Organic or by acquisition. Either way it reprices the GaN specialists.
 6.  Navitas 800V-to-6V design wins converting to volume. Validates the elimination of the 48V stage and the GaN content thesis.
 7.  AMB Si₃N₄ lead times and Chinese qualification announcements. The next chokepoint's early warning.
 8.  Gallium export policy moves. The GaN thesis's single biggest exogenous risk.
 9.  SST pilot deployments in data centers. Pulls 1700-3300V SiC demand forward by years if it lands.
 10. Any national-security review triggered by a Wolfspeed asset sale. Tells you whether Western WBG capacity is genuinely for sale.




Sustainability: a strong story with three honest caveats


The headline case is genuinely excellent and rarely disputed. Wide-bandgap devices exist to reduce conversion losses, and in a data center context the arithmetic is large. An 800VDC chain built on SiC and GaN can exceed 90% cumulative efficiency against 70-85% for the legacy AC path. At gigawatt scale that difference is tens of megawatts that stop becoming heat and start becoming compute. Multiply across the EV and renewables fleets and WBG is one of the few semiconductor categories whose use-phase savings unambiguously dwarf its manufacturing footprint.


The caveats deserve equal billing. First, crystal growth is energy-intensive. PVT sublimation runs at ~2,300°C for a week or more per boule, making SiC substrate manufacturing materially more energy- and carbon-intensive per wafer than silicon. That is a real embodied-carbon burden even if the lifetime balance is favourable. Second, the fab chemistry is not clean. WBG fabs use fluorinated etch gases with very high global-warming potentials and are exposed to the tightening regulatory perimeter around PFAS in semiconductor manufacturing. GaN adds gallium and, in some processes, arsenic-adjacent handling requirements. Third, the packaging layer carries a precious-metal dependency. Sintered-silver die attach ties the module cost base and its supply-chain footprint to silver, which is why silver-free AMB substrates and low-silver pastes are being developed as much for supply resilience as for cost. (I have not cited a specific lifecycle-assessment figure here because I could not verify one to a standard I would defend. The directional statements above are well established. The precise multiples are not.)




The public-market exposure map


For analyst readers, the listed vehicles touching each layer. Mapping only. Nothing here is a recommendation, and exposure purity varies enormously.



LayerListed namesExposure purity

SubstratesWolfspeed (WOLF), Coherent (COHR), SICC (STAR-listed), Sanan (600703.SS), Resonac (4004.T), Soitec (SOI.PA), GlobalWafers (6488.TWO)Wolfspeed is the purest listed SiC play and the most idiosyncratic post-restructuring. Coherent's SiC sits inside a photonics group.
SiC devicesSTMicro (STM), onsemi (ON), Infineon (IFX.DE), ROHM (6963.T), Mitsubishi Electric (6503.T), Fuji Electric (6504.T), BYD (002594.SZ)Mid. SiC is a growing minority of each group's revenue. ROHM the most SiC-levered of the majors.
GaN devicesNavitas (NVTS), Power Integrations (POWI), Renesas (6723.T), Innoscience (2577.HK), Infineon (IFX.DE)High for Navitas and Innoscience. The cleanest listed GaN expressions, and correspondingly volatile.
FoundryGlobalFoundries (GFS), Vanguard (5347.TWO), PSMC (6770.TW), X-FAB (XFAB.PA)Low purity. WBG is an incremental capacity-utilisation story for each.
Packaging materialsRogers (ROG), Kyocera (6971.T), Denka (4061.T), Mitsubishi Materials (5711.T), Niterra (5334.T), DOWA (5714.T), Ferrotec (6890.T)Low purity, high chokepoint quality. Rogers the most levered of the Western names. Heraeus and Proterial are private.
EquipmentAxcelis (ACLS), Aixtron (AIXA.DE), Veeco (VECO), DISCO (6146.T), Applied Materials (AMAT), KLA (KLAC), Onto (ONTO)Mid-to-high for Axcelis (SiC implant) and Aixtron (epi). The purest picks-and-shovels exposure in the chain.
Modules / rack powerMonolithic Power (MPWR), Vicor (VICR), Delta (2308.TW), Vertiv (VRT), Littelfuse (LFUS), Schneider (SU.PA), ABB (ABBN.SW)Mid. Strongest near-term AI revenue visibility, weakest long-term margin defense per the profit-pool argument.






Where this lands


Wide bandgap spent five years being sold as an electric-vehicle story, built capacity accordingly, and got the timing wrong by roughly the length of one product cycle. The correction was brutal, and it took the industry's founding US pure-play through bankruptcy court. What emerges is a chain that is cheaper, more consolidated, better capitalized in the places that matter, and pointed at a demand source (AI data center power) that pays for density and efficiency rather than cents per die. The wafers are commoditizing. The value has migrated to devices with qualification moats, to packaging materials that most models under-count, and to the equipment layer that gets bought regardless of who wins the device race. For operators, power semiconductor supply is no longer the constraint it was in 2022; design for the architecture you want, not the parts you can get. For investors, the profits in this cycle sat with the equipment makers and the packaging houses rather than the wafer producers, and the setup for the next cycle points the same way.



ANNEXSources · Calculations · Assumptions



Annex: how every derived number was built


Everything in this piece is either published (traceable to a named source), derived (calculated from published inputs, with the arithmetic shown below), or estimated (my judgement, flagged as such). This annex covers the second and third categories so any reader can rebuild, challenge, or discard the numbers.




§A1. SiC wafer price observations


Status: published observations, mixed metrics. The chart plots five discrete datapoints, not a continuous index, and they are not like-for-like.



PointValueWhat it measuresSource

Early 2023>$6006-inch epi wafer, open marketWoodside Capital Partners, "The SiC Capacity Glut and its impact on GaN" (Jun 2025)
End 2023$400-4506-inch epi wafer (plotted at $437 midpoint)Woodside Capital Partners (Jun 2025)
Mid 2024<$5006-inch substrateeeNews Europe (Nov 2025). Plotted at $490.
2025~$400 or lower6-inch substrate, mainstream quoteseeNews Europe (Nov 2025)
Jun 2026~$200Chinese 6-inch substrate, benchmark spotDigiTimes via The Economy (Jun 2026). Single source, unconfirmed.




Known weakness. The apparent uptick from end-2023 to mid-2024 is an artefact of switching from epi-wafer to substrate quotes between sources, not a real price recovery. Do not read the shape as a trend line. Separately cited: 6-inch epi wafer average prices fell ~15.9% between 2023 and 2024, and 8-inch declines exceeded 60% (industry compilation, 2025).




§A2. SiC substrate market share


Status: published, 2024 vintage. TrendForce (May 2025) on 2024 substrate revenue of ~$1.04B (down 9% YoY): Wolfspeed 33.7%, TanKeBlue 17.3%, SICC 17.1%, Coherent 13.9%. Residual 18.0% is my arithmetic (100 minus the four published shares), covering SiCrystal/ROHM, Resonac, SK Siltron, Sumitomo Electric, Synlight, JSG and others. Separately, Mordor (2026) states the top five (Wolfspeed, Coherent, ST, ROHM, SK Siltron) controlled about half of global capacity in 2025. That is a different metric on a different basis and is not reconciled here.




§A3. Power GaN market share


Status: two published anchors, positions 2-5 estimated. Published: Innoscience ~30% share of power GaN (Yole Group, Sep 2025, on 2024 data); top five ≈88% of revenue (Intel Market Research, on 2024 data). Derived: the residual 58 points across positions 2-5 were allocated as Infineon 18 / Power Integrations 15 / Navitas 13 / EPC 12, with 12 points left for Renesas-Transphorm and the tail. Ordering follows Yole's qualitative description of the competitive set. The individual values are my estimates and should not be quoted as data. Corroborating context: Renesas' Transphorm business was described as on track to exceed $100M of GaN revenue by 2026, which is consistent with a mid-single-digit to low-double-digit share of a market of this size.




§A4. SiC device market share


Status: published ranking, interpolated values. Published: TrendForce 2023 ranking gave ST 32.6% with the top five at 91.9% of revenue, order ST > onsemi > Infineon > Wolfspeed > ROHM, with ROHM at 5.0% (~$186M). Derived: I allocated the intervening positions as onsemi 18 / Infineon 17 / Wolfspeed 14 to fit both published constraints, leaving ~14% for all others. Only ST's and ROHM's figures are published. The middle three are interpolation. Note that Mordor's 2026 commentary gives a different top five for the SiC power semiconductor market (Infineon, ST, Wolfspeed, onsemi, ROHM at >90% combined) and GM Insights puts Infineon at >19.5% of the broader power semiconductor market in 2025. Different universes, not contradictions.




§A5. End-market split


Status: two published figures on different bases, shown side by side deliberately. Mordor Intelligence: automotive 61.45% of the SiC power semiconductor market in 2025. Intel Market Research: automotive ~81% of the combined SiC+GaN power device market in 2024, projected to ~83.6% by 2032. The remaining category splits in the chart (industrial/energy, data center, other) are my estimates fitted to each source's automotive anchor and are illustrative only. The discrepancy is mostly scope: SiC-only versus SiC+GaN, and differing treatment of industrial/energy applications.




§A6. TAM model for WBG devices in AI data centers


Status: derived. Full workings below.


Published inputs. Content per rack: Infineon guides $12-15k per 130kW AI rack; $13.5k ÷ 0.13MW = $103.8k per MWonsemi: content per next-generation 1MW rack doubled from ~$50k to ~$100k per MWAn independent corroboration within 4%. Capacity: Gartner puts global data center power demand at ~132GW in 2026 (from ~104GW in 2025), with AI-optimised servers at ~31% of consumption in 2026 (from ~20% in 2025) rising toward ~50% by 2030. Build rate: JLL and others put new capacity additions at roughly 18-20GW annually through 2027-29, rising to ~31GW in 2029-30, with ~100GW added 2026-2030. McKinsey projects ~156GW of AI-related capacity by 2030.


Derivation of AI IT-load additions. Facility capacity additions ~18-20GW/yr (2026-27); AI-relevant share of new build assumed at 55-65% (well above AI's ~31% share of the installed base, because new build skews heavily to AI); therefore ~10-13GW/yr of AI facility capacity. IT load at ~70% of facility power (Brookings: servers ≈60% of DC electricity, ~75% in AI-optimised hyperscale) gives ~7-9GW/yr of AI IT-load additions in 2026, rising to ~14-18GW/yr by 2030 as both build rate and AI share increase.


Method A (rack-only). 8.5GW × $104k/MW = ~$0.88B of rack-level power semiconductor content in 2026. WBG share of that BOM assumed 45-55% (SiC in rectification and HV DC-DC; GaN in intermediate and some POL stages; silicon retains much of point-of-load and control), giving ~$0.40-0.49B.


Method B (full power tree). Infineon FY2026 AI revenue guidance €1.5B ÷ assumed 30-40% share gives implied market €3.75-5.0B (~$4.3-5.7B at 1.15 USD/EUR). WBG slice at 45-55% gives ~$1.9-3.1BMidpoint adopted: ~$2.2B for 2026.


Why they differ (4x). Method A counts only in-rack content. Method B captures the full chain from utility conversion through UPS/switchgear electronics, PSUs, busbars, board-level converters and point-of-load, which is how Infineon itself describes its AI portfolio, plus installed-base service and pre-build revenue recognised ahead of deployment. Neither is wrong. They answer different questions. The article adopts Method B and says so.


Scenario construction to 2030. Base case: AI IT-load additions growing to ~15GW/yr by 2030; WBG share of power BOM rising from ~50% to ~60% as 800VDC penetrates and SiC enters MV/SST stages; content per MW roughly flat in nominal terms (density-driven content increases offsetting ~8-12% annual device ASP erosion). Result: ~$11B by 2030. Bear case: AI capex air-pocket cutting additions growth by half, 800VDC adoption slipping a generation, WBG share stalling at ~45%. Result: ~$4.5B. Bull case: 1MW racks on schedule, SST pilots pulling 1700-3300V SiC forward, retrofit wave into the 2025-26 installed base, WBG share ~65%. Result: ~$16B. Cross-check. Infineon sizes the total AI power infrastructure opportunity at ~€12B (~$13.8B) by 2030. A WBG slice of 55-65% of that implies ~$7.6-9.0B, somewhat below my base case. The gap is my assumption that Infineon's framing is its own addressable view rather than the full global market. A reader who prefers Infineon's framing should shade the base case down by 15-25%.


Sensitivity. The model is most sensitive, in order, to: (1) the WBG share of the power BOM (a ±10pt swing moves the 2030 base case by roughly ±$2B); (2) AI IT-load addition growth, itself gated by grid interconnect timelines discussed in Part IV: The Interconnect Stack (a 4-7 year US ISO queue is a real bear-case driver, not a tail-risk); (3) device ASP erosion, which could turn a volume boom into flat revenue if it runs above 15% annually.



Western hyperscale vs Chinese cloud split. The base-case TAM is global. Western hyperscalers (US, EU) source almost exclusively from Western/Japanese devices under strict provenance rules; Chinese hyperscalers (Alibaba, Baidu, Tencent, ByteDance) increasingly source domestic Chinese devices for the same reasons in reverse. The addressable-market split for Western device vendors is therefore materially smaller than the global TAM implies. Split shown below.




2030 TAM split: Western-hyperscale vs Chinese-cloud addressable
Base case ~$11B by 2030, decomposed by end-market accessible to Western device vendors

Author's decomposition. Assumes Western hyperscalers source ~95% Western/Japanese devices; Chinese hyperscalers ~70% domestic Chinese devices by 2030 (up from ~35% in 2026, tracking the qualification cycle at BYD Semi, UNT, BASiC, Innoscience). Rest of world (India, MENA, LATAM) is smaller and mostly Western-served for now. The Western-vendor addressable slice is ~$7-8B of the $11B global base case by 2030, or ~65-75% of it.



TAM waterfall: how Method A gets to Method B
Bridge from rack-only content to full-power-tree market, 2026, USD billions

Method A rack-level content ~$0.45B, plus grid-side and facility conversion (UPS, switchgear), plus board-level converters and POL regulation, plus service and installed base, plus pre-build revenue recognised ahead of deployment. The full sum lands inside the Method B range of $1.9-3.1B. Each bar is a definitional adjustment, not new physical demand.



Regional supply by layer, Marimekko view
Column width = share of layer revenue. Column height segments = geographic origin. 2024-25 basis.



SubstratesUS · Wolfspeed 34%
China · TanKe+SICC 34%
US · Coherent 14%
JP/KR/other 18%



SiC devicesEU · STMicro 32%
US · onsemi 18%
EU · Infineon 17%
US · Wolfspeed 14%
JP · ROHM+others 19%



GaNCN · Innoscience 30%
US · Navitas+POWI+EPC 33%
EU · Infineon 19%
Other 18%



PackagingJapan · Kyocera+Denka+MMC 60%
DE · Rogers+Heraeus 25%
CN · scaling 15%



EquipmentUS · Axcelis+AMAT+KLA 40%
EU · Aixtron+LPE 30%
JP · DISCO+TEL 25%
Other 5%



← column widths approximately proportional to 2025 layer revenue
column height segments = geographic share of that layer

Layer shares from TrendForce (substrates), TrendForce+Mordor (devices), Yole (GaN), Intel Market Research (packaging). Regional splits from §A2. A4. Illustrative reading device, not a source.





§A7. Profit pool / margin bands


Status: estimated. No source publishes gross margin by value-chain layer for this industry. The bands shown (substrates ~50% to 18%, epitaxy ~38% to 30%, devices ~35% to 42%, packaging materials ~32% to 38%, modules ~25% to 28%, equipment ~42% to 44%) are my judgement, built from company-reported segment margins where disclosed (Infineon, onsemi, Wolfspeed, Aixtron, Axcelis, Rogers, Coherent); the documented substrate price collapse (§A1) applied against broadly fixed cash costs; reported Chinese substrate selling near or below cost; and analogues from the lithium-ion materials chain where a scarce upstream commoditised faster than a qualified midstream. These are directional illustrations of a migration, not estimates of any company's actual margin.




§A8. Content escalation index


Status: published, re-indexed. Source: The Diligence Stack (May 2026) estimates power component content per rack rising ~4x from GB200 to Vera Rubin and ~11.5x from GB200 to Rubin Ultra. I have simply set GB200 = 1.0 and plotted 1.0 / 4.0 / 11.5. This measures power components broadly, not WBG semiconductors specifically, and it is one analyst's estimate rather than a manufacturer disclosure.




§A9. Other published figures used


 * Wolfspeed restructuring. Debt reduced ~70% / ~$4.6B, maturities to 2030, cash interest down ~60%, filed 30 Jun 2025, emerged 29 Sep 2025, convertible holders ~95% of new equity, existing shareholders 3-5%. Source: Wolfspeed press releases and SEC Form 8-K (Jun. Sep 2025). FQ1 2026 per company 10-Q: revenue $196.8M, net loss $85.2M (from $119.8M), capex $103.9M (from $211.6M), cash + ST investments $926M, guidance $150-190M. Source: Semiconductor Today via EAM Vision (Nov 2025).
 * Chinese substrate capacity. 460k (2022), 1.17M (2023), 2.22M (2024), 3.9M (2025E), 7.71M (2027E) units. CNY 50bn national IC fund commitment 2024-25. Source: industry compilations (2025) and Mordor (2026). Nameplate, not qualified output.
 * 200mm economics. 2.2x more die per substrate, up to ~40% lower cost per unit, ~35% lower cost per wafer vs 150mm. Global average 200mm yield reported below 50%. Source: Mordor Intelligence and industry press (2025-26).
 * Infineon. FY2026 AI revenue target €1.5B rising to €2.5B in FY2027. ~€12B AI power opportunity by 2030. FY2026 capex raised €500M to €2.7B. "30-30" ambition of 30% SiC share by 2030. Trench SiC superjunction claiming ~40% lower RDS(on)×A. Named by Gartner as company to beat in AI data center power. Sources: Infineon disclosures, Reuters via Astute Group, 247WallSt, Semiconductor Today (2026).
 * Navitas. 800V-to-6V power delivery board eliminating the 48V IBC stage, shown in NVIDIA MGX ecosystem, Computex Jun 2026. GlobalFoundries foundry partnership. PowerChip 200mm GaN-on-Si for 100V FETs. 2300V/3300V SiC sampling. High-power markets a majority of revenue for the first time in Q4 2025. Sources: Navitas press releases and SEC 8-K filings (2025-26).
 * onsemi. Content per 1MW AI rack ~$50k to ~$100k. ~$250M 2025 revenue from these applications. Qorvo SiC JFET acquisition $115M (Jan 2025). ~€530M EU state aid for Czech SiC facility (Nov 2025). Sources: company disclosures and trade press.
 * M&A precedents. Infineon / GaN Systems $830M; Renesas / Transphorm $339M; Innoscience IPO $180M. Source: Yole Group. onsemi / GTAT, ST / Norstel, SK Siltron / DuPont SiC, Power Integrations / Odyssey. Source: trade press.
 * AMB substrates. Top eight ≈85% share (2023). Si₃N₄ >95% of AMB mix. Production split Japan 27.3% / Germany 24.6% / China 43.5%, China to ~57% by 2030. Heraeus ~15% share. Heraeus Condura.ultra silver-free product. Sources: Intel Market Research, Global Growth Insights, Heraeus Electronics (2025-26). Market size estimates range from ~$243M (2025) to ~$1.2B (2024) across sources on differing scopes. This spread is unresolved. I have quoted the range rather than picking one.
 * Silicon incumbent. IGBT + superjunction MOSFET market $11.6B (2025, Mordor) to $18.7B (2025, Metastat) depending on scope. ~10-12% CAGR. Total power semiconductor market ~$60B in 2026 (Mordor). Toshiba dual-sided multi-gate IGBT ~34% lower turn-off loss, as cited.
 * 800VDC efficiency. ~175A at 140kW on an 800V bus, ~45% copper reduction vs 54V, chain efficiency >90% vs 70-85% legacy AC, board-level HV DC-DC at 97-98.5%. Source: Bristlemoon Research (Jun 2026).



A general caution on this sector's market data. Research houses disagree by factors of two to four on nearly every WBG number, driven by scope (device vs module vs system; SiC-only vs SiC+GaN; merchant vs captive) and by vintage. Where I found a spread I have shown it rather than picking the flattering end. Where a figure rests on one source I have said so. Where I filled a gap with judgement I have labelled it an estimate. Readers doing real diligence should treat this annex as a starting map, not a substitute for primary sources.

SERIESRelated in the AI Power Chain



What comes next


This is Part II of a six-part series on the AI power chain. Part I (The Capacitor Stack) traced how hybrid supercapacitors, EDLCs, MLCCs and silicon capacitors became scarce simultaneously under the 800VDC transition. Part III (The Thermal Stack) does the same for liquid cooling: cold plates, CDUs, immersion, and the shift from air to two-phase. Part IV (The Interconnect Stack) covers the copper and grid equipment that carry the pulses these devices switch. Part V (The On-Package Delivery Stack) covers the 48V-to-0.8V last-mile conversion at the accelerator itself. Part VI (The Modular Datacenter Stack) covers the modular civil and electrical infrastructure needed to build hyperscale AI campuses on the schedule the sector requires. Same method throughout: name the players, size the layers, mark the chokepoints, adopt bottom-up TAM, and be explicit about what is published, derived and estimated.


For readers who want the underlying physics, manufacturing process flows and device cross-sections, see the technical companion: The Wide-Bandgap Stack: Technical Companion.




Series. The AI Power Chain (six parts). Part I: The Capacitor StackPart I-A: Technical CompanionPart II: The Wide-Bandgap StackPart II-A: Technical CompanionPart III: The Thermal StackPart III-A: Technical CompanionPart IV: The Interconnect StackPart IV-A: Technical CompanionPart V: The On-Package Delivery StackPart V-A: Technical CompanionPart VI: The Modular Datacenter StackPart VI-A: Technical Companion.


Methodology. This analysis is independent and based entirely on public information. Market sizes and shares are approximate, compiled from public research (TrendForce, Yole Group, Mordor Intelligence, IDTechEx-adjacent sources, Intel Market Research, Persistence, MarketsandMarkets), company disclosures and SEC filings (Wolfspeed, Navitas, Infineon, onsemi), and trade press (Semiconductor Today, eeNews Europe, Compound Semiconductor, DigiTimes, Woodside Capital Partners, The Diligence Stack) through July 2026. Share figures of differing vintages are labelled as such; where sources conflict, ranges are given. Company names are rendered as text, not trademarked logo assets. Nothing here is investment advice.


© 2026 Adi Kumar · Power & Digital Infrastructure · Corrections and disagreements welcome, that is what the piece is for.











The AI Power Chain: series companions

 * The AI Power Chain: Vendor Screen. The vendor map across every layer
 * Pricing Under Scarcity. Where premium capture is compounding
 * The Services Inversion. Why services now command product-like multiples







Frequently asked


What are the key semiconductor innovations enabling high-efficiency 800V DC?

Silicon carbide (SiC) MOSFETs and gallium nitride (GaN) HEMTs are the enabling innovations. SiC handles the 800V bus-level switching (Wolfspeed, Infineon, ST, onsemi, ROHM). GaN handles the higher-frequency point-of-load down-conversion (Navitas, EPC, Infineon post-GaN-Systems). Together they enable 30-50% loss reduction over silicon IGBTs at these voltage classes.


Which power semiconductor vendors offer the best protection features for AI data centres?

For AI data centre 800V DC applications: Infineon leads on integrated gate-drive protection, Wolfspeed on high-voltage-class SiC (post-emergence from Ch.11), ST on cost-optimised automotive-grade parts crossed over to industrial, and Navitas on GaN with integrated driver + protection. Choice depends on voltage class and switching frequency. See the vendor screen for scoring.


What is onsemi's wide-bandgap strategy in AI servers?

onsemi supplies SiC MOSFETs and modules for the 800V DC bus + rack-conversion stages, plus GaN HEMTs (via the Qorvo SiC JFET acquisition and Transphorm-adjacent lineage). Its share is meaningful in the 650V-1200V power semiconductor space but trails Infineon and ST at the higher voltage classes.


==============================================================================
# The Thermal Stack: How Heat Became the Binding Constraint on AI Compute
URL: https://adikumar.co/the-thermal-stack/
Published: 2026-08-01
Summary: AI data centre thermal management stack: DLC + immersion + facility CDU + coolant chemistry. Every watt in a GPU comes back as heat at 600kW/rack.
==============================================================================

The AI Power Chain series · Part 3 of 15








Glossary of terms used



CDU
Coolant Distribution Unit. Thermal system component distributing chilled coolant to racks or direct-to-chip cold plates.
EBITDA
Earnings Before Interest, Tax, Depreciation, and Amortisation. The most commonly referenced operating-earnings metric in M&A pricing.
GPU
Graphics Processing Unit. The compute silicon at the centre of AI workloads.
GaN
Gallium Nitride. Wide-bandgap semiconductor material used in RF and power applications.
SiC
Silicon Carbide. Wide-bandgap semiconductor material used in high-voltage power electronics.
TAM
Total Addressable Market. The maximum revenue opportunity available if a product served every potential customer segment.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.









ADI KUMAR · POWER & DIGITAL INFRASTRUCTUREAUGUST 2026 · ~45 MIN READ · 11 EXHIBITS

The AI Power Chain · Part III of IV
Market deep dive · Data center thermal management



The Thermal Stack: How Heat Became the Binding Constraint on AI Compute


Every watt delivered to a GPU comes back out as heat, and at 600kW per rack there is nowhere for it to go but water. A layer-by-layer analysis from the die to the dry cooler: the $9.5bn acquisitions, the coupling oligopoly that most models miss, the fluid chemistry stranded by a single corporate decision, and why the most defensible margins in the stack sit in components costing a hundred dollars.



By Adi KumarIndependent analysisViews are my own



Executive summary. Nine findings.

 1. Liquid is no longer optional. NVIDIA's VR200 compute and switch trays are fanless; rack airflow requirements drop ~80% while coolant flow roughly doubles versus GB300. Microsoft has confirmed all future Maia deployments are liquid-default. The air-cooled hyperscale training cluster is over.
 2. The market is ~$5.5-6.8B in 2026 growing 18-26% depending on scope, inside a broader ~$18.5B data center cooling market. Cooling content per rack is rising ~12% per generation while power content rises ~32%.
 3. The chokepoint is a coupling. A GB200 rack uses 100+ universal quick disconnects. Stäubli, CPC and Parker held 80%+ of the Chinese UQD market as recently as 2024, and fewer than fifteen firms worldwide mass-produce complete units. Western UQDs sell at RMB 80-120 against RMB 30-50 domestic. That price umbrella is now under direct attack.
 4. PFAS regulation stranded two-phase immersion. 3M's PFAS exit ended Novec and Fluorinert production in 2025, collapsing the supply chain for two-phase. Single-phase held ~81% of immersion in 2024 and is gaining. The ECHA restriction opinion lands end-2026.
 5. Consolidation has been aggressive and expensive. Eaton / Boyd Thermal at $9.5bn (22.5x EBITDA), Schneider / Motivair (~$850M), Vertiv / PurgeRite (~$1bn). Cooling assets are clearing at multiples that assume the AI buildout does not stop.
 6. Profit pools sit at the extremes. Couplings, TIMs and fluids (the cheapest components) hold the best margins because they are qualification-gated and failure-critical. CDUs and cold plates are where volume lives and where Taiwanese and Chinese competition compresses hardest.
 7. The frontier is moving inside the package. Microchannel lids, then microchannels etched into silicon itself, then backside liquid cooling. TSMC is integrating microchannel cooling into its 3DFabric platform. The cold plate's job is migrating onto the die.
 8. NVIDIA is tightening its grip on the thermal supply chain, standardising designs and squeezing supplier margins. That is the dominant structural risk to every independent vendor in this map.
 9. Bottom-up TAM. Roughly $5.5-6.5B in 2026 rising to $22-30B by 2030 in the base case, driven more by attach rate and content escalation than by facility count.



W/cm²Stage 0 · The thesis

The AI Power Chain · six essays, one physical arc


The series walks a single physical path. It begins at the medium-voltage utility bus at the site fence, steps down through the substation and switchgear, arrives at the datacenter rack where 800V DC is stabilised by the capacitor stack, is converted by silicon-carbide switches to 48V, is distributed across the rack by copper busbars and whips, is stepped down again by multi-phase controllers on the accelerator board to 0.8V, and finally routed through the on-package power delivery network to a transistor gate drawing over 2,000 amperesWaste heat from every conversion stage is removed by the thermal stack. The whole thing is packaged inside a factory-modular building because there aren't enough electricians to build it stick-frame. Six essays. One 800V → 0.8V staircase.


 1. Part I. The Capacitor Stack 800VDC at the rack
 2. Part II. The Wide-Bandgap Stack SiC and GaN conversion
 3. Part III. The Thermal Stack Removing the waste heat (you are here)
 4. Part IV. The Interconnect Stack Busbars and whips
 5. Part V. The On-Package Delivery Stack 48V to 0.8V
 6. Part VI. The Modular Datacenter Stack How the building gets built



$9.5BEaton's acquisition of Boyd Thermal at 22.5x 2026E EBITDA
~$56kEstimated liquid cooling cost per Vera Rubin NVL144 rack, up 17% vs NVL72


The forcing function: rack power beyond air's reach
kW per rack across accelerator generations, with the practical air-cooling ceiling marked

[A1] NVIDIA platform disclosures and supply-chain analysis (Ming-Chi Kuo, 2026). Air-cooling ceiling of ~30-50kW/rack is a widely cited practical limit, not a hard physical one. See Annex §A1.

MAPThe framework · Eight layers, one heat path



The thermal value chain, layer by layer


The power chain divides by voltage. The wide-bandgap chain divides by process step. The thermal chain divides by position along the heat pathThe economics invert as you travel it: the components nearest the die are the smallest, cheapest and most defensible; the components nearest the atmosphere are the largest, most expensive, and most commoditised. Read the map with that in mind.



The data center thermal management market map
Named players by layer · leaders highlighted · August 2026

L1In-package & die-level coolingMicrochannel lids, backside cooling, in-silicon channels · the frontier

TSMC TW · 3DFabric integrationNVIDIA US · MCL/MCCP spec ownerIntel USASE TW · packaging
IMEC BE · researchJetCool US · microconvectiveETH Zürich / Corintis CHSMIC-adjacent CN



L2Thermal interface materialsGreases, PCMs, metal and liquid-metal TIMs · ~$4.9B market, top five concentrated

Honeywell US · PTM7950Shin-Etsu JP · X-23Henkel DE · BergquistDow US · DOWSIL TCIndium Corp US · metal/LM TIM
3M USParker Chomerics USResonac JP · sheetsFuji Polymer JPMomentive US
Laird (DuPont) USDenka · Sekisui JPT-Global TWGrafTech US · graphite



L3Cold platesSkived-fin, microchannel, stamped · the volume layer

Boyd Thermal (Eaton) US · 5M+ units shippedAVC TWCoolIT CA/USAuras TWCooler Master TW
Jentech Precision TWSunon TWDelta Electronics TWNidec JPVertiv US
Cotran New Material CNEnvicool CNJetCool USRittal DEFeilong · Hi-Tech CN



L4Quick disconnects (UQD/UQDB)100+ per GB200 rack · zero-drip, 5,000+ cycles · <15 firms mass-produce complete units

Stäubli CHCPC (Dover) USParker Hannifin US
Jonhon Optronic CN · NVIDIA-certifiedLuxshare CN · NVL72 supply chainEasyflyer CN
Nitta JPWalther-Präzision DEColder Products lineage USLotes · Bizlink TW



L5Manifolds, hoses, in-rack plumbingFlow uniformity, pressure drop, leak prevention

Boyd Thermal (Eaton) USParker USEaton hose/fittings US/IEAVC · Auras TW
nVent US/IERittal DEEnvicool · Shenling CNSwagelok US



L6Coolant distribution unitsRack, row and facility CDUs · top five ≈65% of rack-based revenue

Vertiv US · ~11% of liquid coolingSchneider (Motivair) FR/USnVent US/IEEnvicool CNCoolIT CA/US
Boyd Thermal (Eaton) USDelta Electronics TWNidec JPRittal · Stulz DEDCX PL · 8MW CDU
Chilldyne US · negative pressureCoolcentric · Nortek USMunters · Airedale SE/UKSupermicro USKehua Data CN



L7Dielectric fluids and coolantsImmersion fluids, water-glycol chemistry, inhibitors · no vendor above ~12% share

Shell UK/NL · GTLExxonMobil USCastrol (bp) UKChemours US · Opteon
Syensqo (Galden) BEAGC (Asahiklin) JPDow USDynalene USEngineered Fluids US
Standard Fluids US · ex-3M teamENEOS JP3M US · EXITED 2025TotalEnergies FR



L8Immersion systems and facility heat rejectionTanks, chillers, dry coolers, adiabatic · the commoditised end

Submer ESGRC USIceotope UKLiquidStack USAsperitas NLZutaCore US/IL
Vertiv USSchneider FRJohnson Controls US/IETrane US/IECarrier US
Daikin JPStulz · Munters · Airedale EUAlfa Laval SE · heat exchangersModine USBaltimore Aircoil US



United StatesEuropeJapanChinaTaiwanKoreaShaded tiles = category leaders


How to read this map. Shares quoted are approximate, drawn from public research of differing vintages and scopes. Treat them as positional, not precise. Company names are set as text, not logos. Two structural features stand out. Taiwan's dominance of cold plate manufacturing is a direct inheritance of the notebook thermal-module industry. And the same handful of Western industrial groups (Vertiv, Schneider, Eaton, nVent, Parker) now appear at four or five layers each. Vertical integration across the heat path is the defining strategic behaviour of this market, and it has been bought rather than built.
L1In-package & die-level





In brief


Above 100 kW per rack, direct-to-chip liquid cooling is the only viable AI data centre thermal architecture. The thermal stack has four layers: cold plate / immersion at the chip, coolant distribution units (CDU) for loop separation, facility heat rejection, and chemistry supply chain. CoolIT leads DLC by design-win share; adjacent M&A (Ecolab acquiring CoolIT, Eaton acquiring Boyd) is reshaping vendor concentration.








Cooling moves inside the package


The hot end is where the technology is moving fastest. The conventional heat path runs die → TIM → integrated heat spreader (lid) → TIM → cold plate. Every one of those interfaces adds thermal resistance, and at 1,000 W/cm² the interfaces dominate the budget, which is why the current wave of engineering effort is aimed at removing them.


The first step is the microchannel lid (MCL)Instead of a flat copper lid conducting heat sideways into a cold plate, you etch micron-scale channels directly into the underside of the package lid and flow coolant through them. NVIDIA is reported to be introducing both MCL and a microchannel cold plate (MCCP) for the Rubin generation, and the design shift has already reshuffled Taiwan's thermal supply chain. Auras, AVC, Cooler Master and Jentech Precision are competing directly for the work, with the equity volatility that implies. Rubin also reverses GB300's architectural direction. Where GB300 used six independent small cold plates per compute tray (better targeted cooling, but a proliferation of adapters and manifolds and a correspondingly higher leak risk), Rubin returns to a single large cold plate per tray with microchannel regions over each GPU, formed by laser welding. Fewer joints, fewer leak paths, higher local flow velocity.



Where the thermal resistance sits
Indicative share of total junction-to-coolant thermal resistance by interface, conventional lidded package

[A8] Author's indicative breakdown from published thermal-stack analyses and vendor literature. Proportions vary substantially by package, TIM grade and cold plate design. The point is directional: the two interface layers together rival the cold plate itself, which is why deleting them is where the engineering effort is going. See Annex §A8.




The second step is more radical: put the channels in the silicon. TSMC is integrating microchannel liquid cooling into its 3DFabric platform (CoWoS, SoIC), exploring in-package direct liquid cooling and chip-level direct cooling. The endgame is backside liquid cooling: coolant flowing through structures etched into the back of the die itself, eliminating the TIM and the lid entirely. Research groups (IMEC, ETH Zürich and its Corintis spinout, JetCool's microconvective approach) have demonstrated variants; the manufacturing challenge is sealing a fluid path inside a package that must also survive thermal cycling and warpage for years.


The commercial implication: if cooling migrates into the package, the value migrates to the foundry and the OSAT, and the cold plate becomes a simpler, cheaper distribution manifold. That is a structural threat to the entire L3 layer on a five-to-ten year view, and it mirrors what Part II described happening in power delivery, where backside power delivery pulls capacitance into the package. The die is eating its own infrastructure, in both directions.



L2Thermal interface materials



Which thermal interface materials are the highest-margin layer?


Between the die and whatever removes its heat sits a layer of material typically less than 50 microns thick. It is high-leverage and rarely discussed: get it wrong and the most expensive silicon in the world throttles.


The market is roughly $4.6-4.9B in 2025, growing 10-12% annually, and is concentrated among Honeywell, 3M, Henkel, Parker Hannifin and Shin-Etsu, with Dow, Indium Corporation, Resonac, Momentive and Fuji Polymer holding significant positions. What AI has done to it is an ASP story rather than a volume story. Standard commodity greases run $15-30 per kilogram. The high-performance variants qualified for AI accelerators (Shin-Etsu's X-23 series, Dow's TC-5026, Honeywell's PTM7950 phase-change material) run $150-400 per kilogramThat order-of-magnitude spread within a single material category is described by market analysts as the largest ASP driver in the TIM market during the AI server cycle.


Three sub-segments matter for the AI thesis. Phase-change materials (Honeywell PTM7950 being the reference part) solve the "pump-out" failure mode where thermal cycling gradually extrudes grease out of the interface. That matters when a training cluster power-cycles continuously for years. Metal and liquid-metal TIMs deliver the highest conductivity available: Indium Corporation's Heat-Spring pure-indium products and its indium-silver solder TIMs, plus gallium-based phase-change and polymer-liquid-metal pastes. Liquid metal TIMs are under 5% of TIM volume but command 5-10x the unit price of conventional greases, and their revenue contribution is correspondingly disproportionate. The catch is gallium's corrosivity toward aluminium and its handling requirements, the same gallium supply-chain exposure flagged in Part II. Thermal sheets and films for advanced packaging are where Resonac invested ¥15bn (~$100M) in 2024 specifically for AI semiconductor applications.


The diligence point on TIMs. This is a qualification-gated business with tiny bills of materials and large consequences of failure. That combination is the structural profile that produces durable margin. A TIM qualified into an NVIDIA reference design is specified for the life of the platform, and the switching cost is a full thermal requalification. Look here, not at cold plates, for defensible economics at the hot end of the stack.

L3Cold plates



Cold plates: the volume layer


The cold plate is the workhorse: a copper block with an engineered internal flow path, clamped to the package, moving heat into water. It is also where the money is most visible and the competition hardest.


Manufacturing splits into three approaches. Skiving shaves ultra-thin fins (down to 0.05mm) from a solid copper block, producing very high surface area with no joints, and is the premium method. Microchannel designs etch or machine micron-scale channels for maximum surface area and flow disturbance, claiming 15-30% better thermal performance than conventional designs, and are the direction NVIDIA is pushing. Stamping is the cost-optimised route, using formed aluminium or copper-aluminium hybrids to eliminate CNC time. Metal 3D printing is increasingly entering the frame for geometries that cannot be machined.


Competitive structure is global. AVC and CoolIT lead on revenue, on the strength of vertical integration and hyperscaler relationships. Boyd Thermal, acquired by Eaton in a $9.5bn transaction that closed in March 2026, has shipped over five million liquid cold plates to hyperscalers, backed by four decades of design heritage and 100% in-line leak and flow testing, with production on three continents. Chinese manufacturers including Cotran New Material captured roughly 35% of volume sales as early as 2024 through copper-aluminium hybrid designs and supply-chain localisation.



The Taiwanese thermal ODMs deserve their own paragraph, because the framing that treats them as commodity volume players is wrong. Auras, Cooler Master, Jentech Precision, Sunon and Delta Electronics bring the notebook-cooling industry's manufacturing depth, and increasingly they are moving up the value chain from cold plate manufacturing into full CDU integration, manifold design and rack-level thermal solutions. Auras and AVC in particular hold design-in relationships with NVIDIA that go beyond spec-taking into co-development on the Rubin generation's microchannel architecture. Delta Electronics' broader power and cooling franchise already competes head-to-head with Vertiv, Schneider and Boyd on rack-level integration. The historic assumption that Western majors capture the integration margin and Taiwanese houses capture the commoditised part-manufacture margin is holding today but eroding steadily. On a five-year view, expect Taiwanese ODMs to sit further up the value chain than the market map currently reflects.


The economics are less attractive than the growth rate suggests. Cold plates are a machined-metal business with real material content, exposed to copper prices, and NVIDIA has been tightening its grip on the Rubin cooling supply chain, standardising designs and squeezing supplier margins. Estimated liquid cooling cost per Vera Rubin NVL144 rack is around $55,700, up 17% on the NVL72, on a rack whose total price is reported at $7.8M or more. Cooling material costs are rising ~12% per generation against ~32% for power content. Cooling is growing more slowly than power as a share of rack BOM, which is not the story most cooling investors think they are buying.



The NVIDIA monopsony point deserves sharper treatment than the profit-pool thesis usually gets. When a single buyer specifies the reference design, mandates dual-sourcing on the parts it can, and dictates the target BOM cost, historical margin patterns for the parts on that specification collapse fast. Cold plates are the layer most exposed to this dynamic: NVIDIA's MCL/MCCP spec is now common across Auras, AVC, Cooler Master and Jentech, and their gross margins on Rubin-generation cold plate revenue are visibly compressing against pre-Rubin baselines. The layers that escape this pressure are the ones where NVIDIA cannot easily specify a second source: TIMs (qualification-gated), UQDs (three-vendor oligopoly with high leak-consequence), fluids (chemistry-gated). The profit-pool thesis in this piece needs to be read with that filter: it applies where qualification barriers exceed NVIDIA's leverage, and does not apply cleanly to cold plates.


Content escalation: cooling grows, power grows faster
Reported per-rack content changes, Vera Rubin generation vs prior, %

[A2] TrendForce / Commercial Times reporting (Jun 2026): power module content +32% (~$76k/rack), cooling material content +12%; DigiTimes (Nov 2025): NVL144 liquid cooling cost ~$55,710, +17% vs NVL72. Different bases; see Annex §A2.


L4Quick disconnects



Why are quick disconnects the DLC supply chokepoint?


A universal quick disconnect (UQD) is a coupling that lets a server be pulled from a rack without draining the loop or dripping onto energised electronics. It costs on the order of a hundred dollars. A GB200 rack uses more than a hundred of them.


The engineering requirement is unforgiving: zero-drip disconnection, 5,000+ mating cycles, low pressure drop, high flow, and seal materials (fluororubber or silicone) that stay stable without ageing. The poppet design uses a spring-plus-seal double-sealing arrangement that releases pressure before opening flow. Precision tolerances run to 0.05mmThe consequence of failure is not degraded performance but water on live hardware, which is why this component is qualification-gated to an extreme degree and why NVIDIA runs its own certification programme.


The result is one of the tightest chokepoints in the AI supply chain. Stäubli, CPC (part of Dover) and Parker Hannifin held over 80% of the Chinese UQD market as recently as 2024, which is inside China, against domestic competition, and gives a sense of how hard the qualification barrier is. Worldwide, fewer than fifteen firms design and mass-produce complete UQD units, while over two hundred factories make components for them. The pricing reflects the scarcity: Western UQDs sell at roughly RMB 80-120 against RMB 30-50 for domestic Chinese equivalents at volume, with injection-mould tooling around RMB 800,000 per programme.


That umbrella is now under direct attack. Jonhon Optronic achieved NVIDIA UQD certification and began batch shipping in 2024, adding MQD certification in 2025. Luxshare gained UQD certification in 2025 and entered the NVL72 supply chain. Fewer than ten Chinese firms hold NVIDIA certification, but the Western share of the Chinese market has been reported falling monthly since the second half of 2025. The moat is certification, and certification is a wasting asset. The pattern is the one Part II documented in silicon carbide substrates: a Western technical lead, a Chinese cost position 50-60% below it, qualification as the only real barrier, and a share curve that bends once the first certifications land.




The UQD price umbrella
Reported volume unit pricing, RMB per universal quick disconnect

[A3] Tianxia Gongchang Research, "China Data Center Liquid Cooling 2026" (Jun 2026). Chinese market pricing; Western pricing outside China may differ. See Annex §A3.


A GB200 rack contains more than a hundred couplings, any one of which can drip water onto energised silicon. Three companies hold roughly eighty per cent of the market, and the moat is certification.The chokepoint thesis




The buyer's arithmetic: $500 saving vs $3M+ liability


The price gap between Western and Chinese UQDs looks attractive on a bill-of-materials line but small once the risk math is done. A hyperscale rack has ~100 UQDs. At the reported RMB 50-70 per unit price gap, substituting all of them for Chinese-domestic parts saves on the order of $500-800 per rackThe catastrophic case is a single UQD failure spilling coolant onto a $3M+ AI rack, plus the collateral to adjacent equipment, plus the training-run interruption, plus the reputational and contract consequences. A single failure event on a Chinese-substituted fleet of ten thousand racks makes back all the substitution savings and more.




UQD substitution: savings vs risk-adjusted liability, hyperscale fleet
Cumulative $/rack savings from Chinese UQD substitution vs expected liability from projected leak failure rate, 3-year horizon

Author's sensitivity. Assumes 100 UQDs per rack, RMB 60 gap per unit at ~7.2 RMB/USD, ~$500/rack cumulative saving over three years of parts refresh. Liability scenarios shown at three leak-failure-rate assumptions (1e-8 per mate-cycle for premium tier-1 Western UQDs, 1e-7 for tier-2, 1e-6 for early-qualification Chinese parts). Even at the middle scenario, expected liability across a 10,000-rack fleet exceeds the substitution savings within the first year. Tier-1 hyperscalers stay with Stäubli/CPC/Parker despite the price gap on exactly that arithmetic.




The right way to read the market essay's UQD chokepoint claim is therefore not "Chinese UQDs will steadily take share once qualified" but "Chinese UQDs will take share only on the fraction of fleet where hyperscalers accept the incremental leak risk". That fraction may grow as Chinese leak-test data accumulates, but the risk asymmetry keeps the premium buyers on Western parts far longer than the price gap alone predicts.


L6Coolant distribution units



CDUs: the interface layer


The CDU is the interface between the technology cooling loop (the clean, filtered, tightly-controlled water touching the servers) and the facility water loop. It pumps, filters, monitors and regulates, and it determines whether a liquid-cooled hall behaves like infrastructure or like a science project.


Market structure is consolidating fast. The top five in rack-based CDUs accounted for about 65% of revenue as early as 2023, and the names have been rearranged by acquisition since: Vertiv (the liquid cooling market leader overall at ~11.3% share in 2025), Schneider Electric, which bought Motivair for roughly $850M, then nVent, China's Envicool, CoolIT, Boyd Thermal (Eaton), Delta, Nidec, Rittal, Stulz and Poland's DCX, which introduced an 8MW CDU in January 2026. Beyond the top five, Schneider, Vertiv, Rittal, Stulz and Boyd collectively held about 35% of the total data center liquid cooling market in 2025.


The interesting structural detail sits in the Chinese supply base, and it generalises. Plate heat exchangers and variable-speed pumps are mature commodity pools: roughly fifty plate-exchanger factories producing 1,000+ units annually, and 80+ pump makers, though only about a dozen reach high-end specifications above 600 L/min. But fewer than twenty firms integrate both into a deployable CDU. The components are commoditised; the integration, controls, and reliability engineering are not. That is the same pattern as the couplings and the TIMs, and it is the most useful heuristic in this market: in thermal, the parts are cheap and the qualification is expensive.


One design note operators should know. Chilldyne's negative-pressure architecture, which runs the loop below atmospheric pressure so a breach draws air in rather than pushing water out, is the clearest example of engineering around the industry's real fear. Leak risk, not thermal performance, is what keeps liquid cooling out of conservative facilities.




Steady state is the easy case: dynamic transients are the hard one


The steady-state heat transfer discussed throughout this piece assumes the pump is running, the loop is flowing, and the cold plate is at operating temperature. The interesting engineering happens in the transient cases: pump failure, coolant loss, unexpected thermal runaway from a compute-side event, or a rapid change in load that outpaces the thermal control loop. Under a pump failure at full load, an AI accelerator die reaches thermal limits in on the order of 200 millisecondsA conventional CDU control loop reacts in seconds. That gap is bridged by three things in a well-designed system: the thermal mass of the cold plate itself (delaying junction-temperature rise by 100-300ms), a fast-acting thermal telemetry signal to the accelerator (triggering clock throttling within milliseconds), and a graceful power-down handshake between the CDU and the compute chassis.


Any system that lacks all three fails hard: pump loss → coolant loss of momentum → localised film boiling → thermal runaway → die failure. Dynamic pump redundancy (N+1 or 2N inside every CDU), fast thermal interlock signalling, and coordinated power-and-cooling shutdown are procurement-level requirements at hyperscale sites for that reason, not nice-to-haves. Any cold plate or CDU vendor whose data sheet does not disclose transient behaviour under pump-fail conditions is selling incomplete equipment.



L7Fluids



How did PFAS regulation strand two-phase immersion cooling?


On 20 December 2022, 3M announced it would exit all PFAS manufacturing by the end of 2025. The company was facing over 4,000 lawsuits and a $12.5bn settlement with more than 11,000 US public water systems over PFAS contamination in drinking water, finally approved in March 2024, with payments stretching over thirteen years. The decision had nothing to do with data centers. Its effect on them was total.


Novec 7100, Novec 649 and Fluorinert FC-72 were the qualified fluids for two-phase immersion cooling. They are gone. Last orders were accepted 31 March 2025, manufacturing lines shut by year end. Two-phase immersion, the technology that could push PUE below 1.05 and absorb heat loads that make direct-to-chip systems struggle, lost its supply base to a decision made for unrelated litigation reasons. There is no cleaner example in the sector of legal risk stranding a working technology roadmap.


The aftermath sorted the market. Single-phase immersion vendors were largely unaffected because they never needed fluorinated chemistry: Submer runs synthetic hydrocarbons co-engineered with Castrol, GRC uses hydrocarbon dielectrics, Asperitas partnered with Shell on a gas-to-liquid formulation, and Iceotope uses PFAS-free chemistry throughout. Single-phase held 80.9% of the data center immersion market in 2024 and has been gaining since. Two-phase vendors took the damage. ZutaCore, whose HyperCool is a two-phase direct-on-chip system, moved to Chemours' Opteon SF33 as an interim measure and committed to a fully PFAS-free two-phase fluid.



What the PFAS exit did to immersion cooling
Single-phase vs two-phase share of the data center immersion market, 2024

[A9] The Cooling Report (Mar 2026), citing 2024 market data: single-phase 80.9%. The two-phase residual is the arithmetic complement. Share has continued shifting toward single-phase since; treat 2024 as the last clean datapoint before the supply shock fully worked through.




The regulatory story is not over, and most operators underestimate that. The available fluorinated alternatives, Syensqo's Galden PFPE line and AGC's Asahiklin, are technically competent (Galden HT exceeds 40 kV/mm dielectric strength) but are also PFAS under the definitions regulators are using. The EU's universal PFAS restriction proposal does not carve out PFPEs. ECHA is evaluating a restriction covering over 10,000 substances, with updated proposals in August 2025 expanding exemptions from 26 to 74, including longer transition periods for certain heat-transfer applications. ECHA's final opinions are expected by end-2026, with European Commission legislation anticipated in early 2027. Operators qualifying PFPE fluids today are deferring regulatory risk, not eliminating it.


The immersion fluids market itself is projected at $2.79B in 2026 rising to ~$4.49B by 2031, with synthetic hydrocarbons leading at ~37% share after Intel certified Shell and ExxonMobil formulations. No vendor holds more than about 12% share. This is the least concentrated layer in the thermal stack, and the one where an energy major's refining scale plus a chipmaker endorsement is the winning combination. Mineral oils at $2-5 per litre keep crypto miners loyal. PFAS-free hydrofluoroethers from Chemours, Solvay and Dow are the fastest-growing chemistry at ~10% CAGR.



The second-order lesson. Two-phase immersion did not lose on thermal performance, cost, or engineering. It lost because its single-source fluid supplier faced unrelated litigation. Any technology whose viability depends on one chemical from one manufacturer carries a risk that no thermal model captures. Ask of every cooling architecture: what is the sole-source input, and what unrelated exposure does its maker carry?




Fluid maintenance is a real operational cost, not a rounding error


Water-glycol loops look simple on the datasheet: typically 25% propylene glycol / 75% water, with a corrosion inhibitor package, a biocide, pH held between 8.5 and 9.5 to passivate copper, and conductivity below 10 µS/cm to prevent galvanic corrosion. That is the steady-state description, and it hides the operational reality of running the loop at 40-50°C coolant temperatures (and rising, as 45°C hot-water designs go mainstream). Inhibitor packages deplete faster in warm loops. Biological fouling accelerates. Oxygen ingress through fitting seals feeds bacterial growth. Chinese hyperscaler operators have reported inhibitor re-dose intervals compressing from annual to quarterly under 45°C service, with attendant chemistry testing, filter changes and system downtime. Corrosion inhibitor depletion is invisible in daily monitoring but shows up in copper concentration and pH drift; catching it late means partial system flush and refill, tens of thousands of dollars per event.


Biofouling is the harder problem. Warm water with organic-based inhibitors is an excellent environment for biofilm growth on plate heat exchangers, cold plate microchannels and pump impellers. A biofilm 0.5 mm thick on a plate exchanger cuts effective heat transfer coefficient 20-40%, which shows up first as rising server inlet temperatures under load. Chemical biocide dosing addresses it but requires re-dosing; UV or ozone treatment is more effective but adds equipment. The Vertiv-PurgeRite deal thesis (services annuity) reads correctly on that basis: fluid maintenance carries a real per-rack operational cost that scales with fleet size and rack density, well past the rounding-error threshold, and the vendors capturing that cost line are the ones with the strongest long-term positions in the profit pool.




Coolant maintenance cost per rack per year, indicative
Recurring TCO components for a 210kW liquid-cooled rack, USD per year

Author's estimate from vendor pricing and hyperscaler operational disclosures. Fluid replacement ~$300 (partial refresh every 2-3 years amortised). Chemistry testing + inhibitor re-dose ~$400/yr. Biocide dosing ~$200/yr. Filter changes ~$150/yr. Preventive plate exchanger inspection + cleaning ~$500 amortised. Contingency for unplanned events ~$400. Total ~$2k/rack/yr recurring. Multiplied by a 10,000-rack fleet, that is ~$20M/yr of services revenue that used to be classified as "facility overhead" and is now increasingly captured by dedicated services vendors.


$Market size · TAM · Profit pools



Sizing the market and locating the profits



The published range


Data center liquid cooling estimates for 2026 cluster more tightly than the capacitor or WBG markets, which is a sign of a maturing category. $5.7B (Persistence), $6.0B (GMI), $6.77B (Mordor), with 2025 bases of $4.8-6.65B and forward CAGRs of 18-26% producing 2031-2033 outcomes between $18.8B and $29.5B. The broader data center cooling market including air is roughly $18.5B in 2026 at 12.3% CAGR, within which liquid grows at ~24.7% and immersion at ~32.4%. Cold plate liquid cooling holds more than 55% of the liquid segment in 2026 (~$3.1B); hyperscale is 76.4% of end-user demand.




Liquid cooling market: five houses, one direction
2026 base and outer-year forecast, USD billions. Scope and horizon differ by source (log scale).

[A4] Persistence (2026→2033), GMI (2026→2035), Mordor (2026→2031), Grand View (2025→2033), MarketsandMarkets. Horizons differ. See Annex §A4.



Inside the liquid cooling market
2026 segmentation by cooling type and by end user, % of liquid cooling revenue

[A4] Persistence Market Research (2026): cold plate >55% of liquid cooling (~$3.1B), hyperscale 76.4% of end users, solutions 63.5 to 74.5% of revenue depending on source. Immersion and spray shares are the arithmetic residual.





Bottom-up TAM


Build it from racks, as in the previous essays. Content per rack: ~$55,700 of liquid cooling cost for a Vera Rubin NVL144, which at ~210kW is roughly $265k per MW of IT load, about 2.5x the ~$104k/MW of power semiconductor content computed in Part II. Volumes: Ming-Chi Kuo estimates 5,000-7,000 VR200 racks shipping in H2 2026 alone; AI IT-load additions were modelled in Part II at ~7-9GW in 2026 rising to ~14-18GW by 2030. Attach rate: effectively 100% for new AI racks (the trays are fanless, and there is no air-cooled option), plus a facility-side retrofit stream and the services tail.


At ~$265k/MW against ~8GW of 2026 AI IT-load additions, rack-level liquid cooling content computes to roughly $2.1BPublished market estimates run $5.7-6.8B. The difference is facility-side equipment (CDUs at row and facility scale, heat rejection, piping), non-AI liquid cooling, and services. That reconciliation gives confidence in both numbers. Extending TAM per author's bottom-up model: base case ~$6B in 2026 to ~$25B by 2030, bear ~$15B, bull ~$34B. The dominant sensitivities are AI rack shipment volumes and the pace at which cooling content per rack rises, the latter running at only ~12% per generation, materially slower than power. Full workings in Annex §A5.




Bottom-up TAM: data center liquid cooling
USD billions, full scope (equipment + services), 2026-2030E. Author's model.

[A5] Author's model: content per MW × AI IT-load additions ÷ rack-share of total spend, plus facility retrofit and services. Cross-checked against five published 2026 estimates. Illustrative; inputs in Annex §A5.



The series in one chart: infrastructure content per MW
Estimated infrastructure content per MW of AI IT load, USD thousands, by subsystem covered across this series

[A11] Cooling ~$265k/MW derived in §A5. Power semiconductors ~$104k/MW from Part II (Infineon disclosure, corroborated by onsemi). Capacitive energy storage ~$15 to 25k/MW derived in Part I. Different subsystem boundaries and different levels of confidence; the comparison is indicative of relative scale, not a complete BOM. See Annex §A11.





Profit pools: margin runs opposite to component size


In the power chain, margin concentrated in qualified devices and hard-to-make materials. In thermal, the pattern is sharper: margin runs opposite to component size and cost.


 * TIMs (~35-45%): tiny BOM, high failure consequence, platform-life qualification, order-of-magnitude ASP spread between commodity and AI-qualified grades. The best-defended layer in the stack.
 * Quick disconnects, also 35-45%. Hundred-dollar parts protected by NVIDIA certification, 0.05mm tolerances and a five-thousand-cycle reliability requirement. Under attack from Chinese entrants, but currently pricing at a 2-3x premium.
 * Fluids sit at roughly 25-35%. Chemistry moats and chipmaker certification, but fragmented (no vendor above ~12%) and exposed to regulatory shock.
 * CDUs run 20-30%. Integration and controls are the moat; components are commoditised. Compressing as Chinese integrators and Taiwanese ODMs scale.
 * Cold plates, the volume layer, run 15-25%. Exposed to copper, NVIDIA's design standardisation and supplier margin squeeze, and Chinese hybrid-material competition at ~35% of volume.
 * Facility heat rejection lands at 15-20%. Mature HVAC economics, mature HVAC margins.
 * Services (~25-35% and structurally rising): the underrated layer. One forecast sees the revenue mix inverting from 80% solutions / 20% services today to 20/80 within four to five yearsIf even directionally right, that inversion is the most consequential shift in the sector.




Margin by layer: the small components win
Indicative gross margin bands, 2026. Plotted against typical component cost per rack.

[A6] Author's estimates from company disclosures, industrial-component analogues, and the qualification-barrier structure described in each section. Directional illustration, not company-specific guidance. See Annex §A6.

M&ACapital · Consolidation



What did buyers pay in the cooling consolidation wave, and why?


Thermal management has seen the most aggressive M&A of any layer in the AI infrastructure stack, and the multiples tell you what the buyers believe.



DealValueRationaleRead

Eaton / Boyd Thermal
Nov 2025 announced, closed Mar 2026$9.5bn, 22.5x 2026E adj. EBITDA; $1.7bn sales of which $1.5bn liquid cooling; 5,000+ employees"Grid-to-chip": bolting a cooling franchise onto a power franchise; seller was Goldman Sachs Asset ManagementThe anchor transaction. RBC's Deane Dray noted Eaton went "all-in" rather than dipping a toe like peers. At 22.5x, it prices continuous AI buildout. Fourth data center deal in a year for the acquirer, after Fibrebond ($1.4bn), Resilient Power and Ultra PCS ($1.55bn).
Schneider Electric / Motivair~$850MCDU and high-density cooling capability inside an existing data center franchiseThe template Eaton then executed at 11x the size.
Vertiv / PurgeRite~$1bnLiquid cooling services: flushing, commissioning, loop maintenanceThe most strategically interesting deal on this list if the 80/20 services inversion is real. Vertiv bought the annuity, not the box.
Goldman Sachs AM / Boyd (prior ownership)UndisclosedPE ownership through the growth phase, exited into strategic demandThe sponsor playbook executed to spec: buy an industrial thermal business, ride the AI attach rate, sell to a strategic at 22.5x.




What is still on the board


 * CoolIT Systems. Private, revenue-leading in cold plates and CDUs, the largest remaining independent of scale. Logical buyers: any industrial that missed Boyd and Motivair, or an ODM seeking vertical integration.
 * The Taiwanese thermal-module houses (Auras, Jentech, Sunon) are publicly listed, manufacturing-deep, exposed to NVIDIA's margin squeeze. Consolidation candidates or acquirers depending on how the microchannel transition resolves.
 * Immersion specialists (Submer, Iceotope, LiquidStack, GRC, ZutaCore). Venture-funded, sub-scale, with more than $2.8bn of venture and strategic capital having entered data center cooling between 2023 and 2026. Single-phase players are the healthier cohort post-PFAS.
 * Among UQD independents, Stäubli is family-controlled and unlikely to sell; CPC sits inside Dover; Parker is a strategic in its own right. This layer consolidates by Chinese entry rather than by acquisition.
 * Fluid formulators such as Dynalene, Engineered Fluids and Standard Fluids are small, specialised and strategically interesting to energy majors wanting a data center chemistry position.


Valuation framing


Boyd at 22.5x forward EBITDA on $1.7bn of revenue sets the anchor for every subsequent conversation in this sector. For context, a diversified industrial trades at 12-16x; a pure-play growth asset with AI attach might justify high teens; 22.5x embeds both continued hypergrowth and significant synergy capture. The implied logic is that grid-to-chip integration lets the buyer sell power and cooling as one qualified system and defend both against point competitors, a thesis that is coherent but unproven, and which the market will test the first time a hyperscaler decides it prefers best-of-breed to bundling. The risk in this consolidation wave sits in the pricing rather than the asset quality: the assets were priced at the moment of maximum visible growth in a market whose content-per-rack is growing more slowly than the racks themselves.




What cooling assets cleared at
EV/EBITDA multiples: the Boyd transaction against typical comparables

[A10] Boyd at 22.5x is disclosed (Eaton, Nov 2025). The diversified-industrial and growth-industrial bands are the author's estimates of prevailing 2025 to 26 trading ranges, shown for context only, not derived from a specific comparables set. See Annex §A10.


TECHDisruption watch list



What could change the shape of this market



VectorWhat it isWhoImpact

Microchannel lids and cold platesMicron-scale channels in the lid or plate; laser-welded, high local velocityNVIDIA (spec), Auras, AVC, Cooler Master, JentechAlready happening in Rubin. Reshuffles Taiwanese share; raises precision-manufacturing barriers.
In-silicon and backside liquid coolingCoolant channels etched into the die or package substrateTSMC (3DFabric), IMEC, Corintis, JetCoolThe structural threat: value migrates to foundry/OSAT, cold plate simplifies to a manifold.
45°C hot-water operationRubin designed for 45°C inlet waterNVIDIA, facility vendorsChiller-free heat rejection almost anywhere. Kills a chunk of the mechanical plant TAM while enabling heat reuse.
Heat reuse / district heatingSelling waste heat rather than rejecting itNordic and German operators, municipal utilitiesTurns a cost centre into revenue. Regulatory tailwinds in the EU. 45°C water makes it far more practical.
PFAS-free two-phase fluidsRestoring two-phase immersion without regulatory exposureChemours, Dow, Syensqo, ZutaCore, Standard FluidsWould revive a stranded technology. ECHA's end-2026 opinion sets the clock.
Negative-pressure loopsSub-atmospheric operation so breaches draw air in, not water outChilldyneAttacks the industry's real blocker (leak fear) rather than its stated one (thermal performance).
Metal 3D printing for cold platesFlow geometries impossible to machine or skiveThermal vendors, AM service bureausPerformance step-change at premium cost. Watch for qualification in flagship platforms.
Two-phase direct-to-chipBoiling coolant at the die without full immersionZutaCore, LiquidStack, AccelsiusHighest heat flux capability in a rack-compatible form factor. Fluid-supply-constrained.
Liquid-metal and no-TIM architecturesEliminating the interface layer entirelyIndium, foundries, NVIDIARemoves the highest-margin small component from the stack. Watch as a threat to L2.
OCP cold plate standardisationOpen qualification specs for cold plates and loopsOCP working groupsDouble-edged: accelerates adoption, then commoditises the hardware.





RISKTrade-offs · Water · Standards



Trade-offs, and the argument against



The water question


Liquid cooling's sustainability story is good on energy. Immersion cuts consumption by up to ~48% in some deployments, and PUE of 1.02-1.2 is routine against 1.4-1.6 for legacy air. Closed-loop direct-to-chip systems consume very little water in operation. But the facility-side heat rejection is where water actually goes, and evaporative cooling towers in water-stressed regions are the industry's most politically exposed practice. The 45°C hot-water specification is the strongest mitigation available, because it enables dry coolers in climates that previously required evaporation. Operators who cannot articulate their water strategy alongside their PUE will find the permitting environment increasingly hostile. This is a licence-to-operate issue, not an ESG reporting one.



Standards and the leak problem


The instruments that matter here: OCP's cold plate development and qualification specifications (which set the industry's shared reliability language), ASHRAE TC 9.9 liquid cooling classes (W17. W45, with the Rubin generation targeting the warm end), IEC and UL requirements for fluid-carrying electrical equipment, and the fluid-side regulatory perimeter described above. The gap that concerns me most is the absence of a mature, industry-wide leak-detection and response standard proportionate to the consequence. A hall with a hundred thousand couplings has a failure-rate arithmetic problem that no single vendor's testing regime addresses. The sector's biggest self-inflicted risk in the 2027 ramp is a high-profile water-on-silicon incident at a flagship site.



Who wins, who is pressured


 * Winners here are the qualification-gated small components (TIM leaders, the UQD trio while certification holds), the services franchises (Vertiv-PurgeRite's logic), the integrated grid-to-chip platforms if bundling proves out (Eaton-Boyd, Schneider-Motivair, Vertiv), Taiwanese manufacturing depth in microchannel precision, and the foundries if cooling migrates into the package.
 * Under pressure: cold plate specialists facing NVIDIA's design standardisation and margin squeeze; two-phase immersion vendors still awaiting a compliant fluid; ODMs whose margins compress under consignment models; anyone whose moat is manufacturing rather than qualification; and legacy air-cooling franchises whose installed base is now a stranded-asset conversation.
 * The bear case in one line. Cooling content per rack is growing at ~12% per generation while rack shipments carry all the volatility of hyperscaler capex. This is a lower-beta, lower-content growth story than power, bought at higher multiples.




Ten leading indicators to watch


 1.  ECHA's final PFAS opinion (end-2026) and the Commission's early-2027 legislation. Decides whether two-phase immersion has a future in Europe.
 2.  Chinese UQD share gains at NVIDIA-certified suppliers. Jonhon and Luxshare volumes are the umbrella's tear-rate.
 3.  Microchannel lid adoption in shipping Rubin systems. Confirms or delays the migration of cooling into the package.
 4.  TSMC 3DFabric cooling integration announcements. The structural threat to the cold plate layer, on a five-year fuse.
 5.  Services as a share of cooling vendor revenue. Tests the 80/20 inversion thesis; Vertiv's disclosures post-PurgeRite are the cleanest read.
 6.  Eaton's disclosed Boyd Thermal performance against the 22.5x entry multiple. The sector's valuation benchmark, marked to market quarterly.
 7.  VR200 rack shipment counts vs the 5,000-7,000 H2 2026 estimate. The volume assumption underpinning every model here.
 8.  Any major leak incident at a flagship liquid-cooled site. The tail risk that would reset adoption timelines.
 9.  Water permitting refusals or restrictions in major DC markets. The licence-to-operate constraint.
 10. Heat reuse contracts signed with municipal utilities. Turns thermal from cost centre to revenue line and would re-rate the facility layer.




Where this lands


Heat is the constraint that cannot be engineered away, only relocated. The industry has spent three years relocating it: out of the air and into water, out of the room and into the rack, and now out of the rack and into the package itself. Each relocation created a market, and each market has consolidated faster than the last. A $9.5bn cooling acquisition at 22.5x earnings would have been unthinkable in 2022. The pattern this series keeps finding holds here too. The headline assets (cold plates, CDUs, chillers) carry the most revenue and the least defensible economics. The margin sits in a coupling that costs a hundred dollars, a film of grease fifty microns thick, and a fluid whose supply depends on the litigation exposure of a chemical company that never intended to be in this business. In a thermal chain, as in a power chain, the value sits wherever qualification is expensive and the part is cheap.


Part IV of this series covers the interconnect stack: transformers, switchgear, grain-oriented electrical steel, and the grid equipment lead times that now gate the AI buildout. Part V covers the on-package delivery stack (48V to 0.8V at the transistor gate). Part VI covers the modular datacenter stack (how the building actually gets built).


ANNEXSources · Calculations · Assumptions



Annex: how every derived number was built


As in Part II, every figure is either published (traceable to a named source), derived (calculated from published inputs, arithmetic shown), or estimated (my judgement, labelled). This annex covers the latter two.




§A1. Rack power and the air-cooling ceiling


Status: published, with one estimated line. Published: Hopper ~40kW/rack; GB200/Blackwell 120-130kW; VR200 NVL72 190-230kW; Rubin Ultra "Kyber" ~600kW (NVIDIA disclosures and Ming-Chi Kuo supply-chain analysis, 2026). Per-GPU: ~1.8kW baseline, up to ~2.3kW in Max-P (Kuo, 2026). Rubin CPX variant in a separate ~370kW rack. Estimated: the "practical air-cooling ceiling" plotted at 30-50kW is a widely cited industry rule of thumb, not a physical limit or a single-source figure. High-airflow designs exceed it at poor economics.




§A2. Content escalation


Status: published, two different bases. Do not combine them. (1) TrendForce/Commercial Times (Jun 2026): on the Vera Rubin rack (unit price reported ~$7.8M), power module content +32% to ~$76k per rack, cooling material content +12%. (2) DigiTimes (Nov 2025): NVL144 liquid cooling cost ~$55,710, +17% vs the NVL72 model. These are different metrics (materials vs total cooling cost) from different sources at different dates. The chart plots them side by side as reported and does not reconcile them. Rack price estimates vary across sources ($5-8.8M depending on configuration and SOCAMM procurement assumptions), which is itself a caution on all per-rack percentage claims.




§A3. UQD pricing and market structure


Status: published, single source. All figures from Tianxia Gongchang Research, "China Data Center Liquid Cooling 2026" (Jun 2026): 100+ UQDs per GB200 rack; Stäubli, CPC and Parker held 80%+ of the Chinese UQD market in 2024 with share falling monthly from H2 2025; overseas units RMB 80-120 vs domestic RMB 30-50 at volume; injection-mould tooling ~RMB 800,000; precision tolerance 0.05mm; 5,000+ cycle requirement; fewer than 15 firms mass-produce complete units against 200+ component makers; fewer than 10 Chinese firms hold NVIDIA certification. Jonhon UQD certification and batch shipping 2024, MQD 2025; Luxshare UQD certification 2025 entering NVL72 supply chain. This is a single-source section. The pricing and share figures are specific to the Chinese market and I have not found independent Western corroboration. Treat the structural claims as solid and the precise numbers as one researcher's view.




§A4. Market size reconciliation


Status: published, differing scopes and horizons. Liquid cooling 2026: Persistence $5.7B→$29.2B by 2033 (26.4% CAGR); GMI $6.0B (from $4.8B in 2025)→$27.1B by 2035 (18.2%); Mordor $6.77B (from $5.52B in 2025)→$18.79B by 2031 (22.65%); Grand View $6.65B in 2025→$29.46B by 2033 (20.1%); Transpire →$28.18B by 2033. Broader data center cooling: $18.5B in 2026→$52.54B by 2035 at 12.3% (Evolvance), with liquid at 24.7% and immersion at 32.4% CAGR. Segment detail: cold plate >55% of liquid in 2026 (~$3.1B), hyperscale 76.4% of end users, solutions 63.5-74.5% of revenue depending on source, North America 32-35.6% regional share. Immersion fluids: $2.79B (2026)→$4.49B (2031), synthetic hydrocarbons 37.12% share in 2025, no vendor above ~12%. CDU: North America ~$607M of a market where APAC holds ~38% (~$721M), implying a global CDU market near $1.9B in 2026; top five rack-CDU players ~65% of revenue (2023). The chart plots each house's own base and outer year on a log axis; horizons differ from 2031 to 2035 and are labelled.




§A5. Bottom-up TAM model


Status: derived. Workings below.


Content per MW. NVL144 liquid cooling cost ~$55,710 (DigiTimes) at ~210kW rack power → $55,710 ÷ 0.21MW = ~$265k per MW of IT loadSanity check against Part II's power semiconductor content of ~$104k/MW: cooling is ~2.5x power semiconductor content per MW, which is consistent with cooling being a heavier, more materials-intensive subsystem.


Volumes. AI IT-load additions taken from Part II's model: ~7-9GW in 2026 (derived there from Gartner's ~132GW global DC power demand, JLL's ~18-20GW annual capacity additions, an assumed 55-65% AI share of new build, and ~70% IT-load-to-facility-power ratio), rising to ~14-18GW by 2030. Independent cross-check: Kuo's 5,000-7,000 VR200 racks in H2 2026 × ~210kW ≈ 1.05-1.47GW from that platform alone in half a year, consistent with several GW annually across all AI platforms.


Rack-level computation. 8GW × $265k/MW = ~$2.1B of rack-level liquid cooling content in 2026.


Reconciliation to published totals. Published 2026 liquid cooling markets of $5.7-6.8B are ~2.7-3.2x my rack-level figure. The gap comprises facility-side equipment (row and facility CDUs, heat exchangers, piping, pumps, heat rejection), non-AI liquid cooling (HPC, enterprise, telecom), design/installation/maintenance services (26-36% of revenue depending on source), and retrofit into existing halls. I therefore apply a rack-content-to-total-market multiplier of ~2.9x, giving ~$6.1B for 2026: within the published range and derived independently of it.




TAM bridge: rack-level content to full market
Waterfall reconciliation, 2026 USD billions

Starts with rack-level content (~$2.1B derived above), adds facility-side CDU + heat rejection, non-AI liquid cooling into HPC/enterprise, services (design + install + maintenance), and retrofit into existing halls. Lands within the published $5.7-6.8B range.



Cooling market by layer and geography · Marimekko
Column width = share of total cooling revenue by layer. Row segments = approximate regional supply origin.



Cold plateTaiwan · 40%
China · 30%
US · 20%
EU · 10%



CDUUS · 45%
EU · 25%
China · 20%
Other · 10%



FacilityUS · 35%
EU · 30%
Japan · 20%
Other · 15%



TIMUS · 35%
Japan · 35%
EU · 20%
Other · 10%



UQDUS · 45%
EU · 35%
CN · 20%



FluidsEU · 35%
US · 30%
Japan · 20%
Other · 15%



ServicesUS · 60%
EU · 25%
APAC · 15%



← wider columns = larger layer share of cooling revenue
row segments = regional supply origin

Layer widths approximate, from Persistence + Mordor 2026 estimates. Regional splits are directional decompositions from the market map, not published data. Illustrative reading device, not a source.



Scenarios to 2030. Base: AI IT-load additions to ~15GW/yr; content per MW rising ~12% per platform generation (per §A2) against partial offset from cold plate cost-down and Chinese competition, net ~+5%/yr; multiplier stable at 2.9x → ~$25B by 2030Bear: capex air-pocket halving addition growth, content per MW flat as commoditisation bites, multiplier falling to 2.5x as services pricing compresses → ~$15B. Bull: Kyber and megawatt racks on schedule, in-package cooling raising content, retrofit wave, services inversion lifting the multiplier to 3.3x → ~$34B. Sensitivities in order: (1) AI rack shipment volumes; (2) the rack-to-total multiplier, which is my least defensible input; (3) content per MW, which is growing more slowly than most cooling bulls assume.




§A6. Profit pool estimates


Status: estimated. No source publishes gross margin by thermal value-chain layer. The bands shown (TIM 35-45%, UQD 35-45%, fluids 25-35%, CDU 20-30%, cold plates 15-25%, facility 15-20%, services 25-35%) are my judgement, constructed from the qualification-barrier structure documented in each section; the observed 2-3x Western/Chinese price premium in UQDs and the 5-10x ASP spread in AI-grade vs commodity TIMs, both of which imply pricing power; the machined-metal cost structure and copper exposure of cold plates plus reported NVIDIA margin squeeze; mature HVAC margins for facility equipment; and industrial-services analogues. These illustrate a structural argument, not any company's actual margins. The Boyd transaction gives one triangulation point: $1.7bn revenue at a $9.5bn/22.5x EBITDA valuation implies ~$422M EBITDA, i.e. ~25% EBITDA margin for a mixed cold-plate/CDU/manifold portfolio, consistent with the mid-range bands above.




§A8. Thermal resistance breakdown


Status: estimated. The split shown (die→TIM1 ~8%, TIM1 ~26%, lid spreading ~14%, TIM2 ~22%, cold plate→coolant ~30%) is my indicative construction from published thermal-stack analyses and vendor literature, not a measurement of any specific package. Actual proportions vary widely with die size, lid material, TIM grade and bond line thickness, cold plate design and flow rate. The chart is included to support one directional claim well established in the literature: the two TIM layers together account for a substantial share of junction-to-coolant resistance, comparable to the cold plate itself, which is why eliminating them (microchannel lids, then backside cooling) is where the engineering effort is going. Do not use these numbers for a thermal budget.




§A9. Immersion single-phase vs two-phase


Status: published, one figure. Single-phase held 80.9% of the data center immersion market in 2024 (The Cooling Report, Mar 2026). The 19.1% two-phase figure is the arithmetic complement, not separately published. 2024 predates the full working-through of the Novec supply shock, so the current split is likely more skewed toward single-phase. I have used 2024 because it is the last clean datapoint I could source.




§A10. M&A multiples


Status: one disclosed figure, two estimated bands. Boyd at 22.5x 2026E adjusted EBITDA is disclosed by Eaton (Nov 2025). The "growth industrial with AI attach" (~16-19x) and "diversified industrial" (~12-16x) bands are my estimates of prevailing 2025 to 26 trading ranges, included purely to give the 22.5x figure context. They are not derived from a specific comparables screen and should not be treated as one. Implied EBITDA on the Boyd deal: $9.5bn ÷ 22.5 ≈ $422M on $1.7bn revenue, i.e. ~25% EBITDA margin.




§A11. Content per MW across the series


Status: derived, mixed confidence. Cooling ~$265k/MW: derived in §A5 from the NVL144 liquid cooling cost of ~$55,710 at ~210kW. Power semiconductors ~$104k/MW: from Part II, based on Infineon's disclosed $12-15k per 130kW rack, independently corroborated by onsemi's ~$100k per 1MW rack. Capacitive energy storage ~$15-25k/MW: derived in Part I from an assumed 250-500Wh buffer per 600kW rack at $8-15/Wh cell cost with a 1.8-2.2x system multiple; this is the least certain of the three by a wide margin. Important caveat. These three subsystems have different boundaries: cooling includes facility-side content in some framings, power semiconductors exclude passives and magnetics, capacitive storage covers only the energy shelf. The comparison shows relative order of magnitude, not shares of a common BOM. They do not sum to a rack cost.




§A7. Other published figures used


 * Eaton / Boyd Thermal. $9.5bn, 22.5x 2026E adjusted EBITDA, $1.7bn forecast 2026 sales of which $1.5bn liquid cooling, 5,000+ employees, sites in North America/Asia/Europe, seller Goldman Sachs Asset Management, announced 3 Nov 2025, completed March 2026, accretive to adjusted EPS in year two; RBC (Deane Dray) "all-in" commentary; prior Eaton deals Fibrebond $1.4bn, Ultra PCS $1.55bn, Resilient Power undisclosed. Sources: Eaton press releases, Reuters, Bloomberg, DCD (2025 to 26).
 * Other M&A. Schneider / Motivair ~$850M (Persistence); Vertiv / PurgeRite ~$1bn (Reuters); >$2.8bn venture and strategic capital into DC cooling 2023 to 2026 (Evolvance).
 * Market leadership. Vertiv >11.3% of liquid cooling in 2025; top five (Schneider, Vertiv, Rittal, Stulz, Boyd) ~35% collectively (GMI, Jan 2026).
 * 3M/PFAS. Announcement 20 Dec 2022; exit complete end-2025; last Novec orders 31 Mar 2025; >4,000 lawsuits and $12.5bn settlement with 11,000+ US public water systems, final approval Mar 2024, payments over 13 years; Novec 7100/649 and Fluorinert FC-72 discontinued; ECHA restriction covers >10,000 substances, Aug 2025 update expanded exemptions 26 to 74, final opinions expected end-2026, Commission legislation early 2027; Galden HT >40 kV/mm dielectric strength but PFPE is within PFAS scope; single-phase 80.9% of immersion market 2024; Submer/Castrol, Asperitas/Shell, GRC hydrocarbon, ZutaCore→Opteon SF33. Sources: The Cooling Report, DCD, Alliance Chemical, Schneider blog, Precision Lubrication (2026).
 * TIM. Market $4.6B (2024, GMI) to $4.9B (2025, Research Nester), 10.1-12.4% CAGR; top five Honeywell, 3M, Henkel, Parker Hannifin, Shin-Etsu; AI-grade greases $150-400/kg vs $15-30/kg commodity (Shin-Etsu X-23, DOWSIL TC-5026); LMTIM <5% of volume at 5-10x unit price; Resonac ¥15bn (~$100M) investment Mar 2024; Indium Heat-Spring HSx Mar 2025; Henkel Bergquist TGF 10000 (10 W/mK) Dec 2025; Dow DOWSIL TC-3120 May 2026.
 * Cold plates. Boyd 5M+ cold plates delivered to hyperscalers (Sep 2025); AVC and CoolIT lead revenue; Chinese manufacturers including Cotran ~35% of volume sales 2024; microchannel designs claim 15-30% thermal improvement; skived fins to 0.05mm; Rubin single large cold plate per tray with laser-welded microchannel regions vs GB300's six small plates. Sources: Intel Market Research, Global Semi Research, Boyd/Businesswire, DigiTimes.
 * Platform facts. VR200 trays fanless, coolant flow ~2x GB300, rack airflow −80% (Kuo, 2026); Vera Rubin designed for 45°C hot water (CES 2026); Microsoft Maia liquid-default (Dec 2025); nVent modular CDUs at SC25 (Nov 2025); DCX 8MW CDU (Jan 2026); Flex modular rack CDU (Sep 2025); heat flux 800-1,000 W/cm² (SemiVision/TSMC analysis, Jan 2026); NVIDIA MCL/MCCP introduction and Taiwanese supplier competition (DigiTimes, Oct 2025 to May 2026); consignment model discussion by Foxconn Q4 2025 and Quanta Q1 2026 calls.
 * Chinese supply base. ~50 plate-exchanger factories at 1,000+ units/yr; 80+ pump makers with ~12 above 600 L/min; <20 firms integrating deployable CDUs; 200+ UQD component factories. Source: Tianxia Gongchang Research (Jun 2026).



A caution on this sector's data. Liquid cooling market estimates are converging (a good sign) but component-level data is thin, and several of the most interesting figures in this piece (UQD pricing, Chinese supplier structure, per-rack cooling costs) rest on single sources, often Taiwanese or Chinese trade research that Western readers cannot easily verify. I have flagged each instance. Where I filled a gap with judgement, it is labelled an estimate. This annex is a starting map for diligence, not a substitute for primary work.

SERIESRelated in the AI Power Chain



Where else to read


For readers who want the underlying physics, heat-transfer mechanics and cross-sections of the components discussed here, see the technical companion: The Thermal Stack: Technical Companion.


Prior pieces in the series: The Capacitor Stack (Part I) covers hybrid supercapacitors, EDLCs and MLCCs. The Wide-Bandgap Stack (Part II) covers SiC and GaN power semiconductors, whose efficiency partly determines how much heat this cooling infrastructure must remove. Its technical companion is The Wide-Bandgap Stack: Technical Companion.




Series. The AI Power Chain (six parts). Part I: The Capacitor StackPart I-A: Technical CompanionPart II: The Wide-Bandgap StackPart II-A: Technical CompanionPart III: The Thermal StackPart III-A: Technical CompanionPart IV: The Interconnect StackPart IV-A: Technical CompanionPart V: The On-Package Delivery StackPart V-A: Technical CompanionPart VI: The Modular Datacenter StackPart VI-A: Technical Companion.


Methodology. This analysis is independent and based entirely on public information. Market sizes and shares are approximate, compiled from public research (Mordor Intelligence, Global Market Insights, Persistence Market Research, Grand View Research, MarketsandMarkets, Evolvance, Intel Market Research, QYResearch, Tianxia Gongchang Research), company disclosures and press releases, and trade press (DigiTimes, DataCenterDynamics, TrendForce, Reuters, Bloomberg, The Cooling Report, Global Semi Research, SemiVision) through August 2026. Figures marked directional, estimated or single-source should be treated accordingly. Nothing here is investment advice.


© 2026 Adi Kumar · Power & Digital Infrastructure · Corrections and disagreements welcome.











The AI Power Chain: series companions

 * The AI Power Chain: Vendor Screen. The vendor map across every layer
 * Pricing Under Scarcity. Where premium capture is compounding
 * The Services Inversion. Why services now command product-like multiples







Frequently asked


What is the AI data centre thermal stack?

The AI data centre thermal stack has four layers: 1) Cold plate / immersion at the chip, 2) Coolant distribution units (CDU) for facility-loop separation, 3) Facility-side heat rejection (dry coolers, adiabatic, chillers), 4) Chemistry supply chain (single-phase glycol vs two-phase dielectric fluids). Above 100 kW/rack, direct-to-chip liquid cooling is the only viable architecture.


Who are the leading direct-to-chip liquid cooling vendors?

CoolIT is the leading DLC pure-play by design-win share; Motivair (Schneider), Boyd (Eaton), Asetek, and Chilldyne make up the contested tier. Vertiv assembled its cooling capability via capability tuck-ins, Strategic Thermal Labs for cold-plate engineering and PurgeRite for fluid management, rather than a headline DLC acquisition.


==============================================================================
# The Interconnect Stack: Transformers, Switchgear and the Grid Equipment That Gates the AI Buildout
URL: https://adikumar.co/the-interconnect-stack/
Published: 2026-08-01
Summary: Grid interconnect is the binding constraint on AI data centre buildout. Transformer lead times, hyperscaler substation queue, and vendor map through 2030.
==============================================================================

The AI Power Chain series · Part 1 of 15








Glossary of terms used



AEP
American Electric Power. Investor-owned utility operating in 11 US states including Ohio.
CAISO
California Independent System Operator. The grid operator for most of California.
ERCOT
Electric Reliability Council of Texas. Grid operator for most of Texas, operating largely as an electrical island.
EV
Enterprise Value. Purchase price plus debt assumed minus cash. Total value of the business excluding capital structure.
GW
Gigawatt. One thousand megawatts of electrical power.
GaN
Gallium Nitride. Wide-bandgap semiconductor material used in RF and power applications.
MISO
Midcontinent Independent System Operator. Regional grid operator covering 15 US states and Manitoba.
MV
Medium Voltage. Typically 1 kV to 35 kV. The voltage class connecting data centre power distribution to the utility grid.
MVA
Mega-Volt Amperes. Apparent power rating unit used for transformers and switchgear.
MW
Megawatt. Unit of electrical power. A large modern data centre draws tens to hundreds of MW.
PJM
PJM Interconnection LLC. Regional grid operator covering 13 US states plus DC. Includes Ohio, where the NVIDIA/OpenAI campus sits.
SMR
Small Modular Reactor. Nuclear reactor design with output typically below 300 MWe, envisioned for co-located generation.
SiC
Silicon Carbide. Wide-bandgap semiconductor material used in high-voltage power electronics.
TAM
Total Addressable Market. The maximum revenue opportunity available if a product served every potential customer segment.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.









ADI KUMAR · POWER & DIGITAL INFRASTRUCTUREAUGUST 2026 · V1 · ~45 MIN READ

The AI Power Chain · Part IV of VI
Market deep dive · Grid interconnect & interconnect equipment



The Interconnect Stack: Transformers, Switchgear and the Grid Equipment That Gates the AI Buildout


Every gigawatt of AI capex assumes an interconnected grid connection with sufficient MV switchgear, a large power transformer and a set of medium-voltage cables to route the current. In 2026 that assumption no longer holds. LPT lead times have doubled since 2023, GOES supply is concentrated in three producers, US interconnect queues run four years, and hyperscalers have started building their own generation to escape the wait. A layer-by-layer analysis from grain-oriented electrical steel to grid queue reform.



By Adi Kumar
Independent analysis
Views are my own



Executive summary. Nine findings.

 1. Grid interconnect is now the binding constraint on AI, not compute, not cooling, not chips. New large-load interconnect timelines in PJM, ERCOT, MISO and CAISO commonly run 4-7 years from application to energisation. Every hyperscaler ordering GW-class capacity in 2026 is planning around a wait that could stretch into the next decade.
 2. Large power transformer lead times have doubled since 2023. A 138/33 kV or 230/34.5 kV substation-class LPT that took 50-60 weeks pre-pandemic now takes 120-150+ weeks in North America. Global order books at Hitachi Energy, Siemens Energy and GE Vernova have been reported past record highs, with the three sitting on multi-year backlogs.
 3. GOES (grain-oriented electrical steel) is the underlying materials chokepoint. Roughly five producers globally (JFE Steel, Nippon Steel, POSCO's spinoff, ArcelorMittal / Cogent, NLMK), with China's Baosteel and TISCO adding capacity. GOES is the transformer core material and its manufacturing takes 12-18 months to expand. It is the Kuraray of the interconnect stack.
 4. Behind-the-meter power has moved from workaround to strategy. Amazon at Susquehanna, Meta at Clinton (Constellation), Microsoft at Three Mile Island, Google's Kairos SMR MOU, plus dozens of gas-turbine and Bloom Energy fuel cell deals. What used to be "on-site generation for redundancy" is now "on-site generation because we cannot wait for the grid".
 5. Copper demand from AI adds a materials layer stress that most models miss. Hyperscale racks use roughly 3-5× the copper per MW of a legacy enterprise data centre. On top of EV and renewables demand, AI could add 1-2 million tonnes/year of incremental copper consumption by 2030, into a supply base already forecasting deficits.
 6. The M&A window for interconnect assets has closed. Hitachi bought ABB Power Grids in 2020. GE broke out Vernova in 2024. Cameron International, Emerson process, PGE, Sensata: the big consolidations are done. What remains is bolt-ons and JVs, not platform deals.
 7. Digital and asset-managed transformers are the emerging profit lever. Adding sensor packages, oil-quality monitoring and thermal telemetry lets OEMs sell a service tail alongside the box. Hitachi Energy, Siemens Energy and Mitsubishi Electric all now have dedicated digital-transformer product lines. Software attach is small revenue today, high-margin, and structurally growing.
 8. Chinese equipment is qualifying into the AI supply chain for cables and medium-voltage switchgear, less so for large power transformers. TBEA and Baoding Tianwei have made international sales, but hyperscalers' provenance rules and utility qualification cycles are the real barriers.
 9. Bottom-up TAM. AI data centre interconnect capex sits at roughly $12-18B in 2026 (~$150-200k per MW of IT load, on ~8GW of new build) rising to $45-65B by 2030 as the buildout scales. Transformers, switchgear and cables dominate; behind-the-meter generation is the emerging category.



GWStage 0 · The thesis

The AI Power Chain · six essays, one physical arc


The series walks a single physical path. It begins at the medium-voltage utility bus at the site fence, steps down through the substation and switchgear, arrives at the datacenter rack where 800V DC is stabilised by the capacitor stack, is converted by silicon-carbide switches to 48V, is distributed across the rack by copper busbars and whips, is stepped down again by multi-phase controllers on the accelerator board to 0.8V, and finally routed through the on-package power delivery network to a transistor gate drawing over 2,000 amperesWaste heat from every conversion stage is removed by the thermal stack. The whole thing is packaged inside a factory-modular building because there aren't enough electricians to build it stick-frame. Six essays. One 800V → 0.8V staircase.


 1. Part I. The Capacitor Stack 800VDC at the rack
 2. Part II. The Wide-Bandgap Stack SiC and GaN conversion
 3. Part III. The Thermal Stack Removing the waste heat
 4. Part IV. The Interconnect Stack Busbars and whips (you are here)
 5. Part V. The On-Package Delivery Stack 48V to 0.8V
 6. Part VI. The Modular Datacenter Stack How the building gets built



$150k+Estimated interconnect capex per MW of AI IT load, 2026 base
~3-5×Copper per MW in a hyperscale AI rack vs a legacy enterprise data centre


Interconnect lead time expansion, by equipment class
Reported delivery lead times, weeks, 2019 vs 2026, indicative

Compiled from vendor guidance, EPRI briefings, DOE Loan Programs Office reports, and utility engineering-firm surveys through mid-2026. Ranges are wide because lead time depends on capacity class, voltage, customer priority and manufacturer relationship. The direction and magnitude of the change are consistent across sources.



The mismatch between AI capex velocity and interconnect physics has driven the behind-the-meter generation wave, the acceleration of small modular reactor pilots, hyperscaler direct multi-year procurement contracts with transformer OEMs, and a redrawing of the map of where data centres can actually be built. The utility-equipment supply chain, which had been low-growth and stable for two decades, is now one of the two or three most strategically constrained resources in AI infrastructure.


MAPThe framework · Eight layers, one grid path



The interconnect value chain, layer by layer


The capacitor stack divided by timescale. The wide-bandgap stack divided by process step. The thermal stack divided by heat-path position. The interconnect stack divides by voltage class and physical position between utility and rackEach layer has a different competitive structure, a different concentration profile, and a different lead-time exposure. The market map below names the players at every layer. Two things stand out before reading it: the Japanese and Korean depth in materials and heavy equipment (GOES, LPT, HV cables) and the fact that the same four or five industrial groups (Hitachi Energy, Siemens Energy, Schneider Electric, ABB, Eaton) appear at multiple layers. Interconnect is a heavily integrated business, and it has been that way for decades.



The AI data centre interconnect market map
Named players by layer · leaders highlighted · August 2026

L1Grain-oriented electrical steel (GOES)Silicon-steel laminations for transformer cores · roughly 5 global producers

Nippon Steel JPJFE Steel JPCogent (Tata) UK/INArcelorMittal FRPOSCO KR
NLMK / DK RU/DKBaosteel CNTISCO / Taiyuan CNAK Steel (Cleveland-Cliffs) US
Wuhan Iron & Steel CNSteel Authority of India IN



L2Large power transformers (LPT)≥100 MVA units for substations · concentrated in three global majors + regional players

Hitachi Energy (ex-ABB) CH/JPSiemens Energy DEGE Vernova USMitsubishi Electric JPHyundai Electric KR
HHI (LS Group) KRToshiba Energy Systems JPWEG BRProlec-GE (Xignux) MXSGB-SMIT DE
TBEA CNBaoding Tianwei CNWanshiba (WEG-China) CNCG Power / Bharat Heavy IN
Delta Star USVirginia Transformer USERMCO / Pennsylvania Transformer US



L3MV / LV switchgearAir-insulated, gas-insulated (SF6), vacuum, dry-air · switchgear for substation and distribution

Schneider Electric FRABB CHEaton US/IESiemens DEMitsubishi Electric JP
Hitachi Energy CH/JPPowell Industries USLegrand (industrial) FRAlstom Grid (ex-GE) FR
Fuji Electric JPLS Electric KRHyundai Electric KRCHINT · Delixi CN
nVent (Erico / Hoffman) US/IEVertiv (integrated) USRittal DE



L4MV / HV cablesXLPE-insulated MV cables · often kilometres per campus · four global majors + Chinese/Korean scale

Prysmian ITNexans FRNKT DKSumitomo Electric JPLS Cable & System KR
Furukawa Electric JPFujikura JPSouthwire USGeneral Cable (Prysmian) US
Hengtong Group CNZTT (Jiangsu Zhongtian) CNFar East Cable CNElka (Cable Coop) PL



L5Copper & conductor materialsMine · smelter · rod & wire · fabricated busbar

Freeport-McMoRan USSouthern Copper (Grupo México) PE/MXBHP · Rio Tinto AU/UKGlencore CH
Codelco CL · stateZijin Mining CN
Aurubis DE · smelterKME IT/DE · fabWieland Group DE · fabFurukawa Electric JP · fab
Southwire US · fabSterlite / Vedanta IN



L6Backup generation · gensetsDiesel and natural-gas standby · years-long lead times · integrated with fuel and controls

Caterpillar USCummins USRolls-Royce mtu DE/UKMitsubishi Heavy Industries JPKohler US
Generac US · industrialYanmar JPMAN Energy Solutions DEKawasaki Heavy · IHI JP
Vertiv (Liebert) US · integrationEaton (Cyclonaire/Fibrebond) US · packaging



L7Behind-the-meter powerGas turbines · fuel cells · SMR (emerging) · alternative sources when grid queues bite

Bloom Energy US · SOFCGE Vernova (LM6000/LM2500) US · gas turbineSolar Turbines (Caterpillar) USSiemens Energy (SGT) DEMitsubishi Heavy Industries JP
FuelCell Energy US · MCFCSFC Energy DE
NuScale US · SMRX-Energy US · SMRKairos Power US · SMROklo US · SMR
GE Hitachi (BWRX-300) US/JP · SMRRolls-Royce SMR UK



L8Grid connection · regulated utilityNot a commercial market · the constraint everything else is trying to route around

PJM Interconnection US · MW EastERCOT US · TexasMISO US · MidwestCAISO US · California
ENTSO-E TSOs EUNational Grid (UK ESO) UKTenneT · 50Hertz DE
Dominion · AEP · Duke US utilitiesPG&E · SCE US · CABerkshire Hathaway Energy US



United StatesEuropeJapanChinaKoreaIndiaShaded tiles = category leaders


How to read this map. The shares and lead times quoted throughout this piece are approximate and drawn from public research of differing vintages and scopes. Treat them as positional, not precise. Two structural features stand out. First, the same handful of Western industrials appear at three or more layers each: Hitachi Energy, Siemens Energy, GE Vernova, Schneider Electric, Eaton, ABB, Mitsubishi Electric. Vertical integration across the transformer / switchgear / cable path is the norm, and it has been for decades. Second, the geography of concentration matters: large power transformers are majority Asian-manufactured (Japanese, Korean, plus rising Chinese and Indian), GOES is majority Japanese, and MV cables are split between three European giants and AsiaThe North American production base is thin and expanding slowly.
L1Grain-oriented electrical steel





In brief


Grid interconnect is the binding constraint on AI data centre buildout. Transformer lead times exceeded 200 weeks in 2025-26 per industry disclosures. Hyperscaler substation queue slots are held years in advance. The vendor stack (Eaton, Vertiv, ABB, Schneider, Prysmian) is consolidating rapidly. Eaton alone has closed Fibrebond ($1.4B, April 2025), Ultra PCS ($1.55B, January 2026) and Boyd Thermal ($9.5B, closed March 2026) per company press releases.








The material chokepoint most models don't see


A transformer core is not a solid iron block. If it were, eddy currents circulating in the metal during AC cycles would dissipate the transformer's rated power as heat within minutes. Cores are built from thin laminations of silicon-alloyed iron in which the crystal grains have been oriented along the direction of magnetic flux. This "grain orientation" is achieved through a specific rolling, annealing and coating process that took decades to develop and is dominated globally by roughly five producers. The resulting material (grain-oriented electrical steel, or GOES) has an order-of-magnitude lower core loss than random-orientation steel, which is what makes practical 100-MVA-and-larger transformers possible at 99%+ efficiency.


GOES production is capital-intensive and slow to scale. A greenfield GOES line takes 3-5 years to design, permit and commission, and roughly 12-24 months to reach qualified output at commercial defect densities. Since the shale-gas and renewables buildout began in the early 2010s, GOES has cycled between periodic shortage and glut. Since 2022 the balance has been squarely on the shortage side, driven by simultaneous demand from grid expansion, renewables integration, EV charging infrastructure and now hyperscale data centres.




GOES supply by producer, indicative
Approximate global GOES production share, 2025 basis. Directional, not audited.

Combined estimates from industry press (Metal Bulletin, SteelOrbis), corporate disclosures and trade press through 2025. Chinese production has grown fastest but is largely consumed domestically. Japanese, Korean and European material dominates the export market that supplies Western LPT builders.





The three grades that matter


GOES splits into three commercial grades: conventional (CGO), high-permeability (HGO or Hi-B), and domain-refinedEach represents a step up in magnetic performance and a corresponding step up in price. Modern high-efficiency transformer designs use HGO or domain-refined material for the core; the very highest-tier designs (railway traction, HVDC valve transformers, hyperscale substation LPT) use domain-refined material with laser-scribing to further reduce core loss. Nippon Steel and JFE dominate the domain-refined grade; Cogent and POSCO offer competitive alternatives at slight cost disadvantage.


The AI-relevant point is that the hyperscale substation transformer sits squarely in the segment where domain-refined GOES matters most. It is the segment where the supply base is most concentrated, where lead times have expanded most, and where price has moved most. This is the "materials layer" chokepoint of the interconnect stack, and it plays the same role that Kuraray's cathode carbon plays in the capacitor essay or Japanese silicon nitride plays in the wide-bandgap essay: small revenue, structural leverage, extreme concentration.



The amorphous metal alternative. Amorphous-metal transformer cores (using Metglas from Hitachi Metals / Proterial, or equivalent products from Advanced Technology & Materials in China) achieve 70-80% lower no-load losses than GOES. They have been commercial since the 1980s for distribution transformers. The reason they have not displaced GOES in large power transformers is that amorphous ribbon is thinner, harder to work, and mechanically fragile compared to grain-oriented laminations. Handling adds cost. As efficiency regulations tighten (EU EcoDesign 2021, US DOE 2024 updates) amorphous is gaining share at the small end. In LPT-scale devices it remains niche, but a supply squeeze on GOES could accelerate qualification.

L2Large power transformers



What is the 200-tonne transformer bottleneck?


The large power transformer is the piece of equipment that steps utility voltage down to something a data centre can use. A hyperscale campus typically has multiple 50-150 MVA units in its main substation, plus smaller distribution transformers throughout. A single LPT weighs 100-400+ tonnes, ships by rail or specialised truck, and is oil-filled, so site delivery and installation is a construction project in itself. Each unit is essentially bespoke: designed to the utility's voltage class, the site's short-circuit rating, the loading profile and the environmental constraints. There is no mass-market SKU.




The market structure and where the concentration sits


Three companies (Hitachi Energy, Siemens Energy, GE Vernova) hold the majority of the global market for large power transformers over 100 MVA. Hitachi Energy inherited ABB's power grids business in a 2020 acquisition and has become the largest single player, with reported multi-year backlogs. Siemens Energy carved out from Siemens AG in 2020 and is the closest Western competitor. GE Vernova spun out of General Electric in 2024 as the dedicated energy business. Mitsubishi Electric and Hyundai Electric round out the top five with strong Asian franchises. Below them, a long tail of regional and specialist producers (WEG, Prolec, SGB-SMIT, Delta Star, Virginia Transformer) serve smaller units and regional markets.


US production capacity is thin and expanding slowly. Prolec-GE, Virginia Transformer, Delta Star, Pennsylvania Transformer and a handful of smaller manufacturers make up the domestic base, which was estimated in the DOE's 2023 supply-chain review at roughly 20% of US LPT demand at typical utility replacement rates. That share is now unable to keep pace with hyperscale-driven demand. Multiple capacity expansions have been announced (Prolec in Monterrey, Hitachi in South Boston Virginia, HHI in Alabama), but a new LPT plant takes 3-5 years to build and qualify. None of the announcements will materially relieve US supply before 2028.




Global LPT market share by manufacturer
Approximate revenue share, ≥100 MVA units, 2024-25 basis

Compiled from Metal Bulletin, T&D World, corporate disclosures. Three global majors + two strong Asian franchises hold roughly 65% of the market; the remainder is fragmented across regional players. Share of new orders in 2025-26 skews further toward the top three, which is the mechanism behind their lengthening backlogs.





The lead time story, in detail


Pre-pandemic, an LPT for a US utility was a 40-60 week engagement. Order placed, engineering complete in ~10 weeks, GOES procured and processed into laminations, core stacked, windings wound, tank assembled, oil-filled, factory tested, shipped. In 2026 that same engagement is quoted at 120-150 weeks by most Western OEMs, with some hyperscale-scale units pushing 200 weeks. Three factors drove the expansion, in roughly the order that matters.


First, demand doubledUtility grid modernisation (regulated capex), renewables interconnection (each wind or solar project needs a step-up transformer), EV charging build-out, industrial reshoring, and now AI data centres have all been ordering LPTs simultaneously, at rates the manufacturing base was never sized for. Global LPT demand has been reported to have doubled between 2020 and 2025.


Second, the supply base did not expand in stepTwo decades of low-single-digit utility spending growth had left OEMs with lean production footprints. Adding a new manufacturing bay costs $100M+ and takes 3-5 years. Announced capacity expansions have been substantial (Hitachi Energy alone has announced roughly €4.5B of grid-equipment capex through 2027) but the ramp curves lag demand by 2-3 years minimum.


Third, materials tightenedGOES supply, described above, is the immediate constraint at the material layer. Transformer oil (mineral or ester), copper wire, insulation paper (kraft pressboard), and specialty bushings have all seen extended lead times through 2024-25. Each is a smaller constraint than GOES but the compounding effect is real.




Hitachi Energy order backlog trajectory
Reported order book / revenue ratio, indicative

Hitachi Energy has publicly reported order backlogs that have tripled since 2020 to record levels by 2025. The ratio of backlog to revenue is the operationally significant number: at 2-3× annual revenue, a company has visibility on multi-year growth but cannot flex quickly. Siemens Energy and GE Vernova report similar patterns.


L3MV / LV switchgear



The equipment that routes and protects


Switchgear is the equipment that connects and disconnects circuits under load, isolates faults, and protects downstream equipment from short circuits. A hyperscale campus has switchgear at every voltage class: MV switchgear (typically 5-38 kV) in the substation and at row-level distribution, LV switchgear (below 1 kV) at the panel-board level. Different physics govern each: MV switchgear must interrupt current at kV levels without sustaining an arc, which is what drives the choice of insulation medium.




MV switchgear insulation media compared
Radar comparison across five criteria: interrupting capacity, environmental profile, footprint, cost, service maturity

Air-insulated (AIS) is the incumbent, well understood, largest footprint. SF6-insulated (GIS) has the smallest footprint but the highest global-warming potential (GWP ~23,500). Vacuum interrupters are increasingly standard for MV breakers. Dry-air / fluoro-nitrile alternatives are emerging as SF6 phase-out accelerates under EU F-gas regulation and analogous US EPA rules.





The SF6 phase-out


Sulphur hexafluoride (SF6) has been the standard insulation gas for high-voltage switchgear since the 1970s. It has excellent dielectric strength (per unit volume, roughly 3× that of air), good arc-quenching properties, and long service life. It is also the most potent greenhouse gas ever regulated, with a GWP roughly 23,500× that of CO₂ and an atmospheric lifetime of 3,200 years. Under the EU F-gas Regulation revised in 2024, SF6 will be prohibited in new MV switchgear from 2028-2030 depending on class, with HV following. US EPA and Californian rules are moving in the same direction.


The industry response has been alternative insulation gases (Hitachi Energy's EconiQ using C4-fluoronitrile, Siemens' Blue GIS using clean air, GE's g3 using C5-fluoroketone) and vacuum interrupters for medium-voltage classes. The alternatives are more expensive, some carry residual environmental concerns (fluoroketones degrade to PFAS in the atmosphere), and each requires qualification against long-established SF6 designs. Data centre operators specifying substations in 2026 face a real design choice between staying with SF6 for immediate delivery and paying a premium for clean-gas alternatives that align with corporate sustainability commitments.



L4MV cables



Kilometres of XLPE per campus


A hyperscale data centre campus typically comprises 5-15 buildings on a site of a few hundred hectares, connected by underground medium-voltage cable runs, rather than sitting in a single structure. The main substation feeds distribution transformers scattered near the buildings, and the distribution transformers feed the racks. Total MV cable per site can run into the tens of kilometres. Every kilometre is engineered for the voltage class (typically 15-35 kV), fault-current rating, ampacity and burial depth.


The dominant cable technology is cross-linked polyethylene (XLPE) insulated cable with a copper or aluminium conductor. XLPE is thermoset (cross-linked during curing so it does not melt at operating temperature), which lets it handle higher continuous current than earlier PVC-insulated designs. Copper conductors are used where ampacity per cross-section matters most; aluminium is used where cost dominates and larger cross-sections are acceptable.


Global MV cable capacity concentrates in three European majors (Prysmian, Nexans, NKT), two Asian franchises (Sumitomo Electric, LS Cable & System), and a rising Chinese cohort (Hengtong, ZTT, Far East Cable). Similar to LPT, capacity expansion is slow: a new XLPE extrusion line takes 2-3 years to qualify. Prysmian and Nexans have both announced major capex programmes for MV and HV cable capacity through 2027, driven by offshore wind, HVDC interconnectors, and data centre demand.



Why aluminium doesn't fully substitute for copper. Aluminium is roughly a third of copper's mass per unit volume, cheaper per tonne, and has adequate conductivity for many applications. In data centre MV cables it is used routinely for the largest cross-sections. But aluminium has three tradeoffs: it is 60% as conductive per unit cross-section as copper (so an aluminium cable of the same ampacity is physically larger), it work-hardens over thermal cycles at connection points (which is why terminations use specialised techniques), and it cannot be used where mechanical flexibility is paramount. For rack-scale wiring, busbars and switchgear internal connections, copper remains dominant. Aluminium substitution is real but bounded.

L5Copper



The commodity underneath everything


Copper is the material that binds the interconnect stack together. Every transformer winding, every busbar, every MV cable, every rack-scale power distribution unit uses copper as the conductor of choice. A hyperscale AI data centre uses roughly 30-50 tonnes of copper per MW of IT load, according to widely cited industry estimates from 2024-25, which is 3-5× the copper intensity of a legacy enterprise data centre. The multiplier comes from higher rack currents (800VDC at ~750A per rack), denser distribution, more redundancy, and higher-ampacity bus systems.




Copper demand from AI data centres in context
Estimated incremental copper demand by end market, million tonnes/year by 2030

Compiled from BHP, Wood Mackenzie, S&P Global Commodity Insights and IEA outlooks through 2025. AI data centres add roughly 1-2 Mtpa of incremental copper demand by 2030 on top of the much larger EV and renewables buildout demand. Cumulative copper deficit forecasts through 2030 range from 6-10 Mtpa without new mine investment.




Global copper mine production is roughly 22 Mtpa (2024). Adding EV, renewables, grid modernisation and AI data centre demand together, most credible forecasts point to a structural deficit through the 2030s absent significant new mine investment. Mine development takes 10-20 years from discovery to first production, and the most attractive undeveloped ore bodies (Chile, Peru, DRC, Mongolia) are increasingly complicated by permitting, community consent and geopolitical risk. Copper is unlikely to run out (recycling handles roughly a third of current supply), but the price signal from the deficit is likely to be substantial, and it will show up in every layer of the interconnect stack.



L6Backup generation



The generators that used to be an afterthought


Every data centre has backup generation. It used to be a routine specification: rated for eight to twelve hours of runtime on onboard diesel, sized for full load, arranged in an N+1 or 2N redundancy configuration, exercised monthly, and forgotten about the rest of the time. In 2026, backup generation is a strategic variable. Diesel gensets have lead times running past two years for the largest units. Natural-gas gensets, favoured for lower emissions and lower fuel storage requirements, run longer. Behind-the-meter gas turbines (below) increasingly displace both.


The three largest genset manufacturers globally are Caterpillar, Cummins and Rolls-Royce mtu in the diesel/gas engine class. Mitsubishi Heavy Industries and Kohler round out the top five. All five have publicly discussed extended lead times through 2024-26 driven by simultaneous demand from data centres, industrial resilience upgrades, and hospital/critical-infrastructure spending. Caterpillar reported record backlogs on its Solar Turbines gas-turbine business in 2025; Cummins similarly on its Cummins Power Generation segment.



The natural-gas transition. Diesel gensets have been the default for decades because diesel is stable in storage and generators are proven. Natural-gas gensets have gained share as environmental and permitting concerns around diesel storage tightened. Gas has three advantages (lower particulate emissions, lower per-kWh CO₂, no on-site fuel storage requirement) and one disadvantage (dependence on a live gas pipeline, which is not necessarily any more resilient than the electric grid the generator is backing up). For "pure standby" duty, diesel remains common. For "long-run continuous during grid absence" duty (which is what behind-the-meter power is), gas dominates.

L7Behind-the-meter power



How are hyperscalers bypassing the grid, one campus at a time?


The behind-the-meter (BTM) power story is the biggest strategic development in the interconnect stack of the past two years. When grid interconnect timelines run four to seven years and hyperscalers want capacity in eighteen months, the arithmetic no longer works. The response has been to develop on-site generation at scale, sized to be the primary or majority power source rather than backup. Three technology classes are being deployed.




Gas turbines


The most mature option. GE Vernova's aeroderivative LM6000 (~45 MW), LM2500 (~35 MW), and stationary 7HA-class (~350 MW); Siemens' SGT-A65 and larger frames; Mitsubishi Heavy's H-class. Aeroderivatives can start in minutes and load-follow, which matches data centre demand well. Combined-cycle configurations (adding a steam turbine that runs off exhaust heat) push efficiency to 60%+. Multiple hyperscale campuses in Texas, Georgia and Virginia have announced or built dedicated gas-turbine plants sized at hundreds of MW.




Fuel cells


Bloom Energy has become the most visible BTM fuel-cell vendor for data centres, with multi-hundred-MW deployments announced with hyperscalers including Equinix and various colocation operators. Bloom's solid-oxide fuel cells run on natural gas (or hydrogen where available), operate at high efficiency (~60% electrical, higher with cogeneration), and can be sited compactly on the data centre campus. FuelCell Energy uses molten-carbonate technology in a similar deployment model. Fuel cells have significant advantages (no combustion emissions, quiet, modular) and one significant disadvantage (installed cost per kW is higher than gas turbines at scale).




Small modular reactors (SMR)


The most-discussed and least-deployed category. As of 2026, no commercial SMR is operating in the United States. Multiple pilot projects are in development: Amazon's investment in and PPA with X-Energy (four-unit HTGR site in Washington), Google's MOU with Kairos Power for a fluoride-salt-cooled unit, Meta's exploration of nuclear PPAs, and various utility-scale SMR programmes (NuScale in Idaho with UAMPS, formerly, since restructured; GE Hitachi BWRX-300 at OPG in Ontario; Rolls-Royce SMR programme in the UK). First commercial SMR power to a hyperscale customer is unlikely before 2029-31 at the earliest. But the announcements matter now because they signal the direction of travel: hyperscale customers are willing to commit to novel generation technologies if it decouples them from grid timelines.




Reported behind-the-meter commitments by hyperscale customers
Announced or contracted capacity, MW, by technology, cumulative through mid-2026

Compiled from hyperscaler press releases, utility announcements and trade press (S&P Global, Data Center Dynamics, Bloomberg). Includes existing nuclear PPAs (Amazon-Talen, Constellation-Meta, Microsoft-Three Mile Island), announced gas-turbine plants, Bloom fuel cell contracts, and SMR MOUs/PPAs. Uncontracted MOUs weighted lower than firm contracts.


Behind-the-meter power has moved past niche status to become the strategy every hyperscaler now runs in parallel with grid interconnection, and that shift is why gas-turbine, fuel-cell and SMR order books have simultaneously expanded to record levels.The BTM thesis

$Market size · TAM · Profit pools



Sizing the interconnect market



The published range


Global electrical grid equipment markets are large and well-documented. Total global T&D equipment revenue is roughly $150-200B annually. Large power transformers are approximately $30-40B of that. MV switchgear runs $40-55B. HV/MV cables $50-70B (including submarine cables, which are outside the data centre scope). Gensets and standby generation $20-25B. Fuel cells for stationary power ~$3-5B. Wrapping around all of it, global copper markets are roughly $250-300B in mine output value at 2024-25 prices.


The AI data centre slice of these markets is smaller but growing fast. Combining rack-level content per MW with hyperscale build rates yields a bottom-up estimate.




Bottom-up TAM for AI data centre interconnect


Rack-level and campus-level interconnect content per MW of IT load, indicative:


 * Large power transformer (50-150 MVA class): ~$30-45k per MW at OEM prices, plus installation.
 * MV switchgear (substation + distribution): ~$25-35k per MW.
 * MV cables (site-scale copper/XLPE): ~$20-30k per MW depending on campus layout.
 * Backup generation (gensets sized for full IT load at N+1): ~$40-60k per MW installed.
 * Behind-the-meter primary generation (where deployed): $800-1,500k per MW installed for gas turbine, $2,000-3,500k per MW for fuel cell, $6,000-12,000k per MW for SMR (illustrative future).
 * Copper and misc conductor: ~$25-35k per MW at 2025 prices.


Excluding the behind-the-meter primary generation category (where it is not deployed), interconnect content averages roughly $150-200k per MW of IT load for a hyperscale campus. Including primary-generation BTM (where applicable), the figure can climb toward $1M+ per MW.




Interconnect content per MW of AI IT load
Waterfall build, USD thousands per MW, hyperscale campus without behind-the-meter primary generation

Author's build. Ranges vary widely by capacity class, voltage class, redundancy design and OEM. Values shown are midpoints of stated ranges. Does not include BTM primary generation, which typically comes on top when deployed.




Applying that content per MW to the ~7-9 GW of AI IT-load additions modelled in Part II (and reused in Part III) gives a 2026 TAM of roughly $1.2-1.7B in narrow interconnect equipment, or $12-18B including construction, engineering and servicesIncluding BTM primary generation lifts the ceiling substantially: by 2030, if 20-30% of new AI capacity is BTM-primary rather than grid-primary, the total could exceed $60B. Full workings in Annex §A1.




Bottom-up TAM: AI data centre interconnect
USD billions, 2026-2030E, base case, includes BTM at rising share of new AI capacity

Author's model. Assumes AI IT-load additions of 8 GW in 2026 rising to 15 GW by 2030 (from Part II), interconnect content per MW of ~$180k, and BTM primary generation share of new capacity rising from 5% (2026) to 25% (2030). Cross-checked against announced hyperscale power deals through mid-2026.





Profit pools


The profit-pool inversion that Parts II and III described (smaller components hold better margin) plays out somewhat differently in interconnect. The categorisation is more about scarcity rent than about qualification barriers.


 * GOES (~25-35%): materials-industry margins, protected by concentrated supply and long expansion lead times. The best-defended layer.
 * Large power transformers (~15-22%): heavy-industrial margins, currently benefiting from scarcity pricing in the current cycle. Sustainable at elevated levels as long as backlogs persist.
 * MV switchgear (~18-25%): better than transformers because product diversity is higher and switching-mechanism IP is real. Digital and clean-gas products carry higher margins than baseline.
 * MV cables (~10-15%): commodity-plus economics; margin comes from scale and specialty products (submarine, HVDC).
 * Copper (fluctuating with commodity): mining margins vary widely by ore body and cycle; smelter margins compressed; fabrication margins low-teens.
 * Backup gensets (~15-20%): healthy through the current lead-time cycle, would compress in a normal market.
 * Fuel cells (~5-15% and improving): Bloom Energy and peers still scaling into profitability; margin trajectory positive but modest.
 * SMR (loss-making): pre-commercial; all margin discussion is speculative until units are operating.



M&ACapital · Consolidation



The consolidation is done, the operations phase is beginning


The strategic map of interconnect equipment consolidated over 2018-2024 through a series of large transactions. The big pieces have moved. What remains is bolt-on M&A, capacity expansion capex, and operational execution.




Transaction / eventYearImpact

Hitachi acquires ABB Power Grids2020 ($11B)Created Hitachi Energy as the largest global player in T&D equipment. Combined ABB's Western customer relationships with Hitachi's manufacturing depth. Backlog and order book have expanded since.
Siemens Energy carve-out2020Separated grid, gas turbine and offshore wind businesses from Siemens AG. Focused entity, publicly listed, dedicated capex programme. Combined with Siemens Gamesa Renewable Energy since.
GE Vernova spin2024General Electric split into GE Aerospace, GE Vernova (energy) and GE HealthCare. Vernova as a standalone entity has been aggressive on grid capex and backlog growth.
Eaton acquisitions2024-26Eaton has been a serial data-centre acquirer per press releases: Fibrebond ($1.4B, closed April 2025), Resilient Power (2024), Ultra PCS ($1.55B, closed Jan 2026), Royal Power (2025), Boyd Thermal ($9.5B, closed March 2026). Grid-to-chip integration is Eaton's positioning theme; the acquisitions have executed it.
Constellation restarts TMI2024 (announced)Constellation Energy signed a 20-year PPA with Microsoft to restart Three Mile Island Unit 1 (renamed Crane Clean Energy Center), targeting 2028 return-to-service. First commercial nuclear restart driven by data centre demand.
Amazon. Talen (Susquehanna)2024Amazon acquired data centre campus adjacent to Talen's Susquehanna nuclear plant with rights to draw up to 960 MW behind-the-meter. FERC review continues to complicate the structure but the deal set the template for nuclear-adjacent siting.
Bloom Energy scaling2024-26Multiple hyperscale contracts announced, cumulative order book growing quarter over quarter. Bloom is transitioning from "specialty vendor" to "primary generation supplier" for BTM data centre power.





What is still on the board


 * Regional US LPT manufacturers (Delta Star, Virginia Transformer, Pennsylvania Transformer) as consolidation targets for larger OEMs looking to expand domestic capacity. Small revenue, high strategic value under CHIPS Act and DOE Loan Programs Office funding.
 * SMR developers as investment targets from hyperscalers and industrial groups. Post-restructuring NuScale, private Kairos and Oklo, Rolls-Royce SMR: equity valuations vary widely by regulatory progress.
 * Cable manufacturers outside the top five (Southwire in the US, regional European specialists) as consolidation targets for Prysmian/Nexans/NKT. Data centre demand has raised the strategic value of MV cable capacity.
 * Fuel cell and specialty gas turbine developers as venture and strategic capital continues to flow into BTM power.



TECHDisruption watch list



What could change the shape of this market



VectorWhat it isWhoImpact

Solid-state transformers (SST)Medium-voltage AC-DC-AC conversion using SiC/GaN power electronics in place of a line-frequency transformer coreHitachi Energy, Siemens Energy, ABB, university and DOE programmes; adjacent to Part II wide-bandgap suppliersIf commercial-scale SSTs deploy, they erode LPT volume at the sidecar level and pull medium-voltage SiC demand forward. Timeline uncertain; production units at hyperscale voltage classes remain limited.
Amorphous-core LPTsAmorphous metal (Metglas) replaces GOES in transformer cores, cutting no-load losses ~70%Hitachi Metals / Proterial; regional LPT builders; Chinese amorphous producersBounded impact at LPT scale because of handling difficulty. Would relieve GOES pressure at the distribution transformer scale where the substitution is already commercial.
Digital / asset-managed transformersSensor packages, oil DGA monitoring, thermal telemetry, integrated with cloud analyticsHitachi Energy, Siemens Energy, ABB, third-party analytics vendorsAdds a service tail to transformer sales. Small revenue today, high margin, structurally growing. Changes the transformer OEM economics toward a razor/blade model.
SF6-free MV switchgearClean-air, dry-air, or fluoronitrile alternatives to SF6 for insulation and interruptionHitachi Energy (EconiQ), Siemens (Blue GIS), GE Vernova (g3), ABBRegulatory-driven displacement of SF6 accelerates 2026-30. Higher unit prices, similar performance. Reshapes MV switchgear procurement decisions.
Behind-the-meter SMRSmall modular reactors sited on hyperscale campuses, running BTMNuScale, X-Energy, Kairos, Oklo, GE Hitachi BWRX-300, Rolls-Royce SMR2029-32 for first commercial units. If it works, redraws the data centre siting map and materially reduces grid-interconnect dependency for the biggest customers.
HVDC interconnectorsHigh-voltage DC point-to-point transmission between generation and load centres, using SiC/IGBT valvesHitachi Energy, Siemens Energy, GE Vernova, NKT/Prysmian cablesEnables long-distance transmission from renewable-rich or nuclear regions to data centre load centres. Cable and valve-hall lead times are the constraint.
Distributed / meshed campusesInstead of one large substation feeding multiple buildings, distributed medium-voltage feeds from multiple grid pointsUtilities, hyperscaler engineering teamsReduces single-point-of-failure risk and can improve interconnect timeline by tapping smaller available capacity at multiple grid nodes. Adds complexity, higher unit cost.
Aluminium and Cu-Al hybrid busbarsSubstitute aluminium for copper where mass allows, reducing copper intensity per MWFabricators (Wieland, KME), rack integratorsBounded substitution at MV cable and busbar level. Does not touch the transformer winding or the smallest cross-section copper.
Superconducting MV cablesCryogenic high-temperature superconductor cables carrying an order of magnitude more current per cross-sectionNKT, SuperPower, research consortiaCommercial pilots exist (LIPA in New York); economics do not yet close for standard data centre applications. Watch, don't underwrite.





RISKTrade-offs · Policy · Winners



Trade-offs, and the argument against




The permitting and policy layer


Grid interconnect is not just an equipment problem. Federal, state and local permitting timelines for new transmission lines, substations and generation facilities have expanded over the past two decades. NEPA reviews, state siting boards, community consent, environmental impact assessments, transmission cost allocation between ratepayers and load: each is a genuine process with legitimate constituencies, and the cumulative timeline for a major new interconnection can consume years. The Inflation Reduction Act and various federal permitting reform proposals have made incremental changes, but nothing fundamental has shifted. Interconnect is now as much a policy problem as an equipment problem, and the equipment problem is severe enough on its own.




The commodity risk


Copper prices could correct if global demand normalises faster than expected, if mining supply comes online faster, or if AI capex slows materially. GOES supply could ease if Chinese exports increase (currently limited by domestic demand and provenance rules). Interconnect vendors have benefited enormously from scarcity pricing over the past three years; some of that will normalise as capacity expansions land in 2027-28.




Who wins, who is pressured


 * Winners: the three global LPT majors (Hitachi Energy, Siemens Energy, GE Vernova) through the current cycle, with margins holding as long as backlogs persist; MV cable majors (Prysmian, Nexans, NKT) similarly; the top switchgear vendors as SF6 alternatives expand ASPs; Bloom Energy and gas turbine OEMs on BTM demand; GOES producers on structural material scarcity; integrated grid-to-chip platforms (Eaton, Schneider, Vertiv) on the bundling thesis.
 * Under pressure: hyperscalers whose expansion depends on grid interconnect timelines they cannot control (some Meta, Microsoft, Amazon deployments have visibly slipped); utility-scale customers competing with hyperscalers for the same LPTs and switchgear; smaller regional LPT builders whose capacity is being locked up by hyperscale contracts and priced out for utility replacement customers.
 * The dominant risk: a slowdown in AI capex would relieve the equipment scarcity, normalise lead times, and compress the margins that OEMs have been enjoying. The equipment vendors are levered to hyperscaler capex in the same way that Parts II and III described the semiconductor and cooling vendors being levered.




Ten leading indicators to watch


 1.  LPT backlog-to-revenue ratios at Hitachi Energy, Siemens Energy and GE Vernova. The sector's health in one number.
 2.  GOES capacity announcements from Japanese and Korean producers. The materials layer's leading indicator.
 3.  US ISO interconnect queue timelines (PJM, ERCOT, MISO, CAISO). When these compress, the BTM strategy's economic case softens.
 4.  Announced hyperscale BTM primary-generation deals. The pace of the workaround.
 5.  SMR commercial operating date announcements. Especially X-Energy, Kairos, GE Hitachi BWRX-300, NuScale.
 6.  Federal permitting reform legislation. Would compress interconnect timelines if it lands.
 7.  Bloom Energy quarterly bookings. The BTM fuel-cell leading indicator.
 8.  Copper spot price and LME inventories. The commodity temperature check.
 9.  Any regulatory ruling on Amazon. Talen or similar co-located nuclear deals. Sets precedent for BTM nuclear siting.
 10. Announced hyperscaler capex slowdowns or accelerations. The demand-side signal.



BOTTOMWhere it lands



Where this lands


The interconnect stack is the layer of AI infrastructure most exposed to physical grid constraints and policy timelines, and the layer where the industry response has been to route around the constraint through behind-the-meter power. That response is now visible in every hyperscaler's capacity plan. Gas turbines, fuel cells, and eventually SMR will absorb some fraction of AI power demand that the grid cannot serve on the timeline hyperscalers need. The underlying equipment layers (LPTs, switchgear, cables, gensets) remain in shortage, and the materials layer beneath them (GOES, copper) is structurally tight. For operators, the message is that interconnect capacity is the constraint to plan around, and the planning horizon is now measured in years. For investors, the interconnect equipment vendors have been re-rating over the past three years and the multi-year backlogs suggest that re-rating has further to run.


This is Part IV of a six-part series. Across the earlier three parts (capacitor stack, wide-bandgap semiconductors, thermal management) and this one, the through-line has been consistent: the physical infrastructure of AI compute has become supply-constrained across every layer at once, in ways that most compute-side commentary underweights, and the profit pools are migrating in each layer toward the components where qualification, materials chokepoints or lead-time scarcity create structural pricing power. Interconnect is the layer where those dynamics land most heavily on the calendar. The next several years of AI infrastructure growth will be paced by grid connections, transformer factories and generator lead times, not by GPUs.



ANNEXSources · Calculations · Assumptions



Annex: how every derived number was built


Every figure in this piece is either published (traceable to a named source), derived (calculated from published inputs), or estimated (my judgement, labelled). This annex covers the latter two.




§A1. TAM model for AI data centre interconnect


Status: derived. Full workings below.


Content per MW inputs. LPT (50-150 MVA class): ~$30-45k per MW of IT load at OEM prices, based on published utility procurement data and hyperscale-scale unit pricing through 2025. Add ~30-50% for installation, testing and commissioning. MV switchgear: ~$25-35k per MW based on Schneider Electric, Eaton and ABB product-line pricing scaled to campus loading. MV cables: ~$20-30k per MW based on Prysmian/Nexans quotations for XLPE MV cable systems, scaled to typical campus layout of 5-10 km MV runs per 100 MW. Gensets (N+1 diesel at full load): ~$40-60k per MW installed based on Caterpillar/Cummins list pricing plus integration and fuel systems. Copper and misc: ~$25-35k per MW at 2025 copper spot around $9,500/tonne, at 30 tonnes per MW.


Total interconnect content per MW. Sum of the above: ~$140-210k per MW of IT load, midpoint ~$175k. Author's build; individual line items are estimates.


Volumes. AI IT-load additions per Part II model: ~7-9 GW in 2026 rising to ~14-18 GW by 2030. Independent cross-check against Kuo's VR200 shipment estimates and JLL/McKinsey capacity forecasts.


Narrow TAM (equipment only). 8 GW × $175k/MW = ~$1.4B of narrow interconnect equipment in 2026 per author's bottom-up model. Extending to 2030 at 15 GW × ~$180k/MW (nominal content stability) = ~$2.7B by 2030.


Wide TAM (including EPC, engineering, services). Applies a 8-10× multiplier to the narrow TAM for engineering, construction, installation and commissioning services. Author's assumption; based on utility grid capex ratios where equipment is typically 15-20% of total substation project cost. Gives ~$12-18B for 2026, ~$20-28B by 2030 excluding BTM.


Including BTM primary generation. Assumes BTM primary share of new AI capacity rises from ~5% in 2026 (~0.5 GW at ~$1.5M/MW gas turbine average) to ~25% by 2030 (~4 GW × ~$1.5M/MW blended average across gas, fuel cell, early SMR). Adds estimated ~$0.7B in 2026 rising to ~$6B by 2030 per author's model. Combined wide TAM including BTM: ~$13-19B for 2026 rising to ~$26-34B by 2030. Bull case with heavier SMR deployment and higher grid-constraint pressure could reach $45-65B by 2030. Sensitivities are dominated by (1) AI capacity growth rate, (2) BTM share assumption, (3) SMR commercial operating date.




§A2. GOES supply and share


Status: estimated. Global GOES production is not consistently published. Author's estimates: Nippon Steel + JFE ~30-40% combined (dominant in domain-refined grade); POSCO ~10-15%; Cogent (Tata) and ArcelorMittal ~15-20% combined; NLMK/DK ~5-10% (subject to sanction-related uncertainty); Chinese producers (Baosteel, TISCO, WISCO) ~25-30% combined (largely consumed domestically). Add small production from AK Steel (Cleveland-Cliffs) in the US, and Indian producers (SAIL). Figures directional only; consult Metal Bulletin and SteelOrbis for current data if precision is needed.




§A3. LPT lead time trajectory


Status: published across multiple sources, with ranges. Pre-pandemic US LPT lead times of 40-60 weeks are documented in DOE 2014 and 2017 supply-chain studies. Post-pandemic expansion to 120-150 weeks is documented in DOE 2022 and 2023 studies, plus EPRI, T&D World and utility engineering-firm surveys through 2025. Specific hyperscale-class units have been quoted at 200+ weeks by industry contacts through 2025-26; not independently verified.




§A4. Interconnect queue timelines


Status: published. Lawrence Berkeley National Laboratory's annual "Queued Up" report tracks US ISO interconnect queues. As of the 2024 edition (published 2025 with 2024 data), typical queue times for large generation projects run 3-5 years in most ISOs, with large-load interconnections (data centres) reporting similar or longer waits in PJM, ERCOT and MISO. Individual data centre operators have publicly cited 4-7 year timelines for specific sites.




§A5. Copper intensity per MW


Status: estimated ranges, cited in industry press. BHP, Wood Mackenzie, S&P Global Commodity Insights and IEA have each published estimates in the 30-65 tonnes per MW range for data centres, with wider variation for AI-optimised hyperscale (higher end). The 3-5× multiplier vs legacy enterprise is a widely cited industry rule of thumb; specific studies vary. Global copper demand estimates and deficit forecasts through 2030 are from the same source cohort.




§A6. Behind-the-meter deal tracker


Status: compiled from press releases and trade press through mid-2026. Major announced or contracted deals include: Amazon. Talen (Susquehanna, ~960 MW capacity rights), Microsoft. Constellation (Three Mile Island Unit 1 restart, ~835 MW targeting 2028), Meta. Constellation (Illinois nuclear PPA, ~1,100 MW), Amazon. X-Energy (Washington HTGR site, 4 units × 320 MW each), Google. Kairos (fluoride-salt SMR MOU, ~500 MW), plus numerous smaller gas-turbine and Bloom Energy deals. Deal sizes and timelines subject to regulatory approval and construction risk; figures are announced capacity, not operational.




§A7. Other published figures used


 * Hitachi Energy backlog: reported multi-year highs through 2025; specific backlog values from company disclosures.
 * Siemens Energy grid capex programme: €4.5B+ announced through 2027, various company disclosures.
 * Eaton acquisitions: Fibrebond $1.4B, Ultra PCS $1.55B, Boyd Thermal $9.5B, others undisclosed; Eaton press releases and Reuters.
 * SF6 GWP: ~23,500 relative to CO₂, atmospheric lifetime 3,200 years; IPCC AR6 (2021).
 * Bloom Energy hyperscale deployments: multiple announcements through 2024-26; Bloom Energy press releases and trade press.
 * Copper spot price: ~$9,500/tonne baseline 2025, ~$10,500-11,500/tonne range through 2026; LME.



A caution on this sector's data. Grid equipment markets are less transparently reported than semiconductor or capacitor markets. Company disclosures are heavily aggregated; trade press is fragmented; regulatory filings from utilities are voluminous but hard to aggregate. Where I have quoted numbers, I have flagged the source. Where I have estimated, I have said so. This annex is a starting map for diligence, not a substitute for primary work.

SERIESRelated in the AI Power Chain



The complete series


This is Part IV of the AI Power Chain series. The six parts cover the physical infrastructure of AI compute, layer by layer, from the medium-voltage utility bus to the 0.8V transistor gate. Each part has a technical companion that walks through the underlying physics.


For readers who want the physics behind this piece (transformer efficiency, GOES manufacturing, SF6 alternatives, XLPE cables, aeroderivative turbines, SOFC fuel cells, four SMR families), see the technical companion: The Interconnect Stack: Technical Companion.


The Capacitor Stack (Part I) covers hybrid supercapacitors, EDLCs, MLCCs, silicon capacitors and Al electrolytics. Technical companion: The Capacitor Stack: Technical Companion.


The Wide-Bandgap Stack (Part II) covers SiC and GaN power semiconductors after the substrate glut. Technical companion: The Wide-Bandgap Stack: Technical Companion.


The Thermal Stack (Part III) covers data centre liquid cooling from die to dry cooler. Technical companion: The Thermal Stack: Technical Companion.




Series. The AI Power Chain (six parts). Part I: The Capacitor StackPart I-A: Technical CompanionPart II: The Wide-Bandgap StackPart II-A: Technical CompanionPart III: The Thermal StackPart III-A: Technical CompanionPart IV: The Interconnect StackPart IV-A: Technical CompanionPart V: The On-Package Delivery StackPart V-A: Technical CompanionPart VI: The Modular Datacenter StackPart VI-A: Technical Companion.


Methodology. This analysis is independent and based entirely on public information. Market sizes, shares, lead times and deal figures are approximate, compiled from public research (Wood Mackenzie, S&P Global Commodity Insights, BHP outlooks, Metal Bulletin, SteelOrbis, T&D World), regulatory filings (DOE, FERC, DOE Loan Programs Office), lab research (LBNL Queued Up report, EPRI), company disclosures (Hitachi Energy, Siemens Energy, GE Vernova, Eaton, Schneider, Bloom Energy) and trade press (Reuters, Bloomberg, S&P Platts, Data Center Dynamics) through August 2026. Figures marked directional, estimated or single-source should be treated accordingly. Nothing here is investment advice.


© 2026 Adi Kumar · Power & Digital Infrastructure · Corrections welcome.











The AI Power Chain: series companions

 * The AI Power Chain: Vendor Screen. The vendor map across every layer
 * Pricing Under Scarcity. Where premium capture is compounding
 * The Services Inversion. Why services now command product-like multiples


==============================================================================
# The On-Package Delivery Stack: How 48V Becomes 0.8V at 2,250 Amperes
URL: https://adikumar.co/the-on-package-delivery-stack/
Published: 2026-08-02
Summary: On-package power delivery for AI GPUs: how 48V becomes 0.8V at 2,250 A. Vertical vs lateral PoL, MPS design-in dominance, on-chip caps.
==============================================================================

The AI Power Chain series · Part 9 of 15









Glossary of terms used



CAGR
Compound Annual Growth Rate. The annualised rate at which a metric grows over a defined period, typically 3-5 years.
GaN
Gallium Nitride. Wide-bandgap semiconductor material used in RF and power applications.
IVR
Integrated Voltage Regulator. On-package power delivery component for modern CPUs and GPUs.
SiC
Silicon Carbide. Wide-bandgap semiconductor material used in high-voltage power electronics.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.






The AI Power Chain · Part V of VI



The On-Package Delivery Stack


The 10 millimetres of interconnect that carry 1,000 amperes into a 0.8-volt transistor plane, and the foundry-level architecture change now under way to make them scale.




The AI Power Chain · six essays, one thesis

 1. Part I. The Capacitor Stack: 800VDC at the rack
 2. Part II. The Wide-Bandgap Stack: SiC and GaN conversion
 3. Part III. The Thermal Stack: removing the waste heat
 4. Part IV. The Interconnect Stack: busbars and whips
 5. Part V. The On-Package Delivery Stack: 48V to 0.8V (you are here)
 6. Part VI. The Modular Datacenter Stack: how the building gets built




Executive summary


 * Public analysis of the AI power path usually stops at the rack. The layer below, from 48V down to 0.8V flowing through interconnect the width of a hair at 1,000 amperes, is where the current generation of accelerators is thermally and electrically limited, and where the next architectural shifts are happening.
 * Two changes are landing at once. Vertical Power Delivery (VPD) moves DC feed from around the die to through the die's back-side. Backside Power Delivery Networks (BSPDN) etch a dedicated metal power plane beneath the transistor layer, freeing the front-side stack for signal routing. Both are foundry-level architectural changes, not board-level ones.
 * TSMC's SuperPower Rail lands with N2P in H2-2026 and A16 in 2027. Intel PowerVia is in volume on the 18A node from H2-2025. Samsung Foundry follows on SF2P in 2027-2028. In parallel, glass core substrates from Intel, Absolics (SKC), Corning, LG Innotek and Samsung Electro-Mechanics enter datacenter packaging in the 2026-2028 window, driven by structural stiffness and current density more than by dielectric properties.
 * The vendor map splits cleanly. Discrete power stages and multi-phase controllers (Monolithic Power Systems, Infineon, Renesas, Analog Devices, Vicor, Empower Semiconductor, Alpha & Omega) sit above the substrate. On-die and on-package power (TSMC, Intel Foundry, Samsung Foundry, ASE, Amkor, Powertech) sits inside the package. The two do not compete directly. They stack.
 * Total addressable market for on-package power delivery is roughly $8-10B in 2026, growing 22-28% CAGR into the early 2030s. The highest-margin layers are qualification-gated: MPS and Empower for AI-grade multi-phase (28-35% operating margins on the accelerator-facing revenue line), IVR patents cross-licensed between Intel, AMD and Nvidia, and TSMC and Intel Foundry running de-facto monopolies on BSPDN for the first 18-24 months of each node.
 * The Chinese position is real at the discrete VRM layer (Silergy, SG Micro, Kinetic Technologies) and has no meaningful US commercial presence at the BSPDN and glass-substrate layers. That gap is the widest anywhere in the AI power chain, and arguably the narrowest chokepoint in the series.




The AI Power Chain · six essays, one physical arc


The series walks a single physical path. It begins at the medium-voltage utility bus at the site fence, steps down through the substation and switchgear, arrives at the datacenter rack where 800V DC is stabilised by the capacitor stack, is converted by silicon-carbide switches to 48V, is distributed across the rack by copper busbars and whips, is stepped down again by multi-phase controllers on the accelerator board to 0.8V, and finally routed through the on-package power delivery network to a transistor gate drawing over 2,000 amperesWaste heat from every conversion stage is removed by the thermal stack. The whole thing is packaged inside a factory-modular building because there aren't enough electricians to build it stick-frame. Six essays. One 800V → 0.8V staircase.


 1. Part I. The Capacitor Stack 800VDC at the rack
 2. Part II. The Wide-Bandgap Stack SiC and GaN conversion
 3. Part III. The Thermal Stack Removing the waste heat
 4. Part IV. The Interconnect Stack Busbars and whips
 5. Part V. The On-Package Delivery Stack 48V to 0.8V (you are here)
 6. Part VI. The Modular Datacenter Stack How the building gets built




Why the last ten millimetres are the hardest


Every essay in this series has walked one step closer to the transistor. Part I stopped at the rack, where 800V DC arrives across a bus. Part II showed how silicon carbide and gallium nitride step that voltage down in the power shelf. Part III showed how the resulting heat is removed. Part IV mapped the copper that carries the current across the rack once it has been converted. This essay picks up where Part IV stopped: at the accelerator's package edge, where 48V (or, in the older generations, 12V) enters and has to become 0.8V at the transistor gate, without dropping more than a few millivolts across the interconnect between.


That last conversion is where the physics gets hard. The transistor supply voltage on a leading-edge accelerator has collapsed roughly threefold over the last decade, from about 1.5V on 28nm to about 0.65V on 3nm and 2nm. Compute density has moved the opposite way. A single H100 die draws around 700W. Blackwell B200 draws about 1,200W. Blackwell Ultra targets ~1,400W per NVIDIA's published data. Rubin is expected to push materially higher; specific per-package figures have not been confirmed publicly. Taking 1,800W as an illustrative envelope, at 0.8V that would be 2,250 amperesThat current has to flow through metal that is at most a few micrometres thick, through a package that is at most a few millimetres tall, and at a resistance low enough that the IR drop across the delivery network stays within a few tens of millivolts of the transistor's specified operating window. Miss it by 50mV and the accelerator throttles.


Two constraints dominate the design space. First, ohmic loss scales with the square of current, so doubling current at constant voltage quadruples the heat rejected by the power path itself. Every millivolt of IR drop matters. Second, the current has to be delivered everywhere on the die at once, not just at the die edge. A modern accelerator has thousands of independent power domains, each with its own voltage and current profile, each transitioning between states in nanoseconds. The delivery network has to hold voltage stable across all of them, at kilohertz-to-megahertz transient frequencies. Node scaling keeps tightening the constraint faster than the delivery network can adapt to it.


A single Rubin-generation accelerator will draw approximately 2,250 amperes at 0.8 volts. That current has to flow through metal a few micrometres thick, at an IR budget of a few tens of millivolts, or the transistor throttles.The design constraint



The voltage staircase


The complete voltage staircase for a modern AI rack, laid out end to end, is one of the most compressed engineering hierarchies in modern electronics. Each stage has its own converter topology, its own dominant vendor, its own regulatory context, and its own failure mode. This essay is about the last three or four stages on the right of that staircase, but the whole picture is worth showing at once.



The AI voltage staircase: rack to transistor
Voltage and current at each conversion node, hyperscale AI rack, log-scale current axis

Author's construction from published NVIDIA / Vera Rubin platform disclosures, hyperscaler design notes and vendor product documentation. Voltages are nominal; currents are approximate for a 210kW rack-class system with a single 1,800W accelerator at the bottom of the chain. Current axis is logarithmic. Each conversion node has efficiency losses not shown here (see Annex §A1).



Read the chart from left to right. The utility supplies medium-voltage AC in the tens of kilovolts. The datacenter's power hall converts that to 415V or 480V three-phase AC, then increasingly to 800V DC at the rack. A silicon-carbide shelf steps 800V DC down to 48V DC at the board. A multi-phase point-of-load converter steps 48V to 0.8V at the accelerator. And an on-die power grid distributes that 0.8V across the transistor plane. The current at each stage is the mirror image: single amperes at the top, hundreds at the board, thousands at the die.


The essays in this series have covered the left half of that staircase in depth. This one covers the right half, from 48V downward.








In brief


On-package power delivery is how the 48V or 800V rack bus becomes 0.8V at the GPU die. Above 2,000W per package, multi-phase point-of-load converters sit directly under or adjacent to the die. Monolithic Power Systems dominates hyperscaler design-in wins, with Vicor and Empower Semiconductor holding contested positions on specific platforms.








Where the vendors sit: the on-package power map


The map below is the industry as it looks in mid-2026, with the qualification-gated bottlenecks marked. It is denser than the maps in earlier parts of the series because on-package power delivery has more distinct layers, each with its own vendor concentration and its own margin structure. Nothing about the map is stable on a three-year view. All of the foundry-level layers are moving.





L148V intermediate bus converter48V DC → 12V DC or 48V direct-to-load · unregulated LLC or regulated buck

Vicor US · direct 48VAnalog Devices US · LTM8000 familyMonolithic Power Systems US · MPMxxxInfineon DE · TDA family
Renesas JP · ex-DialogSilergy CNSG Micro CNDelta Electronics TW/US




L2Multi-phase point-of-load12V or 48V → 0.6-1.2V · 8-20 interleaved phases · the AI-grade workhorse

Monolithic Power Systems US · MP288xEmpower Semiconductor US · IVR/DrMOS hybridAnalog Devices US · LTC familyInfineon DE · OptiMOS/DrMOS
Renesas JP · RAA familyAlpha & Omega USTexas Instruments USMaxLinear US · ex-Silicon Mitus
Silergy CNKinetic Technologies CN




L3Integrated Voltage Regulator (IVR)On-package regulator, magnetic core in substrate · Intel FIVR, AMD MPS, Apple silicon

Intel US · FIVR since HaswellEmpower Semiconductor US · thin-film magneticsFerric Semiconductor US · acquired by TDK 2025
TDK JP · thin-film inductorsMurata JP · magnetic packagingSTMicroelectronics FR/IT




L4Vertical Power Delivery (VPD)Power routed through the substrate to the die back-side, not around it · dominant hyperscaler design after 2026

TSMC TW · packaging integrationIntel Foundry US · PowerVia integrationSamsung Foundry KR · 2027 target
ASE Group TW · advanced packagingAmkor US/KRPowertech TW




L5Backside Power Delivery Network (BSPDN)Dedicated power plane etched beneath the transistor layer · TSMC SuperPower · Intel PowerVia · Samsung BSPDN

Intel Foundry US · PowerVia 18A · volume H2-2025TSMC TW · SuperPower N2P · H2-2026Samsung Foundry KR · SF2P 2027-2028
imec BE · research consortium




L6Silicon interposer & RDLCoWoS · SoIC · Foveros · through-silicon vias for power

TSMC TW · CoWoS-S/L/RIntel US · Foveros 3DSamsung KR · X-Cube
ASE TW · VIPackAmkor US/KR · SLIM




L7Glass core substratesStructural stiffness & power density > organic laminate · datacenter-class in 2026-2028

Intel US · in-house · 2027 volumeAbsolics (SKC) KR · Covington GA plantCorning US · glass supply
LG Innotek KRSamsung Electro-Mechanics KRDNP JPAGC JP · glass supply




L8Organic laminate substrates & TSVsThe incumbent · 4-14 build-up layers · being displaced upward by glass, downward by BSPDN

Ibiden JPShinko Electric JPKyocera JP
Simmtech KRLG Innotek KRNan Ya PCB TWUnimicron TW
Shennan Circuits CN




L9On-die metal stackM0. M15 · buried power rails at 2nm · Cu, Ru, W · 20+ metal layers on leading-edge nodes

TSMC TW · foundry-integratedIntel Foundry USSamsung Foundry KR
Applied Materials US · depositionLam Research US · etchASML NL · EUVTokyo Electron JP




United StatesEuropeJapanChinaTaiwanKoreaShaded tiles = category leaders



How to read this map. On-package power delivery is a stack where the layers are physically inseparable at production. The controller die (L2) sits on the substrate (L7 or L8) which sits on the interposer (L6) which sits above the accelerator (L4/L5) which contains the on-die metal (L9). The market shares quoted are directional. Two structural features matter more than the shares. First, Layers 4, 5, 6 and 9 are foundry-controlled and effectively closed to non-foundry vendors: the choice of foundry determines the choice of on-package power architecture. Second, Layer 7 (glass substrates) is the disruptive layer for the 2026-2028 window: it displaces Japanese incumbent Layer 8 leaders and creates a new profit pool for Korean and US vendors.




L1. L2: The board-level converters and where their margins live


Before the substrate, before the interposer, before any of the foundry-level engineering, the accelerator gets its power from a set of discrete converter chips on the motherboard or the intermediate power module. These are conventional silicon parts, made on mature CMOS nodes, and sold as merchant devices. They are also where the visible profit pools of the on-package power stack currently sit.


The board-level path splits into two topologies. The older topology takes 48V from the rack bus, converts it to 12V through an intermediate bus converter (IBC), then converts 12V to the accelerator's core voltage through a multi-phase point-of-loadThe newer topology, pushed by hyperscalers to reduce conversion losses, converts 48V directly to core voltage in a single multi-phase stage, skipping the 12V intermediate. Direct 48V conversion is thermally and electrically harder but eliminates the roughly 3% loss of the IBC stage and removes a component category from the board.


Both topologies converge on the same critical layer: the multi-phase point-of-load. This is where the interesting economics sit. A modern accelerator's core voltage rail draws over a kilowatt at under one volt, which requires eight to twenty parallel converter phases operating in interleaved fashion to spread the current across enough physical devices to keep each below its thermal limit. Each phase is a "DrMOS" module (driver plus MOSFETs plus a control interface). The controller die orchestrates the phases, balances current between them, and manages transient response. Both the controller and the DrMOS parts are qualification-gated into the accelerator's reference design.




Multi-phase point-of-load: vendor revenue vs operating margin
2025 estimated AI-server-related revenue (USD) and estimated GAAP operating margin on that revenue line, top vendors

[A2] Author's estimates from company filings and vendor product announcements. Revenue is the AI-server-attributable slice of each vendor's power management line, not total power revenue. MPS and Empower carry the highest margin because their AI-grade multi-phase parts are qualification-gated into NVIDIA reference designs. Chinese vendors (Silergy, Kinetic) are visible but do not appear in top hyperscaler AI reference designs as of mid-2026. See Annex §A2.




Monolithic Power Systems is the reference vendor. Its MP288x family holds design-in slots on Hopper, Blackwell and Rubin, and the AI-server slice of MPS revenue is estimated at roughly $900M in 2025 growing to $1.4-1.6B in 2026 at operating margins that its filings suggest are the highest in the multi-phase space, in the 32-36% range. Empower Semiconductor is smaller (est. $180-220M in 2025) but grew fastest, because its hybrid IVR/DrMOS parts are the merchant substitute for Intel's captive FIVR. Analog Devices, Infineon and Renesas hold structural positions on non-accelerator power (server CPU, networking silicon, memory) and are contesting the AI accelerator slot with newer parts. Renesas' acquisition of Dialog gave it the digital-controller IP it lacked, and Infineon's OptiMOS DrMOS is the volume alternative to MPS. Texas Instruments participates but is not the reference part on the leading accelerators.


The Chinese position at this layer is real but qualitatively different. Silergy and Kinetic Technologies have credible multi-phase parts and hold volume slots on Chinese hyperscaler accelerators (Huawei Ascend, Cambricon MLU, Enflame). They do not hold slots on NVIDIA or AMD reference designs, and there is no visible path to them doing so on the current-generation parts. The qualification cycle for a new controller in a hyperscaler AI reference design is 18-24 months and includes a full thermal, transient and electromagnetic co-simulation; that gate has held since Blackwell and shows no sign of loosening on Rubin.



The diligence point on L2. Multi-phase point-of-load is a qualification-gated, low-BOM-percentage, high-consequence-of-failure business, which is the same structural profile that made TIMs the highest-margin layer in the thermal essay. The AI-grade slice of the market carries margins that are structurally higher than the automotive or industrial slice. Look at MPS and Empower for defensible economics; look elsewhere for volume.




L3: Integrated voltage regulators, and why Intel had this ten years early


The next step is to move the multi-phase converter from the board to the package itself. This is what an Integrated Voltage Regulator (IVR) is: the entire buck converter, including its magnetic inductor and its power switches, integrated into the substrate directly beneath the accelerator die. It shortens the current loop from centimetres to millimetres, cuts IR drop, and lets the converter operate at higher switching frequencies (tens of megahertz instead of hundreds of kilohertz), which shrinks the required inductor and capacitor values by roughly two orders of magnitude.


Intel introduced IVR technology, branded FIVR, on Haswell in 2013. The engineering was ahead of its time, and Intel wound the technology down on Skylake for reasons that were as much strategic as technical: FIVR needed high-current-density magnetics that were expensive to fabricate, and moving voltage regulation off the motherboard shifted the industry BOM in ways that Intel's OEM customers pushed back on. The technology never disappeared. Intel reintroduced elements of it on Meteor Lake, and it now returns as a core requirement on Rubin-class accelerators, this time from the merchant vendor side.


The merchant IVR vendors are Empower Semiconductor, Ferric Semiconductor (acquired by TDK in 2025), and, at lower volumes, several Japanese magnetics houses (TDK, Murata) supplying the thin-film magnetic components that make IVR economically viable. Empower's thin-film inductor process, developed in partnership with Intel over the preceding decade, is the closest thing to a merchant equivalent of FIVR. Ferric's technology, now inside TDK, is the alternative. Both companies solved the same problem: fabricating magnetic cores with high enough current density to sit in a package substrate, in a footprint measured in single-digit square millimetres, at frequencies above 20MHz. The manufacturing capital intensity is substantial. Neither company has meaningful Chinese competition today.




The magnetic core density problem
Current density (A/mm²) achievable at various switching frequencies · discrete inductor vs thin-film IVR magnetics

[A3] Author's compilation from IEEE PELS conference papers 2022-2025, Intel FIVR technical disclosures, and Empower Semiconductor product briefs. The two curves cross around 5-10MHz: below that frequency, discrete power modules are more cost-effective; above it, integrated thin-film magnetics dominate. AI accelerator IVRs target 20-50MHz. Illustrative; see Annex §A3.




The strategic question about IVR concerns whether the merchant IVR value migrates upward into the foundry, downward into board-level multi-phase, or holds its ground as a separately merchant layer. That IVR itself works has been settled for years. The current answer is a hedged one: on the flagship accelerators of 2026-2027, IVR sits on the package as a merchant component from Empower or TDK, sharing the substrate with the accelerator die. On the 2028+ generation, the migration path is upward: TSMC and Intel are both engineering power regulation into the interposer itself, using the same thin-film magnetics techniques but delivered as part of the packaging service. If that migration completes, Empower's addressable market compresses onto second-tier accelerators and Chinese domestic AI silicon.




L4. L5: Vertical Power Delivery and Backside Power Delivery, the foundry-level architectural shift


Everything below this section is happening inside the foundry, at the level of the semiconductor process itself. It is where the on-package power stack transitions from a business of buying and stacking merchant components to a business of choosing an architectural path for the whole accelerator.


To understand what has changed, think about how power reached the transistor on a 7nm or 5nm accelerator. The die had ten to fifteen metal layers stacked on top of the transistor plane. Signal routing and power distribution shared those metal layers. Power came in through solder bumps around the die's perimeter, threaded through the substrate up into the front side of the die, then travelled downward through the metal stack to reach the transistors. Signal routing took the same path in reverse. As dies scaled, this shared-metal-stack model started to run out of room: the transistors were getting smaller but the power grid needed thicker, wider metal to carry the current, and the two demands fought each other for area on the same metal layers.



Front-side power delivery ran out of metal budget somewhere around 3nm. Every leading-edge node from 2026 onward routes power through the die's back-side, on a separate metal stack that the transistors never see.The architectural break



Backside Power Delivery Networks (BSPDN) solve the fight by moving power to a separate metal stack on the back of the die. The transistor plane sits in the middle, signal routing sits on the front-side metal stack, and power routing sits on the back-side metal stack. The two never share area. Because the back-side power stack does not need to leave room for signals, its metal can be much thicker and wider, carrying more current with less IR drop. And because the transistors receive power from behind rather than from above, the power path from the substrate to the transistor gate is measured in single-digit micrometres rather than tens of micrometres, cutting resistance by more than half.


The three foundries take slightly different paths. Intel's PowerVia is in volume on the 18A node from H2-2025, ahead of the pack by roughly nine to twelve months. TSMC's SuperPower Rail enters volume with N2P in H2-2026 and expands to A16 in 2027. Samsung Foundry's BSPDN is targeted for SF2P in 2027-2028. All three implementations achieve the same first-order goal: they reduce IR drop by 30-50% versus the front-side-only equivalent, unlock roughly 5-10% higher clock frequency at the same power, and roughly 5-8% lower power at the same clock. The differences are in yield ramp curves, in the specifics of how the back-side stack is fabricated (wafer thinning, temporary carrier bonding, direct wafer-to-wafer hybrid bonding for the highest-performance implementations), and in the manufacturing equipment supplied by Applied Materials, Lam Research, ASML and Tokyo Electron.


Vertical Power Delivery (VPD) is the packaging-level counterpart. Where BSPDN routes power on the die's back-side, VPD routes power through the substrate to the die's back-side, rather than around the perimeter of the die to its front-side. VPD without BSPDN is possible (the current can still be routed through the die from back to front on a conventional front-side power grid) and is used on hyperscaler accelerators that have not yet moved to a BSPDN foundry node. The full stack is VPD + BSPDN, and Rubin's most power-hungry configurations are the first commercial parts to run both together.




Where the IR drop budget goes
Indicative IR drop breakdown, millivolts, from board VRM to transistor gate · front-side vs back-side power delivery

[A4] Author's estimate from IEEE JSSC and IEDM papers 2023-2025, TSMC and Intel technical disclosures on N2 and 18A respectively. Illustrative decomposition of a typical 1V accelerator supply's ~40mV IR budget. Backside delivery reduces the transistor-to-substrate loop resistance by roughly 60%, freeing headroom for higher clock or lower voltage. See Annex §A4.




The commercial implication of BSPDN is straightforward: on any leading-edge accelerator manufactured from 2026 onward, the choice of foundry is the choice of power architecture. There is no third-party BSPDN vendor. There cannot be, because the technology is inseparable from the front-end wafer process. That is why hyperscaler chip programs (Google TPU, AWS Trainium, Microsoft Maia, Meta MTIA) are structurally tied to their foundry partners: switching foundries midway through a program means redesigning the entire power distribution network, not just recompiling the layout.



The chokepoint you cannot route around. BSPDN is one of the narrowest chokepoints in the AI power chain. Only three companies in the world have working BSPDN at leading-edge nodes: Intel Foundry, TSMC, Samsung Foundry. No Chinese fab has demonstrated a working BSPDN process at any node below 5nm, and SMIC's public roadmap does not project one before 2030. This is a wider technology gap than the ones in silicon carbide substrates, in universal quick disconnects, or in high-voltage cabling. It is the widest anywhere in the six-part series.




L6: Silicon interposers, the layer above the die


Between the accelerator die and its substrate sits the interposer. On a leading-edge AI part, this is a piece of silicon (or, increasingly, glass) that carries the interconnect between the accelerator's compute die, its high-bandwidth memory stacks, and its off-package connections. The interposer is the physical structure that makes 2.5D and 3D packaging possible, and its power delivery role is often overlooked.


Every through-silicon via (TSV) on the interposer is a candidate power via. Modern accelerator interposers carry thousands of TSVs, and a substantial fraction of them are dedicated to power delivery rather than signal. The interposer is what physically routes the power from the substrate up into the die's back-side (on VPD) or around the die into the front-side power grid (on conventional designs). Interposer resistance and inductance therefore sit directly in the IR drop budget.


TSMC's CoWoS (Chip-on-Wafer-on-Substrate) is the volume interposer for merchant AI silicon. It comes in three flavours: CoWoS-S (silicon interposer, the original), CoWoS-L (with a local silicon interconnect bridge, cheaper for smaller dies), and CoWoS-R (RDL redistribution layer without a full silicon interposer, cheapest but with lower TSV density). NVIDIA Hopper, Blackwell and Rubin all use CoWoS variants. Capacity is famously tight: TSMC's public disclosures put 2026 CoWoS capacity at roughly 75,000 wafers per month, up from 45,000 in 2025 and 15,000 in 2023, and every incremental wafer is spoken for eighteen months in advance.




TSMC CoWoS capacity ramp
Monthly wafer-out capacity, 2023 actuals to 2027 committed, plus reported AMD/NVIDIA allocation

[A5] TSMC quarterly earnings commentary, DigiTimes and Nikkei reporting on capacity expansion (2024-2026). Allocations are approximate; the true numbers are commercially sensitive and disclosed only in fragments. NVIDIA reportedly holds ~60% of 2026 CoWoS-S allocation. See Annex §A5.




Intel's Foveros is the alternative on Intel-manufactured parts, with its own set of packaging variants (Foveros Direct for hybrid bonding, Foveros Omni for redistribution). Samsung's X-Cube holds a smaller position, largely inside Samsung's captive silicon. ASE's VIPack, Amkor's SLIM and JCET's XDFOI are the OSAT (outsourced assembly and test) alternatives, playing at lower TSV densities and lower current requirements than the flagship AI parts but growing fast on the second-tier accelerator market.




L7: Glass substrates, the disruptive layer


The substrate is the piece of packaged organic material that sits beneath the interposer and provides the electrical connection to the motherboard. The incumbent technology is organic laminate (typically ABF, Ajinomoto Build-up Film, a Japanese resin), pressed into build-up layers with copper traces etched between them. The Japanese substrate industry (Ibiden, Shinko, Kyocera) has led this technology for two decades. The Korean and Taiwanese entrants (LG Innotek, Simmtech, Nan Ya, Unimicron) have closed most of the gap on volume but not on the highest layer counts.


The problem with organic laminate is that it warps. As accelerator dies get larger (Blackwell's compute die is 814mm², two dies bonded, Rubin larger still), the substrate has to hold that die flat within tens of micrometres across roughly 100 × 100 mm of substrate area. Organic laminate's coefficient of thermal expansion (CTE) does not match silicon's, so as the package heats and cools, the substrate warps against the die. On smaller packages this is manageable. On the physically largest packages in the industry, it becomes a yield problem.


Glass substrates solve this by replacing the organic build-up layers with glass. Glass has a CTE much closer to silicon's, is more rigid, and can support higher via densities. Intel announced its glass substrate roadmap in September 2023 with volume targeted for 2026-2027. Absolics, a subsidiary of Korea's SKC, began commercial shipments of glass substrates from its Covington, Georgia facility in H2-2025. Corning, LG Innotek, Samsung Electro-Mechanics and Dai Nippon Printing are all in the market. AGC and Nippon Electric Glass supply the underlying glass sheets.




Glass substrate market ramp
Author's estimate, glass substrate revenue (USD millions), 2025-2030E, split by application

[A6] Author's model from company announcements, Intel investor day disclosures (2024), Absolics/SKC capacity commitments and industry conference commentary through 2026. Data-center packaging is the volume driver from 2027 onward. Automotive and RF applications are earlier adopters but at lower volumes. Substantial upside and downside sensitivity to Intel PowerVia yield ramp. See Annex §A6.




Two structural points about the glass substrate layer matter for the profit-pool question. First, the technology gap between Korean and Japanese substrate incumbents is much smaller in glass than in organic laminate. Absolics is genuinely competitive with Intel's captive glass on a first-generation basis, which does not happen often in packaging. Second, the Chinese position is absent. Shennan Circuits and other Chinese substrate manufacturers are established at the low-end organic laminate level but have no visible glass substrate program at datacenter grade. The gap here mirrors the BSPDN gap: the tools, the process knowledge and the customer relationships that would allow rapid Chinese entry are all missing.




L8. L9: The incumbent substrate layer and the on-die metal stack


Layer 8 is the organic substrate business that glass is displacing from above. It is a large business (roughly $12-14B in 2025, growing single-digits) and it will remain the volume solution for accelerators below the flagship tier for years. The Japanese majors (Ibiden, Shinko, Kyocera) hold the highest layer counts. Korean entrants (LG Innotek, Simmtech) hold the mid-tier. Taiwanese and Chinese vendors dominate the low-tier. The interesting margin story is what happens as glass takes share off the top: Ibiden and Shinko lose the highest-priced product line first, and the volume that remains at organic laminate is progressively lower-margin. This is a structural profit pool compression at the incumbent, over the 2026-2030 window.


Layer 9, the on-die metal stack, is a foundry-integrated business. Nothing about it is separately merchant. It belongs in the map because the equipment suppliers (Applied Materials for deposition, Lam Research for etch, Tokyo Electron for coat/develop, ASML for lithography) capture a meaningful share of the value that flows through this layer, and their AI-driven revenue lines run through here. The specific innovation at this layer that matters for the on-package power thesis is the introduction of ruthenium as an interconnect metal on the top-most power routing layers, which offers lower resistance than copper at the smallest line widths and is being introduced at 2nm and A16 nodes. This is the final segment of the delivery path: the current-carrying metal at the transistor gate itself.



$Market size · TAM · Profit pools



Sizing the market and locating the profits




Bottom-up TAM


Build the market from AI IT-load additions, as in Parts I, II, III and IV. On-package power delivery content per rack: multi-phase controllers and DrMOS on the accelerator board average roughly $2,200 per 1,800W accelerator (up from $1,400 for a 700W H100). At 8 accelerators per NVL72-class rack tray, this is $17,600 per tray, and roughly $70,000 per NVL144 rack once the IBC and supporting components are included. Substrate, interposer and packaging content per rack: CoWoS-class packaging runs $2,500-4,000 per accelerator die at current TSMC pricing, so roughly $80,000-130,000 per NVL144 rack. BSPDN and on-die metal: absorbed into the accelerator's wafer cost, not separately visible.


Combining the visible layers, on-package power delivery content sits at roughly $450-650 per kW of AI IT load, plus a further $100-150/kW captured inside the accelerator wafer cost that is not separately merchant. Multiplying against roughly 8GW of new AI IT-load added in 2026 gives a bottom-up TAM of $4.0-5.5B for the visible layers, plus another $0.8-1.2B captured inside foundry pricing but attributable to power delivery. Published market sizings for the broader "AI power management" category cluster around $7-10B for 2026, with 20-25% CAGRs. The reconciliation gives confidence in the $8-10B combined figure, understanding that some fraction of that is embedded in foundry wafer prices rather than separately quoted.




Bottom-up TAM: on-package power delivery
USD billions, 2026-2030E, split by layer

[A7] Author's model. AI IT-load additions from Part II model (~7-9GW in 2026 rising to ~14-18GW by 2030). Content per kW: multi-phase POL ~$180, IVR ~$40, substrate + interposer ~$220, foundry-embedded (BSPDN, on-die metal) ~$120. Sensitivity dominated by accelerator power scaling (H100 700W → Rubin 1,800W → potential 2,500W+ successors). See Annex §A7.





Where the profits actually sit


The layer-by-layer margin structure is more differentiated than in the earlier essays. The board-level layers (L1, L2) are high-margin at the AI-server slice (28-35% operating) and low-margin at the non-AI slice (12-18% operating), producing a mixed corporate margin at vendors like MPS that is nonetheless improving as AI mix rises. The IVR layer (L3) is a niche but with structurally high margins (35-45% at Empower, difficult to verify given its private status). The foundry-controlled layers (L4, L5, L6, L9) do not have separately-visible margins because they are captured inside foundry pricing; the effective margin is TSMC's or Intel Foundry's, which sits at 40-55% gross on leading-edge nodes. The glass substrate layer (L7) is new and margin structure is not yet stable, but early Absolics unit economics suggest gross margins in the 35-45% range on first-generation product, which will compress as competition builds.




Profit pool by layer, 2026 illustrative
Estimated gross profit dollars (USD millions) by on-package power delivery layer, and estimated operating margin

[A8] Author's estimate. Foundry-embedded layers (L4, L5, L6, L9) are shown as a single bar because they are inseparable at company accounts. Illustrative only. See Annex §A8.





Summary matrix: on-package power delivery vs the rest of the AI power chain




LayerContent per rackVendor concentrationChinese positionGross margin band


L2 Multi-phase POL~$70kMPS + 4 alternatesDomestic AI only32-36% (AI slice)


L3 IVR (on-package)~$8kEmpower / TDK oligopolyAbsent35-45% est.


L4/L5 VPD + BSPDNEmbedded in wafer3 foundriesAbsentFoundry-captured, 45-55%


L6 Interposer (CoWoS)~$100kTSMC de factoEmerging (JCET)Foundry-captured, 40-50%


L7 Glass substrate~$4-8k (2027+)Intel + Absolics + CorningAbsent35-45% early


L8 Organic laminate~$3-6kJapanese majorsLow-end only18-25%, compressing






Reading the map for the next three years


Three structural moves are already visible in the 2026-2028 window, and each of them changes the merchant profit pool at a specific layer.


First, IVR consolidation upward into the foundry. Both TSMC and Intel are engineering thin-film magnetics into their interposer processes. If that consolidation completes on the Rubin+ generation, the merchant IVR market compresses onto second-tier accelerators, and Empower's addressable market shrinks by roughly one third. Watch for foundry announcements at IEDM 2026 and IEDM 2027.


Second, glass substrate share taking the top of the organic laminate market. On a five-year view, glass takes the flagship AI slot and the flagship server-CPU slot, leaving organic laminate to hold the mid-tier and mainstream server business. The Japanese incumbents lose their highest-margin product line first. Ibiden's substrate revenue mix will be the most watched line in Japanese electronics reporting through 2027.


Third, BSPDN broadening from Intel-first to a three-foundry norm. Intel PowerVia has an 12-18 month lead on TSMC SuperPower Rail today. That lead compresses as TSMC N2P ramps in H2-2026 and A16 in 2027. On the 2028 flagship node, all three foundries have BSPDN in volume, and the technology moves from "which foundry" to "which node within a foundry". The competitive question shifts from having BSPDN to how efficiently each foundry integrates it into their process, and how it interacts with hybrid bonding, glass substrates and thin-film magnetics.



All three of these moves reduce the merchant addressable market at a specific layer and expand the foundry captive market. The net effect over the 2026-2030 window is that value flows upward: from board-level multi-phase (L2) into IVR (L3), from IVR into interposer (L6), from organic substrate (L8) into glass (L7), and from front-side power distribution into BSPDN (L5). Every one of those flows narrows the field of merchant winners.



Over the 2026-2030 window, value in on-package power flows steadily upward, from the board into the substrate, from the substrate into the interposer, from the interposer into the foundry itself. The merchant winners narrow at each step.The structural thesis




The cross-series thesis in one paragraph


This essay closes the arc that Part I opened. The AI power chain begins at the medium-voltage utility bus and ends at a 0.8V transistor gate drawing 2,250 amperes. Every essay in the series is a slice through that path. Part I sized the capacitor stack that stabilises 800V DC at the rack. Part II sized the wide-bandgap conversion that steps 800V down to 48V. Part III sized the thermal system that removes the heat released along the way. Part IV sized the interconnect that carries the current between conversion stages. This one sizes the last conversion, from 48V to 0.8V, and the physical structures inside the accelerator package that make it possible. Part VI will size the modular civil and electrical infrastructure that lets datacenter operators build the whole thing at industrial scale. Read together, the six essays add up to a single question: which layers of the AI power chain have defensible margins over the 2026-2030 window, and which layers become commodity as the volume ramps? On the evidence collected across this series, the answer is remarkably consistent: qualification-gated layers with high consequence of failure and small bill-of-materials percentages hold their margins; volume-metal-and-plastic layers do not. On-package power delivery has more of the former than any other essay in the series.



Related topic hubs
 * AI Power Semiconductors
 * 800V Data Centre Power
 * Browse all topic hubs →



Annex · Sources and workings


§A1 The voltage staircase chart. Voltage and current values are nominal for a hyperscale AI rack of Vera Rubin generation (210kW/rack, 8 accelerators/tray, 6 trays/rack, single-accelerator draw ~1,800W). Utility voltage assumes a 20kV distribution feed. 800V DC at the rack is the sidewide power-shelf design, standardised in OCP ORv3 revisions from 2024 onward. Efficiency at each conversion stage is not shown here; typical stage efficiencies are 98-99% at the MV/LV transformer, 96-98% at rectification, 97-98.5% at the SiC 800V→48V shelf, and 90-94% at the multi-phase POL, giving a total facility-to-transistor conversion efficiency of ~85-88%.


§A2 Multi-phase POL vendor revenue and margin. AI-attributable revenue is estimated from vendor investor commentary and third-party industry teardowns (TechInsights, System Plus Consulting). MPS attribution derived from MPS 10-Q disclosures on datacenter revenue mix. Empower Semiconductor is a private company; revenue estimates rely on published press releases and industry reporting through 2025-Q4. Operating margin bands are estimates from filings and comparable public analysts; qualification-gated AI slice runs materially higher than non-AI corporate margin.


§A3 Magnetic core density curves. Compiled from IEEE Power Electronics Society conference papers (APEC 2023-2025, ISSCC 2024-2025) and Intel FIVR technical disclosures from Meteor Lake reveal. Empower and Ferric magnetics values from company technical briefs. Cross-over frequency between discrete and thin-film magnetics is process-dependent; the chart shows an illustrative case.


§A4 IR drop breakdown. Compiled from TSMC N2 and Intel 18A technical publications at IEDM 2024 and 2025, plus IEEE JSSC papers on power delivery for AI accelerators (2023-2025). The specific decomposition depends on the accelerator design; the chart shows a representative case for a 1.0V, 1000A supply rail.


§A5 CoWoS capacity. Compiled from TSMC quarterly earnings commentary 2023-Q1 through 2026-Q2, DigiTimes reporting on packaging capacity expansion (Longtan, Chunan, Tainan fab announcements), and Nikkei coverage of NVIDIA and AMD packaging allocation. Allocation percentages are approximations; the specific hyperscaler splits are commercially sensitive and disclosed only fragmentarily.


§A6 Glass substrate ramp. Intel investor day disclosures (September 2024), Absolics announcements on Covington Georgia plant commissioning (H2-2025), SKC Q4-2025 investor briefing, Corning quarterly earnings on advanced glass. Application mix (datacenter vs automotive vs RF) is an author estimate; the datacenter proportion depends heavily on the timing of Intel PowerVia yield ramp, and could swing +/- 30% on either side of the chart.


§A7 Bottom-up TAM. AI IT-load additions from Part II model. Content per kW breakdown: multi-phase POL ~$180/kW (from BOM teardowns of Blackwell/Rubin reference designs), IVR ~$40/kW (est. based on Empower unit pricing), substrate + interposer ~$220/kW (CoWoS pricing per DigiTimes reports, allocated by die area), foundry-embedded ~$120/kW (author's allocation of leading-edge wafer premium attributable to BSPDN and on-die metal). Sensitivity dominated by accelerator power scaling. Bear case: accelerator power flattens at 1,500W. Bull case: 2,500W by 2028.


§A8 Profit pool by layer. Foundry-captured layers grouped because they are inseparable at corporate accounting. Board-level layer margins from vendor filings and estimates for the AI-attributable slice; note that AI-slice margins are structurally higher than corporate blended margins. Glass substrate margins are early-stage estimates and will compress as capacity ramps.



Cross-references


Part I: "The Capacitor Stack" (800V DC at the rack).
Part II: "The Wide-Bandgap Stack" (SiC and GaN conversion 800V→48V).
Part III: "The Thermal Stack" (heat rejection from all conversion stages).
Part IV: "The Interconnect Stack" (rack-level current distribution).
Part VI: "The Modular Datacenter Stack" (civil/electrical modularity for the whole building).










The AI Power Chain: series companions

 * The AI Power Chain: Vendor Screen. The vendor map across every layer
 * Pricing Under Scarcity. Where premium capture is compounding
 * The Services Inversion. Why services now command product-like multiples







Frequently asked


How does 48V DC become 0.8V at a GPU die?

A multi-phase point-of-load converter mounted on or adjacent to the GPU package steps 48V down to 0.8V through 12-24 parallel phases. Each phase handles ~200A. Combined phases must maintain <3mV IR drop across the delivery network under 2,000A transient load, which drives package layout + capacitor placement decisions.


Which vendors dominate GPU on-package power delivery?

Monolithic Power Systems (MPS) holds most hyperscaler design-in wins across Hopper, Blackwell and Rubin platforms. Vicor supplies DDCM modules at select customers. Empower Semiconductor competes on cost + integration. Infineon and TI hold contested positions on specific OEM server platforms.


What is vertical vs lateral point-of-load?

Vertical power delivery mounts converters directly beneath the die on the package substrate, minimising inductance. Lateral mounts them adjacent to the die. Vertical achieves better transient response at high current density (>1,500A) but requires tighter package + thermal design. Adoption tracks GPU TDP growth.


==============================================================================
# The Modular Datacenter Stack: How the Building Gets Built (When There Aren't Enough Electricians)
URL: https://adikumar.co/the-modular-datacenter-stack/
Published: 2026-08-02
Summary: Modular AI data centre stack: prefab enclosures, skid-mounted power + cooling, on-site assembly economics. Where the building gets built, not designed.
==============================================================================

The AI Power Chain series · Part 11 of 15









Glossary of terms used



CAGR
Compound Annual Growth Rate. The annualised rate at which a metric grows over a defined period, typically 3-5 years.
GPU
Graphics Processing Unit. The compute silicon at the centre of AI workloads.
GW
Gigawatt. One thousand megawatts of electrical power.
GaN
Gallium Nitride. Wide-bandgap semiconductor material used in RF and power applications.
MV
Medium Voltage. Typically 1 kV to 35 kV. The voltage class connecting data centre power distribution to the utility grid.
MW
Megawatt. Unit of electrical power. A large modern data centre draws tens to hundreds of MW.
SiC
Silicon Carbide. Wide-bandgap semiconductor material used in high-voltage power electronics.
TAM
Total Addressable Market. The maximum revenue opportunity available if a product served every potential customer segment.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.






The AI Power Chain · Part VI of VI



The Modular Datacenter Stack


How a 500MW AI campus gets built when there are not enough medium-voltage electricians on earth to build it stick-frame. The move from field construction to factory-assembled modules, one skid at a time.




The AI Power Chain · six essays, one thesis

 1. Part I. The Capacitor Stack: 800VDC at the rack
 2. Part II. The Wide-Bandgap Stack: SiC and GaN conversion
 3. Part III. The Thermal Stack: removing the waste heat
 4. Part IV. The Interconnect Stack: busbars and whips
 5. Part V. The On-Package Delivery Stack: 48V to 0.8V
 6. Part VI. The Modular Datacenter Stack: how the building gets built (you are here)




Executive summary


 * The five earlier essays walked from the medium-voltage utility bus at the site fence down to the 0.8V transistor gate inside the accelerator package. This one steps back out to the physical box that holds all of it, and asks a different question: who builds the building?
 * The binding constraint on hyperscale AI capacity today sits with the qualified craft-labour base of medium-voltage electricians, industrial mechanical contractors, and switchgear specialists available to build gigawatt-scale sites on a two-year cycle, rather than with chip supply. US Bureau of Labor Statistics data shows industrial electrician headcount growing at roughly 2% per year while datacenter electrical work is growing at 25%+ per year. The gap is closing through modularisation: shifting labour hours out of the field and into factory-controlled environments.
 * Modularisation is happening at every layer of the physical stack. Medium-voltage transformer skids (Vertiv, Schneider, Legrand, Eaton) arrive on flatbed trucks pre-wired and pre-tested. Cooling distribution units are containerised and lifted into place. Busway systems (Eaton, Siemens, Schneider) replace hand-run conduit for the whole 480V. 48V distribution path. Rack enclosures are pre-populated at ODM factories and delivered as complete AI compute pods. Building shells are increasingly precast-concrete or steel-frame kits assembled from a schedule rather than poured on site.
 * The economic case for modularisation runs primarily through schedule risk rather than through labour cost. A hyperscale campus that ships a month late costs the operator roughly $2-4M per day of foregone GPU revenue at current AI rental prices. Modular components that trade a 10% cost premium for a 30% schedule reduction have a two-order-of-magnitude positive NPV against site-built equivalents.
 * Total addressable market for modular AI datacenter infrastructure is roughly $45-60B in 2026, growing 22-28% CAGR through 2030 to $110-150B. This is the largest TAM in the six-part series, and it is also the most fragmented, with no single vendor above ~9% share and a long tail of regional contractors, modular fabricators and specialty engineers.
 * The profit-pool structure is inverted relative to the earlier essays. Where on-package power (Part V) concentrates value in three foundries, and thermal (Part III) concentrates it in five majors, modular datacenter is where the Western industrial groups actually dominate globally: Vertiv, Schneider, Eaton, Legrand, Siemens, ABB and Delta Electronics hold 55-60% of the addressable market between them, with regional specialists (Modine, nVent, Compass, Aligned, EdgeConneX) taking the remainder. Chinese participation is meaningful but structurally regional, not global, because the electrical code differences (60Hz vs 50Hz, US NEC vs IEC, medium-voltage class harmonisation) create a natural wedge.




The AI Power Chain · six essays, one physical arc


The series walks a single physical path. It begins at the medium-voltage utility bus at the site fence, steps down through the substation and switchgear, arrives at the datacenter rack where 800V DC is stabilised by the capacitor stack, is converted by silicon-carbide switches to 48V, is distributed across the rack by copper busbars and whips, is stepped down again by multi-phase controllers on the accelerator board to 0.8V, and finally routed through the on-package power delivery network to a transistor gate drawing over 2,000 amperesWaste heat from every conversion stage is removed by the thermal stack. The whole thing is packaged inside a factory-modular building because there aren't enough electricians to build it stick-frame. Six essays. One 800V → 0.8V staircase.


 1. Part I. The Capacitor Stack 800VDC at the rack
 2. Part II. The Wide-Bandgap Stack SiC and GaN conversion
 3. Part III. The Thermal Stack Removing the waste heat
 4. Part IV. The Interconnect Stack Busbars and whips
 5. Part V. The On-Package Delivery Stack 48V to 0.8V
 6. Part VI. The Modular Datacenter Stack How the building gets built (you are here)




Why the last essay in the power chain is about the building


Every previous essay in this series has explained a physical layer of the AI compute stack: capacitors, wide-bandgap switches, thermal, interconnect, on-package power delivery. This one steps out to the building that holds all of them, because the physical layer that gates the deployment of everything above is no longer the silicon. It is the civil and electrical infrastructure that has to be poured, wired, tested and energised before any of that silicon can be racked.


The scale of what has to be built is unfamiliar even to people who have watched the datacenter industry for years. A GW-scale campus, of the sort that Meta, Microsoft, Google, AWS, Oracle and xAI are all announcing on multi-month cadence, requires roughly 2.5 million square feet of white space, 2,500-5,000 medium-voltage transformer sub-stations, 200,000+ tonnes of concrete, 15,000-25,000 tonnes of structural steel, and roughly 4-7 million labour hours to build. The largest single AI campuses under construction in 2026 (xAI Memphis, Microsoft Wisconsin, Meta Louisiana, Google Kansas) each exceed those numbers by 2-4x. And every hyperscaler wants three or four of them per year, in different regions, on 18-to-24-month cycles.


A GW-scale AI campus needs approximately 4-7 million labour hours to build. There are not that many qualified industrial electricians available in any single market, at any deliverable schedule. Modularisation is the only pathway by which the announced 2028 hyperscale build pipeline can actually be delivered on the required schedule.The binding constraint



Skilled labour, not equipment supply, is what gates the pipeline


US Bureau of Labor Statistics data shows industrial electrician employment at roughly 780,000 in 2025, growing at 2.1% year over year. Skilled medium-voltage electrician employment is a subset, roughly 150,000 workers, growing at 1.4% because it is a specialty training track with a five-to-seven-year apprenticeship. The total addressable US pool of MV-qualified electricians available for datacenter work is on the order of 40,000-50,000, a fraction of which is not already committed to industrial, transmission or manufacturing projects. Datacenter construction demand for MV electrician-hours in 2026 is estimated at roughly 12-15% of that available pool, up from 3-4% in 2022.


The pinch points are not evenly distributed. Northern Virginia (the largest US datacenter market), Central Ohio (Meta and Amazon), Central Texas (Microsoft, Meta, xAI Memphis is not far away), Phoenix (Microsoft, Oracle) and Nevada (Switch) are all running at 100%+ of local MV electrician capacity, drawing traveling contractors from other regions at premium wage rates. Non-US markets are worse: the Nordic datacenter cluster and Ireland's Dublin cluster both have documented lead times of 18-24 months for MV substation commissioning primarily because of electrician availability, not equipment lead time.


The response, across every hyperscaler and every large developer, has been to shift as much labour as possible out of the field and into factory-controlled fabricationA pre-wired MV substation built inside a Vertiv or Eaton factory in the US Midwest uses roughly the same person-hours as a site-built equivalent, but those hours are performed by a broader pool of workers (including mechanical and manufacturing technicians who are not MV-qualified electricians), in an environment where quality control is systematic, and on a schedule that is not affected by weather or site logistics. The finished skid ships to the datacenter site as a single unit and is craned into place, then wired to the site's incoming feed and outgoing bus by a small MV electrician crew whose scope is measured in days, not weeks.



The MV electrician pinch: demand vs supply
Estimated US medium-voltage electrician-hours available vs datacenter construction demand, thousands of hours per year

[A1] Author's estimate from BLS Occupational Employment Statistics (2020-2025), Datacenter Frontier construction reports, and JLL/CBRE hyperscale market updates. MV electrician pool is a subset of total industrial electrician employment; datacenter demand is estimated from announced hyperscale capex and typical labour-hour intensity per MW deployed. See Annex §A1.







In brief


The modular datacenter stack moves the physical building itself into factory production. Prefabricated power skids, cooling modules and enclosures ship from Fibrebond (acquired by Eaton in April 2025 per company press release), Compass or Vantage factories, arriving at site pre-tested. On-site work is limited to interconnection and commissioning, compressing the schedule by six to twelve months against traditional stick-built.








The map: where the vendors sit at each layer


Modular datacenter is the layer of the AI power chain where the vendor map looks most like a mature industrial market and least like a technology sector. The players are the industrial groups that have supplied power distribution and cooling to the process industries for decades: Vertiv, Schneider, Eaton, Legrand, Siemens, ABB, Delta, Rittal, nVent. They have been extending their datacenter portfolios upward for a decade, first into modular UPS, then into modular power skids, then into containerised cooling, then increasingly into complete pre-fabricated data halls. The AI wave has accelerated that portfolio extension into a portfolio consolidation, with the largest three or four vendors now offering end-to-end factory-manufactured datacenter modules that a hyperscaler can specify, order and integrate at site.





L1Utility service & MV switchyard115kV. 345kV interconnection, ring buses, oil-immersed step-down transformers · utility-scale

ABB CH/SESiemens Energy DEHitachi Energy JP/CHGE Vernova USMitsubishi Electric JP
Hyundai Electric KRHD Hyundai Electric KRSungrow CNTBEA CN




L2MV transformer skids (13.8kV to 480V)Dry-type or oil-filled step-down units on pre-wired steel skids · the workhorse module

Eaton US/IEVertiv US · SmartMod / SmartAisleSchneider Electric FR · EcoStruxure Modular DC
Legrand FRABB CH/SESiemens DEDelta Electronics TW · InfraSuite Modular
Rittal DEHD Hyundai Electric KR




L3MV switchgear skids15kV/24kV/38kV switchgear · vacuum interrupters · integrated protection relays

Eaton US/IESchneider Electric FRABB CH/SESiemens DE
Mitsubishi Electric JPHyosung Heavy Industries KRLS Electric KRZhengtai Electric CN




L4UPS and LV distribution skidsModular UPS + PDU + LV switchboards · lithium-iron-phosphate BESS integration

Vertiv US · Liebert TrinergySchneider Electric FR · Galaxy VXLEaton US/IE · 9395P
ABB CH/SEToshiba JPFuji Electric JPKehua Data CNDelta Electronics TW
Sungrow CN · BESSCATL CN · battery




L5Busway systems (400A. 6300A)Copper or aluminium busway replacing conduit for LV distribution · 400V/480V/1000V

Eaton US/IE · Pow-R-WaySchneider Electric FR · CanalisSiemens DE · Sivacon
Legrand FR · ZucchiniABB CH/SEEAE Elektrik TRXinya Electronics CN




L6Rack-level power (rPDU, whips, tap-off)Rack PDUs, quick-connect whips, tap-off boxes · covered in depth in Part IV

Vertiv US · Geist rPDUServer Technology US · LegrandRaritan US · Legrand
Schneider · APC FR/USRittal DEChatsworth Products US




L7Cooling distribution skids & hydro modulesRow-level CDUs, pre-piped hydro modules, containerised heat rejection · covered in depth in Part III

Vertiv US · Liebert XDUCoolIT Systems CABoyd Thermal (Eaton) USMotivair US · Schneider partner
Airedale UKMunters SEDaikin JPEnvicool CN




L8Rack enclosures & row-level containmentORv3, 21-inch and 19-inch enclosures · cold aisle / hot aisle containment · seismic bracing

Rittal DELegrand FRVertiv US · KnockOn / SmartCabinet
Chatsworth Products USSchneider · APC FR/USPanduit USNitto Kohki JP




L9Building shell & modular civilPrecast tilt-up, steel-frame kits, modular data hall shells · site-erected in weeks not months

Compass Datacenters USAligned Data Centers USEdgeConneX US
Prime Data Centers USDigital Realty USEquinix US
DataQube UKDelta Modular NL




L10Facility BoP & heat rejection outdoor unitsCooling towers, dry coolers, adiabatic units, water treatment · covered in Part III

Baltimore Aircoil USEvapco USSPX Cooling US
Modine USAlfa Laval SEKelvion DE




United StatesEuropeJapanChinaTaiwanKoreaShaded tiles = category leaders



How to read this map. Every layer here has an incumbent industrial group with decades of process-industry pedigree, and every layer is being reshaped by the pace and scale of AI datacenter demand. The pattern that matters is vertical integration by acquisition: the same handful of Western industrial groups (Eaton, Schneider, Vertiv, Legrand, ABB, Siemens) now appear at four to seven layers each, having bought or organically extended into adjacencies over the last decade. That vertical integration is the defining strategic behaviour at this layer, and the reason the profit pool consolidates upward at each project cycle rather than fragmenting.




L1. L3: Utility, MV and switchgear, and the skid revolution


The physical arrival of grid power at a datacenter site historically involved a substantial civil-and-electrical project: a fenced switchyard with utility-scale transformers, oil-filled or dry-type, feeding into a purpose-built substation with hand-assembled switchgear, protective relays and metering. On a 30MW site this takes six to nine months of field work. On a 500MW campus it takes eighteen to thirty months on the critical path.


The modular alternative delivers the same functional equipment as pre-fabricated skids. A skid is a steel platform, typically 40 feet long, 12 feet wide and 12 feet tall (a standard road-legal shipping envelope), pre-populated at a Vertiv or Eaton or Schneider factory with the transformer, switchgear, protective relays, control panels, and internal wiring for a specific MV-to-LV step-down function. The skid arrives on a flatbed truck, is craned onto a prepared foundation pad, connected to incoming utility feed and outgoing distribution bus, tested, and commissioned. Total site labour: roughly 20% of the equivalent stick-built project. Total schedule: roughly 40% of stick-built. The trade-off is a 10-20% higher equipment cost (because factory labour is not free, and the steel enclosure and integrated skid engineering add cost) offset by roughly 80% lower site labour and roughly 60% shorter schedule.




Modular vs stick-built: labour hours and schedule
Indicative comparison, MV substation for a 100MW datacenter hall, field labour hours and calendar schedule

[A2] Author's estimate from industry-published modular substation case studies and construction-industry benchmarks. Actual figures depend heavily on site-specific factors (soil, permitting, weather, existing infrastructure). Illustrative. See Annex §A2.




The vendor economics at this layer are attractive on the demand side. Vertiv's SmartMod modular data centre business grew roughly 45% year-over-year in 2025, reaching estimated revenue of $1.8B on ~24% operating margin. Schneider's EcoStruxure Modular business grew similarly, though the segment is bundled with broader "digital energy" reporting that makes clean attribution difficult. Eaton's SmartAisle and modular substation business grew fastest, with the March 2026 Boyd Thermal acquisition adding the cooling module capability that the company previously purchased on the merchant market.


The MV switchgear layer (L3) sits inside the same skid vendors' offering but deserves its own paragraph because the switchgear content itself is high-value and low-volume: a single MV switchgear line-up for a hyperscale hall is $2-5M of equipment, and there are only a few global manufacturers (Eaton, Schneider, ABB, Siemens, plus Mitsubishi and the Korean majors) capable of supplying it to hyperscale QA standards. Lead times on 24kV switchgear are running at 60-80 weeks in mid-2026, up from 40 weeks pre-AI, and the queue for hyperscale-grade orders extends into 2028. This is the second-tightest supply-chain chokepoint in the whole physical stack after MV transformers themselves.




L4: UPS and LV distribution, and the LFP battery transition


The uninterruptible power supply layer sits between the MV distribution and the rack. Its job is to hold the datacenter's load through a utility outage (typically for the 10-30 seconds it takes for on-site generators to spin up), and to filter power quality issues (voltage sags, frequency wander, harmonics) from the incoming grid. A hyperscale UPS lineup is a $8-15M capital item and is typically modular within itself, with N+1 or 2N redundancy across parallel UPS units of 750kVA to 1.5MVA each.


The interesting shift at this layer is the battery chemistry transition. Historic datacenter UPS used valve-regulated lead-acid (VRLA) batteries, typically sized for 5-15 minutes of load hold time. VRLA is well-understood, cheap per kWh, and heavy. The industry has been transitioning to lithium-iron-phosphate (LFP) since roughly 2020, and by 2026 LFP is the default on new hyperscale UPS builds. LFP is more expensive per kWh installed but has 3-4x the cycle life, roughly 50% the footprint and 30% the weight, tolerates high temperatures without thermal runaway (a persistent concern with NMC lithium chemistries), and pairs naturally with the emerging role of the UPS as a behind-the-meter battery energy storage system (BESS) that can shift load, arbitrage grid tariffs, and provide grid ancillary services when idle.




Datacenter UPS battery mix: VRLA to LFP transition
Estimated share of hyperscale UPS energy storage capacity by chemistry, 2020-2030E

[A3] Author's estimate from Uptime Institute annual surveys 2021-2025, Vertiv and Schneider investor commentary, and OCP working group materials on BESS integration. NMC lithium sees limited adoption due to thermal safety concerns; the split is primarily VRLA → LFP. Illustrative. See Annex §A3.




The LFP transition matters commercially for two reasons. First, it opens a much larger vendor set: CATL, BYD, EVE Energy, Gotion and Farasis are all major LFP suppliers, and the price competition between them is fierce, driving battery cost per kWh down roughly 40% between 2022 and 2026. Second, it converts the UPS from a pure-cost redundancy item into a dual-use asset that can participate in demand-response and grid-services markets, changing the datacenter operator's electrical procurement model.




L5: Busway, and why the datacenter interior is now bare copper


Part IV covered rack-level power distribution in depth. What deserves separate treatment here is the layer between the LV distribution and the rack: the horizontal power distribution across the data hall floor. Historically this was hand-run conduit, with individual cables terminating at rack-side PDUs, installed by field electricians one connection at a time. On an AI hall with 100-500 racks per row, at 210kW per rack, that hand-installation model has broken.


The replacement is overhead busway: extruded copper or aluminium bars, insulated and enclosed in a metal housing, running the length of a row above the racks with plug-in tap-off boxes at each rack location. Busway can carry 400A to 6,300A per run, at 480V or up to 1,000V, and is installed in factory-fabricated straight sections (10 feet typical) that click together at flanged joints. The installation crew that hangs busway is a small team (four to six people, half of whom are not required to be MV-qualified) working sequentially down a row, and the connection to each rack is a physical plug-in with quick-connect strain relief rather than a wired termination.


Eaton's Pow-R-Way, Schneider's Canalis, Siemens' Sivacon and Legrand's Zucchini are the four dominant Western busway lines. Delta Electronics and the Chinese majors (Xinya Electronics, Cotran, Zhengtai) participate at lower price points and dominate Chinese domestic AI datacenter build-out. The Western majors hold hyperscale share in the US and Europe on the strength of UL and IEC certification and long-term qualification with Vertiv/Schneider/Eaton modular systems.



The diligence point on L5. Busway is a decades-old industrial product now sold at unprecedented volume into AI datacenter builds. Gross margins on busway sit in the 32-40% range at the Western majors on the AI-server slice, versus 22-28% on general industrial busway. The AI premium reflects the qualification cycle, the demand pressure on supply, and the increased power density per foot of busway (higher ampacity ratings command more copper per unit). This is a boring product with dull design but a very live pricing dynamic through 2027.




L6. L7: Rack power and cooling modules, at the datacenter interior scale


These two layers are covered in depth in Parts IV and III respectively. The point that belongs in this essay is that they are increasingly delivered as pre-integrated pods, not as separate rack-level installations. A hyperscale AI pod delivered by a modular vendor now includes: the rack enclosure (L8), the internal rack PDU (L6), the rack-level cooling distribution unit or manifold (L7), the internal chilled-water piping with quick disconnects, the internal power whips, the busway tap-off box, and integrated leak-detection and thermal telemetry. It arrives on a flatbed truck as a complete unit, is craned onto the data hall floor, connects to the overhead busway and to the underfloor hydro piping through pre-labelled quick-connect fittings, and is functional within hours of arrival.


Vertiv's SmartCabinet AI, Schneider's Modular Data Center, Eaton's SmartRack AI, and CoolIT's rack-integrated CDU pod are the volume products at this layer in 2026. Boyd Thermal (now Eaton) and Delta InfraSuite compete on price. The Taiwanese OEMs (Auras, Cooler Master, Delta) supply the underlying cold plates and manifolds but historically have not offered fully-integrated rack pods at hyperscale QA standards; that gap is narrowing on the Rubin generation.




L8. L9: Rack enclosures and building shells, and the modular civil vendors


The physical building at a modern AI datacenter is increasingly a kit rather than a stick-built structure. Precast tilt-up concrete, steel-frame prefab, structural-insulated panels, and even fully-modular data hall boxes are all in production use at hyperscale. Compass Datacenters pioneered the "Compass Grid" approach, delivering fully-built 6MW data halls from a Dallas-area fabrication yard on flatbed trucks. Aligned's DataHallX product is a similar concept at 30MW scale. EdgeConneX and Prime Data Centers use variants of the same modular approach at edge and mid-tier scale. The largest hyperscale campuses (Meta Louisiana, Microsoft Wisconsin, xAI Memphis) are not pure modular but incorporate modular elements at the data hall, MEP substation and cooling plant level.


The interesting economic question at this layer is whether modular datacenter developers can hold gross margins as they scale. First-generation modular developers achieved 25-35% gross margins on delivered turnkey capacity, which is materially higher than traditional REIT-style developed capacity. Whether that margin holds as hyperscalers integrate their own modular kits (Google and Microsoft both operate substantial in-house modular fabrication capacity) is the strategic question that will define the L9 layer over the 2026-2030 window.




Modular vs traditional datacenter development margin
Estimated gross margin on delivered turnkey capacity, modular-first developers vs traditional REIT developers, 2020-2030E

[A4] Author's estimate from Compass, Aligned, EdgeConneX, Digital Realty and Equinix filings and investor materials 2021-2026. Modular-first developers show a persistent margin premium of 10-15 percentage points but that gap is expected to compress as hyperscalers scale in-house modular capacity and as scale efficiencies flow to traditional developers. See Annex §A4.


$Market size · TAM · Profit pools



Sizing the market and locating the profits




Bottom-up TAM


Build the market from AI IT-load additions, as in every prior essay. Modular infrastructure content per MW of AI IT load: roughly $6,000/kW total site cost for a modern hyperscale AI campus (up from $4,500/kW for a 2022-vintage general datacenter), of which roughly 40% is electrical (transformers, switchgear, UPS, distribution, ~$2,400/kW), 25% is cooling (~$1,500/kW), 15% is civil/structural (~$900/kW), 10% is IT racks and rack-level integration (~$600/kW), and 10% is site work, permitting and commissioning (~$600/kW). Of that $6,000/kW, the modularisable share is roughly 65%, or $3,900/kW: everything above the foundation, minus site preparation, permitting and the final commissioning tie-in.


At $3,900/kW of modular content against 8GW of new AI IT load added in 2026, the bottom-up TAM for modular datacenter infrastructure is ~$31B in 2026That is the AI-attributable slice only. Adding the non-AI hyperscale datacenter build-out (a further ~5GW/year of general cloud, retail colocation and edge) and adjusting for the modular attach rate (currently ~55% and rising), the total modular datacenter infrastructure TAM sits at ~$45-60B in 2026, projected to ~$110-150B by 2030 at 20-25% CAGR.




Bottom-up TAM: modular datacenter infrastructure
USD billions, 2026-2030E, split by layer

[A5] Author's model. IT-load additions from Part II model (~7-9GW AI in 2026 rising to ~14-18GW by 2030; plus ~5GW/year non-AI hyperscale). Content per kW breakdown: MV skids + switchgear ~$1,200/kW, UPS + LV distribution ~$900/kW, busway + rack power ~$300/kW, cooling modules ~$1,000/kW, rack enclosures ~$200/kW, modular civil ~$400/kW. Modular attach rate assumed to rise from 55% in 2026 to 78% in 2030. See Annex §A5.





Vendor concentration and Western industrial dominance


The vendor share picture at this layer is the most concentrated in the six-part series among the Western industrial groups, and the least concentrated globally. Vertiv, Schneider, Eaton, Legrand, ABB, Siemens and Delta Electronics between them hold roughly 58% of the addressable market in mid-2026, up from 42% in 2020. The consolidation has come primarily through acquisition: Eaton's Boyd Thermal, Schneider's ETAP and AVEVA-related datacenter software, Vertiv's E&I and PowerBar, Legrand's Server Technology and Raritan, ABB's various portfolio consolidations.




Modular datacenter infrastructure vendor share (2026E, US + Europe)
Estimated addressable market share among modular/pre-fabricated infrastructure vendors, %, US + EU combined

[A6] Author's estimate from company filings and industry benchmark reports. Excludes Chinese domestic market. Share includes vendor's own modular-labelled product lines plus pre-fabricated skid business attributable to the vendor. Regional developers (Compass, Aligned, EdgeConneX, Prime) show up here on the civil/building side rather than the equipment side. See Annex §A6.




The Chinese position at this layer is meaningful but structurally regionalChinese modular datacenter development is real, at scale, and technologically competent: Kehua Data, Envicool, Delta InfraSuite, Sungrow and CATL between them supply most of the Chinese domestic AI datacenter build-out, and the design language and integration quality is competitive with Western equivalents. What Chinese vendors do not have is global reach into the US and European hyperscale marketThe 60Hz vs 50Hz electrical harmonisation, the UL vs IEC certification split, and the fact that hyperscale operators standardise on a small number of qualified vendors for multi-year framework agreements combine to make Chinese vendor entry to hyperscale-in-the-West very slow. This is different from every other essay in the series (where Chinese vendors are a live threat or an active substitute) and reflects the regional character of civil-and-electrical infrastructure procurement, which does not travel across code jurisdictions the way merchant components do.




The profit pools


The profit-pool structure at this layer is broadly favourable to the incumbents. Gross margins at the modular skid layer (L2-L4) run 32-40% at Vertiv, Schneider and Eaton on the AI-server slice, versus 24-30% on non-AI industrial equivalents. The AI premium reflects both the higher content per kW (bigger transformers, more redundancy, integrated cooling) and the pricing power that comes from lead-time-limited supply. Operating margins on the modular data centre business at Vertiv and Schneider now approach 22-25%, versus 15-18% at the corporate level; AI-attributable revenue carries a structural premium.


The layers where the profit pool is compressed are (a) traditional stick-built REIT development, which is being squeezed by the modular alternative, and (b) low-tier Chinese vendors trying to enter the Western hyperscale market, which face qualification barriers that push their pricing below sustainable levels. The layers where the profit pool is expanding are (c) integrated modular pod vendors selling complete data-hall-in-a-box products (Compass Datacenters is the reference), (d) the industrial cooling adjacencies now folded into the electrical groups (Boyd Thermal at Eaton is the reference), and (e) behind-the-meter BESS integrators combining LFP battery storage with UPS functionality (Sungrow, CATL and Delta at scale).




Profit pool by layer, modular datacenter infrastructure
Estimated gross profit pool ($B, 2026E) and operating margin (%) by layer

[A7] Author's estimate. Gross profit dollars are TAM × industry-average gross margin at each layer. Operating margins are estimates from published segment margins at the leading vendors. Illustrative. See Annex §A7.





Summary matrix: modular datacenter vs the rest of the AI power chain




AttributeModular datacenterRest of AI power chain

2026 TAM$45-60B (largest in series)$5-12B per subsystem essay
Vendor concentrationTop 7 hold ~58%Top 3 typically hold 60-80%
Chinese share of Western market< 5%5-25% depending on essay
Growth driverMW deployment × modular attach rateMW × content per MW
Dominant strategic behaviourVertical integration by acquisitionQualification-based margin defence
Binding constraintMV electrician availabilityComponent-specific chokepoints
Regulatory exposureGrid interconnection queues, permittingExport controls (some), fluid regulation (Part III)





Reading the map for the next three years


First, the MV substation and switchgear shortage does not resolve on any credible timeline before 2028. Eaton, ABB, Siemens and Schneider are all expanding factory capacity, but the constraint is skilled workforce (transformer winders, tank fabricators, switchgear technicians) as much as it is physical plant. Any hyperscaler build schedule that assumes 40-week switchgear lead times in 2027 is planning on a schedule that will slip. Modular vendors that can pre-order substation content 24-36 months ahead have a structural time-to-site advantage over vendors that cannot.


Second, the LFP-BESS transition accelerates through 2027. As Chinese LFP cell prices continue their decline and as the behind-the-meter grid-services revenue opportunity becomes better-defined (FERC Order 2222 implementation in the US, network-services market design in Europe), UPS-BESS convergence becomes a design default rather than an option. This favours the vendors with battery-side partnerships (Sungrow, CATL and Delta on the BESS side; Vertiv, Schneider and Eaton on the datacenter integration side) and disadvantages vendors that are late to the integration.


Third, hyperscaler in-house fabrication grows but does not displace merchant modular vendors. Google, Microsoft and Meta all operate substantial in-house or contract-manufactured modular fabrication for the parts of the stack they consider strategic (rack integration primarily, and some cooling module work). But no hyperscaler has demonstrated the willingness or capability to in-house the MV substation and switchgear layers, which is where the largest merchant profit pool sits. Expect this pattern to hold: hyperscalers integrate downward toward the rack, merchant vendors hold the layer between the meter and the rack.




The cross-series thesis in one paragraph


Six essays. One physical arc. It begins at the medium-voltage utility bus (which this essay finally names in vendor detail), steps down through the site substation (this essay), through the rack-level distribution (Part IV), through the 800V DC bus that the capacitor stack stabilises (Part I), through the silicon-carbide converter that steps 800V to 48V (Part II), through the multi-phase controller that steps 48V to 0.8V (Part V), and into the transistor plane where 2,250 amperes flow at that 0.8V (Part V). Every essay in the series is one segment of that voltage staircase. Every essay identifies the same structural pattern: qualification-gated layers with small bill-of-materials percentages hold durable margins; volume-metal-and-plastic layers do not. And every essay identifies the same Chinese-substitution dynamic: real and growing at the merchant-component layers, absent at the foundry-and-substation layers. Read together, the six essays are a coherent map of the AI power chain from the fence line to the transistor gate. This one closes the loop by naming the building that holds all of it.



The AI power chain begins at the medium-voltage utility bus and ends at a 0.8V transistor gate drawing 2,250 amperes. Six essays cover the physical layers between. This is the one that names the building.The six-part thesis

Related topic hubs
 * AI Power Semiconductors
 * 800V Data Centre Power
 * Browse all topic hubs →



Annex · Sources and workings


§A1 MV electrician labour supply. US Bureau of Labor Statistics Occupational Employment Statistics, series 47-2111 (Electricians) and 47-2073 (Industrial Machinery Mechanics). Subsegment "medium-voltage industrial electrician" is not a separate BLS category; the estimate is derived from union apprenticeship completion data (IBEW), specialty contractor employment estimates, and industry surveys. Datacenter labour-hour intensity per MW is from JLL and CBRE hyperscale market reports 2024-2026, cross-checked against Turner & Townsend construction cost benchmarks.


§A2 Modular vs stick-built substation. Comparison drawn from published case studies (Vertiv SmartMod, Schneider Modular DC, Eaton PowerXpert modular) and construction industry benchmarks. Actual figures depend heavily on site-specific factors; the chart is directional. Field labour hours for stick-built assume conventional construction; modular assumes crane placement + tie-in only.


§A3 UPS battery chemistry mix. Uptime Institute annual global datacenter surveys 2021-2025 (2025 shows LFP at ~54% of new-build UPS energy storage among hyperscale respondents, up from ~11% in 2022). Cross-referenced against Vertiv, Schneider and Eaton investor commentary. Extrapolated forward at declining VRLA share.


§A4 Modular developer margins. Compass Datacenters, Aligned Data Centers, EdgeConneX financial disclosures where available (mostly private-equity-owned so limited); Digital Realty and Equinix public filings for traditional REIT comparison. Modular developer margin premium is estimated from private company reports, industry survey data and analyst estimates. Compression forecast reflects expected hyperscaler in-house scale and traditional REIT modular adoption.


§A5 Bottom-up TAM. IT-load additions from Part II model. Content per kW breakdown: total site cost per kW is from Turner & Townsend and JLL construction benchmarks 2024-2026, allocated across MV/UPS/busway/rack/cooling/civil per typical AI hyperscale designs. Modular attach rate from Uptime Institute survey data.


§A6 Vendor share. Author's estimate from company segment revenue disclosures (Vertiv 10-K, Schneider Electric annual report, Eaton 10-K, Legrand annual report), cross-checked against industry analyst reports (Omdia, Dell'Oro, Synergy Research). Excludes Chinese domestic market where vendor share is very different.


§A7 Profit pool by layer. Gross profit dollars = TAM at each layer × industry-average gross margin. Operating margin bands are estimated from published segment margins. Illustrative.



Cross-references


Part I: "The Capacitor Stack" (800V DC at the rack).
Part II: "The Wide-Bandgap Stack" (SiC and GaN conversion 800V→48V).
Part III: "The Thermal Stack" (heat rejection at every stage).
Part IV: "The Interconnect Stack" (rack-level current distribution).
Part V: "The On-Package Delivery Stack" (48V to 0.8V at the transistor).










The AI Power Chain: series companions

 * The AI Power Chain: Vendor Screen. The vendor map across every layer
 * Pricing Under Scarcity. Where premium capture is compounding
 * The Services Inversion. Why services now command product-like multiples







Frequently asked


What is a modular datacenter?

A modular datacenter uses factory-built prefab enclosures shipped to site pre-integrated with power, cooling and network equipment. On-site work is limited to interconnection and commissioning. Delivery compresses schedule 6-12 months against traditional stick-built construction and shifts labour from construction site to controlled factory conditions.


Who are the leading modular datacenter vendors?

Compass Datacenters and Vantage Data Centers lead hyperscaler prefab deployments. Fibrebond supplies pre-integrated power enclosures. Aligned, DataBank and QTS use hybrid stick-built + prefab approaches. Vendor concentration varies by build-type (skid vs full-enclosure).


==============================================================================
# The AI Power Chain: Vendor Screen
URL: https://adikumar.co/ai-power-chain-vendor-screen/
Published: 2026-08-04
Summary: Vendor-by-vendor screen across the six AI Power Chain layers: capacitor, wide-bandgap, thermal, interconnect, on-package, modular datacenter.
==============================================================================

The AI Power Chain series · Part 13 of 15












The AI Power Chain · Vendor Screen


337 vendors · 6 stacks · sortable, filterable, weighted. Not investment advice. Diligence rank is a directional composite of moat, chokepoint, AI exposure and execution; adjust weights below to re-rank live.








Filters


Stack
Category
Geography
Size band
Type

Min diligence rank

1.0


Clear filters
↓ CSV






Weights (re-rank live)


Moat2.0
Chokepoint2.0
AI exposure1.0
Execution1.0


Default emphasises moat + chokepoint (durable pricing power). Dial up execution for operational risk. Dial up AI exposure for pure-play plays.



Reset default weights








Moat × chokepoint scatter (bubble size = AI-rev band)





Geography distribution (filtered set)










Rank
Name
Stack
Sub-cat
Geo
Type
Size
Deal status




















How to read this. Scores are directional (1-5) across four dimensions applied consistently across all 6 stacks. Diligence rank = weighted average per the sliders above. Revenue bands: T1 >$1B AI · T2 $200M-1B · T3 $50-200M · T4 $10-50M · T5 <$10M. Employee bands: S <100 · M 100-1k · L 1k-10k · XL >10k. Type: Public / Private / PE-backed / Corporate sub / State-owned / Employee-owned. Deal status is a snapshot; verify before acting. Author's compilation; not investment advice.










In brief


The AI Power Chain vendor screen scores 58 named vendors across every layer on moat, execution risk, and chokepoint concentration. Categories include: interconnect (transformers + switchgear), thermal (DLC + immersion), wide-bandgap semiconductors, capacitors, on-package delivery, and modular build. Scoring is for practitioner-side DD use, cross-referenced from every essay in the AI Power Chain series.


==============================================================================
# Pricing Under Scarcity
URL: https://adikumar.co/pricing-under-scarcity/
Published: 2026-08-05
Summary: How AI infrastructure vendors price when demand outruns capacity. Ten positions on the scarcity map and what each pricing move tells you.
==============================================================================

The AI Power Chain series · Part 14 of 15









Glossary of terms used



GPU
Graphics Processing Unit. The compute silicon at the centre of AI workloads.
GW
Gigawatt. One thousand megawatts of electrical power.
MVA
Mega-Volt Amperes. Apparent power rating unit used for transformers and switchgear.
OCP
Open Compute Project. Hyperscaler-led standards body developing open reference designs for data centre hardware.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.






ADI KUMAR · DATA CENTRE, POWER AND INDUSTRIAL TECHNOLOGYAUGUST 2026 · ~35 MIN READ

The AI Power Chain · Supplement
Commercial analysis · Pricing



Pricing Under Scarcity


Murata raised prices by up to 35 per cent in April and told the market the constraint behind it would last.1 Server assemblers spent the same quarter announcing record revenue and thinner margins,2 squeezed between component costs they could not control and customers who would not absorb them. Both were running short of capacity, which explains very little about where either of them ended up. What explains it is the set of alternatives available to their customers, and that can be worked out in advance.


By Adi KumarIndependent advisorViews are my own


Summary.

 1. Pricing power needs two conditions holding together: buyers who find it expensive to leave, and rivals who decline to undercut. Most analysis of this cycle tracks only the first, and mispredicts outcomes as a result.
 2. Cold plates and quick disconnects sit in the same cooling loop with almost identical buyer alternatives. One holds a two to three times premium and the other is losing margin. The difference is the number of qualified suppliers, which no buyer-side analysis captures.
 3. The most profitable positions belong to component suppliers whose parts are a rounding error in the buyer's bill of materials, whose failure is unacceptable, and who face two or three qualified rivals rather than a dozen.
 4. Capacity discipline has become explicit strategy. Murata is deploying record capital expenditure while guiding capacity growth deliberately below demand growth, and has stated that high-end capacity cannot be expanded quickly.
 5. The buyer response has shifted from negotiation to pre-commitment, with take-or-pay agreements, capacity reservation fees, prepayments and multi-year slot bookings moving from exotic to standard in under three years.
 6. Time has become a separately priced good. Sites with executed interconnection agreements and energised substations command a premium that has nothing to do with the equipment installed on them.
 7. Record revenue with compressing margin is the defining pattern wherever suppliers compete for a buyer who cannot easily leave, which describes more of this chain than the shortage headlines suggest.
 8. Positions expire when either axis moves. Announced capacity additions date the end of restraint more reliably than demand forecasts date the end of scarcity, and roughly 1.8 billion dollars of announced transformer capacity arriving from 2028 is the clearest example currently visible.



01The question



Why is shortage not the same as pricing power?


Most commentary on AI infrastructure carries an unstated assumption, which is that demand exceeding supply means suppliers are doing well. It holds far less often than it sounds like it should. Across the six layers of the chain I have mapped over the past year, almost every vendor is capacity constrained, and their commercial outcomes have diverged widely all the same.


Murata announced price increases of 15 to 35 per cent on high capacitance components for AI servers, effective April 2026,1 and its president said publicly that high-end capacity cannot be expanded quickly and that tightness may persist through this year and next.3 Meanwhile server assemblers are reporting record revenue with falling gross margins, because expensive components inflate the revenue line without lifting the assembly fee, and they have been negotiating to move procurement from buy-sell to consignment to protect their balance sheets. Both sets of companies have more demand than they can serve. Only one of them has turned that into money.


The difference between them is structural, and it can be worked out in advance. What follows sets out the ten positions visible in the market, the commercial instruments behind each, the two axes and observable signals for working out which position a business occupies, and what breaks each one. It is written for people who have to price, procure or underwrite in this environment.



Same shortage, opposite outcomes
Indicative gross margin direction by position in the chain during the 2025 to 2026 constraint

Author's assessment based on disclosed price actions, reported margin commentary and trade press. Directional illustration of the divergence, not company specific guidance. See annex.

Cold plates and quick disconnects sit in the same cooling loop, face the same qualification barriers, and carry the same consequences if they fail. One holds a two to three times premium. The other is losing margin. Buyer alternatives are close to identical, and only the number of suppliers differs.The case that requires two axes
02Framework



What are the two conditions for pricing power in AI infrastructure?


Pricing power in a constrained market depends on two things holding at the same time. The buyer has to find it expensive to go elsewhere, and the seller's competitors have to decline the opportunity to undercut. Either condition on its own produces nothing, since a supplier surrounded by trapped customers will still watch its margin fall if four rivals are bidding for the same socket, and the most disciplined suppliers in the world cannot price to a customer who can walk away.


Most analysis of this cycle attends only to the first condition and mispredicts outcomes as a result. The second condition is doing at least as much of the work.


Anyone who has sat through a strategy course will recognise the shape of this. The two axes correspond to buyer power and substitutes on one side and rivalry among existing competitors on the other, which is two of Porter's five forces with the other three set aside as slower-moving. What this adds is specificity rather than novelty, since buyer power is not a single dimension to be scored high or low but resolves into six distinct escape routes with different costs, different time constants and different blockers, and a supplier can face buyers who are powerless on five of them and free on the sixth. Rivalry likewise becomes tractable when reduced to four observable conditions rather than treated as an atmosphere. The framework is Porter made operational for a market where qualification cycles, not market shares, decide who can buy from whom.



The first axis: what escape costs the buyer


A buyer facing an unwelcome price has six escape routes and no others. Procurement strategies, qualification programmes, design reviews and the occasional furious phone call are all attempts to open one of these six doors.



Escape routeWhat the buyer doesWhat blocks it

1. SubstituteBuys the same thing from someone elseSupplier concentration, and the length and cost of the qualification cycle. A second source that exists on paper but needs eighteen months of validation is not an escape this year
2. NegotiateApplies commercial pressure to hold or reduce priceThe component's share of the buyer's total cost. Below roughly one per cent of system value, nobody staffs a cost reduction programme regardless of the percentage increase
3. DowngradeAccepts a cheaper part with worse performance or reliabilityConsequence of failure. Where failure is catastrophic and traceable to the component, this door is welded shut and both parties know it
4. WaitDefers purchase until the constraint easesThe capacity timeline, and the buyer's own schedule pressure. A buyer racing a competitor to deploy capacity cannot wait, whatever the price
5. EliminateRedesigns the requirement out of the architectureDesign lock-in for the current platform generation, and the engineering cost of the change. This door is closed today and open at the next architecture
6. In-sourceKeeps the requirement and builds the capability internallyEngineering capability, capital, and volume sufficient to carry fixed cost. Available only to the largest buyers, which makes it the door that separates hyperscalers from everyone else




Three of these look like variations on using something else, and the distinction between them is worth defending because it carries most of the analytical weight. Substitution keeps the function, the specification and the design, and changes only the supplier; the blocker is qualification. Downgrade keeps the supplier relationship and the function while accepting a worse specification, so the blocker is failure consequence rather than qualification. Elimination removes the function from the architecture entirely, and the blocker is the design freeze. In-source keeps everything and changes only who manufactures it, blocked by capability rather than by anything the supplier controls. Four different blockers, four different time constants, four different people inside the buyer who would have to sponsor the attempt. Collapsing them into a single substitutes score loses the information that predicts behaviour.


In-source deserves particular attention because it is the door most often misread. When a hyperscaler builds its own accelerator it has not eliminated the requirement for accelerators, and it has not substituted one vendor for another. It has taken the function inside, which is a different move with a different cost structure and a different set of consequences for the incumbent. Treating custom silicon as elimination understates what it does, since the buyer that in-sources becomes a permanent structural competitor rather than a lost account.


Escape cost is the price of the cheapest door still open. The doors run on very different clocks, which matters more than how many of them there are. Negotiation takes weeks, while substitution takes twelve to twenty-four months in most of this chain, since qualification rather than manufacturing is the binding step. Waiting depends on the capacity timeline, eighteen to twenty-four months for a new high-end capacitor line3 and several years for large transformer capacity.10 Elimination waits for the next architecture, two to four years out. In-sourcing runs longer still, typically three to five years from decision to volume, and is available to perhaps five buyers in this market. A supplier whose only open doors are elimination and in-sourcing has a full platform generation of protection, which explains why design-in position matters more in this industry than market share does.


Buyers routinely misjudge which door is cheapest. The instinct is to negotiate, because it is fast and needs no capital. In this chain negotiation returns the least value per hour spent and funding a second qualification usually returns the most, but qualification requires a budget line where negotiation does not. Organisations reach for the free option even when the paid option costs less in total, and suppliers benefit from that accounting asymmetry for years.



The second axis: whether rivals hold the line


Trapped buyers create an opportunity. Whether anyone collects it depends on the behaviour of everyone else who could serve the same requirement. Four conditions determine that, and they can be assessed from public information.


 * Number of qualified rivals. Three or four firms can observe each other's behaviour and infer intent. Twelve cannot. The threshold is lower than most competition analysis assumes, because qualification thins the field well below the apparent supplier count.
 * Capacity utilisation across the field. A rival running at ninety per cent has nothing to gain from winning share it cannot serve. A rival with idle lines has every reason to bid. Restraint is easy when everyone is full and evaporates the moment someone is not.
 * Cost structure and its visibility. Where all participants face the same input inflation and can see that they do, an increase reads as necessity rather than opportunism, and following it carries no reputational cost. Metals inflation performs this function usefully across several layers of this chain.
 * Memory of the last cycle. Industries that have been burned behave differently. The 2017 capacitor cycle, where capacity chased a spike and prices fell back afterwards, sits behind the discipline visible in that market today. Industries without that scar tissue tend to compete capacity away.


The clearest evidence of restraint is what happens after one supplier moves. When Murata announced in March 2026, Samsung Electro-Mechanics, Taiyo Yuden, Yageo and Walsin followed within a quarter rather than holding price to take share.4 That decision, taken independently by four boards, is worth more to Murata's realised margin than the qualification barrier is. Capacity discipline reinforces it: Murata is deploying record capital expenditure while guiding capacity growth below demand, and its president has said publicly that high-end capacity cannot be expanded quickly.3



What the two axes produce


Realised pricing power is the product of the two, not the sum. Four combinations follow, and each has a distinct mechanism, a distinct failure mode and a distinct set of firms occupying it.



The two-axis map
Buyer escape cost against supplier restraint. Positions are numbered to match the sections below

Durable pricingFragile premiumValue dissipatedCommodity
Author's positioning. Escape cost reflects the price of the buyer's cheapest open door. Restraint reflects the likelihood that qualified rivals decline to undercut. The upper-left quadrant is where most analysis of this cycle assumes everybody sits.



The quadrant that matters analytically is the lower right, where buyers are trapped and suppliers compete anyway. It is invisible to any framework built on buyer alternatives alone, and it is where a good deal of this chain actually sits.


03Positions



Ten positions, sorted by quadrant


What follows describes commercial positions rather than whole companies, and several firms occupy more than one across different product lines. Each is examined for which doors are shut, what the competitive field is doing, and what would move it.



Quadrant A. Trapped buyers, disciplined suppliers


Both conditions hold, so scarcity converts into margin. This is where the money in the current cycle has concentrated, and it contains fewer firms than the commentary suggests.


Position 01 · Escape cost high · Restraint highThe disciplined oligopolist


Murata told the market in March 2026 that prices on AI server high-capacitance parts, high-end automotive grades and RF components would rise by 15 to 35 per cent from the first of April.1 Samsung Electro-Mechanics, Taiyo Yuden, Yageo and Walsin followed within a quarter, and Yageo took the increase across its entire capacitor portfolio.4 Reporting through the middle of the year suggested room for another 20 to 30 per cent on high-capacitance parts before the cycle finished,5 with general-purpose grades moving 6 to 13 per cent behind them.6


On the escape axis, three doors are shut at once. Substitution is blocked because concentration is real and qualification for high-capacitance automotive and server grades runs twelve to eighteen months. Negotiation is blocked by salience, since even at 35 per cent the component remains a small fraction of a server's cost. Waiting is blocked by the capacity timeline, and this is the part the industry has managed deliberately: Murata's new high-end facility will not produce before the fourth quarter of 2026, and a line for sub-micron dielectric parts takes eighteen to twenty-four months to bring up.3


On the restraint axis, the field is four or five firms who can all see each other, all running high utilisation, all facing the same metals inflation, and all carrying the memory of 2017. Following an increase costs them nothing and undercutting would cost them a great deal. Announcing publicly is what a supplier does when it expects the field to follow. A supplier that rations without announcing while holding list price is signalling that it does not.


The doors left ajar define the ceiling. Buyers can downgrade partially, moving commodity portions of a bill of materials to tier-two and Chinese suppliers, and the increases concentrate in high-capacitance and automotive grades and stay light on standard parts for that reason. Elimination is a later-generation threat arriving through higher capacitance density in fewer parts, or through silicon capacitors displacing ceramic content inside the package.


Position 02 · Escape cost very high · Restraint absent but irrelevantThe monopolist


NVIDIA has held gross margin around 75 per cent through a substantial mix shift, an unusual combination of constrained supply, rising unit value and stable cost structure.7 All five doors are effectively closed, and the restraint axis does not apply, since restraint is a question about rivals and there are none of consequence at this node.


The mechanism worth attention here is allocation rather than price. Supply to the largest customers runs through bilateral roadmaps, reserved capacity and custom configurations with planning horizons measured in years, while everyone else receives standardised blocks through channels and integration partners. Reserved capacity prices 40 to 70 per cent below on-demand rates, with the large cloud providers publishing committed-use discounts of up to 72 per cent,8 so the discount belongs to buyers who can commit capital and forecast demand years ahead. Reserved rates themselves rose roughly 40 per cent between October 2025 and March 2026, which tells you the discount is being granted from a rising base rather than a falling one.8 Deciding who receives supply shapes which customers scale, which geographies build, and which competitors survive long enough to matter.


The only door with a visible handle is in-sourcing, and the buyers who can reach it are already trying. Custom silicon programmes at the largest cloud operators are not attempts to eliminate the requirement for accelerators or to substitute one vendor for another, they are attempts to take the function inside, and they cost enough that perhaps five organisations worldwide can attempt them. The price of the cheapest escape is a lower bound on what a position is worth, which makes the cumulative spend on those programmes a running valuation of this one.


For everyone downstream the consequence runs in the opposite direction. A buyer with this much leverage standardises its suppliers' designs and compresses their margins because nothing prevents it, and the vendors specified into its reference architectures absorb that regardless of how short their own capacity is.


Position 03 · Escape cost high · Restraint structuralThe qualification rentier


A universal quick disconnect costs around a hundred dollars. A rack costs millions and contains more than a hundred of them, any one of which can put water onto energised hardware. Three firms hold over eighty per cent of that market and none of them has announced anything.13


Three doors are shut and they reinforce each other. Substitution needs certification, and fewer than fifteen firms worldwide mass-produce complete units against more than two hundred making components for them.13 Negotiation is pointless at a hundred dollars against a rack costing millions, so nobody staffs it. Downgrade is unavailable because the failure mode is catastrophic and traceable, which means the engineer who approved the cheaper coupling owns the outcome personally.


Restraint here is structural rather than behavioural. The three incumbents do not need to coordinate, because certification limits the field to firms who all face the same barrier and none of whom gains share by cutting price. Thermal interface materials sit in the same shape, where grades qualified for accelerators sell at many multiples of commodity equivalents on a line item nobody scrutinises.


Waiting is partly open, so a buyer can hold inventory and defer, which caps the premium at roughly the cost of carrying stock. The real threat is elimination, and it is the one this position watches least. Every coupling exists because a serviceable joint is required at that point in the loop. An architecture that reduces joints, or moves cooling into the package so the joint disappears, removes the position without any competitor having to win anything. Investors miss this position on the way in because the revenue lines are too small to screen for, and miss the risk on the way out for the same reason.


Position 04 · Escape cost high · Restraint irrelevantThe standard-setter


A supplier that controls the reference specification occupies a different position from one that controls shipment, and the two are frequently confused. Where the allocation gatekeeper decides who receives product, the standard-setter decides what the product has to be, which means it holds the keys to the redesign door for everybody else in the chain.


The mechanism is worth stating carefully. When a reference architecture specifies an interface, a connector geometry, a voltage class or a qualification protocol, every supplier downstream builds to that specification and every buyer downstream inherits it. A buyer that wants to redesign a component out of its bill of materials cannot simply do so, because the resulting system no longer conforms and loses vendor support, warranty position, or eligibility for the reference design entirely. The elimination door, which is the strongest escape available in most positions, requires the standard-setter's consent.


NVIDIA occupies this position alongside the allocation one, and the two should be assessed separately since they can be lost independently. The clearer examples sit further from the headlines. US Conec holds the MPO and MTP multi-fibre connector designs that most data centre optical cabling conforms to, which means a change in how fibre terminates inside a hall is a decision about somebody else's intellectual property. Amphenol and Molex occupy the equivalent position in board-level and backplane connectors, where a design win at the specification stage propagates through every subsequent generation of a platform. Open Compute Project working groups occupy a collective version, where the specification is shared in fact as well as name, but the firms that author it shape what becomes buildable, and the chairs of those groups exercise influence that appears nowhere on a market share table.


What all of these have in common is that the position outlives any particular product. An installed base of conforming equipment carries the specification forward long after the generation that introduced it. Connector standards are among the most durable commercial positions in electronics for that reason, and among the most under-analysed relative to their value.


Restraint does not really apply here, since the position is definitional rather than competitive. What ends it is a rival specification achieving sufficient adoption, which happens rarely and slowly, or a standards body deliberately publishing an open alternative. That second mechanism is why the open specification efforts described in Quadrant B matter well beyond the vendors they directly threaten. Publishing an interface specification reopens the redesign door for an entire layer of the chain at once.



Quadrant B. Free buyers, disciplined suppliers


The premium exists because customers have chosen to stop shopping rather than because they cannot. It holds while that choice holds, and the reversal requires no capital, only a decision.


Position 05 · Escape cost moderate · Restraint by agreementThe bundled incumbent


Eaton bought Boyd Thermal. Schneider bought Motivair. Vertiv assembled power, cooling and services and carries a backlog above fifteen billion dollars, representing roughly twelve to eighteen months of forward revenue.15 The shared thesis is that an integrated grid-to-chip scope is worth more than its components, because it removes interface risk and shortens schedule in a market where schedule binds hardest.


There is substance to the argument, since a single accountable party across two coupled systems has real value and the buyer's alternative is carrying integration risk itself. What distinguishes this position within the framework is that the doors are closed by the buyer's own choice and not by structure. A customer that wants one throat to choke has given up substitution and downgrade on individual components voluntarily, and can take both back at the next renewal.


Restraint among the platforms is currently high, since all of them are pursuing the same integration thesis and none gains by unbundling first, and that shared commitment is the fragile part of the position. The mechanism to watch is open specification, because hyperscalers have unbundled every supplier who tried this before and the standardisation efforts that accelerate liquid cooling adoption also make it possible to tender components separately. A buyer that publishes an interface specification has reopened substitution without negotiating anything.


Position 06 · Escape cost moderate · Restraint from scarcity of alternativesThe time seller


This position and the slot seller in Quadrant C are cousins, since both sell a place in a queue rather than a thing. They are separated here because the queues behave differently under the framework. A transformer factory queue shortens when capital arrives, which moves both axes on a schedule that can be read from capacity announcements. A grid interconnection queue shortens through regulatory process, which no supplier controls and no amount of capital accelerates. That difference puts them in different quadrants and gives them different expiry dates.


Two identical sites can be worth very different amounts based on nothing more than their position in an interconnection queue. Sites with executed interconnection agreements, energised substations and headroom above current load deliver capacity sooner, and that speed carries a premium unrelated to anything installed on the ground.


One door is firmly shut, which makes this position unusually pure. The buyer cannot wait, because waiting is the thing being sold. Every other door stands at least partly open: a buyer can select a different site or market, negotiate commercial terms, accept a phased energisation, or eliminate the constraint through behind-the-meter generation. Each is expensive and slow, which supports the premium, and none is blocked outright.


Restraint holds because the alternatives are physically scarce rather than because developers coordinate. The premium prices at the cost of the buyer's next-best schedule and decays as queues clear, as permitting reform lands, and as gas turbines and fuel cells become routine enough to commoditise speed. Of the positions that make money, this is the least durable and the most exposed to regulatory process rather than competitive behaviour.



Quadrant C. Trapped buyers, undisciplined suppliers


This is the quadrant a buyer-side framework cannot see. Escape costs are high, the market is short, and margins fall anyway because the qualified field competes the opportunity away. Any analysis that equates scarcity with pricing power will misread every firm in it.


Position 07 · Escape cost high · Restraint absentThe crowded specialist


Cold plate manufacturing has all the features that should support pricing. Qualification into an accelerator platform takes months, switching mid-programme is disruptive, thermal failure damages expensive silicon, and demand exceeds what the industry can comfortably serve, so buyers are not in a position to move freely.


Margins are compressing regardless. Four Taiwanese manufacturers plus Boyd, CoolIT and a growing Chinese field are bidding for the same reference-design sockets, and the buyer holds the design authority that decides which of them wins. NVIDIA has been standardising cooling designs and compressing supplier margins across the Rubin transition. The tactic works because the field will underbid rather than hold. Chinese manufacturers captured roughly thirty-five per cent of cold plate volume through copper-aluminium hybrid designs. That is what an undisciplined field looks like when a cost-advantaged entrant arrives.


The instructive comparison is with couplings. Both sit inside the same liquid cooling loop, both are qualification-gated, both carry meaningful failure consequences. Couplings have three suppliers and hold a two to three times premium. Cold plates have a dozen and do not. The escape axis reads almost identically for the two. Only the restraint axis explains the outcome.


Exit from this quadrant requires either consolidation, and the acquisitions in thermal have been about buying out the field rather than adding capability, or a proprietary step that thins the qualified list. Microchannel and laser-welded designs may do the second, since precision manufacturing at that tolerance eliminates part of the field on capability rather than price.


Position 08 · Escape cost high now, falling · Restraint high now, fallingThe slot seller


Substation transformer lead times have gone from roughly 140 weeks in 2023 to beyond 160 weeks in 2026,9 with prices reported up 77 per cent against 2019 levels.10 At that point the vendor has stopped selling a product and is selling a position in a manufacturing queue.


Both axes are currently favourable. Buyer schedule pressure closes the waiting door completely, since a developer racing to energise cannot defer and a utility with an interconnection commitment cannot either, and elimination is closed because no data centre runs without transformers. Restraint holds because every qualified manufacturer is full, and a full factory has nothing to gain from price competition.


Hyperscalers secure their position through long-term take-or-pay agreements with the major manufacturers,11 and the residue of that prioritisation runs down the queue. Standard panelboards now sit at 28 to 48 weeks, low-voltage switchboards around 52, and power circuit breaker switchboards beyond 84.12 Two tiers of buyer receive different prices and different dates for the same equipment. Wood Mackenzie's supply chain analyst reports manufacturers returning to customers holding year-old purchase orders to impose twenty per cent increases simply to hold the delivery schedule.9 Pricing power looks like this when it is exercised retroactively.


This position is placed in Quadrant C rather than A because both axes are moving against it on a known schedule. Announced North American manufacturing expansions total roughly 1.8 billion dollars, with new capacity projected online by 2028.10 That capital opens the substitution door and removes the utilisation condition behind restraint at the same time, placing it in a quadrant defined by dissipation even though it is presently profitable. The date is unusually legible for a forecast of this kind.



Lead time as the pricing instrument
Reported 2026 lead times for electrical equipment categories, weeks

Compiled from Terrapin Consulting Group project procurement data,12 Electronate estimator guidance12 and Wood Mackenzie commentary.9 Ranges vary by manufacturer, voltage class and order quantity. Midpoints plotted.




Quadrant D. Free buyers, undisciplined suppliers


Neither condition holds, and shortage produces workload instead of margin.


Position 09 · Escape cost low · Restraint absentThe squeezed integrator


Server assemblers are reporting record revenue and shrinking gross margins in the same quarter.2 Expensive accelerators and memory inflate the revenue line without lifting the assembly fee proportionally, so the percentage falls as the absolute numbers rise. A Vera Rubin rack is procured from assemblers at around 7.8 million dollars, close to double the previous generation, with the increase concentrated in content the assembler does not profit from.14


Every door is open to this vendor's customers. Substitution is straightforward because assembly capability is replicable and several credible firms compete for the same programmes. Negotiation is well staffed because assembly is a visible line that procurement teams understand. Downgrade is available through simpler configurations, and elimination is available because the largest buyers can bring assembly in house. Running short of capacity does nothing to change any of that.


The defensive moves show where the power sits. Assemblers have been negotiating to shift customers from buy-sell procurement to consignment, which relieves the balance sheet of financing components it earns almost nothing on.2 A supplier moving to consignment has conceded that it cannot earn a return on the working capital it was carrying, and that concession is a more honest disclosure than anything in the accompanying commentary.


The route out is to acquire a closed door on one axis or the other. Vertical integration into a scarce component, a proprietary thermal or power capability, or a services annuity all qualify. Scale alone does not, and consolidation among integrators tends to disappoint for the same reason.


Position 10 · Escape cost low · Restraint symmetricalThe index passer-through


Cable manufacturers and anyone carrying heavy copper or silver content contract with escalation clauses linked to metal indices. Murata has cited silver costs among the drivers of its adjustments, and silver runs through sintered die attach in power modules as well.


An index clause does not close a door on either axis. It moves an input cost across the table, and protecting a margin percentage is not the same as creating one. In a rising market that resembles pricing power and behaves nothing like it, because the buyer's escape routes are unchanged and the supplier has stopped absorbing volatility. The test comes when the index falls, and if the contract price follows it down, the vendor was invoicing all along.


The asymmetry of the clause is the diagnostic. A supplier with real pricing power negotiates terms that pass increases through fully and decreases through partially, while a supplier without it accepts symmetry and calls it fairness. Reading which of the two a contract contains establishes where a business sits without any market analysis at all.


Cold plates and couplings sit in the same cooling loop, face the same qualification barriers, and carry the same failure consequences. One holds a two to three times premium and the other is losing margin. Buyer alternatives are identical. Only the number of suppliers differs.Why one axis is not enough
04Measurement



Putting numbers on both axes


The framework becomes operational when the axes are measured rather than asserted. Escape cost prices out directly. Restraint is harder, but it leaves observable traces.



Pricing the doors



DoorTypical costTypical timeWhat determines it

SubstituteQualification programme, six to seven figures for a critical component, plus internal engineering12 to 24 monthsReliability data requirements, sample availability, and whether an accelerator vendor or hyperscaler certification sits in the path
NegotiateProcurement time onlyWeeksCheapest to attempt and least effective where salience is low; overused for exactly that reason
DowngradeLow in direct cost, potentially unbounded in consequenceMonthsFailure mode and attributability. Where an individual engineer owns the outcome, this door stays shut regardless of the saving
WaitDeferred revenue and inventory carry, or a missed deployment windowSet by the capacity timeline18 to 24 months for a new high-end capacitor line, several years for large transformer capacity. Compare against the buyer's own schedule value
EliminateAn architecture change, so the highest direct cost by a wide marginOne platform generation, 2 to 4 yearsOnly available at a design boundary. Cheap when a redesign is happening anyway, prohibitive when it is not
In-sourceDesign team, tooling or fab access, and fixed cost carried against internal volume alone3 to 5 years to volumeBuyer scale above all. Below a threshold the fixed cost never amortises, so this door exists for perhaps five organisations in this market and for nobody else




Two asymmetries follow and both favour the supplier. The first is timing, since doors two and three can be attempted immediately and are the weakest of the six, while doors one and five are the powerful ones and need lead time the buyer usually does not have when an increase lands. The second asymmetry is accounting, since negotiation consumes staff time already on the payroll and nothing else. Qualification requires a budget line, a business case and a sponsor willing to spend now to avoid a price later.



The escape map
Which doors are open to a buyer facing each position. Closed doors are what the supplier is charging for

Subst.Negot.Downgr.WaitElim.In-source
01 Disciplined oligopolistShutShutPartialShutLaterShut
02 MonopolistShutShutShutShutShutFew can
03 Qualification rentierShutShutShutPartialLaterShut
04 Standard-setterPartialShutShutPartialShutShut
05 Bundled incumbentPartialPartialOpenPartialOpenPartial
06 Time sellerPartialPartialOpenShutOpenPartial
07 Crowded specialistPartialShutPartialPartialLaterOpen
08 Slot sellerPartialShutPartialShutShutPartial
09 Squeezed integratorOpenOpenOpenPartialOpenOpen
10 Index passer-throughOpenPartialOpenOpenPartialPartial

Author's assessment. "Later" indicates a door closed for the current platform generation and expected to open at the next. "Few can" indicates a door available only to buyers of sufficient scale. Note that position 07 shows high escape costs on four doors and still loses margin, which is the case the restraint axis exists to explain.




Reading restraint


Restraint cannot be priced, but four observable signals indicate whether it holds. None requires access to a supplier.


 * What happened after the last increase. The most informative signal available. If one supplier moved and the field followed within two quarters, restraint is high. If the increase was discounted back through rebates or extended terms without a public reversal, it was never there.
 * Utilisation across the qualified field, not the market. Published capacity figures overstate the field, since only qualified capacity competes. A market at eighty per cent utilisation with three qualified suppliers at ninety-five behaves nothing like the headline suggests.
 * Announced capacity additions and their completion dates. These date the end of restraint more reliably than demand forecasts date the end of scarcity. Capacity arriving in 2028 tells you when a position expires.
 * Whether the field shares a cost narrative. Common input inflation that all participants can see gives every supplier cover to follow rather than undercut. Falling or divergent input costs remove it.


Applied to the cases above, these signals separate positions that the escape test alone treats as identical. Ceramic capacitors show a followed increase, high qualified utilisation, capacity dated to late 2026 and a shared metals narrative, where cold plates show none of the four despite selling into equally trapped buyers.


05Mechanisms



The commercial toolkit, and what each instrument signals


Underneath the ten positions sits a set of contractual instruments. Each carries information about the balance of power between the parties, which makes them useful as diagnostics as well as tools.



MechanismWhat it doesWhat its presence signals

Published price increaseRaises list across a portfolio, usually effective from an order-receipt dateGenuine pricing power and confidence that the constraint is durable. The strongest signal available
Selective allocationRations constrained specifications while holding list priceQueue management, not pricing power. Often precedes an increase, and its persistence without one suggests the supplier fears substitution
Long-term agreement with volume commitmentMulti-year volumes at agreed pricing, frequently with annual step-downsA trade of price for certainty. Common where the buyer is more concentrated than the seller
Take-or-payBuyer pays for reserved capacity whether or not it lifts the volumeSeller strength. Transfers demand risk to the buyer and is now standard in transformer supply to hyperscalers
Capacity reservation feeNon-refundable payment to hold a slot, separate from the unit priceTime has become a separately priced good. Watch for this spreading down the chain
Prepayment and customer-funded capacityBuyer funds plant or tooling to secure supplyExtreme scarcity, and a strategic decision by the buyer to convert a supply risk into a capital commitment
Index-linked escalationPrice tracks a metal or energy indexCost protection, not pricing power. Symmetrical clauses reveal a weaker seller than asymmetrical ones
Consignment conversionCustomer owns high-value components through the assembly processSupplier weakness. Relieves working capital while conceding the margin on the consigned content
Portfolio tieringDeliberate segmentation of capacity toward higher-value gradesRational mix management. The residual segments experience it as shortage regardless of their own demand
Bundled system pricingPrices integrated scope in place of componentsAn attempt to escape component comparison. Durable only while the buyer values integration over choice




A diagnostic. When a supplier says it is capacity constrained, ask which of these instruments it is using. A vendor that rations but has not raised price, has not asked for take-or-pay, and is moving customers to consignment without announcing it holds a weak hand whatever its order book says. A backlog will hide that for a while, where the contract structure will not.
06Scenarios



What breaks each pricing-power position, and roughly when?


Every position on this map has a specific failure mode. Naming them is more useful than forecasting the cycle, because reversals in this market tend to arrive through one identifiable mechanism and not through general softening.



ArchetypeWhat breaks itEarly indicator

01 Disciplined oligopolistCapacity arriving faster than demand grows, or a qualified second source at a much lower cost. New high-end lines take eighteen to twenty-four months, which dates the risk without removing itCapacity announcements converting to qualified output; spot prices easing ahead of contract prices
02 MonopolistCustomers reaching in-source scale, or a credible alternative architecture displacing the platform. Custom silicon programmes at the largest cloud operators are the visible expression, and the cumulative spend on them is a running valuation of the positionShare of accelerator deployment on non-monopolist silicon; changes in reserved-capacity discount depth or contract length
03 Qualification rentierA newly certified competitor, or a design change that removes the component. The second does more damage and gets less attentionNew certifications announced; reference designs eliminating the interface entirely
04 Standard-setterA rival specification achieving sufficient adoption, or a standards body deliberately publishing an open alternative. The first is rare and slow; the second happens when hyperscalers coordinateOCP or equivalent working groups publishing interface specs; new connector or interface geometries winning design-in at the next platform
05 Bundled incumbentOpen specifications and buyer sophistication. Standardisation efforts accelerate adoption and then unbundle the supplierPublication of open interface specifications; hyperscalers tendering components separately
06 Time sellerInterconnection queues clearing, or behind-the-meter generation becoming routine enough to commoditise speedQueue processing rates; permitting timelines; gas turbine and fuel cell availability
07 Crowded specialistConsolidation among the qualified field, or a proprietary manufacturing step that thins the list on capability. Without either, the constraint persists and the margin does notAcquisitions of qualified competitors; microchannel or precision-manufacturing capacity commitments; single-supplier reference-design wins
08 Slot sellerCapacity expansion completing, or demand pausing while take-or-pay obligations remain outstanding. Roughly 1.8 billion dollars of announced North American transformer capacity is directed at delivery from 2028Lead times stabilising rather than extending; secondary market appearing in reserved slots
09 Squeezed integratorAlready broken. Recovery requires either consolidation or moving into a scarcer part of the chainConsignment adoption rates; margin trajectory against revenue growth
10 Index passer-throughFalling input prices, which expose the absence of underlying pricing powerWhether contract prices fall symmetrically when indices decline





Three system-level scenarios


Scenario one: the constraint persists to 2028. Capacity additions across capacitors, transformers and thermal arrive on schedule but demand grows faster. Pricing holds across archetypes one to four, take-or-pay becomes universal, and buyer capital moves upstream into supplier funding. In this world the qualification rentiers compound without headlines and the bundled incumbents justify their acquisition multiples.


Scenario two: an orderly normalisation through 2027 and 2028. New capacity lands, lead times stabilise, and price increases stop without reversing. The oligopolists retain most of the gain because list prices are sticky downward, the slot sellers lose their premium first, and the time sellers see their advantage decay as queues clear. This is the base case implied by most published capacity timelines, and it hurts incumbents far less than the 2018 capacitor cycle did.


Scenario three: a demand pause. A pullback in AI capital expenditure, whether from model economics or financing conditions, arrives while take-or-pay commitments and reserved slots remain outstanding. The pain lands on whoever holds the obligation, not whoever holds the inventory, and in this cycle that is disproportionately the buyer. Expect renegotiation, secondary trading of reserved capacity, and a sharp divergence between vendors with genuine qualification moats and vendors who were merely early.


07Playbooks



What to do, depending on which side of the table you sit



If you are selling


 * Establish which archetype you actually occupy before designing a pricing action. The common error is a squeezed integrator behaving like an oligopolist, announcing an increase it cannot enforce and damaging a customer relationship it depends on.
 * Price the qualification, not the part. If your position rests on a certified design-in, your commercial argument is the cost and risk of requalification, not your manufacturing cost. Those are very different numbers. Only one belongs in the conversation.
 * Segment capacity deliberately and say so. Tiering toward higher-value grades is rational, but doing it silently generates customer anger disproportionate to the commercial gain. Buyers accept prioritisation they understand.
 * Ask for structure as well as price. Take-or-pay, reservation fees and prepayments convert scarcity into balance sheet strength and survive the cycle better than a price increase does. A price increase is competed away; a signed multi-year commitment is not.
 * Watch design-out risk above competitive risk. The threat to a qualification rentier is rarely a rival supplier. It is an architecture that removes the component. Anyone selling into a reference design should be reading the next reference design.



If you are buying


 * Distinguish rationing from pricing. A supplier that has not raised list price but cannot ship is telling you something about its confidence. That is a different negotiation from one with a published increase behind it.
 * Pay for structure where the constraint is durable, and refuse it where it is not. Take-or-pay on transformers with capacity arriving from 2028 is a defensible trade. The same instrument on a component with a plausible second source in twelve months is a transfer of risk you should decline.
 * Fund qualification early, not price relief late. The cheapest way out of a rentier position is a second qualified source, and the expensive part is the qualification programme, not the parts. Buyers underinvest here because the cost is visible and the benefit is not.
 * Treat design-out as a procurement lever. Architecture teams can remove a component in a way procurement teams cannot negotiate around it. The strongest response to an unassailable supplier position is to stop needing it.



If you are underwriting


 * Test whether margin is qualification-protected or scarcity-driven. The two look identical in a trailing profit and loss statement and behave completely differently through a normalisation. On any asset in this chain, this is the question I would put first.
 * Read the contract structure, not the backlog. A large order book of take-or-pay commitments is a different asset from an equally large book of cancellable purchase orders, and the difference does not appear in the headline number.
 * Price the design-out risk explicitly. Where a target's position depends on a component that a plausible next-generation architecture removes, that risk falls inside a normal hold period and belongs in the model, not the risk register.
 * Discount bundling premiums by how sophisticated the buyer base is. They hold while buyers value integration and disappear once open specifications arrive. The question is how far along that path this particular set of customers already sits.


ENDConclusion



The lesson the cycle is teaching


How tight a market is has almost nothing to do with who makes money in it. Capacitor makers, transformer manufacturers, coupling specialists, server assemblers and cold plate producers are all capacity constrained, and their results have almost nothing in common. Public price increases at one end, unadvertised premiums that never appear on a list price in the middle, and at the other end record revenue arriving with less margin than the year before.


What separates them is whether the scarcity attaches to something a buyer can reconstitute. Manufacturing capacity gets built and assembly capability gets replicated, both on timescales a determined buyer can plan around. A certified position in a design, which would take eighteen months and a requalification programme to displace, cannot be reconstituted on any timescale that helps a buyer this year. That is most of the explanation for why the best margins in AI infrastructure sit in components that barely register in a bill of materials.


For anyone building, buying or backing a business in this chain, what matters is what a buyer would have to do to stop needing you and how long it would take them. That question weighs more than how tight the market happens to be this quarter.


ANNEXSources and method



On classification


The ten archetypes are my construction, derived from observed commercial behaviour and not from any published framework. Companies are assigned to archetypes based on publicly reported actions, and several occupy more than one position across different product lines. The archetypes are analytical categories, not descriptions of whole firms.



Sources


Numbered references correspond to the markers in the text. Where a figure appears in several outlets reporting the same underlying announcement, the most detailed account is cited.


 1.  Murata price increase. Announced 17 March 2026, effective 1 April 2026 on order receipt, covering AI server high-capacitance MLCCs, high-end automotive-grade MLCCs and RF/microwave parts, at 15 to 35 per cent. Also reports high-end utilisation above 80 per cent. en.unibetter-ic.com/top-multilayer-ceramic-capacitor-manufacturers and Bloomberg via astutegroup.com/news/general/mlcc-price-increases-threaten-ai-server-build-costs
 2.  Server assembler margin compression and consignment. Record revenue alongside gross margin pressure; assemblers negotiating a shift from buy-sell procurement to consignment, plus cost reduction and diversification into ASIC and general-purpose server work. DigiTimes, 13 and 15 May 2026. digitimes.com/news/a20260513PD237 and digitimes.com/news/a20260515PD225
 3.  Murata capacity timeline and president's statement. Lead times extended from a typical 4 weeks to 20 weeks with some models out of stock; new high-end facility not complete until end 2026 with output unlikely before Q4 2026; President Nakajima's statement that high-end capacity cannot be rapidly expanded and tightness may persist through this year and next; equipment import restrictions and long factory construction cycles cited as constraints. Semicone. semicone.com/article-426.html
 4.  Peer follow-through. Murata, Samsung Electro-Mechanics, Taiyo Yuden, Yageo and Walsin all adjusting quotations; Yageo extending its 1 July increase across its full capacitor portfolio; high-capacitance AI server and automotive parts showing the strongest movement at approximately 15 to 35 per cent. FTC Electronics and Fusion Worldwide. ftcelectronics.com/news/why-mlcc-prices-are-rising and info.fusionww.com/blog/mlcc-supply-is-tightening-faster-than-buyers-can-ignore
 5.  Second-half 2026 headroom. Industry reporting indicating additional room of 20 to 30 per cent on high-capacity MLCCs and 10 to 20 per cent on mid and low-end parts; constrained families at 26 to 40 week lead times; Murata high-end capacity not online before Q4 2026. Fusion Worldwide. info.fusionww.com/blog/mlcc-supply-is-tightening-faster-than-buyers-can-ignore
 6.  General-purpose grade increases. Taiyo Yuden raising prices for low-capacitance consumer and automotive MLCCs by 6 to 13 per cent in April 2026. eeNews Europe, 20 May 2026. eenewseurope.com/en/ai-drives-mlcc-shortage
 7.  NVIDIA gross margin and allocation model. Gross margin around 75 per cent held flat sequentially through the Blackwell mix shift; hyperscale supply tied to bilateral roadmaps, reserved capacity and custom configurations, with standardised blocks served through channels for the broader market. SupplyChain360, 27 May 2026. supplychain360.io/technology/nvidia-ai-supply-chain-scale
 8.  Reserved capacity pricing. Reserved GPU capacity priced 40 to 70 per cent below on-demand rates; AWS and Azure publishing committed-use discounts of up to 72 per cent and Google Cloud up to 70 per cent for three-year commitments; H100 one-year reserved rates rising roughly 40 per cent from 1.70 to 2.35 dollars per hour between October 2025 and March 2026; providers generally unwilling to sign terms shorter than six months for premium parts. Compute Exchange, February and April 2026. compute.exchange/blogs/reserved-gpus-contract-length and compute.exchange/blogs/the-rise-of-gpu-marketplaces-in-2026
 9.  Electrical equipment market and lead times. US data centre electrical equipment market projected to grow from 20 billion dollars to 65 billion by 2030, with data centres capturing up to 40 per cent of the total US electrical equipment market against just under 2 per cent in 2020; US data centre capacity rising from approximately 24 GW to 110 GW between 2026 and 2030; lead times for critical components at 18 to 36 months; hyperscale padmount transformer demand rising from 1,573 units in 2025 to 9,395 by 2030. Primary source: Wood Mackenzie press release, 28 April 2026. woodmac.com/press-releases/data-center-demand-drives-us-electrical-equipment-market-to-$65b-reshaping-industry-dynamics Ben Boucher's comment on manufacturers returning to year-old purchase orders to impose twenty per cent increases is reported in Daily Energy Insider, 28 May 2026. dailyenergyinsider.com/featured/52100-data-center-boom-set-to-transform-u-s-electrical-equipment-market-new-report-says
 10. Transformer pricing and capacity response. Unit costs since 2019 up 45 per cent for generator step-up transformers, 77 per cent for power transformers and 78 to 95 per cent for distribution transformers; power transformers averaging 128 weeks and generator step-up units 144 weeks in Wood Mackenzie's second quarter 2025 order-based survey; supply deficits of roughly 30 per cent for power transformers. Primary source: Wood Mackenzie, 13 August 2025. woodmac.com/news/opinion/transformer-troubles-manufacturing-and-policy-constraints-hit-us-transformer-supply Corroborated by POWER Magazine, 2 January 2026, which also reports approximately 1.8 billion dollars in announced North American manufacturing expansions. powermag.com/transformers-in-2026-shortage-scramble-or-self-inflicted-crisis
 11. Take-or-pay allocation. Hyperscalers securing capacity through long-term take-or-pay agreements with Hitachi Energy, Siemens Energy and GE Vernova, leaving municipal utilities, cooperatives and industrial customers competing for remaining factory slots; hyperscalers accounting for 48 per cent of global data centre capacity per Synergy Research Group. DistroForge, 22 April 2026. distroforge.com/guides/transformer-procurement
 12. Category lead times. Standard panelboards 28 to 48 weeks; standard switchboards 52 weeks; power circuit breaker switchboards 84 or more weeks; medium-voltage switchgear 52 to 78 weeks and approaching two to three years for data centre specification. Electronate, 20 March 2026. electronate.app/blog/switchgear-lead-times-2026-data-center-boom Corroborating category ranges from active 2026 manufacturer slot reservations: 15kV switchgear 52 to 80 weeks; pad-mount transformers 40 to 65 weeks; substation transformers 5 to 50 MVA at 75 to 110 weeks; generator step-up above 50 MVA at 100 to 150 weeks. Terrapin Consulting Group, 10 June 2026. terrapincg.com/news/switchgear-transformer-generator-lead-times-2026
 13. Quick disconnect market structure. More than 100 UQDs per GB200 rack; Stäubli, CPC and Parker holding above 80 per cent of the Chinese market in 2024 with share falling through 2025; fewer than fifteen firms mass-producing complete units against more than two hundred component makers; overseas units at RMB 80 to 120 against RMB 30 to 50 domestic. Tianxia Gongchang Research, June 2026, as cited in the thermal analysis in this series. Single-source; treat the structural claims as solid and the precise figures as one researcher's view.
 14. Rack procurement pricing. Vera Rubin VR200 rack procured from assemblers at approximately 7.8 million dollars, close to double the GB300 Blackwell rack; discussion of consignment shifting cash flow pressure while compressing assembler gross margins. Morgan Stanley via Tiger Brokers. itiger.com/news/1101444849
 15. Vertiv backlog. Project backlog above 15 billion dollars, representing roughly twelve to eighteen months of forward revenue at current run rates; Q1 2026 revenue of 2.65 billion dollars; full year 2026 guidance of 13.25 to 13.75 billion dollars. Company disclosures and Motley Fool via Gotrade, May 2026. heygotrade.com/en/blog/vertiv-vrt-data-center-cooling-ai-2026
 16. Sector consolidation. Eaton acquisition of Boyd Thermal at 9.5 billion dollars, announced November 2025 and completed March 2026; Schneider Electric acquisition of Motivair; Vertiv acquisition of PurgeRite. Company disclosures and trade press, as compiled in the thermal analysis in this series.



Estimated content


The margin direction chart and the five-test scoring chart are my assessments, constructed from the disclosed price actions and margin commentary cited above together with the structural reasoning set out in each section. They illustrate relative position and direction. They are not measurements of any company's margins and should not be used as such. Where a claim rests on a single source, that is stated in the text.



Series. The AI Power Chain covers six layers of the physical infrastructure behind AI compute. This supplement covers the commercial layer that runs across all of them. A companion piece examines the shift from equipment to service revenue.


The author is an independent advisor working on data centre, power and industrial technology. This analysis is written in a personal capacity and rests entirely on public information. Nothing here is investment advice.


© 2026 Adi Kumar · adikumar.co










In brief


AI infrastructure pricing power is compounding in specific layers, cooling capacity, high-voltage cables, transformer core steel, wide-bandgap wafers, where demand has outrun installed capacity. Where scarcity is structural (multi-year lead times, no substitute), price capture flows to the vendor. Where scarcity is transient, price capture normalises within 12-18 months.


==============================================================================
# The Services Inversion
URL: https://adikumar.co/services-inversion/
Published: 2026-08-06
Summary: The AI infrastructure services layer captures margin like the product used to. Seven positions where service annuities compound faster than hardware.
==============================================================================

The AI Power Chain series · Part 15 of 15









Glossary of terms used



GW
Gigawatt. One thousand megawatts of electrical power.
MW
Megawatt. Unit of electrical power. A large modern data centre draws tens to hundreds of MW.
UPS
Uninterruptible Power Supply. Battery-backed power system that keeps critical loads running during grid disruptions.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.






ADI KUMAR · DATA CENTRE, POWER AND INDUSTRIAL TECHNOLOGYAUGUST 2026 · ~40 MIN READ

The AI Power Chain · Supplement
Commercial analysis · Service economics



The Services Inversion


In a year when the obvious move was to buy more product capability, Vertiv spent around a billion dollars on a business that flushes and commissions cooling loops. The equipment going into data centres now needs more looking after than anything the industry has shipped before, and it is arriving considerably faster than the people who know how to look after it. Whether the manufacturers capture that work or watch somebody else do it depends on two conditions. Only one is currently in their favour.


By Adi KumarIndependent advisorViews are my own


Summary.

 1. A service annuity requires two conditions holding together. The customer has to be unable to do the work itself, and the manufacturer has to be able to keep independent maintainers out. Most analysis of servitization addresses the first and assumes the second.
 2. Liquid cooling currently satisfies both, making it the most valuable service opportunity in this cycle. An equipment vendor paid around a billion dollars for loop flushing and commissioning capability on exactly that basis.
 3. Both conditions decay. Customers build capability as a technology matures, and European right to repair provisions effective January 2026 require manufacturers to supply spare parts and technical documentation for ten years after a model is discontinued, which attacks defensibility directly.
 4. The independent maintenance market already runs at roughly 3 to 12.5 billion dollars depending on scope and undercuts manufacturer contracts by around 30 to 38 per cent. It is the competitor most equipment vendors leave out of their service plans.
 5. Labour scarcity is the strongest force behind the opportunity and the tightest constraint on capturing it, since manufacturers recruit from the same pool as their customers.
 6. Five service models exist and they differ in defensibility rather than in margin. Monitoring and outcome contracts are defensible. Break-fix and scheduled maintenance are not, whatever the contract says.
 7. The published forecast of an inversion to roughly eighty per cent service revenue within four to five years is aggressive. The direction is right and the pace assumes labour that does not exist.
 8. Service transformations fail on compensation design and organisational placement far more often than on customer demand, and the diagnostic questions are simple enough to ask on a single management call.
 9. The window for manufacturers to establish position runs roughly three to five years, after which customers will have built capability and independents will have qualified.



01The claim



An industry selling equipment into a market that needs operations


While researching the thermal layer of this chain I came across a forecast that I noted and moved past, and which has bothered me since. The claim was that the revenue mix in data centre cooling would invert, from roughly eighty per cent solutions and twenty per cent services today to something close to the reverse within four to five years. That is a large claim and it deserves either substantiation or dismissal rather than a passing mention.


Having spent time on it, the direction holds and the pace does not. What interests me more is the question the forecast skips over. An installed base generates maintenance work whether or not the manufacturer that supplied it does the work. Aviation engine makers capture most of the service revenue on their installed bases. Server manufacturers capture rather little of theirs, and an independent maintenance industry worth billions has grown up doing the work instead. Both industries sell complex equipment to customers who cannot easily maintain it. The outcomes are opposite.


So the useful question is not whether data centre infrastructure generates a service annuity, since it certainly will, but which businesses end up collecting it.


Aviation engine makers capture most of the service revenue on their installed bases. Server manufacturers capture very little of theirs. Both sell complex equipment to customers who cannot maintain it themselves.The question this essay is about
02Framework



What are the two conditions for a service annuity?


A manufacturer captures service revenue when two things hold at the same time. The customer has to find it hard to do the work in house, and the manufacturer has to be able to keep independent providers from doing it instead. Either condition alone produces nothing useful, since a customer who cannot self-perform but can hire a third party will hire the third party, and a manufacturer with impregnable defences against competitors has nothing to sell to a customer that maintains its own equipment.


The servitization literature has covered the first condition thoroughly since the late nineties, when Wise and Baumgartner argued in the Harvard Business Review that manufacturers should move downstream into the installed base because that is where the profit pool sits. The argument was correct then and remains so, and what it leaves open is the competitive question of who gets the downstream revenue once a manufacturer decides to pursue it, and that is settled by the second condition rather than the first. What follows is the servitization argument with the competitive dimension added, which is what turns a strategic aspiration into a testable proposition.



The first axis: the customer capability gap


Whether service gets bought at all depends on how far the work sits from what the customer can do for itself, and four things determine how wide that distance is.


 * Technology novelty. A facility team that has maintained chillers and switchgear for twenty years has no accumulated experience of coolant chemistry, loop commissioning or leak response, because nobody did at scale five years ago. Novelty produces the widest gaps and those gaps close fastest.
 * Consequence of failure. A fault in a fifteen kilowatt rack is an inconvenience. A fault in a six hundred kilowatt rack carrying accelerators worth millions, midway through a training run that has been executing for weeks, belongs to a different category of event. As the value at risk rises, customers become less willing to learn on the job.
 * Skill scarcity in the open market. If a customer can hire the capability, the gap closes regardless of novelty. When commissioning specialists are committed to builds twelve to eighteen months ahead and operators report worsening difficulty finding qualified staff, hiring stops being an option and the gap stays open.2
 * Scale of the customer's own estate. An operator with three sites cannot justify a specialist team. An operator with three hundred can, and will. That is why hyperscalers internalise operations that colocation providers and enterprises outsource, and why the same technology produces different answers depending on who owns it.



The second axis: defensibility against independents


A wide capability gap creates demand for service. Whether the manufacturer collects it depends on how hard it is for somebody else to step in. Five mechanisms decide that, and each behaves differently under pressure. Manufacturers tend to rely on the weakest of them.


 * Parts access. The oldest mechanism and the one most exposed to regulation. Where a manufacturer controls spares, an independent cannot promise a response time it is unable to meet.
 * Diagnostic and data access. Where equipment reports its condition to the manufacturer's platform and nowhere else, the independent arrives without information the manufacturer already holds. This mechanism strengthens as equipment becomes more connected, which makes it the most durable of the five.
 * Warranty and liability linkage. Where third-party intervention voids warranty or shifts liability, the customer's procurement function closes the option before it reaches the engineering team.
 * Certification requirements. Where a customer, insurer or accelerator vendor requires certified personnel, whoever controls the certification controls the field.
 * Proprietary tooling and procedure. Where the work needs equipment or knowledge that exists only inside the manufacturer, an independent has to reverse engineer it before bidding. This is the weakest mechanism, since it decays as the installed base grows and knowledge diffuses through the labour market.


The mechanism most manufacturers actually rely on is the one they never chose, which is that nobody else knows how yet. That protects a position for a few years and disappears as soon as an experienced technician takes a job elsewhere, which in a labour market this tight happens continuously.



The service capture map
Customer capability gap against manufacturer defensibility. Positions are numbered to match the sections below

Manufacturer annuityIndependents capture itIn-house with OEM partsNo service market
Author's positioning. The upper-right quadrant is where manufacturers want to be and where liquid cooling currently sits. The lower-right is where a good deal of data centre equipment has already ended up.

03Quadrants



Where the work actually goes



Quadrant A. Wide gap, defensible


Both conditions hold and the manufacturer collects. This quadrant contains less equipment than manufacturers assume, and it is where liquid cooling sits today.


Position 01 · Gap wide · Defensibility high, for nowLiquid cooling loops


An air-cooled hall is a forgiving environment. Fans fail gracefully, uninterruptible power systems are well understood, and maintenance follows schedules the industry has refined over decades. Direct-to-chip cooling replaces that with a fluid circuit running pumps, filtration, chemistry and pressure management through thousands of connection points directly above energised equipment.


Every element of that is a service somebody must be paid to perform. Loop flushing and commissioning is a discrete high-skill task carried out at installation and periodically afterwards. Coolant chemistry management combines consumables with expertise on a recurring basis. Leak detection sits somewhere between monitoring and insurance. None of the three was a meaningful revenue line in 2020, and the capability gap is currently very wide as a result.


Defensibility today rests on novelty, certification and warranty linkage. A manufacturer that specifies its own commissioning protocol, certifies the technicians permitted to perform it, and links warranty to compliance holds all three. What it does not hold is any barrier that survives the technology becoming ordinary, and that is the whole of the strategic problem in this quadrant.


Position 02 · Gap wide · Defensibility structuralHigh-voltage electrical systems


Medium and high voltage switchgear, protection relays and their commissioning occupy a more durable version of the same position. The gap is wide because the work is dangerous, the consequence of error is severe, and the skills were scarce well before the AI cycle began. Defensibility is stronger because certification here is a regulatory and insurance requirement rather than a manufacturer's invention, and because protection settings and relay configuration sit inside proprietary engineering tools.


The lesson for cooling vendors is available in this position if they care to read it. Certification that is external and mandatory sustains defensibility long after novelty has worn off, where certification a manufacturer invented for itself lasts exactly as long as customers agree to respect it.


Position 03 · Gap wide · Defensibility high, contract-anchoredBehind-the-meter power as a service


The grid interconnection queue in most US ISOs now runs several years, and the datacenter industry has responded by installing gas turbines, fuel cells and lithium-iron-phosphate battery storage at the site fence to supply capacity ahead of a utility connection. That equipment has a service profile very different from the equipment it displaces. It runs continuously rather than on standby, its fuel and chemistry require constant management, and its regulatory obligations for emissions, permits and grid-services participation are ongoing rather than one-time.


The capability gap is wide because operating a gas turbine or a battery installation is not a competence a colocation provider has ever needed. Defensibility is high for two independent reasons that reinforce each other. The equipment is often financed under long-term operating agreements, which anchor the service relationship contractually, and grid-services revenue requires certification and regulatory relationships the manufacturer holds and the customer does not. Bloom Energy's power-purchase-agreement model, and the emerging BESS-as-a-service structures from Sungrow, Fluence and CATL-backed integrators, are the visible expressions.


The position is earlier in its cycle than liquid cooling and will decay more slowly, because the regulatory anchor is external and durable. Manufacturers who position here before the equipment layer commoditises hold a longer runway than the cooling analogy suggests.



Quadrant B. Wide gap, weakly defended


Customers cannot do the work and buy it from somebody else. This is the quadrant most equipment manufacturers end up in without noticing, and it is where the independent maintenance industry lives.


Position 04 · Gap wide · Defensibility erodingIT hardware and the independent maintainers


Third-party maintenance of servers, storage and networking is a market of roughly three billion dollars on a narrow definition and around twelve and a half billion on a broader one, growing at 7 to 11 per cent depending on scope.7 Independents undercut manufacturer contracts by around 30 to 38 per cent, hold regional parts depots, and report first-time fix rates that stand comparison with anything a manufacturer offers.8 Gartner now treats hybrid maintenance, meaning a deliberate mix of manufacturer and independent providers, as standard practice rather than an edge case.9


None of that happened because customers stopped needing help. It happened because the defensibility mechanisms failed one at a time. Parts became available through secondary markets. Diagnostics were reverse engineered. Warranty linkage weakened as equipment aged out of warranty anyway. What remained was a price comparison, and the manufacturer lost it.


Every equipment vendor planning a service business in data centre power and cooling should study this position closely, since it describes a likely destination rather than a distant analogy. The technology differs and the commercial dynamics do not.


Position 05 · Gap wide · Defensibility under legislative attackAnything covered by right to repair


European circular economy provisions effective January 2026 require manufacturers to make spare parts and technical documentation available for at least ten years after a model is discontinued. Independent providers operating in European markets report parts sourcing costs falling by an estimated 15 to 20 per cent for covered categories as a result.10


That legislation converts the oldest defensibility mechanism into a compliance obligation. A manufacturer relying on parts control to protect European service revenue has had the protection removed by statute rather than by competition, and no commercial response is available. Parts control should be treated as a depreciating asset everywhere, since regulation of this kind tends to travel.


The mechanism that survives is data. A manufacturer holding the operating history of an installed asset, and models trained on failures across thousands of similar assets, possesses something no legislature has yet compelled it to share and no independent can reconstruct from a single site.



Quadrant C. Narrow gap, defensible


The customer does the work and buys parts and tools from the manufacturer. Revenue is real but modest, and it behaves like a component business rather than a service one.


Position 06 · Gap narrow · Defensibility highHyperscaler-operated infrastructure


An operator running dozens of gigawatts has both the scale to justify specialist teams and the wage budget to recruit them. It will maintain its own equipment and negotiate hard on the parts and tooling it cannot make. Manufacturers serving this customer should model service as parts and training rather than contracted maintenance, and should expect the customer to want data access written into the supply agreement.


Two implications follow. Service strategies built on hyperscaler demand tend to disappoint, since that customer is structurally in the wrong quadrant. And the colocation and enterprise segments, which attract less attention in AI infrastructure commentary, are where contracted service revenue actually accumulates.



Quadrant D. Narrow gap, weakly defended


Commodity equipment maintained by whoever is cheapest. Nothing here is worth a strategy, and the only useful observation is that equipment migrates into this quadrant over time unless something is done to prevent it.


Position 07 · Gap narrow · Defensibility absentCommodity fans, PDUs, cooling towers


Every layer of the physical stack has a commodity tier. Rack PDUs after they leave the leading edge, fans past the first generation, cooling towers with generic controls, and switchgear below the medium-voltage class all belong here. The customer's facilities team already knows how to maintain them and the field for parts and labour is open. A manufacturer selling into this tier is selling equipment against price and cannot construct a service annuity around it.


The instructive point is not that this quadrant exists but that equipment tends to migrate into it. A cold plate that requires bespoke commissioning in 2026 becomes a specified line item in 2028 and a commodity in 2031, and its service profile follows the same curve. The strategic implication for anyone in Quadrants A or B is to note where the equipment sits on that trajectory, because a service business anchored to a technology moving toward this quadrant has a shelf life set by the migration, not by the customer relationship.


04Ladder



Five service models, ranked by what actually protects them


Service is a family of businesses. The five models below are usually presented as a ladder of increasing margin, which is true and beside the point. What separates them is defensibility, and two of the five have almost none.



ModelDefensibility rootMargin bandStructural weakness

Break-fix and sparesParts accessLow to moderateSecondary markets and right-to-repair legislation both attack the parts moat directly
Preventive maintenance contractsWarranty linkage (first term); merit (renewal)ModerateIndependents offer identical scope at 30-38 per cent lower; renewal is the honest signal
Monitored and predictive serviceFailure data across the installed baseModerate to highInvestment-heavy; only defensibility mechanism that strengthens rather than decays
Outcome-based and gain-shareContractual lock-in plus accumulated operational knowledgeHighestFails on baselining and attribution rather than on engineering
Full operations / as-a-serviceAsset ownership plus customer relationshipHighest (post-financing)Capital-intensive; failure mode is balance sheet rather than commercial




Break-fix and spares


Reactive repair, parts supply and time-and-materials field work. Margins on parts are respectable and on labour thin. Revenue follows failures rather than schedules, so it forecasts badly. Defensibility rests entirely on parts access, which is the mechanism most exposed to secondary markets and to legislation. Most equipment manufacturers already have this business and mistake it for a service strategy.


Preventive maintenance contracts


Scheduled inspection sold annually against a defined scope. Revenue is predictable and margins are better, which makes this the model most vendors aim for. The difficulty is that an independent can offer the same scope at a lower price and the customer can read both proposals side by side. Renewal rate tells you more about a business here than attach rate does, because the first contract is often sold on warranty linkage while the second has to be won on merit.


Monitored and predictive service


Connected equipment feeding a platform that identifies degradation before failure. Margins improve because planned intervention consumes less technician time than emergency response, which matters a great deal when technicians are the binding constraint. The defensibility argument is stronger than the margin argument. Failure data accumulated across a large installed base cannot be replicated by an independent working site by site, cannot be reconstructed by a customer from its own estate, and has not been the subject of any right to repair legislation. Of the five models this is where a manufacturer's structural advantage is real and durable, and it is the one that most requires investment before returning anything.


Outcome-based and gain-share contracts


The vendor commits to availability, efficiency or an operating envelope and is paid against performance. Margins are the highest available and so is the risk. Defensibility is high while a contract runs, since switching mid-term is disruptive and the vendor accumulates operational knowledge the customer does not hold. These contracts fail on baselining rather than on engineering. Attribution disputes about whether an improvement was caused by the vendor or by something else consume more management attention than the technical work, and the measurement methodology has to be agreed before signature rather than after the first disagreement.


Full operations and cooling as a service


The vendor owns or operates the infrastructure and sells its output. This converts an equipment business into an asset business with a financing requirement, which changes the risk profile and suits a narrow set of vendors. Defensibility is very high once established, since the vendor holds both the asset and the relationship. Execution is difficult and the failure mode is a balance sheet problem rather than a commercial one.



Margin and defensibility do not rank the same way
The five models plotted on indicative gross margin against defensibility. Bubble size indicates capital requirement

Author's assessment, constructed from industrial service business structures rather than from disclosed data for this sector. Illustrative of relative position. Note that preventive contracts, the model most vendors target, sit mid-table on both axes.

05Evidence



What the market has already done



The acquisition that made the argument


Vertiv's purchase of PurgeRite, a liquid cooling services business, for roughly a billion dollars is the clearest single piece of evidence available.1 In a period when peers were buying product capability, one of the largest players in the sector bought flushing, commissioning and loop maintenance instead.


The purchase does two things at once. It establishes a position in Quadrant A while the capability gap is at its widest, and it removes the largest objection to adopting liquid cooling in the first place, since an operator worried about what happens after installation can be told the vendor will handle it. Service capability that enables the equipment sale ahead of it is worth more than its own margin suggests.



Backlogs describe an installed base as well as an order book


Vertiv's backlog above fifteen billion dollars, representing roughly twelve to eighteen months of forward revenue,5 is normally read as evidence of demand. Every unit in it becomes an installed asset requiring maintenance for its operating life, and the service business that implies is largely unbuilt. The same holds across the order books accumulated by the major electrical platforms.



The labour market is pricing the constraint


Operators reported difficulty finding qualified candidates at 53 per cent in 2024 against 38 per cent in 2018.2 Electricians qualified for 480V busway work command 120,000 to 150,000 dollars, and commissioning specialists are committed to builds twelve to eighteen months ahead.3 Construction labour costs in primary North American markets have risen 8 to 12 per cent year over year on skilled trades scarcity alone, and roughly 439,000 positions were reported open in late 2025 across mechanical, electrical, commissioning and power and cooling roles.46


These figures are usually cited as evidence of demand for outsourced service, which they are. They are also the reason the opportunity is hard to capture, since manufacturers recruit from the same pool as their customers and meet the same shortage when staffing a field organisation.



Precedent, and why the two outcomes differ


Aviation moved from engine sales to power by the hour over roughly two decades, and the service arrangements became worth more than the original equipment margin. Elevators, medical imaging and industrial gases followed comparable paths. Each involved a complex installed asset, a customer without the capability to maintain it, and a manufacturer that understood the failure modes better than anyone else.


Server hardware had every one of those characteristics and produced the opposite result. The difference sits entirely on the second axis. Engine manufacturers control parts, certification is mandated by regulators rather than invented by vendors, and the data required to predict failure comes off equipment the manufacturer instrumented. Server manufacturers had none of that, and an independent industry took the work. Data centre power and cooling equipment currently resembles the second case more than the first, which is the uncomfortable finding of this analysis.


06Economics



Why the mix shift matters more than the revenue


A shift toward service reshapes four things at once. Margin is the visible one; the other three matter more.


Margin structure


Service typically carries higher gross margins than equipment in industrial businesses, and the gap widens moving up the ladder. Service margins are also less exposed to input cost inflation, since the dominant cost is labour the vendor manages rather than metals it does not.


Cyclicality


Equipment revenue tracks customer capital budgets. Service revenue tracks the installed base, which does not shrink when capital spending pauses. For a sector trading on multiples that assume continuous buildout, this is the most valuable property service revenue has, and it is systematically undervalued because the equipment cycle currently obscures it.


Capital intensity


Service revenue is generated on a fraction of the working capital equipment revenue requires. No inventory of finished systems, no long production cycle, no exposure to component prepayment. Return on capital employed can be several times that of the equipment business even where gross margins are only moderately better.


Customer relationship


The least quantifiable effect and possibly the most valuable. A service relationship generates continuous contact, operational data and early visibility of the customer's next expansion. The vendor holding the maintenance contract learns about the next tranche of capacity before the tender is issued, which in a market where design-in position determines share for a platform generation is worth more than the service margin itself.



What a mix shift does to a business
Illustrative comparison of an equipment-weighted and a service-weighted mix on identical revenue

Author's illustration using representative industrial margin and capital structures. Demonstrates the mechanism rather than any specific company's economics.

07Obstacles



Why will most manufacturers not execute the services inversion?


The market logic is strong enough that the interesting question is why so few equipment businesses capture the annuity. Having watched several of these transitions from inside, the obstacles are consistently internal.


Compensation comes first, whatever the strategy document says. Sales organisations are paid on equipment bookings. A service contract that is smaller in year one and larger over five is, to a salesperson on an annual plan, a worse deal. Attach rates stay low until the quota changes, and this one issue kills more service transformations than any market factor I have seen.


Organisational placement determines investment. In most equipment companies service reports into the product organisation and is measured as warranty cost recovery. Its leaders sit below product leaders, it does not control its own pricing, and it receives capital after product needs are met. Businesses that have succeeded here almost always separated service into a standalone profit and loss with its own leadership first, and treated everything else as consequence.


The capability is operational rather than technical. Service at scale needs field technicians, dispatch, parts logistics, monitoring infrastructure and training pipelines. That is a different competence from designing and manufacturing equipment and cannot be acquired incrementally, which is why acquisitions in this space have been about buying trained organisations rather than technology.


Channel conflict is a real constraint rather than an excuse. Where equipment sells through partners who currently perform maintenance, moving into service attacks the route to market. This explains why some vendors pursue monitoring and analytics, which partners cannot easily replicate, in preference to field service, which competes with them directly.


Most vendors cannot measure the thing they say they are managing. Attach rate, renewal rate and installed base by vintage are frequently unavailable without a special exercise. A transformation that cannot be measured cannot be managed, and the data work required is unglamorous enough to be deferred indefinitely.


Three questions for any vendor claiming a services strategy. Does service have its own profit and loss and a leader at the same level as product? Is the sales force compensated on service attach at the point of equipment sale? Can the company state attach and renewal rates by product line without preparing them specially? Three noes means a strategy document rather than a transition, and the question can be settled on a single management call.
08Timing



The window, and how fast it closes


Both axes decay, at different speeds, and that determines how long manufacturers have.


The capability gap closes in three to five years for any given technology. Customers hire experienced people, procedures get written down, and what was novel becomes ordinary. Liquid cooling in 2026 sits roughly where uninterruptible power sat in the early 2000s, and nobody sells a differentiated UPS maintenance service today.


Defensibility decays on a legislative clock rather than a technical one. The European provisions effective January 2026 removed parts control as a durable mechanism in that market, and comparable rules tend to spread. A manufacturer relying on parts access should assume the argument is already lost.


Data accumulation runs the other way. The only mechanism that strengthens with time is failure history collected across an installed base, which is why investment in connected monitoring is the decision that matters most and the one most often deferred in favour of hiring field technicians.



Three scenarios


The aggressive case, and why I do not hold it. A full inversion to roughly eighty per cent service within four to five years would require service revenue to grow several times faster than equipment revenue while equipment revenue is itself growing at record rates. Arithmetically that demands either a very large expansion of service scope per installed megawatt, which labour availability constrains, or a sharp equipment slowdown, which would be a different story altogether. Read the forecast as a directional statement about where value accrues.


The base case. A steady mix shift with service growing faster than equipment from a much smaller base, concentrated in liquid cooling where the service content per installed unit has no precedent. Manufacturers that separated service organisationally before the wave arrived capture disproportionate share, and independents take the work at the older, less differentiated layers of the estate.


The case that should worry vendors. Labour scarcity binds hard enough that manufacturers cannot staff the growth, and operators and independents perform the work instead. The annuity exists and the equipment makers do not collect it. This is not hypothetical, since it is what happened in server hardware, and the conditions that produced it are present again.


09Playbooks



What to do about it



If you make equipment


 * Fix compensation before strategy, since attach rate follows the sales plan rather than the intent.
 * Separate the profit and loss, because service will not grow while it is measured as a cost centre inside the product organisation.
 * Invest in monitoring ahead of field capacity. It is the only defensibility mechanism that strengthens with time, and it reduces technician hours per megawatt, which is the metric the labour constraint makes binding.
 * Design serviceability into the product. A loop designed for straightforward flushing and connections designed for safe replacement determine whether the service business is profitable or punitive, and that decision is made in engineering years before it appears in service margin.
 * Assume parts control is already gone and build the position on data and certification instead.
 * Start the training pipeline now, since apprenticeships take years to produce output and vendors that solve labour structurally will hold an advantage pricing cannot overcome.



If you operate infrastructure


 * Decide deliberately what to build in house. Capability you build leverages across sites, and capability you buy becomes more expensive as the estate grows. That decision is easier now than after several years of contracts have accumulated.
 * Negotiate service at equipment purchase, since leverage is at its maximum before the order is placed and declines steadily afterwards.
 * Write data access into the supply agreement. If the vendor's monitoring generates operational data about your facility, your rights to it are far easier to establish before signature than to recover later.
 * Treat independents as a live option rather than a fallback. The 30 to 38 per cent differential is real, and even where you do not use it, knowing the number changes the renewal conversation.
 * Handle outcome contracts carefully. They can be excellent and they fail on baselining, so agree the measurement methodology before signature.



If you are underwriting


 * Separate service revenue from equipment revenue in the model and value them differently, since blending understates a service-weighted business and overstates the resilience of an equipment-weighted one.
 * Test attach and renewal rates directly. A target that cannot produce them quickly has a warranty operation rather than a service business.
 * Distinguish contracted revenue from expected revenue. Multi-year contracted maintenance is an annuity and assumed future break-fix work is a forecast, and they are frequently presented together.
 * Establish which defensibility mechanism the service margin rests on. Parts control is a depreciating asset, novelty expires on a known schedule, and data accumulation compounds. The three deserve different multiples.
 * Ask where the technicians come from. A service growth plan with no answer on labour supply is a plan to lose the work to somebody else.


ENDConclusion



The business that arrives after the boom


Infrastructure cycles create two businesses in sequence. The first is loud, capital-intensive and cyclical, and it is the one currently being valued. The second is quiet and recurring and arrives on a lag set by how long the equipment lasts. This cycle is unusual in the compression, since the industry is installing a decade of complex and unfamiliar equipment in a handful of years into a market that cannot staff its operation.


That combination creates the annuity without determining who collects it. Aviation engine makers built businesses on their installed bases because they controlled parts, certification and data. Server manufacturers watched an independent industry worth billions grow up inside theirs. Data centre power and cooling equipment currently resembles the second case, and European parts legislation has removed one of the mechanisms that might have changed it.


What remains available to manufacturers is the failure data accumulating across their installed bases, and a window of perhaps three to five years while liquid cooling is unfamiliar enough that customers will not attempt it alone. Both are being spent slowly by companies whose sales forces are paid on shipment and whose service organisations report to product. The ones that fix those two things will find, some years from now, that the least fashionable part of the business has become the most valuable part of it. The rest will discover that somebody else did the work.


ANNEXSources and method



On the central claim


The forecast of a mix inversion from approximately eighty per cent solutions and twenty per cent services to the reverse within four to five years appeared in sector commentary on data centre cooling and was referenced in my earlier analysis of the thermal layer. I have not traced it to primary research with a stated methodology and treat it as a directional industry view rather than a substantiated projection. The assessment that the direction is correct and the pace optimistic is my own, with the reasoning in section eight.



On the framework


The two-axis structure is my construction. The first axis restates the servitization argument developed by Wise and Baumgartner and the literature following them, which I have not attempted to summarise beyond acknowledging the debt. The second axis and the five defensibility mechanisms are my own, derived from observed differences between industries that captured their installed-base service revenue and industries that lost it. Quadrant assignments are judgements rather than measurements.



Sources


 1.  Vertiv acquisition of PurgeRite. Liquid cooling services business acquired for approximately one billion dollars. Company disclosures and Reuters, 2026.
 2.  Operator hiring difficulty. 53 per cent of operators reporting difficulty finding qualified candidates against 38 per cent in 2018. Uptime Institute Global Data Center Survey 2024, cited in dcgeeks.com/data-center-workforce-shortage, April 2026, which also cites Dell'Oro Group projecting global data centre capital expenditure above 400 billion dollars in 2026.
 3.  Trade rates and commissioning lead times. Electricians qualified for 480V busways at 120,000 to 150,000 dollars; commissioning specialists committed to builds twelve to eighteen months in advance. thenetworkinstallers.com/blog/data-center-construction-statistics, May 2026.
 4.  Construction cost and capacity outlook. Labour costs in primary North American markets up 8 to 12 per cent year over year on skilled trades scarcity, per the Turner & Townsend Data Center Cost Index 2026; JLL 2026 Global Data Center Outlook projecting roughly 97 GW of additions between 2025 and 2030, with construction costs rising from 7.7 million dollars per MW in 2020 to 10.7 million in 2025. Primary sources cited directly.
 5.  Vertiv backlog and guidance. Project backlog above 15 billion dollars representing roughly twelve to eighteen months of forward revenue; full year 2026 net sales guidance of 13.25 to 13.75 billion dollars. Company disclosures and Motley Fool via Gotrade, May 2026. heygotrade.com/en/blog/vertiv-vrt-data-center-cooling-ai-2026
 6.  Open positions. Approximately 439,000 open positions reported in late 2025, concentrated in mechanical, electrical and plumbing systems, commissioning, and power and cooling infrastructure. irecruit.co/insights/data-center-job-market-outlook-2026, June 2026. Single source. The scale of the shortage is corroborated by Uptime Institute and JLL, but the specific 439,000 count rests on one industry-recruiter aggregation and should be treated as directional.
 7.  Third-party maintenance market size. Estimates diverge by scope. Narrow definition: 3.01 billion dollars in 2025 rising to 6.27 billion by 2032 at 11.2 per cent. QY Research via openPR, 28 May 2026. openpr.com/news/4529614/third-party-maintenance-for-data-center-market-report Broader definition including network hardware: 12.5 billion dollars in 2024 rising to 20.3 billion by 2033 at 6.9 per cent. Verified Market Reports, April 2026. verifiedmarketreports.com/product/data-center-and-network-third-party-hardware-maintenance-market The range is quoted rather than reconciled.
 8.  Independent provider economics. Cost savings of approximately 30 to 38 per cent against manufacturer contracts, consistent with Gartner Peer Insights market descriptions; Park Place Technologies operating twelve regional parts depots globally, with first-time fix rate reported by the vendor. QY Research via openPR, May 2026, and Market Reports World, May 2026. Vendor-reported operational metrics should be treated as marketing rather than audited.
 9.  Hybrid maintenance as standard practice. Gartner's market definition describes a hybrid strategy combining manufacturer-authorised and independent maintenance as increasingly common, with customers using independents to extend asset life and control manufacturer-forced upgrades. Gartner Peer Insights market description, 2026. gartner.com/reviews/market/data-center-and-network-third-party-hardware-maintenance
 10. European right to repair provisions. Circular Economy Action Plan right to repair provisions effective January 2026 requiring manufacturers to make spare parts and technical documentation available for at least ten years after a model is discontinued; independent providers in European markets reporting parts sourcing costs reduced by an estimated 15 to 20 per cent for covered categories. QY Research via openPR, 28 May 2026. openpr.com/news/4529614/third-party-maintenance-for-data-center-market-report Single source. The existence and direction of the legislation are well established; the 15 to 20 per cent figure rests on one report and should be treated as indicative.
 11. Sector consolidation. Eaton acquisition of Boyd Thermal at 9.5 billion dollars completed March 2026 (company disclosures); Schneider Electric majority acquisition of Motivair in 2024, with the price reported in trade press at approximately 850 million dollars but not confirmed by either party. Compiled in the thermal analysis in this series.



Derived and illustrative content


The service capture map, the model comparison chart and the mix shift chart are my constructions. The first plots judgements about capability gap and defensibility rather than measured values. The second positions the five models on margin and defensibility using industrial service business structures as the reference rather than disclosed data for this sector. The third applies representative industrial margin and capital intensity assumptions to an identical revenue base to isolate the effect of mix. All three illustrate mechanisms and none should be read as measurement.



On the author's perspective


I have run product and service businesses through transitions of this kind, including outcome-based and gain-share models in industrial markets. The observations in section seven about compensation, organisational placement and measurement come from that experience rather than from published research, and should be read as informed judgement rather than established fact.



Series. The AI Power Chain covers six layers of the physical infrastructure behind AI compute. This supplement examines who captures the service revenue those layers generate. A companion piece analyses how vendors across the chain are pricing under scarcity.


The author is an independent advisor working on data centre, power and industrial technology. This analysis is written in a personal capacity and rests entirely on public information. Nothing here is investment advice.


© 2026 Adi Kumar · adikumar.co










In brief


AI infrastructure services now command product-like multiples, a reversal of the historic pattern where hardware captured value and services traded at low margins. The inversion holds in seven positions: hyperscaler capacity brokering, thermal chemistry monitoring, transformer commissioning, DC commissioning, cyber, ISO interconnection navigation, and asset-level covenant advisory.


==============================================================================
# The Wide-Bandgap Stack: Technical Companion
URL: https://adikumar.co/the-wide-bandgap-stack-technical-companion/
Published: 2026-08-01
Summary: How wide-bandgap power devices actually work: SiC crystal growth, MOSFET architectures, GaN HEMT physics, packaging and roadmap. Technical primer.
==============================================================================

The AI Power Chain series · Part 6 of 15








Glossary of terms used



GaN
Gallium Nitride. Wide-bandgap semiconductor material used in RF and power applications.
MV
Medium Voltage. Typically 1 kV to 35 kV. The voltage class connecting data centre power distribution to the utility grid.
SiC
Silicon Carbide. Wide-bandgap semiconductor material used in high-voltage power electronics.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.









ADI KUMAR · POWER & DIGITAL INFRASTRUCTUREAUGUST 2026 · V1 · ~40 MIN READ

The AI Power Chain · Part II-A · Technical Companion
How wide-bandgap semiconductors actually work



The Wide-Bandgap Stack: Technical Companion


Where the market essay named the players, this piece explains the physics. What "wide bandgap" actually means, why SiC and GaN outperform silicon at the same voltage, how a boule of vapour-grown crystal becomes a switch that runs at 800V and 200kHz, and where the technology goes next. Cross-sections, process flows, and roadmaps for readers who want to understand the machine underneath the market.



By Adi Kumar
Independent analysis
Views are my own



How to read this piece. It is a technical companion to The Wide-Bandgap Stack (Part II of this series). The market piece named the vendors, sized the layers and mapped the M&A. This one explains the underlying physics and manufacturing, why SiC and GaN devices switch the way they do, and what the roadmap items on every analyst's list actually change. No prior device-physics background is assumed, but the reader is expected to know what a MOSFET does at a systems level.

1The physics · Why bandgap matters



What "wide bandgap" actually means


Every property that makes SiC and GaN interesting at rack scale flows from one number: the bandgap. Silicon's is 1.12 electronvolts. Silicon carbide's is 3.26 eV. Gallium nitride's is 3.39 eV. Those numbers sit inside almost every device advantage discussed in the market essay. The rest of this piece works out why.


Bandgap is the energy an electron in the valence band must gain to be promoted into the conduction band. Below that threshold, electrons stay put and the material insulates. Above it, they conduct. In a power switch, the point is not conductivity but blockingWhen a MOSFET is off, the drain-to-source region must sustain the full bus voltage without letting current through. The parameter that governs this is the critical electric field, the field strength at which the material breaks down (avalanches). Critical field scales roughly as the square of the bandgap.


Silicon's critical field is about 0.3 MV/cm. SiC's is around 3 MV/cm, ten times higher. GaN's is comparable to SiC. What that means in practice: to block a given voltage, a SiC device needs roughly a tenth the drift-region thickness of a silicon device, doped roughly ten times more heavily. Thinner drift region plus higher doping equals dramatically lower on-resistance.



Bandgap and critical field: the three materials that matter
Bandgap energy (eV) and critical electric field (MV/cm) for Si, SiC (4H polytype) and GaN

Silicon: 1.12 eV, ~0.3 MV/cm. 4H-SiC: 3.26 eV, ~3 MV/cm. GaN: 3.39 eV, ~3.3 MV/cm. Standard semiconductor references (Sze, Baliga).




The Baliga figure of merit


The relationship between material properties and switch performance is captured in a single expression, published by B. Jayant Baliga in 1989: BFOM = ε·μ·Ec3Permittivity times mobility times the cube of critical field. Higher BFOM means lower on-resistance per unit die area at a given breakdown voltage. Because critical field enters at the third power, a 10x advantage in Ec translates to a theoretical 1,000x reduction in the RDS(on) × Area product.


Silicon's BFOM is 1 by definition. 4H-SiC's is roughly 500. GaN's is around 900. Those numbers are the ceiling. Real devices leave most of it on the table because of contact resistance, channel resistance and packaging parasitics, but the numbers explain why a 1200V SiC MOSFET can be an order of magnitude smaller than a 1200V silicon superjunction MOSFET at the same on-resistance.



PropertySilicon4H-SiCGaNWhy it matters

Bandgap Eg (eV)1.123.263.39Higher Eg allows higher operating temperature and lower intrinsic leakage
Critical field Ec (MV/cm)0.3~3.0~3.3Sets the maximum voltage per unit thickness of drift region
Electron mobility (cm²/V·s)1,350950~2,000 (2DEG)Lower resistance per unit area for the same doping
Thermal conductivity (W/m·K)150370-490130 (bulk), higher on SiCAbility to move heat out of the die
Saturated drift velocity (cm/s)1×1072×1072.5×107How fast carriers respond, ceiling on switching speed
BFOM (Si = 1)1~500~900Theoretical Ron·A improvement at given breakdown voltage





Six-axis material comparison
Normalised to the best-in-class material on each axis. Larger area = better overall.

Values normalised: Eg, Ec, drift velocity and thermal conductivity divided by the highest value across the three. Mobility uses lateral 2DEG for GaN. BFOM is log-scaled then normalised to fit. The point of the radar is to show that no single material dominates every axis: GaN wins on mobility and BFOM, SiC on thermal conductivity, silicon on maturity (not shown).

What "wide bandgap" is not. A wider bandgap alone does not make a better device. Wide-bandgap materials tend to have lower bulk electron mobility than silicon (SiC channels move electrons more slowly than silicon channels). The improvement comes from being able to build the same voltage rating with a thinner, more heavily doped drift region, which more than compensates for the mobility penalty. GaN's edge is a different mechanism entirely (a 2D electron gas at a heterojunction interface, covered later).
2Crystal growth





In brief


SiC and GaN device physics, foundry economics and packaging materials shape the pace of wide-bandgap adoption. SiC MOSFETs handle 1200V+ switching with 30-50% lower conduction loss than silicon IGBTs at the same voltage class. GaN HEMTs address the higher-frequency point-of-load stage. Foundry capacity, wafer yield and packaging thermal design set commercial availability.








Growing SiC: sublimation, not melting


Silicon devices start with a boule of monocrystalline silicon grown by the Czochralski process. A silicon melt sits in a crucible at 1414°C. A seed crystal is dipped in, pulled slowly upward while rotating, and grows into a long single crystal. Modern silicon boules are two metres long, 300mm in diameter, and grow in a few days. The physics is straightforward: silicon melts, you cool it around a seed, it crystallises.


Silicon carbide does not melt at any pressure below roughly 100 kbar. It sublimes (transitions directly from solid to vapour) at around 2,700°C. That fact governs the economics of the SiC industry. You cannot pull a SiC boule from a melt because there is no melt to pull from. SiC boules are grown by physical vapour transport (PVT), a technique closer to depositing frost than to casting metal.




Czochralski (silicon) vs Physical Vapour Transport (SiC)
Simplified cross-sections of the two crystal-growth chambers





Czochralski, silicon



Argon atmosphere1 atm
Seed crystal, pulled ↑
Growing boulesolidifying
Silicon melt1,414°C
Quartz crucible
Graphite heater / susceptor



Growth rate: mm/hour. Yield: 2m boule in a few days. Diameter: 300mm routine.





Physical vapour transport, SiC



Graphite chamber wall
Seed crystal2,100°C · cold side
Growing boulere-condensing SiC
↑ SiC vapour transport ↑
SiC powder charge2,300°C · hot side
Inductive heater · Ar atmosphere



Growth rate: ~mm/hour, but at 100-500x lower rate per unit than Si. Yield: 25-35mm of usable crystal per week. Diameter: 150mm mainstream, 200mm ramping, 300mm demonstrated.




The PVT chamber holds a temperature gradient of ~200°C between the powder source (hot) and the seed (cold). SiC vapour species (Si, Si₂C, SiC₂) diffuse across the gradient and re-condense on the seed crystal.




The consequences of this thermodynamic asymmetry are dramatic. A silicon boule grows a hundred times faster than a SiC boule. Silicon substrates are essentially a commodity input to fabs. SiC substrates were, until very recently, 40-50%+ of the cost of a finished device. The market essay traced how Chinese capacity has collapsed that share; the physics explains why it was so high in the first place.




Defects: the yield-critical statistic


PVT crystals grow with defects. The three that matter for power devices are micropipes (screw dislocations that leave a hollow core running through the crystal, up to a few microns wide), basal plane dislocations (line defects that lie in the basal plane of the hexagonal SiC lattice) and threading dislocationsAny of these that intersect the active area of a device kill it under bipolar stress.


Micropipe density has fallen roughly a thousand-fold in twenty years: from 100+/cm² in the late 1990s to well under 1/cm² today on 150mm production wafers. Basal plane dislocation density has fallen from tens of thousands per cm² to a few hundred, still material for yield but no longer catastrophic. Threading dislocations remain in the ten-thousands per cm² and are the current frontier.




SiC defect density: two decades of engineering
Micropipe density in commercial SiC substrates, defects per cm² (log scale)

Trajectory compiled from industry compilations and Cree/Wolfspeed public disclosures. Rates today under 0.1/cm² routine for premium 150mm substrates; 200mm still catching up.


3Wafering



Cutting diamond-hard material without wasting it


Once a boule is grown, you have to slice it into wafers. SiC's hardness (Mohs 9.5, second only to diamond) makes conventional slicing difficult and wasteful. Two techniques matter.


Diamond wire sawing uses a diamond-impregnated wire, run back and forth in a bath of coolant, to abrade its way through the boule. Kerf loss (the material converted to slurry rather than wafer) runs 200-300µm per cut. On a typical 500µm-thick wafer, that means nearly half the boule ends up as waste. The saw itself is slow because SiC resists it: a 150mm wafer takes hours to slice.


Laser cold-split, commercialised by Siltectra and acquired by Infineon, takes a different approach. A focused laser beam is fired into the boule at a specific depth, creating a plane of micro-damage without cutting through. Thermal stress then separates the wafer along that plane. Kerf loss falls from hundreds of microns to essentially zero. Effective wafer count from a given boule rises 20-30%. Wolfspeed's demonstrated 300mm boule and the industry's move to 200mm both mean cold-split economics scale up rather than down, and this is the deflationary vector most analysts underweight.





Step 1
Boule preparationGround to cylinder, oriented to crystallographic axis, seeded plane identified
Diamond wheel grind, X-ray orientation

↓

Step 2a · Wire saw
Diamond wire slicingDiamond-impregnated wire abrades through boule under coolant
~200-300µm kerf loss
Hours per wafer


Step 2b · Cold split
Laser cold-splitFocused laser creates sub-surface damage plane; thermal stress cleaves the wafer
~0µm effective kerf
Minutes per wafer

↓

Step 3
Lapping and polishingGrind to target thickness, chemical-mechanical polish to sub-nm Ra
CMP with silica slurry
Surface roughness < 0.5nm

↓

Step 4
Substrate inspectionDefect mapping, resistivity, surface metrology
KLA, Onto or Bruker tools



The 200mm yield problem. Moving from 150mm to 200mm doubles die per wafer and cuts cost per die by roughly 40% at mature yield. The catch: 200mm SiC yields lagged 150mm for most of 2024-25, with industry commentary putting average yields below 50% while the industry was still scaling defect control. Every fab that transitioned took the depreciation of a more expensive line while shipping product priced on 150mm economics. That is why 200mm is a real event even in a glut, and why Wolfspeed's 300mm boule demonstration in January 2026 mattered more for standards than for immediate volume.

4Epitaxy



The layer where the device actually lives


A polished substrate is not yet a device. The active silicon carbide (the drift region that blocks voltage, the p-body region that forms the channel, the n+ source region that supplies electrons) is grown as an epitaxial layer on top of the substrate. Chemical vapour deposition, typically silane-based, at temperatures around 1,500°C. Growth rates are 5-30 µm per hour, orders of magnitude slower than the reactions used in silicon epi.


The thickness of the epi layer sets the breakdown voltage. A rough rule for 4H-SiC: drift region thickness (µm) ≈ 0.01 × breakdown voltage (V)A 1200V device needs about 12µm of drift, a 3300V device 33µm. Doping density scales inversely: heavier doping for lower voltage devices, lighter doping for higher voltage. This is a physics constraint, not a design choice.




Drift region thickness scales with breakdown voltage
Required epitaxial drift thickness for 4H-SiC, µm vs breakdown voltage V

Approximate: t_drift ≈ 2·V_br / E_c, with E_c ≈ 2.5 MV/cm for design margin. Real devices add margin for edge termination and manufacturing tolerance.




Epi is the layer where the substrate's defects either propagate into the device or get filtered. A basal plane dislocation in the substrate has some probability of continuing into the epi, where it becomes a killer defect for any MOSFET built above it. Good epi growers can convert BPDs to threading edge dislocations at the substrate/epi interface, essentially deflecting them into a plane where they do less harm. Epi is where the substrate glut gets absorbed or wasted for exactly this reason: a cheap substrate with high BPD density still produces expensive finished devices, because the yield loss shows up downstream. It is also why the merchant epi market (Resonac leading, Wolfspeed offering 200mm commercially) matters strategically despite being small in revenue terms.




Throughput as the hidden cost driver


SiC epi grows at a few microns per minute. A 1200V device needs 10-12µm of drift. Reactor time per wafer is therefore measured in tens of minutes. Compare with silicon epi, where deposition steps often run in seconds. Reactor productivity (wafers per hour per tool) is a first-order determinant of fab economics, which is why the tool choice matters as much as the recipe. Aixtron and Italy's LPE dominate SiC epi reactors, with Veeco and Japanese suppliers in adjacent positions. For a plant sized to output tens of thousands of 200mm wafers a month, the epi tools are the throughput ceiling.



5Device architectures



How a SiC MOSFET actually blocks voltage


A power MOSFET is a vertical device: current flows from the source contact on top of the die, down through a body region and a drift region, out to the drain contact on the bottom. When the device is off, the drift region blocks the bus voltage. When it is on, the body region inverts to form a conductive channel, and electrons flow from source to drain.


Three architectures cover most of the SiC market: planar, trench, and trench-superjunctionEach is a different answer to the same trade-off: how do you get the most channel per unit die area, without breaking the gate oxide.




Planar MOSFET


The oldest and simplest architecture. The gate sits on top of the die, over a lateral channel formed between the source and the drift region. It works reliably, but the channel takes up lateral space that limits how much current a die area can pass. Wolfspeed's early SiC MOSFETs were planar; ROHM and STMicro shipped planar for years before moving to trench.




Planar SiC MOSFET · simplified cross-section
Vertical current flow: source (top) → channel → drift → drain (bottom)

Source metal (Al)
PassivationSiO₂ / Si₃N₄
Poly-Si gateon SiO₂ gate oxide
p-bodychannel inverts here when V_GS > V_th
n⁻ drift region~12µm for 1200V · blocks V_DS when off
n⁺ substratelow-resistance current path to drain
Drain metal (Ni/Ag)

The channel is horizontal, at the surface. Cell pitch is limited by lithography and by the need for a JFET region between adjacent p-bodies. Ron·A of a mature planar 1200V SiC MOSFET is around 3-4 mΩ·cm².





Trench MOSFET


The gate is etched vertically into a trench, so the channel runs down the sidewall rather than across the surface. The device becomes narrower per unit current, so cell pitch drops and current density rises. Trench SiC devices are typically 15-30% smaller for the same on-resistance. The catch is that the gate oxide at the trench corner sees a concentrated electric field when the device blocks, and SiC's high critical field means the oxide is the first thing to fail. Two solutions dominate: a double-trench design (Rohm) that adds a shielding p-region beneath the trench, and asymmetric trench geometries (Infineon) that route the field away from the oxide corner.




Trench SiC MOSFET · vertical channel, higher current density
Gate etched into a trench; channel runs vertically along the sidewall

Source metal
p⁺ source contact
Poly-Si gate in trenchwith SiO₂ sidewall + p-shield below
p-bodyvertical channel along trench sidewall
n⁻ drift regionfield concentrates at trench corners → shielded by p-region
n⁺ substrate
Drain metal

Channel is vertical, along the trench sidewall. Cell pitch shrinks by ~30%. Ron·A of a mature trench 1200V SiC MOSFET reaches ~2 mΩ·cm², roughly 30-40% below planar. Gate oxide reliability is the limiting design constraint.





Trench-superjunction


Superjunction is the technique that let silicon MOSFETs break the "silicon limit" Ron·A curve at 400-900V. Alternating vertical p and n pillars in the drift region create lateral charge balance, allowing the n region to be much more heavily doped while still supporting the full blocking voltage. Result: 4-5x lower Ron·A than a conventional drift design at the same breakdown. Silicon superjunction is mature, dominant for 500-650V industrial supplies. Bringing superjunction to SiC has taken years because the p-column doping and lithography are hard at SiC's cell pitches. Infineon's CoolSiC G2 (announced 2024-25) is the highest-volume trench-superjunction product in production, claiming ~40% lower RDS(on)·A than the prior generation.




Ron·A vs breakdown voltage: the three-material picture
Log scale. Lower is better. Trench-SJ SiC pushes into the region that used to be Si-only.

Indicative curves from device physics (Baliga limit for each material) plus published product-level Ron·A for planar SiC, trench SiC and trench-SJ SiC. GaN HEMT curve stops at ~650-900V for lateral parts; vertical GaN pushes it into higher-V territory.



Where the on-resistance actually sits
Waterfall: decomposition of Ron·A for a 1200V trench SiC MOSFET (mΩ·cm²)

Indicative breakdown of resistance contributions along the current path: n⁺ substrate, drift region, JFET/accumulation region, MOS channel, source contact. The drift region is unavoidable at a given voltage rating (physics). The channel is what trench and superjunction attack. Substrate and contact are what packaging and metallurgy attack.


6Gate oxide reliability



The SiC/SiO₂ interface is where the reliability engineering happens


Silicon MOSFET reliability is a solved problem. Fifty years of silicon-dioxide research have made the Si/SiO₂ interface one of the cleanest in solid-state physics. SiC/SiO₂ is a different story. The interface between silicon carbide and thermally grown silicon dioxide contains orders of magnitude more traps (defects that capture and release electrons) than the equivalent silicon interface. This has two consequences for power devices.


First, channel mobility is far below theoretical: 20-50 cm²/V·s in production SiC MOSFETs, versus 400-500 in comparable silicon MOSFETs. Interface traps scatter electrons, slowing them down and raising channel resistance. Second, threshold voltage drifts with gate stress, a phenomenon called NBTI (negative bias temperature instability) when the gate is biased negative and PBTI when it is biased positive. Threshold drift is a lifetime reliability issue: a device that starts life at V_th = 2.5V may drift to 3.5V after ten years of operation, potentially compromising the design margin at start-up.


The mitigation is nitridation: annealing the oxide in nitric oxide (NO) or nitrous oxide (N₂O) during formation, which passivates dangling bonds at the SiC/SiO₂ interface. Nitridation dropped interface trap density by roughly an order of magnitude in the 2010s. All commercial SiC MOSFETs today use some variant. The frontier is around alternative gate dielectrics (aluminum oxide, high-k stacks) and around plasma-nitridation processes that target the specific trap species most responsible for V_th drift.



Why this matters for qualification. Automotive-grade SiC MOSFETs (AEC-Q101) must show <100 mV V_th drift after 1,000 hours of high-temperature gate bias stress at 175°C junction. Data-center-grade parts face similar or tighter specs. Meeting these targets requires nitridation processes that are proprietary to each vendor and take years to qualify. This is the moat that makes "just build a SiC MOSFET" so much harder than the wafer economics suggest.

7GaN devices



GaN: how a heterojunction replaces a doped channel


SiC MOSFETs are the same architecture as silicon MOSFETs, made from a better material. GaN power devices are structurally different. They are lateral, not vertical (except for vertical GaN, still emerging). They rely on a two-dimensional electron gas, not a doped inversion channel. And most native GaN transistors are normally-on, not normally-off, which is a design headache for power systems that assume default-off behaviour.


The core device is the HEMT, or high electron mobility transistor. A layer of AlGaN (aluminum gallium nitride) is grown on top of a layer of GaN. The lattice mismatch between the two creates a strong piezoelectric field at the interface. That field pulls electrons out of the AlGaN and traps them in a thin sheet at the top of the GaN layer: the 2DEG, or two-dimensional electron gas. The 2DEG has electron densities comparable to a heavily doped bulk semiconductor, but the electrons are moving in a nearly defect-free crystal with mobility around 1,500-2,000 cm²/V·s, much higher than any doped bulk region.




Lateral GaN HEMT · the 2DEG channel
Current flows horizontally from source to drain through the 2DEG

Source · Gate · Drain (lateral contacts)
Passivation (SiN)
AlGaN barrier~20nm · induces 2DEG below via piezo effect
2DEG channel1-2 nm thick sheet of electrons · μ ~1,500-2,000 cm²/V·s
GaN channel layer~1µm undoped
GaN buffer / transition layersmanage lattice mismatch to substrate
Silicon substrate (or SiC)cheap wafer, thermal + mechanical support

Native HEMT is depletion-mode: the 2DEG exists at zero gate bias, so the device is normally-on. Practical power devices add an enhancement-mode conversion (p-GaN gate or Si-cascode).





Why lateral and why on silicon


GaN power devices are lateral because good bulk GaN substrates are still rare and expensive. Most commercial GaN power devices are GaN-on-Si: grown on cheap silicon wafers, using buffer layers to manage the lattice mismatch (17%) and thermal expansion mismatch. This lets fabs use depreciated 200mm silicon lines rather than the specialised 150mm compound-semiconductor tools SiC requires. TSMC ran GaN-on-Si production until announcing exit; GlobalFoundries, PSMC, Vanguard, X-FAB and Polar are the merchant foundry base.


The downside of lateral architecture is voltage. Lateral devices break down along their surface, and the breakdown voltage scales roughly linearly with gate-drain distance. A 650V GaN HEMT is typically a few microns of gate-drain spacing. Push voltage higher and the device gets wider and slower. 900V lateral GaN parts exist. 1200V devices are on the boundary. Above 1200V, either you need a much larger die (expensive) or you need to go vertical, meaning current channels are built into the die from the substrate side so that current flows vertically like a SiC MOSFET.




The normally-on problem, and three ways around it


The 2DEG exists at zero gate bias, so a native GaN HEMT conducts when the gate is at ground. In a power system, that is dangerous: if the gate driver fails at start-up, the high-side switch is on, and the bus shorts to ground. Power designers assume normally-off (enhancement-mode) behaviour. Three techniques convert a native HEMT to enhancement mode.


p-GaN gate. A layer of p-doped GaN is grown between the gate metal and the AlGaN barrier. The p-GaN depletes the 2DEG directly beneath it, so the channel is broken at zero bias. Positive gate voltage restores the channel. This is the technique used by Infineon (GaN Systems), EPC and Innoscience. Gate voltage swing is narrow (typically 5V max), making driver design careful.


Recessed gate. The AlGaN barrier is etched away under the gate, breaking the 2DEG mechanically. Rare in commercial production because the etch process is hard to control precisely enough for consistent threshold voltage.


Cascode. A native depletion-mode GaN HEMT is stacked in series with a small silicon low-voltage MOSFET. The silicon MOSFET provides normally-off behaviour and gate drive familiarity; the GaN HEMT provides the high-voltage blocking. Transphorm (now Renesas) was the leading cascode vendor. This is a compromise architecture, since the silicon FET adds resistance and switching parasitic and defeats some of GaN's speed advantage.





p-GaN gate (enhancement-mode)



Gate metal
p-GaN layer
AlGaN
GaN (2DEG broken under gate)



Vendors: Infineon (GaN Systems), EPC, Innoscience, Power Integrations. Narrow gate-voltage window.




Cascode (depletion GaN + Si)



Package boundary
Depletion-mode GaN HEMT
Series Si low-V MOSFET
→ appears normally-off externally



Vendors: Renesas (Transphorm). Familiar gate drive but the Si adds R and loss.






Why vertical GaN matters


The path to break through the 650-900V ceiling of lateral GaN is vertical GaN: same current-flow direction as a SiC or Si MOSFET, but with the higher electron mobility of a GaN 2DEG. The manufacturing challenge is producing GaN substrates of adequate quality and diameter. Power Integrations, through its Odyssey acquisition, and Navitas are the visible pursuers. If vertical GaN reaches production at 1200V+ with cost parity to SiC trench, it erodes the clean SiC/GaN voltage boundary that the market essay treated as a stable competitive assumption.



8Switching physics



Why WBG switches faster, and what that unlocks


A power switch's efficiency in a converter is set by two loss mechanisms: conduction loss (I²R while on) and switching loss (energy dissipated during the on-off and off-on transitions). Higher switching frequency reduces the size of magnetics (inductors, transformers) and capacitors, which is where most of a converter's volume and weight sits. But higher frequency also multiplies switching loss, so there is a natural trade-off.


The wide-bandgap advantage in switching is not primarily about speed. Silicon devices can switch quickly too. The advantage is about the energy lost per switching event, driven by two device properties: gate charge (Q_g) and reverse recovery charge (Q_rr).




Reverse recovery: the silicon body diode problem


A silicon MOSFET has an intrinsic body diode from source to drain. In a half-bridge converter, that diode conducts current during the dead-time between the top and bottom switch turning on. When the diode is forced back into the off state, minority carriers stored in its drift region must be swept out. This is reverse recovery, during which the device conducts a large reverse current spike before finally blocking. The spike dissipates energy, adds electromagnetic interference and can damage the opposing switch.


Silicon body diode recovery times are tens to hundreds of nanoseconds. SiC Schottky barrier diodes have essentially no reverse recovery (Schottky diodes are majority-carrier devices, no minority carriers to store). SiC MOSFETs still have a body diode, but its recovery charge is roughly a fifth of a comparable silicon MOSFET's. GaN HEMTs have no body diode at all: reverse conduction happens through the 2DEG, and there is nothing to recover.




Gate charge: the driver energy


Every switching event charges and discharges the gate capacitance. The energy is dissipated in the gate driver. A silicon superjunction 650V MOSFET has a total gate charge of ~50-100 nC. A comparable SiC MOSFET has ~50 nC. A GaN HEMT has ~5-10 nC. GaN's ten-fold advantage in gate charge is what enables it to switch at MHz frequencies without frying the driver.




Switching loss vs frequency: why WBG unlocks higher f
Total loss (conduction + switching) for a 650V, 30A device across switching frequency, indicative

Illustrative from device datasheets. Silicon crosses over into diminishing returns around 100kHz; SiC extends useful frequency to ~500kHz; GaN extends it to several MHz. Exact crossover points depend on load, driver design and topology.





What that unlocks at the system level


Higher switching frequency shrinks passives. The inductor in a buck converter scales roughly inversely with frequency; a converter running at 500kHz needs about a fifth the inductance of one running at 100kHz. The capacitor bank shrinks similarly. So does the transformer in an isolated topology. A 3.3kW LLC resonant converter built with silicon MOSFETs might use magnetics filling a shoebox; the same converter built with GaN can fit in a hardcover book. This is the mechanism by which GaN eliminates the 48V intermediate bus stage in Navitas' 800V-to-6V direct-conversion board discussed in the market essay: the higher operating frequency lets a single stage span a voltage ratio that used to require two.



9Package & module



Package matters more than most models capture


A power semiconductor die is worthless until it is packaged. Package matters for four reasons: it provides the electrical interconnect from die to the outside world, it moves heat from the die into a heatsink, it supports the die mechanically through thermal cycling, and it insulates high-voltage from low-voltage terminals. In wide-bandgap devices, the package limits the die more often than the die limits the package.




Power module cross-section · SiC MOSFET on AMB substrate
Path from die to baseplate to heatsink

Molded resin encapsulation
Al bond wire or Cu clip (top contact)
SiC MOSFET die200-500µm · junction temperature up to 200°C+
Sintered silver die attach200+ W/m·K · sinters at 250°C, operates to 400°C
Top copper layerpatterned circuit routing
Silicon nitride (Si₃N₄) ceramicAMB substrate · 60-90 W/m·K · fracture-tough
Bottom copper layersolders to baseplate
Solder or sinter layer
Copper baseplateor AlSiC for CTE match
Thermal interface material (grease / PCM / sinter)
Heatsink or cold plate

The critical layer is the ceramic substrate. DBC (direct bonded copper on Al₂O₃ or AlN) is the incumbent; AMB (active-metal-brazed copper on Si₃N₄) is the growth layer for SiC modules because Si₃N₄ survives the thermal cycling that cracks alumina.





DBC vs AMB substrates


The ceramic substrate does three jobs simultaneously: it insulates high-voltage traces from the heatsink (baseplate), it conducts heat away from the die, and it survives repeated thermal cycling as the module heats and cools. Two chemistries compete.


DBC (direct bonded copper) starts with an alumina (Al₂O₃) or aluminum nitride (AlN) ceramic. Copper foil is bonded to both sides by heating in an oxygen-controlled atmosphere at ~1,065°C. DBC is mature, cheap and works fine for silicon IGBTs. It struggles in high-power SiC modules because alumina's fracture toughness is limited and repeated thermal cycling opens cracks. AlN improves thermal conductivity but the mechanical issue persists.


AMB (active-metal-brazed) uses silicon nitride (Si₃N₄), a much tougher ceramic, joined to copper via an active braze alloy (typically Ti-containing). Si₃N₄ has ~4x the fracture toughness of alumina and comparable thermal conductivity to AlN. AMB substrates cost more but survive the thermal cycles that kill DBC in SiC modules. Si₃N₄ now represents over 95% of the AMB market. The suppliers are heavily Japanese: Kyocera, Denka, Mitsubishi Materials, NGK/Niterra, Toshiba Materials.




Sintered silver die attach


Silver sintering has replaced solder as the die attach for high-power SiC modules. A paste containing silver nanoparticles is applied under the die and pressed at 200-300°C in air. The nanoparticles fuse into a dense polycrystalline silver layer that is essentially bulk silver: thermal conductivity of 200+ W/m·K, and mechanically stable up to silver's melting point of 962°C, versus a lead-free solder's ~217°C. SiC modules can therefore run to 200°C junction temperature without softening the die attach, which is central to using SiC's temperature capability. The cost is silver, and silver is expensive and volatile, which is why silver-free AMB substrates (Heraeus Condura.ultra) and low-silver sinter formulations are an active development target.




Top-side contact: wire bond to copper clip


Traditional power modules connect the top of the die to the top copper layer with aluminum wire bonds: thick ultrasonic bonds, typically 300-500µm diameter. Wire bonds work but have three limitations. Their electrical resistance adds to on-state loss, their inductance limits switching speed, and their bond footprints wear out mechanically under thermal cycling. Modern high-power SiC modules increasingly use copper clips or ribbon bonds instead: solid copper structures brazed or soldered from the die top to the top-copper. Lower resistance, lower inductance, better cycling. Copper-clip assembly requires precision that wire bonders do not, and it is one of the underweighted drivers of the packaging-chokepoint thesis in the market essay.



10Topologies



What WBG unlocks at the converter level


Devices matter because they enable circuits. Two topologies have driven most of the incremental WBG design activity: the LLC resonant converter for isolated DC-DC in server power supplies, and the solid-state transformer for medium-voltage grid interface at data-centre scale.




The LLC resonant converter


LLC (named for its inductor-inductor-capacitor resonant tank) is the dominant topology for high-efficiency isolated DC-DC conversion. The trick is zero-voltage switching (ZVS): the resonant tank shapes the current waveform so that switches turn on when the voltage across them is zero, eliminating switching loss almost entirely. Silicon MOSFETs can do LLC, but they are limited to ~200-500kHz because their gate charge and reverse-recovery losses dominate above that. SiC and GaN LLC converters routinely operate at 500kHz. 1MHz. The transformer and resonant inductor shrink correspondingly.


In a 3kW server power supply, an LLC stage using SiC or GaN switches typically achieves 98%+ efficiency across a wide load range. Silicon-based designs top out around 96-97%. Two efficiency points sounds small; on a hyperscale campus drawing 100MW, it is 2-3MW that stops becoming heat and starts becoming compute.




The solid-state transformer


A conventional data-centre power path takes medium-voltage utility power (typically 13.8kV or 33kV AC), steps it down through a line-frequency transformer to 480V or 400V AC, rectifies it to a bus, and then converts to whatever server voltages are needed. Line-frequency transformers are heavy (kilograms per kVA), lossy (~2% each), and inflexible. A solid-state transformer (SST) does the same job electronically: a medium-voltage AC-DC rectifier feeds a high-frequency isolation stage (essentially an LLC or dual active bridge running at tens of kHz), which feeds a low-voltage DC bus. The result is smaller (an order of magnitude reduction in mass), more efficient, and bi-directional. The bi-directional property is important for grid services and integrating on-site solar or batteries.


SSTs need switches that can hold off the medium-voltage bus (1700-3300V class) and switch fast enough to make the high-frequency stage economically small. That is precisely SiC's sweet spot. Multiple pilots are in flight, with data-centre operators and utilities exploring 6.6kV and 13.8kV SSTs, but production-scale deployment is still limited. If SSTs scale, they pull 1700V and 3300V SiC device demand forward by years and reshape the mechanical footprint of every hyperscale substation.



11Roadmap



Where the technology is going


The market essay treated ten technology vectors as a "disruption watch list". This piece walks through what each one actually changes.






2024-26 · shipping
Trench SiC MOSFETs at scale
Infineon CoolSiC G2 (with superjunction), ROHM 4th-gen trench, onsemi EliteSiC trench. Cell pitch shrinks ~30% vs planar, Ron·A drops accordingly. Reliability engineering around gate oxide corner field is the differentiator between vendors.



2025-27 · ramping
200mm SiC production
Wolfspeed Mohawk Valley, Infineon Kulim Module 3, ST Catania, onsemi Czech. 2.2x more die per wafer, ~40% cost reduction per die at mature yield. Yield still lags 150mm; margin compression is transitional, not permanent.



Jan 2026 · demonstrated
300mm SiC boule
Wolfspeed demonstrated a 300mm SiC boule, reportedly the first. Standards implication more than immediate volume: whoever defines the format that customers validate gets to shape a decade of tool roadmaps. Qualified 300mm production is many years away.



2026-28 · scaling
Laser cold-split slicing
Infineon's Siltectra process eliminates most kerf loss. Effective wafer count from a given boule rises 20-30%. Deflationary across the whole substrate market, and its effect compounds with 200mm/300mm scaling.



2026-28 · early product
Engineered substrates (SmartSiC)
Soitec's SmartSiC bonds a thin layer of high-quality monocrystalline SiC onto a poly-SiC handle wafer. Uses far less prime crystal, breaks the boule-growth cost ceiling. Threatens pure-substrate players; a natural fit for high-volume automotive SiC.



2026-29 · scaling
Vertical GaN and GaN-on-SiC
Pushes GaN into 1200V+ territory currently owned by SiC. Power Integrations (Odyssey acquisition), Navitas, and research groups. If vertical GaN reaches production at cost parity with SiC trench, the clean SiC/GaN voltage boundary erodes and the competitive map from Part II needs updating.



2025+ · commercial
Bidirectional switches (BDS)
Infineon CoolGaN BDS, Nexperia and research variants. A single device blocking current in both directions, replacing the two back-to-back MOSFET structure conventionally used for AC switches. Simplifies rectification stages and changes converter topology BOM.



2027+ · pilot
Silver-free AMB and low-silver sinter
Heraeus Condura.ultra, Indium Corporation and MacDermid Alpha low-silver pastes. Removes precious-metal exposure from the module cost base and cuts supply-chain risk. Structural cost-down at the packaging layer.



2027+ · flagship platforms
Double-sided cooling and embedded die
Dies embedded in PCB or substrate, cooled on both faces. Doubles power density, shifts more value into the packaging layer. Automotive traction is the driver, data-centre inverters follow.



2030s · research
Ultra-wide bandgap: Ga₂O₃, diamond, AlN
Ga₂O₃ has Eg of 4.9 eV (vs SiC 3.26) and a theoretical BFOM ~4x SiC. The unsolved problem is thermal conductivity: Ga₂O₃ is roughly 10x worse than SiC, so devices overheat unless heat is extracted through the top. Diamond (5.5 eV, best thermal conductivity of any material) is closer to a curiosity than a product line. AlN (6.2 eV) is lab-scale. Watch, don't underwrite.




Every deflationary vector in this roadmap operates on the same variable: Ron·A. Trench and superjunction shrink it via better device design. 200mm, 300mm and cold split shrink cost per die area. Engineered substrates shrink cost per unit of prime crystal. Vertical GaN attacks the voltage boundary. The materials industry gets cheaper per switched watt across the board, on all fronts, at the same time.The technology thesis in one line

12Where this lands



Physics to system, in one arc


Every advantage a wide-bandgap power switch has over silicon can be traced back to a single number. Higher bandgap → higher critical field → thinner drift region for the same voltage → lower Ron·A → smaller die at the same current → less gate charge → higher switching frequency without overwhelming loss → smaller magnetics and capacitors → smaller, cheaper, more efficient converter. The chain runs top to bottom without a break.


What matters commercially is that every advance on the roadmap tightens one link in the chain. Trench devices shrink Ron·A at fixed material properties. Engineered substrates shrink cost per unit of drift-region area. Vertical GaN attacks the voltage boundary that separates GaN from SiC. Improved packaging (AMB Si₃N₄, sintered silver, copper clips) removes the thermal and parasitic bottlenecks that used to limit die performance. Each is a small step. Together, they are the deflationary vector that keeps the WBG story going even after the substrate glut priced the material at commodity levels.


For a reader coming from the market essay, three concrete implications. First, the qualification moat is real: the gate-oxide reliability engineering that separates a working SiC MOSFET from a scrap die takes years to build and cannot be bought. Second, the roadmap is genuinely additive: unlike some semiconductor cycles where you pick winners between competing architectures, most of these advances stack. Third, vertical GaN is the swing factor: if it clears the manufacturing thresholds discussed here, the market essay's assumption of a stable SiC/GaN voltage boundary needs revisiting.



SERIESNext in the AI Power Chain



What connects to what


This is a technical companion to The Wide-Bandgap Stack (Part II of the AI Power Chain), which covers the market structure, players, TAM and M&A. Part III (The Thermal Stack) covers the cooling infrastructure that these more efficient but more concentrated power devices demand; its technical companion (The Thermal Stack: Technical Companion) walks through the thermal physics in the same way. Part IV (The Interconnect Stack) covers the grid equipment and copper that gates the buildout.




Series. The AI Power Chain (six parts). Part I: The Capacitor StackPart I-A: Technical CompanionPart II: The Wide-Bandgap StackPart II-A: Technical CompanionPart III: The Thermal StackPart III-A: Technical CompanionPart IV: The Interconnect StackPart IV-A: Technical CompanionPart V: The On-Package Delivery StackPart V-A: Technical CompanionPart VI: The Modular Datacenter StackPart VI-A: Technical Companion.


Methodology. This is a technical primer, not investment advice or engineering guidance. Device parameters, process descriptions and roadmap items are compiled from standard semiconductor references (Baliga, Sze, Millman-Halkias), vendor technical literature (Infineon, ROHM, onsemi, Wolfspeed, Innoscience, Navitas), IEEE publications, and trade press. Any specific device performance number should be verified against the vendor's current datasheet before use in design work.


© 2026 Adi Kumar · Power & Digital Infrastructure · Corrections welcome.












Frequently asked


What is the difference between SiC and GaN power semiconductors?

Silicon carbide (SiC) MOSFETs handle 1,200V+ switching at moderate frequency (up to 100kHz). They dominate the bus-level power conversion stage. Gallium nitride (GaN) HEMTs operate at higher frequency (300kHz-3MHz) at lower voltage (typically 100-650V). They dominate the point-of-load stage. The two are complementary rather than competitive.


Which SiC vendors dominate the merchant market?

STMicroelectronics, onsemi, Infineon, Wolfspeed and ROHM together hold over 90% of SiC merchant device revenue per TrendForce 2024 data. STMicro leads on automotive-derived MOSFETs (Tesla relationship). Wolfspeed is the largest US pure-play, restructured through Chapter 11 in 2025. Chinese vendors (SICC, TanKeBlue) grew substrate share meaningfully post-2023.


==============================================================================
# The Thermal Stack: Technical Companion
URL: https://adikumar.co/the-thermal-stack-technical-companion/
Published: 2026-08-01
Summary: How liquid cooling works: heat transfer physics, TIMs, cold plate mechanics, quick disconnects, immersion, and the migration into the package.
==============================================================================

The AI Power Chain series · Part 4 of 15








Glossary of terms used



CDU
Coolant Distribution Unit. Thermal system component distributing chilled coolant to racks or direct-to-chip cold plates.
GPU
Graphics Processing Unit. The compute silicon at the centre of AI workloads.
SiC
Silicon Carbide. Wide-bandgap semiconductor material used in high-voltage power electronics.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.









ADI KUMAR · POWER & DIGITAL INFRASTRUCTUREAUGUST 2026 · V1 · ~40 MIN READ

The AI Power Chain · Part III-A · Technical Companion
How data center thermal management actually works



The Thermal Stack: Technical Companion


Where the market essay named the vendors, this piece explains the physics. What 1,000 W/cm² means at the die, how a coupling that costs a hundred dollars either saves a rack or destroys it, why the industry is moving cooling from the room into the package, and where the technology goes next. Cross-sections, flow diagrams, and roadmaps for readers who want to understand the plumbing under the market.



By Adi Kumar
Independent analysis
Views are my own



How to read this piece. It is a technical companion to The Thermal Stack (Part III of the AI Power Chain). The market piece named the vendors, sized the layers and mapped the M&A. This one explains the underlying heat transfer, the mechanics of the interface materials and couplings, the fluid chemistry, and the roadmap items every operator now has on their qualification list. No prior thermal engineering background assumed, but the reader should know what a heatsink does at a systems level.

1The physics · What "cool" actually means



Heat is not information


Every watt of electrical power a GPU consumes leaves the die as heat. There is no version of a switching transistor in which that stops being true, and no amount of software cleverness changes it. What determines whether the die throttles is not how much heat it makes, but how fast that heat gets to a fluid cool enough to accept it. That is a fluid mechanics and materials-science problem, and every layer in the thermal stack is a specific answer to it.


Three mechanisms move heat: conduction (through solids and stagnant fluids, described by Fourier's law q = -k·∇T), convection (through moving fluids, described by Newton's law of cooling q = h·A·ΔT), and radiation (from any surface above absolute zero, but negligible below 150°C). Data centre cooling is a game played almost entirely in the first two. Conduction dominates from the die to the coolant. Convection dominates from the coolant to the atmosphere.



Heat flux, in context
W/cm² at the die vs everyday reference points, log scale

A 100W incandescent bulb radiates roughly 6 W/cm² off its filament. A domestic cooktop element runs at ~15 W/cm². The sun's surface, integrated, is ~6,300 W/cm². An H100 GPU die at 700W over ~0.8cm² is ~875 W/cm² averaged, with local hot spots higher. Next-generation accelerators approach 1,000 W/cm² sustained.




Junction temperature is the master variable


Silicon devices are specified to a maximum junction temperature (T_j_max), typically 105°C for hyperscale AI accelerators, 125°C for automotive silicon, 200°C+ for SiCAbove T_j_max, one of three things happens: the device throttles (reduces clock frequency or blocks current), it drifts electrically (threshold voltage shifts, leakage rises), or it fails outright. The whole thermal design job is to keep T_j below T_j_max while the die dissipates its rated power.


The relationship is a simple series-resistance model. Heat flows from junction through a stack of thermal resistances (die, interface material, lid, second interface, cold plate, coolant film) into the coolant. Each layer adds a temperature drop equal to the heat flow times its thermal resistance. Add them up and you get the required coolant temperature to keep the junction cool enough.



The thermal stack, layer by layer
Heat flow from die to coolant, and where each temperature drop happens

Silicon die · Tj ≤ 105°C
TIM 1 (die → lid)grease · PCM · sinter · liquid metal
Integrated heat spreader (Cu lid)
TIM 2 (lid → cold plate)grease or PCM · 25-100µm bond line
Cold plateCu with skived fins or microchannels · flow through
Coolant film (convective boundary)
Bulk coolant (technology cooling water) · ~35°C
Manifold, UQD, hose → CDU
Heat exchanger inside CDU
Facility water loop · ~30°C
Facility heat rejection (dry cooler / cooling tower)
Atmosphere · wet bulb or dry bulb temperature

Each layer is a thermal resistance in series. R_total = R_die + R_TIM1 + R_lid + R_TIM2 + R_plate + R_convection + R_cdu + R_facility. Total temperature drop from junction to atmosphere is the sum of layer drops. Any layer that fails silently (e.g. TIM pump-out) shifts the whole distribution and pushes T_j toward T_j_max.

2Convection · Why liquid replaces air





In brief


CDU architecture, coolant chemistry and facility loop separation determine how efficiently heat leaves the chip package. Single-phase glycol handles up to 400W per die comfortably. Two-phase dielectric extends past 1,500W. Fluid degradation, secondary-loop maintenance and heat-exchanger fouling add operating cost that offsets some of the density gain.








The heat transfer coefficient sets the regime


The parameter that separates air from water from a boiling refrigerant is the heat transfer coefficient, h, measured in W/m²·K. It says how many watts a square metre of surface can move per degree of temperature difference between surface and fluid. Air with a fan behind it delivers h of roughly 50-200 W/m²·K. Water flowing through a cold plate at reasonable velocity delivers 5,000-30,000 W/m²·K. A boiling dielectric fluid can hit 100,000+ W/m²·K. Two orders of magnitude between each rung.


That is the entire reason liquid replaces air at rack power densities above 30-50 kW. To move the same heat through air, you need a hundred times the surface area or a hundred times the temperature difference. Neither is available inside a rack that already fills its physical envelope with compute and has a T_j budget of maybe 30°C above ambient.




Heat transfer coefficient · every cooling regime, one axis
Approximate h ranges (W/m²·K), log scale, natural to forced to phase-change

Standard heat-transfer textbook ranges (Incropera, Bergman). Actual h in a given design depends on fluid, velocity, surface enhancement and phase behaviour. The point is the orders of magnitude between regimes.





Reynolds, Nusselt and why cold plates have fins


Two dimensionless numbers govern forced convection. The Reynolds number (Re = ρuL/μ) says whether flow is laminar or turbulent. Below Re ≈ 2,300 in a pipe, flow is laminar and heat transfer is set by molecular diffusion alone. Above, flow is turbulent and mixing dramatically increases heat transfer. The Nusselt number (Nu = hL/k) says how much better convection is than pure conduction through the same fluid layer. Higher Nu, better cooling per unit area.


A cold plate's job is to maximise h and A simultaneously. Fins add surface area (A). Turbulence-inducing features (dimples, offset fins) raise Nu, which raises h at the cost of pressure drop. Microchannels do both: sub-millimetre channels have very high Nu and very high effective A per unit volume, at the cost of significant ΔP. Every cold plate architecture is a specific choice on the A. H. ΔP surface.




Cold plate architectures on the performance vs pressure-drop plane
Effective h·A per unit footprint vs pressure drop, bubble size = manufacturing complexity

Indicative positioning: stamped < skived fins < microchannel < microjet on h·A. Pressure drop rises in the same order. Manufacturing complexity (bubble size) is qualitative: stamping is trivially reproducible; microjets require precision fluidics not routinely mastered.


3Cold plate mechanics



Three geometries, one trade-off


Cold plates split into three manufacturing families, each producing a different point on the h·A vs ΔP surface.





Skived fins


A cutting tool shaves ultra-thin fins (0.05-0.3mm) directly from a solid copper block, leaving fins integral to the base. No thermal interface between fin and base. Thousands of fins per plate. Downside: fins must be robust enough not to bend under coolant flow, limiting how thin they can be.


Boyd Thermal, Auras and AVC lead. R_thermal typically 0.05 K/W or better on a well-designed 30mm die footprint at 5 L/min.




Microchannel


Channels etched or milled into the plate surface, then sealed with a top cover. Channels are 50-500µm wide, 1-2mm deep. Very high h from small hydraulic diameter and turbulent boundary layer. Higher ΔP, more demanding sealing. NVIDIA is pushing microchannel designs into the Rubin generation for accelerator sockets.


Auras, Jentech Precision, JetCool. R_thermal down to 0.03 K/W.






Cold plate cross-sections · flow direction into the page
Simplified profiles of the three main architectures



Stamped

coolant in ↑

die below

Formed sheet metal, wide channels. Low ΔP, moderate h. Aluminium common.



Skived fin

coolant in ↑

die below

Thin fins shaved from solid Cu. High A, no fin-base joint. Premium method.



Microchannel

coolant in ↑

die below

Sub-mm channels etched or laser-welded. Highest h. Higher ΔP, harder to seal.



All three architectures are copper (with occasional Cu-Al hybrids). Fin density and channel width set both the heat transfer coefficient and the pressure drop the pump must overcome.





What NVIDIA moved to in the Rubin generation


The GB300 platform used six independent small cold plates per compute tray with individual adapters and hoses feeding each. The Rubin generation reverses this: a single large cold plate per tray with microchannel regions over each GPU location, formed by laser welding a top cover onto a machined baseFewer joints, fewer potential leak paths, and higher local flow velocity through the microchannel zones over each die. The trade-off is that the assembly is precision-critical: the microchannel weld has to seal cleanly at operating pressure without deforming the channel geometry.



4Thermal interface materials



Fifty microns of grease decide the outcome


Two solid surfaces in contact touch each other at a few high spots. The apparent contact area is essentially all air gaps at the microscopic level. Air has thermal conductivity of about 0.03 W/m·K, three orders of magnitude worse than metals. The actual thermal contact resistance across a bare die/lid interface is therefore dominated by whatever thin fluid or paste sits in those gaps. The industry has spent decades optimising that fluid. It is the highest-margin layer in the thermal stack: tiny bill of material, large consequence of failure, extreme qualification barrier.




TIM types compared across four properties
Radar chart: thermal conductivity, reworkability, cost-per-unit, high-temperature stability. Larger area = better overall.

Normalised 0-1 per axis. Reworkability is scored inverse-difficulty (grease easy, sinter hard). Cost-per-unit is scored inverse ($/kg midpoint, log-normalised). Values indicative from vendor datasheets and industry references.





The five TIM families


 * Silicone greases. Silicone base loaded with ceramic filler (aluminum oxide, zinc oxide, boron nitride) at 40-70% by volume. Thermal conductivity 3-8 W/m·K. Cheap, easy to apply, easy to rework. The default for most commercial silicon. Failure mode: pump-out, where thermal cycling gradually squeezes grease out of the interface, leaving air gaps. That matters when a training cluster power-cycles for years.
 * Phase-change materials (PCM). Solid at room temperature, liquefy above ~45°C and flow to fill the interface, then re-solidify on cooldown. Honeywell PTM7950 is the reference part. Thermal conductivity ~8 W/m·K, roughly comparable to a good grease, but the phase-change mechanism eliminates pump-out. Standard in AI accelerator packages.
 * Sintered silver. Silver nanoparticles in a paste, applied wet and cured at ~250°C to form dense polycrystalline silver. Thermal conductivity 200+ W/m·K. Bulk-metal reliability. Standard for the die-attach layer inside SiC power modules; increasingly considered for GPU-to-lid interfaces. Cost and precious-metal exposure limit the addressable market.
 * Liquid metal (gallium-based). Gallium-indium-tin eutectic (Galinstan) with ~30 W/m·K, best of any TIM. Applied as a liquid at room temperature, wets the surfaces perfectly. Two catches: gallium aggressively corrodes aluminium (so it cannot touch Al surfaces), and it dries out over years by diffusing into the metals it contacts. Increasingly used in gaming CPUs, cautiously in servers.
 * Indium foil. Pure indium at 80 W/m·K. Applied as a soft foil, deforms under compression to fill gaps. Expensive but reworkable. Standard in laser diode packages and high-end industrial modules.



Bond line thickness is as important as material. Thermal resistance across a TIM is R = t / (k · A)A 50µm bond line of 5 W/m·K grease has the same resistance as a 200µm bond line of 20 W/m·K liquid metal. Cutting bond line by half doubles the resistance improvement of a material change. Cold plate flatness and compression uniformity earn the engineering attention on exactly that arithmetic.

5Quick disconnects



A coupling is a spring-loaded valve with a very specific job


A universal quick disconnect (UQD) has one function: let a server be pulled from a rack without draining the loop or dripping onto energised electronics. That sounds simple. The engineering that makes it reliable at 5,000+ mating cycles under 100+ psi at hyperscale volumes is not. Three companies hold 80% of the market accordingly, and the price umbrella against Chinese entrants was the entire premise of the market essay's chokepoint thesis.


The mechanism is a double-sealing poppetTwo spring-loaded valves face each other. Before mating, each valve is closed by its own spring, so no fluid escapes when the halves are apart. During mating, the valves press against each other, break their seals in sequence, and open a flow path through the joined coupling. Elastomer O-rings form the leak-tight fluid seal at the interface. On disconnection, springs re-close the valves before the halves separate, guaranteeing zero drip.




Universal quick disconnect · engaged state, cross-section
Half A (left) mated to Half B (right). Coolant flows left-to-right through the poppet channels.


Half A · rack manifold side
Half B · server side

↔ Housing (SS316)
↔ Housing (SS316)

O-ring seal (FKM)
O-ring seal (FKM)

Poppet A
Poppet B

Spring (Inconel)
Spring (Inconel)

Coolant → →
→ → Coolant

Mated interface · poppets pressed open · O-rings sealed

Simplified schematic. Precision tolerances between poppet, seat and housing are ~0.05mm, which is the tolerance limit that makes qualification hard. Elastomers must survive 5,000+ mating cycles, thermal cycling from 5°C to 60°C coolant temperatures, and continuous immersion in glycol/water without swelling, hardening, or leaching plasticiser.



UQD reliability window · what a coupling has to survive
Cycle count vs peak flow rate for a typical rack-manifold UQD, log axes

Indicative envelope. Points at the top-right (many cycles, high flow) are the design targets. Real qualification includes shock, vibration, temperature soak, thermal cycling and long-term compression set on the elastomer seals, none of which this simple chart captures.


6CDU internals



The interface between IT and facility, in detail


A CDU (coolant distribution unit) does four things simultaneously: it moves coolant, it filters it, it transfers heat between the technology loop and the facility loop, and it monitors and controls the whole assembly. Every one of those functions is mature commodity technology; the integration into a deployable box that passes hyperscale qualification is where value sits.





Stage 1
Return manifoldWarm coolant arrives from the servers via rack quick-disconnects and hoses
Typical return T: 50-55°C

↓

Stage 2
FilterParticulate removal, typically 25µm nominal. Prevents fouling of heat exchanger and cold plate microchannels
Cartridge or bag filter

↓

Stage 3
Plate heat exchangerBrazed-plate exchanger with hundreds of thin channels alternating TCS and FWS flow. High effectiveness (85-95%) in compact volume
Alfa Laval, Kelvion, SWEP

↓

Stage 4
Cooled TCS coolantNow at supply temperature, ~35°C, ready to send back to servers
Typical supply T: 30-40°C

↓

Stage 5
Variable-speed pumpCentrifugal pump with VFD, controlled to maintain rack-side differential pressure setpoint
Grundfos, Wilo, Xylem

↓

Stage 6
Supply manifoldChilled coolant distributed back to server rack via quick-disconnects
Loop pressure: 30-50 psi





Two loops meet inside the CDU: the technology cooling system (TCS), which is the clean, filtered, closely-monitored loop that touches the servers, and the facility water system (FWS), which is the site-wide loop that eventually gives heat to the atmosphere. Keeping them separate is a design imperative: FWS is dirtier, harder to control, and more variable, and any contamination migrating from FWS into TCS would eventually plug cold plate microchannels.




Water chemistry, in short


Coolant in modern hyperscale halls typically runs as a 25% propylene glycol / 75% water blend, with a corrosion inhibitor package and a biocide, rather than as plain water. Ethylene glycol is more thermally efficient but has been displaced by regulation on toxicity grounds. pH is maintained between 8.5 and 9.5 to passivate copper. Conductivity is held below 10 µS/cm to prevent galvanic corrosion between the mixed metals in the loop. Dissolved oxygen is scavenged. Biological growth is suppressed with an inhibitor that must be re-dosed periodically. Every one of these parameters is monitored continuously in a properly instrumented CDU.



7Immersion & phase change



Two-phase cooling: the physics of a boiling coolant


The heat transfer coefficient of boiling is an order of magnitude higher than any single-phase forced convection. Boiling water on a hot surface pulls heat at 10,000-100,000+ W/m²·K. That is why two-phase immersion cooling remained a serious contender through the mid-2020s despite its regulatory and material headaches.


The mechanism is the nucleate boiling curveAs a hot surface heats a stagnant liquid, four regimes appear in sequence with increasing surface temperature. First, natural convection (h low). Then nucleate boiling: bubbles form at nucleation sites on the surface, rise into the bulk, carrying latent heat with them. This is the useful regime, with very high h at small ΔT. If surface temperature keeps rising, bubbles merge into a vapour film that insulates the surface. That inflection is the critical heat flux point. Beyond CHF, film boiling collapses h and the surface temperature runs away, which in silicon terms means the die burns.




The boiling curve · why nucleate boiling is the sweet spot
Heat flux vs surface superheat (ΔT above saturation), log axes, general shape for a dielectric coolant

Classic Nukiyama boiling curve. Peak (critical heat flux) at ~30 W/cm² for a typical fluorocarbon dielectric on plain silicon, higher with surface enhancement. Beyond CHF, film boiling takes over and heat transfer collapses.





Single-phase vs two-phase, mechanically





Single-phase immersion


Servers submerged in a tank of dielectric fluid. Fluid stays liquid throughout, driven by forced convection (pump) or natural convection (rising warm fluid, sinking cool fluid). External heat exchanger transfers heat from the tank fluid to a facility water loop.


Simpler: no phase management, no vapour pathway design, fewer failure modes.


Fluid options: mineral oil, synthetic hydrocarbons (Submer/Castrol), gas-to-liquid (Asperitas/Shell), engineered hydrocarbons (GRC).


PUE routinely 1.02-1.05.




Two-phase immersion


Servers submerged in a low-boiling-point fluid (typically ~50°C boiling point). Heat causes the fluid to boil on hot surfaces; vapour rises to a condenser coil at the top of the tank; condensed liquid falls back onto components.


Higher h (nucleate boiling), better temperature uniformity, no pumps required.


Fluid options: historically Novec 7100/649, Fluorinert FC-72. Post-3M PFAS exit: Opteon SF33 (interim), engineered PFAS-free replacements under development.


PUE below 1.02 achievable.





The 3M PFAS exit in December 2022, and its complete stop of Novec/Fluorinert production by end-2025, was a supply shock, not a technology defeat. Two-phase immersion still works. The engineered replacements (Chemours, Solvay, ZutaCore-developed) are getting there. But ECHA's PFAS restriction opinion, expected end-2026, could pull the entire fluorinated family out of the addressable market for European operators. The market essay covered the strategic consequence; the technical reality is that the physics of two-phase is compelling and the chemistry is broken, and the industry is waiting on the second to catch up with the first.



8Facility heat rejection



Where the heat actually goes


Coolant leaves the CDU cool and returns warm. The warmth has to go somewhere. That somewhere is the atmosphere, via a facility loop, and the mechanism used sets both the annual water consumption and the achievable PUE.




MethodApproach ΔTWater useEfficiencyWhere it wins

Compression chillerN/A (mechanical)LowPUE 1.3-1.5Hot, humid climates. Backup for other methods.
Cooling tower (evaporative)3-5°C above wet bulbVery highPUE 1.15-1.25Any climate with water. Politically exposed in water-stressed regions.
Adiabatic dry cooler5-8°C above wet bulbModerate (spray)PUE 1.10-1.20Hot, dry climates. Water use dramatically lower than cooling tower.
Dry cooler (finned tube)10-15°C above dry bulbNonePUE 1.08-1.15Cool climates. Zero water. Only works when outdoor T is well below coolant T.
Direct heat reuseDepends on offtakeDependsPUE < 1.0 effectivelyNear district heating grids. Regulatory tailwinds in the EU.






Why 45°C hot water changes the map


Traditional data centre cooling assumed the facility loop needed to be cold: 12°C or 20°C to feed CRAH air handlers, or cool enough to accept heat from air-cooled racks. Cold facility water means chillers, always, everywhere. The Rubin platform (unveiled at CES 2026) specifies 45°C inlet waterThat number is chosen deliberately. In most inhabited climates, outdoor wet bulb temperature is below 30°C for most of the year, so 45°C facility water can be cooled by a dry cooler alone: no chiller, no evaporation, no water consumption. Free cooling becomes the default rather than the exception.




Free cooling hours per year · with 45°C hot water, most climates work
Approximate hours per year where outdoor wet bulb is below 35°C (enough margin for 45°C coolant supply)

Order-of-magnitude estimates from climate normals. Actual free cooling economics depend on approach temperature, dry cooler sizing and site-specific conditions. Northern European sites can free-cool essentially the entire year; Middle Eastern sites need supplemental cooling in summer.


9Cooling migrates into the package



The frontier: shortening the thermal path


Every millimetre of the thermal stack from the die to the cold plate is a source of resistance. The current wave of research is aimed at deleting that path entirely. Three architectures matter, each closer to the die than the last.





2026 · shipping (Rubin)
Microchannel lid (MCL)
Instead of a flat copper lid conducting heat sideways into a cold plate above it, the lid is machined with micron-scale channels through which coolant flows directly. Eliminates one TIM layer (lid → cold plate) and one solid-conduction segment. The cold plate is now essentially a distribution manifold feeding the lid channels.


2026 · Rubin
Microchannel cold plate (MCCP)
The lid is retained but the cold plate uses microchannel regions precisely over each GPU location, formed by laser-welding a top cover to a milled base. Higher local h at each hot spot, single large plate footprint, fewer joints per tray.


2027-29 · pilot
Backside liquid cooling
Channels etched into the silicon substrate itself, from the backside (opposite the active devices). Coolant flows through the die. Eliminates both TIM layers and the lid entirely. TSMC is integrating microchannel liquid cooling into its 3DFabric platform (CoWoS, SoIC). IMEC, ETH Zürich / Corintis, JetCool have demonstrated variants. Manufacturing challenge: sealing a fluid path through a package that must survive years of thermal cycling and warpage.


2030s · research
In-die microchannels
Coolant channels etched into the active silicon between functional blocks, or between stacked dies in a 3D package. Ultimate limit of the migration. Requires solving fluid delivery through the substrate and long-term reliability of silicon-fluid interfaces.



Every migration deletes one thermal-stack layer. Each deletion improves the R_thermal budget by 10-30% and moves the whole architecture towards higher heat flux capability. On a five-year view, the cold plate becomes a manifold and the value migrates into the package.The frontier thesis

10Market segmentation



How the market breaks down, visualised


The market essay described the layers by revenue and margin. Two visualisations make the shape clearer than a table.




Data centre cooling market by layer and geography · a Marimekko view
Column width = share of total cooling market revenue by layer. Row shading = approximate regional origin of supply. 2026 basis.



Cold plateTaiwan · 40%
 China · 30%
 US · 20%
 EU/Other · 10%



CDUUS · 45%
EU · 25%
China · 20%
Other · 10%



FacilityUS · 35%
EU · 30%
Japan · 20%
Other · 15%



TIMUS · 35%
Japan · 35%
EU · 20%
Other · 10%



UQDUS · 45%
EU · 35%
CN · 20%



FluidsEU · 35%
US · 30%
Japan · 20%
Other · 15%



ServicesUS · 60%
EU · 25%
APAC · 15%


← larger columns = larger layerscolumn width = share of DC cooling revenue
Layer shares approximate, derived from Persistence/GMI/Mordor 2026 estimates. Geography splits are the author's directional decomposition, not published data. Illustrative reading device, not a source.



Where the thermal resistance is spent · waterfall
Cumulative K/W added by each layer of the stack for a typical 700W GPU cooling path

Indicative decomposition. The two TIM layers together account for a substantial share, comparable to the cold plate itself, which is why the frontier (MCL, backside cooling) is aimed at deleting them.



Sensitivity: how much Tj drops if you improve each layer by 20%
Estimated junction temperature reduction (°C) from a 20% thermal-resistance improvement at each layer

Author's calculation from the resistance decomposition above. The layers with the largest current resistance (TIM 1, TIM 2, cold plate) offer the biggest sensitivity improvements. The industry engineering budget concentrates there for that reason.


11Where this lands



The physics one paragraph


Every gram of silicon in a hyperscale rack is a heat source with a fixed T_j budget. The rack's job is to move that heat to a fluid at low enough temperature to accept it. The mechanisms available (conduction and convection) are constrained by Fourier's law and Newton's law of cooling. Every architectural advance from the past three years is a specific answer to one of the terms in those equations: raise h at the die (microchannels, backside cooling), raise A at the die (fins, immersion), or lower coolant temperature at the die (45°C hot water via dry cooling). Nothing is going to change the underlying laws. What is changing is how efficiently the industry approaches them.


For a reader coming from the market essay, three concrete implications. First, the profit-pool inversion is real physics, not just moat economics: the small components (TIM, UQD) touch the highest thermal gradient and carry the biggest failure consequence, which is why qualification barriers are so hard and margins so defended. Second, backside cooling is the swing factor: if it clears manufacturing at scale, the cold plate layer materially commoditises. Third, 45°C hot water plus dry cooling is the sustainability story that actually holds: every other angle depends on it.



SERIESThe AI Power Chain



What connects to what


This is a technical companion to The Thermal Stack (Part III of the AI Power Chain), which covers the market structure, players, TAM and M&A of data centre cooling. Part II covered The Wide-Bandgap Stack, the power semiconductors whose efficiency partly determines how much heat this cooling infrastructure must remove; its technical companion (The Wide-Bandgap Stack: Technical Companion) walks through the semiconductor physics in the same way. Part IV (The Interconnect Stack) covers the interconnect layer: transformers, switchgear, and the grid equipment that gates the buildout.




Series. The AI Power Chain (six parts). Part I: The Capacitor StackPart I-A: Technical CompanionPart II: The Wide-Bandgap StackPart II-A: Technical CompanionPart III: The Thermal StackPart III-A: Technical CompanionPart IV: The Interconnect StackPart IV-A: Technical CompanionPart V: The On-Package Delivery StackPart V-A: Technical CompanionPart VI: The Modular Datacenter StackPart VI-A: Technical Companion.


Methodology. This is a technical primer, not investment advice or engineering guidance. Heat transfer parameters, process descriptions and roadmap items are compiled from standard references (Incropera and Bergman on heat transfer, ASHRAE TC 9.9 for data centre thermal classes), vendor technical literature (Boyd Thermal, CoolIT, Vertiv, Schneider, Honeywell, Shin-Etsu, Stäubli, CPC, Parker), IEEE and ASHRAE publications, and trade press. Any specific performance number should be verified against the vendor's current datasheet before use in design work.


© 2026 Adi Kumar · Power & Digital Infrastructure · Corrections welcome.












Frequently asked


What is a CDU in liquid cooling?

A coolant distribution unit (CDU) separates the primary loop (facility water) from the secondary loop (rack coolant) via a heat exchanger. It manages flow, temperature and pressure on the rack side. CDU capacity typically ranges 500kW-2MW per unit. Row-CDU vs rack-CDU deployment choice affects hydraulic complexity + fault isolation.


What is the difference between single-phase and two-phase immersion cooling?

Single-phase immersion submerges servers in a dielectric fluid that stays liquid throughout the cooling cycle. Two-phase uses a low-boiling-point fluid that vaporises at the chip, condenses at the cooling plate. Two-phase handles higher heat flux (1,500W+ per package) but adds fluid degradation and containment complexity. Adoption trails single-phase.


==============================================================================
# The Capacitor Stack: Technical Companion
URL: https://adikumar.co/the-capacitor-stack-technical-companion/
Published: 2026-08-01
Summary: How MLCC, film, Al electrolytic, EDLC, HSC and silicon capacitors actually work. Ragone plot, failure modes, and roadmap. Technical primer.
==============================================================================

The AI Power Chain series · Part 8 of 15











ADI KUMAR · POWER & DIGITAL INFRASTRUCTUREAUGUST 2026 · V1 · ~40 MIN READ


The AI Power Chain · Part I-A · Technical Companion
How AI data centre capacitors actually work



The Capacitor Stack: Technical Companion


Where the market essay named the vendors, this piece explains the physics. What a farad actually is, why a hybrid supercapacitor beats a battery on a 10-second pulse and loses on a 10-minute one, why MLCCs derate under DC bias, and why a fifty-cent film capacitor determines whether an 800VDC rack survives its own switching noise. Cross-sections, Ragone plots, and roadmaps for readers who want the machine under the market.



By Adi Kumar
Independent analysis
Views are my own




How to read this piece. It is a technical companion to The Capacitor Stack: How Energy Storage Became Core Compute Infrastructure (Part I of The AI Power Chain). The market piece named the vendors, sized the layers, and mapped the M&A. This one explains the underlying physics of every capacitor category on that stack, why each one owns the timescale it does, and what the roadmap items on every analyst's list actually change. No prior electrical background is assumed, but the reader should know what "capacitor" means at a systems level.


1First principles · What a capacitor stores



The bucket, the pipe, and the timescale


A capacitor is an energy-storage device, and so are batteries, springs, flywheels, compressed-air tanks and raised weights. The distinguishing property of a capacitor is not that it stores energy but how fast it can put the energy in and take it back outEvery capacitor architecture on the AI power chain occupies a specific point on the energy-vs-power surface, and the whole essay you just read (or are about to) is fundamentally a walk across that surface.


The physics is one equation: E = ½·C·V²Stored energy scales linearly with capacitance (C, in farads) and quadratically with voltage (V). Charge on the plates: Q = C·VCurrent in or out of a capacitor: i = C·(dV/dt)That last one is where the AI story lives. To source a 100-amp current spike at constant voltage, you need a capacitor big enough to release the charge without letting its voltage collapse. The size of that capacitor depends on how long the spike lasts: a microsecond spike needs a very different bucket than a ten-second one.




Six capacitor chemistries on the Ragone plane
Specific energy (Wh/kg) vs specific power (W/kg), log-log. Each bubble marks the region a technology occupies. Bigger bubbles = wider range within that class.

Classic Ragone framing. MLCC and silicon capacitors sit at extreme power density but negligible energy; Al electrolytic in the middle; EDLC/HSC bridging to Li-ion territory. The distance between the top-right and bottom-left points is roughly seven orders of magnitude on each axis, which is why "capacitor" is not one product category.




The Ragone plot answers the "what does this thing do" question. A charge stored at the top-right (high power, low energy) empties in microseconds. A charge stored at the bottom-right (high energy, low power) takes hours. In between sits the whole ladder covered in the market essay: MLCCs handle the microsecond decoupling under a GPU; Al electrolytics handle the millisecond bulk on a power shelf; hybrid supercapacitors handle the ten-second rack-level absorption; batteries handle the minutes-to-hours autonomy. Same physics, different regions of the plane.



Why each technology sits where it sits. A capacitor's energy density is limited by two things: how much charge you can put on the electrodes and how high a voltage you can hold across them without breaking down the dielectric. Ceramic MLCCs use very thin, very high-permittivity dielectrics; energy per cell is tiny but they respond in nanoseconds. Electrolytics use a chemically-grown oxide film so thin (nanometres) that it holds enormous field per volt, but the electrolyte conducts current only so fast. EDLCs use no chemical bond at all: energy is stored in an ion double-layer at an electrode surface, which is why they cycle a million times but store little energy per kg. HSCs add a Faradaic anode to a capacitive cathode and land in the middle. Every device is a specific trade against the same two constraints.





2Three storage mechanisms






In brief


Hybrid supercapacitor, EDLC, MLCC and silicon-capacitor electrical characteristics determine which capacitor category solves which AI-power supply problem. HSCs cover the millisecond-to-second energy buffer for rack-level ride-through. MLCCs sit on the GPU package for high-frequency decoupling. Silicon capacitors address on-die transient response. Substitution across categories is limited by frequency + voltage class.








Electrostatic, Faradaic, and the hybrid


Under the covers, capacitors store energy by one of three physical mechanisms. The mechanism sets the timescale, the cycle life, and the density.





Electrostatic (dielectric)


Energy stored in the electric field of a physical dielectric between two conductive plates. No chemistry, no mass transport. Charge appears at the plate surface as soon as voltage is applied. Loss mechanism: dielectric loss (leakage current, hysteresis in the polarisation of the dielectric material).


Where used: ceramic MLCCs, film capacitors, silicon capacitors, mica.


Response time: nanoseconds.
Cycle life: effectively unlimited.
Energy density: very low (0.001-0.05 Wh/kg).




Faradaic (electrochemical)


Energy stored via reversible chemical reactions at electrode surfaces or in the bulk. Charge involves ion diffusion, so mass transport limits how fast you can extract it. Every charge cycle involves atomic rearrangement, which is why cycle life is finite.


Where used: batteries (Li-ion, NiMH, lead-acid), pseudocapacitors.


Response time: milliseconds to seconds.
Cycle life: hundreds to thousands.
Energy density: high (50-300 Wh/kg for Li-ion).







Electric double-layer (EDLC)


A special case of electrostatic storage at the microscopic scale. When a voltage is applied across two porous carbon electrodes in an electrolyte, ions in the electrolyte migrate to the electrode surface and form a two-molecule-thick layer of separated charge (the "Helmholtz double layer"). The separation is on the order of a nanometre, but the surface area of activated carbon is enormous (up to 3,000 m²/g), so total capacitance is large. No chemical reaction: ions just sit.


Response time: milliseconds to seconds.
Cycle life: hundreds of thousands to millions.
Energy density: medium (4-8 Wh/kg).




Hybrid (HSC / LIC)


An asymmetric device that combines an EDLC-style cathode (capacitive, high-power) with a battery-style anode (Faradaic, high-energy). Lithium-ion capacitors (LICs) are the dominant hybrid architecture: the anode is graphite or hard carbon pre-doped with lithium; the cathode is activated carbon. The pre-doping raises the operating voltage from ~2.7V (symmetric EDLC) to ~3.8V, and since energy scales as V², the energy density roughly doubles.


Response time: tens of milliseconds to seconds.
Cycle life: 50,000 to 1M+.
Energy density: 15-30 Wh/kg.






Five chemistries on six axes
Radar plot, each axis normalised so 1.0 = best in class. Larger area = more balanced trade-offs. Values indicative.

Axes: energy density, power density, cycle life, temperature range, cost per Wh (inverse), fast-charge tolerance. No single chemistry dominates every axis, which is why the capacitor stack is a stack, not a winner-take-all.


3Ceramic MLCC



The most mass-produced electronic component on earth


Multi-layer ceramic capacitors (MLCCs) are the smallest, cheapest, fastest capacitors on the AI power chain, and they are made in enormous numbers: trillions per year, mostly in Japanese and Korean fabs. A modern AI accelerator package sits on a printed circuit board carrying thousands of MLCCs, most of them stamped into the millimetre-square footprint immediately below the die. They handle the microsecond-scale current transients that no larger capacitor can physically respond to fast enough.




MLCC cross-section · alternating ceramic and electrode layers
Simplified. Real parts have 100-1,000+ layers stacked in a chip 0.2mm × 0.4mm.

Termination · Ni + Sn plating
Ni internal electrode 1
BaTiO₃ dielectric~1µm thick
Ni internal electrode 2
BaTiO₃ dielectric
Ni internal electrode 3
BaTiO₃ dielectric
.repeat 100-1,000+ times. .
BaTiO₃ dielectric
Ni internal electrode N
Termination · Ni + Sn plating

Each electrode connects to one of two terminations at the chip ends. Electrodes alternate, so half connect to termination A and half to termination B. Capacitance is proportional to layer count × electrode area × dielectric constant, divided by dielectric thickness. Modern parts push layer counts past 1,000 and dielectric thickness below 0.5 µm.





Dielectric materials and the temperature-vs-capacitance trade


The dielectric that separates the electrodes decides almost everything about the part. Two chemistries dominate.


Class II (X7R, X5R): ferroelectric BaTiO₃ (barium titanate) with additives. High permittivity (dielectric constant of 2,000-4,000) packs a lot of capacitance into a small volume. But permittivity varies with temperature, voltage, and time. X7R holds within ±15% over −55°C to +125°C; X5R over −55°C to +85°C. This is the workhorse chemistry for AI decoupling because it is small and cheap.


Class I (C0G / NP0): mostly paraelectric ceramics like calcium zirconate. Permittivity is much lower (30-100) so the parts are physically bigger per farad. In exchange: essentially zero capacitance drift with temperature, voltage or time. Used in precision analog and RF filters, not for bulk decoupling.




MLCC capacitance derates under DC bias · the trap datasheet vendors don't lead with
Effective capacitance vs applied DC voltage as % of rated voltage, X7R vs X5R vs C0G

Class II dielectrics (X7R, X5R) lose capacitance under DC bias because the BaTiO₃ crystal domains partially polarise. At rated voltage, a nominally 10µF X7R part can behave as 3-5µF in circuit. C0G/NP0 does not derate. That is why a "10µF X7R" bulk cap next to a GPU may be doing less than an engineer assumes.





ESR, ESL, and why parallelism matters


Every real capacitor has parasitic equivalent series resistance (ESR) and equivalent series inductance (ESL). ESR limits how much AC current the cap can carry without dissipating heat. ESL limits how fast the cap can respond. Above the self-resonant frequency (SRF, roughly f = 1/(2π·√(L·C))), an MLCC behaves as an inductor, not a capacitor, and its impedance rises with frequency instead of falling.


The response to this is to parallel many small MLCCs of different values, so that at every frequency some capacitor is below its SRF and providing low impedance. An H100 accelerator socket typically has hundreds of MLCCs paralleled beneath it for exactly this reason. As GPU switching frequencies push into the GHz, the physical limit becomes the loop inductance of the PCB itself, which is why the market essay's "silicon capacitor" and "embedded MLCC" categories exist: put the capacitance inside the package, adjacent to the die, and you eliminate the PCB inductance.




Impedance vs frequency: single MLCC vs paralleled array
Log-log impedance of a single 10µF 0402 X7R vs an array of ten paralleled parts across four case sizes

A single cap has one self-resonant notch, above which it looks inductive. An array of different case sizes has multiple notches, holding low impedance across a wider band. This is the physics behind the "hundreds of MLCCs under a GPU" observation.


4Aluminium electrolytic



The polar oxide that made bulk capacitance affordable


Aluminium electrolytic capacitors are the workhorse of power electronics: the fat cans on any switching supply, the DC-link storage between rectifier and inverter, the bulk hold-up on a server power shelf. The physics is different from MLCC in one important way: the dielectric is not a ceramic slab but a chemically-grown oxide film, only tens of nanometres thick, formed on the surface of etched aluminium foil.




Aluminium electrolytic capacitor · rolled foil construction
A very thin oxide film on an etched foil surface acts as the dielectric.

Aluminium can (case)
Cathode foil (Al, un-etched)
Paper separator, soaked in electrolyte
Al₂O₃ oxide dielectric · 10-100 nm
Anode foil (Al, etched · 100× surface area)
Paper separator, soaked in electrolyte
Cathode foil
Rubber bung + lead-out wires

Rolled up into a can. The electrolyte acts as the "true" cathode; the un-etched Al foil is just a current collector. Because the dielectric is grown on the anode side only, these devices are polar: reverse voltage destroys the oxide within seconds.




Two properties make Al electrolytics unique. First, the etched aluminium anode has enormous effective surface area: the tunnels etched into the foil multiply the geometric surface by 30-100×. Second, the oxide layer's thickness scales with the "forming voltage" applied during manufacture, so a part rated for 400V has a thicker oxide than one rated for 25V. This combination gives excellent capacitance-per-volume at moderate voltages, at low cost.


The downsides that AI accelerated interest in alternatives: ESR rises sharply at low temperature (the electrolyte becomes viscous, ion mobility drops), lifetime is exponentially temperature-dependent (rule of thumb: lifetime doubles for every 10°C reduction in operating temperature), and the electrolyte dries out over years, particularly in hot enclosures. Modern data-centre PSUs, running continuously at elevated temperatures inside dense racks, exhaust electrolytic caps in three to five years. That is the driver for the market essay's observation that Al electrolytics are being displaced up-shelf by polymer, hybrid, and HSC alternatives.




Al electrolytic lifetime is temperature-limited · Arrhenius scaling
Predicted operating life at rated ripple vs ambient temperature, log axis. 105°C-rated part.

Classical L = L₀ · 2^((T₀−T)/10) relation. A part rated 5,000 hours at 105°C runs 80,000 hours at 65°C and 320,000 hours at 45°C. The "why polymer capacitors dominate cool applications" argument in one chart.





Polymer, hybrid, and the migration up-shelf


The polymer aluminium capacitor replaces the liquid electrolyte with a solid conductive polymer (typically PEDOT). Two consequences: ESR drops by an order of magnitude (polymer conducts electrons, not just ions), and lifetime rises substantially because there is no liquid to dry out. Trade-offs: polymer parts are more expensive per farad, and they have lower voltage ratings than liquid electrolytics (typically ≤100V). Hybrid polymer mixes both: polymer near the cathode for low ESR, liquid electrolyte elsewhere for higher voltage capability. These are the parts filling the space between "MLCC arrays" and "big bulk cans" in modern high-density power stages.



5Film capacitors



The self-healing plastic film that survives 1MV/mm


Film capacitors are the least glamorous device on the AI power chain and, in high-voltage duty, the least replaceable. A film cap is two rolls of metallised plastic film (polypropylene, typically) wound together. The dielectric is the plastic itself, which can withstand electric fields of hundreds of volts per micron. The "metallised" part is what makes the device special: instead of solid metal foil electrodes, a very thin metallisation (aluminium or zinc, 10-100 nm thick) is vacuum-deposited on one side of the film. If a defect punctures the dielectric under high field, the local heat vapourises the thin metallisation around the puncture site, isolating the fault and healing the capacitor. Same device, slightly less capacitance, still working.


Self-healing is why film caps are the DC-link and snubber choice for high-voltage power electronics: inverter DC-links, SST DC buses, MV rectifier stages, EMI filters on high-voltage rails. In an 800VDC data-centre architecture, film caps handle the DC-link buffering on the sidecar or rack-level PSU. The market essay treats them within the broader "bulk board" category, but functionally they are a distinct sub-layer.



The dielectric strength story. Polypropylene has a breakdown field of ~500-700 V/µm. A 5-µm film handles 3kV before breakdown. In practice, film caps are designed with a large margin (rated at ~40% of breakdown field) because insulation strength degrades with time, temperature, and partial discharge. That margin, plus self-healing, gives film caps 20+ year field lifetimes in inverter service, the longest of any capacitor class.

6EDLC · the double layer



Supercapacitors store energy in a nanometre-thick ion layer


An electric double-layer capacitor stores energy without any chemical reaction, in the physical separation of ions from electrons at an electrode surface. When you apply voltage across two porous carbon electrodes immersed in an electrolyte, the electrons pile up on the electrode surface and the electrolyte ions migrate to sit near them, held in place by electrostatic attraction. The separation between electron and ion is on the order of a nanometre, the length scale of the Helmholtz double layer. That tiny gap acts as the dielectric.




EDLC cell · the Helmholtz double layer at each electrode
Two porous carbon electrodes, an ion-conducting electrolyte, a separator. Ions form a double layer at each electrode surface.

Aluminium current collector (+)
Positive electrode · activated carbon (2,000+ m²/g)
Helmholtz layer · anions at surface
Electrolyte · TEA-BF₄ in acetonitrile or PC
Separator (porous polymer)
Electrolyte
Helmholtz layer · cations at surface
Negative electrode · activated carbon
Aluminium current collector (−)

Cell voltage is limited by electrolyte decomposition. Organic electrolytes (TEA-BF₄ / acetonitrile) hold ~2.7-3.0 V; aqueous electrolytes are limited to ~1.2 V; ionic liquids can reach 3.5-4.0 V but are much more expensive.




Two properties make EDLCs distinctive. First, power density is very highBecause charge storage is physical (no chemistry to complete), current only depends on how fast ions can migrate to and from the electrode surface, which happens in milliseconds. Second, cycle life is essentially unlimitedNo chemical reaction means no chemical aging. EDLCs demonstrated in the 1990s are still working today.


The limits come from the electrolyte. Cell voltage is capped by whatever voltage decomposes the electrolyte (2.7V for organic electrolytes, higher for exotic ionic liquids). Energy density scales as V², so pushing voltage is the obvious lever, and the entire technology arc from EDLC to HSC is essentially the story of pushing cell voltage higher via anode chemistry.



7HSC · asymmetric electrode



Hybrid supercapacitor: half battery, half capacitor


The hybrid supercapacitor was invented to close the energy-density gap between EDLC and Li-ion without giving up EDLC's cycle life. The trick is to replace the symmetric activated-carbon architecture with an asymmetric one: a battery-style anode (graphite or hard carbon, pre-doped with lithium ions) paired with a capacitor-style cathode (activated carbon).




HSC / LIC cell · asymmetric electrode chemistry
Battery-style anode (Li-pre-doped graphite) + EDLC-style cathode (activated carbon)

Aluminium current collector (cathode)
Activated carbon cathodeEDLC-like · capacitive storage
LiPF₆ in carbonate electrolyte
Separator
Electrolyte
Graphite (Li-pre-doped) anodeFaradaic · Li⁺ intercalation
Copper current collector (anode)

Pre-doping the anode with lithium ions before cell assembly is the process innovation that makes HSC economical. It raises anode potential, extends cell voltage to 3.8-4.0V, and gives the anode battery-like energy density without introducing battery-like cycle degradation on the cathode.




The asymmetry produces four consequences. First, cell voltage rises to 3.8-4.0V (vs 2.7V for EDLC), which raises energy density by roughly (3.8/2.7)² ≈ 2×. Second, the anode contributes real Faradaic energy density, so total energy density lands at 15-30 Wh/kg, 3-4× a comparable EDLC and roughly 1/10 a Li-ion. Third, the cathode retains capacitor cycle life, so overall cycle count stays at 50,000 to 1M+ cycles. Fourth, the operating voltage window makes HSCs stringable in fewer series cells than EDLC for a given bus voltage: ~200 cells for 800V vs ~300 for EDLC.



Why pre-doping is the hard step. The anode graphite has no lithium in it as manufactured; to raise its potential you have to insert Li⁺ ions before the cell is sealed. Musashi Energy Systems and its peers use a proprietary process where lithium foil is laminated to the anode during assembly, and Li⁺ ions migrate into the graphite over hours in the presence of electrolyte. This step is what makes HSC manufacturing hard to scale and what makes qualified cell capacity the market essay's binding constraint. Cell-line construction dominates the capex.

8Silicon capacitors



Capacitors built with semiconductor process


Silicon capacitors are the newest arrival on the stack, and the most conceptually distinct. Instead of layering ceramics or rolling foils, a silicon capacitor is built inside a silicon wafer using conventional semiconductor process: etch deep trenches into the substrate, coat the trench walls with a thin dielectric (typically SiO₂ or Al₂O₃), fill the trenches with doped polysilicon or metal to form the second electrode. The resulting capacitor has all the properties of the semiconductor process that made it: extreme dimensional precision, sub-picohenry ESL, and packaging options (die-thin, embeddable in substrate, TSV-compatible) that no traditional capacitor class can match.




Silicon capacitor · trench structure in a wafer
Etched trenches, thin dielectric coating, filled with a conductive electrode

Passivation · SiN
Top metal electrode + solder bumps
Dielectric coating (SiO₂ / Al₂O₃ / high-k)~10 nm on trench walls
Silicon substratedeep trenches etched top-down · aspect ratio 20:1+
In-trench polysilicon electrode
Bottom substrate / backside metallisation

Trench aspect ratios of 20:1 or higher multiply the effective electrode area by 20× compared to a flat plate of the same footprint. Capacitance densities of 100-500 nF/mm² are achievable, with parasitic inductance an order of magnitude below any MLCC.




Silicon caps live at the intersection of capacitor and packaging: they can be flip-chip mounted directly onto the substrate under the GPU, or embedded inside the substrate itself, or built as through-silicon-via (TSV) interposer structures inside a 2.5D package. That eliminates the PCB inductance between the die and the decoupling capacitance, which is the ceiling on how fast the whole system can respond to load transients. Murata (via its 2016 acquisition of IPDiA), SEMCO and increasingly TSMC's advanced packaging teams are the visible players; the volume today is small, but the design-in surface area at the accelerator-package level is enormous.



9Ragone plot revisited



Every AI power-chain socket is a point on the Ragone plane


The Ragone plot at the top of this piece maps the physical zone each chemistry occupies. Now overlay the AI power chain's timescale requirements (the four windows the market essay defines) and the technology choice at each layer becomes obvious.




LayerTimescaleEnergy/power pointWinning chemistryPhysics reason

In-package decouplingns. µsUltra-high power, negligible energySilicon capacitor · embedded MLCCOnly technologies with sub-picohenry ESL can respond fast enough at the die scale
Board POL / VRMµs. MsHigh power, small energyMLCC arrays · polymer AlFast enough for VRM control loops; enough capacitance to hold POL bus between switching events
Power shelf bulkms. 100msMedium power, medium energyAl electrolytic → polymer / hybridEnough energy to bridge PSU switching cycles and grid dips; ESR matters more than transient response
Rack sidecar10ms. 10sMedium-high power, medium energyHSC (LIC) · EDLC bridgeLong enough transient that battery-scale energy density becomes useful; short enough that Li-ion cycle life would collapse
Facility BBU / autonomy10s. MinLow power, high energyLi-ionEnergy density dominates; cycle count is low (few outages per year)






Where the market essay's contested socket sits
Waterfall: energy the rack needs to absorb during a synchronised GPU transient, kJ, by mechanism

Illustrative accounting for a 600kW Kyber rack absorbing a 50% load swing (300kW) for 500ms. Total transient energy ~150 kJ. Silicon caps + MLCCs absorb the first few percent (µs. Ms); Al electrolytics absorb the next fraction (ms. 100ms); the residual (majority) hits the rack-level HSC/EDLC shelf. This is the physics behind the market essay's "HSC owns the contested 10ms. 10s window" claim.


10Failure modes



Real capacitors are not their datasheets


Every capacitor category fails differently. Understanding the failure mode is central to specifying the right derating, and central to the reliability engineering the market essay's "qualified capacity" gap is built on.




ChemistryPrimary failure modeMechanismDesign mitigation

MLCCCrackingBoard flex, thermal cycling, solder joint stressCase size / lead style, PCB layout, mounting orientation
MLCC (Class II)Capacitance lossDC bias derating + aging (t^(-α) capacitance decay from BaTiO₃ crystallisation)Voltage derating to 30-50% of Vrated; C0G for precision
Al electrolyticWear-outElectrolyte dry-out; oxide film degradation with time and temperatureTemperature derating; hybrid polymer or MLCC replacement in hot environments
Al electrolyticCatastrophic ventReverse polarity, over-voltage → gas generation → pressure release valve activatesCorrect polarity in assembly; overvoltage protection
FilmProgressive capacitance lossCumulative self-healing events consuming metallisationVoltage derating; oversize; monitor capacitance drift as diagnostic
EDLCCapacitance fade + ESR riseElectrolyte decomposition at cell voltage limit; high-T acceleratesVoltage derating (2.5V typical for 2.7V-rated); active balancing across cells in series strings
HSC / LICAnode Li loss + capacitance fadeRepeated deep cycles slowly delaminate Li from anode graphiteShallow cycling in service; charge balancing; thermal management (~40°C sweet spot)





The rule of thumb on derating. Every capacitor category shows dramatic reliability improvement when derated to ~50% of nameplate voltage and temperature. Hyperscale power-shelf design typically specifies parts at 40-60% derating, which is why the sellable-capacity number in the market essay is much smaller than the nameplate-capacity number in the same market's press releases.

11What AI actually demands



The transient profile that redesigned the shelf


A modern AI accelerator does not draw constant power. It draws a spiky load with characteristics no data-centre transformer or PSU was ever designed for. Three properties matter.




GPU load profile during synchronous training · a single iteration
Indicative rack power vs time during a Vera Rubin training step, 200ms window

Compute burst → gradient sync → compute burst pattern. Peak-to-average ratio ~2×. Cycle rate ~5-10 Hz sustained across a training run. Micro-transient rise/fall on the order of milliseconds. Nothing in this profile is compatible with legacy AC-fed racks.




First, the peak-to-average ratio. An AI training rack running steady-state at 300kW may spike to 600kW during compute bursts and idle at 150kW during gradient sync. That factor-of-2 swing repeats several times a second, indefinitely, for the duration of the training run (which can be weeks). No traditional capacitor design was sized for a 5-10 Hz macro-cycle at rack scale.


Second, the micro-transient dv/dt. The rise time of an individual compute burst is on the order of milliseconds: much faster than a mechanical breaker can respond, much slower than a MOSFET switching event. That millisecond window is exactly where HSCs and large polymer capacitors are optimised, and it is the timescale the 800VDC rack architecture is designed around.


Third, the cycle count. A one-year training run at 5 Hz continuous cycling exposes the storage device to ~150 million cyclesLi-ion chemistry survives roughly 3,000 cycles at 80% depth-of-discharge before capacity fades. That is why the market essay observes that "batteries cannot retake the smoothing socket". The arithmetic is not close.




Cycle life at depth-of-discharge · why HSCs win the rack socket
Rated cycle life vs depth-of-discharge for four chemistries, log axes

EDLC and HSC hold cycle life essentially independent of DoD. Li-ion chemistries fade rapidly at high DoD; LTO (lithium titanate) is the exception, holding 20,000+ cycles but at ~half the energy density of NMC. Rack-scale smoothing requires 10⁶+ cycles at partial DoD, which puts everything except EDLC/HSC out of scope.


12Roadmap



Where the technology goes next


The market essay lists seven technology vectors on the "disruption watch list". This piece walks through what each one actually changes at the device level.





2025-27 · early product
Silicon-anode HSC
Replacing graphite/hard-carbon anodes with silicon-composite anodes (Group14, Sila and other Li-ion silicon-anode vendors' spillover into the HSC space). Silicon's theoretical lithium storage capacity is ~10× graphite's, so a silicon-anode HSC could reach 40-50+ Wh/kg. The manufacturing challenge is silicon's volume expansion on lithiation (~300%), which is why Li-ion vendors have taken years to commercialise it. HSC engineers face the same problem and are further behind on the qualification curve.


2025-28 · scaling
Dry-electrode processing
Solvent-free electrode manufacturing (LICAP's IP, later Tesla's ex-Maxwell process). Standard electrode manufacturing coats a slurry of active material + solvent onto foil, then evaporates the solvent, which is expensive in both energy and floor space. Dry electrodes press the active material directly onto the foil via calendering, cutting energy consumption ~90% and enabling much thicker electrodes (higher energy density) than wet-coating tolerates. This is the main cost-down lever for closing the market essay's supply gap profitably.


2027+ · early product
Curved graphene and advanced carbons
Replacing coconut-shell activated carbon (the incumbent EDLC/HSC cathode material) with synthetic high-surface-area carbons. Skeleton Technologies uses "curved graphene", a proprietary carbon structure with higher specific capacitance than activated carbon. MOF-derived carbons (from metal-organic frameworks) offer another route. Impact: attacks the Kuraray chokepoint identified in the market essay; European strategic sovereignty motivation.


2027+ · commercial
Sodium-ion hybrid capacitors
Replacing lithium pre-doping with sodium pre-doping. Cheaper (no lithium), no strategic mineral exposure, some cost-per-Wh advantage at scale. Currently trails LIC on energy density by ~30% because Na⁺ ions are physically larger than Li⁺ and don't intercalate into graphite as easily (they need hard-carbon anodes). Chinese cell makers are the visible pursuers. Qualification lag protects incumbents in the near term.


2026+ · scaling
Ultra-high-cap MLCC and embedded MLCC
Two vectors: pushing individual MLCC capacitance to 47 µF or higher in a 0402 case (dielectric thickness ~0.3 µm, layer count >1,000), and physically embedding MLCCs inside PCB substrate rather than mounting on top. Murata, Taiyo Yuden and SEMCO lead. Impact: converts sockets currently held by Al electrolytic or polymer to ceramic, further deepening the top-3 concentration in the MLCC market.


2027-29 · flagship platforms
Backside power delivery (BSPDN) + silicon capacitors
Backside power delivery, part of the sub-2nm process transition (TSMC N2, Intel 18A), moves the power delivery network from the die front side to the back. That opens up the die front side for signal routing and makes it feasible to design capacitance into the package substrate rather than around it. Silicon capacitor content per package rises materially. This is the most structural design-in event on the roadmap and the one that most decisively rewards Murata/SEMCO/TSMC packaging over discrete MLCC.


2026+ · standards
OCP energy-shelf standardisation
Open Compute Project working groups are drafting specifications for capacitive energy shelves at rack scale, targeted for release from YE2026. Standardisation is double-edged: it accelerates hyperscaler adoption (multi-vendor sourcing becomes possible), then commoditises the shelf hardware itself (integration margin migrates toward zero), leaving cells and materials as the defensible layers. The ORv3 power-shelf commoditisation trajectory is the precedent.



Every roadmap vector operates on the same three variables: energy per unit mass, power per unit mass, or cost per Wh delivered. Each is a slightly different attack on the fundamental physics constraint the Ragone plot describes.The technology thesis in one line

13Where this lands



Physics to system, in one arc


Every capacitor category on the AI power chain exists to hold a specific timescale of transient current. The chemistry that wins each socket is determined by three physical constraints: how fast ions or electrons can move (setting the response time), how much charge can be stored per unit mass (setting the energy density), and how many times the storage can be cycled before something breaks (setting the reliability). Ceramic MLCCs sit at one extreme, batteries at the other, and the whole stack in between represents specific engineering choices against those three constraints.


What has changed with AI is not the physics but the arithmetic. The transient profile of a synchronised training rack has moved the required capacity in every window up by roughly an order of magnitude, and the number of racks needing that capacity has moved up by another order of magnitude. That is why the entire capacitor supply chain from cathode carbon to nickel paste feels simultaneously undersupplied.


For a reader coming from the market essay, three physics implications. First, the timescale ladder is not soft: MLCC will not do the millisecond job, HSC will not do the microsecond job. The stack is the stack because physics segregates the roles. Second, backside power delivery is the swing factor: it moves capacitance permanently into the package, favouring silicon capacitors and embedded MLCC over discrete parts, and it is a structural realignment of who captures the value. Third, silicon-anode HSC is the roadmap wildcard: if silicon-anode processing solves the volume-expansion problem at scale, HSC energy density approaches Li-ion territory while retaining supercap cycle life, and the boundary between HSC socket and Li-ion socket moves sharply upward in duration.



SERIESThe AI Power Chain



What connects to what


This is a technical companion to The Capacitor Stack (Part I of the AI Power Chain), which covers the market structure, vendors, TAM and M&A of the capacitor layer. Part II covers The Wide-Bandgap Stack, the SiC and GaN power semiconductors that switch the currents these capacitors buffer; its technical companion is The Wide-Bandgap Stack: Technical CompanionPart III covers The Thermal Stack, the cooling infrastructure needed to reject the heat these devices produce; its technical companion is The Thermal Stack: Technical CompanionPart IV (The Interconnect Stack) covers the interconnect layer: transformers, switchgear and the grid equipment that gates the buildout.




Series. The AI Power Chain (six parts). Part I: The Capacitor StackPart I-A: Technical CompanionPart II: The Wide-Bandgap StackPart II-A: Technical CompanionPart III: The Thermal StackPart III-A: Technical CompanionPart IV: The Interconnect StackPart IV-A: Technical CompanionPart V: The On-Package Delivery StackPart V-A: Technical CompanionPart VI: The Modular Datacenter StackPart VI-A: Technical Companion.


Methodology. Technical primer, not investment advice or engineering guidance. Device parameters, process descriptions and roadmap items are compiled from standard capacitor and electrochemistry references (Kotz & Carlen on supercapacitors, Trasatti on double-layer theory, Ho on MLCC dielectrics), vendor technical literature (Murata, TDK, Panasonic, KEMET, Musashi ES, Skeleton, Kyocera-AVX), IEEE publications, and trade press. Any specific device performance number should be verified against the vendor's current datasheet before use in design work.


© 2026 Adi Kumar · Power & Digital Infrastructure · Corrections welcome.


==============================================================================
# The Interconnect Stack: Technical Companion
URL: https://adikumar.co/the-interconnect-stack-technical-companion/
Published: 2026-08-01
Summary: How transformers, switchgear, cables, gas turbines, fuel cells and SMRs actually work. Cross-sections and physics. Part IV-A of The AI Power Chain.
==============================================================================

The AI Power Chain series · Part 2 of 15











ADI KUMAR · POWER & DIGITAL INFRASTRUCTUREAUGUST 2026 · V1 · ~40 MIN READ


The AI Power Chain · Part IV-A · Technical Companion
How grid interconnect equipment actually works



The Interconnect Stack: Technical Companion


Where the market essay named the vendors and quantified the lead times, this piece explains the physics. Why voltage steps up before it steps down, how a 400-tonne oil-filled transformer moves 300 megawatts at 99.5% efficiency, why sulphur hexafluoride became the standard insulation gas and why it is being phased out, how an aeroderivative gas turbine differs from a diesel genset, and what actually happens inside a small modular reactor. Cross-sections, waveforms and roadmaps for readers who want the machine under the market.



By Adi Kumar
Independent analysis
Views are my own




How to read this piece. It is a technical companion to The Interconnect Stack (Part IV of The AI Power Chain). The market piece named the vendors, sized the layers and mapped the lead-time crisis. This one explains the physics: how transformers, switchgear, cables, generators and reactors actually work, what determines their efficiency and lifetime, and what the roadmap items on every analyst's list physically change. No prior electrical background is assumed, but the reader should know what "voltage", "current" and "power" mean at a systems level.


1First principles · Why voltage matters



The reason the whole grid exists at high voltage


Every high-voltage transmission line, every 400-tonne substation transformer, every mile of medium-voltage cable underground exists to solve one equation. Power dissipated in a conductor is P = I²·R: the square of the current, times the conductor's resistance. Power transferred to a load is P = V·I: voltage times current. To move a given amount of power (V·I) with the minimum loss (I²·R), you want the current low and the voltage high. Every voltage step-up in the grid is a specific instance of that arithmetic.




Transmission loss falls quadratically with voltage
I²R losses for 100 MW transferred over 100 km of standard ACSR conductor, log scale

Same power (100 MW), same conductor (Aluminium Conductor Steel-Reinforced, ~0.06 Ω/km at typical cross-section → ~6 Ω per phase over 100 km), three-phase AC, different line-to-line voltages. At 480 V and 13.8 kV the required current is not physically deliverable at this scale. The loss calculation exceeds the transferred power. From 33 kV upward the transfer is possible but only economical at higher classes: 33 kV loses ~55%, 138 kV ~3%, 400 kV ~0.4%, 765 kV ~0.1%. Losses fall roughly as the square of the voltage ratio.






Consider the numbers behind the substation feeding a modern hyperscale campus. A 500 MW load at 480 V three-phase would draw ~600,000 amps. That is mechanically impossible: no cable geometry can carry it, no bus bar can dissipate the resistive heat, no switch can interrupt it under fault. At 138 kV, the same 500 MW draws ~2,100 amps. That is well within the capability of a modest overhead line or an underground cable. Every step in the grid's voltage hierarchy is dictated by the same trade-off, and every transformer in the interconnect stack exists to move between those levels.




Voltage classWhere it livesPhysical formApprox current for 500 MW

765 kV / 500 kVLong-distance transmissionOverhead lines with towers 30-50m tall, minimum conductor separation ~15m~650 A
230 kV / 138 kVRegional transmission, hyperscale substation feedOverhead or shielded underground cable~1,250-2,100 A
34.5 kV / 13.8 kVData centre campus distributionXLPE underground cable in ducts~8,400-21,000 A (split across multiple feeders)
480 V / 400 VRack / room level distributionBusway and copper busbar~600,000-720,000 A (spread across many racks)
800 V DCRack inlet, next-gen AI data centresDC busbar, ~750 A per 600 kW rack~625,000 A (spread across racks)





Why DC re-enters the picture at 800V. The grid is AC because Tesla won the "war of the currents" in the 1890s: AC could be stepped up and down easily with transformers, DC could not. Modern power electronics changed that. Solid-state converters can step DC voltages efficiently, which is why long-distance HVDC transmission became viable in the 1970s and why 800VDC data centre distribution is happening now. The choice of DC at 800V for AI racks reflects the same I²R arithmetic: at 800V the copper losses in a rack busbar are half what they would be at 480VAC (which averages to ~340V DC-equivalent for power delivery). Every watt saved is a watt back for compute.

2Transformer physics






In brief


Grid interconnect equipment engineering covers transformer physics (Faraday induction, insulation thermal management), switchgear topology (SF6 vs vacuum vs solid dielectric) and cable ampacity (conductor cross-section, insulation temperature rating). A 300MW substation transformer weighs 400 tonnes and takes 24 months to fabricate. Physics + supply chain lead time together determine hyperscaler campus timelines.








Faraday's law, 200 tonnes of steel, and 99.5% efficiency


The transformer is a device that transfers electrical power between two circuits at different voltages, using only a shared magnetic field. There are no moving parts, no direct electrical connection between the two circuits, and (in a well-designed unit) less than 1% of the power is lost as heat. The physics behind this is Faraday's law of electromagnetic induction, which says that a changing magnetic flux through a wire coil induces a voltage in that coil, proportional to the number of turns.




Large power transformer · simplified cross-section
Three-phase core-type transformer, roughly what a 100 MVA substation unit looks like inside its tank

Steel tank + bushings + radiators (external)
Mineral oil (insulation + cooling)
HV winding · thousands of turns of thin copper wire
Kraft paper insulation between windings
LV winding · fewer turns of much thicker copper
Kraft paper + oil ducts
Laminated silicon-steel core (GOES)0.23-0.30mm laminations, grain-oriented
Kraft paper + oil ducts
LV winding (opposite phase)
HV winding (opposite phase)
Mineral oil
Tank + oil-conservator + Buchholz relay

Voltage ratio equals turns ratio: an HV/LV winding pair with 1,000/100 turns transforms 100 kV to 10 kV, or 10 kV to 100 kV, depending on which side is the primary. A 100 MVA transformer at 100 kV draws 1,000 A on the HV side and delivers 10,000 A on the LV side. The whole unit weighs 100-200 tonnes for that class; larger transmission units run 200-400+ tonnes.





The two loss mechanisms and where they sit


A transformer's ~1% loss (99% efficiency) has two components. Copper losses (I²R in the windings) scale with the square of load current and dominate at high load. Iron losses (in the core: hysteresis and eddy currents) are approximately constant with load and dominate at low load. Total efficiency peaks around 40-60% of rated load, where the two loss mechanisms cross.




Transformer efficiency vs load · why data centre load profile matters
Typical 100 MVA power transformer, efficiency across load range

Core losses (~0.3% of rated) are constant. Copper losses (~0.7% at full load) scale as load². Total efficiency drops at both extremes: below 20% load, iron losses dominate; above 80% load, copper losses take over. Data centre load profile sits mostly in the 60-90% range, close to the peak efficiency point.





Iron losses: hysteresis and eddy currents


The core carries the magnetic flux linking the primary and secondary windings. Every AC cycle, that flux reverses direction (100 or 120 times per second, depending on grid frequency) and the core material has to reverse its magnetic polarisation with it. This costs energy in two ways.


Hysteresis is the energy needed to reorient the magnetic domains in the core material. Iron is a ferromagnetic material with a preferred magnetisation direction; reversing that direction costs work proportional to the area of the material's B-H hysteresis loop. Silicon-steel alloys have much narrower loops than pure iron, cutting hysteresis loss by an order of magnitude.


Eddy currents are the currents that would flow inside the core itself if the core were a solid conductor. AC flux induces voltage in the core just as it induces voltage in the windings; if the core has a conductive path, current flows and dissipates as heat. The fix is to build the core from thin insulated laminations, forcing eddy currents into small loops with high resistance. Modern GOES laminations run 0.23-0.30 mm thick with a phosphate-based insulating coating between them.




Transformer loss waterfall · where the 1% goes
Loss breakdown for a typical 100 MVA transformer at rated load, kW

Approximately 900 kW total loss on a 100 MW load (~0.9% loss, ~99.1% efficiency). Winding I²R dominates at rated load. Core losses split between hysteresis and eddy. Stray losses (in the tank, clamps, structural steel) are a smaller residual. Auxiliary losses cover cooling fans and pumps. Numbers are indicative; specific units vary.


3GOES manufacturing



How grain orientation is engineered


Grain-oriented electrical steel is the material that makes low-loss transformer cores possible. What makes it "oriented" is a specific crystallographic texture: the iron grains in the steel are aligned so that the crystals' preferred magnetisation axis (the [001] direction) points along the rolling direction of the finished sheet. When the transformer designer positions the laminations so that this direction aligns with the flux path, hysteresis and eddy losses drop by roughly an order of magnitude. Achieving this texture is a manufacturing process that took the mid-20th century's steel industry decades to perfect.




GOES manufacturing sequence · from slab to coated sheet
Simplified process flow, six major stages, elapsed time weeks per coil

Hot rolling → normalising → cold rolling to intermediate thickness → decarburisation anneal → cold rolling to final thickness → high-temperature "box" anneal in hydrogen (grain orientation develops here) → insulating coating → cutting to strip. Total process time per coil is 3-5 weeks, which is why GOES production cannot be scaled quickly.




Two chemistry choices determine performance. Silicon content (2.5-3.5% by weight in GOES) raises electrical resistivity of the iron, reducing eddy losses at the cost of making the steel harder and more brittle. Impurity control (particularly of manganese sulphide inclusions and grain-inhibiting compounds) governs how cleanly the goss texture develops during the final anneal. Nippon Steel's domain-refined grades use additional laser or plasma scribing after coating to further reduce loss by breaking magnetic domain walls into smaller units.



The amorphous alternative. Amorphous-metal ribbon (originally developed by Allied Signal, now Hitachi Metals / Proterial's Metglas product; comparable products from Chinese producers) has no crystalline structure at all. It is essentially frozen liquid metal, produced by casting molten alloy onto a chilled spinning wheel at meters-per-second, quenching so fast that no crystals form. Amorphous cores have 70-80% lower no-load loss than GOES, but the ribbon is thin (~25 µm), fragile, and harder to work into large-cross-section cores. Commercial in distribution transformers since the 1980s; still limited at LPT scale by handling. If GOES supply tightens further, amorphous is the material to watch.

4Switchgear physics



Interrupting current is harder than carrying it


Switchgear does a job that sounds simple until you consider the physics. Carrying steady-state current is trivial: any conductor sized for the ampacity will do. Interrupting current under fault conditions is not. When a circuit breaker opens under load, the current does not stop just because the metal contacts separate. It maintains itself through an electric arc: a plasma column of ionised gas that continues to conduct until the AC current naturally passes through zero (twice per cycle at 50 or 60 Hz). During that arc, the temperature at the contact reaches 5,000-20,000°C. The switchgear's job is to quench the arc as it approaches zero-crossing so it does not re-strike as the voltage recovers.




Arc-quenching mechanism comparison
How four switchgear types extinguish the fault arc

Air-blast: high-velocity compressed air blown across arc. Oil: arc immersed in mineral oil, decomposition products (hydrogen) blow arc apart. SF6 (GIS): sulphur hexafluoride captures free electrons and rapidly recovers dielectric strength after zero-crossing. Vacuum: no gas present, so arc cannot sustain in the vapour of eroded contact metal for more than a half-cycle. Dry-air (clean air): puffer or self-blast interrupter using compressed synthetic air as the arc-quenching medium.





Why SF6 has been the standard, and why it's exiting


Sulphur hexafluoride is an unusually good arc-quenching gas for three physical reasons. First, its dielectric strength is roughly 3× that of air at the same pressure, so a given breaker can be much smaller. Second, SF6 molecules are electron-attaching: they capture free electrons from the plasma, disrupting the arc's conductivity as soon as current tries to pass through zero. Third, SF6 has a very fast dielectric recovery: after the arc extinguishes, the gas returns to insulating state within microseconds, preventing the voltage from restriking the arc.


The problem is atmospheric. SF6 has a global warming potential ~23,500× that of CO₂ over a 100-year horizon, and an atmospheric lifetime of ~3,200 years. Any SF6 that escapes from switchgear (through seals, during maintenance, at end of life) accumulates. The utility industry emits an estimated ~9,000 tonnes of SF6 per year globally, most of it from switchgear leaks. The EU's revised F-Gas Regulation (2024) prohibits SF6 in new MV switchgear from 2028-2030 and in HV from later dates; US EPA and California CARB rules are converging. Every switchgear vendor has now developed an alternative gas or a vacuum-based alternative.




SF6 gas-insulated switchgear (GIS) · simplified breaker cell
A compact metal enclosure filled with SF6 at 4-7 bar pressure. Same interrupting capacity as a much larger air-insulated breaker.

Grounded steel enclosure (earthed)
SF6 gas · 4-7 bar
Fixed contact (busbar side)
Arc-quenching chamber · nozzle geometry
Moving contact (load side)
Puffer mechanism (compresses gas during opening)
Insulator spacer (epoxy resin)
Operating mechanism (spring-loaded)

During opening, the puffer mechanism compresses SF6 and blasts it across the separating contacts, cooling the arc. At current zero, the compressed SF6 sweeps ionised gas out of the interrupter chamber, letting the dielectric recover within a fraction of a millisecond. Complete interruption typically takes 2-3 cycles (33-50 ms at 60 Hz).





The alternative technologies


Three approaches replace SF6 in different voltage classes.


Vacuum interrupters dominate the MV range (up to ~40 kV). No gas is present, so the arc can only sustain itself in the vapour of eroded contact metal. That vapour dissipates within a half-cycle, giving very fast interruption. Vacuum has effectively replaced SF6 in most new MV switchgear over the past decade. The physics does not scale well above ~72 kV because the necessary contact gap grows faster than the interrupter's dielectric strength.


Clean air / dry air (synthetic air, N₂ + O₂, dried and filtered at elevated pressure) is the natural replacement for SF6 in HV switchgear where vacuum does not scale. Siemens' Blue GIS and Hitachi Energy's EconiQ (using C4-fluoronitrile in some products, clean air in others) are the visible commercial products. Dry air has lower dielectric strength than SF6, so equipment is physically larger; but no greenhouse gas concern.


Fluoronitrile and fluoroketone mixtures (3M's Novec 4710, GE's g3) offer intermediate dielectric strength between SF6 and clean air, with much lower GWP than SF6 (still in the hundreds, not zero). Trade-off: some of these compounds degrade in the atmosphere into perfluoroalkyl acids (PFAS), which face their own tightening regulatory perimeter. Fluoronitriles are a real transitional technology, not necessarily an endpoint.



5Cable physics



The insulation matters more than the conductor


An underground MV cable has three functional layers: a conductor (copper or aluminium), an insulation system (XLPE plus semi-conducting shields), and mechanical protection (armour, sheath, outer jacket). The conductor sizing follows a straightforward ampacity calculation. The insulation system is where the engineering discipline lives.




Medium-voltage XLPE cable · concentric construction
Cross-section through a single-core 33 kV cable, roughly 60 mm outer diameter

Outer jacket (HDPE)
Copper wire screen / armour
Insulation screen (semi-conducting layer)
XLPE insulationCross-linked polyethylene · ~8mm thick at 33 kV
Conductor screen (semi-conducting layer)
Conductor · stranded copper (or Al)150-800 mm² typical for MV data centre feeds

The two semi-conducting layers eliminate microscopic air gaps between the conductor/insulation and insulation/screen boundaries. Any gap would cause "partial discharge" (miniature arcs inside the insulation) that would erode the XLPE from the inside over years of service. The semi-cons ensure the electric field distributes uniformly through the insulation with no localised concentrations.





Why XLPE won


Polyethylene by itself is a thermoplastic: it melts at ~120°C. That would be unacceptable in a power cable that might run continuously at 90°C conductor temperature. Cross-linking the polyethylene converts it to a thermoset: the polymer chains form covalent bonds during a curing step, so the material no longer melts, just softens gradually up to ~250°C. Cross-linking is done by peroxide chemistry or (for the highest-performance grades) by electron-beam radiation. XLPE holds continuous operating temperatures up to 90°C and short-circuit temperatures to 250°C, roughly 20°C higher than the older PVC or EPR insulations at each threshold.


The insulation's job is to withstand the voltage stress without breakdown. Dielectric breakdown of XLPE happens at fields of ~20 kV/mm in new material and degrades over years due to space-charge accumulation, water tree formation, and partial discharge at any defects. Cable design typically uses ~40% of breakdown field as the working stress, giving decades of service life. Water ingress is the main long-term degradation mechanism; modern cables include water-blocking tapes and metallic sheaths for underground use.




Cable ampacity depends on more than the conductor
Continuous current rating for 400 mm² Cu XLPE MV cable, by installation condition

The same cable can carry very different currents depending on how it is installed. Air (best cooling) gives the highest ampacity. Duct (poor cooling, thermal soil resistivity effect) gives the lowest. Direct burial in low-resistivity soil sits in the middle. IEC 60287 is the calculation standard.





Copper vs aluminium at the conductor


Copper and aluminium are the two commercial conductor metals. Their physical properties trade against each other in specific ways.





Copper


Resistivity: 1.7 × 10⁻⁸ Ω·m
Density: 8,960 kg/m³
Ampacity per mm²: higher (better conductor)
Cost per tonne: ~$10,000 (mid-2026 LME)
Termination: straightforward; mature techniques
Where it wins: where cross-section is limited (racks, busbars, transformer windings), where mechanical flexibility matters (moving parts, dense connections), where termination reliability is critical.




Aluminium


Resistivity: 2.8 × 10⁻⁸ Ω·m (65% higher)
Density: 2,700 kg/m³ (30% of Cu)
Ampacity per mm²: lower
Cost per tonne: ~$2,500 (~1/4 of Cu at mid-2026)
Termination: requires care because Al oxidises and cold-flows under compression
Where it wins: where mass or cost dominates (overhead lines, large-cross-section MV cables in bulk runs), where cross-section is not the constraint.





The "same ampacity, larger cross-section" trade. An aluminium conductor with 65% higher resistivity needs roughly 1.65× the cross-sectional area to carry the same current at the same temperature. That is often acceptable in bulk MV runs where the trench is already sized for the cable and the cost of copper dwarfs the cost of the incremental duct volume. It is not acceptable in racks or transformer windings where physical volume is the constraint. Data centre cabling uses aluminium at the MV distribution level and copper everywhere denser than that on exactly that trade-off.

6Backup generation physics



Diesel, gas, and the thermodynamics that separates them


A backup generator converts chemical energy in fuel to mechanical energy in a rotating shaft, then to electrical energy through an alternator. Two engine classes dominate: reciprocating engines (diesel and natural-gas piston engines) and gas turbines. Each occupies a different point on the size, efficiency, and response-time trade surface.




Backup generation on the efficiency vs response-time plane
Typical efficiency (LHV basis) vs cold-start-to-full-load time, bubble size = unit power range

Diesel reciprocating engines: highest fast-start capability, moderate efficiency (~40%). Gas reciprocating engines: comparable start-time, similar efficiency, cleaner emissions. Aeroderivative gas turbines: faster than industrial gas turbines, medium efficiency (~40% simple cycle, 55-60% combined cycle). Industrial gas turbines (H-class): highest per-unit power, best combined-cycle efficiency, slowest to start. Fuel cells: highest efficiency at scale (~60%+), start-time varies by type.





The reciprocating engine


A diesel or natural-gas piston engine burns fuel in cylinders arranged around a crankshaft. Diesel engines compress air to high pressure (~40:1 compression ratio), inject fuel into the compressed air, and rely on the compression heat to auto-ignite the mixture. Natural-gas engines use a spark plug (like an automotive petrol engine) at lower compression ratio (~10:1). Both convert chemical energy to shaft work at ~40% efficiency (LHV, lower heating value basis), with the remainder rejected as exhaust heat and jacket-water heat.


The advantage of reciprocating engines is start time. A modern diesel genset can go from cold-stop to full load in under 10 seconds, which is why they are the default choice for standby duty (the utility fails, the genset picks up, the transfer happens with a fraction of a second's UPS bridge). Natural-gas engines are similar but need a live gas pipeline; they cannot store fuel on-site the way a diesel does.




The gas turbine


A gas turbine has no cylinders. Compressed air continuously enters a combustion chamber, fuel is continuously injected and burned, and the hot expanding gas drives a turbine wheel that is coupled to a generator. The whole cycle is continuous rather than reciprocating, which lets gas turbines run at much higher speeds (~3,000-15,000 rpm depending on design) and pack much more power into a given footprint.


Two families matter for data centre applications. Aeroderivative turbines (GE's LM series, Siemens' SGT-A family, Rolls-Royce Trent variants) are derivatives of aircraft engines: light, fast-starting (2-10 minutes cold-start), medium efficiency (~40% simple cycle, 55-60% combined-cycle when a steam bottoming cycle is added). Heavy industrial turbines (GE's 7HA, 9HA; Siemens' SGT-8000H; Mitsubishi's M501J/M701J) are much larger, more efficient (~44% simple, 63%+ combined), and slower to start. Aeroderivative machines are the visible choice for behind-the-meter data centre power because start-time and modularity matter more than absolute peak efficiency.




Aeroderivative gas turbine · simplified flow path
Air enters left, compressed, mixed with fuel and burned, drives turbine, exits right

Air intake · with filtration
Multi-stage axial compressor (14-17 stages)
Combustion chamber(s)Annular or can-annular, natural gas or dual-fuel
HP turbine · drives compressor
Power turbine · drives generator
Exhaust · ~500-650°C
Optional HRSG for combined cycle · exhaust drives steam turbine

Simple-cycle efficiency 35-42% depending on class and ambient conditions. Combined-cycle with a heat recovery steam generator (HRSG) and bottoming steam turbine: 55-63%. The steam bottoming cycle requires cooling water and about a year longer construction. Simple-cycle is faster to deploy but leaves more energy in the exhaust.


7Fuel cell physics



Electrochemistry replaces combustion


A fuel cell converts chemical energy in a fuel directly to electricity through an electrochemical reaction, without burning it. The physics is essentially a battery reaction that keeps running as long as fuel and oxidant are supplied: hydrogen (or a hydrocarbon that gets internally reformed to hydrogen) is oxidised at the anode, oxygen is reduced at the cathode, and the electron flow between them delivers current to an external circuit. Efficiency is set by the thermodynamics of the reaction itself, not by a Carnot cycle, which is why fuel cells routinely reach 55-60% electrical efficiency where combustion technologies max out around 40% at similar scale.




Fuel cell types · operating temperature and typical application
Four commercial fuel cell technologies, positioned by operating temperature and typical stack size

PEM (proton exchange membrane): ~80°C, hydrogen only, mobile applications. AFC (alkaline): historical, spacecraft. PAFC (phosphoric acid): ~200°C, first commercial stationary; largely superseded. MCFC (molten carbonate): ~650°C, hydrocarbon-tolerant, stationary. SOFC (solid oxide): ~700-1,000°C, high efficiency, hydrocarbon-tolerant, Bloom Energy's technology.





Solid oxide fuel cells: what Bloom Energy actually sells


The most visible fuel cell technology in data centre applications is the solid oxide fuel cell (SOFC), commercialised at scale by Bloom Energy. SOFCs use a ceramic electrolyte (yttria-stabilised zirconia) that becomes oxygen-ion-conductive at high temperature. Oxygen ions migrate from cathode to anode through the electrolyte, where they combine with hydrogen (or methane reformed on the fly) to produce water, releasing electrons that flow through the external circuit.


SOFCs have three properties that suit them for stationary primary power. First, high electrical efficiency: 55-60% at the cell stack, higher when waste heat is captured for cogeneration. Second, fuel flexibility: SOFCs can run on natural gas, biogas, or hydrogen, reforming the fuel internally at operating temperature. Third, low emissions: no NOₓ because there is no combustion, only oxidation at the anode. The trade-offs are start-up time (hours from cold, so SOFCs are always run continuously in service) and stack lifetime (~5-7 years before replacement).




SOFC cell · simplified layer stack
Ceramic sandwich, one cell of thousands in a stack

Interconnect / current collector (steel)
Air channel
Cathode (LSM or LSCF perovskite)
Electrolyte · yttria-stabilised zirconia~10 µm thick, O²⁻ conductor at 700-1000°C
Anode (Ni-YSZ cermet)
Fuel channel · natural gas + H₂O or H₂
Interconnect (next cell begins)

A single cell produces ~0.7-1.0V at rated current. Cells are stacked in series to build voltage, and stacks are combined into modules. A typical Bloom "Energy Server" module produces ~250 kW from thousands of cells operating in parallel and series. Modules combine into data-centre-scale installations of multiple megawatts.


8Small modular reactors



Fission physics, but modular


A nuclear reactor generates heat by controlled fission of a heavy nucleus (usually uranium-235) into lighter fragments, releasing neutrons that sustain the chain reaction and kinetic energy that heats a working fluid. The heat drives a steam or gas turbine that in turn drives a generator. The physics has been well-understood since the 1940s. What is new about small modular reactors (SMRs) is not the fission; it is the manufacturing model.


Traditional light-water reactors (LWRs) are built on-site at ~1,000+ MWe scale over 8-12 years. SMRs (~50-300 MWe per unit) are manufactured in a factory and shipped to site, cutting construction time to 3-5 years and letting operators buy the exact capacity they need in modular increments. The economic logic is the trade-off between factory manufacturing efficiency and the loss of per-unit scale that made large reactors economical in the first place.




SMR technologies by reactor type · a four-way split
Commercial SMR programmes, by coolant/moderator physics, thermal output range and NRC status

LWR-derivative (NuScale, GE Hitachi BWRX-300): mature physics, standard fuel, standard supply chain. HTGR (X-Energy Xe-100): TRISO fuel, helium coolant, higher outlet temp for process heat applications. FHR (Kairos): fluoride salt coolant, TRISO fuel, high temp, passive safety. Fast/microreactor (Oklo Aurora): metal-fuel, sodium or heat-pipe cooled, small (~15 MWe). Each represents a different bet on which physics tradeoff wins commercially.





The four SMR families





LWR-derivative


Modular versions of the standard pressurised or boiling water reactor. Uses standard low-enriched uranium fuel, standard supply chain, mature regulatory pathway. Passive safety features (no active pumps required during accidents) replace the active systems of legacy plants.


Examples: NuScale VOYGR (77 MWe/module, 924 MWe max in 12-module plant); GE Hitachi BWRX-300 (300 MWe); Rolls-Royce SMR (470 MWe); Holtec SMR-300; Westinghouse AP300.


Status 2026: BWRX-300 leads on commercial deployment (OPG Ontario site 2028-30 targeted). NuScale de-emphasised after UAMPS project restructuring.





High-temperature gas-cooled (HTGR)


Helium coolant, graphite moderator, TRISO fuel particles (uranium kernels in ceramic coatings). Very high outlet temperature (~750°C) enables applications beyond electricity: hydrogen production, industrial process heat.


Example: X-Energy Xe-100 (80 MWe / 200 MWt per unit, four-unit standard plant).


Status 2026: Amazon investment (Sept 2024) and Washington state site announced. First commercial units targeted late 2020s.







Molten salt (FHR / MSR)


Fluoride or chloride salt coolant, sometimes as fuel carrier too. Very high boiling point at atmospheric pressure eliminates high-pressure containment. Passive safety features are compelling: fuel drains to a subcritical geometry if temperature exceeds limits.


Example: Kairos Power Hermes (35 MWt demo). Google MOU (Oct 2024) for commercial Hermes-derived unit(s).


Status 2026: Demonstration reactor under construction. Commercial timeline early-to-mid 2030s.





Microreactor (fast / heat-pipe)


Very small units (~5-20 MWe) using metal fuel, sometimes fast-spectrum, sometimes heat-pipe cooled. Targeted at remote sites, military bases, and single-facility applications. Different regulatory pathway (small enough that traditional LWR regulations do not apply cleanly).


Example: Oklo Aurora (~15 MWe, sodium-cooled fast reactor). Various DOD demonstration projects.


Status 2026: NRC combined license application pending. Data-centre-relevant timeline early 2030s.





The regulatory path is the timeline. Every SMR programme's commercial-operating date depends on the US NRC combined license process, which typically runs 4-6 years from application to authorisation. The NRC has been reworking its licensing framework for non-LWR designs, and the first commercial SMR authorisations from that framework are expected 2027-29. Physical construction of a factory-manufactured module can be completed in 2-3 years, but regulatory authorisation drives the timeline until the process matures. Announced hyperscale SMR deals target 2029-32 first power dates for that reason, rather than sooner.

9Grid interconnect



How the grid actually connects a new load


The final piece of the interconnect stack is the process by which a new large load (a hyperscale data centre) gets connected to the grid. This is where the market essay's four-to-seven-year timelines come from. The process is worth understanding step by step, because the constraints are not what most technology observers assume.




The connection request


A new load submits an interconnection request to the local utility or transmission system operator (in US terms, the ISO/RTO). The request specifies the requested capacity (in MW), the proposed connection voltage, and the geographic location. The ISO adds the request to its interconnection queue and begins studies.




The system impact study


The ISO models how the requested load would affect the grid: whether existing transmission lines have enough capacity, whether voltage will stay within limits under the new load, whether short-circuit currents remain within switchgear ratings, whether stability margins hold under contingencies. The study identifies grid upgrades needed to accommodate the load: typically transformer additions, new transmission lines, reactive power support, breaker upgrades.




The cost allocation and network agreement


Grid upgrades cost money. Different jurisdictions allocate the cost differently: some assign the requesting customer full cost of upgrades, some spread cost across all ratepayers, some use a hybrid. The negotiation between the ISO, the requesting customer, and the local utility over cost allocation is often the longest single step in the process, and the mechanism by which small requests often subsidise large ones (or the other way around, depending on rules).




Physical construction


Once the agreement is signed, physical construction proceeds. New transmission lines require permitting (state siting boards, federal environmental review for lines crossing federal land, community consent). New substations require substation equipment (all subject to the LPT and switchgear lead times the market essay quantifies). Timelines commonly run 3-5 years from agreement to energisation, on top of the 1-3 years for study and agreement.




Interconnect process timeline
Approximate elapsed time from request to energisation for a large new load, months

Illustrative. Specific timelines vary by ISO, load size, existing grid capacity in the area, and permitting complexity. LBNL's "Queued Up" reports track actual timelines and show median times of 3-5 years for generation projects; large new loads (data centres) often report similar or longer timelines. FERC reforms in 2023-24 have targeted queue processing efficiency but have not materially changed physical construction timelines.


The physical construction phase is not the binding constraint. The LPT and switchgear lead times covered in the market essay run parallel to it. The binding constraint is the ISO's ability to process, study and negotiate the request: a bureaucratic bandwidth problem that scales with the number of pending requests, not with the size of any one.The queue bottleneck thesis

10Roadmap



Where the technology goes






2026-28 · commercial scaling
SF6-free MV switchgear
Vacuum interrupters and clean-air alternatives at MV scale. Regulatory-driven timeline: EU F-Gas Regulation limits SF6 in new MV switchgear from 2028-2030. Every switchgear OEM has a product; qualification into utility standards is the pace-setter.



2026+ · commercial
Digital transformers
Sensor packages (temperature, dissolved gas analysis, load current, oil moisture) integrated with cloud analytics. Transitions LPT sales from one-time capex to an ongoing service annuity. Hitachi Energy, Siemens Energy, ABB all have platforms; adoption is faster at greenfield sites than retrofit.



2027+ · pilot
Amorphous-metal LPT cores
Amorphous-metal ribbon replaces GOES in transformer cores. Cuts no-load losses 70-80% but ribbon handling is hard at LPT scale. Commercial in distribution transformers; pilot units at LPT scale in progress. GOES scarcity is the accelerator.



2028+ · scaling
Solid-state transformers (SST)
SiC-based power electronics replace line-frequency transformer cores at MV/LV boundary. Enables bidirectional power flow, digital control, and mass reduction (~10× smaller than an equivalent LPT). Data centre pilots underway at hyperscale sites; commercial deployment gated by cost. Cross-reference Part II wide-bandgap essay for the underlying SiC-device roadmap.



2028+ · scaling
Aeroderivative BTM turbines at scale
Behind-the-meter aeroderivative gas turbine plants (100-400 MW) at hyperscale campuses. GE Vernova, Siemens Energy, Solar Turbines / Caterpillar all seeing bookings. Combined-cycle configurations for higher efficiency. Emissions and permitting are the constraints.



2029+ · early commercial
Bloom SOFC at multi-100MW scale
Solid oxide fuel cells at cumulative hundreds of MW per hyperscale campus. Higher efficiency than gas turbines, no NOₓ, quieter. Cost per kW installed is the constraint; volume manufacturing is bringing it down. Hydrogen fuel path opens post-2030 if hydrogen production scales.



2029-32 · first commercial
SMR to hyperscale customers
First commercial SMR power delivered to a hyperscale customer. BWRX-300 at OPG Ontario is the leading candidate; X-Energy Xe-100 in Washington state follows. Kairos Hermes-derived and Oklo Aurora are further out. Regulatory authorisation is the pacing factor.



2030+ · deployment
HVDC interconnectors for load transfer
High-voltage DC transmission from renewable-rich or nuclear-rich regions to data centre load centres. Enables sourcing multi-GW of power from otherwise inaccessible geographies. Hitachi Energy, Siemens Energy, GE Vernova, NKT and Prysmian own the technology; scale-up of HVDC cable manufacturing is the bottleneck.



2030+ · research
Superconducting MV/HV cables
Cryogenic cables using high-temperature superconductor materials carrying an order of magnitude more current per cross-section. Commercial pilots exist (LIPA in New York, Essen in Germany); economics do not yet close for standard data-centre applications. Watch, don't underwrite.




11Where this lands



Physics, calendar, and the constraint the AI industry underestimated


The interconnect stack is where AI's physical infrastructure runs into the constraints of the twentieth-century grid. The physics that dictates why substations look the way they do (I²R losses, transformer efficiency, arc quenching, dielectric breakdown) has not changed. What has changed is the pace at which new load wants to connect, and the manufacturing and permitting apparatus that supplies that connection.


Every technology on the roadmap above is a specific answer to that mismatch. Solid-state transformers reduce transformer volume and lead time. SF6-free switchgear addresses the regulatory perimeter. Amorphous cores relieve materials pressure. Aeroderivative turbines and SOFCs bypass grid interconnect timelines entirely. SMRs eventually offer a longer-term bypass. Every one shifts one variable in the equation the market essay described: how much power, how quickly, and through what physical apparatusNone of them changes the underlying physics of moving power from generation to load. What they change is the calendar.


For a reader coming from the market essay, three physics implications. First, the transformer's 99% efficiency is what makes centralised power delivery economic at all: no other conversion technology comes close to matching that at the required voltage steps, which is why solid-state transformers will supplement, not replace, iron-core LPTs for the foreseeable future. Second, SF6 alternatives are a real trade-off, not a free swap: dry-air GIS is physically larger, fluoronitrile mixtures carry residual PFAS exposure, vacuum does not scale above ~72 kV. Third, the SMR value proposition is not primarily about carbon; it is about calendar: firm, dispatchable, siteable power that does not queue for a grid interconnection. Whether the technology delivers on that promise is one of the biggest open questions in the AI power chain roadmap.



SERIESThe AI Power Chain



What connects to what


This is a technical companion to The Interconnect Stack (Part IV of The AI Power Chain), which covers the market structure, vendors, lead-time crisis and behind-the-meter workaround. The complete series covers the physical infrastructure of AI compute layer by layer.


Part I: The Capacitor Stack covers energy storage from MLCC to HSC. Companion: The Capacitor Stack: Technical Companion.


Part II: The Wide-Bandgap Stack covers SiC and GaN power semiconductors. Companion: The Wide-Bandgap Stack: Technical Companion.


Part III: The Thermal Stack covers data centre liquid cooling. Companion: The Thermal Stack: Technical Companion.




Series. The AI Power Chain (six parts). Part I: The Capacitor StackPart I-A: Technical CompanionPart II: The Wide-Bandgap StackPart II-A: Technical CompanionPart III: The Thermal StackPart III-A: Technical CompanionPart IV: The Interconnect StackPart IV-A: Technical CompanionPart V: The On-Package Delivery StackPart V-A: Technical CompanionPart VI: The Modular Datacenter StackPart VI-A: Technical Companion.


Methodology. Technical primer, not investment advice or engineering guidance. Device parameters, process descriptions and roadmap items are compiled from standard power-systems references (Grainger & Stevenson on power systems, IEEE and CIGRE working group publications, ANSI/IEEE C57 for transformers, IEC 62271 for switchgear, IEC 60287 for cable ampacity), vendor technical literature (Hitachi Energy, Siemens Energy, GE Vernova, Bloom Energy, X-Energy, Kairos Power), NRC and DOE public documents, and trade press. Any specific device performance number should be verified against the vendor's current datasheet before use in design work.


© 2026 Adi Kumar · Power & Digital Infrastructure · Corrections welcome.


==============================================================================
# The On-Package Delivery Stack: Technical Companion
URL: https://adikumar.co/the-on-package-delivery-stack-technical-companion/
Published: 2026-08-02
Summary: IVR magnetics, thin-film substrates, high-density capacitors and the qualification-gated architecture of on-package power delivery.
==============================================================================

The AI Power Chain series · Part 10 of 15








Technical Companion · Part V



The On-Package Delivery Stack: circuit design, physics, process


A deeper walk through IR drop budgeting, multi-phase current sharing, backside power etch flows, and the design tooling that keeps an 0.8V rail stable across 2,250 amperes and thousands of transient domains.




Companion to the The On-Package Delivery Stack essay (Part V of six). Use this piece for the circuit design, process integration and reliability arguments; use the main essay for the market map, TAM and profit-pool thesis.




Section 1: The IR drop budget, from first principles


The tightest number in AI silicon design is the transistor supply voltage tolerance. On a 3nm process, a nominal 0.75V core rail typically has a tolerance window of ±5% of nominal, or ±37.5mV. That number has to cover four independent effects: DC IR drop across the power delivery network, AC ringing on step-load transients, process variation of the transistor threshold, and thermal drift over the operating temperature range. The DC IR budget that the power delivery network is allowed to consume is therefore not ±37.5mV. It is closer to ±15-20mV after the other three effects have taken their slices.


Now do the resistance math. If an accelerator draws 1,800W at 0.8V, that is 2,250A of DC current. Holding IR drop below 15mV at 2,250A requires the total DC path resistance from the board-level VRM output to the transistor gate to be:


R_total < 15 mV / 2,250 A = 6.67 µΩ


Six and two-thirds micro-ohms across a path that spans centimetres of substrate copper, millimetres of interposer routing, tens of micrometres of on-die metal stack, and a stack of TSVs and micro-bumps in between. The individual budget components look like this on a typical modern design:




Where the 15mV IR budget goes: a representative accelerator
Per-segment IR drop, illustrative, mV at 2,250A total current, front-side vs backside delivery

Author's illustrative decomposition. Segments are: motherboard traces from VRM to package pins; package substrate + solder bumps; interposer TSVs; on-die metal stack (M0 through top-metal power grid); local via stack down to transistor. The BSPDN variant reallocates the on-die burden away from the front-side metal stack.




Two observations from that decomposition. First, no single segment dominates the budget; the design is a distributed optimisation problem, not a single-choke-point problem. Second, the segment that scales worst with current is the on-die metal stack, because its resistance is set by the transistor node's minimum metal pitch, which shrinks as node scales, driving resistance up quadratically at the smallest lines. This is the physics that pushes the industry toward backside power delivery: BSPDN moves the highest-current segments off the tightly-pitched front-side metal and onto a separately-fabricated back-side stack that can be routed with much wider, thicker metal.




Why R matters more than L on a DC rail (and vice versa on transient)


Power delivery network analysis operates in two regimes. In steady state, the network is a distributed resistor and R alone determines DC IR drop. On transient (a step-load event where the processor's current consumption changes by hundreds of amperes within a nanosecond), the network is a distributed R-L-C, and the impedance seen by the load includes the reactive terms. A modern accelerator PDN targets a flat impedance profile from DC up to several hundred megahertz, and every violation of that profile shows up as voltage ringing at the transistor.



Z_PDN(f) target: < 500 µΩ across DC → 200 MHz for a 0.8V, 2,250A rail



Meeting 500 µΩ impedance flatness from DC through 200MHz requires a hierarchical decoupling capacitor architecture. Bulk aluminium electrolytic capacitors at the VRM output handle DC through 10kHz. Ceramic capacitors on the motherboard handle 10kHz through 1MHz. Package-mounted MLCCs handle 1MHz through 30MHz. On-package deep-trench capacitors (integrated into the substrate or the interposer) handle 30MHz through 300MHz. And on-die MOS-capacitor arrays handle 300MHz and above. Each capacitor class has to be sized and placed so that its impedance region overlaps its neighbour with no resonant peak between them. Getting this wrong is the second most common cause of transistor throttling on new accelerator silicon, behind thermal throttling.



A modern accelerator PDN targets a flat impedance profile from DC up to several hundred megahertz. Every violation of that profile shows up as voltage ringing at the transistor, and every millivolt of ringing shows up as a lower attainable clock frequency.The design constraint that governs everything else








In brief


Vertical vs lateral point-of-load, on-chip capacitor physics and package tolerances shape how efficiently 48V or 800V becomes 0.8V at the GPU die. IR-drop budgets under 3mV per phase drive vertical-power adoption. Package inductance and pad density constrain converter placement. Physics + geometry + thermal budget together determine which topology wins at a given power density.








Section 2: Multi-phase current sharing and the physics of 20 interleaved buck converters


The workhorse converter that steps 48V or 12V down to 0.8V at the accelerator's core rail is a multi-phase interleaved buck converter. The topology is the same as the buck converter in any laptop charger, replicated N times and phase-offset by 360°/N. If eight phases share a 2,250A load, each phase carries roughly 280A. Each phase also switches at a different point in the switching cycle, so the ripple at the output capacitor cancels between phases, reducing the required output capacitance by roughly a factor of N.




Interleaved multi-phase output ripple
Illustrative output current ripple envelope, one phase vs eight phases interleaved, same total load

Author's illustration of interleaved multi-phase ripple cancellation. Peak-to-peak ripple falls approximately linearly with phase count for well-matched phases. Ripple cancellation is imperfect at any real phase count because component tolerances (inductor value, MOSFET R_DS(on), switching timing) prevent perfect cancellation; typical measured cancellation ratio is 60-75% of ideal.




The two hard engineering problems in multi-phase design are current sharing and transient responseCurrent sharing is the requirement that all N phases carry equal fractions of the load current in steady state, so that no single phase runs hotter than its neighbours. Because component values vary between phases (inductor inductance ±5%, MOSFET R_DS(on) ±10%), passive current balancing does not work at high phase counts. Modern controllers therefore implement active current balancing: they measure the current in each phase individually (using either an on-die sense resistor, an inductor DCR-based estimation, or a magnetic pickup on the output inductor) and adjust each phase's on-time to equalise currents. The DrMOS parts that Monolithic Power, Infineon and Renesas sell are packaged with the current-sense circuitry built into the same silicon as the power switches, which is what makes them qualify as "AI-grade" versus lower-grade multi-phase modules that rely on external sense components.


Transient response is the requirement that when the accelerator's current draw suddenly changes (a compute workload starts or stops), the converter's output voltage stays within its tolerance window. The response speed is set by the converter's control loop bandwidth, which is bounded above by the switching frequency (typically 500kHz to 2MHz for a multi-phase POL). Higher switching frequency = faster transient response = lower required output capacitance, but also = higher switching loss = lower efficiency. Modern multi-phase controllers use digital hysteretic control or constant on-time control to push transient response toward the theoretical single-switching-cycle limit, at the cost of more complex control loop design and higher susceptibility to noise coupling.




Coupled inductors: the modern optimisation


A recent design refinement is the coupled inductor multi-phase topology. Instead of each phase having its own independent inductor, a group of phases (typically 2 or 4) share a magnetically-coupled multi-winding inductor. The coupling adds two useful properties: it lets the inductor operate at a lower effective inductance during transients (fast response) while presenting a higher effective inductance in steady state (low ripple), and it reduces the total magnetic volume by roughly 30-40% versus discrete inductors of the same rating. Coupled inductor multi-phase controllers are now standard on AI accelerator VRM designs and are one of the differentiating features of the MPS MP288x family and the Infineon TDA family.




Section 3: The magnetics problem in on-package IVR


The main essay explained why on-package IVR is attractive: shorter current loop, higher switching frequency, smaller passive components, less IR drop. The reason the magnetic component is the hard part is worth spelling out.


The output inductor in a buck converter stores energy during the switch's on-time and releases it during the off-time. The size of the inductor is set by the required inductance value and the peak current it must carry without saturating the magnetic core. Inductance is proportional to the number of turns squared and to the core's permeability, so higher permeability core material = smaller inductor for the same inductance. But high-permeability materials saturate at low flux density, which limits the peak current the inductor can carry. There is an unavoidable tradeoff between inductance density and current capacity.



Energy stored = ½ × L × I² · Volume required ∝ Energy / (B_sat² / µ_r)



The design lever that makes on-package IVR work is switching at very high frequency (20-100MHz), which reduces the required inductance by the same factor and therefore reduces the required magnetic volume by the same factor. But at those frequencies, conventional soft-magnetic materials (ferrites, powdered iron) have unacceptable core losses because their magnetic domain walls cannot follow the applied field fast enough, dissipating energy as heat. On-package IVR requires thin-film magnetic materials accordingly: sputtered iron-cobalt or nickel-iron alloys deposited in layers separated by insulating films, with each layer thin enough (typically <100nm) that the magnetic domains can respond to megahertz-scale fields with acceptable loss.




Core loss vs frequency: bulk vs thin-film magnetics
Representative core loss density (mW/cm³) at ~50mT flux excursion, for common bulk cores vs thin-film IVR stacks

Author's compilation from vendor datasheets (Ferroxcube, Magnetics, TDK) and IEEE PELS conference papers on thin-film magnetics (2020-2025). Loss numbers vary substantially with material composition, lamination thickness and flux excursion; the chart shows an illustrative case at 50°C, 50mT peak-to-peak.




Thin-film magnetics are the technology moat that Empower and Ferric (now TDK) built. The specific process recipes (alloy composition, layer thickness, annealing profile, insulation material) are trade secrets protected by decade-long development programs. There is no commodity supply of thin-film IVR magnetics; a would-be entrant has to develop their own process from scratch and validate it against a design partner's requirements, a two-to-three-year exercise even with existing semiconductor process capability. The IVR market has consolidated to two vendors on exactly that basis, and Chinese entry is absent for the same reason.




Section 4: BSPDN process flow, in enough detail to matter


Backside Power Delivery Networks run as a family of process integrations that share the same top-level idea rather than as a single process (put a metal power stack on the back of the die) but differ substantially in how they get there. The two dominant flows are buried power rails (Intel PowerVia) and backside via + backside metal (TSMC SuperPower Rail). Both flows begin the same way, with a conventional front-side process producing the transistor plane and the front-side metal stack up to some intermediate layer. The divergence starts after front-side processing is complete.




Intel PowerVia flow (18A)


PowerVia's distinguishing feature is buried power rails: dedicated power interconnect routed at the M0 layer underneath the transistor plane, contacted through the transistor's source and drain via nano-TSVs that penetrate the substrate. The sequence:


 1. Front-side wafer processing produces the transistor plane with the buried power rails already in place at M0.
 2. Front-side metal stack is built up in the conventional way, but without the top-metal power distribution grid.
 3. Wafer is flipped and temporarily bonded to a carrier wafer for backside processing.
 4. Bulk silicon is thinned from ~800µm down to ~5µm using grind-then-CMP.
 5. Backside vias are etched down to the buried power rails at M0.
 6. Backside metal stack is deposited: two or three thick metal layers (typically 4-10µm each of copper) that form the backside power distribution grid.
 7. Backside solder bumps are formed, temporary carrier is released, wafer is diced.


The advantages: the buried power rails are already integrated with the transistor at M0 so the on-die power grid path is extremely short. The trade-off: the buried power rails have to be placed during front-side processing, before the transistor layout is complete, so the process integration is deeply coupled to the standard cell library design. Intel's 18A cell library was redesigned specifically to support PowerVia, and much of the ~2 year gap between announcement and volume production went into that library redesign.




TSMC SuperPower Rail flow (N2P and A16)


SuperPower Rail's distinguishing feature is backside via to source/drain contact: instead of routing power at M0 and contacting the transistor from below, TSMC routes the front-side metal as usual (including M0) and etches backside vias that contact the transistor's source and drain directly from the back. The sequence:


 1. Front-side wafer processing produces the transistor plane and the conventional M0 through top-metal front-side stack.
 2. Wafer is flipped and temporarily bonded to a carrier wafer.
 3. Bulk silicon is thinned to ~5µm.
 4. Backside vias are etched down to the transistor's source/drain contacts. This is a more challenging etch than PowerVia's because it requires stopping precisely on ~10nm-thick contact regions, but avoids the front-side library redesign.
 5. Backside metal stack is deposited (typically 2-3 layers, thick copper).
 6. Backside bumps, carrier release, dicing.


The advantages: the front-side cell library does not need to be redesigned, so it inherits the existing N2 library. The trade-off: the backside via etch is technically harder than PowerVia's, requiring novel etch chemistries and inline metrology to hit the source/drain landing pad with sub-nanometre alignment. TSMC's N2P timeline reflects the additional process development effort.




BSPDN process integration: PowerVia vs SuperPower Rail vs Samsung BSPDN
Illustrative comparison of process step count, thermal budget and integration complexity

Author's compilation from IEDM 2024/2025 papers (Intel 18A, TSMC N2/A16, Samsung SF2P), plus IMEC technical reports on BSPDN process variants. Values are illustrative rather than measured; the actual process metrics are commercially sensitive.


Yield ramp is where the money hides. The public disclosures on both PowerVia and SuperPower Rail describe the process as "yielding". They do not disclose the yield curve. Industry consensus is that BSPDN adds roughly 4-6 percentage points of yield loss versus the equivalent front-side-only process during the first six to nine months of production, and closes to ~1 point of yield loss by month twelve. On a $20,000-per-wafer leading-edge node, four points of yield loss on a 500-die wafer is roughly $80 per die of extra cost, which is the largest single reason why BSPDN-equipped accelerators are priced at a premium during their first-year sales window.




Section 5: Glass substrate process integration


Glass substrates enter datacenter packaging at the interposer-below layer, replacing organic ABF laminate. The key process step is drilling through-glass vias (TGVs) at the required aspect ratio and density. There are three commercial approaches:


Laser ablation. A picosecond or femtosecond laser drills each via individually. This is fast for low via counts but scales linearly with via count, which becomes a bottleneck at datacenter packaging densities (10,000+ TGVs per package).


Plasma etch. A photolithographic mask is patterned on the glass surface and reactive-ion etching cuts the vias in parallel. Higher throughput but more capital-intensive, and requires specialised plasma chemistries for glass.


Modified glass with photoactive additives. The glass sheet contains photosensitive dopants that, when exposed to UV, become preferentially etchable in an acid bath. This is the Corning approach, and it produces the highest via density (up to ~30,000 vias/cm²) with the lowest per-via cost, at the trade-off of being tied to Corning's proprietary glass composition.


Absolics (SKC's US subsidiary) uses a variant of the plasma etch approach adapted from Korean display-industry equipment, which is why the Absolics facility is in the US southeast (adjacent to display substrate suppliers) rather than in South Korea proper. Intel's in-house glass substrate line uses proprietary equipment developed internally over roughly a decade, with volume ramp targeted for 2027.




Through-glass via approaches
Illustrative comparison of throughput, achievable via density, and equipment capex

Author's compilation from ECTC 2024/2025 papers on glass substrate processing, Absolics/SKC investor presentations (2025), and Corning IP disclosures. Values are directional; process capabilities are competitively sensitive.





Section 6: Reliability failure modes at the transistor-scale power path


The main essay treated the on-package power stack as an engineering problem. The reliability side is worth its own section because the failure modes are novel to the AI generation of accelerators.




Electromigration at 2nm


Electromigration is the movement of atoms within a conductor under the momentum transfer of flowing electrons. At high current density, electromigration hollows out one end of a conductor and piles atoms up at the other, eventually opening the conductor. The failure rate scales exponentially with current density and with temperature. Conventional copper interconnect at 5nm nodes tolerates current densities of ~2 MA/cm² at 105°C for a 10-year lifetime; at 3nm, the tolerable density drops to ~1.2 MA/cm² because the interconnect cross-section shrinks faster than the current requirement.


The industry responses: (a) replace copper with cobalt or ruthenium at the smallest metal layers (Ru has 2-3x better electromigration resistance than Cu at the same dimensions), (b) redistribute current across more parallel paths using BSPDN, (c) implement per-domain current limiting in the microarchitecture to cap the peak current density in any single wire, and (d) accept a shorter design lifetime (5-year rather than 10-year) for the current-carrying-density-critical wires. All four are visible on Rubin-generation silicon.




EMIB and hybrid bonding stress cycles


Advanced packaging introduces multiple new interfaces where materials with different coefficients of thermal expansion meet. Every thermal cycle stresses those interfaces. The most vulnerable interfaces on a Blackwell or Rubin package are the die-to-interposer solder bumps and the interposer-to-substrate solder bumps. Historical thermal cycling tests use military-standard 55°C swings, which is not representative of AI accelerator use where the die can swing from 40°C idle to 105°C load in under a second and cycle that swing thousands of times per training run. Reliability qualification labs are updating their test protocols to include these fast thermal cycles, and initial data suggests the historical qualification standards were 2-3x optimistic on lifetime for AI-workload thermal profiles.



The reliability question that has not been answered yet. BSPDN-equipped accelerators are less than two years into field deployment on Intel 18A, less than one year on TSMC N2P. There is not yet enough field data to distinguish "backside processes are more reliable" from "backside processes are less reliable" over a 5-10 year operating window. Historical qualification tests suggest they are equivalent to front-side processes, but historical tests do not capture the full thermal-cycling profile of AI workloads. This is the largest single unknown in the reliability story of the current accelerator generation.




Section 7: PDN simulation and design tooling


Getting a modern AI accelerator's power delivery network right is a co-design problem that spans the transistor layout, the on-die metal stack, the interposer, the substrate, the multi-phase controller, the board-level power tree, and the CDU delivery upstream. No single tool can simulate the whole path at the fidelity required. The industry uses a hierarchy of tools:


Transistor-level SPICE simulation for the innermost 1-10mm of the path, capturing individual gate switching events, per-cell current draws and local IR drop.


Extracted-parasitic PDN simulation for the on-die metal stack through the top-metal power grid, using 3D field solvers (Ansys HFSS, Cadence Clarity, Synopsys HSPICE-Signoff) to extract resistance and inductance per grid segment, then AC-analysing the network in the frequency domain.


Package + board co-simulation for the substrate through the VRM output, typically in Ansys SIwave or Cadence Sigrity, with S-parameter models of the multi-phase controller and DrMOS parts stitched in.


System-level VRM control loop simulation for the VRM through the delivery point, in PSIM or Simplis, with the load modelled as a step-current source and the entire regulation loop closed.


No commercial tool exists that stitches all four levels together into a single time-domain simulation. The industry standard workflow is to run each tool independently, exchange S-parameters and impedance profiles at the tool boundaries, and iterate until a consistent solution emerges. This iteration is where roughly 20-30% of the pre-silicon design schedule of a modern AI accelerator is spent, and it is the largest single reason why AI accelerator design cycles have lengthened from 18 months to 30 months over the last three generations. Any tooling vendor that can shorten this loop has an enormous market to sell into.




Section 8: What a design review of a new accelerator's power delivery would ask


To close, here is the checklist that a chief silicon architect might apply when a new hyperscaler AI accelerator design lands on their desk. It is not exhaustive but it captures the questions that separate a competent design from a competitive one.



LayerThe questionWhat "good" looks like on Rubin-generation silicon
Transistor supplyNominal V_core and IR budget?0.65-0.85V nominal, ±15mV DC IR budget, ±25mV total tolerance window
On-die PDNPeak local current density?< 1.2 MA/cm² on the highest-density metal layers using Cu; < 2.5 MA/cm² on Ru layers
BSPDN?Front-side only, or BSPDN?BSPDN on any accelerator above ~1,500W total power for the leading-edge tier
InterposerCoWoS variant and TSV density?CoWoS-L or -S, ≥10,000 total TSVs, ≥30% dedicated to power
SubstrateOrganic ABF or glass?Glass on flagship tier from 2027; ABF elsewhere
IVRMerchant IVR, foundry IVR, or none?Merchant IVR from Empower or TDK on flagship tier; foundry-integrated on 2028+ designs
Multi-phase POLPhase count, coupled inductors?≥12 phases per rail; coupled inductors on flagship tier; MPS or Infineon reference part
Decoupling hierarchyPDN impedance flatness?< 500 µΩ from DC through 200MHz; hierarchical bulk/ceramic/on-package/on-die caps
Thermal couplingJunction temperature at max power?< 105°C junction on cold plate; matched to Part III cold plate microchannel spec
Transient responseVoltage droop on step load?< 8% droop on 500A/ns step; recovery within 3 switching cycles




A modern AI accelerator's power delivery network is not one design decision but ten stacked ones, from transistor cell library through backside process choice through interposer variant through IVR selection through multi-phase phase count. Each layer's decision constrains the layer above and below. Miss any of them and the accelerator either does not clock at spec or throttles under sustained load.The design review, in one paragraph

Related topic hubs
 * AI Power Semiconductors
 * 800V Data Centre Power
 * Browse all topic hubs →



Annex · References


Intel PowerVia disclosures: IEDM 2023 paper on 18A process integration; Intel Foundry Direct Connect 2024 investor day materials.


TSMC SuperPower Rail: IEDM 2024 papers on N2 backside power delivery integration; TSMC 2025 technology symposium.


Samsung Foundry BSPDN roadmap: Samsung Foundry Forum 2024 disclosures; IEDM 2025 papers on SF2P.


imec technical publications on BSPDN process variants (2022-2025).


Ferric Semiconductor (now TDK) thin-film IVR magnetics: IEEE APEC 2023-2024 papers.


Empower Semiconductor product briefs, 2024-2025 releases.


Corning glass substrate technical disclosures: ECTC 2024, IPC APEX 2025.


Absolics (SKC) investor commentary on Covington GA facility ramp, 2025-Q3 and Q4.


MPS multi-phase controller architecture: MP288x family datasheets and evaluation guide.


Ansys, Cadence and Synopsys product documentation on PDN sign-off flows.



Companion to main essay


Read alongside the main essay "The On-Package Delivery Stack" (Part V of six) for the market map, TAM and profit-pool arguments. This companion focuses on the engineering physics; the main essay focuses on the vendor and margin structure.


==============================================================================
# The Modular Datacenter Stack Technical Companion
URL: https://adikumar.co/the-modular-datacenter-stack-technical-companion/
Published: 2026-08-02
Summary: Skid architectures, BESS integration, MV switchgear, busway and rack topologies for the modular hyperscale datacenter build.
==============================================================================

The AI Power Chain series · Part 12 of 15









Glossary of terms used



CAISO
California Independent System Operator. The grid operator for most of California.
ERCOT
Electric Reliability Council of Texas. Grid operator for most of Texas, operating largely as an electrical island.
GW
Gigawatt. One thousand megawatts of electrical power.
ISO
Independent System Operator. Regional grid operator (e.g. ERCOT, MISO, PJM, CAISO). Manages transmission access and wholesale market operations.
MISO
Midcontinent Independent System Operator. Regional grid operator covering 15 US states and Manitoba.
MV
Medium Voltage. Typically 1 kV to 35 kV. The voltage class connecting data centre power distribution to the utility grid.
PJM
PJM Interconnection LLC. Regional grid operator covering 13 US states plus DC. Includes Ohio, where the NVIDIA/OpenAI campus sits.
TAM
Total Addressable Market. The maximum revenue opportunity available if a product served every potential customer segment.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.






Technical Companion · Part VI



The Modular Datacenter Stack: engineering the site


Grid interconnection queues, MV substation design, factory-test protocols, structural floor loading for 3,000-pound racks, and the water and wastewater arithmetic that governs where a 500MW campus can actually be built.



Companion to the The Modular Datacenter Stack essay (Part VI of six). Use this piece for the interconnection, protection, structural and site-engineering arguments; use the main essay for the market map, vendor structure and TAM.



Section 1: Grid interconnection queues, not construction schedules, set the pace


The published construction schedule for a hyperscale AI campus, from ground-breaking to first-power, is 18-24 months. That schedule assumes the campus already has an approved grid interconnection agreement with the local transmission operator. If it does not, the effective timeline is 4-6 years, because the queue for large-load interconnection studies is that long in most US ISOs and most European transmission areas.


A grid interconnection request for a 500MW load triggers, at minimum, a feasibility study, a system impact study, and a facilities study. Each of these takes 6-12 months of transmission-planner time. The queue for those studies is now filled with wind, solar, battery storage, industrial electrification, and increasingly, datacenter load. In PJM (the largest US ISO by capacity), the interconnection queue has grown from ~200GW in 2020 to over 1,100GW in 2025. Approval timelines have stretched from 18 months in 2015 to 4-5 years in 2025. MISO, ERCOT, ISO-NE, CAISO all show similar patterns. Europe (ENTSO-E territory) has parallel queues at national TSOs; Ireland's EirGrid moratorium on new Dublin datacenter interconnection (2022 onward) is the sharpest visible example of the constraint.



US ISO interconnection queue growth
Active interconnection queue (GW of requested generation + load), major US ISOs, 2020-2025

FERC Form 715 and Lawrence Berkeley National Laboratory interconnection queue reports (2021-2025). PJM data from PJM Queue Points reports. Approval timelines from LBNL "Queued Up" annual reports.

The datacenter siting decision that everyone underestimates. The choice of where to build a hyperscale campus turns primarily on which utility service territory has grid capacity that can be interconnected within the schedule the operator can tolerate, more than on land, tax incentives, or labour cost. Sites in slow-queue territories (PJM, CAISO, Ireland) are being deferred or downsized. Sites in fast-queue territories (ERCOT, SPP, parts of the US Southeast, Malaysia, some Middle East locations) are being oversized. The "modular" advantage in this context is that a modular MV substation can be pre-built while the interconnection study runs, and can be energised within weeks of study approval, which is a form of schedule risk arbitrage.






In brief


Prefab enclosure economics turn on skid weight, transport constraints and on-site assembly tolerances. A 5MW power skid weighs 40-60 tonnes and requires specialised transport. Interface standards for water, power and control connections at the skid boundary determine how quickly multi-vendor systems can be commissioned. The engineering choice sits at that boundary.








Section 2: MV substation design, in enough detail to matter


The MV substation is where the utility feed transitions into the datacenter's distribution system. On a hyperscale campus, it typically operates at 115kV, 138kV or 230kV incoming, stepping down through one or more transformer stages to 34.5kV or 13.8kV for on-site distribution, then to 480V or 400V at the individual data hall's PDU. The substation contains the following primary equipment:


 1. Utility metering and disconnect (owned and operated by the utility, physically on the utility side of the property line)
 2. Incoming line breakers (SF6 or vacuum, 115kV+ class)
 3. Primary transformers (typically two, in an N+1 configuration, each rated for 100% of load; 60-120 MVA each is typical for a hyperscale site)
 4. Distribution switchgear (13.8kV or 34.5kV class, with vacuum interrupter breakers, protective relays and metering)
 5. Secondary transformers (13.8kV to 480V, typically 2.5-3.5 MVA each, one per data hall)
 6. Low-voltage switchboards (feeding UPS, mechanical loads, general facility load)


The modular version of this substation prefabs the entire secondary transformer + LV switchboard block on a single steel skid. Some vendors go further and prefab the primary switchgear + secondary transformer + LV switchboard as a single "e-house" (an enclosure large enough to walk inside). The primary transformer is typically not modular because a 60 MVA transformer is too large to fit on a road-legal shipping envelope; it arrives as a single piece on a specialised multi-axle transport.




Modular vs traditional substation delivery
Illustrative critical path for MV substation commissioning, weeks from equipment order to energisation

Author's estimate from industry-published modular substation case studies (Eaton, Vertiv, Schneider) and typical stick-built substation project timelines. Actual figures vary by utility approval process, site conditions and equipment specifications.





Transformer sizing and the redundancy question


The primary transformers on a hyperscale substation are the most expensive piece of electrical equipment on the site (roughly $2-4M each for a 60-120 MVA unit) and have the longest lead time (currently 80-120 weeks for hyperscale-grade transformers from Hitachi Energy, ABB, Siemens Energy, GE Vernova, Mitsubishi, or the Korean majors). The choice of redundancy strategy at this layer determines both cost and resilience:


N configuration: one transformer, sized for 100% load. Cheapest, but a transformer failure means full site outage until the transformer is repaired or replaced. Unacceptable for hyperscale AI.


N+1 configuration: two transformers, each sized for 100% load. Standard for hyperscale. One transformer can be taken out of service for maintenance without site impact. Requires additional switchgear bay and space, but no capacity penalty.


2N configuration: two fully-independent lineups, each with its own transformer, switchgear and distribution. Used on the highest-reliability tiers. Doubles the substation cost. Common for financial services datacenters, rare for hyperscale AI where the workload is more tolerant of individual-site outage.


Hyperscale AI campuses are typically N+1 at the primary transformer level and N+1 or 2N at the secondary transformer and UPS levels. The design choice is driven as much by the operator's overall fault-tolerance philosophy (application-level resilience vs infrastructure-level resilience) as by any specific reliability calculation.




Section 3: Protection coordination and the arc-flash question


Every distribution system needs a protection scheme: a coordinated set of protective relays and circuit breakers that isolates faults quickly enough to prevent equipment damage and personnel injury. On a hyperscale datacenter, the protection scheme has to satisfy four constraints simultaneously:


 1. Fault clearance time short enough that the fault current does not damage upstream equipment (typically <100ms for MV, <50ms for LV)
 2. Selectivity such that a fault is cleared by the nearest upstream breaker, not by a further-upstream breaker that would take out a larger portion of load
 3. Arc-flash mitigation such that if a fault does occur, the incident energy released is low enough to be survivable for personnel wearing appropriate PPE
 4. Selective ride-through for load-side transients such that the UPS does not trip on a downstream fault that clears normally


The protection design is embodied in the protective relay settings, and it is one of the most technically demanding pieces of a datacenter's electrical engineering. Modern protective relays (SEL, GE Multilin, ABB Relion, Siemens Reyrolle, Beckwith) are microprocessor-based devices with dozens of protection functions per relay, communicating over IEC 61850 GOOSE messaging for high-speed protection interlocks. The settings coordination for a hyperscale campus involves modelling every possible fault location and every possible source of fault current (utility, on-site generators, UPS output, BESS), and verifying that the selected relay curves produce the required clearance time and selectivity at every fault location.




Arc-flash incident energy: unmitigated vs mitigated
Indicative arc-flash incident energy (cal/cm²) at typical datacenter equipment, with and without modern arc-flash mitigation

Author's illustrative comparison based on IEEE 1584 arc-flash calculations. Mitigation includes fast-clearing relays, arc-flash-resistant switchgear, and light-sensing arc-flash detection. Actual incident energy depends on fault current, clearance time, working distance and enclosure type.




Modular substations from Vertiv, Eaton, Schneider and ABB are increasingly delivered with the protection coordination already engineered and the relay settings pre-programmed at the factory. This is a significant time-savings versus a stick-built substation where the protection engineer has to develop the settings on-site during commissioning. Factory-set protection also reduces the risk of on-site settings errors, which are a persistent source of protection mis-coordination and unexpected trips.




Section 4: Structural loading, and why the floor slab has become the problem


Historical general-purpose datacenters designed floor slabs for roughly 250 pounds per square foot (~1,200 kg/m²) uniform live load, which was sufficient for typical air-cooled 4-8kW racks. Modern hyperscale AI racks weigh 2,500 to 4,000 pounds each (~1,100 to 1,800 kg), concentrated on a footprint of roughly 4 square feet. That is a point load of over 1,000 pounds per square foot, more than 4x the historical design basis.


The load is not evenly distributed. Racks are placed in rows with narrow aisles between them, so the aisle floor sees ~50 psf while the rack row sees ~1,200 psf. This concentrated loading pattern requires either thickened slab design (poured slabs 12-18 inches thick, with reinforcement designed for the concentrated load pattern) or reinforced pier foundations at each rack location. Both are more expensive than a conventional datacenter slab, and both take substantially longer to install.




Datacenter floor loading: historical vs AI
Design floor load (pounds per square foot) for general datacenter vs modern AI racks vs point loads

Author's compilation from datacenter structural design guides (Uptime Institute, ASHRAE TC 9.9), Meta Louisiana campus disclosures on structural design, and industry benchmark reports on AI rack weights (NVIDIA GB200 NVL72 documentation, OCP ORv3 specification).


The modular civil answer. Compass Datacenters, Aligned and other modular-first developers address the structural loading problem by manufacturing the floor slab and rack pier system as part of the pre-fabricated data hall, with structural engineering done once at the fabrication yard rather than repeatedly at each site. This is one of the largest single time-savings that modularisation delivers on a hyperscale build, because the structural pour and cure cycle can run in parallel with other fabrication work rather than sitting on the site critical path.




Section 5: Water and wastewater arithmetic, and where datacenters can actually be built


The Part III essay treated water use as a Direct-Liquid-Cooling design question. This companion treats it as a facility siting question, because for hyperscale AI campuses the water-and-wastewater arithmetic frequently determines whether a site is even feasible.


An adiabatic-cooled hyperscale AI campus consumes roughly 1.5-2.5 gallons of water per kWh of IT load, primarily for evaporative cooling. At a 500MW site running 24/7 with a PUE of 1.2, that is roughly 600GWh/year of IT energy and roughly 1-1.5 billion gallons of water per yearThat volume is comparable to the annual water consumption of a small city (~50,000 people). Only certain sites can supply that water on a sustainable basis without imposing on regional water supplies.


The siting responses:


 1. Air-cooled hyperscale: zero water consumption, but roughly 10-15% higher power consumption than adiabatic. Used in water-constrained regions (US Southwest, parts of Middle East). Requires higher rack power density and more aggressive DLC.
 2. Closed-loop DLC with dry-cooler heat rejection: minimal water consumption (only make-up for evaporative losses in the closed loop). The default new-build design in Northern Europe, parts of Canada, US Pacific Northwest. Requires cooler ambient temperatures to work economically.
 3. Reclaimed water: some sites contract with municipal wastewater treatment plants for treated reclaimed water rather than potable. Reduces regional-supply impact but requires additional treatment for datacenter use.
 4. Seawater or brackish water cooling: coastal sites use seawater directly for heat rejection, with heat exchangers isolating the corrosive seawater from the internal chilled-water loop. Common in Nordic and Middle East coastal sites.




Datacenter water intensity by cooling architecture
Water consumption (litres per kWh IT load) and PUE by cooling architecture, hyperscale AI campus

Author's compilation from Meta, Microsoft and Google sustainability reports (2023-2025), Uptime Institute water-use survey data, and vendor cooling architecture disclosures. Actual water intensity varies substantially with climate, dry-cooler design and ambient temperature. Illustrative.





Wastewater disposal is the second binding constraint


The wastewater side is often overlooked. Blowdown from evaporative cooling towers concentrates dissolved solids and requires disposal, typically to municipal sewer at rates that many small municipalities cannot absorb at hyperscale volumes. Chemical additives to cooling loops (biocides, corrosion inhibitors) trigger industrial wastewater discharge permits with monitoring and treatment requirements. Some jurisdictions (parts of Arizona, Ireland, Netherlands) have specific datacenter wastewater regulations that impose higher costs and monitoring than industrial norms. Site selection increasingly considers wastewater permit feasibility alongside water supply.




Section 6: Commissioning and factory acceptance testing


The last engineering topic to cover in this companion is the shift from field commissioning to factory acceptance testing (FAT). On a stick-built substation, commissioning is done on site after all equipment is installed and connected: functional testing, protection settings verification, load-bank testing, integrated system testing. On a modular substation, most of this testing happens at the factory before the equipment ships. The site work is limited to reconfirming settings, running end-to-end integrated tests with the utility supply and downstream load, and formally accepting the equipment.


The FAT process for a hyperscale modular substation is a documented multi-day protocol that verifies:


 1. All wiring is correct per drawing
 2. All protective relays are set per the commissioning specification and trip within tolerance
 3. All breakers cycle correctly under manual and automatic control
 4. SCADA and remote monitoring communicates correctly with the datacenter's building management system
 5. Insulation resistance, contact resistance, transformer turns ratio and other factory tests confirm equipment integrity
 6. Load-bank testing at full rated current confirms thermal performance


A typical FAT protocol for a 15 MVA modular substation takes 4-6 days of engineering time at the factory, replacing what would be 15-20 days of on-site commissioning work. The customer's commissioning engineer typically travels to the factory for the FAT, which is a substantial time saving on aggregate for a hyperscaler building multiple substations per year.



Factory acceptance testing shifts the commissioning risk out of the field. The failure mode of a stick-built substation is that a wiring error or setting error is discovered during on-site commissioning, requiring rework that delays energisation by weeks. The failure mode of a factory-built substation is that the same error is discovered at the factory, where it can be corrected on the manufacturing floor without impacting site schedule.The engineering value of modularity




Section 7: What a design review of a new hyperscale AI campus would ask


The design review checklist for a hyperscale AI campus, from the perspective of a chief facility engineer:



DomainThe questionWhat "good" looks like on a 2026 hyperscale campus
Grid interconnectionStudy approved? Schedule?Feasibility study complete; system impact study in progress; energisation within 18 months of study approval
MV substationModular or stick-built? Redundancy?Modular where possible; N+1 primary transformers; N+1 secondary; 2N UPS
Transformer supplyOrder placed? Delivery date?Primary transformers ordered 24-36 months in advance of energisation date
SwitchgearVacuum or SF6? Arc-flash mitigation?Vacuum interrupter with arc-flash-resistant enclosure; light-sensing arc-flash detection
UPS + BESSLFP chemistry? Grid-services integration?LFP battery; behind-the-meter BESS integration where market conditions allow
BuswayAmpacity? Tap-off box density?≥ 1,600A rated per row; tap-off box per rack; UL 857 or IEC 61439-6 certified
Floor structuralDesign point load?≥ 1,200 psf point load capability; reinforced pier per rack row where slab thickness is insufficient
Water supplySustainable annual withdrawal?Reclaimed water or dry-cooler primary; potable make-up limited to <10% of annual demand
WastewaterDischarge permit? Treatment?Industrial wastewater permit in place; blowdown treatment sized for peak concentration
CommissioningModular content FAT'd at factory?≥ 80% of factory-buildable content commissioned at factory; site commissioning limited to integrated tests




Related topic hubs
 * AI Power Semiconductors
 * 800V Data Centre Power
 * Browse all topic hubs →



Annex · References


FERC Form 715 transmission planning data.


Lawrence Berkeley National Laboratory, "Queued Up" annual interconnection queue reports (2021-2025).


Uptime Institute Global Data Center Survey 2023, 2024, 2025 editions.


ASHRAE TC 9.9 Thermal Guidelines for Data Processing Environments.


OCP Open Rack v3 specification.


IEEE 1584 Guide for Performing Arc-Flash Hazard Calculations.


IEC 61850 substation automation communication protocol series.


Vertiv, Eaton, Schneider Electric modular substation product documentation and case studies (2023-2025).


Meta, Microsoft and Google sustainability reports on datacenter water use (2023-2025).


Turner & Townsend and JLL hyperscale market cost benchmarks (2024-2026).



Companion to main essay


Read alongside the main essay "The Modular Datacenter Stack" (Part VI of six) for the vendor map, TAM and profit-pool arguments. This companion focuses on the engineering side of substation, structural, water and commissioning; the main essay focuses on the market structure.


==============================================================================
# How the AI buildout is financed
URL: https://adikumar.co/how-the-ai-buildout-is-financed/
Published: 2026-08-10
Summary: How the AI buildout is financed. DDTL, hyperscaler ABS, private-market equity, nuclear PPAs, sale-leaseback: the ten capital-structure stacks in one map.
==============================================================================

The Investment Layer series · Part 1 of 8








The Investment Layer · Part I of VIII



How the AI buildout is actually financed


The $500B headline hides six funding pools with different costs of capital, different return paths, and different constraints. Which pool each layer of the physical chain draws from is the first thing to know before pricing any position in it.




The Investment Layer · eight essays for capital allocators

 1. Part I. How the AI buildout is financed (you are here)
 2. Part II. The M&A map, past and forward
 3. Part III. The build-to-lease flip
 4. Part IV. The coverage asymmetry
 5. Part V. The PE playbook, layer by layer
 6. Part VI. Underwriting AI infrastructure debt
 7. Part VII. The three downside cases
 8. Part VIII. Sovereign capital and industrial policy




Executive summary


 * The reported figure for global AI-infrastructure capex over 2025-2028 is around $500-700 billion per year. That number is an aggregation across at least six distinct pools of capital, each with its own cost, return horizon, and constraint set. Treating it as one figure produces incoherent analysis.
 * The largest single pool is hyperscaler operating cash flow. Google, Meta, Microsoft, Amazon and Oracle together fund roughly one-third of annual AI-infrastructure capex from their own free cash. The cost of that capital is the hyperscaler's own equity cost of capital, which is unusually low. It buys equipment they will own, on schedules they control.
 * The second largest pool is REIT equity and debt, funding wholesale colocation capacity that hyperscalers lease. Digital Realty and Equinix together carry over $300 billion of enterprise value at cap rates that have compressed steadily since 2022. This pool funds the buildings and the land, not the equipment inside them.
 * OEM balance-sheet financing is a third pool that most analysis misses. Vertiv, Eaton, Schneider, Legrand and their peers extend working capital, deferred billing and consignment arrangements at scale, effectively financing customer inventories from their own credit lines. The size of this pool sits inside receivables and payables lines that do not aggregate cleanly.
 * Structured infrastructure debt is a fourth pool that has grown fastest. Project finance around behind-the-meter power, dedicated substations, and modular data-hall developments now supports single deals above $2 billion. The lenders are infrastructure debt funds, insurance company balance sheets, and export credit agencies where equipment sourcing crosses borders.
 * Sovereign capital is the fifth and least transparent pool. Gulf funds, East Asian state entities and European industrial policy vehicles each direct capital into specific parts of the chain for reasons that are strategic before they are financial. This pool moves the marginal deal in several sub-sectors.
 * Private equity growth and buyout capital is the sixth pool. It has concentrated in modular developers, thermal services businesses, and specialty electrical equipment. Cheque sizes range from $100 million to $3 billion. Exit paths are narrower than most funds assume.
 * Each layer of the six-part physical chain draws primarily from one or two of these pools. Knowing which pool funds which layer is the first analytical step in every downstream question the following essays address.




How does the $500 billion AI buildout figure decompose?


Every summary of the AI buildout begins with an aggregate capex number. Bloomberg and the sell-side have converged on $500-700 billion of annual capital spending across the hyperscaler and colocation segments during 2025-2028. The direction and order of magnitude are correct. The number is analytically unhelpful because it aggregates capital with different costs, different holding periods, different return requirements, and different sensitivities to the same event.


A single-tenant hyperscale build funded from Google's operating cash flow, a modular data hall developed by Compass Datacenters on Ontario Teachers' equity and Blackstone debt, a merchant substation project financed by Brookfield Infrastructure debt with off-taker credit substitution, and a Gulf sovereign wealth fund's direct investment into a UAE campus with accelerator supply guarantees are all counted in the same headline. They price completely differently. When AI capex slows, they behave completely differently. When rates rise, they revalue completely differently.


The rest of this essay decomposes the aggregate into six pools and describes how each behaves. The following essays in the series build on this decomposition. Anyone underwriting, advising or investing in this space should be able to name which pool a given position sits inside before they price anything else about it.



Chart 1. Six pools, dollar share of annual AI-infrastructure capex, 2026 estimate


The $500-700 billion aggregate is not one pool. Hyperscaler cash flow is the largest single source, but the other five together are larger, and each behaves differently under stress. The share numbers are directional and rest on a synthesis of company disclosures and industry estimates.



Sources: hyperscaler capex from Alphabet, Meta, Microsoft, Amazon, Oracle FY2025 filings and 1H2026 guidance; REIT enterprise values from company filings; OEM balance sheet estimates from Vertiv, Eaton, Schneider quarterly disclosures; structured debt from IJGlobal Q2 2026; sovereign and PE from Preqin, PitchBook, and public transaction records. Total sums to approximately $575 billion at midpoint.



POOL 1Hyperscaler operating cash flow



Hyperscaler operating cash flow


The five hyperscalers (Google, Meta, Microsoft, Amazon, Oracle) between them fund roughly 30-35% of global AI-infrastructure capex directly from operating cash flow. On 2026 forward guidance, that is a combined figure of $180-220 billion per year in AI-attributable capex funded internally, alongside similar amounts of general-purpose data-centre and cloud capex. The number is disclosed to the extent that each firm chooses to disclose it, and the AI-attributable split is estimated with wide error bars.


The cost of this capital is the hyperscaler's own weighted average cost of capital, which is unusually low. Google's WACC sits in the 7-9% range depending on the calculation. Meta's is comparable. Microsoft is lower. Oracle is higher, reflecting its debt load. For comparison, a REIT developer's blended cost of capital sits at 8-12% (higher equity cost, offset partly by cheaper debt), a modular-first PE-owned developer at 12-16%, and an infrastructure debt fund's underlying returns are 5-7% on debt itself but require pre-lease-backed revenue certainty.


This means hyperscalers can rationally build assets that a modular developer cannot economically build for them at the price they would pay to lease it. That is one reason the ownership share of hyperscale capacity has not fallen further than it has: for their most strategic capacity, hyperscalers have a cost-of-capital advantage that no third party can match.


The constraint sits in management attention and organisational capacity rather than in cost of capital. Hyperscalers self-develop where the site is strategic, the design is proprietary, or the technology is at the frontier. They lease where speed matters more than customisation and where the developer's construction organisation is larger than the hyperscaler's internal team can staff.


Reader implication: any analysis of hyperscale capacity that treats the ownership and leasing splits as fixed misses the underlying dynamic. The split reflects a make-versus-buy calculation that shifts by geography, technology generation and organisational capacity, and it has moved meaningfully over 2020-2026 without a single headline event to mark it.


POOL 2REIT equity and debt



REIT equity and debt


Data centre REITs and REIT-adjacent operators fund the buildings and the land that host the equipment. Digital Realty (DLR) and Equinix (EQIX) together carry over $300 billion of enterprise value on 2026 figures. Chindata (delisted), Keppel DC REIT, and NextDC round out the internationally listed pool. The private REIT-like operators (Vantage, Aligned, Compass, Prime, STACK) fund from PE equity plus infrastructure debt on comparable but not identical terms.


REIT capital funds a specific slice of the buildout: the shell, the land, the power interconnection, and the base electrical and cooling infrastructure. It does not fund the accelerators, the memory, the servers or the network gear inside the halls. Those sit on the tenant's balance sheet (for owned hyperscale) or on the lease structure (for wholesale colocation). This distinction sounds obvious. It is regularly missed in coverage that talks about "REIT AI exposure" without decomposing which parts of an AI rack the REIT actually owns.


Cap rates in the sector have compressed from around 6% in 2022 to closer to 4.5% in 2026 for stabilised hyperscale-tenanted assets in top-tier markets. The compression reflects three things: rate-cutting expectations that have partly reversed, tenant concentration risk being repriced downward as hyperscaler credit continues to strengthen, and the scarcity premium on interconnection-ready sites. That last factor is the interesting one. Two identical sites can trade at cap rates 100-150 basis points apart based only on the presence of an executed interconnection agreement. This is the same finding the pricing essay reached from a different angle: time has become a separately priced good.


REIT debt sits alongside REIT equity in the capital stack. Investment-grade REIT bonds trade at 100-180 basis points over Treasuries depending on issuer and tenor. Structured debt on individual campuses trades wider. The overall pool has grown from around $220 billion in 2022 to over $340 billion by 2026 across public and private developers, on very rough estimates.


POOL 3OEM balance-sheet financing



OEM balance-sheet financing


The third pool sits inside working-capital lines on the equipment vendors' balance sheets and is systematically undercounted in aggregate AI capex figures. Vertiv, Eaton, Schneider, Legrand, ABB, Siemens Energy and their peers extend supplier credit, deferred billing, consignment arrangements, and multi-year fulfilment schedules that effectively finance customer inventories.


The clearest visible signal of this pool is the acceleration in the vendors' own receivables balances. Vertiv's trade receivables grew from around $1.2 billion at end-2022 to over $2.8 billion by mid-2026. Eaton's electrical segment shows a comparable pattern. The vendors are extending progressively longer payment terms and larger reserved-slot positions, funded from their own credit facilities and commercial paper programs. In effect, the OEM is lending its investment-grade balance sheet to its hyperscaler customer at spreads well below what the customer would pay in the market.


The pool is contingent capital in a way that headline capex figures do not capture. It can be pulled back rapidly if credit conditions tighten or if the vendor's own leverage begins to matter to its rating, which means the effective availability drops before any capex number changes. It is also a competitive weapon: the OEM most willing to extend the balance sheet wins the reservation, and the willingness varies across vendors in ways procurement teams track carefully. And the financing is priced into the equipment sale rather than delivered separately, so the customer pays for it in the unit price whether or not the receivable is explicit.


Sizing this pool precisely requires reconstruction from vendor balance sheets, and the number would be an estimate in any case. Reasonable range: $60-90 billion of embedded customer financing across the major electrical and thermal OEMs by end-2026. That figure has roughly tripled from the 2022 baseline.


POOL 4Structured infrastructure debt



Structured infrastructure debt


The fourth pool has grown fastest. Project finance and structured infrastructure debt now supports single AI-infrastructure deals above $2 billion, an order of magnitude larger than the typical data-centre construction loan of five years ago. The pool includes traditional bank-syndicated project debt, infrastructure debt funds, insurance company balance sheets deployed as private credit, and export credit agency-backed financing where equipment sourcing crosses borders.


The deal that established the template was the Meta-Sabey campus financing in Louisiana in 2025, structured as a pre-lease-backed construction and term loan facility of over $3 billion at an all-in rate 250 basis points inside comparable general-purpose infrastructure debt. Since then, dozens of deals have followed the same shape: a hyperscaler pre-lease commitment substitutes for the traditional off-taker credit that supports transmission or gas-pipeline debt.


The construction is analytically interesting. In classical infrastructure debt, an asset generates cash flow from a diversified user base over a long life (20-40 years), and lenders underwrite the asset. In AI-infrastructure debt, the asset generates cash flow from one hyperscaler tenant over a shorter effective life (7-15 years), and lenders effectively underwrite the tenant. The tenant's balance sheet is the true security. This is fine while hyperscaler credit is investment-grade. It becomes uncomfortable if a downside case pressures any single hyperscaler's cash generation, which Part VII treats in detail.


Behind-the-meter power projects (SMR PPAs, gas turbine peakers, LFP BESS installations at datacenter sites) are financed on a related but distinct structure, closer to classical merchant power finance with contracted revenue from the adjacent hyperscaler load. The Amazon-Talen Energy nuclear PPA for the Susquehanna campus is the visible template. Comparable deals with GE Vernova and Bloom Energy are structured on similar principles.


Total 2026 flow into this pool is estimated at $80-120 billion, growing 30-40% annually. The distinction between the debt and the underlying tenant credit becomes important in downside cases and is a specific focus of Part VI.


POOL 5Sovereign capital



Sovereign capital


The fifth pool is the least transparent and increasingly the marginal buyer. Middle Eastern sovereign wealth funds, East Asian state entities, and European industrial policy vehicles each direct capital into specific parts of the AI-infrastructure chain for reasons that are strategic before they are financial. The pool is not an aggregate market: it is a set of specific investment programs with specific mandates, and the aggregate is only meaningful because the sums are large enough to move sector-level flows.


MGX, backed by the Abu Dhabi government, has committed capital across AI accelerator supply, data centre development in the UAE and outside, and vertically integrated AI campuses. Public Investment Fund of Saudi Arabia has taken similar positions with a stated preference for infrastructure adjacent to the Kingdom's own accelerator supply. Both funds have deployment horizons that run into decades and cost-of-capital assumptions that treat strategic and financial returns as blended. Their arrival in the deal-making conversation over 2024-2026 has repriced several assets that would otherwise have traded at strictly financial multiples.


East Asian state capital works differently. Japanese, Korean and Taiwanese public and quasi-public funds have deployed at scale into equipment vendors within their national industrial bases (JIC's role in the Shinko take-private attempt is illustrative), and into strategic long-term supply relationships. Chinese state capital has funded the parallel domestic buildout at scale, largely outside the Western dealflow, and modifies the global supply-demand picture in ways that affect Western pricing without directly funding Western assets.


European industrial policy vehicles (EU Innovation Fund, national strategic funds in France, Germany and the Netherlands) direct capital into sovereignty-flagged assets: European foundry capacity, European AI compute, European semiconductor packaging. The sums are smaller in absolute terms but influence which specific projects reach financial close in Europe.


Total sovereign deployment into AI-infrastructure across all three regional flavours is estimated at $40-80 billion per year by 2026, with wide variance. The important characteristic is that this capital sets prices for the deals it participates in, and its withdrawal from a market can move sector valuations rapidly.


POOL 6Private equity growth and buyout



Private equity growth and buyout


The sixth pool is conventional PE capital deployed as growth equity, minority stakes, and leveraged buyouts across the equipment, services and modular-developer parts of the chain. It is the smallest of the six pools in aggregate but the most active in generating dealflow visible to a typical PE reader.


Cheque sizes range from around $100 million for growth-equity minority stakes in specialised private vendors, through $500 million to $2 billion for mid-market buyouts of established equipment or services firms, up to $3 billion or more for platform acquisitions of modular developers. Compass Datacenters (Brookfield and Ontario Teachers, 2021), STACK Infrastructure (IPI Partners, then Blue Owl 2024), Vantage (DigitalBridge and majority PE), Aligned (Macquarie and BlackRock), CoolIT Systems (KKR majority 2024), and PurgeRite (acquired by Vertiv from PE ownership) are all illustrative deals.


Deployment has concentrated in a small number of sub-sectors: modular datacenter development, thermal services, specialty electrical equipment, and CDU and cooling integration. PE has largely stayed out of the substrate, equipment and semiconductor layers where check sizes, technology risk, and exit paths do not fit typical mandates. The exit paths for the platform investments that have been made are narrow. Strategic sale to one of the Western industrial majors (Eaton, Schneider, Vertiv), IPO into a receptive market, or secondary sale to a larger infrastructure fund. Each has been executed but each has capacity limits, and the queue of PE-owned platforms approaching exit windows over 2026-2028 is uncomfortably long.


Beyond the exit-path constraint, PE-backed platforms compete for capital and talent with the equipment vendors' own service organisations. That competition affects the pricing of both, and is the reason many recent PE roll-ups in cooling and electrical have hired directly from OEM service teams. The dynamic is treated in Part V's PE playbook per layer.


Total 2026 flow from this pool is estimated at $30-60 billion of new commitments and follow-ons.


Six pools. Different costs of capital, different return horizons, different constraints. The layer-by-layer allocation is the map that every subsequent essay in this series uses.The framework in one line



How do the six capital pools map onto the six-layer physical chain?


The final table below allocates each of the six layers of the physical chain (from the AI Power Chain series) to its primary funding pools. The mapping is directional and any specific project will draw from more than one pool. The pattern matters more than the precision.



LayerPrimary funding poolSecondary poolNotable characteristic

Capacitor stack (HSC, MLCC, silicon caps)OEM balance sheet + PE growthSovereign (Japanese state capital into passives)Small BOM per system, large aggregate value, qualification-gated. Not a PE mega-deal layer.
Wide-bandgap stack (SiC, GaN devices)OEM balance sheet + PE growthPublic market equity for the merchant device makersWolfspeed's restructure exemplifies the risk of over-capitalised single-technology bets.
Thermal stack (cold plates, CDUs, UQDs, fluids)PE growth + OEM balance sheetStrategic capital from Western majors buying into servicesFastest-consolidating layer. PE has active deployment. Exit paths active.
Interconnect stack (transformers, switchgear, GOES, generators)REIT-adjacent + structured infrastructure debtOEM balance sheet, sovereign for national-security-sensitive assetsThe layer where structured infra debt has grown fastest. Lead times gate the deal calendar.
On-package delivery (multi-phase POL, IVR, glass substrates, foundries)Public market equity + sovereignOEM balance sheet at the merchant device layerFoundry-anchored: three foundries hold the strategic capital allocation decisions.
Modular datacenter stack (skids, UPS + BESS, busway, rack enclosures, modular civil)PE growth (modular developers) + REIT equity + structured debtSovereign for flagship international sites; hyperscaler CFO for owned buildsThe layer where all six pools appear. Most exposed to cost-of-capital shifts.




Two observations from the table. First, the layers most exposed to PE capital (thermal, modular DC) are also the layers with the most active M&A. The mapping is not accidental: PE deploys where exit paths exist, and exit paths exist where strategic buyers are consolidating. This is the topic of Part II.


Second, the layers most exposed to structured infrastructure debt (interconnect, modular civil) are also the layers where the tenant-credit substitution described earlier is most visible. In a scenario where hyperscaler credit deteriorates, these are the layers where the debt stack revalues first. This is the topic of Part VI.



What this decomposition changes about analysis


Anyone who wants to make sensible statements about the AI-infrastructure buildout needs to know which of these six pools they are talking about. The aggregate figure is not just imprecise; it is misleading in specific and predictable ways.


Rate sensitivity is the clearest example. A 100 basis point move damages Pool 4 (structured debt) severely, damages Pool 2 (REIT equity) meaningfully through cap rate expansion, damages Pool 6 (PE growth) via multiple compression, and barely touches Pool 1 (hyperscaler cash flow). Any aggregate figure that masks the split cannot say anything useful about what happens when rates move.



Chart 2. Estimated valuation impact of a sustained +100 bps rate move, by pool


The pools have very different rate sensitivities. Structured debt and REIT equity absorb almost all the damage from a rate move; hyperscaler cash flow is largely insulated. The estimates are directional and reflect typical portfolio composition within each pool.



Author estimates from disclosed cap rate sensitivities (Green Street Advisors), infrastructure debt spread duration (industry benchmarks), and PE multiple compression precedents. Individual assets vary considerably around the pool-level figures.



Return attribution matters equally. A given asset in the chain is funded by a specific pool, and its returns accrue to that pool. The public conversation about "who is making money on AI infrastructure" often conflates the vendor selling equipment, the developer building the shell, the REIT owning the shell, and the PE fund holding the developer. Each captures a different slice; none captures all of them.


The forward outlook is where the aggregation is most damaging. Pool 4 (structured debt) is doubling every 18-24 months. Pool 5 (sovereign) is repricing individual sub-sectors as it enters. Pool 6 (PE) is approaching a wall of exit obligations described in Part II. These trajectories point in different directions, and averaging them into a single capex figure erases the underlying moves that any allocator needs to see.



Chart 3. Pool growth trajectory, 2022-2026, indexed to 2022 = 100


The six pools have grown at very different rates. Structured debt is the fastest-growing (3.6x since 2022); OEM balance sheet has tripled; sovereign has doubled. Hyperscaler cash flow has grown alongside underlying revenue but at a lower multiple. Read the essay as a decomposition of where the acceleration has actually come from.



Author estimates from company disclosures, Preqin, PitchBook, IJGlobal, and industry reports. Indexed to a 2022 baseline that was itself already elevated by the 2020-2021 acceleration in datacenter capex. The relative growth rates matter more than the specific index values.



The remainder of this series builds on the six-pool decomposition. Part II describes the M&A dealflow that follows Pool 6's exit obligations and Pool 3's strategic consolidation. Part III walks through what Pool 1's shift from own to lease has done to Pool 2. Part IV maps what public-market analysts miss because they cover Pool 2 heavily and Pool 3 sparsely. Part V applies the framework to specific PE opportunities across the six physical layers. Part VI treats the emerging Pool 4 debt product in detail. Part VII names three scenarios that reprice Pools 2, 4 and 6 differently. Part VIII treats Pool 5 in the depth it deserves.




Glossary of terms used



CDU
Coolant Distribution Unit. Thermal system component distributing chilled coolant to racks or direct-to-chip cold plates.
GaN
Gallium Nitride. Wide-bandgap semiconductor material used in RF and power applications.
HSC
Hybrid Supercapacitor. Energy storage technology bridging batteries and traditional capacitors.
IPO
Initial Public Offering. A private company's first sale of shares to the public market.
IVR
Integrated Voltage Regulator. On-package power delivery component for modern CPUs and GPUs.
PPA
Power Purchase Agreement. Long-term contract locking in electricity price per MWh over a defined term, typically 10-15 years.
SMR
Small Modular Reactor. Nuclear reactor design with output typically below 300 MWe, envisioned for co-located generation.
SiC
Silicon Carbide. Wide-bandgap semiconductor material used in high-voltage power electronics.
UPS
Uninterruptible Power Supply. Battery-backed power system that keeps critical loads running during grid disruptions.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.





Sources and method


Capex aggregates for 2025-2028 draw on Bloomberg NEF Q2 2026 data centre capex tracker; McKinsey "Global Infrastructure Outlook 2026" edition; Dell'Oro Group's data centre capex report. Numbers are rounded and directional.


Hyperscaler capex and cash flow figures from FY2025 and 1H2026 disclosures (Alphabet 10-K and 10-Q; Meta 10-K and 10-Q; Microsoft 10-K; Amazon 10-K; Oracle 10-K). AI-attributable capex splits are estimated from company commentary and are not audited disclosures.


REIT enterprise values from Digital Realty and Equinix filings and market data as of mid-2026. Cap rate direction from Green Street Advisors and CBRE data centre reports 2022-2026.


OEM balance sheet analysis from Vertiv, Eaton, Schneider Electric, Legrand, ABB, Siemens Energy quarterly filings 2022-2026. Trade receivables growth figures directly from balance sheet reporting.


Structured infrastructure debt deal templates from IJGlobal Infrastructure and Project Finance database, PitchBook, and Preqin. The Meta-Sabey Louisiana structure and Amazon-Talen Susquehanna PPA are drawn from company disclosures and IJGlobal reporting.


Sovereign capital deployment from MGX, PIF, JIC and comparable fund disclosures where available, and Reuters and Financial Times reporting for undisclosed transactions.


PE deployment figures from Preqin data centre and digital infrastructure category, Q2 2026 update.



Series


The Investment Layer (eight parts). Part I (this piece): How the AI buildout is financed. Part II: The M&A map. Part III: The build-to-lease flip. Part IV: The coverage asymmetry. Part V: The PE playbook per layer. Part VI: Underwriting AI infrastructure debt. Part VII: The three downside cases. Part VIII: Sovereign capital and industrial policy. This series is a commercial companion to the AI Power Chain (six parts plus six technical companions) and the earlier supplements Pricing Under Scarcity and The Services Inversion.


The author is an independent advisor working on data centre, power and industrial technology. This analysis is written in a personal capacity and rests entirely on public information. Nothing here is investment advice.







See also · The Investment Layer

 * The AI infrastructure M&A map
 * The build-to-lease flip
 * The coverage asymmetry





Further reading + sources
Related essays

 * Underwriting AI infrastructure debt. the specific credit lens
 * Sovereign capital and industrial policy. the international-capital layer
 * The PE playbook layer by layer. private-equity strategy specifics
 * Financing I. DDTL and contract-collateralised debt. the DDTL primer


Primary sources

 * S&P Global Market Intelligence. private-credit market data
 * US Treasury. sovereign wealth + CFIUS overlay







In brief


The AI buildout is financed through ten distinct capital structures, from delayed-draw term loans backed by hyperscaler contracts to sovereign wealth deployments. DDTL structures at CoreWeave, Nscale and Crusoe reached approximately $95B in aggregate committed capital by mid-2026. Private-market equity, hyperscaler-contract ABS, nuclear PPAs, sale-leaseback and REITs make up the rest.








Frequently asked


What is a DDTL structure in AI infrastructure financing?

A delayed-draw term loan (DDTL) lets an operator draw debt against specific customer contracts as they are signed. Each tranche matches specific contract commitments. Debt service comes from contracted cash flows. CoreWeave, Nscale and Crusoe use DDTL structures for their GPU-cloud buildout, with aggregate committed capital reaching approximately $95B by mid-2026.


How is the AI buildout different from traditional infrastructure financing?

Traditional infrastructure debt underwrites long-dated regulated assets with utility-like cash flows. AI infrastructure debt underwrites contracted revenue from a handful of concentrated hyperscaler counterparties. Contract-specific covenants (change-of-control, capacity flex, cure periods) matter more than typical utility credit metrics. Recovery scenarios in default differ sharply.


==============================================================================
# The AI infrastructure M&A map
URL: https://adikumar.co/the-ai-infrastructure-m-a-map/
Published: 2026-08-11
Summary: AI infrastructure M&A map: every material deal 2023-2026, four through-lines that explain them, fifteen forward hypotheses on where consolidation goes next.
==============================================================================

The Investment Layer series · Part 2 of 8








The Investment Layer · Part II of VIII



The AI infrastructure M&A map


Every material deal 2023-2026 across the six layers, the four through-lines that describe them, and fifteen forward candidates by name for the next 24 months.




The Investment Layer

 1. Part I. How the AI buildout is financed
 2. Part II. The M&A map (you are here)
 3. Part III. The build-to-lease flip
 4. Part IV. The coverage asymmetry
 5. Part V. The PE playbook, layer by layer
 6. Part VI. Underwriting AI infrastructure debt
 7. Part VII. The three downside cases
 8. Part VIII. Sovereign capital and industrial policy




Executive summary


 * Consolidation across the AI infrastructure chain has followed four recognisable patterns rather than a scattered set of one-offs. Recognising the pattern is the fastest way to predict the next transaction.
 * The four through-lines are: buying out the qualified field to convert scarcity into pricing power, buying the trained organisation because operational capacity cannot be built incrementally, buying certification barriers that regulation is otherwise eroding, and buying installed-base access for the services annuity that follows.
 * The Eaton acquisition of Boyd Thermal at $9.5 billion (closed March 2026 per Eaton press release) is the largest deal on the map. Vertiv's purchase of PurgeRite at approximately $1 billion, Schneider's majority acquisition of Motivair, and the Ferric take-out by TDK are smaller in headline size and comparable in strategic weight.
 * The forward map identifies fifteen firms as likely acquisition targets over the next 24 months, distributed across the thermal, on-package delivery, modular datacenter, and specialty electrical layers. The list is names rather than categories.
 * Three exit paths dominate the private-market side: strategic sale to one of the six Western industrial majors, IPO into a market that has become receptive to specialty industrials with AI exposure, and secondary sale to a larger infrastructure fund. Each has capacity limits, and the queue is approaching them.
 * The layer with most M&A activity by dollar volume is thermal, followed by modular datacenter. The layer with least activity is on-package delivery, where foundry ownership of the strategic value pool leaves little for consolidators to acquire.
 * Chinese domestic M&A activity is running at comparable dollar volumes to Western dealflow but operates in a largely separate market. Cross-border consolidation into or out of China has been shrinking since 2023 and is priced accordingly.




Chart 1. AI-infrastructure M&A deals by size and quarter, 2023 Q1 to 2026 Q3


Each bubble is a transaction from the deal ledger. Bubble area scales to deal value; colour indicates the through-line pattern (see next chart). The Eaton/Boyd deal in Q4 2025 dominates the size distribution; the frequency has accelerated through 2024-2026 without any single deal matching its scale.



Sources: PitchBook, Bloomberg, Financial Times, Reuters, and company disclosures. Undisclosed deal values estimated for chart placement.




The four through-lines


Consolidation across the AI-infrastructure chain over 2023-2026 has followed four recognisable strategic patterns rather than a scattered set of one-offs. Each pattern justifies a different price and produces a different post-deal outcome. Reading a fresh transaction requires first placing it in one of the four buckets.



Through-line 1. Buying out the qualified field


Where a small number of firms hold the qualification barrier for a critical component, acquiring one or more of them converts an oligopoly into a monopoly or a strong duopoly. The buyer captures pricing power that the standalone target could not have captured on its own, because the pricing question was always about the field's discipline rather than any single firm's cost position.


The Yageo acquisition of KEMET in 2020 is the reference case. Yageo bought a tantalum and polymer capacitor leader, consolidating the field on top of its MLCC and passive component franchise. The strategic value sat in the removal of a competitor from the qualified list on customer specifications rather than in KEMET's specific product portfolio (which Yageo could have replicated organically) where switching costs were high.


The Ferric acquisition by TDK in 2025 fits the same pattern. Ferric was one of two merchant IVR magnetics vendors globally (Empower is the other). Bringing Ferric inside TDK converts the merchant IVR market from a duopoly to effective single-vendor concentration in a strategically valuable customer segment. The deal price (undisclosed but reported around $250 million) is a fraction of what the strategic value implies, because Ferric's standalone financials did not yet reflect the position it was moving into.


Reader implication: any target that is one of three or fewer qualified providers to a critical AI-infrastructure specification prices higher on strategic value than on standalone financials. Precedent supports a 2-4x uplift over comparable non-qualified peers.



Through-line 2. Buying the trained organisation


The AI-infrastructure buildout is labour-constrained more than capital-constrained. Field service technicians, commissioning specialists, MV electricians, and thermal loop engineers cannot be recruited at the pace the buildout requires. Acquiring a firm that already employs and has trained hundreds of qualified staff produces immediate scale that could not be built organically in less than five to seven years.


The Vertiv acquisition of PurgeRite at approximately $1 billion in 2024 is the clearest recent example. PurgeRite's technical IP is real but narrow. Its trained workforce of commissioning and loop maintenance specialists is what commanded the premium. Vertiv paid for the ability to promise a commissioning service alongside every liquid cooling installation, at scale, immediately.


The Vantage-CyrusOne-style consolidation of trained development and construction management teams by PE-backed developers over 2022-2025 fits the same pattern. The bought firms had property portfolios and pipelines, but the price premiums reflected the value of intact development organisations that could execute at the pace hyperscaler pre-lease commitments required.


Reader implication: any target with trained field organisation of over 500 qualified specialists in a scarce trade prices at 3-6x revenue rather than the 1.5-2.5x revenue multiples that industrial services businesses typically command.



Through-line 3. Buying certification and regulatory barriers


Where a certification, qualification programme, regulatory approval or reference-design position is the durable moat, acquiring the certified entity locks that position in place ahead of any legislative or standards-body erosion. This pattern is most visible in electrical and thermal categories where a small number of firms hold hyperscaler or accelerator-vendor certifications that took years to earn.


The Schneider majority acquisition of Motivair in 2024, at a reported price around $850 million but not officially confirmed, fits this pattern. Motivair's CDU designs held reference-position certification with the two dominant hyperscaler thermal specifications for direct-to-chip cooling. Schneider could have developed a competing product line, but not before Motivair's certification position had matured into an incumbent advantage difficult to displace.


The Prysmian acquisition of Encore Wire in 2024 at $4.4 billion fits the same pattern on the electrical side. Encore Wire held UL and NEC certification positions across a full US electrical cable and conduit portfolio, plus a US manufacturing footprint that is strategically important given onshoring pressure. Prysmian added regulatory position that would have taken it years to develop, in a market where the AI buildout was already accelerating.


Reader implication: certification-holder targets price on the option value of the certification's remaining useful life, which typically extends 5-10 years from acquisition. Right-to-repair legislation and open specification movements shorten this window, and are treated in detail in The Services Inversion.



Through-line 4. Buying installed-base access for the services annuity


Any equipment class with an installed base measured in the tens of thousands of units generates a maintenance and services stream that either accrues to the manufacturer or accrues to an independent maintenance industry. Acquiring the equipment vendor, or the services business already serving that installed base, secures the annuity ahead of independent competition.


The Eaton acquisition of Boyd Thermal at $9.5 billion (closed March 2026 per Eaton press release) is the largest deal on the map and rests on this pattern more than the others. Boyd's cold plate manufacturing business is meaningful but the strategic weight is the installed base of over five million cold plates already deployed across hyperscale and colocation sites, each requiring service and refresh work over its operating life. Eaton is buying access to that annuity while the independent maintenance market for liquid cooling has not yet emerged at scale.


Vertiv's purchase of Geist earlier and PurgeRite subsequently followed the same logic. The equipment installed by others creates the maintenance requirement; Vertiv acquired the firms best positioned to capture that maintenance work over the next decade.


Reader implication: installed-base access targets price on the projected services revenue stream, which typically implies 15-25x recurring services EBIT even where equipment margins are moderate.


Four patterns explain the M&A. Buying the qualified field, buying the trained organisation, buying the certification, buying the installed base. Each has a different price, a different post-deal outcome, and a different implication for the next transaction.The strategic map in one line



Chart 2. Dollar volume of AI-infrastructure M&A 2023-2026, by through-line


The four through-line patterns account for very different dollar volumes. Installed-base plays dominate value because the Eaton/Boyd transaction sits inside that bucket. Certification-driven deals are the second-largest by dollar volume; qualified-field plays are the largest by deal count but smallest by dollar volume because individual targets are small.



Author's classification of published deals against the four through-line patterns. Some deals blend patterns; primary through-line used for chart placement.




The deal ledger, 2023 through mid-2026


The table below summarises the material transactions across the six layers of the AI-infrastructure chain in the sub-3-year window covered. It excludes non-material tuck-ins, IPOs without change of control, and financing rounds that did not change ownership. Cross-border deals into and out of China are shown only where a Western buyer or seller was involved.



DealAnnouncedReported valueLayerThrough-line

Eaton / Boyd ThermalNov 2025, closed Mar 2026$9.5 billionThermalInstalled base + trained org
Prysmian / Encore WireApr 2024, closed Nov 2024$4.4 billionInterconnect (cables)Certification + US onshoring
Vertiv / PurgeRiteSep 2024~$1 billionThermal servicesTrained organisation
Schneider Electric / Motivair (majority)Jun 2024~$850 million reportedThermal (CDU)Certification
Infineon / GaN SystemsNov 2023$830 millionWide-bandgap (GaN)Qualified field
Renesas / TransphormFeb 2024, closed Aug 2024$339 millionWide-bandgap (GaN)Qualified field
TDK / Ferric2025Undisclosed, ~$250M reportedOn-package delivery (IVR)Qualified field
Blue Owl / STACK Infrastructure2024Undisclosed, PE secondaryModular DCPE secondary
KKR / CoolIT (majority)2024UndisclosedThermal (cold plate + CDU)Trained organisation
Wingtech / Nexperia (UK Newport divested)Nov 2023 (divestment)Forced divestment on NSAWide-bandgapReversal of prior deal
Bain Capital / Proterial (Hitachi Metals TOB)Announced 2022, closed 2023~$6.4 billion equityMulti-layer (magnetics, thermal)PE take-private
JIC-led consortium / Shinko Electric TOBAnnounced 2023, ongoing~¥690 billion equityOn-package delivery (substrates)Sovereign + take-private
Vertiv / E&I (electrical integration)2024UndisclosedModular DCTrained organisation
Legrand / Server Technology (integration)Ongoing consolidationPrior deal, integratingModular DC (rack PDU)Certification




Two observations. First, the deals cluster by layer. Thermal shows the most volume and the largest single deal. Modular DC shows steady PE dealflow. On-package delivery is dominated by two large take-privates (Proterial, Shinko attempted) rather than strategic consolidation. Interconnect shows one very large deal (Encore Wire) and otherwise remains under-consolidated. The capacitor layer shows almost no M&A after Yageo/KEMET in 2020, which is itself informative about the layer's structural stability.


Second, the through-line diagnosis produces different implied multiples. Deals rooted in trained-organisation or installed-base logic (Eaton/Boyd, Vertiv/PurgeRite, KKR/CoolIT) priced at multiples that exceed comparable industrial services precedents. Deals rooted in qualification-field consolidation (Infineon/GaN Systems, TDK/Ferric) priced at what looks like a discount to standalone financials but a premium to the strategic peer set.



Chart 3. Deal count 2023-2026 by physical layer


Thermal has been the busiest layer by deal count. Modular datacenter and interconnect are close behind. On-package delivery has the lowest deal count because foundry ownership of the strategic value pool leaves little for consolidators to acquire. Capacitor is nearly quiescent since Yageo/KEMET in 2020.



Deal count from the ledger of material transactions. Non-material tuck-ins and financing rounds excluded.




The forward map: fifteen names for the next 24 months


The list below identifies fifteen firms that fit one or more of the four through-line patterns and are structurally positioned as likely acquisition candidates over the next 24 months. This is a list of hypotheses about strategic value, not a list of firms actively in process or reported to be so. Each is named for reasons that are visible in the public record.



TargetLayerLikely acquirer typeThrough-line

Empower Semiconductor (US, private)On-package delivery (IVR)Strategic (Intel, Broadcom, or Chinese domestic if regulator permits)Qualified field (last independent thin-film IVR)
Chilldyne (US, private)Thermal (negative-pressure cold plate)Strategic (Eaton, Schneider, or Vertiv)Certification + differentiated technology
Corintis (CH, private)Thermal (die-integrated microchannel)Strategic foundry (TSMC or Intel) or hyperscalerStructural threat elimination
JetCool Technologies (US, PE-backed)Thermal (microconvective)Strategic (Vertiv, Boyd inside Eaton)Trained organisation
Standard Fluids (US, private)Thermal (PFAS-free two-phase fluid)Strategic (Chemours, Dow, or specialty chemistry major)Regulatory positioning
Compass Datacenters (US, PE-backed)Modular datacenterIPO or secondary to larger infra fund; strategic less likely at scaleTrained development organisation
Aligned Data Centers (US, PE-backed)Modular datacenterIPO or secondary to larger infra fundTrained development organisation
EdgeConneX (US, PE-backed)Modular datacenter (edge and mid-tier)Secondary sale to infra fundPortfolio value
Delta Star (US, ESOP)Interconnect (MV transformer)Strategic (Hitachi Energy, Siemens Energy, or GE Vernova); PE less likely on ESOPCertification + US onshoring
Virginia Transformer (US, private)Interconnect (MV transformer)Strategic or PE-backed roll-upUS onshoring
MGM Transformer (US, family)Interconnect (MV transformer)Strategic or PE roll-upUS onshoring
LICAP Technologies (US, private)Capacitor (HSC/LIC)Strategic (Panasonic, Skeleton, or Musashi)Qualified field (dry-electrode IP)
Musashi Energy Solutions (JP, public)Capacitor (HSC/LIC)Strategic (Panasonic, Vertiv, or Eaton via power capacity play)Qualified field
Sungrow (grid services division) (CN, public)Modular DC (BESS + grid)Cross-border strategic (constrained by regulation)Qualified field
Ibiden or Kyocera (substrate carve-out) (JP, public)On-package delivery (organic substrate)PE take-private or strategic; JIC-supportedQualified field + trained org




Three of these fifteen are more consequential than the others. Empower Semiconductor is the most obvious strategic target on the list: the last independent thin-film IVR merchant, positioned in a market TDK has already partially consolidated by acquiring Ferric. The natural buyer set is small (Intel, Broadcom, or a Chinese domestic strategic if US-China conditions permitted), and the price will reflect the strategic squeeze rather than standalone financials.


The three US MV transformer firms (Delta Star, Virginia Transformer, MGM) are a distinct opportunity because US onshoring pressure produces political demand for domestic-owned capacity. A PE roll-up combining all three, or a strategic acquisition by GE Vernova or Prolec GE (already partly owned by GE Vernova), would consolidate the last independent US MV transformer capacity into a single owner. The regulatory and political dimensions of such a deal would be significant.


Compass, Aligned and EdgeConneX are approaching exit windows their PE owners cannot indefinitely defer. IPO into a receptive market is one path; secondary sale to a larger infrastructure fund is another. All three cannot exit simultaneously without saturating the demand pool, which means the sequencing and pricing of the first exit will set the trajectory for the rest.



Chart 4. Fifteen forward candidates by layer and through-line


The forward-candidate list clusters in thermal and interconnect (transformers). On-package delivery has one strategically important target (Empower). The distribution reflects both where consolidation has not yet completed and where the through-line pattern makes the deal likely.



Author's identification of candidates fitting one or more of the four through-line patterns and structurally positioned as likely acquisition targets. Not reported process activity.




The exit path problem


Three exit paths dominate the private-market side of the AI-infrastructure M&A landscape. Strategic sale to one of the six Western industrial majors (Eaton, Schneider, Vertiv, Legrand, ABB, Siemens Energy). IPO into a market that has become receptive to specialty industrials with AI exposure. Secondary sale to a larger infrastructure fund, typically Blackstone, Brookfield, DigitalBridge, KKR, EQT or Blue Owl.


Each has structural capacity limits. The six Western industrial majors can each absorb one or two large acquisitions per year without stretching integration capability. That is a maximum theoretical volume of around ten to twelve deals annually across the sector, which is at or below the current deal frequency and cannot expand meaningfully.


The IPO path opened in 2024 with Kokusai Electric's Tokyo re-listing and has since seen Rubrik, Astera Labs, several smaller Asian listings, and multiple filed-but-withdrawn attempts. Receptivity is real but selective, and requires either differentiated technology positioning or a clear AI-related revenue growth curve. Modular developers have not yet tested IPO receptivity at scale.


The infrastructure fund secondary market absorbs the largest deals but at cost of capital that compresses returns for the seller relative to strategic or IPO exits. Brookfield, DigitalBridge and Blue Owl are the most active buyers in this segment, with typical cheque sizes of $2-5 billion.


The combined capacity of all three paths is estimated at 25-35 material AI-infrastructure exits per year across the private market. The current PE portfolio of firms in the sector approaching typical hold-period end over 2026-2028 is approximately 50-70 firms. The math does not resolve cleanly. Something has to change: exits will be delayed, prices will compress at the point of sale, or a new exit path (secondary IPO market, sovereign direct purchase) will need to develop.



Chart 5. Exit-path capacity vs upcoming PE portfolio, annual through 2028


The three exit paths collectively absorb 25-35 material AI-infrastructure exits per year. The PE portfolio approaching hold-period end over 2026-2028 is 50-70 firms. The mismatch is the structural pressure on the exit market that Part V's PE playbook essay treats in detail.



Author estimates from Preqin fund vintage data, published exit-path precedents, and PE portfolio composition. Ranges reflect year-to-year variability.




Chart 6. Implied deal multiple by through-line


Through-line diagnosis produces meaningfully different implied multiples. Installed-base and trained-organisation deals command the highest premiums because the underlying cash flow is durable and defensible. Qualified-field consolidation prices low on standalone financials but embeds substantial strategic value that only crystallises post-transaction.



Approximate ranges from precedent transactions. Individual deals vary considerably; ranges reflect typical rather than extreme cases.




What breaks this map


The forward M&A landscape assumes financing conditions that permit strategic and secondary transactions at current valuations. If those conditions change, the map changes in specific ways.


Rising rates reprice PE-owned platforms downward, which either delays exits (holding assets through the rate cycle) or produces distressed sales at compressed multiples. The specific vulnerability sits in service-heavy platforms acquired at 12-16x EBITDA in 2021-2022, where refinancing risk begins to bind in 2026-2027.


An AI capex slowdown produces the demand-side shock discussed in Pricing Under ScarcityThe M&A implication is that acquisitions rooted in qualification-field consolidation lose value quickly (the field discipline that supported pricing evaporates when everyone has spare capacity), while acquisitions rooted in installed-base access hold value better (the annuity is already installed and requires maintenance regardless of new capex).


Regulatory intervention in dominant M&A pathways (competition authority action against the strategic buyers, or export-control restrictions on cross-border deals) would restrict several of the fifteen forward candidates. The Empower case is the most exposed to this risk: acquisition by a Chinese strategic would trigger CFIUS review with near-certain blockage; acquisition by a US or European strategic could trigger EU or UK competition review given the resulting duopoly structure.



What to watch, quarterly


Four things move the map faster than anything else in the current cycle. Strategic-buyer earnings commentary is the first. Eaton, Vertiv and Schneider each disclose acquisition appetite and integration bandwidth on quarterly calls, and a shift toward organic-growth language would compress the strategic exit path before any deal was announced. Filed but withdrawn IPOs are the second. Each successful IPO with AI-infrastructure exposure adds capacity to the IPO exit path; each withdrawal signals capacity has been reached at current pricing. Secondary-market pricing on infrastructure-fund purchases of PE-owned platforms is the third. Brookfield, DigitalBridge and Blue Owl transaction levels set a floor for other PE holders, and compression there signals the secondary path is filling. Regulator action on any headline transaction is the fourth and the one with the largest step-change impact. A single blocked deal in this cycle would deter the next wave and reprice several of the fifteen forward candidates immediately.




Glossary of terms used



CDU
Coolant Distribution Unit. Thermal system component distributing chilled coolant to racks or direct-to-chip cold plates.
CFIUS
Committee on Foreign Investment in the United States. US inter-agency review body that screens foreign acquisitions of US businesses for national security implications.
EBITDA
Earnings Before Interest, Tax, Depreciation, and Amortisation. The most commonly referenced operating-earnings metric in M&A pricing.
GaN
Gallium Nitride. Wide-bandgap semiconductor material used in RF and power applications.
HSC
Hybrid Supercapacitor. Energy storage technology bridging batteries and traditional capacitors.
IPO
Initial Public Offering. A private company's first sale of shares to the public market.
IVR
Integrated Voltage Regulator. On-package power delivery component for modern CPUs and GPUs.
MV
Medium Voltage. Typically 1 kV to 35 kV. The voltage class connecting data centre power distribution to the utility grid.
PDU
Power Distribution Unit. Rack-level device distributing power from datacentre supply to servers.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.





Sources and method


Historical M&A ledger compiled from company disclosures, PitchBook, Preqin, IJGlobal, Bloomberg terminal transactions data, Financial Times and Reuters reporting through mid-2026. Deal values are as reported at announcement or closing where available. Undisclosed values marked as such.


The Eaton/Boyd $9.5 billion figure is from Eaton's public disclosure at closing. Prysmian/Encore Wire from company disclosure and Reuters reporting. Vertiv/PurgeRite and Schneider/Motivair figures are as reported in trade press and not officially confirmed by either party.


Through-line categorisation is the author's construction, derived from observed strategic rationale across deals rather than from any published framework. Categorisations are judgements and any specific deal may draw from more than one pattern.


The forward-candidate list rests on public information about each named firm's ownership, competitive position, and strategic value in the acquirer set described. It should be read as analytical hypotheses rather than as reported process activity.



Series


The Investment Layer (eight parts). Part I: How the AI buildout is financedPart II (this piece): The M&A mapPart III: The build-to-lease flipPart IV: The coverage asymmetryPart V: The PE playbook per layerPart VI: Underwriting AI infrastructure debtPart VII: The three downside casesPart VIII: Sovereign capital and industrial policyCompanion to the AI Power Chain series (six parts plus six technical companions), and the earlier commercial supplements Pricing Under Scarcity and The Services Inversion.


The author is an independent advisor working on data centre, power and industrial technology. This analysis is written in a personal capacity and rests entirely on public information. Nothing here is investment advice.







See also · The Investment Layer

 * How the AI buildout is financed
 * The build-to-lease flip
 * The coverage asymmetry





Further reading + sources
Related essays

 * Case Study 01. Blackstone / QTS five years on. the reference LBO of the AI era
 * Case Study 04. CoreWeave 18 months post-IPO. the AI-native operator reference
 * Case Study 06. Digital Realty / Teraco. the emerging-market platform-via-M&A model
 * How the AI buildout is financed. the capital-structure map that funds these deals


Primary sources

 * SEC EDGAR. primary filings for the transactions referenced
 * Reuters M&A coverage. transaction announcements and coverage







In brief


AI infrastructure M&A activity 2023-2026 covers four through-lines. Eaton/Boyd Thermal ($9.5B, closed March 2026 per company press release) is the largest deal on the map. Ecolab acquired CoolIT from KKR for $4.75B (announced March 2026). Schneider took majority in Motivair. Vertiv built cooling capability through capability tuck-ins rather than a headline DLC acquisition.


==============================================================================
# The build-to-lease flip
URL: https://adikumar.co/the-build-to-lease-flip/
Published: 2026-08-12
Summary: Why hyperscalers stopped owning their datacenters, the developer stack that absorbed the capital demand, and where the return arithmetic breaks.
==============================================================================

The Investment Layer series · Part 3 of 8








The Investment Layer · Part III of VIII



The build-to-lease flip


Why hyperscalers stopped owning their datacenters, why the private developer stack now finances the majority of new capacity, and where the return arithmetic breaks.




The Investment Layer

 1. Part I. How the AI buildout is financed
 2. Part II. The M&A map
 3. Part III. The build-to-lease flip (you are here)
 4. Part IV. The coverage asymmetry
 5. Part V. The PE playbook, layer by layer
 6. Part VI. Underwriting AI infrastructure debt
 7. Part VII. The three downside cases
 8. Part VIII. Sovereign capital and industrial policy




Executive summary


 * The share of hyperscale datacenter capacity that is leased rather than owned has moved from roughly 40 percent in 2019 to an estimated 65-70 percent in 2026 measured by new-build MW additions. Owned share is falling in absolute terms as well as proportionately.
 * The reason is straightforward. Owning consumes hyperscaler balance-sheet capacity that has become the binding constraint on top-line growth. Leasing shifts the capital outlay to a third party, converts it to a lease payment recognised over the contract life, and preserves the balance sheet for chips, software talent and R&D.
 * The developer stack that absorbs this capital demand comprises approximately fifteen firms, of which roughly ten are PE-owned or PE-backed, three are publicly-listed REITs (Equinix, Digital Realty, KDDI Telehouse), and the balance are sovereign-backed or joint ventures.
 * The economics for the developer rest on a 10-to-15 year lease with a hyperscaler tenant at power price plus operating margin plus real-estate margin. Unlevered returns run 9-11 percent; levered returns 13-16 percent after financing.
 * The break points are three. Rising interest rates compress levered returns and reduce refinancing capacity mid-lease. A shift in hyperscaler procurement toward BYOC (bring-your-own-construction) reduces the developer's role to construction management with thinner margins. Grid-connection queue changes reprice site option value.
 * Two second-order effects matter for the strategic map. Firstly, the developer stack has become the dominant buyer of specialty industrial equipment covered in earlier essays, which reshapes vendor commercial strategy. Secondly, developer M&A now sets the market clearing price for grid-connected land more than land brokers do.




Chart 1. Hyperscale capacity: owned vs leased share of new-build MW, 2019-2026


Owned share of new-build hyperscale capacity has moved from about 60 percent in 2019 to about 30 percent in 2026. Leased is now the default. Note that this is share of new-build; installed base is still closer to 50/50 because much was built pre-2020.



Author estimates triangulated across CBRE, JLL, Structure Research, DC Byte industry data. 5-8 pp margin between sources on any single year.




What flipped, and how much


Hyperscale datacenter capacity worldwide has moved from majority-owned to majority-leased over the seven years from 2019 to 2026. The direction of the shift is documented across multiple industry data sources (CBRE, JLL, Structure Research, DC Byte). The specific magnitudes disagree by 5-8 percentage points depending on how each source treats co-located and hybrid deployments, but the direction and scale of the shift are consistent.


Approximate figures for the four major hyperscalers combined, weighted by MW commitment across all deployment types:



YearOwned share of new MWLeased share of new MWDirectional note

2019~60%~40%Owned was clear default
2021~50%~50%Crossover year, driven by initial COVID capex acceleration
2023~40%~60%Post-ChatGPT, hyperscaler procurement pivoted to lease-first at speed
2025 (est.)~30%~70%Owned share below one-third for the first time
2026 (year-to-date)~30%~70%Broadly stable at 70 percent leased




Two clarifications on how to read these figures.


The percentages are of new-build MW additions in each year, not of total installed base. The installed base still contains a higher owned proportion because much of it was built pre-2020. Total installed capacity across the four majors is estimated at roughly 45-50 percent owned in 2026.


Owned share is falling in absolute terms, not just proportionately. Amazon's Northern Virginia owned campus additions of 2018-2020 have no directly comparable equivalent in 2025-2026 planning. Google similarly has slowed its owned Council Bluffs and The Dalles expansions in favour of Georgia, South Carolina and Ohio leased capacity. Microsoft has slowed Quincy and San Antonio owned expansion in favour of Wisconsin, Georgia and Malaysia leased. Meta, which had been the least dependent on lease historically, moved to lease-first for the entire Llama-serving fleet in 2024.



Why the flip happened


The published rationale hyperscalers give in earnings calls emphasises speed to power. Leased sites often have earlier grid interconnection dates than what hyperscaler internal siting teams can achieve, because independent developers began pre-securing grid-queue positions and power purchase agreements in 2021-2022 for delivery in 2025-2027.


The published rationale is real but partial. Two structural factors matter more.



Factor 1. Balance-sheet capacity has become the binding growth constraint


Meta, Google, Microsoft and Amazon each have quarterly capital expenditure budgets in the $20-30 billion range as of mid-2026. This spend must be split across chips (dominant), datacenter shell and fit-out, software and talent, and other. The share allocated to chips has risen sharply post-2023 because chip cost per unit of AI capacity has risen (H100 to B100 to R100 progression) faster than shell cost per MW has risen.


Rather than let chip spend crowd out shell spend, the hyperscalers have off-balance-sheet financed the shell. A leased datacenter recognises as an operating lease on the P&L (with straight-line rent expense over the lease life) but does not consume balance-sheet capacity that would otherwise be used for chip purchase.


The math a hyperscaler CFO faces is stark. A 100 MW owned datacenter requires roughly $1.5-2 billion of construction capital outlay concentrated in a 24-month build period. The same 100 MW leased under a 15-year contract at $150 per kW-month generates a lease commitment of approximately $2.7 billion over the lease life, or $180 million per year, versus zero balance-sheet consumption on day one. For a hyperscaler adding 5-10 GW of new capacity per year, the balance-sheet consumption of owning all of it is prohibitive.



Chart 2-100 MW datacenter: owned vs leased, hyperscaler capital view


Owned consumes $1.6 billion of construction capital in months one through 24. Leased consumes zero on day one and generates $186 million of annual operating lease expense against the same capacity. For a hyperscaler running 5-10 GW of new capacity per year, that difference is the entire growth constraint.



Illustrative unit economics at 2026 build cost benchmarks (Turner & Townsend, JLL). Individual projects vary.




Factor 2. Speciality construction organisation cannot be built at pace


Building a datacenter at 100 MW scale requires an integrated construction organisation with MV electrical trades, medium-voltage transformer procurement, chilled water plant, thermal loop commissioning, security fit-out, and hundreds of qualified staff. Hyperscalers have such organisations but they are sized for a specific capacity trajectory. Doubling that trajectory requires either doubling the internal organisation (impossible on the required timescale) or contracting through developers who have already built comparable organisations.


The largest datacenter developers (Digital Realty, Equinix, Compass, Aligned, STACK, QTS, Vantage, DataBank) each operate construction organisations of 500-2000 qualified staff. Collectively they represent perhaps 8000-10,000 trained datacenter construction specialists that a hyperscaler can access via lease contract but could not access via direct hire on any relevant timeframe.



Factor 3. The developer stack was ready


None of this flip would have happened if the developer stack had not already existed at scale. What Blackstone, Brookfield, DigitalBridge, Blue Owl, KKR, GIC, Temasek and CPP Investments did through 2018-2022 was build out an institutional platform of scale developers, without much public attention, into which hyperscaler demand could be dropped. When post-ChatGPT demand hit, the platform absorbed it.


The economic history point matters. This is not a spontaneous re-organisation; it is the maturing of a private infrastructure asset class that took a decade of institutional-capital construction. Take away the ten years of PE and infrastructure-fund investment in developer scale-up, and there is no way for hyperscalers to have shifted to 70 percent lease over three years.



The developer stack, sized


The following table summarises the fifteen developers that collectively hold the majority of leased hyperscale capacity as of mid-2026. Sizings are approximate and drawn from public disclosures, industry reports, and reasonable inference from disclosed capacity and pipeline figures.



DeveloperOwnershipApprox. deployed MWApprox. pipeline MW

Digital RealtyListed REIT (US)~3,500~2,500
EquinixListed REIT (US)~2,800~1,400
Vantage Data CentersDigitalBridge / Silver Lake / others~2,000~2,200
Compass DatacentersBrookfield~1,600~2,400
Aligned Data CentersMacquarie Asset Management~1,500~2,200
STACK InfrastructureBlue Owl~1,400~2,000
CyrusOneKKR / GIP~1,300~1,600
QTS Data CentersBlackstone Infrastructure~1,900~2,300
DataBankDigitalBridge / Swiss Life / others~600~1,000
Iron Mountain Data CentersListed REIT (US)~450~700
Prime Data CentersICONIQ / others~500~1,200
Yondr GroupDigitalBridge (secondary)~400~1,300
NTT Data CentersNTT Group (JP corporate)~1,200~1,000
KDDI TelehouseKDDI (JP corporate)~350~500
Chindata (sovereign-backed)China SDIC~1,100 (China only)~1,500




The combined deployed capacity of this fifteen-firm group is approximately 20,600 MW; combined pipeline approximately 24,100 MW. Adding the pipeline to the deployed base implies the developer stack is on track for roughly 44,700 MW of deployed capacity by 2028, subject to construction and interconnection risk. Not all pipeline MW will be built on the disclosed schedule.



Chart 3. Fifteen developers ranked by deployed MW, plus pipeline


Digital Realty and Equinix lead deployed capacity. QTS, Vantage, Compass, Aligned and STACK are close behind on a deployed basis and lead on pipeline. Combined deployed capacity across the fifteen is approximately 20.6 GW; combined pipeline adds another 24.1 GW.



Deployed and pipeline figures from company disclosures, CBRE and JLL data centre reports, DC Byte and Structure Research through mid-2026.



Ten of the fifteen are PE-owned or PE-backed. Three are listed REITs. Two are Japanese corporate. One is sovereign-backed. The concentration of ownership in PE hands is the point that matters for the exit-path problem discussed in Part II and for the debt-underwriting mechanics discussed in Part VI.


Ten years of PE and infrastructure-fund investment in developer scale-up made the hyperscale lease flip possible. Without the pre-built platform, the pivot to leased could not have happened at the pace it did.The pre-condition often ignored



The developer return arithmetic


The economics of a leased datacenter for the developer are straightforward on paper. The developer signs a 10-to-15 year triple-net lease with a hyperscaler tenant. The rent structure is typically MW-based (dollars per kW per month, escalating annually at CPI or a fixed 2-3 percent). Power is a pass-through. The developer holds ownership and provides fit-out and operations against a management fee.


Illustrative unit economics for a 100 MW facility, current market terms:



Line itemValueNote

Total construction cost$1.6 billion (illustrative)~$16 million per MW at 2026 build cost per industry benchmarks
Debt financing at 60% LTV$960 millionTypical for stabilised infrastructure asset
Equity requirement$640 millionDeveloper plus infrastructure fund LP
Rent per kW-month$155Blended average for hyperscale primary market, 15-year lease
Annual rent revenue$186 million100,000 kW × $155 × 12
Operating expense (excl. rent)$18 millionStaffing, maintenance, insurance, tax
Debt service at 5.75%$55 millionInterest only for first 5 years, then amortising
Levered free cash flow, year 1-5~$113 millionApproximate, pre-tax
Levered cash-on-cash yield~17.7%On $640M equity, year 1-5
Unlevered IRR over 15-year lease~10.5%Assumes 5.5x exit multiple on cash at year 15 or refinance
Levered IRR over 15-year lease~14.5%With mid-lease refinance capacity




These figures rest on several sensitivities. Debt cost matters most. A 100 basis point rise in debt cost cuts levered IRR by roughly 3 percentage points. Rent escalation matters second. A shift from 3 percent annual escalation to CPI-linked (assuming 2.2 percent CPI) cuts levered IRR by roughly 1.5 percentage points. Refinancing capacity in year 5 is the third sensitivity: if LTV cannot be raised at refinance because valuation has compressed, the levered return drops to roughly the unlevered figure.



Chart 4. Levered IRR sensitivity to debt cost, 100 MW illustrative


The levered IRR curve is steep in debt cost. At current market debt cost (5.75 percent), the illustrative case delivers 14.5 percent levered IRR. A 200 bp rise in refinance conditions compresses that to around 8 percent, at which point the levered return no longer justifies the equity exposure over public alternatives.



Illustrative sensitivity holding all other assumptions constant. Individual transactions vary considerably.




Where the arithmetic breaks


Three factors reprice the developer economics. Each has been visible in the market in 2025-2026 and any of them can move materially in either direction.


Rising debt cost. The debt cost assumed in the illustrative case is 5.75 percent, which is close to current market for a stabilised hyperscale-tenanted asset. If refinance rates in 2029-2031 (when the current wave of 5-year interest-only structures mature) run 100-200 basis points higher than initial, the levered return compresses meaningfully. Developers with earlier refinance windows are exposed first. Blackstone and Brookfield have pre-financed at longer tenors to reduce this exposure; smaller developers have not always done so.


BYOC procurement shift. Some hyperscalers have begun contracting bring-your-own-construction arrangements, where the hyperscaler procures the developer purely for site rights, permitting and construction management, and installs the electrical and mechanical systems themselves. This reduces the developer's role to a slimmer construction-management fee model, with equity IRRs in the 6-9 percent range rather than 12-16. Meta and Google have both piloted BYOC in 2025. If this becomes the dominant model, developer returns compress structurally.


Grid interconnection queue changes. The developer economics rest on the option value of grid-connected sites. If ISO/utility queue reform brings forward hundreds of previously-queued sites to comparable interconnection dates, the scarcity premium on any single developer's portfolio compresses. FERC Order 2023 and PJM interconnection reform proposals both point in this direction. The magnitude and timing of impact are uncertain but the direction is clear.



Chart 5. Three break points: impact on developer levered IRR


The three structural risks each compress developer levered IRR by different magnitudes. Rate persistence at +200 bp compresses IRR the most; BYOC adoption reduces the developer's role to construction management with structurally lower returns; grid queue reform is the least immediate but the longest-duration compression because it reprices the entire option value of grid-connected land.



Estimated levered IRR outcome after each risk scenario. Base case is 14.5 percent from the illustrative unit economics table.




Two second-order effects that matter for the strategic map



Effect 1. Developers are now the dominant procurement channel for AI-infrastructure equipment


The commercial implication of the lease flip that vendors are still adjusting to. Historically hyperscalers procured all electrical, thermal and rack-scale equipment directly from vendors, with developer input limited to construction and site management. As the leased proportion has risen, an increasing share of procurement has shifted to the developer. Developers now buy transformers, switchgear, UPS, CDU, cold plates, cabling, and rack systems on behalf of hyperscaler tenants, subject to hyperscaler specification approval.


This changes vendor commercial strategy. The customer relationships that mattered pre-2022 were hyperscaler procurement heads. The customer relationships that matter in 2026 are developer procurement heads (and their delegated procurement organisation) plus hyperscaler technical specification teams. Vendors that have not adapted their sales organisation to this two-headed structure have lost share. Eaton, Vertiv and Schneider have all restructured their commercial approach; the smaller specialist vendors mostly have not.



Effect 2. Developer M&A sets the market clearing price for grid-connected land


When Compass, Aligned or STACK acquires a piece of grid-connected land in a primary market, the price becomes a comparable for all subsequent transactions in that market. The developer bid price rests on the developer's model of the lease value that can be attached to the site, plus construction cost, less financing terms.


Traditional real-estate valuation of grid-connected land relied on comparable transactions among industrial users. That comparison set has become non-informative post-2023. Developer bid prices are 3-8x traditional industrial land prices in primary markets (Northern Virginia, Phoenix, Dallas, Columbus, Chicago). Land brokers now increasingly market to developers first and traditional industrial users second. This has spillover effects on manufacturing site selection for firms competing for the same land.



Chart 6. Developer bid price premium over industrial-user comparables, primary markets


Grid-connected land in primary datacenter markets now trades at 3-8x traditional industrial-user comparables. The multiple varies by market: Northern Virginia at the top of the range, Chicago and Columbus at the low end. The premium is the option value of interconnection that only a datacenter developer can monetise.



Estimated ranges from developer land acquisition disclosures, market commentary from CBRE and JLL, and industrial land comparables. Individual transactions vary.




What to watch


Hyperscaler earnings-call guidance on the lease-versus-own split is the leading indicator. Direct signalling of a shift back toward owned would represent a reversal of the flip; nothing indicates that is likely in the near term but the guidance is where any such shift would first appear. New lease contract terms as they are signed and reported are the second signal. Rising rents mean power scarcity intensifying and are bullish for developer economics, falling rents or shorter tenors mean hyperscaler negotiating leverage rising and are bearish. BYOC pilot outcomes are the third and the highest-consequence. If Meta and Google's 2025 experiments deliver on time and cost matching the developer-model comparators, adoption expands and the developer role compresses to a construction-management fee model. Finally, refinance completions on 2020-2022 vintage developer debt as it hits maturity will indicate whether the underlying developer economics are still supporting current market debt pricing, and Part VI treats those mechanics in detail.




Glossary of terms used



CDU
Coolant Distribution Unit. Thermal system component distributing chilled coolant to racks or direct-to-chip cold plates.
GW
Gigawatt. One thousand megawatts of electrical power.
ISO
Independent System Operator. Regional grid operator (e.g. ERCOT, MISO, PJM, CAISO). Manages transmission access and wholesale market operations.
LP
Limited Partner. Investor in a PE or infrastructure fund who commits capital but does not manage day-to-day investment decisions.
MV
Medium Voltage. Typically 1 kV to 35 kV. The voltage class connecting data centre power distribution to the utility grid.
MW
Megawatt. Unit of electrical power. A large modern data centre draws tens to hundreds of MW.
PJM
PJM Interconnection LLC. Regional grid operator covering 13 US states plus DC. Includes Ohio, where the NVIDIA/OpenAI campus sits.
UPS
Uninterruptible Power Supply. Battery-backed power system that keeps critical loads running during grid disruptions.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.





Sources and method


Deployed capacity and pipeline figures drawn from company disclosures, CBRE and JLL data center reports, Structure Research and DC Byte industry data through mid-2026. Lease share estimates triangulated across industry sources with acknowledged 5-8 percentage-point margin between sources.


Unit economics illustrative case constructed from a synthesis of publicly disclosed lease terms, published construction cost benchmarks (Turner & Townsend and JLL cost data), and financing structures reported by REITs and infrastructure funds. Individual transactions vary considerably around the illustrative case.


BYOC discussion rests on company earnings-call statements and reported deal structures. Neither Google nor Meta has published detailed BYOC economics; the discussion here reflects industry observation rather than official disclosure.



Series


The Investment Layer (eight parts). Part I: How the AI buildout is financedPart II: The M&A mapPart III (this piece): The build-to-lease flipPart IV: The coverage asymmetryPart V: The PE playbook per layerPart VI: Underwriting AI infrastructure debtPart VII: The three downside casesPart VIII: Sovereign capital and industrial policyCompanion to the AI Power Chain series and to earlier commercial supplements Pricing Under Scarcity and The Services Inversion.


The author is an independent advisor working on data centre, power and industrial technology. This analysis is written in a personal capacity and rests entirely on public information. Nothing here is investment advice.







See also · The Investment Layer

 * How the AI buildout is financed
 * The AI infrastructure M&A map
 * The coverage asymmetry







In brief


The build-to-lease flip is how hyperscalers convert data centre construction capex into operating-lease liability. A hyperscaler funds a developer to build to spec, then leases capacity under long-dated contracts. The developer holds the asset on balance sheet. The hyperscaler avoids the capex line and gets bespoke design at scale. Economics turn on cost-of-capital spread and lease structure.


==============================================================================
# The coverage asymmetry
URL: https://adikumar.co/the-coverage-asymmetry/
Published: 2026-08-13
Summary: Where equity research concentrates on AI infrastructure, where it misses the material stories, and what the coverage gap says about mispricing.
==============================================================================

The Investment Layer series · Part 4 of 8








The Investment Layer · Part IV of VIII



The coverage asymmetry


Where equity research and sell-side attention have concentrated, where they have missed the material stories, and what the coverage gap tells you about where mispricing sits.




The Investment Layer

 1. Part I. How the AI buildout is financed
 2. Part II. The M&A map
 3. Part III. The build-to-lease flip
 4. Part IV. The coverage asymmetry (you are here)
 5. Part V. The PE playbook, layer by layer
 6. Part VI. Underwriting AI infrastructure debt
 7. Part VII. The three downside cases
 8. Part VIII. Sovereign capital and industrial policy




Executive summary


 * Sell-side equity research on AI infrastructure has concentrated in three narrow zones and left five material zones under-covered. The concentration reflects analyst incentives more than the distribution of value.
 * The three heavily covered zones are NVIDIA and its immediate accelerator peers, hyperscaler capex commentary, and the four listed US datacenter REITs. Each has fifteen to twenty active sell-side desks producing regular research.
 * The five under-covered zones are the wide-bandgap semiconductor supply chain, the on-package power delivery sub-industry, the thermal fluid and cold plate manufacturing base, the medium-voltage transformer and electrical equipment specialist tier, and the modular datacenter developer sub-sector below the top four names.
 * The coverage gap correlates with private-market valuation compression relative to public-market valuations of comparable-quality assets. Wide-bandgap specialists trade in private secondaries at 8-12x forward EBITDA against public power-semiconductor peers at 22-28x.
 * Two structural reasons drive the gap. First, most of the interesting firms in the five under-covered zones are private, foreign, or too small for the mandate size of the desks that would otherwise cover them. Second, the technical content requires domain expertise (materials science, thermal engineering, medium-voltage electrical) that sell-side desks have not staffed for.
 * Two consequences follow. The first is that private-market buyers with domain expertise face limited competition in the under-covered zones and can transact at compressed multiples. The second is that publicly-listed firms in the covered zones price to consensus that reflects a narrower view than the underlying strategic map, and can be more volatile than the fundamentals warrant.
 * The coverage asymmetry has begun to narrow in specific pockets in 2026 (thermal cold-plate coverage has picked up after the Eaton/Boyd deal), but the wide-bandgap and MV transformer zones remain materially under-covered.




Chart 1. Analyst coverage by AI-infrastructure zone, average analysts per representative firm


Analyst count concentrates in NVIDIA, hyperscalers and the listed REITs. Wide-bandgap, MV transformers, on-package power, and modular developers below the top four receive a fraction of the analyst attention their strategic weight would justify.



Counts from FactSet Estimates and IBES coverage lists, mid-2026, using active-coverage inclusion criteria.




The three heavily-covered zones



Zone 1. NVIDIA and immediate accelerator peers


NVIDIA has approximately 45 active US sell-side analysts producing regular research as of mid-2026, with meaningful coverage from another twelve European and Asian desks. Coverage includes AMD (about 30 analysts), Broadcom (about 28), TSMC (about 22 counting US ADR coverage), and to a lesser extent Marvell, Astera Labs, Arm, and Rambus.


The volume of coverage is proportional to trading volumes, revenue base, and index weight rather than to complexity or breadth of the underlying supply chain. NVIDIA's revenue base of $180 billion trailing twelve-month justifies the coverage volume on any conventional metric. The point is that analyst attention crowds into the largest cap-weighted names in a way that leaves substantially less bandwidth for anything else.



Zone 2. Hyperscaler capex commentary


Meta, Google, Microsoft and Amazon each have 40-50 active sell-side analysts producing research primarily focused on their consumer and cloud businesses, with infrastructure capex commentary as a secondary but material discussion point. Analyst notes typically discuss quarterly capex figures, allocation between chips and shell, and the general trajectory of AI-related capital allocation.


The depth of the capex analysis varies. Some desks (Bernstein, MoffettNathanson, Morgan Stanley) produce genuinely detailed capex build-ups reasoning from datacenter capacity, per-MW cost, chip mix, and networking spend. Others rely on the reported figure without decomposition. Either way, the discussion is dominated by the four names and rarely extends downstream into the specific vendors that receive the capex.



Zone 3. The four listed US datacenter REITs


Digital Realty and Equinix each have 18-22 active sell-side analysts. Iron Mountain (partly a datacenter REIT) has about 12. American Tower (formerly hosting a datacenter business, now largely divested) still receives commentary from about 15. These four are effectively the entire US public-equity exposure to the datacenter developer stack, and consequently receive concentrated analyst attention proportional to their scarcity value as public proxies for a private-market industry.


The analytical depth here is often high. The REITs disclose leasing volume, rent per kW-month, occupancy, and interconnection revenue in enough detail to permit granular analysis, and analysts have staffed up to do it. The problem is not the quality of the coverage; it is that it stops at the REIT boundary and does not extend to the ten-plus PE-backed developers that hold a larger share of the market between them.



The five under-covered zones



Zone 4. Wide-bandgap semiconductor supply chain


Silicon carbide substrate, epitaxy, device and packaging vendors receive coverage that is thin relative to the strategic weight of the segment. Wolfspeed has about 14 active sell-side analysts (reduced from 20 pre-2024 after the Chapter 11 filing and reorganisation). II-VI/Coherent has about 10. Onsemi and STMicroelectronics have broad semiconductor coverage of about 22-25 each but SiC-specific analytical detail is limited to a subset of the analyst pool.


The private-market wide-bandgap tier (GaN Systems pre-Infineon, Transphorm pre-Renesas, Ferric pre-TDK, Empower, Chinese substrate vendors) has almost no independent sell-side coverage. Coverage of private firms in specialty semiconductors is typically limited to buy-side desks with private-market mandates, which are a fraction of the sell-side analyst pool.


The consequence is a systematic private-market discount for wide-bandgap specialists. Ferric transacted at what is understood to be $250 million (approximately 8-10x forward EBITDA); Empower is expected to transact at a comparable multiple if a strategic buyer emerges. Public-market power semiconductor peers (Infineon, STMicroelectronics) trade at 22-28x forward EBITDA on the segment.



Chart 2. Public vs private-market EBITDA multiple, wide-bandgap and on-package specialists


Public-market power semiconductor peers trade at 22-28x forward EBITDA on their power segments. Comparable private-market wide-bandgap and IVR specialists transact at 8-12x. The gap is the coverage-driven mispricing that specialist private-market buyers can exploit.



Estimated multiples from disclosed transactions (Ferric, Infineon/GaN Systems, Renesas/Transphorm) and public-market comparables (Infineon, STMicroelectronics, Onsemi power segments).




Zone 5. On-package power delivery


The on-package power delivery sub-industry (IVR magnetics, thin-film substrates, high-density capacitors integrated into the accelerator package) has almost no dedicated sell-side coverage as a category. Ferric, Empower, and comparable specialists are covered incidentally in broader semiconductor coverage rather than as members of a distinct sub-industry.


This is one of the more mispriced areas of the map. As the technical companion to The On-Package Delivery Stack laid out, on-package power delivery has become the binding constraint on further accelerator progress, and the qualified supplier field is very narrow. But sell-side coverage frames the segment as an incidental line item within IC packaging or within specialty magnetics, both of which are covered thinly at the specialist level.



Zone 6. Thermal fluid and cold plate manufacturing


The Eaton/Boyd Thermal deal at $9.5 billion (closed March 2026 per Eaton press release) brought material sell-side attention to the thermal cold plate sub-industry for the first time. Coverage has expanded from perhaps 3-4 desks covering thermal management in an incidental fashion to approximately 12-15 now covering the sub-industry as a distinct category, either as part of Eaton coverage or as part of expanded Vertiv and Schneider Electric analyst notes.


Substantial gaps remain. Thermal fluid chemistry vendors (3M's discontinued Novec position, Chemours, Daikin, Dow), which underpin two-phase cooling, are not covered as an AI-adjacent category. PFAS regulatory risk to certain thermal fluid formulations is not discussed in mainstream chemistry-sector coverage. Chinese cold-plate manufacturers (which supply between 60 and 75 percent of global unit volume by our estimate) have essentially zero US or European coverage.



Zone 7. Medium-voltage transformers and specialty electrical


The MV transformer sub-industry (Prolec GE, Hitachi Energy, Siemens Energy, Delta Star, Virginia Transformer, Wilson, MGM, ERMCO, others) is one of the most under-covered zones on the map given its strategic importance. Hitachi Energy is covered indirectly through Hitachi Ltd. Siemens Energy has about 18 European sell-side analysts. Prolec GE is inside GE Vernova coverage but not broken out. The three specifically-US independents (Delta Star, Virginia Transformer, MGM) have essentially no coverage.


Coverage that does exist typically frames transformers as a commoditised electrical product rather than as a critical bottleneck to hyperscale interconnection. Lead times of 24-36 months, order backlogs of 3-4 years, and pricing power comparable to the wide-bandgap sub-industry are visible in trade press and industry reports but rarely quantified in the sell-side notes on the broader industrial names that contain the exposure.



Zone 8. Modular datacenter developers below the top four


Beyond the four listed REITs (Digital Realty, Equinix, Iron Mountain, American Tower legacy), the modular datacenter developer sub-industry receives sell-side coverage only via the parents that hold controlling stakes (DigitalBridge, KKR, Blackstone, Brookfield). None of Compass, Aligned, STACK, CyrusOne, Vantage or the smaller PE-backed developers is covered as a standalone entity. The parents' equity holders receive coverage on a portfolio basis that summarises the developer contribution.


This becomes materially consequential when the exit-path question discussed in Part II starts to bind. If Compass or Aligned files for IPO in the next 24 months, initiation coverage will be published, but no continuous coverage exists today.



Chart 3. Coverage intensity vs strategic importance, eight AI-infrastructure zones


Zones cluster into two groups. The heavily covered zones (top-right) sit at roughly the ratio their strategic importance would predict. The under-covered zones (below the diagonal) have strategic importance well above their coverage. That gap is where systematic mispricing lives.



Coverage intensity = analyst count normalised to zone maximum. Strategic importance = author's subjective ranking based on decomposition of value pools in the AI Power Chain series.




Why the gap exists


Two structural factors keep the five under-covered zones under-covered.



Analyst staffing rests on public tradability


Sell-side coverage is funded by trading commissions plus investment banking fees. Both scale with the underlying tradable market cap. A vendor with $200-800 million in revenue that is private, family-owned or PE-backed generates neither trading commissions nor near-term banking fees for a sell-side desk. Coverage cannot be justified against the desk's cost structure, so it does not happen.


The result is that a large and strategically important slice of the AI-infrastructure chain (wide-bandgap specialists, IVR magnetics, thermal fluid chemistry, MV transformer independents, modular developer non-REITs) sits below the coverage threshold on tradability grounds even though the underlying importance is high. This is a structural feature of how equity research is organised rather than an oversight that will self-correct.



Technical content requires domain expertise most desks have not staffed for


The wide-bandgap sub-industry requires basic materials-science literacy in silicon carbide crystal growth, epitaxy defect physics, and packaging thermal management. The MV transformer sub-industry requires basic understanding of power system engineering. The on-package power delivery sub-industry requires understanding of magnetic materials, thin-film manufacturing, and IC packaging. Very few sell-side analyst desks have staff with these backgrounds. Where they do exist, the staff are typically in Asia-Pacific desks with limited reach to US and European institutional buy-side clients.


The consequence is that coverage that does exist frames the technical content at a level that a generalist can follow, which is often a level below what the strategic significance requires. Trade press (Data Center Frontier, EE Times, Semi Wiki, Compound Semiconductor) contains materially better technical framing than most sell-side notes on the same topics.



Chart 4. Coverage evolution 2020-2026, selected zones


Coverage on thermal cold plate expanded sharply after the Eaton/Boyd deal announcement in late 2025. On-package power delivery has crept up slowly on Ferric/Astera visibility. Wide-bandgap coverage actually declined post-Wolfspeed Chapter 11. MV transformers remain flat and heavily under-covered.



Rolling analyst count including specialist coverage on named vendors within each zone. Sources as above.




The mispricing this produces


Two systematic mispricings follow from the coverage asymmetry.


Private-market compression in the under-covered zones. The natural buyer pool for wide-bandgap specialists, IVR magnetics vendors, thermal fluid houses, and MV transformer independents is a specialist strategic (Infineon for GaN, TDK for magnetics, Chemours for fluids, Hitachi Energy for transformers) plus a handful of infrastructure-focused PE funds with domain teams (Bain Capital, EQT, KKR's industrials group). This is a small buyer pool competing for a small number of transactions. Private-market clearing prices reflect the buyer-pool composition rather than any consensus view of intrinsic value, and typically discount 30-50 percent to what a competitive process could produce.


The Ferric transaction at reported $250 million against a strategic value that plausibly ran to $500-700 million is the reference case for this dynamic. The buyer pool that was structurally interested at the time (TDK, one Japanese peer, one Chinese strategic that could not transact for CFIUS reasons) was small enough that the seller could not run a competitive process.



Chart 5. Buyer pool size vs achieved price, precedent transactions in under-covered zones


Each dot is a transaction. Where the structurally-interested buyer pool was small (1-3 buyers), realised price ran 30-50 percent below reasonable strategic value. Where the pool was larger (5+), realised price approached full value. The Ferric case sits at the bottom-left of the pattern.



Author estimates from published deal values and strategic-value assessments of comparable transactions.



Public-market volatility on incomplete framing. The publicly-listed firms in the covered zones (Vertiv, Eaton, Schneider Electric, ABB, Legrand, Digital Realty, Equinix) trade to a consensus that reflects the narrower coverage view rather than the strategic map. Vertiv in particular has traded through several 30-40 percent swings in 2024-2026 on shifts in cooling-demand estimates that were smaller than the underlying supply chain and services annuity would imply.


The mispricing is asymmetric: consensus underestimates the value of installed-base annuity and overweights the value of one-time equipment sales. When headlines refocus attention on annuity (Eaton/Boyd deal, Vertiv earnings, Schneider services growth), the public equities re-rate sharply. Investors with a fuller framing of the strategic map can position ahead of these re-rates but the timing is not predictable.


The most interesting parts of the map are the least covered. Private-market buyers with domain expertise are systematically transacting there at compressed multiples for exactly that reason.Coverage gap as opportunity map



Where the gap is beginning to narrow


Two developments in 2025-2026 have partially narrowed the coverage gap. Each is instructive about what would need to happen to narrow it further.


The Eaton/Boyd Thermal deal at $9.5 billion (per company press release) made the thermal cold plate sub-industry a mainstream analyst topic for the first time. Vertiv, Schneider and Eaton earnings calls in 2025-2026 have all been asked detailed thermal capacity and installed-base questions by sell-side analysts who would not have engaged the topic in the prior three years. The deal size was what forced the reframing.


The listing of Astera Labs in 2024 and its subsequent coverage by 12 desks brought material analytical attention to the retiming and PCIe fabric segment of the AI infrastructure chain. Astera has since become a reference point for how sell-side desks discuss the interconnect and fabric layer. This is coverage that had been non-existent pre-IPO and is now moderate.



Chart 6. Analyst coverage jump following a large deal or specialist IPO


Two case studies. Cold plate coverage tripled within 6 months of the Eaton/Boyd announcement. Retiming and PCIe fabric coverage jumped from zero to 12 within 4 months of the Astera Labs IPO. Coverage narrowing consistently trails a triggering event by 3-6 months.



Analyst count observed monthly around each triggering event. The point is directional: coverage follows the deal.



The pattern: coverage follows either a very large deal or a specialist IPO. It rarely appears in advance of the transactional event. Investors who position ahead of the coverage narrowing capture the re-rating; those who wait for coverage to catch up pay for it.



What to watch


A Compass, Aligned or STACK IPO in 2026-2027 would open coverage on the modular developer sub-sector below the top four and reprice several PE-backed developer peers upward via visibility alone. An Empower Semiconductor transaction (whether IPO or acquisition by TDK, Broadcom or another strategic) would open dedicated coverage on the IVR magnetics segment. A Delta Star, Virginia Transformer or MGM transaction would open MV transformer coverage as a distinct analyst topic and would reprice Prolec GE, Siemens Energy transformer segment, and comparable exposures. None of these events is on any published calendar; all are visible as candidates within the M&A map of Part II. The pattern for investors is that coverage narrowing generally follows deals rather than preceding them, and the value accrues to whoever positioned ahead of the coverage.




Glossary of terms used



CFIUS
Committee on Foreign Investment in the United States. US inter-agency review body that screens foreign acquisitions of US businesses for national security implications.
EBITDA
Earnings Before Interest, Tax, Depreciation, and Amortisation. The most commonly referenced operating-earnings metric in M&A pricing.
GaN
Gallium Nitride. Wide-bandgap semiconductor material used in RF and power applications.
IC
Investment Committee. The decision-making body inside a PE firm or investor that approves or rejects a proposed transaction.
IPO
Initial Public Offering. A private company's first sale of shares to the public market.
IVR
Integrated Voltage Regulator. On-package power delivery component for modern CPUs and GPUs.
MV
Medium Voltage. Typically 1 kV to 35 kV. The voltage class connecting data centre power distribution to the utility grid.
MW
Megawatt. Unit of electrical power. A large modern data centre draws tens to hundreds of MW.
SiC
Silicon Carbide. Wide-bandgap semiconductor material used in high-voltage power electronics.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.





Sources and method


Analyst count figures compiled from FactSet Estimates and IBES coverage lists through mid-2026, verified against Bloomberg ANR and against desk-published research on the named companies. Counts are approximate and rest on active-coverage inclusion criteria (at least one published note in the trailing twelve months).


Private-market valuation compression figures rest on transaction disclosures where available (Ferric acquisition, Infineon/GaN Systems, Renesas/Transphorm) and on estimated private-market comparables where public disclosure is limited. Valuation multiple comparisons are approximate.


Trade press citations are illustrative rather than exhaustive; the substantive point is that specialist technical publications materially outperform generalist sell-side coverage on the technical content of the under-covered zones.



Series


The Investment Layer (eight parts). Part I: How the AI buildout is financedPart II: The M&A mapPart III: The build-to-lease flipPart IV (this piece): The coverage asymmetryPart V: The PE playbook per layerPart VI: Underwriting AI infrastructure debtPart VII: The three downside casesPart VIII: Sovereign capital and industrial policyCompanion to the AI Power Chain series and to earlier commercial supplements Pricing Under Scarcity and The Services Inversion.


The author is an independent advisor working on data centre, power and industrial technology. This analysis is written in a personal capacity and rests entirely on public information. Nothing here is investment advice.







See also · The Investment Layer

 * How the AI buildout is financed
 * The AI infrastructure M&A map
 * The build-to-lease flip







In brief


Sell-side coverage of AI infrastructure trails the buildout by roughly 18 months. Sector analysts still price data centre operators against 2023-era rack density and 2019-era power economics. Underpriced areas include thermal supply chain, wide-bandgap semiconductor consolidation, and hyperscaler-nuclear PPAs. Overpriced areas include commoditised colocation and legacy 415V AC power distribution.


==============================================================================
# The PE playbook, layer by layer
URL: https://adikumar.co/the-pe-playbook-layer-by-layer/
Published: 2026-08-14
Summary: Five recognisable PE playbooks across the six AI infrastructure layers, their return profiles, and the three platform theses in construction.
==============================================================================

The Investment Layer series · Part 5 of 8








The Investment Layer · Part V of VIII



The PE playbook, layer by layer


How private equity has approached each of the six layers of the AI infrastructure chain, which strategies have worked, and where the next platform theses are being constructed.




The Investment Layer

 1. Part I. How the AI buildout is financed
 2. Part II. The M&A map
 3. Part III. The build-to-lease flip
 4. Part IV. The coverage asymmetry
 5. Part V. The PE playbook, layer by layer (you are here)
 6. Part VI. Underwriting AI infrastructure debt
 7. Part VII. The three downside cases
 8. Part VIII. Sovereign capital and industrial policy




Executive summary


 * PE strategies across the AI infrastructure chain group into five recognisable playbooks: platform roll-ups, take-privates, developer scale-outs, structured minority stakes, and specialty carve-outs. Each has a distinct returns profile and each fits a specific layer of the chain better than the others.
 * The developer scale-out has been the dominant playbook by dollar volume. Blackstone (QTS), Brookfield (Compass), KKR/GIP (CyrusOne), DigitalBridge (Vantage, Yondr, DataBank), Blue Owl (STACK) and Macquarie (Aligned) have collectively deployed over $75 billion into datacenter developer equity between 2021 and 2026.
 * Platform roll-ups have been the second-most-common playbook. Vertiv's assembly under Platinum Equity before its 2020 SPAC listing is the reference case; other examples include Advent's roll-up of specialty cooling assets and Bain's Proterial take-private that unlocked a multi-year roll-up of Japanese specialty industrials.
 * Take-privates of listed specialty industrials with material AI exposure have accelerated in 2024-2026. Roper, Emerson (post-carve-out), Pentair, and Watts Water are all candidates that domain-focused funds are believed to be evaluating. None has closed yet at the time of writing.
 * Structured minority stakes have become the mechanism through which sovereign and pension capital participates without triggering strategic-review scrutiny. GIC, CPP Investments, ADIA, PIF, Temasek and NBIM have all taken minority positions in developers and specialty industrials at scale over 2023-2026.
 * The specialty carve-out playbook has been more common than expected. Emerson's electrical carve-out (AspenTech and other spin-outs), Honeywell's segment separation, and Schneider Electric's continued portfolio adjustment have all produced carve-out candidates that PE has evaluated.
 * The strategy-to-layer fit is not uniform. Developer scale-outs fit only the modular datacenter layer. Platform roll-ups fit the thermal and electrical layers. Take-privates fit specialty industrials with public listing. Structured minority stakes fit any layer where control transactions are politically sensitive. Specialty carve-outs concentrate in the electrical and thermal layers.




Chart 1. Committed PE and infrastructure-fund capital by playbook, 2021-2026


Developer scale-out dominates dollar volume by a wide margin. Platform roll-ups are the second largest bucket, driven by Vertiv-era assembly and comparable European precedents. Take-privates, structured minority stakes, and specialty carve-outs are meaningful in count but individually smaller.



Estimated commitments to identifiable AI-infrastructure transactions across the five playbook patterns, aggregated 2021 through mid-2026.




Five recognisable playbooks


Private-equity strategies across the AI infrastructure chain group into five patterns. Each has been executed at scale by identifiable funds, produces a distinct returns profile, and fits certain layers of the chain better than others. Reading a new PE announcement requires placing it in one of the five buckets to understand the intended return path and likely exit route.



Playbook 1. Developer scale-out


PE takes majority or full ownership of a modular datacenter developer platform, provides equity capital and structured debt to fund a multi-year pipeline of new-build capacity, negotiates long-term leases with hyperscaler tenants, and either holds the platform through the lease-cash-flow period or exits via IPO or secondary sale to a larger infrastructure fund.


Blackstone/QTS is the reference case. Blackstone took QTS private in June 2021 at $10 billion, at a time when the developer's pipeline was thin. By 2026 QTS has approximately 1900 MW deployed and 2300 MW additional pipeline, having become the largest hyperscale developer in the world. Blackstone's implicit MoIC on the transaction is in the 1.8-2.4x range at reasonable holding-period valuation. The IRR is comparable to a public-market REIT return but on materially larger deployed capital, which is what matters for a fund of Blackstone Infrastructure's mandate size.


Brookfield/Compass follows the same pattern with more measured pace. KKR/GIP's CyrusOne acquisition (closed March 25 2022, $15 billion including debt per company press release) has appreciated further with pipeline maturation. DigitalBridge has assembled Vantage, Yondr, and DataBank as a diversified portfolio of developer platforms rather than a single platform. Blue Owl's acquisition of STACK from IPI Partners in 2024 fits the same model.


The strategy works because the developer economics laid out in Part III have supported it. Levered IRRs in the 13-16 percent range on a stable operating asset with contracted cash flows are attractive to infrastructure LPs. The strategy will continue to work as long as those underlying economics hold. It compresses if BYOC procurement takes hold or if debt refinancing conditions deteriorate materially, both of which are discussed in Part III and Part VII.


Applicability: this playbook fits only the modular datacenter layer. The equipment layers (thermal, electrical, semiconductor) do not have the underlying real-estate asset base that supports the strategy.



Playbook 2. Platform roll-up


PE acquires a mid-size specialty industrial platform, uses it as the base to acquire a series of adjacent smaller vendors, integrates them into a consolidated platform with broader product portfolio and larger commercial reach, and exits via IPO or strategic sale.


Vertiv is the reference case. Platinum Equity acquired Vertiv (the former Emerson Network Power) for $4 billion, per Emerson 8-K (agreement announced Aug 2016, closed Dec 1 2016). Platinum then executed roll-up acquisitions of E&I Engineering, Geist, and a series of smaller specialty vendors, and exited via SPAC listing in February 2020 at approximately $5.3 billion (initial market cap; the listing has since appreciated to roughly $60 billion). The Vertiv playbook combined operational improvement (margin expansion from 6-8 percent EBITDA to 16-19 percent EBITDA over the hold period) with platform assembly (broader product portfolio for cross-sell into the developer stack).


Advent International has executed a comparable playbook in specialty cooling through the assembly of ProKlima and adjacent European liquid cooling specialists. Bain Capital's take-private of Proterial (Hitachi Metals) in 2023 at approximately $6.4 billion equity has been described as a roll-up of Japanese specialty industrials (magnetics, thermal materials, electrical steel) that Bain intends to consolidate under a unified operating platform.


The strategy works when the target industry is fragmented, has multiple mid-size vendors with complementary product portfolios, and where commercial reach into the developer/hyperscaler procurement channel is a scarce resource. It does not work where the industry is already concentrated (limited targets) or where the operational integration challenges outweigh the platform synergies (which was the risk in the Vertiv case that Platinum navigated successfully).


Applicability: this playbook has been executed at scale in the thermal layer (cooling, cold plate, CDU) and the electrical layer (integrated power distribution). It is beginning to see execution in specialty magnetics and specialty capacitors. It has not been executed in the semiconductor layers because the industries are structurally more concentrated already.



Playbook 3. Take-private of listed specialty industrials


PE identifies a publicly-listed specialty industrial with meaningful AI-related exposure where public-market valuation is compressed by non-AI segments, complexity discount, or short-term earnings volatility. PE takes the firm private, executes operational improvement and strategic focus, and exits via IPO, strategic sale, or trade sale to another PE holder.


Bain/Proterial in 2023 is the closest recent example, though it combined take-private and roll-up characteristics. JIC's proposed take-private of Shinko Electric in Japan (still pending as of mid-2026 after several regulatory and shareholder delays) fits the same pattern applied to on-package substrates.


Several publicly-listed specialty industrials have been discussed in the market as candidates for take-private on comparable grounds. Roper Technologies (application software plus specialty industrial exposure). Emerson (post its Aspen and other portfolio actions, the residual industrial core has been discussed). Pentair (specialty water plus some datacenter-adjacent segments). Watts Water (thermal fluid distribution and control). None of these has moved beyond discussion at time of writing, and each carries specific factors that may or may not support a take-private economically.


Applicability: this playbook fits publicly-listed specialty industrials in the thermal and electrical layers, plus specialty semiconductor firms with material AI-related revenue that trade at compressed multiples. The Wolfspeed situation is a variant in which the take-private has occurred through Chapter 11 reorganisation rather than a conventional PE process, but the outcome (private ownership, longer investment horizon) is comparable.



Playbook 4. Structured minority stakes


Sovereign wealth or major pension capital takes a minority (5-40 percent) stake in a developer, specialty industrial, or infrastructure asset, typically without control rights but with defined liquidity provisions and information rights. The stake provides growth capital to the target while allowing the LP to gain exposure to the AI infrastructure buildout at institutional scale without triggering the political-review scrutiny that a full acquisition might invite.


GIC's minority position in Digital Realty joint ventures, Vantage, and other developer platforms is a reference case. CPP Investments' co-investment in Compass, Aligned and QTS-related infrastructure debt is another. Temasek's minority holdings across the Asian portion of the developer stack. ADIA and NBIM's positions in Blackstone Infrastructure and other fund vehicles that hold developer exposure. PIF's structured commitments in US developers via co-investment vehicles that provide exposure without direct-holding sensitivity.


The playbook has become more common in 2024-2026 as trade-restriction and national-security concerns have complicated direct control transactions across borders. A US sovereign or Middle-East sovereign acquiring a controlling stake in a US developer would trigger CFIUS review; a 15 percent stake in a joint venture holding vehicle typically does not.


Applicability: this playbook fits any layer where control transactions are politically sensitive, which includes essentially all the layers when the buyer is a sovereign or has sovereign backing. It is most common in the developer and specialty industrial layers where the transactions are large enough to interest sovereign scale but small enough to permit meaningful minority participation.



Playbook 5. Specialty carve-out


PE acquires a discrete segment of a diversified industrial conglomerate through a carve-out process, either through direct negotiation with the parent or through a competitive process the parent runs. The carve-out then operates as a standalone platform, often with roll-up additions, and eventually exits.


The most active carve-out sellers over 2023-2026 have been Emerson (which has been executing a multi-year portfolio simplification and has separated software and specialty industrial businesses), Honeywell (which is executing its own segment separation), and Schneider Electric (which is continually pruning non-core positions to focus on datacenter and energy management). Each has produced carve-out opportunities that PE has evaluated.


Executed transactions include Advent's acquisition of certain specialty electrical assets from a European conglomerate, KKR's carve-out purchase of a specialty magnetics business from a Japanese parent (undisclosed price), and multiple smaller carve-outs in cooling and thermal management.


Applicability: this playbook fits primarily the electrical and thermal layers because those are the layers with the most conglomerate exposure. It is less common in semiconductor and developer layers because those industries have fewer conglomerate parents with divestible AI-adjacent segments.



Playbook fit by layer


The five playbooks map onto the six layers of the AI infrastructure chain with distinct concentrations. The following matrix summarises the fit and offers examples where applicable.



LayerPlaybook fit (heaviest first)Example executed transaction

Modular datacenterDeveloper scale-out (dominant); structured minority stakesBlackstone/QTS; GIC/Digital Realty JV
Thermal (cooling, cold plate, CDU)Platform roll-up; specialty carve-out; take-privatePlatinum/Vertiv; KKR/CoolIT; Advent/ProKlima
Electrical (MV distribution, transformer, switchgear)Platform roll-up; specialty carve-out; structured minorityPlatinum/Vertiv (E&I add); GIC/Prysmian; CPP/Hitachi Energy
Interconnect (cabling, connector, optical)Take-private; specialty carve-outPrysmian/Encore Wire (strategic); Bain/specialty carve-outs
Wide-bandgap semiconductorStructured minority stakes; take-privateBain/Proterial (adjacent); Chapter 11 restructuring for Wolfspeed
On-package power deliveryTake-private (via strategic combination); specialty carve-outJIC/Shinko (pending); TDK/Ferric (strategic)




Read a new PE announcement in one of five buckets: developer scale-out, platform roll-up, take-private, structured minority stake, specialty carve-out. Each has a distinct return path and exit route.The PE strategy taxonomy



Chart 2. Playbook fit by AI-infrastructure layer


The fit matrix. Developer scale-out fits only the modular datacenter layer. Platform roll-ups and specialty carve-outs dominate the thermal and electrical layers. Take-privates concentrate where public listings exist. Structured minority stakes are the pattern of choice where control transactions face political friction.



Author's fit assessment based on precedent transactions and structural characteristics of each layer. Darker shading indicates stronger fit.




Returns profile by playbook


The five playbooks produce distinct returns profiles. The following is an approximate summary of gross unlevered and levered IRR expectations, based on precedent transactions and observable outcomes where available.



PlaybookTypical holdUnlevered IRRLevered IRR (gross)Exit route

Developer scale-out7-12 years9-11%13-16%IPO, secondary to larger infra fund, or perpetual hold
Platform roll-up5-8 years13-17%19-24%IPO or strategic sale
Take-private4-7 years15-20%22-30%Strategic sale, secondary IPO
Structured minority5-15 years7-10%N/A (unlevered)Liquidity provisions, IPO, secondary
Specialty carve-out4-6 years18-24%26-35%Strategic sale, IPO




The return profile ordering matters for the LP strategy question. LPs with liquid-market alternatives that generate 7-10 percent unlevered are typically not attracted to the structured minority playbook unless the strategic exposure is otherwise unavailable. LPs with long-duration liability matching (Canadian pensions, sovereign wealth) are more attracted to the developer scale-out and structured minority playbooks. LPs focused on absolute return maximisation are more attracted to the platform roll-up, take-private, and specialty carve-out playbooks.


The composition of the LP base for each of these playbooks reflects this segmentation. Developer scale-out funds are typically anchored by Canadian pensions, sovereign wealth, and long-duration insurance. Take-private and carve-out funds are typically anchored by US endowments, private wealth, and corporate pensions with higher return targets.



Chart 3. Return profile by playbook: unlevered and levered IRR ranges


The five playbooks produce very different return profiles. Specialty carve-out and take-private target the highest returns. Structured minority is the lowest-return but the largest deployable capacity. Developer scale-out sits in between with stable, contracted cash-flow-driven returns.



Approximate ranges from precedent transactions and observable outcomes. Individual transactions vary considerably.




Chart 4. Typical hold period by playbook


Structured minority stakes are the longest-duration positions, often perpetual. Developer scale-out involves long holds because the underlying cash flow is contracted for 10-15 years. Take-privates and platform roll-ups are the shortest-hold playbooks, with clearer exit windows.



Ranges from published fund vintage data and typical infrastructure fund hold conventions.




Chart 5. LP composition typical to each playbook


The five playbooks attract distinct LP bases. Developer scale-out and structured minority draw from long-duration liability-matched capital (Canadian pensions, sovereign, insurance). Platform roll-up, take-private, and carve-out draw from higher-return-target capital (US endowments, private wealth, corporate pensions).



Author estimates from publicly-disclosed fund LP registers and typical infrastructure/PE fund composition patterns.




Where the next platform theses are being constructed


Three specific PE platform theses appear to be in construction as of mid-2026, each of which would produce material transactions over the next 24 months if executed.



Thesis A. US medium-voltage transformer roll-up


The proposition: acquire Delta Star, Virginia Transformer and MGM Transformer (or two of the three) as a base platform, execute a domestic US MV transformer roll-up, add capacity through greenfield US manufacturing supported by IRA and CHIPS Act eligibility, and exit via IPO or strategic sale to GE Vernova or Hitachi Energy within 5-7 years.


The economic argument is strong. US MV transformer lead times of 24-36 months constrain hyperscale interconnection more than almost any other factor. Domestic manufacturing capacity has been under-invested for two decades and is now the political priority of both parties. A consolidated domestic platform would be strategically valuable to a strategic acquirer at exit and would benefit from continuing federal and state support for domestic manufacturing.


The candidates: two of the three US independents (Delta Star, Virginia Transformer, MGM) are publicly discussed as sale candidates, with the third variably discussed depending on family ownership dynamics. The natural PE acquirers are Bain Capital's industrials group, KKR's industrials platform, EQT Infrastructure, and possibly a specialist industrial fund with more concentrated positioning.



Thesis B. Thermal fluid chemistry platform


The proposition: acquire a specialty PFAS-free thermal fluid chemistry business (Standard Fluids is the closest available example, though there are others in pre-commercial stage) and use it as the base to acquire adjacent specialty chemistries (dielectric fluids, coolants, high-temperature working fluids) into a consolidated platform positioned specifically for AI-driven thermal demand.


The economic argument rests on the regulatory pressure against legacy PFAS-based fluids combined with the projected growth in liquid-cooled deployments. A dedicated thermal fluid platform with regulatory-compliant chemistries could capture the transition from legacy fluids over 5-10 years. Total addressable market is smaller than in some other theses but the concentration is high.


The candidates: this thesis is more speculative than the transformer thesis because the target set is thinner. Standard Fluids is the most concrete candidate but is early-stage. Multiple traditional specialty chemistry firms could pivot into this space through acquisition. The natural PE acquirers are specialty chemistry-focused funds and Bain's industrials group.



Thesis C. Modular datacenter secondary consolidation


The proposition: consolidate two or three of the mid-size modular datacenter developers (Compass, Aligned, STACK, DataBank, EdgeConneX) through secondary transactions into a single larger platform, either via direct combination or via one fund acquiring another fund's holdings. The resulting scale would be comparable to the top three developers and would command a valuation premium at exit.


The economic argument rests on the exit-path capacity problem discussed in Part II. If the market cannot absorb all developer exits simultaneously, some consolidation among the developers themselves becomes attractive. Secondary consolidation reduces the number of exit events required and allows the combined platform to exit at scale.


The candidates: Compass (Brookfield), Aligned (Macquarie), STACK (Blue Owl), DataBank (DigitalBridge) are all approaching typical hold-period end within 24-36 months. A combination of any two of them would create a top-3 developer by scale. The transaction economics are complex because it involves inter-PE negotiation on relative valuation, but the strategic logic is attractive to LPs of both parties.



Chart 6. Three forward platform theses: dollar sizing and likelihood


The US MV transformer roll-up is the largest addressable opportunity by dollar sizing and has strong policy tailwind. Modular DC secondary consolidation is the most probable given the exit-path capacity math. Thermal fluid chemistry is the most speculative but the most concentrated on regulatory pressure.



Author's rough sizing and probability estimates from public information about each named target set and typical PE process dynamics.




What breaks these playbooks


Two broad conditions would compress returns across all five playbooks and force strategy revision.


Rising rates and refinancing conditions. All five playbooks rely on debt financing at multiple points. Developer scale-outs rely on 60 percent LTV construction debt refinanced at stabilisation. Platform roll-ups typically involve leveraged buyout financing at 5-6x EBITDA. Take-privates typically involve high-yield debt financing. Structured minority stakes rely on the underlying holding entity's financing. Specialty carve-outs typically involve LBO financing.


A sustained rate environment 200 basis points above current would compress levered returns by 4-8 percentage points across all playbooks and would materially delay exits as refinancing conditions become less favourable. This is not a scenario that PE holders can hedge fully; they can prefund some of it through longer-tenor debt but not all of it.


AI capex slowdown. All five playbooks rely on a continued or expanding AI infrastructure buildout. If hyperscaler capex flattens or contracts (discussed in Part VII as one of three downside cases), the demand assumptions underlying developer scale-outs, thermal roll-ups, and specialty carve-outs weaken. The severity of impact varies by playbook (developer scale-outs are most exposed because their capacity commitments are locked in whether or not demand shows up; specialty carve-outs are least exposed because they involve smaller absolute investments) but the direction is the same.



What to watch


Announced or completed PE take-privates of listed specialty industrials with AI exposure are the first signal. Each such transaction validates the take-private playbook for the specific layer involved and typically triggers competitive processes on comparable candidates. Developer secondary transactions among the PE holders themselves are the second: a Blue Owl sale of STACK to Brookfield, or a Macquarie sale of Aligned to Blackstone, would confirm the secondary consolidation thesis is being executed at scale. Sovereign minority stakes in specialty industrials are the third. Each such transaction validates the structured minority playbook for a layer that had not previously seen sovereign participation. And IPO filings by developer platforms are the fourth. A Compass or Aligned IPO filing would validate the developer IPO exit path and reduce pressure on the secondary and strategic exit paths, which is the specific bottleneck the exit-path capacity math points to.




Glossary of terms used



CDU
Coolant Distribution Unit. Thermal system component distributing chilled coolant to racks or direct-to-chip cold plates.
CFIUS
Committee on Foreign Investment in the United States. US inter-agency review body that screens foreign acquisitions of US businesses for national security implications.
EBITDA
Earnings Before Interest, Tax, Depreciation, and Amortisation. The most commonly referenced operating-earnings metric in M&A pricing.
IPO
Initial Public Offering. A private company's first sale of shares to the public market.
JV
Joint Venture. Business arrangement where two or more parties pool resources for a specific project, with separate governance and defined economic sharing.
LBO
Leveraged Buyout. Acquisition financed primarily with debt secured against the target's own cash flows.
LP
Limited Partner. Investor in a PE or infrastructure fund who commits capital but does not manage day-to-day investment decisions.
MV
Medium Voltage. Typically 1 kV to 35 kV. The voltage class connecting data centre power distribution to the utility grid.
MW
Megawatt. Unit of electrical power. A large modern data centre draws tens to hundreds of MW.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.





Sources and method


Transaction figures and PE deal history compiled from PitchBook, Preqin, IJGlobal, Bloomberg terminal data, S&P Capital IQ, and company disclosures through mid-2026. Deal values are as reported at announcement or closing where available.


Returns profile figures are approximate and derived from a synthesis of published fund returns (where available), PitchBook LP disclosures, industry survey data, and reasonable inference from precedent transactions. Individual transactions vary considerably around the summary figures.


The three platform theses discussed in the final section are the author's construction of plausible strategies visible from public information. They do not reflect any specific PE fund's disclosed strategy or reported process activity. Any resemblance to specific transactions in market is coincidental.



Series


The Investment Layer (eight parts). Part I: How the AI buildout is financedPart II: The M&A mapPart III: The build-to-lease flipPart IV: The coverage asymmetryPart V (this piece): The PE playbook per layerPart VI: Underwriting AI infrastructure debtPart VII: The three downside casesPart VIII: Sovereign capital and industrial policyCompanion to the AI Power Chain series and to earlier commercial supplements Pricing Under Scarcity and The Services Inversion.


The author is an independent advisor working on data centre, power and industrial technology. This analysis is written in a personal capacity and rests entirely on public information. Nothing here is investment advice.







See also · The Investment Layer

 * How the AI buildout is financed
 * The AI infrastructure M&A map
 * The build-to-lease flip





Further reading + sources
Related essays

 * The AI infrastructure M&A map. the transaction landscape
 * Underwriting AI infrastructure debt. the debt-side view
 * Case Study 01. Blackstone / QTS. the reference LBO
 * The build-to-lease flip. the specific structure this playbook exploits


Primary sources

 * Blackstone. the reference sponsor for the AI-infra thesis
 * Brookfield Infrastructure Partners. the reference infra fund







In brief


The PE playbook for AI infrastructure sponsor investments varies by layer of the power chain. Blackstone bought QTS at the developer layer. KKR + GIP bought CyrusOne at the same layer. Vendor-layer sponsors (Advent/Ultra PCS, Bain/Proterial) target industrial technology exposed to AI infra without direct data centre ownership. Return profile + hold period differ sharply between layers.








Frequently asked


How does the PE playbook differ by AI infrastructure layer?

Developer-layer sponsors (Blackstone/QTS, KKR-GIP/CyrusOne) buy platforms and compound through follow-on capital. Vendor-layer sponsors (Advent/Ultra PCS, Bain/Proterial) target industrial technology with AI infra exposure. Cash flow, hold period, and exit path differ by layer. Return profile is not uniform across the sector.


==============================================================================
# Underwriting AI infrastructure debt
URL: https://adikumar.co/underwriting-ai-infrastructure-debt/
Published: 2026-08-15
Summary: How the $200 billion of AI infrastructure debt actually gets priced: structures, spreads, covenants, rating criteria, and where underwriting is soft.
==============================================================================

The Investment Layer series · Part 6 of 8








The Investment Layer · Part VI of VIII



Underwriting AI infrastructure debt


The structures, spreads, covenants and rating agency criteria that determine how the $200+ billion of AI infrastructure debt outstanding in 2026 actually gets priced, and where the underwriting is soft.





The Investment Layer

 1. Part I. How the AI buildout is financed
 2. Part II. The M&A map
 3. Part III. The build-to-lease flip
 4. Part IV. The coverage asymmetry
 5. Part V. The PE playbook, layer by layer
 6. Part VI. Underwriting AI infrastructure debt (you are here)
 7. Part VII. The three downside cases
 8. Part VIII. Sovereign capital and industrial policy





Executive summary


 * Debt financing supporting the AI infrastructure buildout across all instruments and layers exceeds $200 billion outstanding as of mid-2026. Composition: investment-grade corporate debt of developer parents (~$60 billion), infrastructure project debt (~$45 billion), CMBS-adjacent structures (~$25 billion), REIT unsecured debt (~$40 billion), plus lease-adjacent structures and other private credit (~$30 billion).
 * The core underwriting question is the same across most instruments: how robust is the anticipated lease cash flow from hyperscaler tenants, and what covenants and reserves protect the lender if that cash flow does not materialise as expected?
 * Spreads on stabilised hyperscale-tenanted debt run 150-200 basis points over comparable-tenor Treasury for investment-grade issuers, 275-375 basis points for BB-rated project financings, and 450-650 basis points for structured private credit against sub-investment-grade developers. Spreads have compressed 40-80 basis points since early 2024 as institutional appetite has grown.
 * Rating agency criteria (S&P, Moody's, Fitch) for datacenter project debt have converged in 2024-2026 on a common framework that emphasises tenant credit quality, lease tenor, geographic diversification, and structural protections. The framework produces investment-grade outcomes at roughly 55-65 percent LTV on stabilised hyperscale-tenanted assets.
 * Three areas of soft underwriting stand out. Refinancing risk in years 5-7 of interest-only construction debt is not fully captured in current covenant packages. Concentration exposure to single hyperscaler tenants across a portfolio is often not measured. Grid interconnection contingencies are frequently assumed rather than reserved against.
 * The construction debt segment is the most exposed to any tightening in credit conditions. Approximately $30-40 billion of construction debt was originated in 2020-2022 at rates significantly below current market. The refinancing wave in 2026-2028 will test both borrower resilience and lender capacity to absorb the reset.
 * Chinese domestic AI infrastructure debt runs in parallel with Western structures, mostly on sovereign or policy-bank balance sheets, with limited direct connection to Western credit markets. Cross-border syndication of AI infrastructure debt has been minimal since 2023.





Chart 1. AI infrastructure debt composition, mid-2026, $B outstanding


Western AI-infrastructure debt outstanding across all instruments sits at about $220 billion. Investment-grade REIT and corporate debt is the largest bucket. Project debt and structured private credit have grown fastest. Chinese domestic policy-bank debt runs in parallel at roughly comparable scale but is largely disconnected from Western credit markets.



From company disclosures, Bloomberg BVAL, Moody's ABS/CMBS reports, S&P Structured Finance data through mid-2026. Individual instrument totals are approximate.





The composition of AI infrastructure debt


Total debt outstanding supporting the AI infrastructure buildout across all instruments and layers is approximately $200-220 billion as of mid-2026. The composition breaks down as follows:




Instrument typeApprox. outstandingTypical issuerTypical use of proceeds

Investment-grade corporate debt$60 billionDigital Realty, Equinix, Iron Mountain, listed REIT and industrial parentsBalance-sheet financing for portfolio and development
Infrastructure project debt (bank + institutional)$45 billionProject SPVs holding single or small-portfolio datacenter assetsDevelopment and construction of specific projects
CMBS-adjacent structures$25 billionREIT or developer-sponsored trustsRefinancing of stabilised operating portfolios
REIT unsecured debt$40 billionDigital Realty, Equinix, plus international REIT peersGeneral corporate purposes, acquisitions
Structured private credit and lease-adjacent$30 billionBlackstone, Blue Owl, KKR direct lending; leasing structuresBridge financing, construction financing to non-REIT developers
Equipment financing (thermal, electrical, IT)$5 billionOEMs (Eaton, Vertiv, Schneider) and equipment financiersWorking capital extension into vendor sales
Chinese domestic policy-bank debt~¥800-1000 billion (~$110-140 billion equivalent)China Development Bank, Export-Import Bank of China, ICBCDomestic datacenter and semiconductor buildout
Sovereign-backed infrastructure debt (ex-China)$15 billionMiddle East, Southeast Asia sovereign vehiclesNational datacenter programmes





The Western totals (excluding Chinese domestic) approximate $220 billion. The Chinese domestic total on comparable purpose approaches $130 billion equivalent, though the underwriting mechanics are materially different from Western structures and the two pools have limited direct connection.




The core underwriting question


Almost all AI infrastructure debt underwriting reduces to the same fundamental question: how robust is the anticipated lease cash flow from hyperscaler tenants, and what covenants and reserves protect the lender if that cash flow does not materialise as expected?



Four sub-questions structure the analysis:




Sub-question 1. Tenant credit quality


The dominant lease counterparties are Meta (A+/A1), Google (AA+/Aa2), Microsoft (AAA/Aaa), Amazon (AA/A1), plus Oracle (BBB+/Baa2), Apple (AA+/Aaa) and a small number of others at investment grade. Tenant credit quality is the strongest single factor in AI infrastructure debt underwriting because the lease cash flow determines the debt service.



The concentration issue is material. Approximately 65 percent of Western hyperscale leased capacity is contracted with just four tenants (Meta, Google, Microsoft, Amazon). Diversification within a single portfolio typically means selling to 2-3 of the four rather than genuinely different customer types. Underwriting frameworks generally do not penalise this concentration heavily because each of the four is investment grade, but a scenario in which any one hyperscaler slowed procurement or reduced its commitment would produce material concentration risk that is under-priced in current spreads.




Sub-question 2. Lease tenor and escalation


Typical lease tenor is 10-15 years, with 15-year leases being most common for hyperscale primary-market deployments and 10-year leases for secondary markets and specialised buildings. Escalation is typically 2.5-3 percent annually fixed, or CPI-linked with a floor and ceiling. Longer tenor supports higher LTV and lower spreads; shorter tenor with rollover risk pushes spreads wider.



The underwriting question that is less well-addressed: what happens at lease expiry, when the leased building becomes a specialised asset with limited alternative use? A hyperscale datacenter without a hyperscaler tenant is not straightforwardly re-leasable to a comparable tenant at comparable terms. Some structures reserve against re-lease risk explicitly (mostly the CMBS-adjacent structures); many do not.




Sub-question 3. Geographic and grid diversification


Concentration in specific markets (Northern Virginia, Phoenix, Dallas, Columbus, Chicago, Frankfurt, Singapore) has been increasing rather than decreasing as developers concentrate on the markets where interconnection and land are most accessible. Portfolios that are diversified across 6-8 US metro markets receive better ratings than portfolios concentrated in 2-3.



The grid-side dimension is not always well-captured. Two datacenter assets in Northern Virginia may be on the same substation and share the same interconnection queue exposure. Rating frameworks have not typically drilled into grid concentration within a market; they have focused on metro-market diversification.




Sub-question 4. Structural protections


Reserve accounts (typically 6-9 months of debt service), cash sweeps at defined performance triggers, tenant credit-support requirements (guarantees or letters of credit from the hyperscaler parent, though not always available), completion guarantees during construction phase, and change-of-control provisions are the standard structural protections. Package strength varies considerably across issuers.



The soft area: many private credit structures rely more heavily on covenant-lite structures than the equivalent investment-grade or high-yield public debt would tolerate. This is not intrinsically wrong (private credit lenders have more direct information rights and control over the borrower than public debt investors do) but it does mean that reported spreads on private credit under-reflect the covenant risk.




Chart 2. Tenant concentration across Western hyperscale leased capacity


Four tenants (Meta, Google, Microsoft, Amazon) account for approximately 65 percent of Western hyperscale leased capacity by MW. The remaining 35 percent is spread across Oracle, Apple, TikTok/ByteDance, and enterprise/smaller cloud tenants. The concentration is the tenant credit risk that no single portfolio can diversify away.



Estimated share from disclosed leased capacity and industry allocation data. Concentration figure varies by 5 percentage points depending on how mixed-use and secondary-market leases are attributed.





Spread benchmarks as of mid-2026


The following benchmarks are approximate mid-market spreads over comparable-tenor Treasury for representative issuers and structures as of mid-2026. Individual transactions vary considerably around these benchmarks based on structure specifics, tenant mix, and market timing.




InstrumentRatingApprox. spread over USTSpread evolution since Q1 2024

Investment-grade REIT unsecured, 10-yearA- to BBB+135-165 bps~50 bps tighter
Investment-grade REIT unsecured, 30-yearA- to BBB+175-205 bps~40 bps tighter
Infrastructure project debt, stabilised hyperscaleBBB / BBB-200-260 bps~60 bps tighter
CMBS-adjacent datacenter portfolioVarious tranchesSenior A: 175-210 bps; Sub B: 350-450 bps~40-70 bps tighter
Infrastructure project debt, construction phaseBB / BB+275-375 bps~80 bps tighter
Structured private credit, sub-IG developerN/A (unrated)450-650 bps~100 bps tighter
Chinese domestic policy-bank rate (5-year LPR+)N/A (sovereign)N/A (policy rate)Relatively stable





The tightening of 40-100 basis points across most Western instruments over the past 24 months reflects institutional appetite growth for the sector. Long-duration LPs (pensions, sovereign wealth, insurance) have increased allocations to infrastructure debt with AI-infrastructure exposure meaningfully, and the flow-based tightening has priced in a good scenario that assumes continued hyperscaler demand.



The concern is symmetric to the appetite. If institutional appetite reverses (whether from AI-capex slowdown, from a specific tenant credit event, or from a broader risk-off move), the spreads could widen by 100-200 basis points relatively quickly. The market has not been tested with a downside scenario since 2022.




Chart 3. Spread over Treasury by instrument, mid-market as of mid-2026 (bps)


Spreads run from about 150 bps for investment-grade REIT unsecured 10-year, out to 550 bps for structured private credit against sub-investment-grade developers. Spreads have compressed 40-100 bps across most instruments since Q1 2024 as institutional appetite has grown. That compression is what a demand-side or credit shock would reverse.



Mid-market spreads over comparable-tenor UST. Individual issuances vary. Sources as above plus Bloomberg curves for representative issuers.





Rating agency criteria and outcomes


S&P, Moody's, and Fitch have converged on a common analytical framework for datacenter project debt over 2024-2026. The framework produces investment-grade outcomes on stabilised hyperscale-tenanted assets at approximately 55-65 percent loan-to-value, with lower LTVs required for portfolios with shorter lease tenors, smaller tenant credit quality, or higher geographic concentration.



The key inputs the frameworks assess:



Tenant credit rating and lease tenor produce the "contracted cash flow" score. A portfolio with 90 percent AA-rated tenants on 15-year leases receives the highest score; 70 percent BBB-rated on 10-year leases the lowest for hyperscale portfolios (below that threshold the asset would not typically qualify as hyperscale-tenanted for framework purposes).



Geographic diversification, sub-portfolio concentration, and grid diversification produce the "structural resilience" score. A US portfolio across six or more metros receives the highest; a portfolio concentrated in two metros receives the lowest.



Reserve accounts, cash-sweep triggers, and change-of-control provisions produce the "structural protection" score. A robust package receives investment-grade uplift; a covenant-lite structure does not qualify for investment grade at any LTV.



The three scores combine into an overall assessment that maps to a rating on the standard scale. The framework produces BBB outcomes at 55-65 percent LTV for typical portfolios and A- outcomes at 45-55 percent LTV. Higher LTVs (65-75 percent) are achievable only with tenant credit guarantees or specific structural enhancements.



The frameworks have been criticised by some market participants for being insufficiently sensitive to tenant concentration and refinancing risk. The critique is not without foundation. A portfolio contracted 90 percent with a single AAA-rated tenant would receive investment-grade rating despite carrying tenant concentration that would attract material scrutiny in most other credit frameworks. The rating agencies' response is that AAA counterparty credit dominates the concentration concern, which is true up to the point where the AAA tenant is not AAA any longer.



All AI infrastructure debt underwriting reduces to one question: how robust is the anticipated hyperscaler lease cash flow, and what protects the lender if it does not materialise? Answering it well matters more than benchmarks.The underwriting frame that matters




Chart 4. Rating agency framework: LTV vs achievable rating on hyperscale-tenanted debt


Investment-grade outcomes are typically achievable at 55-65 percent LTV on stabilised hyperscale-tenanted assets with typical tenant mix and diversification. Sub-investment grade at higher LTV. Higher LTV outcomes require tenant credit guarantees or specific structural enhancements.



Author's summary of published methodologies from S&P (Q3 2024 revision), Moody's (Q2 2025 update), Fitch (2024). Specific criteria vary by rating agency; the pattern is directional.





Three areas of soft underwriting


Three areas of AI infrastructure debt underwriting are consistently softer than the risk profile warrants.




Soft area 1. Refinancing risk on 2020-2022 vintage construction debt


Approximately $30-40 billion of construction debt was originated between 2020 and 2022 at all-in coupons in the 3.5-4.5 percent range. Most of this debt was structured as 5-7 year interest-only with a balloon at maturity. The refinancing wave hits 2025-2028, with 2026-2027 as the peak period.



Current market refinancing rates for comparable structures are 5.75-6.75 percent all-in, which is 200-250 basis points higher than the original coupons. The debt service coverage assumption in the original underwriting was typically 1.20-1.35x. At the refinance rate, coverage ratios move to 1.00-1.15x, which is below most covenant thresholds and often below the refinancing lender's minimum requirement.



The scenarios that resolve this: (a) rates drop by 100+ basis points before refinance (possible but not central case); (b) developer equity infusion at refinance to reduce LTV; (c) lease amendment to increase rent (which requires tenant consent and typically comes at cost to the developer); (d) refinancing at a higher spread and shorter tenor at compressed coverage ratios. Most transactions will likely require some combination of (b), (c), and (d).



Current covenants do not require developers to build reserves against refinancing risk of this magnitude, which is a structural underwriting weakness. The lenders on the current-vintage debt are largely insulated (they get paid at maturity) but the refinancing lenders will inherit the compressed coverage.




Soft area 2. Cross-portfolio tenant concentration


A given developer's portfolio may show acceptable tenant diversification (say, 30 percent Meta, 25 percent Google, 20 percent Microsoft, 15 percent Amazon, 10 percent others). But across the developer stack as a whole, the same four tenants account for the substantial majority of leased capacity. A lender holding debt against multiple developers may inadvertently have combined tenant exposure that exceeds what any single-portfolio analysis would show.



Underwriting frameworks do not typically require cross-portfolio tenant exposure reporting. Insurance company debt portfolios and pension debt portfolios often carry very concentrated tenant exposure without being sized against tenant-specific stress scenarios. Institutional risk management has not fully caught up with the specific concentration issue that AI infrastructure debt has created.




Soft area 3. Grid interconnection contingencies


Construction debt is typically structured with milestone-based drawdown that assumes grid interconnection is achieved on schedule. Delayed interconnection triggers cost overruns (extending the pre-revenue period) and can trigger completion covenant issues.



Current market practice reserves against typical interconnection delays of 3-6 months. Actual delays on complex hyperscale interconnections have run 12-24 months on approximately 15-20 percent of projects since 2023, as detailed transformer supply constraints, substation upgrade requirements, and ISO/utility queue backlogs have proven materially worse than pre-2023 planning assumptions.



The lenders on the delayed projects have generally worked with borrowers to extend interest-only periods or to fund incremental completion costs from equity. This has been possible because the sector has not been under broader stress. If broader stress emerged, the same delays would translate to covenant defaults and forced restructurings rather than accommodations.




Chart 5. Debt service coverage ratio at refinance, by original coupon and current refinance rate


The coverage compression math on 2020-2022 vintage debt. Debt originated at 3.5-4.5 percent coupon with 1.20-1.35x coverage moves to 1.00-1.15x coverage at 5.75-6.75 percent refinance rates. Below 1.05x, most covenant packages require equity infusion, lease amendment, or renegotiation.



Author's estimate of coverage evolution as rates move from origination to refinance level.





The construction debt refinancing wave


The 2026-2028 refinancing wave for construction debt originated in 2020-2022 is the most concrete stress test the AI infrastructure debt market will face. The wave has three characteristics that make it distinct from typical refinancing cycles.



Volume concentration. Approximately $30-40 billion of debt refinances over a 24-month window, which is a meaningful proportion of the total sector debt. Absorbing this volume without spread widening requires continued institutional appetite growth.



Rate reset magnitude. The 200-250 basis point coupon reset described above is larger than any single event the current-vintage lenders' underwriting anticipated. Coverage compression at refinance is the norm rather than the exception across this vintage.



Borrower resilience variability. Some borrowers (Blackstone-backed QTS, Brookfield-backed Compass) have pre-financed longer tenors, hedge portfolios, and equity capacity to absorb the reset with minimal accommodation required. Others (smaller PE-backed developers and some earlier-stage developers) have not, and will require material lender accommodation to complete refinancing.



The most likely outcome is orderly refinancing at compressed coverage ratios, with 5-10 percent of transactions requiring material equity infusion or lease amendment to complete. A minority of transactions may require restructuring, but the base case does not envision a broad restructuring wave.




Chart 6. Construction debt refinancing wave, 2020-2022 vintages by maturity


Roughly $30-40 billion of construction debt originated in 2020-2022 hits refinancing over 2025-2028, with 2026-2027 as the peak. This is the single most-concentrated stress test the AI infrastructure debt market will face, and its outcome will indicate whether current market pricing survives the rate-reset environment.



Estimated volume of construction debt hitting maturity by year, based on typical 5-7 year interest-only structures on 2020-2022 vintages.





What breaks this market


Three specific shocks would materially widen spreads and reduce debt availability across the sector.



A tenant credit event. If one of the four dominant hyperscaler tenants experienced a material credit event (rating agency downgrade below investment grade, earnings deterioration, or governance-driven credit repricing), spreads on portfolios contracted with that tenant would widen sharply, and portfolios diversified across multiple tenants including the downgraded one would experience meaningful spread widening. Given the concentration, no single portfolio is well-diversified against this specific risk.



A regulatory intervention. If US or EU regulators imposed antitrust-style restrictions on hyperscaler cloud market power (which has been discussed in policy circles but not concretely proposed), the resulting revenue disruption to the hyperscalers could translate to lease-covenant issues for datacenter debt.



A capex slowdown. Discussed in more detail in Part VII, but the debt-market implication is that construction debt originated in 2024-2026 against forward pipelines could face material take-out issues if hyperscaler demand for the underlying capacity does not materialise as underwritten.




What to watch


Successful refinancing of 2020-2022 vintage construction debt is the leading signal. Completion of major refinancings (which will begin in earnest in Q4 2026) will indicate whether the market is absorbing the rate reset as expected; failed or restructured refinancings would signal broader stress and reprice the sector. Spread evolution on new-issue project debt is the second signal. Continued tightening confirms institutional appetite growth, widening signals appetite reversal, and any 50+ basis point movement in either direction over a quarter is meaningful. Rating agency framework revisions are the third: any tightening in tenant concentration criteria or refinancing reserve requirements would constrain debt availability at existing spreads even without any underlying credit event. And hyperscaler credit rating moves or watch actions are the fourth and highest-consequence. Any negative watch or downgrade on Meta, Google, Microsoft, or Amazon would ripple through datacenter debt spreads immediately given the concentration issue.





Sources and method


Debt outstanding figures compiled from company disclosures, Bloomberg BVAL data, Moody's ABS/CMBS reports, S&P Structured Finance data, and industry syndication reports through mid-2026. Individual instrument totals are approximate.


Spread benchmarks derived from Bloomberg BVAL curves for representative issuers, cross-referenced against private market indications from ratings agency reports and industry survey data. Individual transaction spreads vary based on structure, timing, and market conditions.


Rating agency framework discussion rests on published methodologies from S&P (Data Center Property Methodology, Q3 2024 revision), Moody's (Data Center Real Estate methodology update Q2 2025), and Fitch (Digital Infrastructure Rating Criteria, 2024). Specific criteria and outcomes are the author's summary of the published frameworks rather than direct quotations.


Refinancing wave analysis is the author's estimate based on public disclosure of 2020-2022 vintage debt structures at origination and reasonable inference from typical structure conventions. The specific volume estimates should be read as directional rather than precise.



Series


The Investment Layer (eight parts). Part I: How the AI buildout is financedPart II: The M&A mapPart III: The build-to-lease flipPart IV: The coverage asymmetryPart V: The PE playbook per layerPart VI (this piece): Underwriting AI infrastructure debtPart VII: The three downside casesPart VIII: Sovereign capital and industrial policyCompanion to the AI Power Chain series and to earlier commercial supplements Pricing Under Scarcity and The Services Inversion.


The author is an independent advisor working on data centre, power and industrial technology. This analysis is written in a personal capacity and rests entirely on public information. Nothing here is investment advice.









See also · The Investment Layer

 * How the AI buildout is financed
 * The AI infrastructure M&A map
 * The build-to-lease flip





Further reading + sources
Related essays

 * How the AI buildout is financed. the capital-structure frame
 * The coverage asymmetry. the specific covenant-side lens
 * Financing I. DDTL. the DDTL primer this essay assumes
 * Financing structure, covenants, closing risk. the closing-risk lens


Primary sources

 * S&P Global Ratings. rating methodology for AI-infra debt
 * Moody's Ratings. alternative rating perspective







In brief


AI infrastructure debt underwriting requires a different question set than typical infrastructure lending. Hyperscaler credit backs most contracted revenue, but contract-specific covenants (change-of-control, cure periods, capacity flex) reshape recovery in default. DDTL structures pledge specific customer contracts to specific debt tranches, matching debt-service coverage to contracted cash flows.








Frequently asked


What credit questions matter most for AI infrastructure DDTL structures?

Change-of-control clauses on the underlying hyperscaler contract determine debt recovery if the operator is acquired. Cure periods (typically 30-180 days) affect debt-service coverage under stress. Capacity-flex clauses let the hyperscaler downsize commitments under specified conditions, which resets the underlying cash flow. Standard corporate credit metrics understate these risks.


==============================================================================
# The three downside cases
URL: https://adikumar.co/the-three-downside-cases/
Published: 2026-08-16
Summary: Three scenarios that compress the AI infrastructure investment map: demand shock, supply reversal, financing shock. Probabilities and exposures.
==============================================================================

The Investment Layer series · Part 7 of 8








The Investment Layer · Part VII of VIII



The three downside cases


Three distinct scenarios in which the AI infrastructure investment map compresses, ranked by structural probability, with the specific transmission mechanisms and asset exposures for each.




The Investment Layer

 1. Part I. How the AI buildout is financed
 2. Part II. The M&A map
 3. Part III. The build-to-lease flip
 4. Part IV. The coverage asymmetry
 5. Part V. The PE playbook, layer by layer
 6. Part VI. Underwriting AI infrastructure debt
 7. Part VII. The three downside cases (you are here)
 8. Part VIII. Sovereign capital and industrial policy




Executive summary


 * Three distinct downside scenarios could compress the AI infrastructure investment landscape over the next 24 to 60 months. Each has a different probability, a different transmission mechanism, and a different pattern of asset exposure.
 * Downside 1 is a demand-side shock: hyperscaler capex flattens or contracts because AI monetisation lags AI capital deployment, or because a model architecture shift reduces the marginal compute intensity per user. Our estimated probability over the next 24 months is roughly 25 percent for a mild version and 8-12 percent for a severe version.
 * Downside 2 is a supply-side reversal: interconnection queue reform, transformer capacity build-out, or a technology substitution reduces the scarcity premium that supports current developer and equipment valuations. Probability roughly 20 percent for a partial reversal in the 24-month window, higher over a 60-month window.
 * Downside 3 is a financing shock: rate persistence, refinancing failure at 2020-2022 vintage vintages, or a specific credit event that widens spreads across the sector. Probability roughly 30-35 percent for material impact in the 24-month window.
 * The three scenarios are not fully independent. Downside 1 partially triggers Downside 2 (supply catches up faster if demand slows) and partially amplifies Downside 3 (compressed rents make refinancing harder). Downside 3 partially triggers Downside 1 (financing shocks compress hyperscaler capex directly). Compound scenarios are more probable than either downside case alone.
 * Asset exposure to the three downside cases is asymmetric. Developer platforms are most exposed to Downside 1 (rent compression) and Downside 3 (refinancing). Specialty equipment vendors with strong installed-base annuities are least exposed to Downside 1 (annuity persists). Specialty semiconductor vendors are most exposed to Downside 2 (technology substitution).
 * The signals that would validate any of the three cases in progress are observable in specific data series (hyperscaler capex guidance, interconnection queue clearance rates, credit spread evolution). Position adjustment ahead of the confirming data is possible for institutional investors but requires accepting some optionality cost.




Chart 1. Three downside cases: probability and severity, 24-month horizon


The three downside cases sit at very different probability-impact points. Financing shock has the highest 24-month probability. Demand-side shock has the largest sector-wide impact if it materialises severely. Supply-side reversal has moderate probability but longest-duration compression because the affected valuations do not recover.



Author's subjective probability estimates. Impact is sector-wide equity impact estimate, not any specific asset.




Downside 1. Demand-side shock: AI monetisation lags AI capex


The upside case for AI infrastructure rests on continued hyperscaler capex at current or growing pace, driven by anticipated AI monetisation via improved cloud margins, new consumer products, and enterprise revenue. If monetisation lags well behind the capex pace, the four dominant hyperscalers will either compress capex growth to match observed monetisation (mild version) or contract capex sharply to preserve free cash flow (severe version).



Mechanism


The transmission from monetisation lag to infrastructure demand runs through hyperscaler earnings pressure. Analyst focus on AI monetisation has intensified since 2024, with quarterly commentary on AI-specific revenue growth becoming a standard question. If two consecutive quarters of underperformance versus AI revenue expectations occur, capex guidance typically follows within one to two quarters.


The mild version: capex growth compresses from current 20-30 percent year-over-year to 5-10 percent. This translates to reduced pipeline pressure on developers, which relaxes the scarcity premium on grid-connected sites, which compresses new-build rents on marginal contracts. Existing 10-15 year leases remain intact but new contracts price weaker.


The severe version: capex flattens or contracts by 10-15 percent year-over-year for one or two years. This compresses new developer deals sharply, forces some developer pipeline into permanent shelved status, and can trigger lease amendment discussions on newer contracts (particularly those signed at peak pricing in 2024-2025).



Probability assessment


The mild version has meaningful probability over the next 24 months. AI monetisation across consumer products (co-pilot integration, search, chat) and enterprise cloud has been slower than the capex pace anticipated. Google's Gemini and Microsoft's Copilot suite have both underperformed initial internal projections. Enterprise adoption of AI-agent services has been meaningful but not the transformational pace some had assumed. The gap between capex growth and monetisation growth has been noted in analyst commentary throughout 2025-2026.


The severe version has lower but non-negligible probability. It would require either a specific model architecture breakthrough that reduces compute intensity per user (which cannot be predicted but has precedent) or a broader macro or governance event that forces hyperscaler capital preservation. The severe version is more likely as a compound with Downside 3 (financing shock) than as a standalone event.


Approximate probability weighting: mild version 25 percent probability over 24 months; severe version 8-12 percent. Combined at least mild: 30-35 percent.



Asset exposure


Most exposed: modular datacenter developers with high proportion of forward pipeline (Compass, Aligned, STACK, EdgeConneX). Pipeline-heavy developers depend on continued lease flow to justify the equity model. If lease flow compresses, forward pipeline gets shelved and equity returns compress sharply.


Moderately exposed: publicly-listed REITs (Digital Realty, Equinix). Established portfolios continue generating rent from existing leases, but new-build and re-lease pricing compresses. Equity valuations reprice on the growth outlook rather than immediate cash flow.


Moderately exposed: specialty equipment vendors with heavy new-build revenue mix (Vertiv, Schneider Electric equipment segment). New-build revenue compresses; services and installed-base revenue continues.


Least exposed: specialty equipment vendors with heavy installed-base services mix (Eaton post-Boyd, Vertiv Services, Schneider Services), and speciality semiconductor vendors on qualified supplier lists for existing chip programmes (TDK for Ferric magnetics, Onsemi for SiC devices in current programme wins).



Recovery path


The mild version recovery would come through renewed monetisation improvement or through model architecture developments that increase compute intensity per user again. Historical precedent suggests recovery windows of 12-24 months if the fundamental AI product-market fit case remains intact. The severe version recovery would take 24-48 months and would require more substantial revenue proof-points.



Downside 2. Supply-side reversal: scarcity premium compresses


The current AI infrastructure valuation landscape rests on scarcity in multiple specific inputs: grid interconnection, transformer capacity, thermal loop specialists, on-package power delivery, and specialty semiconductor manufacturing. If any of these scarcities compresses through capacity addition, the pricing power of the constrained suppliers compresses with it.



Mechanism


Supply-side reversal is not a single event but a series of specific capacity releases that could compound over 24-60 months. Each has its own probability profile.


Grid interconnection queue reform. FERC Order 2023 and comparable state and regional ISO reforms aim to clear stalled interconnection queues by imposing readiness deposits and cost allocation rules. If successful, previously-queued sites become interconnected faster than developer expectations built for, and the scarcity premium on any single developer's grid-connected portfolio compresses.


Transformer manufacturing capacity build-out. US and European utilities have committed to substantial transformer capacity investment through 2028, backed by IRA and comparable European incentive structures. If the capacity commitments deliver on the anticipated timeline, transformer lead times of 24-36 months could compress to 12-18 months by late 2027, which reduces the pricing power of MV transformer specialists.


On-package power delivery technology substitution. If IVR technology from Empower, Ferric, or a new entrant reaches broader qualification faster than expected, or if a competing approach (integrated voltage regulator embedded in silicon, capacitor-only power delivery) matures, the concentrated qualified-supplier field expands and pricing compresses.


Thermal specialist scale-up. Vertiv's post-PurgeRite integration, KKR's post-CoolIT expansion, and Eaton's post-Boyd platform are all building thermal service organisations at scale. If the combined capacity meets or exceeds pipeline demand growth, the scarcity premium on thermal specialists compresses.


Chinese domestic capacity release into export markets. Chinese specialty industrial capacity in transformers, cooling equipment, and specialty semiconductors has been building rapidly and is primarily serving domestic demand. If regulatory conditions permit Chinese export capacity to reach Western datacenter buyers at scale, Western pricing compresses meaningfully.



Probability assessment


Each of the five specific supply-side reversals has 15-25 percent standalone probability over 24 months and higher probability over 60 months. Combined probability that at least one meaningfully compresses (grid queue OR transformer capacity OR any of the others) over 24 months is roughly 45-55 percent. The severity of impact varies by which specific reversal materialises.


Grid interconnection queue reform is the most consequential single lever because it affects the developer stack broadly. Transformer capacity build-out is second-most consequential because it affects the electrical layer broadly.



Asset exposure


Most exposed: specialty semiconductor vendors with concentrated qualified-supplier positions where alternative suppliers or alternative technologies could displace. Wolfspeed and comparable SiC substrate vendors are exposed to Chinese substrate entry. Ferric and IVR magnetics are exposed to technology substitution.


Moderately exposed: MV transformer specialists (Hitachi Energy, Siemens Energy transformer segment, Prolec GE, Delta Star, Virginia Transformer, MGM). US-specific onshoring positioning provides some protection against Chinese import competition even in a supply-side reversal, but does not fully insulate against domestic capacity growth.


Moderately exposed: developer stack more broadly. Grid interconnection queue reform would compress the scarcity premium on developer portfolios, though not eliminate it. Existing leases remain intact; new-build economics compress.


Least exposed: specialty vendors with technical moats that are not easily displaced by capacity addition. Cold plate specialists benefit from installed base annuity that persists regardless of new capacity. Thermal fluid chemistry with regulatory-compliant PFAS-free positioning benefits from regulatory pressure that supply-side reversal does not undo.



Recovery path


Supply-side reversals tend to be structural rather than cyclical. Once transformer capacity or interconnection queue capacity is built, it typically remains in the market. The affected valuations do not recover unless new demand growth exceeds the new supply capacity, which requires either continued strong AI infrastructure demand or a new demand source. Recovery windows for supply-side compression are typically longer than for demand-side or financing shocks.



Downside 3. Financing shock: rates, refinancing, or credit event


The AI infrastructure buildout relies on debt financing at multiple points as detailed in Part VI. A financing shock (rate persistence at higher levels, failed refinancing at 2020-2022 vintages, or a specific credit event) would widen spreads across the sector and compress developer and specialty industrial equity returns.



Mechanism


Three distinct financing shocks have material probability over the 24-month window:


Rate persistence. If US and European policy rates remain at current levels through 2027-2028 rather than easing on the current market-implied path, refinancing conditions for 2020-2022 vintage construction debt (approximately $30-40 billion) become substantially more difficult. Developer equity infusion or lease amendment becomes the norm for most refinancings.


Refinancing failure. A minority of the 2020-2022 vintage refinancings could fail outright (rather than complete at compressed coverage) if developer equity capacity is exhausted and tenants decline to amend leases. Failed refinancings translate to distressed sale processes or restructurings. A visible failure event (a mid-tier developer forced into restructuring) would trigger broader spread widening across the sector.


Tenant credit event. Any of the four dominant hyperscaler tenants (Meta, Google, Microsoft, Amazon) experiencing a material credit deterioration (rating agency negative watch, downgrade, or governance-driven repricing) would ripple through datacenter debt and equity valuations immediately. The concentration issue discussed in Part VI means no portfolio is fully insulated against this specific risk.



Probability assessment


Rate persistence at higher levels is the most probable of the three shocks; 40-50 percent probability that policy rates remain at or above current levels through 2027. Not all of this constitutes a shock (the market has partially priced it in), but the specific refinancing challenge described in Part VI does become more likely as rates persist.


A visible refinancing failure has lower standalone probability (15-20 percent over 24 months) but higher probability conditional on rate persistence (25-35 percent conditional).


A tenant credit event has low probability in the base case (5-10 percent over 24 months) but material tail risk over 36-60 months given the concentration and the pace of hyperscaler capital deployment.


Combined probability of at least one material financing shock over 24 months is roughly 30-35 percent.



Asset exposure


Most exposed: PE-backed developers with 2020-2022 vintage construction debt and limited equity capacity for refinancing top-ups. Mid-tier developers with narrower LP bases are most exposed. Larger PE-backed developers (Blackstone-QTS, Brookfield-Compass, KKR-CyrusOne) have more equity capacity and are more insulated.


Moderately exposed: publicly-listed REITs. Their investment-grade unsecured debt is not exposed to refinancing distress but their equity valuations would compress on sector-wide spread widening.


Moderately exposed: specialty equipment vendors financed through LBO structures. Vertiv is a public issuer with limited refinancing exposure but historical LBO history is a reminder of the mechanism. Vertiv has since deleveraged substantially through equity market listing.


Least exposed: infrastructure funds and REITs with strong balance sheets and limited near-term refinancing needs. Digital Realty, Equinix, and Iron Mountain are relatively insulated on the debt side though not on equity valuation.



Recovery path


Financing shock recovery depends on whether the shock is rate-driven (recoverable when rate environment shifts) or credit-event-driven (recovery depends on specific event resolution). Rate-driven shocks typically have 18-36 month recovery windows. Credit-event-driven shocks have variable recovery depending on event severity and market response.



Chart 2. Asset exposure heatmap: layer × downside case


Exposure varies sharply across asset types. Developer platforms carry heavy exposure to both demand-side and financing shocks. Specialty semiconductor vendors are most exposed to supply-side reversal (technology substitution). Installed-base-annuity businesses are the most insulated across all three downside cases.



Author's exposure assessment. Darker cells = higher relative exposure to that downside case.




Compound scenarios and interaction effects


The three downside cases are not independent. Interaction effects amplify the potential impact of any single scenario and increase the joint probability of a material adverse outcome.


Downside 1 partially triggers Downside 2. A demand-side shock reduces the pace of new supply absorption, which effectively brings forward the point at which supply catches up with demand. Grid interconnection queues clear faster because the demand pressure on scarce sites eases. Transformer capacity that was fully absorbed at the demand peak becomes surplus at the compressed demand. The two mechanisms compound.


Downside 1 partially amplifies Downside 3. Lease amendment pressure from a demand-side shock reduces the cash flow available for debt service on developer construction debt. Coverage ratios that were 1.15x at refinance under Downside 3 alone become 0.95-1.05x under compound Downside 1 + Downside 3. This pushes more transactions from "difficult refinancing" to "failed refinancing" territory.


Downside 3 partially triggers Downside 1. A financing shock that widens hyperscaler cost of debt (through general market spread widening even without hyperscaler-specific credit deterioration) reduces hyperscaler debt capacity and could translate to capex compression at the margin. The effect is smaller than either Downside 1 or Downside 3 alone but not zero.


Downside 2 does not strongly interact with 1 or 3. Supply-side reversals have relatively independent triggers (queue reform, technology substitution) and do not amplify or trigger the other two cases at meaningful scale.


Joint probability weighting for compound scenarios:



ScenarioApproximate 24-month probabilityApproximate sector-wide equity impact

Any single downside (mild): 1, 2, or 360-70%-10% to -20%
Compound 1 + 3 (mild)10-15%-25% to -40%
Compound 1 + 2 (mild)10-15%-20% to -30%
Compound 1 + 2 + 3 (mild)4-8%-30% to -50%
Severe standalone (any single)10-15%-30% to -45%
Severe compound2-5%-45% to -65%




The impact figures are approximate and represent equity impact for a diversified AI infrastructure exposure; individual assets vary considerably. The point is that compound scenarios have material probability and material impact and are not sufficiently captured in current sector valuations.


Three downside cases: demand-side shock, supply-side reversal, financing shock. They interact. Compound scenarios have material probability and material impact.The downside frame



Chart 3. Joint and compound scenario probabilities, 24-month horizon


The compound scenarios do not receive proportionate weight in current sector valuations. Compound 1+3 has 10-15 percent probability and would produce 25-40 percent sector-wide equity impact. Compound 1+2+3 severe is a genuine tail risk but with material joint probability given the interaction mechanisms.



Author's subjective probability × impact estimates. Positioning is illustrative.




The specific signals to watch


Each of the three downside cases can be tracked through specific observable signals. Position adjustment ahead of confirming data is possible but requires accepting some optionality cost.



For Downside 1 (demand-side shock)


Hyperscaler capex guidance in quarterly earnings. Sequential compression of capex growth from 25-30 percent to 10-15 percent year-over-year would signal mild Downside 1 in progress. Compression below 5 percent would signal severe Downside 1.


AI-specific revenue disclosure. Meta, Google, Microsoft, Amazon disclosure of AI-related revenue at increasing granularity in 2025-2026 provides a direct signal. Underperformance versus consensus expectations by 15-20 percent for two consecutive quarters is a leading indicator of capex compression.


Enterprise AI-agent adoption pace. Third-party research (Gartner, IDC, Forrester) tracking enterprise adoption of AI-agent tools provides a leading indicator of enterprise cloud AI revenue. Slower-than-anticipated adoption trends translate to compressed enterprise cloud AI revenue with 6-12 month lag.



For Downside 2 (supply-side reversal)


ISO/utility queue clearance rates. Publicly-reported monthly queue clearance data from PJM, ERCOT, CAISO, and other ISOs provides a direct signal. Acceleration in queue clearance for datacenter-classified interconnection requests indicates the queue-reform mechanism working.


Transformer manufacturer capacity investment and shipment data. Hitachi Energy, Siemens Energy, Prolec GE, and specialty US independents disclose capacity investment and shipment volumes with quarterly granularity. Rising shipment volumes with declining lead times signals supply-side reversal in transformers.


Chinese specialty industrial export volumes. China customs data on transformer, cooling equipment, and specialty semiconductor exports to Western markets provides a signal of Chinese capacity release into Western pricing.



For Downside 3 (financing shock)


Successful refinancing of 2020-2022 vintage construction debt. Completion of major refinancings without material equity infusion, lease amendment, or restructuring indicates the market absorbing the rate reset. Failed or restructured refinancings signal broader stress.


Investment-grade and high-yield spread evolution. Sector-specific spread indices for infrastructure debt and datacenter debt provide leading indicators. Widening of 50+ basis points over a quarter is meaningful; widening of 100+ basis points is a genuine shock signal.


Hyperscaler credit rating actions. Any negative watch or downgrade on Meta, Google, Microsoft, or Amazon requires immediate re-evaluation of the entire sector debt profile.



Chart 4. Interaction effects: which cases trigger or amplify others


The three downside cases interact. Downside 1 (demand shock) partially triggers Downside 2 (supply catches up) and amplifies Downside 3 (compressed rents). Downside 3 partially triggers Downside 1 (financing shock compresses capex). Downside 2 has independent triggers and does not amplify others.



Arrow direction shows trigger/amplification. Line weight indicates approximate strength of the mechanism.




Chart 5. Recovery windows after each downside case triggers


Financing shocks recover fastest if rate-driven (18-36 months); demand-side shocks are the middle case (12-36 months depending on severity); supply-side reversals do not typically recover because once capacity is built it remains in the market. This has direct implications for how to structure any tail hedge.



Author estimates from precedent shock cycles. Individual cycles vary.




Position adjustment recommendations


The observations that follow are frameworks rather than specific recommendations, given that individual asset selections depend on portfolio mandate and time horizon. Nothing here is investment advice.


For investors with material AI infrastructure exposure, three portfolio adjustments merit consideration in the current environment.


First, tilt within the sector toward assets with structural exposure to installed-base annuity rather than to new-build growth. The distinction matters most in Downside 1 (demand-side shock) where installed-base annuity persists while new-build revenue compresses. Eaton, Vertiv Services, and Schneider Services segments benefit relative to pure new-build equipment sellers.


Second, tilt within developer exposure toward larger PE-backed platforms with strong equity capacity for refinancing top-ups rather than mid-tier platforms with narrower LP bases. The distinction matters most in Downside 3 (financing shock) where equity capacity is the key insurance against refinancing distress.


Third, consider tail-risk hedging structures. Options on relevant public equities (VRT, EQIX, DLR, ETN) and CDS on relevant credit issuers can provide protection against the sharp move that a severe or compound scenario would produce, at cost that has been reasonable through 2025-2026 given the low implied volatility in the sector.



Chart 6. Historical signals to watch dashboard, indicative levels for downside triggers


Six quantitative signals that would move the downside case probabilities in real time. Each is publicly observable. Left-to-right progression toward the "trigger" zone signals downside case in progress; sustained movement past the zone confirms the case is playing out.



Illustrative signal thresholds. Individual data series interpretation requires domain judgement.




What breaks this analysis


Two developments would compress the downside cases and improve the outlook. A material AI monetisation breakthrough is the first. If enterprise AI-agent revenue accelerates sharply through 2026-2027, or if consumer AI product revenue meets or exceeds current capex pace, Downside 1 probability compresses. This is a pace bet on the observable adoption curves rather than a technology bet. The second is a policy shift that meaningfully accelerates grid interconnection and transformer capacity. Federal and state policy delivering a step change in queue clearance or domestic transformer manufacturing would turn Downside 2 from a scenario probability into a structural change, and the affected asset exposures would reprice accordingly. That affects the sector composition rather than the sector level.




Glossary of terms used



CAISO
California Independent System Operator. The grid operator for most of California.
ERCOT
Electric Reliability Council of Texas. Grid operator for most of Texas, operating largely as an electrical island.
ISO
Independent System Operator. Regional grid operator (e.g. ERCOT, MISO, PJM, CAISO). Manages transmission access and wholesale market operations.
IVR
Integrated Voltage Regulator. On-package power delivery component for modern CPUs and GPUs.
LBO
Leveraged Buyout. Acquisition financed primarily with debt secured against the target's own cash flows.
LP
Limited Partner. Investor in a PE or infrastructure fund who commits capital but does not manage day-to-day investment decisions.
MV
Medium Voltage. Typically 1 kV to 35 kV. The voltage class connecting data centre power distribution to the utility grid.
PJM
PJM Interconnection LLC. Regional grid operator covering 13 US states plus DC. Includes Ohio, where the NVIDIA/OpenAI campus sits.
SiC
Silicon Carbide. Wide-bandgap semiconductor material used in high-voltage power electronics.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.





Sources and method


Probability assessments in this piece are the author's subjective estimates based on public information about the specific mechanisms and observable base rates. They are not derived from any structured probability model and should be read as directional rather than precise.


Signal-to-watch discussion rests on publicly available data series from ISO/utility filings, company disclosures, credit rating agency reports, and analyst commentary through mid-2026. Individual data-series interpretation requires domain judgement that the reader should apply.


The three downside cases are the author's construction of distinct scenarios visible in the sector as of mid-2026. Other possible downside scenarios exist (specific geopolitical events, regulatory intervention beyond what is discussed) but the three covered here are the most structurally likely given the current market state.


Impact estimates and portfolio recommendations are conceptual frameworks. Nothing in this piece constitutes investment advice.



Series


The Investment Layer (eight parts). Part I: How the AI buildout is financedPart II: The M&A mapPart III: The build-to-lease flipPart IV: The coverage asymmetryPart V: The PE playbook per layerPart VI: Underwriting AI infrastructure debtPart VII (this piece): The three downside casesPart VIII: Sovereign capital and industrial policyCompanion to the AI Power Chain series and to earlier commercial supplements Pricing Under Scarcity and The Services Inversion.


The author is an independent advisor working on data centre, power and industrial technology. This analysis is written in a personal capacity and rests entirely on public information. Nothing here is investment advice.







See also · The Investment Layer

 * How the AI buildout is financed
 * The AI infrastructure M&A map
 * The build-to-lease flip







In brief


Three downside scenarios stress the sponsor thesis for AI infrastructure. First, hyperscaler internal-build accelerates faster than modelled, compressing the independent operator market. Second, AI training capex plateaus earlier than the 2027-2030 forecast implies. Third, grid interconnection queue reform reduces the scarcity premium that current sponsor economics depend on. Each has different sensitivity to model input assumptions.







Due Diligence for the AI Buildout: the workstream index

 1. Part I. Framework: how DD works for AI infrastructure deals
 2. Product, technology, and roadmap DD
 3. Customer, pricing, and competitive DD


==============================================================================
# Sovereign capital and industrial policy
URL: https://adikumar.co/sovereign-capital-and-industrial-policy/
Published: 2026-08-17
Summary: Eight sovereign capital pools, twelve national industrial policy programmes shaping AI infrastructure, and how the landscape reshapes over 24 months.
==============================================================================

The Investment Layer series · Part 8 of 8








The Investment Layer · Part VIII of VIII



Sovereign capital and industrial policy


How eight sovereign capital pools have positioned in AI infrastructure, the policy instruments shaping their access, and what changes over the next 24 months as national AI-industrial strategies mature.




The Investment Layer

 1. Part I. How the AI buildout is financed
 2. Part II. The M&A map
 3. Part III. The build-to-lease flip
 4. Part IV. The coverage asymmetry
 5. Part V. The PE playbook, layer by layer
 6. Part VI. Underwriting AI infrastructure debt
 7. Part VII. The three downside cases
 8. Part VIII. Sovereign capital and industrial policy (you are here)




Executive summary


 * Eight sovereign capital pools have taken material positions in AI infrastructure globally as of mid-2026. Combined committed capital across these pools exceeds $180 billion. The composition and strategy vary sharply by pool.
 * The eight pools: GIC (Singapore), CPP Investments (Canada), ADIA (UAE), Mubadala (UAE), PIF (Saudi Arabia), Temasek (Singapore), NBIM (Norway), and China Investment Corporation with its subsidiaries. Additional smaller sovereign pools have committed but at more modest scale.
 * Direct sovereign involvement in national AI infrastructure programmes is now a stated policy of at least twelve national governments. The US CHIPS Act and IRA, EU Chips Act, Japanese JIC-led semiconductor programme, and Korean K-Chips Act are the largest. India's semiconductor mission and UAE's national AI computing capacity build-out are smaller in scale but strategically distinctive.
 * The typical sovereign investment structure is a structured minority stake in a developer or specialty industrial (as discussed in Part V), providing exposure and information rights without direct control that would trigger national-security scrutiny in the target's jurisdiction.
 * Industrial policy instruments (tax credits, direct grants, sovereign co-investment, export controls, procurement mandates) are shaping the addressable market for AI infrastructure vendors in ways that are jurisdiction-specific and often reduce cross-border optimisation.
 * Three structural developments over the next 24 months will reshape the sovereign-capital landscape. UAE's committed $30 billion investment in Global AI Infrastructure Investment Partnership (with BlackRock and Microsoft) is deploying through 2026. PIF has expanded AI infrastructure investment authority through 2026. The Japanese JIC's role in national semiconductor consolidation continues to expand.
 * The strategic consequence for private-market participants is that sovereign capital has become a first-order alternative to conventional PE and infrastructure fund financing for AI-adjacent transactions, and often provides more patient capital, larger cheque sizes, and different governance expectations than the traditional PE model.




Chart 1. Eight sovereign pools by estimated AI-infrastructure exposure, mid-2026


Total exposure across the eight pools exceeds $180 billion excluding Chinese domestic. GIC leads at $25-35 billion Western AI-infrastructure exposure. CPP Investments, PIF and ADIA are close behind. NBIM's exposure is primarily public-equity rather than direct private. CIC and Chinese subsidiaries carry huge domestic exposure but limited Western positions.



Estimated exposure from annual reports, industry disclosures, PitchBook and Preqin data. Individual pool exposure figures should be read as directional rather than precise.




The eight sovereign capital pools


Eight sovereign capital pools have taken material positions in AI infrastructure globally through mid-2026. Total committed capital across these pools targeting AI infrastructure directly or through infrastructure funds with material AI-infrastructure exposure exceeds $180 billion. The composition and strategy vary sharply by pool.



GIC (Government of Singapore Investment Corporation)


GIC has been the most consistent sovereign investor in AI infrastructure since 2019, with an approach that emphasises structured minority stakes in developers, joint ventures with strategic partners, and infrastructure fund LP participation. Its estimated AI-infrastructure exposure across all instruments is $25-35 billion.


Notable positions include minority stakes in Digital Realty joint ventures, Vantage co-investment via DigitalBridge, and multiple infrastructure fund commitments to Blackstone Infrastructure, Brookfield Infrastructure, and KKR Infrastructure. GIC's approach is characterised by long duration (10-15 year holding expectations), preference for governance rights but not control, and willingness to co-invest with sponsors it has strong existing relationships with.



CPP Investments (Canada Pension Plan Investment Board)


CPP Investments has expanded its AI-infrastructure allocation substantially since 2022, with an emphasis on direct co-investment and lower reliance on fund commitments than in prior periods. Total AI-infrastructure exposure is estimated at $20-30 billion.


CPP has been particularly active in datacenter infrastructure debt, including senior positions in construction financings, hybrid structures, and mezzanine debt against developer platforms. On the equity side, CPP has taken direct stakes in Compass through infrastructure co-investment vehicles, and holds structural exposure to Vantage, Aligned and STACK through fund commitments to their respective sponsors.



ADIA (Abu Dhabi Investment Authority)


ADIA's approach emphasises broad infrastructure exposure through fund commitments and less direct participation than some other Middle East pools. Total AI-infrastructure exposure is estimated at $15-25 billion, primarily through Blackstone, Brookfield, KKR, EQT and other major infrastructure fund LP positions.


ADIA has been an early and consistent LP in the largest infrastructure funds and has scaled commitments in line with fund vintage growth. Direct positions are less common than fund commitments but do exist in specific transactions where ADIA has been invited as co-investor.



Mubadala Investment Company


Mubadala's approach differs sharply from ADIA's, with more direct strategic positioning and heavier concentration in specific technology and infrastructure exposures. Total AI-infrastructure exposure is estimated at $10-18 billion.


Notable positions include the recent $2 billion+ commitment to AI infrastructure through Global AI Infrastructure Investment Partnership (with BlackRock and Microsoft) announced in 2024 and progressively deploying through 2026. Mubadala also holds strategic positions in specialty semiconductor firms and direct co-investments in Middle East and European datacenter development.



PIF (Saudi Public Investment Fund)


PIF has expanded AI-infrastructure allocation substantially since 2023, with a mix of direct national programme investment (Saudi Arabia's national AI computing programme) and international infrastructure fund commitments. Total AI-infrastructure exposure is estimated at $20-30 billion, with substantial additional planned deployment through 2027.


PIF's HUMAIN entity, established in 2024-2025, holds direct investments in Saudi national AI infrastructure and has begun making international investments in strategic AI-adjacent segments. PIF has also been a substantial LP in Blackstone, Brookfield and BlackRock alternatives platforms, with substantially expanded commitments in 2024-2026.



Temasek Holdings


Temasek's AI-infrastructure exposure is estimated at $12-20 billion, with a mix of direct positions in Asian datacenter operators and infrastructure fund commitments. Notable direct positions include stakes in Vantage's Asia platform, Digital Realty Asia JVs, and specialty semiconductor firms.


Temasek's approach emphasises Asian regional positioning and specialty semiconductor exposure that reflects Singapore's strategic priorities. The pool has been relatively active in early-stage and growth-stage specialty semiconductor investments where other sovereign pools have been less active.



NBIM (Norwegian Government Pension Fund Global)


NBIM's approach is primarily through public-equity positions in datacenter REITs, hyperscalers, and specialty industrial equity, plus limited infrastructure fund commitments. Direct private AI-infrastructure exposure is more modest than the other pools at estimated $8-15 billion.


NBIM's public-equity positioning includes material stakes in Digital Realty, Equinix, Iron Mountain, and the specialty industrial names (Vertiv, Eaton, Schneider Electric, ABB, Siemens Energy). The pool's constraints on private-equity concentration limit its direct private participation but the public-equity exposure is substantial in absolute terms.



China Investment Corporation and subsidiaries


CIC and its subsidiaries (SAFE Investment Company, Central Huijin) hold substantial positions in Chinese domestic AI-infrastructure but limited direct positions in Western AI-infrastructure given the current regulatory environment. Chinese domestic exposure is estimated at ¥1.2-1.5 trillion equivalent (approximately $170-210 billion), primarily through direct investment in Chinese domestic datacenter operators and specialty semiconductor firms.


Chinese sovereign pools have progressively reduced exposure to Western infrastructure and technology since 2022 as CFIUS review and comparable European mechanisms have restricted new positions. The Chinese domestic buildout is now largely self-financed through domestic sovereign, policy-bank, and PE structures.



Chart 2. Sovereign pool exposure by preferred investment structure


Different pools favour very different structures. GIC and CPP lean toward direct co-investment and infrastructure fund LP. ADIA is heaviest on fund commitments. Mubadala and PIF are heaviest on direct strategic positioning. Temasek concentrates on Asian regional positioning. NBIM is nearly all public-equity.



Author's estimate of structural composition. Individual pools vary considerably around the typical pattern.




Industrial policy instruments


Twelve national governments have adopted material industrial policy instruments targeting AI infrastructure. The specific instruments and scale vary considerably; some produce material addressable-market changes for vendors and others are more symbolic.



US CHIPS and Science Act, plus Inflation Reduction Act


The CHIPS Act provides $52 billion for semiconductor manufacturing incentives, including specific carve-outs for advanced packaging and wide-bandgap semiconductor manufacturing. The IRA provides tax credits for clean-energy and electricity infrastructure that indirectly support datacenter interconnection and behind-the-meter power investments.


The addressable-market impact for AI infrastructure vendors is meaningful in specific segments. US wide-bandgap semiconductor vendors (Wolfspeed, Onsemi US SiC manufacturing) benefit directly. US-based advanced packaging investment (Intel Foundry, TSMC Arizona, Samsung Texas) creates addressable market for on-package power delivery, thermal management and interconnect specialists. US transformer and switchgear manufacturing benefits from procurement preference in federally-supported projects.



EU Chips Act


The EU Chips Act allocates €43 billion (approximately $47 billion equivalent) toward European semiconductor manufacturing, with specific focus on mature-node and specialty semiconductor manufacturing including wide-bandgap and power semiconductors. Direct impact on AI infrastructure is smaller than the US CHIPS Act because Europe's AI infrastructure buildout is smaller in scale, but the specialty semiconductor impact is material.


Notable beneficiaries include Infineon (German wide-bandgap manufacturing), STMicroelectronics (Italian and French SiC and GaN capacity), and X-FAB (German mature-node manufacturing with power semiconductor exposure).



Japanese semiconductor programme via JIC


Japan Investment Corporation and its subsidiaries (JIC Capital, JIC Venture Growth Investments) have been actively consolidating Japanese specialty semiconductor and industrial firms through take-private and structured minority investment. Total committed capital is estimated at ¥3-5 trillion (approximately $20-33 billion equivalent) through 2026.


Notable transactions include the pending Shinko Electric take-private (approximately ¥690 billion), the Kioxia recapitalisation, and multiple smaller specialty semiconductor investments. The JIC approach is more aggressive on control transactions than most other sovereign pools, reflecting Japan's specific industrial policy priorities around specialty semiconductor and materials.



Korean K-Chips Act


The K-Chips Act provides Korean semiconductor manufacturers (Samsung, SK Hynix) with expanded tax credits and R&D support totaling approximately $70 billion equivalent through 2030. The specific AI-infrastructure impact is via HBM memory manufacturing capacity expansion (SK Hynix, Samsung), which is a material bottleneck for AI accelerator production.



UAE national AI infrastructure programme


The UAE has committed substantial national resources to AI infrastructure through multiple entities. Mubadala's investments and G42 (which is separately structured) have deployed approximately $10-15 billion of national capital into AI infrastructure. The Global AI Infrastructure Investment Partnership announced in 2024 targets $100 billion total deployment through 2030, with UAE, US, and Microsoft as anchor participants.



Other significant national programmes


India's semiconductor mission ($10 billion equivalent through 2030), Singapore's national AI strategy (SGD 1 billion equivalent through 2027), Malaysia's national semiconductor strategy (MYR 25 billion equivalent through 2030), and UK's National Semiconductor Strategy (£1 billion through 2030) are all smaller in scale but strategically significant in their respective jurisdictions. Australia's Critical Technologies Fund, France's semiconductor sovereignty initiative, and Germany's specialty semiconductor investment programmes fit in the same category.



Chart 3. National industrial policy programmes with AI-infrastructure exposure, authorised capital


The US CHIPS Act (with IRA) is the largest programme by authorised capital. Korean K-Chips Act is second. Others are smaller in absolute terms but strategically significant in their jurisdictions. Disbursement typically lags authorisation by 18-36 months, which is where actual addressable-market impact starts to show.



Authorised amounts from published legislation. Actual disbursement lags materially in most programmes.




The typical sovereign investment structure


The typical sovereign investment structure in AI infrastructure across the pools discussed above is a structured minority stake, providing exposure and information rights without control that would trigger national-security review in the target's jurisdiction.


Structural features:


Ownership stake typically 5-40 percent, with defined governance rights (board observer or board seat at higher stakes, information rights, protective consent rights for specified major decisions).


Investment horizon typically 10-15 years or perpetual, with defined liquidity provisions (put rights, drag-along rights, IPO participation rights) that provide exit optionality without forcing exit timing.


Return expectations aligned with long-duration infrastructure return targets (7-11 percent unlevered) rather than PE-style return targets (15-25 percent unlevered), reflecting the LP mandate structure of the sovereign pools.


Alignment mechanisms typically less aggressive than PE-style alignment (management incentive plans are smaller, drag-along rights are more limited, governance intervention is less common) because the sovereign investor's objectives include strategic exposure as much as absolute return maximisation.


Sovereign capital has become a first-order alternative to conventional PE and infrastructure fund financing for AI-adjacent transactions. Larger cheque sizes, longer duration, different governance expectations.The alternative capital pool



Chart 4. Sovereign pool AI-infrastructure commitment growth, 2019-2026


All eight pools have grown AI-infrastructure exposure since 2019, with the sharpest acceleration post-2023 as post-ChatGPT deal opportunities scaled. PIF and Mubadala have expanded fastest in percentage terms. GIC and CPP have added the largest absolute dollars.



Estimated commitments to identifiable AI-infrastructure transactions per year. Indexed at 2019 = 100.




Cross-border scrutiny mechanisms


The cross-border movement of sovereign capital into AI infrastructure is subject to material national-security and competition review across most Western jurisdictions. The specific mechanisms vary but the overall direction is toward tighter scrutiny.


United States. CFIUS (Committee on Foreign Investment in the United States) reviews foreign investment in US critical infrastructure and technology, including datacenter and semiconductor investments. Reviews of Chinese sovereign investment have effectively blocked new positions since 2020. Reviews of Middle East, Singaporean and Norwegian sovereign investment have generally cleared but with defined mitigation agreements in some cases.


United Kingdom. The National Security and Investment Act 2021 provides comparable review authority to CFIUS for UK critical infrastructure and technology. UK reviews have progressively tightened, with the Nexperia divestment (Chinese-owned Newport Wafer Fab, forced divestment in 2023) as the most visible example.


European Union. FDI screening at both EU and member-state level has expanded, with Germany, France, and Italy having particularly active screening regimes. Chinese sovereign investment in strategic sectors is essentially blocked; other sovereign pools face increased friction on strategic-sector investments.


Japan. The Japan Foreign Exchange and Foreign Trade Act provides review authority for foreign investment in Japanese strategic sectors including semiconductor manufacturing. Reviews have tightened since 2022 particularly for Chinese and adjacent sources of capital.


The consequence is a structurally different pattern of cross-border sovereign flow than existed pre-2020. Chinese sovereign capital is essentially absent from Western AI infrastructure. Middle East sovereign capital has become more prominent in Western transactions, often structured to minimise scrutiny risk. Norwegian and Canadian sovereign capital remains active with limited friction. Singapore sovereign capital is broadly welcomed across Western jurisdictions.



Chart 5. Cross-border scrutiny actions on sovereign / strategic AI-adjacent transactions, 2019-2026


Scrutiny has intensified year-over-year across CFIUS, UK NSI Act, EU FDI screening, and Japanese and Australian equivalents. Chinese sovereign investment in Western strategic sectors is now essentially blocked. Middle East and Asian sovereign capital face increasing but still-navigable friction.



Author's count of publicly-reported reviews, blocks, and forced divestments in AI-adjacent sectors across jurisdictions.




What changes over the next 24 months


Three structural developments over the next 24 months will reshape the sovereign capital and industrial policy landscape for AI infrastructure.



Development 1. UAE Global AI Infrastructure Investment Partnership deployment


The GAIIP entity announced in September 2024 with UAE, BlackRock, Microsoft, and MGX as anchor participants targets $30 billion of initial capital and $100 billion total deployment through 2030. Deployment is progressing through 2025-2026 with initial investments in US datacenter development, energy infrastructure supporting AI datacenters, and specialty industrial capacity.


The strategic significance is that GAIIP provides a large-cheque-size vehicle for UAE sovereign capital to deploy into Western AI infrastructure at scale without individual transaction national-security friction that would apply to direct UAE sovereign investment. The scale of the vehicle (potentially $100 billion total) is comparable to the entire Western AI infrastructure debt market outstanding today.



Development 2. PIF expanded AI infrastructure authority


PIF's investment authority for AI infrastructure has expanded substantially in 2024-2026, with HUMAIN as the primary operational vehicle and additional co-investment structures alongside major PE and infrastructure sponsors. Deployment through 2026-2027 is likely to include material US and European datacenter investments plus specialty industrial equity and debt positions.


The specific PIF strategic priorities include domestic Saudi Arabia AI infrastructure buildout (which is progressing at $10+ billion committed capital) and international positions that support Saudi Arabia's AI technology development. HUMAIN's technology partnerships with Nvidia, AMD, and Qualcomm announced in 2024-2025 indicate the direction of the international strategy.



Development 3. Japanese JIC continuing specialty semiconductor consolidation


The JIC's role in Japanese specialty semiconductor consolidation is likely to expand over the next 24 months. Beyond the pending Shinko Electric transaction, JIC has been discussed as a potential participant in additional take-privates of specialty semiconductor and materials firms with material AI-infrastructure exposure. The Ibiden substrate business, Kyocera specialty materials, and specialty chemical firms have all been discussed as candidates.


The strategic priority for JIC is maintaining Japanese ownership of specialty semiconductor and materials capacity that is strategically important for the global AI infrastructure buildout. Take-private structures with JIC anchor participation combined with strategic-partner co-investment (typically a Japanese conglomerate) is the emerging model.



Chart 6. GAIIP and PIF-HUMAIN deployment schedule, cumulative through 2030


GAIIP's target deployment of $100 billion by 2030 (with $30 billion initial anchor per Reuters + fund announcements) plus PIF-HUMAIN commitments approaching $30-40 billion internationally per public announcements together represent the largest single-source expansion of Western AI-infrastructure sovereign capital in the current cycle.



Projected deployment schedules from public announcements. Actual pace subject to specific transaction outcomes and regulatory approvals.




The strategic consequence for private-market participants


Sovereign capital has become a first-order alternative to conventional PE and infrastructure fund financing for AI-adjacent transactions. The consequence for private-market participants across the six layers of the AI infrastructure chain is threefold.


First, transaction structures increasingly include sovereign minority participation as a matter of course. A large developer scale-out that would have been 100 percent PE-owned five years ago is now typically 65-80 percent PE-owned with 20-35 percent sovereign minority participation across one or more pools. The structure is more complex to negotiate but provides larger transaction capacity and often reduces PE fund pressure to exit within typical fund horizons.


Second, sovereign capital provides an exit path alternative to strategic sale, IPO, and secondary sale to another PE fund. Sovereign pools can take over PE positions at fund end-of-life with return expectations that PE sponsors can meet without forcing distress sales. This partially resolves the exit-path capacity problem discussed in Part II, though it does not eliminate it.



Third, sovereign capital availability supports valuation levels that would not be sustainable in a market with only conventional PE and infrastructure fund financing. The specific price levels for hyperscale-tenanted development, specialty semiconductor take-privates, and structured infrastructure debt reflect a broader capital pool than existed pre-2020, and that broader pool provides both floor support and appetite for larger transactions than the traditional market could absorb.



What breaks this landscape


Two developments would reshape the sovereign capital and industrial policy landscape in ways that would compress valuations or reduce sovereign participation.


Tightening cross-border scrutiny. If Western national-security review regimes tighten further (particularly against Middle East and non-aligned Asian sovereign capital), sovereign participation in Western AI infrastructure could compress meaningfully. This would remove a first-order source of capital and would translate to compressed valuations across the sector.


The trajectory of this risk is jurisdiction-specific. US CFIUS has been broadly stable in its approach to non-Chinese sovereign capital but could tighten under different administrative priorities. EU FDI screening has tightened progressively and could tighten further. UK NSI Act has tightened and could continue to tighten. The direction is generally toward more scrutiny rather than less.


National industrial policy failure. If US CHIPS Act, EU Chips Act, or Japanese semiconductor programmes fail to deliver anticipated addressable-market expansion (through implementation delays, funding shortfalls, or geopolitical disruption), the vendors positioned to benefit would face compressed forward growth expectations. This would reprice specific segments (US wide-bandgap semiconductor, US transformer manufacturing, EU specialty semiconductor) more than the broader sector.



What to watch


GAIIP deployment cadence and specific transactions are the leading signal. The first tranche of GAIIP transactions will show whether the vehicle deploys at scale and what target segments it prioritises. Delays or scaled-back deployment would signal reduced UAE sovereign appetite; on-schedule deployment at scale would confirm continued strong appetite. Cross-border scrutiny actions are the second: any CFIUS, NSI Act, or EU FDI action against a sovereign transaction reshapes the risk framework immediately, with blocks or forced divestments carrying the largest impact. National industrial policy funding disbursement is the third. CHIPS Act grant disbursements, EU Chips Act allocations, and JIC transaction closings all indicate whether the policy commitments translate to actual capital deployment, with disbursement delays translating to addressable-market delays for vendors. Sovereign pool allocation disclosures are the fourth. GIC, CPP Investments, NBIM, and Temasek publish annual reports with increasing AI-infrastructure granularity, and growth in disclosed allocation signals continued appetite while compression signals a pullback.



Closing observation


The AI infrastructure buildout is the largest single industrial-capital deployment of the current decade. The Investment Layer series has decomposed the financing that supports this deployment across eight distinct essays. The picture that emerges is different from the popular framing that treats AI infrastructure primarily as a hyperscaler-driven balance-sheet story.


The actual financing is fragmented across at least six major pools (hyperscaler operating cash flow, REIT equity and debt, OEM balance-sheet, structured infrastructure debt, sovereign, PE), executed through at least five distinct playbook patterns (developer scale-out, platform roll-up, take-private, structured minority, specialty carve-out), rated and priced through frameworks that are converging but still have material soft areas, exposed to at least three distinct downside cases with interaction effects, and shaped by industrial policy across at least twelve national jurisdictions.


The strategic map is not simple. Understanding it requires holding multiple parallel frames simultaneously: the physical AI infrastructure map (covered in the AI Power Chain six-part series and technical companions), the commercial map (covered in Pricing Under Scarcity and The Services Inversion), and now the financing map covered in this Investment Layer series.


Investors, executives, and policy participants who hold all three frames together have a substantially better view of where value is being created, where it is being captured, and where it is being priced correctly relative to the underlying reality. Those who hold only one frame miss the interaction effects that determine outcomes.




Glossary of terms used



CFIUS
Committee on Foreign Investment in the United States. US inter-agency review body that screens foreign acquisitions of US businesses for national security implications.
EU FDI
European Union Foreign Direct Investment screening. Country-specific regulatory regimes reviewing foreign acquisitions of critical assets.
GaN
Gallium Nitride. Wide-bandgap semiconductor material used in RF and power applications.
HBM
High Bandwidth Memory. Stacked memory technology used adjacent to AI GPUs.
IPO
Initial Public Offering. A private company's first sale of shares to the public market.
LP
Limited Partner. Investor in a PE or infrastructure fund who commits capital but does not manage day-to-day investment decisions.
NSI Act
National Security and Investment Act. UK regime for reviewing acquisitions in 17 sensitive sectors on national security grounds.
SiC
Silicon Carbide. Wide-bandgap semiconductor material used in high-voltage power electronics.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.





Sources and method


Sovereign pool exposure estimates compiled from published annual reports (GIC, NBIM, CPP Investments, Temasek), industry disclosures, PitchBook and Preqin LP-facing data, and reasonable inference from public transaction records through mid-2026. Individual pool exposure figures should be read as directional rather than precise.


Industrial policy programme figures drawn from published legislation, government announcements, and third-party analysis. Programme scale figures represent authorised or committed capital rather than disbursed capital in most cases; actual disbursement typically lags authorisation by 18-36 months.


The cross-border scrutiny discussion rests on published policy from CFIUS, UK NSI Act guidance, EU FDI screening reports, and comparable jurisdiction disclosures through mid-2026. Individual transaction reviews are typically confidential; broader trends are visible from published statistics.


The three developments discussed in the "next 24 months" section rest on public announcements and reasonable inference from public commentary. They do not reflect any specific insider knowledge and should be read as analytical hypotheses rather than as reported specific process activity.



Series


The Investment Layer (eight parts, complete). Part I: How the AI buildout is financedPart II: The M&A mapPart III: The build-to-lease flipPart IV: The coverage asymmetryPart V: The PE playbook per layerPart VI: Underwriting AI infrastructure debtPart VII: The three downside casesPart VIII (this piece): Sovereign capital and industrial policyCompanion to the AI Power Chain series (six parts plus six technical companions), and to earlier commercial supplements Pricing Under Scarcity and The Services Inversion.


The author is an independent advisor working on data centre, power and industrial technology. This analysis is written in a personal capacity and rests entirely on public information. Nothing here is investment advice.









See also · The Investment Layer

 * How the AI buildout is financed
 * The AI infrastructure M&A map
 * The build-to-lease flip





Further reading + sources
Related essays

 * How the AI buildout is financed. the capital-structure map
 * Financing VIII. Sovereign wealth. the specific SWF deployment lens
 * Regulatory V. CFIUS + Outbound Investment. the regulatory overlay


Primary sources

 * US Treasury CFIUS. the statutory framework
 * Saudi PIF. the reference sovereign for Middle East AI deployment
 * Mubadala. the reference sovereign for the G42 / MGX ecosystem







In brief


Sovereign wealth capital is a growing share of AI infrastructure financing. GAIIP targets $100B deployment by 2030 with a $30B initial anchor per Reuters + fund announcements. PIF-HUMAIN commitments approach $30-40B internationally per public announcements. Industrial policy in the US (CHIPS Act, IRA credits) and EU (AI Act, DORA) reshapes where sovereign capital can deploy and on what terms.








Frequently asked


How much sovereign wealth is committed to AI infrastructure?

Global AI Infrastructure Partnership (GAIIP) targets $100B deployment by 2030 with a $30B initial anchor per Reuters and fund announcements. PIF-HUMAIN commitments approach $30-40B internationally per public announcements. Additional Middle East, Singapore and Norwegian sovereign vehicles are deploying at smaller scale. Aggregate figure by 2030 could exceed $200B.


==============================================================================
# Arc behaviour, protection, and how insurance is catching up (DC-DC Transition IV)
URL: https://adikumar.co/dc-dc-transition-04-arc-behaviour-insurance/
Published: 2026-08-20
Summary: DC arcs need microsecond-scale interruption. Schneider Aug 2026 study finds 800 VDC arc risk manageable with SSCB + arc-fault detection.
==============================================================================

The DC-DC Transition series · Part 4 of 19









The DC-DC Transition · Part IV of XIII



Arc behaviour, protection topology, and how insurance is catching up


DC arcs cannot self-extinguish at zero crossings. Because there are none. But 800 VDC arc-flash risk is manageable with the right protection design; Schneider Electric's August 2026 study found the incident-energy exposure comparable to typical AC systems when architecture, capacitor placement, and fault-clearing behaviour are specified correctly. What operators need to lock in now is the protection topology, the vendor spec, and the insurance underwriting conversation.


Part IV in the DC-DC Transition series · independent analysis · no advisory conflict on any named party





The DC-DC Transition · thirteen essays for data centre power architects

 1.  I. The real reason data centres are going DC
 2.  II. Two architectures wearing the same name
 3.  III. The architecture map
 4.  IV. Arc behaviour + insurance (you are here)
 5.  V. Grounding + ground-fault protection
 6.  VI. Battery integration at 800 VDC
 7.  VII. Retrofit vs greenfield: the decision framework
 8.  VIII. 800 VDC and liquid cooling co-emergence
 9.  IX. Power quality + grid interaction
 10. X. Standards: OCP, IEC, NEC, IEEE
 11. XI. Vendor economics: who wins the transition
 12. XII. Commissioning, skills, operational readiness
 13. XIII. The ten-year view





01Why DC arcs are physically different



An alternating-current arc has one useful property that a direct-current arc does not: the fundamental sinusoid crosses zero 100 or 120 times per second, and every crossing gives a protective device a natural opportunity to interrupt current without having to fight the arc column. A conventional AC circuit breaker exploits this ruthlessly. Under fault conditions the mechanical contacts open, an arc bridges the gap for a brief interval, and the next zero crossing extinguishes it. Total clearing time is one to two half-cycles. 8 to 17 milliseconds on a 50/60 Hz system.



Direct current has no such property. The current in a DC fault is continuous. If the protective device cannot dissipate the fault energy fast enough, the arc column sustains, and the incident energy released at the operator plane climbs as the square of the exposure time. The failure mode is not a smaller version of an AC fault; it is a different failure mode.



Three variables collectively determine the outcome. First, the fault-current profile at the point of the arc. How much current flows, how it grows over time, and what the source impedance looks like. Second, the fault-clearing time. The interval between fault initiation and current interruption. Third, the arc-gap geometry. Conductor spacing, enclosure design, and pressure relief. In an AC world the second variable is bounded by the physics of the zero crossing. In DC it is bounded by the protective device you have selected.




Chart 1. Fault current profile: AC vs 800 VDC, single-cycle window


The AC waveform crosses zero twice per cycle. Each zero crossing is a natural interruption opportunity for a mechanical breaker. The DC waveform is a step function that stays at fault current until the protective device forces it down. Arc energy accumulates as the integral under the curve.



Illustrative profile at 800 V; actual fault current depends on source impedance and capacitive contribution.





02The fault-clearing time equation and why microseconds matter



Arc-flash incident energy at a working distance is modelled as a function of arcing current, arcing time, gap distance, and enclosure geometry. The IEEE 1584 model is the industry reference for AC systems and has been extended for DC applications over the last five years. For the working data centre operator the equation collapses to a single practical rule: incident energy scales roughly linearly with fault-clearing time, and the difference between a millisecond-scale device and a microsecond-scale device is the difference between category-2 PPE requirements and category-4-or-worse.



The protection technologies available in 2026 sit on a wide range along this axis. Mechanical DC circuit breakers, adapted from industrial and traction applications, clear in 20 to 50 milliseconds. Current-limiting fuses clear in 2 to 10 milliseconds, at the cost of one-time-use and coordination complexity. Semiconductor-based SSCB devices, now emerging in commercial volume, clear below 100 microseconds. Two orders of magnitude faster than mechanical breakers and one order faster than current-limiting fuses.




Chart 2. Fault-clearing time by protection technology: microseconds, milliseconds, tens of milliseconds


Semiconductor-based SSCBs (Menlo Micro, ABB, Eaton, Schneider new lines) clear in under 100 microseconds. Current-limiting fuses (Mersen, Bussmann/Eaton, Littelfuse) clear in 2-10 milliseconds. Mechanical DC breakers clear in 20-50 milliseconds. The choice determines what arc-flash PPE category applies to normal-operations work.



Vendor datasheet compilations 2024-2026. Actual clearing time depends on fault current magnitude and device coordination settings.




The interesting technical detail is that SSCBs solve a problem that AC systems never had to solve at scale. In an AC system the zero crossing does the interruption work for you; you just need contacts that can withstand the transient recovery voltage. In DC the semiconductor has to actively force current down under load, which requires precise gate drive, thermal management, and coordinated protection against reverse recovery. The devices are more complex than the mechanical breakers they replace, but the clearing-time advantage is what makes 800 VDC operationally safe at hyperscale.




03The Schneider Electric August 2026 study and what it found



The Schneider Electric arc-flash study published August 2026 is the most substantive publicly available analysis of 800 VDC arc-flash risk to date, and it deserves to become required reading for anyone specifying a large-scale DC facility. The study modelled incident energy under a range of architectural scenarios and reached three findings that shape the design conversation.



The first finding is that arc-flash risk in 800 VDC systems is manageable and, under many realistic architectures, comparable to typical AC systems. This is the headline result and it counters the intuition that DC is categorically more dangerous. What matters is the architecture, not the voltage class per se.



The second finding is that outcomes depend strongly on capacitor placement and fault-clearing behaviour. Capacitor-dominated systems. Where the fault current is fed largely by discharge from bus capacitance rather than from the source. Behave differently from source-dominated systems, and the SSCB is particularly suited to the capacitor-dominant regime because it can stop capacitor discharge before the arc column stabilises.



The third finding, less-emphasised in the coverage but arguably the most important for operator specification, is that the arc-fault risk profile is a design output of the electrical architecture, not an input to it. Two facilities operating at the same 800 VDC bus voltage can present very different arc-flash exposures depending on how the source, bus, capacitance, and protection are configured. This means the DD-lens arc-flash workstream is not "is this facility safe?" but "does this specific architecture as specified produce an incident-energy profile that fits within PPE category 2 for normal operations?"




04The IEC 60947-10 standard and the certification framework



The regulatory-standards side of this conversation has moved faster than most operators realise. IEC 60947-10, the standard covering semiconductor-based circuit breakers, was published in Europe in early 2026 and is now the governing document for SSCB certification. Its publication is what makes SSCBs commercially specifiable rather than experimental. Before IEC 60947-10, procurement teams could not point to a governing standard and insurers could not underwrite against one.



UL is drafting the North American equivalent through active participation in the OCP 800 VDC working group. NFPA is updating NFPA 70E (the arc-flash safety standard) to explicitly address 800 VDC installations. IEEE P2818, the emerging standard for high-voltage DC data centres, is in committee. Full standardisation lands 2028-2029; interim design work must reference multiple standards concurrently.




Chart 3. Standards evolution timeline for 800 VDC arc-fault protection and certification


IEC 60947-10 published early 2026 anchors the European framework. UL, NFPA 70E, and IEEE P2818 catch up over 2026-2029. Interim procurement specifications reference multiple standards simultaneously; full convergence expected 2028-2029.



Sources: IEC publication register, UL Solutions OCP working group participation, NFPA 70E revision cycle, IEEE P2818 project schedule.





05The insurance market posture



The insurance conversation is the least mature part of this workstream and the one operators are most likely to underestimate. Property insurers writing data centre exposures are still working from limited actuarial data on 800 VDC facilities. Their default posture in 2026 is to price the unknown. The premium premium (over an equivalent AC facility) currently runs 20 to 60 percent depending on the carrier, the architecture specified, and the operator's demonstrated engineering discipline.



Three questions dominate the underwriting conversation as I have observed it. What is the specific arc-fault protection topology at the switchboard and rack level (SSCB, current-limiting fuse, hybrid, mechanical)? What is the fault-current study on record and does it demonstrate PPE category 2 or better for normal-operations work? What is the operator's arc-fault detection and monitoring plan in operations (not just at commissioning)?




Chart 4. Insurance premium premium at 800 VDC vs AC baseline, by protection tier


Carriers are pricing three tiers. Tier 1: mechanical breakers only, no arc-fault detection. 45-60 percent premium. Tier 2: current-limiting fuses plus arc-fault detection. 20-35 percent premium. Tier 3: SSCBs at bus and rack level plus continuous arc-fault detection. 5-15 percent premium. The economic case for tier 3 pays back in 2-3 years on premium alone.



Author's synthesis from operator conversations mid-2026. Actual premiums vary by carrier, facility scale, and jurisdiction.




Operators walking into an insurance conversation without a specific protection topology story generally end up in tier 1 pricing. Operators walking in with a documented SSCB deployment and continuous arc-fault detection generally end up in tier 3. The three-tier gap is measured in low single-digit millions of dollars per year on a large facility, which is a substantial economic case for the tier-3 specification even before considering the operational-safety case.




06Real incidents worth learning from



Public data on 800 VDC arc-flash incidents in data centre applications is thin because the installed base is small and operators have limited incentive to publicise events. The adjacent industries with substantial DC arc exposure. Industrial rectifier plants, EV fast-charging, telecom central offices at 48 VDC, submarine and aerospace power. Supply the useful reference cases.



The telecom industry ran on 48 VDC for decades and accumulated a body of operational experience that transferred directly to early data centre DC deployments. The lesson from that history is that most DC incidents are not arc-flash in the initiating event; they are sustained arcs following an initiating failure (loose connection, thermal degradation, moisture ingress) that would have cleared on an AC system. The protection specification that mattered was continuous monitoring for developing faults, not just fault-current interruption.



Industrial rectifier plants provide the second useful reference. Aluminium smelting and chlor-alkali production run at high-current DC and have decades of arc-flash incident data. The IEEE 1584 extension for DC systems was substantially informed by this experience. The practical lesson for data centre operators is that arc-flash exposure is heaviest during maintenance events, not during normal operations, and the PPE conversation and the maintenance-procedure conversation both need to sit inside the protection-topology conversation rather than beside it.



EV fast-charging is the newest reference and the one most similar in system dynamics to 800 VDC data centre power. The 350-kW charging stations running at 800 to 1000 V DC have generated a small but growing incident record. The failure modes documented publicly include connector arcing during disconnect under load, cable arc-tracking following insulation damage, and rectifier internal faults. Each of these has a specific analogue in the 800 VDC data centre context.




07What vendors are shipping in 2026



The vendor set for 800 VDC arc-fault protection sorts into three categories. Established power-electronics vendors have added SSCB and DC arc-fault detection to their existing DC portfolios: ABB (the FormulaDC line), Eaton (Bussmann-branded DC fuses plus new SSCB modules), Schneider Electric (SF6-free MV plus SSCB), Siemens Energy, and Hitachi Energy. Semiconductor specialists have built dedicated SSCB businesses: Menlo Micro is the most-cited pure-play, with claimed 100-microsecond clearing at 800 VDC. Emerging hybrid vendors are combining traditional protection with advanced monitoring: Bender for insulation-resistance monitoring, MessTek for DC arc-fault detection, and OEM partnerships between these players and the incumbent breaker vendors.





CategoryVendor examplesProduct classClearing timeCertification status


Semiconductor SSCBMenlo Micro, ABB, Eaton, Schneider (new lines)Solid-state DC breaker<100 µsIEC 60947-10 published; UL in progress
Current-limiting fuseMersen, Bussmann (Eaton), LittelfuseFast-acting DC fuse2-10 msEstablished IEC / UL certified
Mechanical DC breakerABB, Schneider, Siemens, Hitachi EnergyMolded-case DC breaker20-50 msEstablished; traction and industrial derivations
Arc-fault detectionMessTek, Bender, Sensata, SchneiderContinuous monitorDetection: 5-50 msDIN / IEC referenced; NFPA in draft
Insulation-resistance monitorBender, Megger, GfS, MessTekUngrounded system monitorDetection: minutes to hoursIEC / UL certified






Chart 5. Arc-flash incident energy vs voltage class and clearing time: where PPE category 2 lives


The Category-2 PPE threshold (roughly 4-8 cal/cm²) is the operational boundary for normal-work exposure without specialised arc-flash suits. At 800 VDC with SSCB protection and typical fault current, Category-2 is achievable. At 800 VDC with mechanical breakers only, the same architecture crosses into Category 3 or 4.



Illustrative surface based on IEEE 1584 DC extension. Actual incident energy requires facility-specific fault-current study.





08Operator checklist: eight items to verify before commissioning



The practical output of this workstream is a checklist an operator can walk into vendor and insurance conversations with. Eight items cover the substance.



 1. Fault-current study on record. Facility-specific study demonstrating the arcing current, arc-gap geometry, and incident energy at every operator work location. Not a template; a study done on the actual as-built.
 2. Protection topology specified at bus and rack level. SSCB, current-limiting fuse, mechanical breaker, or hybrid. With coordination settings that ensure the fastest device operates first.
 3. Clearing-time budget documented per protection tier. Total clearing time from fault initiation to current interruption, per typical fault scenario. Should demonstrate PPE Category 2 or better for normal-operations work.
 4. Continuous arc-fault detection deployed. Not just fault-current interruption; monitoring that detects developing faults before they initiate an event.
 5. Insulation-resistance monitoring on ungrounded portions. Bender-style monitors where the topology is IT-System; trending on grounded systems as well.
 6. Maintenance procedure documented for de-energised work. Lock-out / tag-out procedure that accounts for capacitor bus energy, not just the source-side disconnect.
 7. PPE inventory and training current. Category-2 arc-flash suits at minimum; commissioning training completed for all electricians who will work on the facility.
 8. Insurance underwriting file complete. All of the above documented in a form insurance underwriters can review; specific carrier conversation begun 6-9 months before commissioning, not the week before.




09The reframe for anyone specifying a facility now



The intuition that 800 VDC is categorically more dangerous than AC is wrong at the architectural level. It is more precisely true that 800 VDC is a design responsibility that AC is not. An AC facility inherits protection behaviour from the physics of the zero crossing. An 800 VDC facility inherits nothing; its protection behaviour is entirely a function of the specific topology and vendor selection.



The operational takeaway for anyone specifying a facility now is that the arc-fault workstream cannot be deferred to Stage 4 electrical engineering. It has to be scoped at architecture selection (essay 2) because the choice between ±400 V and 800 V single-ended interacts with the protection topology in ways that change insurance premium tiers by 40+ percentage points. Operators that treat arc-fault protection as a checkbox at the end of the electrical design will pay for that treatment on the insurance premium line for the life of the facility.



Part V of this series moves into grounding and DC ground-fault protection. The second half of the DC protection story, and one that interacts with the arc-fault design in ways that are not obvious until both are specified together.





Glossary of terms used



ACAlternating Current. Electrical current that periodically reverses direction, in most public grids 50 or 60 times per second.
AFCIArc-Fault Circuit Interrupter. Protective device designed to detect arc faults and interrupt the circuit.
DCDirect Current. Electrical current flowing continuously in one direction, without the periodic reversal of AC.
IECInternational Electrotechnical Commission. Global standards body for electrical and electronic technologies, headquartered in Geneva.
IEEEInstitute of Electrical and Electronics Engineers. Global professional association publishing power and communications standards including P2818 for high-voltage DC data centres.
NECNational Electrical Code. US electrical installation code published by NFPA, updated on a three-year cycle.
NFPANational Fire Protection Association. US organisation publishing fire and electrical safety codes including the NEC and NFPA 70E.
OCPOpen Compute Project. Hyperscaler-led standards body developing open reference designs for data centre hardware.
PPEPersonal Protective Equipment. Arc-flash suits, gloves, and face shields rated to specific incident-energy levels per NFPA 70E.
SSCBSolid State Circuit Breaker. Semiconductor-based protective device that interrupts DC fault current in microseconds rather than the milliseconds required by mechanical breakers. Governed by IEC 60947-10 published early 2026.
ULUnderwriters Laboratories. US safety certification body that tests and certifies electrical equipment against applicable codes.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.




Related topic hubs
 * 800V Data Centre Power
 * AI Power Semiconductors
 * Browse all topic hubs →


Method and sources. Public information only. Schneider Electric August 2026 arc-flash study cited from Schneider Electric newsroom and Data Center Knowledge coverage. IEC 60947-10 publication cited from IEC publication register. Vendor product coverage cited from public datasheets, press releases, and OCP working-group participation lists. Insurance premium ranges are the author's synthesis from operator conversations mid-2026 and are directional; actual premiums require carrier-specific quote. No advisory relationship with any named party.


Primary sources.
Schneider Electric 800 VDC arc-flash study (Aug 2026) ·
Schneider WP219: DC Arc Flash Analysis ·
ARC Advisory Group commentary ·
Data Center Richness follow-up ·
Siemens DC protection technical paper



Series footer. Part IV in The DC-DC TransitionRelated reading: Part II on the two 800 VDC architectures, Part III on the six-layer architecture map, Part V on grounding and ground-fault protection (next), Part X on standards evolution. Companion context: Due Diligence for the AI Buildout Part XI on regulatory / environmental / cyber DD, The AI Power Chain Part IV on the Interconnect Stack.


Written in a personal capacity. No advisory conflict on any named party. Nothing here is investment advice.













The DC-DC Transition: start with the foundation

 1. The real reason data centres are going DC (four forces)
 2. Two architectures wearing the same name (800V DC vs OCP Mt Diablo)
 3. The six-layer architecture map


==============================================================================
# Grounding, bonding, and DC ground-fault protection at 800 VDC (DC-DC Transition V)
URL: https://adikumar.co/dc-dc-transition-05-grounding-ground-fault/
Published: 2026-08-20
Summary: TN-S vs IT-System at 800 VDC, insulation-resistance monitoring at scale, ground-fault detection technologies, coordination with arc-fault protection.
==============================================================================

The DC-DC Transition series · Part 5 of 19









The DC-DC Transition · Part V of XIII



Grounding, bonding, and ground-fault protection at 800 VDC


Grounding topology at 800 VDC is not a Stage-4 electrical-engineering afterthought. The TN-S / TT / IT-System choice interacts with the arc-fault protection design (Part IV) and with the insurance framework in ways that make it a Stage-1 architecture decision. Miss the decision at architecture selection and the facility either cannot achieve continuous availability or cannot pass insulation-resistance monitoring at commissioning.


Part V in the DC-DC Transition series · independent analysis · no advisory conflict on any named party





The DC-DC Transition · thirteen essays for data centre power architects

 1.  I. The real reason data centres are going DC
 2.  II. Two architectures wearing the same name
 3.  III. The architecture map
 4.  IV. Arc behaviour + insurance
 5.  V. Grounding + ground-fault protection (you are here)
 6.  VI. Battery integration at 800 VDC
 7.  VII. Retrofit vs greenfield: the decision framework
 8.  VIII. 800 VDC and liquid cooling co-emergence
 9.  IX. Power quality + grid interaction
 10. X. Standards: OCP, IEC, NEC, IEEE
 11. XI. Vendor economics: who wins the transition
 12. XII. Commissioning, skills, operational readiness
 13. XIII. The ten-year view





01Why grounding at 800 VDC is a different problem



Grounding in an AC system is a well-mapped problem with a small set of topology choices and a large body of installed-base experience. Grounding at 800 VDC borrows the topology vocabulary from AC (TN-S, TN-C, TT, IT-System) but inherits none of the operational reflexes. Three specific properties change how the topology behaves.



First, ground-fault current at DC does not have a natural zero crossing (see Part IV). A first-fault event that would trip an AC ground-fault interrupter within one half-cycle can persist for hundreds of milliseconds on DC before conventional protection responds. The failure mode is thermal, not electrical. The fault current heats the conductor path or the insulation until something else fails.



Second, insulation degradation at DC accumulates differently. Continuous DC voltage stress produces electrochemical corrosion at insulation defects at rates that AC systems do not experience because AC voltage stress reverses direction. The five-year insulation-resistance profile of an 800 VDC system is not the same shape as its AC-equivalent.



Third, ground fault at high DC voltage produces sustained arc column risk in ways ground fault at low DC (48 V telecom systems) does not. This means the ground-fault protection topology choice interacts directly with the arc-fault protection topology from Part IV. The two workstreams cannot be specified independently.




02The four topology options and where each historically lives



The four options for 800 VDC grounding are the same nomenclature as AC but the trade-offs are different.





TopologyWhere the source connects to earthHistorical use in DCTrade-off at 800 VDC


TN-SNeutral solidly earthed at source; separate PE conductor to loadRare in DC (AC convention)Ground-fault current is high and predictable; conventional protection can operate; large fault-clearing energy required
TN-CCombined neutral/PE conductor from source to loadNot applicable in DC (bipolar architectures)Not recommended for 800 VDC data centre applications
TTSource and load earthed at separate electrodesOccasionally in DC industrialSimpler wiring; ground-fault current lower; residual current monitoring required
IT-System (Isolated)No intentional earth connection; monitored insulation resistanceStandard for submarine, aerospace, industrial critical DCFirst fault does not trip; continuous availability preserved; insulation-resistance monitoring is complex and expensive at MW scale





The interesting question for 800 VDC data centres is which of these scales to hyperscale. IT-System is the historical DC default for critical applications precisely because a first ground fault does not immediately interrupt the load. On a submarine or in a chemical plant, the alternative to riding through a first fault is losing the load, which is unacceptable. On a hyperscale AI training cluster, the alternative is also losing the load, which is also unacceptable. So IT-System is architecturally attractive. But insulation-resistance monitoring at hundreds of MW of DC bus becomes a substantial engineering problem.




03Insulation-resistance monitoring: why IT-System is hard at scale



An IT-System depends on continuously monitoring the insulation resistance between the ungrounded DC bus and earth. When resistance drops below a threshold (indicating a developing fault), the monitoring device alarms and the operator schedules corrective action. The system continues to operate on the healthy pole while the fault is investigated.



The physics of this monitoring is straightforward at small scale. A Bender-style device injects a low-frequency test current and measures the resulting voltage; the ratio gives the insulation resistance. At 400 V and a few tens of kW, the device works well and is standard on hospital operating rooms, submarine electrical systems, and industrial rectifier plants.



At 800 V and hundreds of MW of connected load, the same technique runs into two engineering problems. First, the parallel capacitance of a large DC bus (rack-level bulk capacitors, DC bus capacitance, EMI filter capacitance) is high enough that the injected test signal is masked by the bus capacitance response. Second, the density of connected devices means the location resolution of the insulation-resistance measurement drops. The monitor tells you resistance has dropped somewhere, but not where among 20,000 racks.




Chart 1. Insulation-resistance monitoring: signal quality vs bus capacitance


At low bus capacitance (small facility, tens of kW), Bender-style monitors give clean insulation-resistance readings. As bus capacitance rises with facility scale, the injected test signal is masked by capacitive response, and effective resolution drops. Above ~100 MW of connected load, the technique needs augmentation with distributed local monitoring at rack or row level.



Illustrative curve based on Bender IRDW technical documentation and industrial deployment experience.




The practical response to this at hyperscale is distributed insulation-resistance monitoring: a hierarchical monitoring architecture where the main DC bus has one monitor, each row has a monitor, each rack cluster has a monitor, and fault localisation is done through a network of coordinated measurements. This is engineering-solvable but adds cost and complexity that the TN-S alternative does not have.




04TN-S at 800 VDC: the emerging default and its failure modes



The emerging default for hyperscale 800 VDC is TN-S with fast ground-fault protection at the pole level. The reasoning is that IT-System does not scale cleanly, and TN-S allows conventional protection topologies (SSCBs, current-limiting fuses) to respond to ground faults on the same time scale as they respond to short-circuit faults.



Three failure modes matter most. The first is high ground-fault current in the TN-S topology. A solidly earthed neutral means ground faults produce fault current comparable to short-circuit current, which stresses the protective device more than the IT-System equivalent. The protection specification has to be sized accordingly.



The second is common-mode current on the protective earth conductor. In a switching power-electronic environment (which is what a data centre is), common-mode currents on the PE conductor are non-zero even in the absence of fault. If ground-fault protection is set too sensitive, it triggers on nuisance common-mode current; if set too coarse, it misses developing faults. The tuning is not trivial.



The third is transient overvoltage from lightning or switching. TN-S provides a clear path for surge current to ground, which is desirable. But the transient voltage that appears across insulation during a strike can be several kilovolts even at nominal 800 V, and the BSL rating of connected equipment has to be specified against the actual expected transient envelope, not just the nominal bus voltage.




05Ground-fault detection technologies in 2026



The vendor landscape for DC ground-fault detection has three categories. Passive protection uses conventional overcurrent devices with settings calibrated for ground-fault current. Active differential current sensing uses Hall-effect or fluxgate sensors on both polarity conductors and detects the imbalance. Continuous insulation-resistance monitoring uses Bender-style techniques as described above.




Chart 2. Ground-fault detection technologies: response time vs sensitivity


Passive overcurrent protection is fast but insensitive. Misses low-current developing faults. Active differential sensing is sensitive but has moderate response time. Insulation-resistance monitoring detects faults before they draw fault current, but the detection window is minutes to hours rather than milliseconds. Best-in-class designs use all three in a layered scheme.



Vendor datasheet compilations from Bender, Sensata, MessTek, Schneider Electric, Eaton 2024-2026.





Chart 3. Named vendor products by category, mid-2026


Bender dominates insulation-resistance monitoring. Sensata and MessTek are the specialists on active differential sensing at DC. Schneider Electric and Eaton have integrated ground-fault detection into their SSCB product lines. ABB provides both traction-derived DC breakers and a new generation of DC ground-fault protection modules.



Vendor category positioning from public product literature 2024-2026.





06Transient overvoltage handling: the surge-protection layer



Transient overvoltage protection at 800 VDC is a design layer often deferred to the end of the electrical design, where it should be scoped alongside grounding. The protection devices. SPDs using MOVs or gas-discharge tubes. Have to be specified against the specific expected transient envelope for the facility location and topology.



The interaction with grounding topology matters. In a TN-S system, SPDs typically connect from line to earth and clamp the transient overvoltage against the PE conductor. In an IT-System, SPDs connect between poles rather than to earth, changing the coordination requirements. Getting the coordination between SPDs and downstream ground-fault protection wrong is a documented cause of nuisance tripping and premature SPD failure in real deployments.




Chart 4. SPD coordination by grounding topology


TN-S: SPDs connect line-to-earth, coordinated with the main bonding conductor. IT-System: SPDs connect line-to-line, with additional line-to-earth via a spark gap that operates only above IT-System withstand. TT: hybrid approach depending on distribution characteristics.



Coordination schematics from IEC 61643 series (SPD standards) and Bender/MessTek application notes.





07Interaction with arc-fault protection (Part IV)



The grounding topology and the arc-fault protection topology (Part IV) are not independently specifiable. Three specific interactions matter.



The first is that arc-fault current profile depends on grounding topology. In TN-S with solidly earthed neutral, an arc-fault-to-earth draws high fault current that fast SSCBs can interrupt in microseconds. In IT-System, the same fault does not immediately draw fault current from the source (because there is no return path through earth), so arc-fault detection has to trigger on other signatures. Voltage disturbance, radiated emission, or acoustic signature.



The second is that ground-fault protection has to coordinate with arc-fault protection to avoid nuisance tripping. If both operate on the same fault, one has to operate first with the other as backup, and the timing coordination is tight enough at DC clearing times that it has to be simulated at design.



The third is that the operator's insurance underwriting file (Part IV) needs both topologies documented together. Insurers are not underwriting arc-fault protection or ground-fault protection separately; they are underwriting the combined electrical safety architecture. Operators presenting the two workstreams as separate design outputs get pushed back to consolidate.




08Vendor procurement checklist



The operator-level checklist for the grounding workstream is shorter than the arc-fault checklist but has to be executed at Stage 1.



 1. Grounding topology decision documented at architecture selection. Not deferred to Stage 4 electrical engineering. The TN-S vs IT-System choice interacts with arc-fault protection (Part IV) and with insurance premium tiering.
 2. Insulation-resistance monitoring plan (if IT-System) or ground-fault protection tuning (if TN-S) sized against actual facility scale. Not a template; a study.
 3. Distributed monitoring architecture if the facility exceeds 100 MW connected load on any DC bus. Single-point Bender monitors will not resolve faults at hyperscale.
 4. SPD coordination documented against expected transient envelope. Includes lightning, switching, and equipment-derived transients specific to the site.
 5. Ground-fault + arc-fault coordination study on record. The two workstreams do not compose additively; they have to be coordinated at design.
 6. Insurance underwriting file consolidated across both electrical safety workstreams. Carriers underwrite the combined architecture, not the two separately.




09The reframe for anyone specifying now



Grounding at 800 VDC is a Stage-1 architecture decision because it interacts with three other Stage-1 decisions: the two-architecture choice (Part II, ±400 V bipolar vs 800 V single-ended), the arc-fault protection topology (Part IV), and the insurance underwriting framework. Any project treating grounding as a Stage-4 electrical-engineering output ends up with an architecture that cannot be adequately protected or insured on the timeline the rest of the project assumes.



Part VI moves into battery integration. The layer where the DC transition changes UPS architecture, BMS complexity, and battery chemistry choice simultaneously, and where the interactions with grounding and arc-fault protection (Parts IV and V) determine what actually gets specified.





Glossary of terms used



BSLBasic Insulation Level. The withstand voltage rating of electrical equipment for transient events such as lightning strikes.
IECInternational Electrotechnical Commission. Global standards body for electrical and electronic technologies.
IT-SystemIsolated grounding topology where no supply conductor is intentionally connected to earth; used historically in critical DC systems including submarine, aerospace, and industrial applications.
MOVMetal-Oxide Varistor. Nonlinear resistor used in surge protection devices to clamp transient overvoltage.
PEProtective Earth. The dedicated conductor bonding equipment enclosures to ground for safety.
PCCPoint of Common Coupling. The electrical connection point between a customer facility and the utility grid.
RCMResidual Current Monitor. Device detecting current imbalance between supply conductors indicative of a ground fault.
SPDSurge Protective Device. Component protecting equipment against transient overvoltage from lightning, switching, or fault events.
SSCBSolid State Circuit Breaker. Semiconductor-based protective device that interrupts DC fault current in microseconds. Governed by IEC 60947-10 published early 2026.
TN-STerra Neutral-Separate. Grounding topology with separate neutral and protective earth conductors from source to load.
TN-CTerra Neutral-Combined. Grounding topology with combined neutral and protective earth conductor.
TTTerra Terra. Grounding topology with source and load earthing at separate ground electrodes.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.




Related topic hubs
 * 800V Data Centre Power
 * AI Power Semiconductors
 * Browse all topic hubs →


Method and sources. Public information only. Vendor product positioning from published datasheets and application notes (Bender IRDW documentation, Sensata Hall-effect and fluxgate sensor datasheets, MessTek DC arc-fault detection, Schneider Electric and Eaton SSCB product literature, ABB DC breaker portfolio). Standards references from IEC 60947-10 (published early 2026), IEC 61643 (SPD series), NFPA 70E revision cycle. Distributed monitoring architecture derived from hyperscaler-published OCP working-group discussions on 800 VDC infrastructure. No advisory relationship with any named party.


Series footer. Part V in The DC-DC TransitionRelated reading: Part IV on arc behaviour and insurance (interacts directly with grounding topology), Part II on the two 800 VDC architectures, Part X on standards evolution, Part XII on commissioning. Companion context: Due Diligence for the AI Buildout Part V on product technology DD, Part XI on regulatory / environmental scope.


Written in a personal capacity. No advisory conflict on any named party. Nothing here is investment advice.













The DC-DC Transition: start with the foundation

 1. The real reason data centres are going DC (four forces)
 2. Two architectures wearing the same name (800V DC vs OCP Mt Diablo)
 3. The six-layer architecture map


==============================================================================
# Battery integration at 800 VDC: UPS restructured, not replaced (DC-DC Transition VI)
URL: https://adikumar.co/dc-dc-transition-06-battery-integration/
Published: 2026-08-20
Summary: Native 800 VDC UPS architecture, BMS complexity at 220-cell series stacks, LFP vs NMC vs LTO chemistry choice, HSC role for sub-second transients…
==============================================================================

The DC-DC Transition series · Part 6 of 19









The DC-DC Transition · Part VI of XIII



Battery integration at 800 VDC: UPS restructured, not replaced


The transition to 800 VDC does not eliminate uninterruptible power supply from the data centre; it restructures how it is built. Native 800 VDC UPS architectures collapse three or four conversion stages into one or two, connect the battery directly to the DC bus, and change which battery chemistry wins. The BMS complexity increases sharply, but end-to-end efficiency climbs and steady-state footprint drops.


Part VI in the DC-DC Transition series · independent analysis · no advisory conflict on any named party





The DC-DC Transition · thirteen essays for data centre power architects

 1.  I. The real reason data centres are going DC
 2.  II. Two architectures wearing the same name
 3.  III. The architecture map
 4.  IV. Arc behaviour + insurance
 5.  V. Grounding + ground-fault protection
 6.  VI. Battery integration at 800 VDC (you are here)
 7.  VII. Retrofit vs greenfield: the decision framework
 8.  VIII. 800 VDC and liquid cooling co-emergence
 9.  IX. Power quality + grid interaction
 10. X. Standards: OCP, IEC, NEC, IEEE
 11. XI. Vendor economics: who wins the transition
 12. XII. Commissioning, skills, operational readiness
 13. XIII. The ten-year view





01The AC UPS architecture and its inefficiency



The traditional double-conversion online UPS is a workhorse of the AC data centre. Utility AC comes in, a rectifier converts it to DC, the DC bus feeds an inverter that produces conditioned AC out, and a battery attaches to the DC bus for backup. The topology is well-understood, well-standardised, and well-maintained by a mature vendor set (Vertiv, Eaton, Schneider Electric, ABB, Riello, Piller, Delta).



The inefficiency is structural. In the AC-in / AC-out chain, the load ultimately consumes DC (every server has an internal PSU converting AC to the 12 V or 48 V the components use). So the round trip runs AC → DC → AC → DC, with conversion losses at every stage. Total end-to-end efficiency for a modern AC-UPS-plus-server-PSU chain lands around 90-92 percent. Each conversion stage costs 2-4 percentage points of efficiency and adds capex, weight, and thermal load.




02The native 800 VDC UPS: what changes



A native 800 VDC UPS collapses the topology substantially. Utility AC comes in through an AC-DC rectifier that produces 800 V DC directly. The DC bus feeds the racks at 800 V (via DC distribution. See Part III on the six-layer architecture map). The battery attaches directly to the 800 V DC bus without any intermediate conversion. The rack-level PSU steps down from 800 V to 48 V or lower for the actual compute silicon.




Chart 1. UPS topology comparison: AC double-conversion vs native 800 VDC


Traditional AC UPS: 4 conversion stages, ~90-92 percent end-to-end efficiency. Native 800 VDC UPS: 2 conversion stages, ~96-97 percent end-to-end efficiency. The gain compounds at scale. A 200 MW facility saves 8-14 MW of load loss, which is 800-1400 racks worth of compute at 10 kW per rack.



Efficiency data from Delta Electronics, Vertiv Liebert, and Schneider Electric public technical papers 2024-2026 on native DC UPS architectures.




The efficiency gain is the headline. The topology-simplification gain is often larger in dollar terms. Fewer components mean less capex, less footprint, less thermal load, fewer failure points. But the topology simplification is not free; it is paid for on the BMS side and the battery-chemistry side.




03BMS complexity increases sharply



A lithium-ion battery cell has a nominal voltage of roughly 3.6 V (NMC) or 3.2 V (LFP). To reach an 800 V nominal DC bus, you need a series stack of roughly 220-250 cells. Each cell has to be monitored for voltage, temperature, and state-of-charge. Cell-to-cell balancing has to be maintained during charge and discharge. A single cell drifting out of tolerance can propagate through the string.



The BMS complexity increases roughly as the cell count. A 48 V stack has 12-15 cells; the BMS runs on a small microcontroller. An 800 V stack has 220-250 cells; the BMS is a distributed sensor network with communication overhead and safety-certified isolation between cell-monitoring domains.




Chart 2. BMS component cost per kWh at different voltage classes


BMS cost as a share of installed battery capex rises with voltage class. At 48 V, BMS is roughly 8-12 percent of installed capex. At 800 V, BMS is 15-22 percent. The offsetting factor is that the overall stack is smaller in capacity terms. Fewer parallel strings deliver the same power. But the per-kWh economics of BMS deteriorate with voltage.



Composite from BYD Battery-Box, LG Energy Solution RESU, and CATL EnerC container BMS spec sheets 2024-2026.





04Battery chemistry choice at 800 VDC



The chemistry winner at 48 V (NMC dominant for its higher energy density, LFP for cost/safety-sensitive applications) does not automatically win at 800 V. Three considerations shift the balance.



The first is safety. Higher-voltage series stacks accumulate more stored energy per module. LFP's flatter voltage curve and lower thermal-runaway risk become more attractive as the failure consequence grows. The 800 V UPS market has been trending toward LFP faster than the EV market has.



The second is cycle life. Data centre UPS applications draw shallow discharge cycles frequently (every utility disturbance) plus rare deep discharge cycles. LFP's superior cycle life at partial state-of-charge cycling favours it for the UPS duty profile. NMC handles fewer cycles at the same depth-of-discharge.



The third is second-life battery integration. EV batteries retired from vehicle use (typically at 70-80 percent of original capacity) are becoming an increasingly viable supply channel for stationary storage. The chemistry mix in retired EV packs skews heavily toward NMC in Western markets and LFP in Chinese markets. At 800 V DC bus, second-life battery integration requires additional cell-monitoring headroom, which changes the BMS spec.




Chart 3. Battery chemistry TCO at 800 VDC data centre UPS over 10-year hold


LFP has ~30 percent higher upfront capex per kWh than NMC but ~50 percent lower TCO over ten years due to longer cycle life and lower fire-safety insurance premium. LTO commands 2-3x upfront capex and is only competitive where cycle count is extreme (grid-services applications). Second-life NMC undercuts fresh-LFP on year-1 capex but carries 20-40 percent capacity uncertainty.



TCO model assumptions: 20-year facility life, 200 MW facility, 15-minute backup duration at rated load. Chemistry-specific cycle life from CATL, LG ES, BYD, Toshiba SCiB public technical documentation.





05Hybrid supercapacitor role: bridging the 0.1-10 second gap



Batteries do not handle sub-second and few-second discharges well. Extreme peak-power events (GPU cluster synchronisation transients at the millisecond scale, brief utility ride-through requirements at the sub-second scale) accelerate battery degradation disproportionately relative to the energy delivered. HSCsHybrid supercapacitors, sometimes called lithium-ion capacitors. Fill this gap.



The AI Power Chain Capacitor Stack essays cover HSC technology in depth. The specific role at 800 VDC UPS is that HSCs sit on the DC bus in parallel with the battery, absorbing the sub-second transients that would otherwise be shed onto the battery. The result is longer battery cycle life at a moderate capex adder, and the ability to specify smaller-and-cheaper battery capacity because the peak-power headroom moves to the HSC.




Chart 4. HSC vs battery discharge duration coverage


HSCs cover 0.1-10 second discharges at high power density; batteries cover minutes-to-hours at high energy density. The overlap window (roughly 1-30 seconds) is where the two technologies compete on TCO. Hybrid deployments with both technologies dominate on TCO for typical UPS duty profiles.



Discharge duration ranges from HSC vendors (Maxwell/Skeleton, JSR Micro) and battery vendors (CATL, LG ES). Typical UPS duty profile from Vertiv Liebert and Eaton 9395 technical documentation.





06Storage capex composition and where each technology lands by 2028



The capex composition of an 800 VDC battery installation is not just cell cost. Cells are roughly 40-50 percent of installed capex; BMS 15-22 percent; enclosure and mounting 8-12 percent; thermal management 8-14 percent; power electronics interface 10-15 percent; commissioning and controls 5-10 percent. The percentages shift with chemistry (LFP tolerates less aggressive thermal management, so thermal capex drops) and with voltage class (BMS goes up with voltage as noted above).




Chart 5. Storage capex composition per kWh installed at 800 VDC


Cells dominate but do not overwhelm. BMS + power electronics interface together approach parity with cell cost at 800 VDC. Thermal and commissioning are the "invisible" line items that separate best-in-class deployments from average.



Composite from BloombergNEF energy storage cost surveys 2024-2026, adjusted for 800 VDC UPS-application specifics using CATL EnerC, LG ES RESU/ESS, BYD Battery-Box publications.





Chart 6. Battery chemistry market share in data centre UPS applications 2022-2028E


LFP share has been climbing steadily in data centre UPS from ~15 percent in 2022 to a projected 55-65 percent by 2028. NMC declines in share (though total volume grows). LTO holds a niche in grid-services / high-cycle-count applications. Second-life battery integration reaches notable share by 2027.



Share estimates from Wood Mackenzie storage market tracker, BloombergNEF energy storage outlook 2024-2026, and vendor commentary.





07Named vendor landscape at 800 VDC UPS





VendorProduct linePositioningChemistry focus


VertivLiebert EXL S1 + eBoost + native DC UPS in developmentIncumbent leader, transitioning full line to 800 VDC nativeLFP-forward with NMC option
Eaton9395P series + DC-native product line launched 2025Incumbent leader, aggressive DC-native positioningLFP-forward
Schneider ElectricGalaxy VXL + BATTERY MODULAR CX at 800 VDCModular DC UPS at hyperscaleLFP-forward with second-life offering
Delta ElectronicsUltron DPS + Modulon DPHEfficiency leader; strong at DC-native from Taiwan/China market historyLFP + NMC dual
ABBMegaFlex + DPA + industrial-derived DC systemsIndustrial-heritage strengthLFP + LTO for extreme cycles
Riello / PillerNiche premium UPSSlower to DC-native transitionNMC legacy + LFP additions
CATL / BYD / LG ESContainer-scale storage integrated into UPSBattery-vendor-led competition to traditional UPS vendorsLFP dominant
Skeleton / MaxwellHSC modules for peak-power augmentationComplementary rather than competingN/A (supercap-based)






08Interaction with arc-fault and grounding (Parts IV, V)



Battery integration at 800 VDC interacts with the two prior essays in specific ways. The battery is a fault-current source in its own right. A shorted 800 V battery stack can deliver kilo-amp fault current in microseconds. The arc-fault protection topology (Part IV) has to account for this contribution; a battery-side fault has a different current profile from a source-side fault.



The grounding topology (Part V) also interacts. IT-System grounding places the battery on an ungrounded bus, which means the first fault does not draw fault current from the source but can draw fault current from the battery. Insulation-resistance monitoring has to distinguish between source-side and battery-side developing faults.



The practical implication for the operator is that the battery specification, the arc-fault protection spec, and the grounding spec have to be produced by the same engineering workstream at Stage 1, not by three separate workstreams that meet at commissioning. Operators that try to procure batteries against a boilerplate spec without the arc-fault and grounding context typically discover the mismatch at commissioning, at which point the fix is expensive.




09Operator procurement checklist



 1. Chemistry decision documented against cycle-life duty profile. LFP is the default for typical UPS duty; NMC only where density is critical and cycle-life tolerance is well-understood; LTO for extreme-cycle grid-services applications.
 2. BMS spec sized for the actual cell count at 800 V. Distributed cell monitoring, safety-certified isolation, redundant temperature sensing per module.
 3. HSC augmentation modelled against actual peak-power profile. If the GPU cluster produces sub-second transients above the battery's rated peak, HSC augmentation pays back on battery cycle life.
 4. Second-life battery option evaluated where capital constraint favours it. Requires additional BMS headroom and separate insurance underwriting.
 5. Fault-current contribution documented alongside arc-fault protection design (Part IV). Battery-side fault current is not the same as source-side.
 6. Grounding topology decision integrated with battery topology. IT-System behaves differently from TN-S on battery-side fault detection (Part V).
 7. Thermal management sized for chemistry-specific requirements. LFP is more thermally tolerant than NMC; commissioning skip on this line item causes accelerated degradation.
 8. End-of-life recycling plan committed at procurement. Regulatory posture on battery recycling is tightening; take-back agreements at procurement reduce end-of-life cost.




10The reframe for anyone specifying now



The 800 VDC UPS operates as a different architecture from the AC UPS, with different capex composition, different chemistry economics, different BMS complexity, and different interactions with the rest of the electrical protection design, rather than as a smaller version of the AC UPS in different wiring. Operators specifying a 200+ MW facility on the assumption that battery integration is a Stage-4 procurement decision consistently discover in commissioning that the battery topology, the arc-fault protection, and the grounding scheme were designed against inconsistent assumptions. The fix ranges from expensive to project-delaying.



Part VII moves into retrofit vs greenfield. The strategic decision that determines whether an operator gets the benefits of the 800 VDC transition on their existing facility footprint or has to build new capacity to capture them.





Glossary of terms used



BMSBattery Management System. The electronics and software controlling battery pack cell balancing, thermal management, and state-of-charge estimation.
HSCHybrid Supercapacitor. Energy storage technology bridging batteries and traditional capacitors, well-suited to sub-second peak-power discharge.
LFPLithium Iron Phosphate. Battery chemistry favoured for stationary storage due to safety and cycle life, at some density penalty vs NMC.
LTOLithium Titanate Oxide. Battery chemistry with very high cycle life and fast-charge capability, at density penalty.
NMCNickel Manganese Cobalt. Battery chemistry favoured for high energy density in EVs and premium stationary storage.
PSUPower Supply Unit. Rack-level device that converts input power to the DC voltage the server consumes.
UPSUninterruptible Power Supply. Battery-backed power system that keeps critical loads running during grid disruptions.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.




Related topic hubs
 * 800V Data Centre Power
 * AI Power Semiconductors
 * Browse all topic hubs →


Method and sources. Public information only. Vendor product coverage from public datasheets and technical papers (Delta Electronics Ultron and Modulon lines, Vertiv Liebert EXL S1 and eBoost, Eaton 9395P and DC-native launches, Schneider Galaxy VXL and BATTERY MODULAR CX, ABB MegaFlex and DPA, Riello, Piller). Battery chemistry technical parameters from CATL, LG ES, BYD, Toshiba SCiB, and Skeleton/JSR Micro published specifications. TCO modelling from BloombergNEF energy storage cost surveys 2024-2026 and Wood Mackenzie storage market tracker. HSC context from the AI Power Chain Capacitor Stack essays. No advisory relationship with any named party.


Series footer. Part VI in The DC-DC TransitionRelated reading: Part IV on arc behaviour (battery fault-current contribution), Part V on grounding (battery-side vs source-side fault detection), Part VIII on cooling co-emergence (battery thermal management), Part X on standards. Companion context: The AI Power Chain Part I on the Capacitor Stack and HSC technology, Due Diligence for the AI Buildout Part V on product technology DD.


Written in a personal capacity. No advisory conflict on any named party. Nothing here is investment advice.













The DC-DC Transition: start with the foundation

 1. The real reason data centres are going DC (four forces)
 2. Two architectures wearing the same name (800V DC vs OCP Mt Diablo)
 3. The six-layer architecture map


==============================================================================
# Retrofit versus greenfield: the DC-DC decision framework (DC-DC Transition VII)
URL: https://adikumar.co/dc-dc-transition-07-retrofit-greenfield/
Published: 2026-08-20
Summary: Retrofit vs greenfield TCO for 800 VDC data centres. Break-even by facility age, cooling readiness, real-estate constraint.
==============================================================================

The DC-DC Transition series · Part 7 of 19









The DC-DC Transition · Part VII of XIII



Retrofit versus greenfield: the DC-DC decision framework


Retrofitting an existing AC facility to 800 VDC is technically possible but rarely economic. Break-even TCO analysis under typical assumptions favours greenfield by 30-50 percent over a ten-year hold. Retrofit only makes sense in three narrow cases. Most operators considering retrofit should scope a satellite greenfield instead.


Part VII in the DC-DC Transition series · independent analysis · no advisory conflict on any named party





The DC-DC Transition · thirteen essays for data centre power architects

 1.  I. The real reason data centres are going DC
 2.  II. Two architectures wearing the same name
 3.  III. The architecture map
 4.  IV. Arc behaviour + insurance
 5.  V. Grounding + ground-fault protection
 6.  VI. Battery integration at 800 VDC
 7.  VII. Retrofit vs greenfield (you are here)
 8.  VIII. 800 VDC and liquid cooling co-emergence
 9.  IX. Power quality + grid interaction
 10. X. Standards: OCP, IEC, NEC, IEEE
 11. XI. Vendor economics: who wins the transition
 12. XII. Commissioning, skills, operational readiness
 13. XIII. The ten-year view





01Why retrofit is structurally hard



An existing AC-native data centre is a set of physical constraints an 800 VDC upgrade has to work around. Four constraints dominate: the existing MV-to-LV substation infrastructure sized for AC distribution, the cooling plant sized for the prior rack-density profile, the real-estate footprint fixed at build time, and the operational continuity requirement that constrains what can be de-energised for how long.



Each of these translates to a specific retrofit cost. Rebuilding the MV-to-LV distribution to feed 800 VDC racks requires either wholesale replacement of the LV switchgear (expensive) or interposition of an AC-to-800 VDC conversion stage between the existing LV and the new DC bus (defeats much of the efficiency argument). Increasing rack density to the level 800 VDC economics require means the existing cooling plant is undersized. Typically by 3-5x if the retrofit is targeting 100 kW+ per rack. The real-estate footprint constrains whether new equipment fits at all; some retrofits require external containerised additions. And the continuity requirement means the retrofit is typically phased over 12-24 months with the facility running throughout, which multiplies commissioning cost.




02Retrofit capex composition



Retrofit capex is dominated by the electrical distribution rebuild, not by the rack-level equipment upgrade. Understanding this composition is what separates a well-scoped retrofit business case from a poorly-scoped one.




Chart 1. Retrofit capex composition per MW at 800 VDC, illustrative


Electrical distribution rebuild dominates at 35-45 percent. Cooling plant upgrade is the second-largest bucket at 20-30 percent. Rack-level equipment (the visible part of the upgrade) is only 12-18 percent. Commissioning premium. The cost of executing the retrofit while the facility operates. Is 8-15 percent of total.



Composite from published retrofit case studies (Digital Realty, Iron Mountain, Global Switch retrofits 2024-2026) and vendor engineering estimates.




The counterintuitive line item is the commissioning premium. Executing electrical work on an operational facility requires isolation procedures, temporary power arrangements, work-window scheduling around production maintenance, and additional safety inspection. Vendor commissioning quotes for retrofit work typically run 30-50 percent above the equivalent greenfield quote for the same equipment scope.




03Greenfield capex composition



Greenfield capex has a categorically different composition. The electrical distribution can be sized natively for 800 VDC without accommodating existing infrastructure. The cooling plant can be sized for the target rack-density profile without oversizing to accommodate legacy air-cooling. The real-estate footprint can be optimised for the target facility layout. And the commissioning happens once, on an empty building, without operational-continuity constraints.




Chart 2. Greenfield capex composition per MW at 800 VDC, illustrative


Site preparation and shell construction dominate at 25-35 percent (land, foundations, building envelope). Electrical distribution runs 20-25 percent. Cheaper than retrofit because sized-natively. Cooling plant sized to target density runs 15-20 percent. Rack-level equipment is 15-20 percent. Commissioning is 5-8 percent of total, roughly a third of retrofit commissioning as a share.



Composite from PORTS-Pike-scale project economics, Modular Datacenter Stack essay estimates, and vendor engineering references.





04Ten-year TCO comparison



The headline retrofit-vs-greenfield economics show at ten-year total cost of ownership rather than at day-one capex. Retrofit day-one capex per MW is typically 60-80 percent of greenfield day-one capex, and this appears to be the reason some operators evaluate the two on capex alone and choose retrofit. That evaluation is incomplete.



Adding opex. Energy losses from the retrofit's inherent inefficiencies (extra conversion stages, undersized cooling), maintenance premium on hybrid AC/DC operation, insurance premium on a facility that carries both architectures during phased retrofit. Retrofit TCO per MW over ten years lands 30-50 percent above greenfield TCO. The break-even is not at day one; it is roughly year 3-4 of hold, after which greenfield pulls sharply ahead.




Chart 3. TCO retrofit vs greenfield over ten-year hold at 800 VDC target


Retrofit shows favourable day-one capex but the cumulative TCO curve crosses at year 3-4 and diverges thereafter. By year 10 greenfield TCO is 30-50 percent lower than retrofit TCO for equivalent 800 VDC capacity. The specific delta depends on the retrofit's efficiency penalty, its insurance premium, and its operational-continuity cost.



Illustrative TCO model. Base case assumptions: 200 MW facility, 10-year hold, retrofit efficiency penalty 6 pp, retrofit insurance premium 25 pp above equivalent greenfield.





05Break-even sensitivity: when retrofit makes sense



Three specific conditions make retrofit economically defensible: the facility is under 5 years old (so remaining useful life justifies the retrofit capital); the cooling plant is already liquid-cooling-ready (avoiding the largest single retrofit cost); the real-estate is constrained (greenfield alternative is not physically available). If all three hold, retrofit TCO can beat greenfield-plus-decommissioning. If only two hold, the case is marginal. If only one, retrofit is almost certainly the wrong choice.




Chart 4. Retrofit break-even sensitivity: facility age at retrofit decision


A facility retrofit at age 2 delivers positive NPV over the remaining useful life. At age 5, the calculation is marginal. At age 8, retrofit TCO exceeds greenfield-plus-decommissioning under most reasonable assumptions. The exception is where real-estate scarcity dominates. Permit-constrained urban markets where greenfield is unavailable.



NPV sensitivity model with 10-year hold, 10 percent discount rate, and typical retrofit efficiency penalty.





06Phased retrofit: the approach that sometimes works



A phased retrofit executes the upgrade one rack cluster (or row, or hall) at a time while the rest of the facility continues to operate. The advantage is operational continuity. No full shutdown, no revenue interruption. The disadvantage is a 25-40 percent timeline premium and 15-25 percent capex premium relative to a shutdown-and-retrofit approach, plus the complexity of running two architectures side-by-side during the transition period.



The specific cases where phased retrofit works well are hyperscaler campuses with multiple halls (retrofit one hall at a time while others handle the load), colocation facilities with staged tenant migrations (align retrofit with tenant renewals or moves), and facilities where the retrofit is targeting only a specific rack-cluster type (AI training racks, for example, while the rest of the facility remains AC).




07Named case studies



Public disclosure on data centre retrofits is thin because operators have limited incentive to publicise operational disruption. The following are the retrofit cases with useful public information as of mid-2026.





Operator / facilityRetrofit scopeReported outcomeLessons per public commentary


Digital Realty (Northern Virginia)Partial DC-native rack rows retrofit into existing AC halls~18-month project, on budgetCooling plant compatibility was the enabling factor; would not have been feasible without prior liquid-cooling investment
Iron Mountain (Denver)Retrofit of colocation racks to 400 VDC (bipolar) for a specific hyperscaler tenant~24-month project, cost overrun 15-20 percentPhased execution added 6+ months; tenant-specific spec reduced retrofit ROI to marginal
Global Switch (London)Discussed retrofit; ultimately chose greenfield satellite insteadRetrofit business case did not closeFacility age (12+ years) and cooling constraint drove the greenfield decision
Hyperscaler-A campus (undisclosed)Phased 800 VDC retrofit of a 3-hall campus, one hall per yearIn progress, first hall commissioned Q1 2026Success dependent on other halls absorbing load during commissioning windows






08Decision framework for operators evaluating both



A defensible decision framework has three gates. Gate 1: is the facility under 5 years old? If yes, proceed. If no, greenfield unless real-estate scarcity is binding. Gate 2: is the cooling plant already liquid-cooling-ready or easily converted? If yes, proceed. If no, greenfield unless retrofit budget can absorb 40+ percent cooling capex. Gate 3: is there real-estate for a greenfield satellite within acceptable distance of the retrofit target? If yes, greenfield-plus-decommission-old. If no, retrofit becomes the default.




Chart 5. Retrofit vs greenfield decision flowchart, three gates


Gate 1: facility age. Gate 2: cooling readiness. Gate 3: real-estate availability. Only projects clearing all three gates should proceed as retrofits. Everything else defaults to greenfield.



Framework synthesised from public retrofit case studies and DD-lens application to 12+ evaluations 2024-2026.





Chart 6. Per-MW capex scatter of documented retrofit and greenfield 800 VDC projects


Retrofit projects cluster around $12-18M per MW installed. Greenfield 800 VDC projects cluster around $8-12M per MW. The overlap band ($11-13M per MW) is where retrofit-well-executed meets greenfield-poorly-executed, and where the decision-framework detail matters.



Named projects from public disclosure plus author's DD engagements 2024-2026. Specific project data anonymised where not publicly disclosed.





09The reframe for anyone evaluating now



The default assumption that retrofit is cheaper than greenfield is wrong on TCO terms even when it appears right on capex terms. Operators evaluating an 800 VDC upgrade should build the ten-year TCO model, not the day-one capex model. The retrofit case survives that scrutiny only when facility age is below 5 years and cooling plant is compatible and real-estate scarcity is binding. In the majority of evaluations, the correct answer is greenfield satellite with staged decommissioning of the legacy facility.



Part VIII moves into cooling. The co-emerging technology that determines whether the retrofit or greenfield decision even matters. Above 100 kW per rack, 800 VDC and direct-to-chip liquid cooling are not independent architecture choices.





Glossary of terms used



800 VDC800 Volts Direct Current. The emerging standard voltage class for AI-scale data centre power distribution.
DDDue Diligence. The workstream discipline of testing a target's claims before committing to a transaction.
LVLow Voltage. Typically below 1 kV, the class historically used for data centre power distribution below the MV substation.
MVMedium Voltage. Typically 1 kV to 35 kV, the voltage class connecting data centre power distribution to the utility grid.
NPVNet Present Value. Discounted cash-flow metric used to compare projects with different timing and duration.
TCOTotal Cost of Ownership. Capex plus opex over a defined hold period, discounted to present value.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.




Related topic hubs
 * 800V Data Centre Power
 * AI Power Semiconductors
 * Browse all topic hubs →


Method and sources. Public information only. Retrofit case studies from published operator commentary (Digital Realty investor presentations, Iron Mountain data centre operations disclosures, Global Switch strategic communications 2024-2026). Capex composition estimates from vendor engineering quotes (Vertiv, Schneider Electric, Eaton, Delta Electronics public engineering references) and Modular Datacenter Stack essay work. TCO model assumptions are the author's; specific projects require project-specific modelling. No advisory relationship with any named party.


Series footer. Part VII in The DC-DC TransitionRelated reading: Part III on the six-layer architecture map (retrofit inventory), Part VIII on cooling co-emergence (cooling is the largest retrofit constraint), Part IV on arc-fault protection (retrofit interacts with insurance premium), Part XI on vendor economics (retrofit vs greenfield vendor mix differs). Companion context: The AI Power Chain Part VI on Modular Datacenter Stack economics, Due Diligence for the AI Buildout Part VIII on capex + growth stress-testing.


Written in a personal capacity. No advisory conflict on any named party. Nothing here is investment advice.













The DC-DC Transition: start with the foundation

 1. The real reason data centres are going DC (four forces)
 2. Two architectures wearing the same name (800V DC vs OCP Mt Diablo)
 3. The six-layer architecture map


==============================================================================
# 800 VDC and liquid cooling: co-emerging, not independent (DC-DC Transition VIII)
URL: https://adikumar.co/dc-dc-transition-08-cooling-co-emergence/
Published: 2026-08-20
Summary: Rack density trajectory forces liquid cooling above 100 kW per rack. Vendor consolidation (Vertiv/PurgeRite, Eaton/Boyd, Schneider/Motivair) reshaped the…
==============================================================================

The DC-DC Transition series · Part 8 of 19







The DC-DC Transition · Part VIII of XIII



800 VDC and liquid cooling: co-emerging, not independent


Above 100 kW per rack, the choice of 800 VDC delivery and the choice of direct-to-chip liquid cooling are not independent. Air cooling caps at roughly 40-60 kW per rack at reasonable PUEDirect-to-chip liquid cooling is required for hyperscale AI compute. The vendors best positioned for the combined transition are the ones with credible cooling capability alongside 800 VDC hardware. That is the reason Vertiv/PurgeRite, Eaton/Boyd Thermal, and Schneider/Motivair happened when they did.


Part VIII in the DC-DC Transition series · independent analysis · no advisory conflict on any named party




The DC-DC Transition · thirteen essays for data centre power architects

 1.  I. The real reason data centres are going DC
 2.  II. Two architectures wearing the same name
 3.  III. The architecture map
 4.  IV. Arc behaviour + insurance
 5.  V. Grounding + ground-fault protection
 6.  VI. Battery integration at 800 VDC
 7.  VII. Retrofit vs greenfield
 8.  VIII. 800 VDC and liquid cooling co-emergence (you are here)
 9.  IX. Power quality + grid interaction
 10. X. Standards: OCP, IEC, NEC, IEEE
 11. XI. Vendor economics: who wins the transition
 12. XII. Commissioning, skills, operational readiness
 13. XIII. The ten-year view




01Rack density trajectory and the cooling implication


Rack density has been rising through 2022-2028 on a trajectory that no prior generation of data centres anticipated. Typical enterprise racks in 2022 drew 5-15 kW. Modern hyperscaler AI training racks in 2025 draw 60-100 kW. Purpose-built AI training racks announced for 2027-2028 target 400 kW to 1 MW per rack. This trajectory forces both the DC delivery decision (see Part II on the two 800 VDC architectures) and the cooling architecture decision.



Chart 1. Rack density trajectory 2020-2028E: from 5 kW to 1 MW per rack


Enterprise racks in 2022 drew 5-15 kW. Modern AI training racks in 2025-2026 draw 60-100 kW. Announced 2027-2028 racks target 400 kW to 1 MW. The trajectory is driven by GPU power draw, HBM stacked-memory density, and interconnect optical/electrical requirements per rack.



Density estimates from OCP Mt. Diablo target specifications, NVIDIA GB200/GB300 platform disclosures, hyperscaler earnings-call commentary 2024-2026.




02The air-cooling ceiling


Air cooling. Where computer room air handlers move chilled air through the raised floor and across the front of racks. Has been the workhorse data centre cooling architecture for four decades. It works well up to a specific density ceiling and degrades sharply above it. The ceiling is roughly 40-60 kW per rack at PUE below 1.4, and the degradation above that ceiling is not gradual; PUE deteriorates rapidly as the air-handling infrastructure fights to move enough mass flow of air to remove the heat.


The physics is straightforward. Air has low specific heat compared to water or dielectric coolant. Removing 100 kW of heat with air requires roughly 20,000 CFM of air flow at reasonable delta-T. That volume of air requires substantial fan power, and fan power itself dissipates as heat. Above roughly 60 kW per rack, the fan-power-to-cooling-delivered ratio starts to exceed 25 percent, and total facility PUE degrades past 1.5. Above 100 kW per rack, air cooling becomes physically impractical.



Chart 2. PUE degradation with rack density: air cooling ceiling


PUE stays flat below 40 kW per rack. Between 40-60 kW PUE begins to rise slowly. Above 60 kW the rise accelerates. Above 100 kW air cooling is impractical at any reasonable PUE. The vertical axis PUE crosses 2.0 at roughly 120-130 kW per rack.



PUE degradation curve synthesised from ASHRAE TC 9.9 thermal guidelines, Vertiv Coolcentric technical publications, and hyperscaler-published PUE data.




03Direct-to-chip liquid cooling: the required architecture above 100 kW


Direct-to-chip liquid cooling (also called cold-plate liquid cooling) routes chilled coolant through cold plates mounted directly on the GPU, HBM, and CPU packages. The coolant absorbs heat at the source and carries it to a rack-level or row-level heat exchanger. The heat exchanger transfers the heat to a facility water loop that ultimately dumps it to ambient via cooling towers, chillers, or heat-reuse systems.


The economics work above roughly 100 kW per rack because the coolant carries orders-of-magnitude more heat per unit volume than air. The engineering complexity is real. Coolant chemistry, leak detection, cold-plate integration with the silicon package, pump sizing, redundancy for pump failure. But each of these is a solved problem individually. What has changed in 2024-2026 is the vendor set that ships integrated solutions.


The alternative is immersion cooling. Submerging the entire server in dielectric fluid. Immersion has better thermal performance and simpler electrical interfaces, but higher operational complexity (service events require draining and cleaning), higher fluid cost, and slower vendor maturation. Direct-to-chip has won the near-term deployment race in hyperscale AI even though immersion may win in the long run for the highest-density workloads.



Chart 3. Cooling architecture per MW capex: air vs direct-to-chip vs immersion


Air cooling has the lowest capex per MW at low density but the cost gap closes above 60 kW per rack and inverts above 80. Direct-to-chip has the middle capex band and dominates at 80-400 kW per rack. Immersion has the highest capex but the lowest facility PUE, competitive for 400+ kW deployments where facility PUE dominates opex.



Capex estimates from Vertiv Liebert XDU, CoolIT Systems reference designs, Submer immersion vendor data, and hyperscaler-published thermal architecture references.




04Coolant chemistry and DC electrical isolation


Coolant chemistry choice at 800 VDC interacts with the electrical protection design. The primary coolants in use in 2026 are water-glycol mixtures (common in direct-to-chip), engineered dielectric fluids (3M Novec derivatives and successors, in immersion), and specialty fluids from vendors like Solvay and Chemours designed for compatibility with high-voltage DC operations.


The 800 VDC-specific consideration is coolant electrical conductivity. A coolant that becomes conductive under contamination or ion accumulation can create a fault path across the DC bus. Water-glycol mixtures are electrically conductive by design and require dielectric barriers between the coolant loop and the DC electrical system. Dielectric fluids are non-conductive but expensive and require monitoring for degradation. Specialty low-conductivity fluids sit in between on cost and complexity.


Leak detection becomes a new failure mode. A pinhole leak in a direct-to-chip cold plate that would be an inconvenience in a 48 V system can create an arc-fault initiator in an 800 VDC system if the leaked coolant bridges energised conductors. Continuous leak detection sensors, drip pans with conductivity monitoring, and rapid isolation protocols are all Stage-1 design decisions, not Stage-4 monitoring add-ons.



05Vendor consolidation dynamics


The vendor set for combined 800 VDC + liquid cooling has consolidated sharply through 2024-2026 as hyperscaler customers have made clear they want both from one vendor. Three named transactions illustrate the pattern.


Vertiv acquired PurgeRite in 2023 for approximately $1B plus earnout. PurgeRite's mechanical flushing, purging, and filtration services for liquid cooling systems support hyperscale AI deployments. The Vertiv 800 VDC power portfolio (Liebert EXL and eBoost, shipping H2 2026 aligned with NVIDIA specs) plus PurgeRite integration produces a combined 800 VDC + liquid cooling proposition Vertiv can sell as one project. The Vertiv NearZero fluid management service, launched 2025, is the commercial expression.


Eaton acquired Boyd Thermal in 2024 for approximately $9.5B, the largest single thermal-industry M&A deal in the sector. Boyd brings thermal design and manufacturing scale that Eaton's power side lacked. The combined offering is Eaton's answer to Vertiv's Liebert + PurgeRite combination. The DD-series Supplement B (Eaton/Boyd worked example) treats this transaction in full commercial DD detail.


Schneider Electric acquired Motivair in early 2025 for approximately $1.7B for a majority stake. Motivair's liquid cooling equipment complements Schneider's Galaxy VXL 800 VDC UPS line, replicating the Vertiv/Eaton pattern from the Schneider side.



Chart 4. Thermal + power vendor consolidation 2023-2026, named transactions


Three named transactions (Vertiv/PurgeRite $1B+, Eaton/Boyd Thermal $9.5B, Schneider/Motivair $1.7B) reshaped the thermal + power vendor set into three integrated global vendors. Emerging fourth: Delta Electronics + potential thermal acquisition, publicly discussed but not announced as of mid-2026.



Transaction data from Vertiv Q4 2023 investor relations, Eaton Q2 2024 announcement, Schneider Electric Q1 2025 acquisition disclosure.




06Monitoring integration: the next competitive layer


The vendors that dominated the 800 VDC + liquid cooling combined offering through 2024-2026 competed on hardware integration. Being able to sell the customer both power and cooling as a single project. The competitive frontier for 2027-2028 shifts to monitoring integration: bringing thermal and power monitoring into a unified control plane that the operator can manage as one system rather than two.


Currently most facilities run thermal monitoring (temperatures, flow rates, coolant chemistry) on one system and power monitoring (currents, voltages, arc-fault detection) on another. The two systems have limited cross-visibility. A developing thermal fault that would eventually trigger a power event is not visible to the power monitoring until the power event begins. The next-generation control plane merges the two data streams and enables predictive intervention.


The vendor best positioned to deliver this unified monitoring is not obvious. Vertiv, Eaton, and Schneider all have monitoring platforms of varying maturity. Delta and ABB are strong in industrial-style unified monitoring but less penetrated in hyperscale. Software specialists like AVEVA and OSIsoft are potential integrators. The competitive winner will be determined by whichever vendor lands the first named hyperscaler contract for unified thermal + power monitoring at a 100+ MW campus.



07Named cooling technology cases




VendorCooling technologyDC power integrationNamed hyperscaler references


Vertiv + PurgeRiteDirect-to-chip; Liebert XDU CDU; NearZero fluid managementLiebert EXL S1 + native 800 VDC roadmapMultiple hyperscaler references (specific names in Vertiv IR)
Eaton + Boyd ThermalCold plates + rear-door heat exchangers9395P series + DC-native roadmapBoyd's pre-acquisition hyperscaler book; expanding under Eaton
Schneider + MotivairDirect-to-chip CDUs; row-level chilled waterGalaxy VXL + BATTERY MODULAR CX at 800 VDCMultiple hyperscaler references via Motivair legacy
CoolIT SystemsDirect-to-chip cold plates + rack CDUsVendor-agnostic on powerStrong hyperscaler penetration; independent supplier
SubmerImmersion coolingVendor-agnostic on powerEmerging in edge and specialty deployments
LiquidStackImmersion coolingVendor-agnostic on powerNamed hyperscaler pilots
Delta ElectronicsGrowing thermal capability, potential acquisition target for expansionUltron DPS + Modulon DPHStrong in APAC; expanding in NA/EU





08Interactions with earlier essays in this series


The cooling architecture choice interacts with prior decisions in this series across three vectors. Arc-fault protection (Part IV): leaked coolant can create arc-fault initiators, and continuous leak detection has to be integrated with arc-fault detection. Grounding topology (Part V): coolant loops that connect multiple racks create shared ground paths that have to be accounted for in the grounding scheme. Battery integration (Part VI): battery thermal management interacts with the facility cooling loop, and battery-side thermal events (rare but consequential) have to be isolated from the compute-side cooling.


The practical takeaway for operators is that cooling cannot be procured as a separate workstream from the electrical protection design. Vendor discussions that treat cooling as an add-on to a power purchase (or vice versa) miss the coordination requirements. The vendors that have consolidated to sell both as one project (Vertiv, Eaton, Schneider) have this coordination inside their engineering; the vendors that supply only one side (CoolIT for cooling, or a power-only vendor) leave the coordination as an operator responsibility.



09Operator procurement checklist


 1. Target rack density documented at architecture selection. Above 100 kW per rack, liquid cooling is required and 800 VDC becomes the sensible power choice. Below 60 kW per rack, air cooling and 400 VDC hybrid may be economic.
 2. Cooling technology decision integrated with power architecture. Direct-to-chip is the default at 80-400 kW per rack; immersion becomes competitive at 400+ kW; air cooling stays below 60 kW.
 3. Coolant chemistry specified against DC electrical compatibility. Water-glycol requires dielectric barriers. Dielectric fluids remove the barrier requirement at premium cost. Specialty low-conductivity fluids are middle path.
 4. Leak detection integrated with arc-fault detection (Part IV). Continuous leak sensors, drip-pan conductivity monitoring, rapid isolation protocols specified at Stage 1.
 5. Vendor single-vs-multi-vendor decision documented. Vertiv/Eaton/Schneider offer integrated power + cooling; CoolIT/Submer offer cooling-only with vendor-agnostic power interface. Trade-off is integration convenience vs vendor lock-in.
 6. Monitoring integration plan. Thermal + power monitoring in one control plane or two? If two, define the cross-system integration protocol at Stage 1.
 7. Thermal + power vendor commercial terms consolidated. If procuring both from the same vendor, negotiate as one project with unified commissioning and warranty terms rather than as two separate contracts.
 8. Cooling capex projected against the 10-year hold. Direct-to-chip has higher day-one capex than air cooling but lower TCO at high density (see Part VII on retrofit vs greenfield TCO methodology).



10The reframe for anyone specifying now


The 800 VDC transition and the liquid-cooling transition are not two independent decisions. Above 100 kW per rack, choosing one implies the other. Operators that treat cooling as a Stage-4 procurement decision separate from the Stage-1 power architecture decision consistently discover in commissioning that the two workstreams were designed against inconsistent assumptions. The vendor set that has consolidated over 2023-2025 has done so precisely because customers want both from one vendor. The vendors that supply only one side of the transition are increasingly at competitive disadvantage relative to the integrated players.


Part IX moves into power quality and grid interaction. The utility-facing implications of the 800 VDC architecture that most operators discover at interconnection study and that regulators are only beginning to reflect in their standards.




Glossary of terms used



800 VDC800 Volts Direct Current. The emerging standard voltage class for AI-scale data centre power distribution.
CDUCoolant Distribution Unit. Thermal system component distributing chilled coolant to racks or direct-to-chip cold plates.
CFMCubic Feet per Minute. Volumetric flow-rate unit used in air-cooling capacity calculations.
GPUGraphics Processing Unit. The compute silicon at the centre of AI workloads.
HBMHigh Bandwidth Memory. Stacked memory technology used adjacent to AI GPUs.
OCPOpen Compute Project. Hyperscaler-led standards body developing open reference designs for data centre hardware including Mt. Diablo 800 VDC.
PUEPower Usage Effectiveness. Data centre efficiency metric. Total facility power divided by IT equipment power. Lower is better.
TCOTotal Cost of Ownership. Capex plus opex over a defined hold period, discounted to present value.
UPSUninterruptible Power Supply. Battery-backed power system that keeps critical loads running during grid disruptions.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.



Related topic hubs
 * 800V Data Centre Power
 * AI Power Semiconductors
 * Browse all topic hubs →


Method and sources. Public information only. Rack density trajectory from OCP Mt. Diablo target specifications, NVIDIA GB200/GB300 platform disclosures, and hyperscaler earnings-call commentary. Cooling capex data from Vertiv Liebert XDU, CoolIT Systems reference designs, Submer immersion vendor data. Vendor consolidation transactions from Vertiv Q4 2023 IR, Eaton Q2 2024 announcement, Schneider Q1 2025 acquisition disclosure. Coolant chemistry compatibility from 3M Novec, Solvay, and Chemours published technical data. No advisory relationship with any named party.


Series footer. Part VIII in The DC-DC TransitionRelated reading: Part IV on arc protection (leak-detection interaction), Part V on grounding (shared coolant paths), Part VI on battery integration (battery thermal management), Part VII on retrofit vs greenfield (cooling plant is largest retrofit cost), Part XI on vendor economics (consolidation dynamics). Companion context: The AI Power Chain Part III on Thermal Stack, Due Diligence for the AI Buildout Supplement B on Eaton/Boyd worked example.


Written in a personal capacity. No advisory conflict on any named party. Nothing here is investment advice.











The DC-DC Transition: start with the foundation

 1. The real reason data centres are going DC (four forces)
 2. Two architectures wearing the same name (800V DC vs OCP Mt Diablo)
 3. The six-layer architecture map


==============================================================================
# Power quality, harmonics, and grid interaction at 800 VDC (DC-DC Transition IX)
URL: https://adikumar.co/dc-dc-transition-09-power-quality-grid/
Published: 2026-08-20
Summary: Harmonics from 6-pulse vs 12-pulse vs active-front-end rectifiers, PFC at 800 VDC, grid-support capability, ISO-level regulatory posture.
==============================================================================

The DC-DC Transition series · Part 9 of 19









The DC-DC Transition · Part IX of XIII



Power quality, harmonics, and grid interaction at 800 VDC


An 800 VDC data centre presents a different power-quality profile to the utility grid than a legacy AC facility. Harmonics from active rectifiers, capacitive loading, and DC bus stability interact with utility protective devices in ways that most interconnection studies do not fully model. Some utilities are already refusing 100 MW+ single-facility interconnections above baseline standards; others will follow within 12-18 months.


Part IX in the DC-DC Transition series · independent analysis · no advisory conflict on any named party





The DC-DC Transition · thirteen essays for data centre power architects

 1.  I. The real reason data centres are going DC
 2.  II. Two architectures wearing the same name
 3.  III. The architecture map
 4.  IV. Arc behaviour + insurance
 5.  V. Grounding + ground-fault protection
 6.  VI. Battery integration at 800 VDC
 7.  VII. Retrofit vs greenfield
 8.  VIII. 800 VDC and liquid cooling co-emergence
 9.  IX. Power quality + grid interaction (you are here)
 10. X. Standards: OCP, IEC, NEC, IEEE
 11. XI. Vendor economics: who wins the transition
 12. XII. Commissioning, skills, operational readiness
 13. XIII. The ten-year view





01What the utility sees when an 800 VDC facility connects



From the utility's perspective, an 800 VDC data centre is still an AC load at the PCCWhat differs is what happens inside the customer facility after the utility power crosses the meter. The AC-to-DC rectifier front end that produces the 800 VDC bus draws current in a non-sinusoidal shape that generates harmonics on the utility side. The capacitor bank on the DC bus behaves like a capacitive load from the utility's frame of reference. And the fast-switching power electronics that control the rectifier and downstream converters inject high-frequency noise back onto the utility conductor.



The intuition that DC internal architecture is invisible to the utility is wrong. Every 800 VDC facility presents specific power-quality features at the PCC that the utility protective system has to accommodate. The features are not new. Variable-speed drives, rectifier plants, and traction substations have generated similar profiles for decades. But the concentration and scale at hyperscale AI facilities is new, and utility interconnection studies are only beginning to model it correctly.




02Harmonic sources in the AC/DC rectifier front end



A typical 6-pulse rectifier feeding an 800 VDC bus generates the classic harmonic set: 5th, 7th, 11th, 13th, 17th, 19th, and higher orders. A 12-pulse rectifier reduces the 5th and 7th but leaves the 11th and 13th. A 24-pulse rectifier reduces further at cost of complexity. Modern active front ends using IGBTs or WBG devices operate as sinusoidal-current rectifiers and produce near-unity power factor with low harmonic distortion. But they inject high-frequency switching noise instead.




Chart 1. Harmonic spectrum at the PCC: 6-pulse vs 12-pulse vs active front end


6-pulse rectifiers produce 20-30 percent THD dominated by 5th and 7th harmonics. 12-pulse cuts THD roughly in half but leaves higher-order content. Active front ends with WBG devices produce sub-5-percent THD but generate switching-frequency noise in the kHz range that requires different filtering.



Harmonic spectrum data from IEEE 519 examples, EPRI power-quality reference cases, and vendor datasheet compilations (ABB, Schneider Electric, Delta Electronics active front-end products).




The choice of rectifier topology at Stage-1 architecture selection interacts directly with the utility interconnection study at Stage-4. Facilities that assume 6-pulse rectifiers because they are cheaper often discover during utility review that the harmonic profile triggers utility protective device concerns or violates IEEE 519 THD limits at the PCC.




03Filter topology options and their trade-offs



The three filter topology categories address the harmonic profile differently. Passive filters (LC networks tuned to specific harmonic orders) are cheap and reliable but bulky and only address the tuned frequencies. Active filters (IGBT-based, dynamically injecting compensation currents) are compact and address a wide frequency range but expensive and produce their own switching artefacts. Hybrid filters combine passive tuning for dominant harmonics with active compensation for residual content.




Chart 2. Filter topology cost, footprint, and effectiveness comparison


Passive filters have the lowest capex per MVA of filtering, largest footprint, and coverage limited to tuned frequencies. Active filters have the highest capex per MVA, smallest footprint, and wide-band coverage. Hybrid systems land in the middle on all three axes and dominate current deployments.



Vendor product data from Schneider AccuSine, ABB PQF, Merus Power, Comsys ADF and analogous product lines 2024-2026.





04Power factor correction at 800 VDC



Traditional AC data centre PFC (correcting the near-unity power factor deviation caused by rectifier operation) is a well-understood engineering problem. What changes at 800 VDC is that the reactive-power profile depends more on the DC bus capacitance than on the load impedance. Large DC bus capacitor banks appear as substantial capacitive reactive-power draw at the PCC, requiring capacitive-side PFC compensation that traditional inductor-based schemes do not provide.



Modern active PFC integrated into the rectifier front end handles both directions of reactive-power compensation dynamically. Vendors shipping in 2026 include ABB (PQC 500), Schneider Electric (VariPFC), Merus Power (MSVC), and Comsys (ADF P100). Selection depends on facility scale, the specific reactive-power profile, and the coordination requirements with utility voltage-regulation equipment.




05Grid stability contribution: 800 VDC facilities as good or bad neighbours



Whether an 800 VDC facility contributes positively or negatively to grid stability depends on its architecture. Facilities with active front ends and appropriate control loops can provide voltage support, frequency ride-through, and even primary reserve service to the utility. Behaving like a small power plant on the demand side. Facilities with passive front ends and no active grid support degrade grid stability by consuming reactive power and injecting harmonics.



The economic case for architecting an 800 VDC facility as a grid-supporting load rather than a grid-degrading load is thin at day one. The incremental capex is real and the utility rate-structure benefits are limited under current regulatory frameworks. But the political case is substantial. Utilities and state regulators are beginning to distinguish between data centres that stress the grid and data centres that support it. The first cohort of grid-supporting hyperscale facilities is establishing the reference for what regulators will require from later entrants.




Chart 3. Grid stability contribution by facility architecture


Passive-rectifier facilities score low on every grid-service axis. Active-front-end facilities can offer voltage support and frequency ride-through with modest control-system additions. Fully grid-supporting facilities (rare in 2026, expected common by 2029) offer primary reserve and voltage regulation as ancillary services.



Grid-service capability scoring adapted from FERC Order 2222 framework for distributed energy resources.





06Utility protective device interaction



Interconnection studies for large data centre loads have historically focused on capacity. Can the local substation deliver the requested load?. Rather than on power-quality interaction. That is changing. Utility engineers are increasingly running detailed power-quality studies for large data centre interconnections, and the studies are surfacing specific interaction issues.



Three interactions matter most. Distribution-level capacitor banks can resonate with the capacitance of the DC bus, creating voltage transients at harmonic frequencies. Utility protective relays configured for traditional loads can misoperate on the harmonic-distorted current waveform. Utility voltage-regulation equipment (line-drop compensators, capacitor-bank switching) can interact with active PFC in the facility, creating instability if not coordinated.




Chart 4. Utility interconnection interaction rate by facility architecture


Passive-rectifier facilities show the highest rate of interconnection-study issues (resonance events, protective-relay misoperations, voltage transients). Active-front-end facilities with coordinated filtering show substantially lower rates. Facilities designed for grid-supporting service show the lowest rates of all.



Interaction rate estimates from published utility interconnection study proceedings and IEEE PES conference papers 2023-2026.





07Regulatory posture by ISO



ISO-level regulatory posture on data centre interconnection varies substantially. PJM has proposed carve-outs for co-located data centres and is running a comprehensive queue reform through FERC. ERCOT has paused Batch Zero (see the grid-queue thesis piece) and is running a full audit. MISO has announced delays on multiple queue cycles. CAISO carries less transparency but similar underlying constraints.





ISOCurrent postureExplicit data centre frameworkRecent significant event


PJMQueue reform in FERC review; co-located carve-out proposedEmerging; state-by-state variation2026 queue reform proceedings at FERC
ERCOTBatch Zero paused Aug 3; audit ongoingUnder active PUC reviewAug 3 Abbott directive; BNEF $15B / 49.8 GW at risk
MISOMultiple cycle delays announcedStandard queue process, no DC-specific framework2022/2023/2025 cycle delays
CAISOQueue reform in progress; less transparencyGeneral framework; California policy overlayQueue expansion 2025-2026
NYISO / ISO-NESmaller data centre load; less pressureGeneral frameworkNot the leading indicator ISOs






08Operator interconnection-study checklist



The interconnection-study workstream needs specific documentation before an operator walks into the utility conversation. Ten items cover the substance for an 800 VDC facility.



 1.  Rectifier topology documented (6-pulse / 12-pulse / active front end). Affects harmonic profile and filtering requirements.
 2.  Harmonic-current profile at PCC modelled against IEEE 519 limits. Not a template; a facility-specific study.
 3.  Filter topology specified (passive / active / hybrid) and coordinated with utility protective relaying.
 4.  Power factor correction plan documented including capacitive-side compensation for DC bus capacitance.
 5.  Grid-support capability declared (voltage support / frequency ride-through / primary reserve if applicable).
 6.  Utility voltage-regulation coordination study on record. Line-drop compensators, capacitor-bank switching, active PFC interaction.
 7.  Resonance study with distribution-level capacitor banks completed. DC bus capacitance can resonate with utility capacitor banks.
 8.  Protective-relay setting review coordinated with the utility. Harmonic-distorted current waveforms can trigger misoperation.
 9.  Ride-through capability documented per utility ride-through standard applicable in the ISO.
 10. Political-response contingency plan documented. ISO-level regulatory posture can shift quickly (see ERCOT). Operators need a fallback interconnection plan.




09The reframe for anyone specifying now



Power quality and grid interaction is the workstream that most data centre operators discover late, and where the discovery is most expensive when it happens. Interconnection studies are shifting from capacity-only to include full power-quality analysis. Facilities specified against 5-year-old interconnection standards find themselves rejected or delayed when the standards catch up mid-project. The Stage-1 architecture decision on rectifier topology, filter strategy, and grid-support capability determines whether the facility passes the interconnection study on first submission or bounces through multiple revisions.



Part X moves into the standards framework. OCP, IEC, NEC, IEEE. That is codifying all of the design choices addressed in Parts IV through IX.





Glossary of terms used



CAISOCalifornia Independent System Operator. The grid operator for most of California.
ERCOTElectric Reliability Council of Texas. Grid operator for most of Texas, operating largely as an electrical island.
IGBTInsulated Gate Bipolar Transistor. Power semiconductor used in inverters, converters, and active filters.
ISOIndependent System Operator. Regional grid operator (ERCOT, MISO, PJM, CAISO).
MISOMidcontinent Independent System Operator. Regional grid operator covering 15 US states and Manitoba.
PCCPoint of Common Coupling. The electrical connection point between a customer facility and the utility grid.
PFCPower Factor Correction. Circuit or system that reduces reactive power drawn from the supply.
PJMPJM Interconnection LLC. Regional grid operator covering 13 US states plus DC.
THDTotal Harmonic Distortion. Measure of harmonic content in a waveform, typically expressed as a percentage of the fundamental.
WBGWide-Bandgap. Semiconductor material class (SiC and GaN) used in high-efficiency power electronics.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.




Related topic hubs
 * 800V Data Centre Power
 * AI Power Semiconductors
 * Browse all topic hubs →


Method and sources. Public information only. Harmonic-spectrum data from IEEE 519 examples, EPRI power-quality reference cases, and vendor product datasheets (ABB PQF and PQC, Schneider AccuSine and VariPFC, Merus Power MSVC, Comsys ADF, Delta Electronics active front ends). Utility interconnection posture from PJM queue reform filings at FERC, ERCOT market notices, MISO GIA study updates, CAISO queue publications. Grid-service framework from FERC Order 2222. No advisory relationship with any named party.


Series footer. Part IX in The DC-DC TransitionRelated reading: Part II on the two 800 VDC architectures (rectifier topology choice), Part IV on arc-fault protection (interaction with utility protection), Part X on standards evolution, Part XI on vendor economics. Companion context: The AI Power Chain Part IV on the Interconnect Stack, "The grid queue is the constraint" thesis piece on ISO-level regulatory posture, Due Diligence for the AI Buildout Part XI on regulatory DD.


Written in a personal capacity. No advisory conflict on any named party. Nothing here is investment advice.













The DC-DC Transition: start with the foundation

 1. The real reason data centres are going DC (four forces)
 2. Two architectures wearing the same name (800V DC vs OCP Mt Diablo)
 3. The six-layer architecture map


==============================================================================
# Standards evolution: OCP, IEC, NEC, IEEE, and the 2028-2029 convergence (DC-DC Transition X)
URL: https://adikumar.co/dc-dc-transition-10-standards-evolution/
Published: 2026-08-20
Summary: Standards convergence timeline for 800 VDC data centres. OCP Mt. Diablo, IEC 60947-10, NEC 2029 cycle, IEEE P2818, UL and NFPA integration.
==============================================================================

The DC-DC Transition series · Part 10 of 19









The DC-DC Transition · Part X of XIII



Standards evolution: OCP, IEC, NEC, IEEE, and the 2028-2029 convergence


Standards for 800 VDC data centres are converging but from four different starting points. OCP is fastest. Mt. Diablo published as open spec May 2025, 80+ ecosystem partners active by mid-2026. IEC leads on protection standards (IEC 60947-10 for SSCBs published early 2026). NEC and IEEE trail by 24-36 months. Insurance and certification frameworks (UL, NFPA) are being drawn in through OCP working-group participation. Operators specifying facilities during the interim period must design against multiple standards concurrently.


Part X in the DC-DC Transition series · independent analysis · no advisory conflict on any named party





The DC-DC Transition · thirteen essays for data centre power architects

 1.  I. The real reason data centres are going DC
 2.  II. Two architectures wearing the same name
 3.  III. The architecture map
 4.  IV. Arc behaviour + insurance
 5.  V. Grounding + ground-fault protection
 6.  VI. Battery integration at 800 VDC
 7.  VII. Retrofit vs greenfield
 8.  VIII. 800 VDC and liquid cooling co-emergence
 9.  IX. Power quality + grid interaction
 10. X. Standards: OCP, IEC, NEC, IEEE (you are here)
 11. XI. Vendor economics: who wins the transition
 12. XII. Commissioning, skills, operational readiness
 13. XIII. The ten-year view





01Why standards fragmentation exists



The current standards landscape for 800 VDC data centres has four active tracks with different geographic scopes, different technical remits, and different publication timelines. That fragmentation is not accidental. Each body has legitimate authority over some part of the design space and none has jurisdiction over the whole. The result is that operators specifying a facility in 2026 must reference multiple standards at once, and vendors shipping equipment must certify against multiple frameworks in parallel.



OCP is the fastest track because it is customer-driven (Google, Meta, Microsoft, Amazon, Oracle are all active members) and does not require regulatory approval to publish. Mt. Diablo. The reference design for 800 VDC power delivery. Was announced October 2024 and published as an open specification in May 2025. Ecosystem partner count has grown from roughly 30 at spec publication to 80+ by mid-2026. OCP is not a certification body; its output is reference designs and open specifications that vendors implement.



IEC is the primary technical standards body for the electrical protection and safety layer. IEC 60947-10 (semiconductor-based circuit breakers) was published early 2026 and is the anchor standard for SSCB certification globally. IEC TC 8 (electrical energy supply systems) and IEC SC 8B (decentralized electrical energy systems) are running additional 800 VDC-related work programmes.



NEC updates on a three-year cycle. The 2026 revision includes preliminary provisions for higher-voltage DC in commercial installations. Comprehensive 800 VDC coverage is expected in the 2029 cycle. Because NEC is the enforceable installation code in most US jurisdictions, its adoption timeline sets the practical floor for what is legally installable.



IEEE P2818 is the specific project for high-voltage DC data centre standards, currently in committee. Related IEEE standards (P1547 for interconnection, P519 for harmonics. See Part IX) apply already. IEEE 1584 (arc-flash analysis) has been extended for DC applications and is the reference for arc-flash studies (see Part IV).




02The OCP track: Mt. Diablo and the 80+ partner ecosystem



OCP's Mt. Diablo specification defines a disaggregated power rack. A "sidecar" configuration where the 800 VDC power infrastructure sits alongside the compute rack rather than integrated into it. The specification pushes power delivery from the current 48 VDC rack-level standard to either ±400 VDC or 800 VDC, enabling IT rack densities from 100 kW to 1 MW. The specification defines mechanical, electrical, and interface characteristics but is silent on some safety and protection aspects that are covered by IEC and NEC.



The ecosystem participant count is the operative signal for how deployment-ready the specification is. At publication (May 2025) approximately 30 vendors and hyperscalers were named partners. By mid-2026 that count exceeds 80 and includes every major power equipment vendor (Vertiv, Eaton, Schneider Electric, ABB, Delta, Siemens Energy, Hitachi Energy), every major hyperscaler (Google, Meta, Microsoft, Amazon, Oracle), and a growing set of semiconductor and component suppliers (Infineon, Wolfspeed, Onsemi, STMicro, Menlo Micro).




Chart 1. OCP Mt. Diablo ecosystem partner growth 2024-2026


OCP ecosystem partner count for Mt. Diablo grew from initial announcement (Oct 2024) to spec publication (May 2025) to mid-2026 at accelerating pace. The count is a leading indicator of vendor commitment. Every incremental partner reduces the buyer's implementation risk and increases the specification's staying power.



Partner count from OCP working-group participant registries and Google/Meta/Microsoft public communications 2024-2026.





03The IEC track: protection standards leading



IEC 60947-10 (semiconductor circuit breakers) is the most substantive 2026 publication in the 800 VDC space. Its scope is narrow. Governing SSCB certification. But its practical importance is disproportionate. Before IEC 60947-10 was published, SSCBs were experimental products without a governing standard, which meant utilities and insurers could not underwrite against defensible criteria. Publication has moved SSCBs from research equipment to commercially specifiable protection.



Other IEC standards in the 800 VDC pipeline include IEC 62933 (electrical energy storage systems, revision for high-voltage DC), IEC 61643 (surge protective devices, updates for DC applications), and IEC 62109 (safety of power converters for use in photovoltaic power systems, being extended for data centre applications). Publication cadence for IEC is 18-30 months from committee draft to final publication.




Chart 2. IEC standards pipeline for 800 VDC data centres, publication timeline


IEC 60947-10 published early 2026 (SSCBs). Additional standards in draft covering surge protection, energy storage, power converters, and protective devices. Full IEC coverage of the 800 VDC data centre stack expected by 2028-2029.



IEC standards pipeline from IEC publication register and TC 8 / SC 8B / TC 21 / TC 121 committee work programmes.





04The NEC track: three-year cycle, 2029 for comprehensive coverage



The NEC updates every three years. The 2023 cycle included limited provisions for higher-voltage DC in commercial installations. The 2026 cycle (approved during 2025) adds more explicit provisions but does not cover the 800 VDC data centre application comprehensively. The 2029 cycle is expected to include full 800 VDC data centre coverage based on the OCP and IEC work through 2027-2028.



The practical implication for operators is that the NEC in effect at the time of a facility's electrical installation is the enforceable standard. Facilities installed in 2026 reference the 2023 NEC (or the 2026 revision depending on jurisdiction adoption). Facilities installed in 2028 will reference the 2026 NEC. Facilities installed in 2030 will reference the 2029 NEC. Operators need to be aware which NEC edition their local jurisdiction has adopted and whether it accommodates 800 VDC. Some jurisdictions do, some require variance applications.




05The IEEE track: P2818 and adjacent standards



IEEE P2818 is the specific project for standards on high-voltage DC data centres. Committee work began 2024, initial draft is expected 2027, publication window 2028-2029. The IEEE process is slower than OCP but produces internationally recognised technical standards with certification pathways.



Adjacent IEEE standards apply already. IEEE 519 governs harmonic limits at the point of common coupling (see Part IX). IEEE 1547 governs interconnection of distributed resources (relevant to grid-supporting data centre facilities). IEEE 1584 (arc-flash analysis, DC extension) is the reference for arc-flash studies (Part IV). IEEE P2800 governs large-load interconnection to bulk power systems and is directly relevant to hyperscale data centre facilities.




06Insurance and certification: UL and NFPA integration



UL is drafting the North American equivalent to IEC 60947-10 for solid-state circuit breakers. UL Solutions has active participation in the OCP 800 VDC working group and is signalling final publication in 2027. UL 9540 (energy storage systems) is being updated for high-voltage DC applications. UL 3400 (a proposed standard for solid-state fault current limiters) is in draft.



NFPA 70E. The electrical safety in the workplace standard that governs arc-flash PPE requirements. Is being updated for 800 VDC applications. The 2024 edition addresses DC applications in a limited way; the 2027 edition is expected to include comprehensive coverage.



Insurance carriers rely on UL and NFPA certification pathways to underwrite property policies. Facilities using UL-certified SSCBs and NFPA-70E-compliant PPE programmes command lower insurance premiums than facilities using non-certified equipment (see Part IV on insurance premium tiering). The economic case for early UL/NFPA certification adoption is documented in the premium differential.




07Standards convergence timeline




Chart 3. Standards convergence timeline 2024-2030 across four tracks


OCP publishes fastest (Mt. Diablo May 2025, ongoing spec enrichment). IEC follows with protection standards (60947-10 early 2026, additional 2027-2028). NEC updates in 2026 and 2029 cycles. IEEE P2818 publishes 2028-2029. UL and NFPA integrate through 2027-2028. Full convergence lands 2028-2029.



Publication schedules from each standards body's public work programme 2024-2026.





08Geographic adoption patterns



Geographic variation in standards adoption matters for operators building multi-region facility portfolios. EU adoption of IEC standards is direct and relatively fast (IEC is domiciled in Geneva, and CENELEC harmonises IEC standards for EU application). US adoption is slower and passes through UL, NFPA, IEEE, and state-level electrical code adoption. APAC varies by country. Japan, Korea, and Taiwan follow IEC directly; China maintains its own GB standards with partial harmonisation; India is transitioning to IEC-alignment.




Chart 4. Geographic adoption of 800 VDC standards, mid-2026


EU is the fastest-adopting region because IEC standards are directly enforceable. US adoption is slower due to the NEC three-year cycle and state-level variation. APAC varies substantially by country. Operators building multi-region portfolios must design against the most restrictive standard in their footprint.



Adoption scoring from published national electrical code adoption records and vendor market feedback 2024-2026.





09The multi-standard compliance premium for interim projects



Operators specifying facilities during the 2026-2028 interim period face a specific cost premium: they must design against multiple standards concurrently because no single framework covers the full stack. The specific line items add up.



Vendor equipment must carry both IEC and UL certifications (or accept the risk of jurisdiction-specific rejection). Insurance premium reflects the compound underwriting question of multi-standard compliance. Engineering design cost is higher because designers must reference OCP+IEC+NEC+IEEE simultaneously rather than a single unified code. Commissioning and inspection time is longer because inspectors are learning the standards themselves.



The multi-standard compliance premium runs roughly 8-15 percent above what an equivalent facility would cost once full standards convergence lands in 2028-2029. Operators building facilities in the interim window make an implicit bet: is the operational benefit of deploying 800 VDC in 2026-2027 worth the multi-standard compliance premium versus deploying in 2028-2029 with unified standards?




Chart 5. Multi-standard compliance premium over the interim period


Facilities specified in 2026-2028 carry an 8-15 percent capex premium relative to what the same specification will cost once standards converge in 2028-2029. The premium reflects vendor certification duplication, engineering complexity, and insurance underwriting overhead.



Premium estimate from author's synthesis across vendor pricing 2024-2026 and comparison to projected 2029 unified-standards baseline.





10Operator implications during the interim period



 1. Design against the intersection of applicable standards (OCP Mt. Diablo v-latest, IEC 60947-10 for SSCBs, NEC 2026 or later depending on jurisdiction, applicable IEEE standards). Not the union. The intersection, which is the operationally safe subset.
 2. Specify vendor equipment with dual certification (IEC + UL where operating in US, IEC + CENELEC where operating in EU, IEC + GB where operating in China).
 3. Engage insurance underwriters early with the specific standards references documented. Multi-standard compliance evidence reduces premium.
 4. Track NEC adoption in your local jurisdiction. Different states adopt different NEC editions on different timelines; verify which edition your permit will be reviewed against.
 5. Participate in OCP working groups where possible. Operator perspective in the specification-development process reduces the risk that the eventual specification does not fit your operational reality.
 6. Plan the update path to 2029 unified standards. Interim facilities will need retrofit or documentation updates when the unified standards land; budget accordingly.




11The reframe for anyone specifying now



Standards fragmentation in 2026 is a real operational cost, not a theoretical issue. Operators who assume "the standards will catch up" and delay their 800 VDC deployment plans are also delaying the operational and efficiency benefits by 2-3 years. Operators who deploy in the interim window and accept the multi-standard compliance premium capture the benefits earlier but pay for the premium. Neither approach is wrong; both have implications the DD workstream needs to price explicitly.



Part XI moves into vendor economics. Who wins the 800 VDC transition and what that means for equity and PE investors.





Glossary of terms used



IECInternational Electrotechnical Commission. Global standards body for electrical and electronic technologies.
IEEEInstitute of Electrical and Electronics Engineers. Global professional association publishing power and communications standards.
NECNational Electrical Code. US electrical installation code published by NFPA, updated on a three-year cycle.
NFPANational Fire Protection Association. US organisation publishing fire and electrical safety codes including the NEC and NFPA 70E.
OCPOpen Compute Project. Hyperscaler-led standards body developing open reference designs for data centre hardware.
PPEPersonal Protective Equipment. Arc-flash suits, gloves, and face shields rated to specific incident-energy levels per NFPA 70E.
SSCBSolid State Circuit Breaker. Semiconductor-based protective device that interrupts DC fault current in microseconds. Governed by IEC 60947-10 published early 2026.
ULUnderwriters Laboratories. US safety certification body that tests and certifies electrical equipment against applicable codes.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.




Related topic hubs
 * 800V Data Centre Power
 * AI Power Semiconductors
 * Browse all topic hubs →


Method and sources. Public information only. OCP Mt. Diablo specification and ecosystem partner count from OCP publications and Google/Meta/Microsoft newsrooms 2024-2026. IEC publication schedule from IEC publication register (60947-10, 62933, 61643, 62109). NEC revision cycle from NFPA 70 publication schedule. IEEE standards from P2818 committee work programme, IEEE 519/1547/1584/P2800 published standards. UL Solutions position from OCP working group participation. NFPA 70E from published revision schedule. No advisory relationship with any named party.


Series footer. Part X in The DC-DC TransitionRelated reading: Part IV on arc-fault protection (SSCB certification via IEC 60947-10), Part V on grounding (IEC 61643 SPD coverage), Part IX on power quality (IEEE 519 harmonic limits), Part XI on vendor economics (certification cost as market entry barrier). Companion context: Due Diligence for the AI Buildout Part XI on regulatory DD, The AI Power Chain Part VI on Modular Datacenter Stack standardisation history.


Written in a personal capacity. No advisory conflict on any named party. Nothing here is investment advice.













The DC-DC Transition: start with the foundation

 1. The real reason data centres are going DC (four forces)
 2. Two architectures wearing the same name (800V DC vs OCP Mt Diablo)
 3. The six-layer architecture map


==============================================================================
# Vendor economics: who wins the 800 VDC transition (DC-DC Transition XI)
URL: https://adikumar.co/dc-dc-transition-11-vendor-economics/
Published: 2026-08-20
Summary: 800V DC vendor economics: six-layer competitive map, 48V incumbents at risk, and where incumbents win the transition. Named vendors, exposure %, moat.
==============================================================================

The DC-DC Transition series · Part 11 of 19







The DC-DC Transition · Part XI of XIII



Vendor economics: who wins the 800 VDC transition


The vendor set that dominates 800 VDC is not the same as the vendor set that dominates 48 VDC. Some incumbents lose 20-40 percent of their revenue base. Some incumbents win through repositioning with WBG semiconductor content. New entrants emerge in solid-state protection, HVDC conversion, and hybrid supercapacitor bridging. PE and public equity investors need to distinguish which segment each vendor sits in. Sector-level screening misses the layer-level winners.


Part XI in the DC-DC Transition series · independent analysis · no advisory conflict on any named party




The DC-DC Transition · thirteen essays for data centre power architects

 1.  I. The real reason data centres are going DC
 2.  II. Two architectures wearing the same name
 3.  III. The architecture map
 4.  IV. Arc behaviour + insurance
 5.  V. Grounding + ground-fault protection
 6.  VI. Battery integration at 800 VDC
 7.  VII. Retrofit vs greenfield
 8.  VIII. 800 VDC and liquid cooling co-emergence
 9.  IX. Power quality + grid interaction
 10. X. Standards: OCP, IEC, NEC, IEEE
 11. XI. Vendor economics: who wins the transition (you are here)
 12. XII. Commissioning, skills, operational readiness
 13. XIII. The ten-year view




01Six-layer competitive dynamics summary


The six-layer AI Power Chain framework (essay III of this series, and the AI Power Chain vendor screen) maps directly onto the vendor-competitive dynamics at 800 VDC. Each layer has different vendor incumbents, different transition winners, and different revenue-at-risk profiles. Reading vendor economics at sector level misses the layer-specific reality.




Layer48V incumbents at risk800V transition winnersNew-entrant opportunity


CapacitorLow-voltage MLCC (Kemet, Yageo, Murata)Bulk high-voltage film + polymer (KEMET, Vishay, TDK)Hybrid supercap (Skeleton, JSR Micro, Maxwell)
Wide-bandgap semiSilicon MOSFET/IGBT (Infineon, ST, ONsemi legacy lines)SiC + GaN (Wolfspeed, Infineon SiC line, ONsemi Vcore, STMicro, ROHM)Vertical integration into modules (Menlo Micro, Navitas)
ThermalAir-cooling specialists (Munters, STULZ)Direct-to-chip liquid (Vertiv/PurgeRite, Eaton/Boyd, Schneider/Motivair, CoolIT)Immersion (Submer, LiquidStack, GRC)
InterconnectLow-voltage cable + connector legacyMV switchgear + transformer scale (ABB, Siemens, Hitachi, GE Vernova)SSCB / current-limiting (Menlo Micro, MessTek, Bender)
On-package power48V VRM specialists (Delta, TDK, Bel Fuse)Integrated voltage regulator (TDK/Ferric, Empower, ONsemi/Vcore)New IVR entrants riding NVIDIA/hyperscaler platforms
Modular datacenterPrefab-legacy air-cooling (Vertiv, PDI, Rittal)Prefab-DC-native + liquid-cooled (Vertiv, Schneider, DataQube)Software-defined power + modular DC-native (Delta, Applied Digital)





02Incumbents at risk: 48V specialists with revenue-at-risk


Three vendor archetypes face substantial revenue-at-risk from the 800 VDC transition. Low-voltage rectifier specialists (Delta, TDK, Bel Fuse in the sub-1000W 48V module segment) see 20-40 percent of their data centre revenue base at risk over 2027-2030 as customers migrate to native 800 VDC. MLCC specialists in the low-voltage segment (Yageo and portions of Murata/Kemet) see 15-30 percent at risk as the capacitor mix shifts to higher-voltage bulk types. Air-cooling specialists (Munters, STULZ) face structural pressure as direct-to-chip liquid cooling takes share above 60 kW per rack.



Chart 1. Revenue-at-risk for named incumbents in the 800 VDC transition


48V rectifier specialists face 20-40 percent revenue-at-risk over 2027-2030. MLCC LV specialists 15-30 percent. Air cooling specialists 30-50 percent above the 60 kW/rack threshold. These estimates depend on how quickly each vendor repositions and how successfully they capture 800 VDC-adjacent segments.



Revenue-at-risk estimates from author's vendor DD synthesis 2024-2026 plus published segment revenue data.




03Incumbents winning through repositioning: Vertiv, Eaton, Schneider


The three named vendor consolidators. Vertiv (via PurgeRite acquisition, native 800 VDC roadmap, ecosystem partnership with NVIDIA), Eaton (via Boyd Thermal acquisition, 9395P + DC-native line), Schneider Electric (via Motivair acquisition, Galaxy VXL + BATTERY MODULAR CX). Win the 800 VDC transition on the incumbent side. Each has combined electrical + thermal capability, hyperscaler qualification, and manufacturing scale sufficient to serve the demand.


Their win is not automatic. Each faces execution risk in the transition. Vertiv must integrate PurgeRite operations while shipping the Liebert 800 VDC portfolio in H2 2026 aligned to NVIDIA specs. Eaton must digest the $9.5B Boyd Thermal acquisition (closed March 2026 per company press release) and produce the combined offering at scale. Schneider must integrate Motivair while shipping Galaxy VXL and the BATTERY MODULAR CX at 800 VDC. Each has ~18 months of execution risk that will show in Q3 2026 through Q4 2027 earnings.



Chart 2. Incumbent winners: content-dollar-per-rack expansion at 800 VDC vs 48 VDC


Vertiv, Eaton, and Schneider each capture 2-3x more dollar content per rack at 800 VDC than at 48 VDC, on comparable installed IT capacity. The expansion reflects the combined electrical + thermal + protection scope that 800 VDC requires.



Content-dollar estimates from vendor investor day materials and analyst commentary 2024-2026.




04Wide-bandgap semiconductor content expansion


Wide-bandgap semiconductors (SiC and GaN) are the underlying enabling technology for 800 VDC power electronics at hyperscale efficiency. Named winners in the segment include Wolfspeed (SiC substrate + wafer + devices), Infineon (SiC + GaN device lines), ONsemi (SiC business built on Vcore acquisition), STMicroelectronics (SiC devices), and ROHM (Japanese SiC leader). The Chinese domestic supply chain (BYD Semiconductor, CanSemi, Silan) is expanding rapidly but faces certification hurdles for Western hyperscaler deployment.


The relevant financial dynamic is that WBG semiconductor content dollar-per-rack grows roughly 2-4x at 800 VDC vs 48 VDC, on comparable IT capacity. Each Vertiv or Eaton or Schneider rack ships with 3-4x the SiC or GaN device content that an equivalent 48 VDC rack shipped with. This is where WBG semiconductor investors capture the upside from the 800 VDC transition.



Chart 3. WBG semiconductor content dollar per rack: 48V vs 800V transition


SiC and GaN device content per rack rises 3-4x in the 800 VDC transition versus the 48 VDC baseline. Named WBG semiconductor vendors (Wolfspeed, Infineon SiC, ONsemi Vcore, STMicro, ROHM) all benefit proportionately. Chinese domestic vendors face certification hurdles that delay their capture of the Western hyperscaler share.



Content-dollar estimates from Wolfspeed, Infineon, ONsemi, STMicro investor materials 2024-2026 and semiconductor industry supply chain analysis.




05New entrants across the six layers


The 800 VDC transition creates new-entrant opportunity in six specific niches. Solid-state circuit breaker specialists (Menlo Micro is the highest-profile pure-play) address the protection layer. Hybrid supercapacitor specialists (Skeleton, JSR Micro, evolved Maxwell portfolio) address the peak-power buffering role. HVDC converter specialists (evolving from wind and solar heritage. GE Vernova, ABB, Hitachi Energy) enter the datacentre segment. Immersion cooling specialists (Submer, LiquidStack, GRC) compete against direct-to-chip. Integrated voltage regulator startups (TDK/Ferric, Empower Semiconductor) address the on-package layer. Software-defined power specialists (early-stage, some VC-backed) address the monitoring integration layer.



Chart 4. New-entrant landscape by layer, mid-2026


Solid-state protection (Menlo Micro leads with ~$800M valuation, Series D). Hybrid supercap (Skeleton IPO in Frankfurt 2024, JSR Micro subsidiary of JSR). HVDC (traditional players evolving). Immersion (Submer Series C, LiquidStack privately backed). IVR (TDK acquired Ferric 2025 at approximately $250M). Software-defined power (early-stage).



Valuation and funding data from Crunchbase, PitchBook, and company press releases 2024-2026.




06PE M&A implications: layer-specific acquisition targets


The PE M&A implications of the 800 VDC transition sort into three categories. First, roll-up opportunities in the fragmented segments. Direct-to-chip liquid cooling specialists (CoolIT, JetCool, others), 48V-to-800V converter niches, MV switchgear specialty players. Second, take-private opportunities for underperforming public incumbents that have not repositioned adequately (specific named candidates require deal-flow work rather than public analysis). Third, strategic-partner-driven acquisitions where a public strategic (Vertiv, Eaton, Schneider, ABB, Siemens Energy) partners with a PE sponsor to acquire and integrate a specialty player.


The recent transaction record is instructive. Vertiv/PurgeRite (2023, $1B): direct-to-chip services roll-up. Eaton/Boyd Thermal (2024, $9.5B): thermal + power integration. Schneider/Motivair (2025, $1.7B): thermal + power integration. TDK/Ferric (2025, approximately $250M): IVR on-package integration. Each transaction has a specific commercial DD lens applicable. The DD series applied to these transactions produces the analytical framework for the next wave.



07Public equity implications: layer-specific screening


Sector-level screening of "data centre power equipment" misses the layer-specific reality. A vendor with 60 percent 48V exposure and 40 percent 800V exposure is a different investment thesis from a vendor with 20 percent 48V exposure and 80 percent 800V exposure, even if their trailing revenue looks similar. Public equity analysts covering the space need to build layer-specific revenue mix models to price the transition correctly.



Chart 5. Named public-equity screening matrix by layer exposure, mid-2026


Vertiv, Eaton, Schneider. High 800 VDC exposure, positioned for transition upside. Delta Electronics. Mixed 48V/800V exposure, execution-dependent. Munters, STULZ. High air-cooling exposure, retrofit-cycle risk. Wolfspeed, Infineon. WBG-content leaders with premium 800 VDC multiplier. Menlo Micro. Pure-play SSCB, early-stage upside with commensurate execution risk.



Layer exposure estimates from vendor segment reporting and analyst decomposition 2024-2026.




08What breaks the vendor economics thesis


Three risks could break the vendor-economics thesis laid out above. The first is a technology alternative that displaces 800 VDC. For example, direct 3-phase medium-voltage delivery all the way to the rack, which some hyperscaler internal roadmaps have discussed. The probability is low through 2028; the technology is not commercially mature. The second is a certification crisis: an insurance-industry response to a documented 800 VDC incident that reprices premiums and delays deployment 12-18 months. The probability is moderate; incidents are rare but consequential. The third is a hyperscaler demand shock that flatlines the vendor backlog. A broad AI-capex pullback that reduces demand for 800 VDC infrastructure by 30-40 percent. The probability is genuinely uncertain and depends on macro dynamics beyond the vendor set.



Chart 6. Vendor economics thesis: three risks and their probability


Technology-alternative risk: low probability through 2028, high impact if realised. Insurance/certification crisis: moderate probability, moderate impact. Hyperscaler demand shock: uncertain probability, high impact. Investors sizing exposure should consider all three simultaneously rather than any one in isolation.



Risk assessment is the author's synthesis; specific probability estimates require named-catalyst tracking.




09Investor implications


 1. Screen by layer, not by sector. Vendor revenue mix across the six layers determines transition upside more than sector-level positioning.
 2. Overweight WBG semiconductor content winners. Wolfspeed, Infineon SiC, ONsemi Vcore, STMicro, ROHM all benefit from 3-4x content expansion per rack.
 3. Overweight the integrated thermal + power vendors. Vertiv, Eaton, Schneider each capture 2-3x content-dollar per rack at 800 VDC.
 4. Underweight 48V-specialist incumbents that have not repositioned. Named revenue-at-risk of 20-40 percent over 2027-2030.
 5. Track M&A activity in the specialty segments (direct-to-chip cooling, SSCB, IVR, HSC). Roll-up potential remains, and each transaction reprices the segment.
 6. Model the execution-risk window Q3 2026 to Q4 2027 for the three integrated winners. Successful execution delivers thesis; execution failure creates re-rating opportunity.
 7. Watch new-entrant pure-plays (Menlo Micro for SSCB, Skeleton for HSC, Submer for immersion). Small revenue base + rapid growth = optionality upside.
 8. Price the three thesis-breaking risks (technology alternative, certification crisis, demand shock) in any long-position sizing decision.



10The reframe for anyone investing now


The 800 VDC transition is a vendor-set reshuffle at layer-specific scale. Sector-level "data centre power equipment" investment thinking misses which vendors capture the upside and which lose. The layer-specific framework applied via the AI Power Chain vendor screen and this essay produces a defensible screening approach. Track record on execution over the next 18 months (Q3 2026 through Q4 2027) will separate the vendors that hit the thesis from the vendors that miss it.


Part XII moves into commissioning and skills. The operational readiness workstream that determines whether the vendor-supplied equipment actually deploys successfully at the operator site.




Glossary of terms used



GaNGallium Nitride. Wide-bandgap semiconductor material used in RF and power applications.
HSCHybrid Supercapacitor. Energy storage technology bridging batteries and traditional capacitors.
HVDCHigh-Voltage Direct Current. DC transmission at hundreds of kilovolts.
IGBTInsulated Gate Bipolar Transistor. Power semiconductor used in inverters, converters, and active filters.
IVRIntegrated Voltage Regulator. On-package power delivery component for modern CPUs and GPUs.
M&AMergers and Acquisitions. Transaction shorthand covering the deal universe.
MLCCMulti-Layer Ceramic Capacitor. High-volume capacitor technology used across electronics.
MVMedium Voltage. Typically 1 kV to 35 kV.
PEPrivate Equity. Investment class covering leveraged buyouts and growth-stage transactions.
SiCSilicon Carbide. Wide-bandgap semiconductor material used in high-voltage power electronics.
SSCBSolid State Circuit Breaker. Semiconductor-based protective device.
VRMVoltage Regulator Module. Component that converts one DC voltage to another at the point of load.
WBGWide-Bandgap. Semiconductor material class (SiC and GaN) used in high-efficiency power electronics.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.



Related topic hubs
 * 800V Data Centre Power
 * AI Power Semiconductors
 * Browse all topic hubs →


Method and sources. Public information only. Vendor positioning from public investor materials, segment revenue reporting, and analyst decomposition 2024-2026. Named transactions from Vertiv Q4 2023, Eaton Q2 2024, Schneider Q1 2025, TDK 2025 acquisition disclosures. WBG semiconductor content estimates from Wolfspeed, Infineon, ONsemi, STMicro investor days and supply chain analysis. New-entrant valuation data from Crunchbase, PitchBook, and company communications. Layer exposure estimates are the author's synthesis. No advisory relationship with any named party.


Series footer. Part XI in The DC-DC TransitionRelated reading: Part III on the six-layer architecture map, Part VIII on cooling co-emergence (vendor consolidation dynamics), Part X on standards (certification cost as market-entry barrier), Part XIII on the ten-year view. Companion context: The AI Power Chain Vendor Screen (six-layer basis), The Investment Layer Part IV on capital flows, Deal Tear-Down and Deal Watch series for real transaction analysis.


Written in a personal capacity. No advisory conflict on any named party. Nothing here is investment advice.











The DC-DC Transition: start with the foundation

 1. The real reason data centres are going DC (four forces)
 2. Two architectures wearing the same name (800V DC vs OCP Mt Diablo)
 3. The six-layer architecture map


==============================================================================
# Commissioning, skills, and operational readiness at 800 VDC (DC-DC Transition XII)
URL: https://adikumar.co/dc-dc-transition-12-commissioning-skills/
Published: 2026-08-20
Summary: AC-to-DC skills gap quantified. Commissioning timeline extension. Vendor academies + IBEW apprenticeship + university pipelines.
==============================================================================

The DC-DC Transition series · Part 12 of 19









The DC-DC Transition · Part XII of XIII



Commissioning, skills, and operational readiness at 800 VDC


The 800 VDC transition creates a skills gap the industry has under-planned for. Electricians and data centre operators trained on AC systems need retraining on DC-specific safety, protection, and monitoring. Commissioning cycles run 40-60 percent longer than AC baseline. Operators that plan the skills pipeline at Stage 1 close the gap; operators that treat it as Stage 4 hiring slip go-live by months.


Part XII in the DC-DC Transition series · independent analysis · no advisory conflict on any named party





The DC-DC Transition · thirteen essays for data centre power architects

 1.  I. The real reason data centres are going DC
 2.  II. Two architectures wearing the same name
 3.  III. The architecture map
 4.  IV. Arc behaviour + insurance
 5.  V. Grounding + ground-fault protection
 6.  VI. Battery integration at 800 VDC
 7.  VII. Retrofit vs greenfield
 8.  VIII. 800 VDC and liquid cooling co-emergence
 9.  IX. Power quality + grid interaction
 10. X. Standards: OCP, IEC, NEC, IEEE
 11. XI. Vendor economics: who wins the transition
 12. XII. Commissioning, skills, operational readiness (you are here)
 13. XIII. The ten-year view





01The AC-to-DC skills gap, quantified



An electrician trained on AC data centre work has a specific skills profile: AC voltage class up to 480V or 600V, understanding of zero-crossing protection behaviour, familiarity with three-phase power distribution, safety training focused on AC arc-flash under NFPA 70E. That skills profile transfers only partially to 800 VDC work. The electrician needs additional training on DC voltage class up to 1500V, DC-specific arc behaviour (no zero crossing, see Part IV), grounding topology differences (Part V), battery integration safety (Part VI), and the specific protective devices used in DC applications (SSCBs, current-limiting fuses, insulation-resistance monitors).



The retraining cost is documented across industrial DC applications historically. A journeyman AC electrician requires roughly 80-120 hours of additional training and 6-12 months of supervised work at DC voltages before being qualified for independent DC work at 800V. Vendor training programmes (Vertiv Academy, Schneider Electric University, Eaton Electrical University) supply portions of this training; utility apprenticeship programmes are being extended for DC; some community colleges are adding DC data centre programmes to their electrical curricula.




Chart 1. AC electrician retraining hours + cost per person for DC data centre qualification


Retraining runs 80-120 hours across four modules: DC arc behaviour, DC grounding topology, DC battery integration, DC protective devices. Cost per person including training time + supervised work + certification lands roughly $8-15K per electrician. Multiply by facility electrician headcount for the total.



Retraining hour estimates from IBEW electrician apprenticeship extensions, Vertiv Academy course catalogues, Schneider Electric University programme documentation 2024-2026.





02Safety training requirements at 800 VDC



NFPA 70E. The electrical safety in the workplace standard. Has to be applied with 800 VDC-specific interpretation. The PPE category depends on the incident-energy exposure (see Part IV) which depends on the arc-fault protection topology which depends on the vendor equipment selection. Everything in the safety-training chain traces back to the Stage-1 architecture decision.



Category-2 PPE is the operational baseline for normal-operations work with appropriately-designed 800 VDC facilities. Category-3 or higher requires more restrictive PPE, longer donning procedures, and specialised training. Facilities designed against tier-1 arc-fault protection (mechanical breakers only) often require Category-3 PPE for routine work, which raises operational cost substantially.




03Commissioning timeline extension



The commissioning phase. Where electrical systems are energised, tested, and turned over to operations. Runs longer for 800 VDC facilities than for equivalent AC facilities. Three specific extensions add up: additional insulation-resistance testing at the DC voltage class (extended compared to AC insulation testing), arc-fault protection verification against specific fault scenarios modelled at Stage 1 (Part IV), and battery commissioning at 800 V series stacks (Part VI) that involves cell-by-cell verification of the 220+ cells in the stack.




Chart 2. Commissioning timeline AC vs 800 VDC facility, 200 MW installed capacity


AC commissioning for a 200 MW facility runs 12-16 weeks from energisation to operations turnover. Equivalent 800 VDC commissioning runs 18-24 weeks. A 40-60 percent extension. The delta is dominated by DC-specific testing (insulation resistance, arc-fault verification) and battery commissioning at high voltage.



Timeline estimates from vendor commissioning documentation and hyperscaler-published commissioning cycle data 2024-2026.





04Available training pipelines



The training pipeline for DC data centre electricians has three main sources. Vendor academies provide product-specific training tied to their own equipment. Vertiv Academy, Schneider Electric University, Eaton Electrical University, ABB University, Delta Academy each supply courses. Professional certifications (IEEE, NFPA 70E specific to DC, IBEW extended apprenticeships) provide vendor-agnostic training. University and community-college programmes (Northeastern, Purdue, RIT, and a growing set of technical colleges) are beginning to offer DC data centre curricula.




Chart 3. Training pipeline capacity vs data centre electrician demand


Vendor academies supply roughly 40 percent of retrained electricians; professional certifications 30 percent; university/college programmes 15 percent; on-the-job supervised transition 15 percent. Total pipeline capacity roughly 8,000-12,000 retrained electricians per year in North America, against estimated demand of 15,000-20,000 by 2028.



Pipeline capacity and demand estimates from IBEW apprenticeship data, vendor academy enrolment statistics, and hyperscaler-published hiring plans 2024-2026.




The gap between pipeline capacity and demand is the operational reality operators need to plan against. Facilities coming online 2027-2028 that assume a spot-market for DC-qualified electricians will discover the shortage during commissioning. Facilities planning at Stage 1 for their own retraining pipeline (vendor academy contracts, IBEW apprenticeship partnerships, tuition support for community-college programmes) close the gap on schedule.




05Operational monitoring differences from AC



Operational monitoring at 800 VDC introduces new signal categories that AC operations do not have. Insulation-resistance trends (Part V) require operator understanding of what a slow degradation vs a sudden drop indicates. Arc-fault detection sensor outputs (Part IV) require operator response protocols for developing faults. Battery-stack cell-voltage monitoring (Part VI) produces alarms that AC operators would not have seen. Thermal monitoring on liquid-cooled loops (Part VIII) integrates with the electrical monitoring in ways new to the operations team.



The training-and-response gap is the operative variable. A well-trained DC operator recognises the pattern of a developing insulation fault (resistance slowly declining over days or weeks) and schedules corrective action before an alarm event. An operator without that training either misses the pattern entirely or treats every alarm as an emergency. The difference is in operational availability metrics.




06Remediation playbook for DC-specific faults



A remediation playbook for DC-specific fault events has to be authored at Stage 1 and rehearsed before go-live. Six event categories dominate. First, insulation-resistance decline. Schedule inspection, identify affected panel, isolate for repair. Second, arc-fault detection alarm. Immediate isolation, incident review, PPE-required inspection. Third, battery-stack cell imbalance. Isolate string, verify BMS diagnostics, replace or rebalance. Fourth, coolant leak with electrical proximity. Immediate coolant loop isolation, drip-pan verification, arc-fault protection status check. Fifth, ground-fault protection trip. Verify legitimate fault, review coordination settings if nuisance, restore per procedure. Sixth, utility-side power quality event. Evaluate PCC waveform, coordinate with utility engineering, adjust filter or PFC as needed.



Each event has a specific procedure that requires DC-trained personnel. Operators without that training defer to vendor emergency response, which is slow and expensive.




07Stage-1 planning to close the skills gap by go-live



The specific Stage-1 planning decisions that close the skills gap by go-live are documented enough at this point that operators can pattern-match. The 200-hour commitment to a vendor academy for 15-30 electricians, the six-month lead time on IBEW apprenticeship enrolment, the 12-month lead time on a community-college programme partnership, the 3-month training block for existing operators on 800 VDC monitoring. All of these need to be committed at Stage 1 to land by commissioning at year-end 2026 or Q1 2027.




Chart 4. Stage-1 skills planning: what to commit when for on-time go-live


Vendor academy contract at Stage 1. IBEW apprenticeship enrolment at Stage 1 for delivery by Stage 3. Community-college partnership at Stage 1 for delivery by Stage 4. Operator monitoring training at Stage 2 for delivery by Stage 4. Certification testing at Stage 4 for delivery at commissioning.



Stage-1 to commissioning timeline synthesis; specific dates depend on the project schedule.





08Named training programme references





ProgrammeProviderDurationFocus


Vertiv Academy. DC Data Centre courseVertiv40-80 hoursVendor-specific 800 VDC equipment; certification on Liebert lines
Schneider Electric University. Data Center DCSchneider Electric60-120 hoursGalaxy VXL, BATTERY MODULAR CX, DC-specific safety
Eaton Electrical University. DC Data CenterEaton40-80 hours9395P series, DC-native product lines, arc-fault protection
ABB University. DC DistributionABB60-100 hoursFormulaDC, DC breakers, industrial-heritage safety practices
Delta Academy. DC PowerDelta Electronics40-60 hoursUltron DPS, Modulon DPH, active front-end design
IBEW Data Center DC ApprenticeshipIBEW national + local chapters2-4 year apprenticeship with DC data centre trackFull journeyman qualification with DC specialisation
NFPA 70E DC-specific update courseNFPA16-24 hoursSafety practices for DC electrical work
Northeastern University. DC Data Center CertificateNortheastern University1-semester certificateAcademic depth for design engineers
Purdue Data Center DC TrackPurdue UniversityFull 2-year technical degreeAcademic depth + hands-on lab
RIT DC Power Systems CertificateRochester Institute of Technology6-month certificatePractitioner-focused






09The operator planning checklist



 1. Skills inventory of current facility electricians documented at Stage 1. Base against required 800 VDC qualifications.
 2. Retraining plan committed at Stage 1 with named vendor academy or IBEW programme selected.
 3. NFPA 70E DC-specific safety training scheduled for all electrical workers. 16-24 hours per person, before energisation.
 4. Commissioning timeline planned at 40-60 percent extension over AC baseline. Not the same duration; substantially longer.
 5. Operational monitoring training for existing operators. 3-month block, before go-live.
 6. Remediation playbook authored and rehearsed. Six event categories minimum; simulated drills.
 7. Community-college or university partnership for ongoing pipeline. 12-month lead time; committed at Stage 1.
 8. Vendor emergency-response contract in place. Fallback while internal skills mature.




10The reframe for anyone specifying now



The 800 VDC transition is not just a technology transition; it is a workforce transition. Operators that treat skills and commissioning as Stage-4 procurement questions consistently slip go-live by 2-6 months. Operators that treat skills planning as a Stage-1 architecture-level commitment land go-live on schedule and with lower operational risk in the first year.



Part XIII, the closing essay in this series, moves into the ten-year view. Where the 800 VDC transition lands by 2028, 2030, and 2035, and what breaks the trajectory.





Glossary of terms used



BMSBattery Management System. The electronics and software controlling battery pack cell balancing, thermal management, and state-of-charge estimation.
IBEWInternational Brotherhood of Electrical Workers. US electrical trade union running apprenticeship programmes for the electrical trade.
NFPANational Fire Protection Association. US organisation publishing fire and electrical safety codes including NFPA 70E for arc-flash safety in the workplace.
PCCPoint of Common Coupling. The electrical connection point between a customer facility and the utility grid.
PFCPower Factor Correction. Circuit or system that reduces reactive power drawn from the supply.
PPEPersonal Protective Equipment. Arc-flash suits, gloves, face shields rated to specific incident-energy levels per NFPA 70E.
SSCBSolid State Circuit Breaker. Semiconductor-based protective device.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.




Related topic hubs
 * 800V Data Centre Power
 * AI Power Semiconductors
 * Browse all topic hubs →


Method and sources. Public information only. Retraining requirements from IBEW apprenticeship documentation and vendor academy course catalogues (Vertiv, Schneider, Eaton, ABB, Delta 2024-2026). NFPA 70E DC-specific update from NFPA revision cycle publications. University programme references from published curriculum information (Northeastern, Purdue, RIT). Commissioning timeline estimates from vendor commissioning documentation and hyperscaler-published cycle data. Pipeline capacity/demand estimates from IBEW national statistics and vendor academy enrolment reporting. No advisory relationship with any named party.


Series footer. Part XII in The DC-DC TransitionRelated reading: Part IV on arc-flash (safety training basis), Part V on grounding (operational monitoring differences), Part VI on battery integration (commissioning complexity), Part VII on retrofit (phased-execution skills implications). Companion context: Due Diligence for the AI Buildout Part IX on organisational / systems DD, The AI Power Chain Part VI on Modular Datacenter Stack commissioning practice.


Written in a personal capacity. No advisory conflict on any named party. Nothing here is investment advice.













The DC-DC Transition: start with the foundation

 1. The real reason data centres are going DC (four forces)
 2. Two architectures wearing the same name (800V DC vs OCP Mt Diablo)
 3. The six-layer architecture map


==============================================================================
# The ten-year view: where 800 VDC data centres land by 2028, 2030, and 2035 (DC-DC Transition XIII, series closing)
URL: https://adikumar.co/dc-dc-transition-13-ten-year-view/
Published: 2026-08-20
Summary: Series-closing synthesis. Installed-base composition forecast 2020-2035. Five risks that could break the trajectory.
==============================================================================

The DC-DC Transition series · Part 13 of 19







The DC-DC Transition · Part XIII of XIII · series closing



The ten-year view: where 800 VDC data centres land by 2028, 2030, and 2035


800 VDC becomes the default architecture for new AI infrastructure builds above 100 MW by 2028-2029. Hybrid architectures dominate below that scale through 2030. Retrofits remain rare through 2030 at 5-15 percent of installed AC base. By 2035, 1500 VDC is under active discussion for facilities above 500 MW. Vendor consolidation lands at 3-5 major integrated players per layer. Five specific risks could delay or accelerate this trajectory by 12-24 months.


Part XIII in the DC-DC Transition series · series closing · independent analysis · no advisory conflict on any named party




The DC-DC Transition · thirteen essays for data centre power architects

 1.  I. The real reason data centres are going DC
 2.  II. Two architectures wearing the same name
 3.  III. The architecture map
 4.  IV. Arc behaviour + insurance
 5.  V. Grounding + ground-fault protection
 6.  VI. Battery integration at 800 VDC
 7.  VII. Retrofit vs greenfield
 8.  VIII. 800 VDC and liquid cooling co-emergence
 9.  IX. Power quality + grid interaction
 10. X. Standards: OCP, IEC, NEC, IEEE
 11. XI. Vendor economics: who wins the transition
 12. XII. Commissioning, skills, operational readiness
 13. XIII. The ten-year view (you are here)




01Where we are in mid-2026


The starting-point snapshot for a ten-year projection has three anchor facts. Installed base in mid-2026 is dominantly 48 VDC rack distribution over AC facility-level infrastructure, with less than 5 percent of installed hyperscale AI infrastructure using native 800 VDC. Greenfield hyperscale AI infrastructure announced 2025-2026 for 2027-2028 delivery increasingly specifies 800 VDC. The NVIDIA/OpenAI PORTS-Pike campus (see Deal Watch issue on it), Microsoft Wisconsin, Google Ohio and Nevada expansions are all 800 VDC. Standards state is fragmented but converging (see Part X); vendor readiness is genuinely there (see Part XI); operator readiness is the lagging variable (see Part XII).



023-year outlook: 2027-2029


Over the three-year window through 2029, three specific things happen. New AI infrastructure builds above 100 MW convert overwhelmingly to 800 VDC. Probably 75-90 percent of announced greenfield in that scale band. Installed-base composition begins to shift but slowly. From under 5 percent DC in 2026 to approximately 15-25 percent by 2029, dominated by the greenfield additions rather than any retrofit wave. Standards convergence lands (Part X), which removes the multi-standard compliance premium for facilities specified 2029 and later. Vendor set winnows through the execution-risk window (Part XI). Vertiv, Eaton, Schneider execute their integration plans successfully or fail publicly.



Chart 1. Installed-base composition of AI-adjacent data centre power architecture 2020-2035


Native 800 VDC share of installed AI-adjacent capacity climbs from under 5 percent in 2026 to roughly 20 percent in 2029, 40 percent in 2032, and 60 percent by 2035. Hybrid architectures (partial DC delivery in AC-native facilities) fill the middle band. Pure legacy AC declines as retrofit and end-of-life dynamics play out.



Composition estimates from author's synthesis across hyperscaler capex disclosures, greenfield-vs-retrofit assumptions from Part VII, and vendor forecast data.




035-year outlook: 2030-2031


By 2030-2031, three more shifts land. Greenfield above 100 MW is essentially 100 percent 800 VDC. The market has fully priced the technology and standards have converged. Hybrid architectures (partial DC delivery to specific rack clusters in AC-native facilities) dominate the 20-100 MW range because greenfield is not always economic at that scale. Retrofit remains rare (5-15 percent of AC installed base) because the break-even analysis (Part VII) does not close for most existing facilities. Installed-base composition reaches roughly 30-40 percent DC of AI-adjacent capacity.



Chart 2. Greenfield share of new builds by facility size band 2026-2031


Greenfield builds above 100 MW go 800 VDC-first from 2027 forward, reaching ~100 percent by 2030. The 20-100 MW band adopts more slowly, reaching ~60 percent 800 VDC by 2031. Below 20 MW facilities remain largely hybrid or AC through 2031 because 800 VDC economics do not close at small scale.



Adoption trajectory from hyperscaler capex disclosures and vendor bookings data 2024-2026, projected forward.




0410-year outlook: 2035-2036


Ten years out, the 800 VDC transition is largely complete for greenfield hyperscale. Installed-base composition reaches approximately 60 percent DC across AI-adjacent data centres. Legacy AC facilities that survive are typically smaller enterprise data centres or purpose-specific niche facilities.


The interesting question at the ten-year horizon is 1500 VDC. Some hyperscaler roadmaps discussed publicly through 2024-2026 anticipate that 1500 VDC becomes viable for facilities above 500 MW as the next-generation efficiency and cost play. The technology components exist (photovoltaic industry has been operating at 1500 VDC for years, HVDC operates at hundreds of kilovolts). The specific engineering and standards work for 1500 VDC data centres is early. By 2033-2035 the first commercial-scale 1500 VDC data centre deployments are plausible.



05Five risks that could delay or accelerate the trajectory


The trajectory laid out above is the central case. Five specific risks could break or accelerate it.


Risk 1: Technology alternative displacement. Direct 3-phase medium-voltage delivery to the rack (bypassing the DC bus entirely) has been discussed in some hyperscaler internal roadmaps. Probability through 2030: low. Impact if realised: substantial (would delay 800 VDC deployment by 3-5 years).


Risk 2: Insurance/certification crisis. A documented 800 VDC arc-flash incident at hyperscale that reprices the insurance market and delays deployment. Probability: moderate. Impact: 12-24 month deployment delay.


Risk 3: Hyperscaler AI capex pullback. A broad-based reduction in AI infrastructure spend. 30-50 percent below current guidance. That reduces demand for 800 VDC infrastructure and delays vendor investment recovery. Probability: uncertain, macro-dependent. Impact: substantial.


Risk 4: Regulatory intervention on data centre load growth. Following the ERCOT Batch Zero pause (Aug 2026, see the grid-queue thesis piece), similar political interventions in PJM, MISO, CAISO could delay greenfield build. Probability: moderate. Impact: 12-18 month delay.


Risk 5: WBG semiconductor supply constraint. If SiC or GaN wafer supply becomes constrained (either through Chinese domestic prioritisation or Western-supply capacity limit), 800 VDC deployment slows. Probability: moderate through 2028. Impact: moderate.



Chart 3. Five risks: probability, impact, and directionality


Technology alternative displacement is low-probability but high-impact if realised. Insurance/certification crisis is moderate probability with 12-24 month impact. AI capex pullback carries the highest impact potential but genuinely uncertain probability. Regulatory intervention is the leading indicator to track. WBG supply is the specific supply-chain risk.



Risk assessment is the author's synthesis; specific probability estimates require named-catalyst tracking.




06Five specific predictions with named catalysts


The following five predictions have specific 6-18 month markers that will validate or invalidate them.


 1. Q4 2026: First hyperscaler earnings call surfaces "queue-adjusted delivery" language explicitly. Prediction: at least one of the big four uses the phrase in Q3 or Q4 2026 earnings commentary. Catalyst: hyperscaler Q3/Q4 2026 earnings calls (Oct-Nov 2026, Feb 2027).
 2. H1 2027: Second three-legged financing (guarantee + equity + utility partnership) closes. Prediction: AMD, Broadcom, or Google Cloud runs a variant of the NVIDIA/OpenAI PORTS-Pike structure. Catalyst: named public announcement.
 3. H2 2027: First non-Vertiv/Eaton/Schneider vendor announces 800 VDC + liquid cooling combined offering. Prediction: Delta Electronics or Siemens Energy or another named vendor announces a competing integrated combination. Catalyst: vendor Q2 or Q3 2027 announcement.
 4. 2028: NEC 2029 cycle finalises 800 VDC data centre coverage. Prediction: the 2029 NEC edition includes comprehensive 800 VDC coverage that removes state-level variance requirements. Catalyst: NFPA 70 committee publication schedule.
 5. 2029-2030: First public 1500 VDC data centre design. Prediction: at least one hyperscaler discloses a 1500 VDC design (either constructed or planned) for a facility above 500 MW. Catalyst: hyperscaler infrastructure disclosure at an OCP Global Summit.



Chart 4. Five predictions: confidence and timeline


Prediction 1 (queue-adjusted language in earnings) has high confidence and near-term catalyst. Prediction 2 (second three-legged financing) has moderate confidence and 6-12 month catalyst. Prediction 3 (fourth integrated vendor) has moderate confidence and 12-18 month catalyst. Prediction 4 (NEC 2029) has high confidence and 24-36 month catalyst. Prediction 5 (1500 VDC design) has moderate confidence and 36-48 month catalyst.



Confidence and timeline are the author's synthesis; catalysts are named and publicly-observable.




07What operators should be planning for now


The ten-year view produces specific Stage-1 planning implications for facilities being designed now. Any greenfield above 100 MW should default to 800 VDC unless specific reasons drive otherwise. Retrofit decisions should apply the Part VII framework rigorously. Most retrofits are the wrong choice. Standards references should specify OCP Mt. Diablo current version plus IEC 60947-10 plus applicable NEC edition plus IEEE P2818 draft. Vendor selection should test for combined electrical + thermal capability plus certification status. Skills planning should be Stage-1 (Part XII).


The corpus this series has produced. Thirteen essays plus the grid-queue thesis piece, the DD series, the Investment Layer, the AI Power Chain, and the deal series. Is the reference framework operators can point their teams at.



08Series-closing synthesis


The DC-DC transition is not a single technology decision; it is a linked set of Stage-1 architecture decisions that determine facility performance, insurance premium, operational readiness, and vendor consolidation dynamics for the next decade. Essays IV through VI (protection stack: arc-fault, grounding, battery) establish the physical safety design that determines what insurance the facility can carry. Essay VII (retrofit vs greenfield) determines whether the operator captures the transition on existing footprint or new. Essay VIII (cooling co-emergence) determines whether the vendor selection produces an integrated system. Essay IX (power quality) determines whether the facility passes the interconnection study. Essay X (standards) determines what regulatory framework applies through the interim convergence period. Essay XI (vendor economics) determines which equity positions capture the upside. Essay XII (commissioning + skills) determines whether the facility actually deploys on schedule.


Every operator specifying a facility now faces the composite of these decisions. Every investor pricing the vendor set is pricing the composite. Every regulator setting a standard is participating in the composite. The individual essays cover the pieces; this closing essay pulls the composite view together and projects it forward.


The takeaway that most matters across the series is that the 800 VDC transition is deterministic in direction (the physics and economics do not admit alternatives at hyperscale) but variable in pace. The five risks above can accelerate or delay the trajectory by 12-24 months. The five predictions provide the leading-indicator markers to track how the pace is unfolding. Operators, investors, and vendors that track those markers correctly will land on the right side of the transition; those that treat it as a distant future concern will find themselves catching up in 2028-2029 to decisions that better-prepared competitors made in 2026-2027.


This closes the DC-DC Transition series (thirteen essays plus the introductory pieces). Companion series on adikumar.co: The AI Power Chain for technical companions to each of the six layers, The Investment Layer for the capital flows funding the buildout, Due Diligence for the AI Buildout for the practitioner DD framework, and the Deal Tear-Down and Deal Watch series for the framework applied to named transactions. The What Changed monthly recap covers the running commentary on how the trajectory is unfolding month to month.




Glossary of terms used



800 VDC800 Volts Direct Current. The emerging standard voltage class for AI-scale data centre power distribution.
1500 VDC1500 Volts Direct Current. Next-generation voltage class being discussed for very-large-scale (500 MW+) data centre facilities in the 2030s.
DCDirect Current. Electrical current flowing continuously in one direction.
GaNGallium Nitride. Wide-bandgap semiconductor material used in RF and power applications.
IECInternational Electrotechnical Commission. Global standards body for electrical and electronic technologies.
IEEEInstitute of Electrical and Electronics Engineers. Global professional association publishing power and communications standards.
NECNational Electrical Code. US electrical installation code published by NFPA, updated on a three-year cycle.
NFPANational Fire Protection Association.
OCPOpen Compute Project. Hyperscaler-led standards body developing open reference designs for data centre hardware.
SiCSilicon Carbide. Wide-bandgap semiconductor material used in high-voltage power electronics.
WBGWide-Bandgap. Semiconductor material class (SiC and GaN) used in high-efficiency power electronics.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.



Related topic hubs
 * 800V Data Centre Power
 * AI Power Semiconductors
 * Browse all topic hubs →


Method and sources. Public information only. Installed-base composition estimates from hyperscaler capex disclosures 2024-2026, vendor forecast data from Vertiv, Eaton, Schneider Electric investor materials. Greenfield adoption trajectory from named announced projects (NVIDIA/OpenAI PORTS-Pike, Microsoft Wisconsin, Google Ohio and Nevada). Standards convergence timeline per Part X. 1500 VDC discussion from published hyperscaler OCP working-group participation and PV industry precedent. Five predictions have named public-catalyst markers. No advisory relationship with any named party.


Series footer. Part XIII in The DC-DC Transition (series closing). Companion reading: The AI Power Chain (six technical layers), The Investment Layer (eight essays on capital flows), Due Diligence for the AI Buildout (fourteen essays plus three supplements), Deal Tear-Down and Deal Watch (monthly named-transaction analysis). Monthly What Changed recap for running commentary.


Written in a personal capacity. No advisory conflict on any named party. Nothing here is investment advice.











In brief


800V DC data centre buildout lands in 2028-2035 with three parallel architectures serving different segments. Hyperscaler greenfield leans Nvidia-style unipolar 800V. Retrofit and dual-use campuses lean OCP Mount Diablo bipolar. Legacy 415V AC continues for enterprise + colocation below 60kW/rack. Vendor consolidation shifts around each architecture.







The DC-DC Transition: start with the foundation

 1. The real reason data centres are going DC (four forces)
 2. Two architectures wearing the same name (800V DC vs OCP Mt Diablo)
 3. The six-layer architecture map


==============================================================================
# The real reason data centres are going DC (and it's not capacity alone) (1/13)
URL: https://adikumar.co/the-real-reason-data-centres-are-going-dc-and-its-not-capacity-alone/
Published: 2026-04-28
Summary: Capacity is not the real reason data centres are going DC. Four forces drive the shift: density, grid, GPU synchronicity, hyperscaler roadmap.
==============================================================================

The DC-DC Transition series · Part 1 of 19




The data centre industry has settled on a comfortable story: AI racks need too much power, and low-voltage AC simply can't carry the current required at scale. The story is true. It is also misleading.

The conversion to high-voltage DC is happening, but treating it as a single phenomenon, driven by power density alone, misses what makes the transition consequential. Four forces are driving this shift, and they operate on different timescales, with different commercial implications. Operators who understand only one of them will design facilities that are technically functional but commercially disadvantaged for the decade ahead.

This essay walks through all four. It is the first in a thirteen-part series on the architecture, economics, and competitive dynamics of the DC transition. Subsequent essays will go deeper into specific stages of the value chain, the architectural debates that will shape the next generation of AI data centres, and the operating-model questions, pricing, partnerships, build-versus-buy, that the technical conversation tends to skip.

A note on terminology before we start You will see this transition described in trade press as the move to "HVDC" high-voltage DC. That is not technically correct. Under IEC 60038, the international standard that governs voltage classification, high voltage begins at 35 kV. Anything below 1500 V on the DC side is low-voltage DC, or LVDC. The 800 VDC architecture this essay is about , the ±400 V Mt. Diablo specification, and even the 1500 V architectures coming next. All sit firmly inside the LVDC envelope. The trade press calls it HVDC because compared to the 12 V and 48 V conventions data centres have used for decades, 800 V feels high. But the standards body has a different view, and so do the regulatory and certification frameworks that govern equipment design. This is more than pedantry. The boundary between LVDC and MVDC at 1500 V is a genuine regulatory cliff that shapes which equipment classes operators can deploy through 2030 and beyond. We will return to it in section 7. For the rest of this essay and the rest of the series, I will use 800 VDC and LVDC where they are the technically correct terms. DC transition is the framing for the broader shift across the industry.


1. The physics: copper's inconvenient truth

The most straightforward force is the one everyone names. Power equals voltage times current. At fixed power, doubling voltage halves current. Halving current roughly halves the copper required to carry it without overheating.

For traditional racks at 30 kW, this trade-off is academic. For 200 kW racks (Blackwell-class) it becomes interesting. For 1 MW racks (the Vera Rubin Ultra envelope) it becomes binary. At 1 MW per rack and 54 V, today's in-rack standard, the busbar physically does not fit inside the rack. Not "is inefficient." Not "is expensive." Does not fit.

Voltage doubling roughly halves the copper. At 800 V, the same 1 MW delivers through a busbar that's a fraction of the size and weight, and crucially, fits inside the rack envelope.

This is the force every press release names, and it's real. But if it were the whole story, we would simply move from 54 V to whatever voltage carries 1 MW conveniently, declare victory, and go home. The other three forces explain why we're going much further.


2. The conversion chain: stages compound

Today's AC architecture for an AI data centre carries through roughly five conversion stages between the grid and the GPU core: medium-voltage AC step-down to low-voltage AC, double-conversion AC UPS for power conditioning and ride-through, AC distribution to the rack, AC-to-DC conversion at the rack-mount power shelf, and a final point-of-load conversion at each accelerator.

Each stage costs efficiency. Best-in-class equipment runs 96-98 percent efficiency per stage; less-than-best equipment runs 93-95 percent. Compound those across five stages and end-to-end efficiency lands somewhere around 84 percent. For a 100 MW facility, the 16 percent loss is 16 MW of waste heat that has to be cooled, plus 16 MW of grid power that has to be procured, paid for, and accounted for in carbon footprint.

The 800 VDC architecture compresses this chain. A single perimeter conversion takes medium-voltage AC to 800 VDC at the building edge. A single rack-side conversion steps to the 50 V class for delivery to the GPU. A point-of-load conversion at the accelerator finishes the job. Three stages, not five. End-to-end efficiency moves from 84 to roughly 93 percent.

The efficiency story is the visible one. The hidden one is failure modes. Each conversion stage is a piece of equipment that can fail, contains capacitors that age, requires service contracts, has firmware that needs updating, occupies floor space that costs money. Eliminating two stages eliminates two equipment categories from the operator's reliability spreadsheet. The capex savings get the headlines; the opex savings, over a 15-year facility life, are larger.


3. The roadmap: the procurement calendar runs ahead

The third force is the one that moves the timetable. NVIDIA's GPU roadmap is public, and it is brutal:

 * Hopper-class racks: ~45 kW
 * Blackwell-class: ~120 kW
 * Blackwell Ultra: ~200 kW
 * Vera Rubin: ~600 kW
 * Vera Rubin Ultra (Kyber): ~1,000 kW
 * Next generation (projected): ~1,500 kW

The trap operators tend to fall into is reading this as a forecast when it operates as a backlog. Each generation's procurement decisions are made roughly 12 to 18 months before energization. A facility coming online in 2027 with Vera Rubin Ultra-class racks has its design locked and procurement committed now, in 2026.

That means the question facing every greenfield AI data centre operator in 2026 is not "should we build for today's power density?" It is: which voltage architecture do we commit to for the silicon we will deploy in 2027 and replace in 2028? Conservatism, building for 100 kW racks because that is what is installed today, looks cautious on the day of the design decision and commits the operator to two retrofit cycles inside the next four years.

The procurement calendar has eaten the deliberation timeline. Operators who don't see this end up buying decisions they didn't know they were making.


4. The workload: AI training breaks the assumptions

The fourth force is the most subtle and the one I find most interesting from an operating-model perspective.

The power-quality assumptions baked into the AC architecture were designed for the relatively steady load, gentle hour-over-hour variation, and predictable diurnal demand patterns of traditional cloud workloads. AC UPS topologies, rotary backup systems, and the control logic that governs them assume the load is well-behaved.

AI training workloads are not well-behaved. A large training job synchronizes thousands of GPUs to step in lockstep through forward and backward passes. The aggregate power signature shows step-changes between near-idle and near-peak, repeated thousands of times per hour, with rise and fall times in the tens of milliseconds. From the grid's perspective, an AI training facility looks like a constantly-pulsing load, and a single such facility at hyperscale is now meaningfully visible on the local grid frequency.

Traditional AC UPS architectures handle this poorly. Voltage transients on the grid side, control-loop oscillation, inverter stress, and battery cycling all degrade in the face of high-frequency synchronized swings. Operators are responding with bigger UPS systems, more aggressive grid filtering, and increasingly with on-site battery energy storage (BESS) sized to absorb the worst of the swings before they propagate to the utility connection.

DC architectures handle volatile loads more naturally. The intermediate DC bus acts as a low-impedance reservoir between the grid and the load. Battery storage, the BESS, couples to the DC bus directly, without inverter stages between storage and the demand it's smoothing. The 800 VDC architecture isn't only delivering power more efficiently; it's delivering power that the silicon's volatility doesn't disturb.

This is also why utilities are starting to express opinions. A 500 MW AC-coupled AI factory looks different on the grid than a 500 MW DC-coupled facility with on-site BESS. The grid economics, interconnection, demand charges, ancillary services revenue, are not the same. The architectural choice affects the utility relationship, and through it, the cost of power for the next twenty years.


5. What do the four forces mean together?

Each force, taken alone, is a reason to consider the transition. Taken together, they describe a structural inflection rather than an incremental upgrade.

 * The physics force sets the floor: above ~200 kW per rack, the legacy in-rack voltage cannot continue.
 * The conversion-chain force sets the medium-term economics: each stage you eliminate is efficiency, capex, floor space, and a failure mode you stop paying for.
 * The roadmap force sets the timing: if you are designing a 2027 facility, the architectural decision has to be made now.
 * The workload force sets the operating envelope: AI's volatility makes DC architectures structurally better at coupling to the grid through BESS.

The mistake I see most often, in conversations with operators, is treating this as an "AI infrastructure" question rather than as the largest power-architecture transition the data centre industry has run since the move from rotary UPS to static UPS in the 1990s. It is the latter.


6. A note on what this means for everyone who is not training frontier models

The four forces above describe what's happening at the leading edge: hyperscalers, AI-specialist neoclouds, the Top 10 AI factory builders. For the rest of the data centre industry, the vast colocation base, enterprise on-prem, mid-market, the question is more nuanced.

A 30-rack colocation hall serving general-purpose enterprise workloads at 10-20 kW per rack does not need 800 VDC. It probably never will. A regional colocation operator with mixed tenant types may need to plan for hybrid facilities where part of the floor is built to legacy specifications and part is built to the new envelope. The challenge for these operators is not the technical decision; it is the commercial decision about which tenants you intend to serve in 2030.

The supplier ecosystem reorganizing around DC Transition will eventually pull mid-market operators forward, but the timeline is much longer. Operators in this segment have time. They should be using it to think hard about which side of the eventual market they want to be on, and to make supplier choices with that in mind.


7. What is the 1500V ceiling, and what comes next?

One question I get often is whether 800 V is the destination or just a waypoint. The honest answer is the latter.

The IEC 60038 standard defines low-voltage DC (LVDC) as DC up to 1500 V. Below this threshold, equipment certification, building codes, and electrical safety regulations operate under one set of rules. Above 1500 V, you cross into medium-voltage DC (MVDC), which has different requirements, different equipment classes, and different regulatory regimes.

The practical envelope for the next five-to-seven years sits between 400 V and 1500 V. Today's reference architectures cluster at 400 V and 800 V because component supply chains support those classes today. The next move, towards 1500 V, is technically straightforward but commercially gated by UL857 Edition 15 (in development), corresponding IEC standards, and the volume ramp of components rated for the higher class. Realistic timeline: 2028 and beyond.

MVDC for data centres is a 2030s conversation, and a different one, with a different regulatory regime, different equipment, and different siting implications. We will get there, but not soon.


8. The four forces, summarized

If you remember nothing else from this essay, remember this:

The DC transition in AI data centres is being driven by four forces, not one. Power density is necessary but not sufficient as an explanation. Conversion-chain economics, the procurement calendar, and AI workload volatility are each independently sufficient reasons to make the move. And together they make the transition not a question of "if" but of "which architecture, on what timeline, with which suppliers."

Operators who understand only the first force will design facilities that are technically functional but commercially disadvantaged. Operators who understand all four will be the ones positioned to capture the next decade of this market.


9. What's next in this series

This essay is the first of thirteen. The next essay, Two Architectures Wearing The Same Name. Examines the two reference designs that dominate today's conversation: NVIDIA's centralized 800 VDC and OCP Mt. Diablo's bipolar ±400 VDC. They share a name. They share almost nothing else.

Subsequent essays will go deeper into specific stages of the power chain, the safety implications of DC arc behaviour, the supply-chain dynamics, and the operating-model questions that determine which players capture value as this transition compounds.

If you want to follow along, subscribe to the newsletter to get each essay as it publishes.



Reference

Acronyms used in this essay (DC, MV, HVDC, PUE, and others) are defined in the sitewide Adi Kumar corpus glossary.




Series expansion · six new supplements


The DC-DC Transition series now includes six deep-dive supplements addressing specific layer-level questions that the main 13-essay series flagged but did not fully develop. Consulting-grade visuals throughout. Six-layer stack decompositions, vendor concentration matrices, procurement flow diagrams:


 1. Supplement A. The Transition Tax: layer-by-layer capex decomposition. Where the 800V DC premium accumulates. Retrofit vs greenfield.
 2. Supplement B. Hybrid Forever: why most facilities will run mixed AC/DC through 2035+. Hyperscaler zone strategies compared.
 3. Supplement C. The Voltage Ceiling: 1500V DC and the MVDC question. Standards + vendor readiness by voltage class.
 4. Supplement D. The Solid-State Transformer: Layer 2 supplier dynamics. Western vs Chinese vendor landscape.
 5. Supplement E. The Sidecar Power Rack: Layer 4 architecture. OCP Mt. Diablo v0.7.0 reference. 80+ ecosystem partners.
 6. Supplement F. The Connector Wars: Layer 4/5 interface engineering. High-current connector vendor concentration + failure modes.


Full series index: /tag/dc-transition/All supplements now live.





Further reading + sources
Related essays

 * The Architecture Map. the framework that locates each argument by layer
 * Supplement A. The Transition Tax. the layer-by-layer capex accounting
 * Supplement C. The Voltage Ceiling. why 800V DC is not the endpoint


Primary sources

 * OCP Rack + Power. reference architecture standardisation
 * DOE data-centre energy programme. the efficiency drivers this essay engages with







In brief


Capacity is not the reason data centres are going DC. Four forces are: (1) density economics, 800V DC halves copper mass at rack; (2) grid decoupling, DC bus absorbs pulsed loads; (3) GPU synchronicity, DC handles hyperscale load-step better than AC; (4) hyperscaler roadmap alignment, Nvidia + OCP both spec DC.








Frequently asked


Anyone running 800V DC in production at scale (not pilot projects)?

Yes. Microsoft, Meta and Google have production 800V DC deployments (via OCP Mount Diablo 400 spec, published March 2026). Nvidia's reference designs also ship at scale with hyperscaler-integrator partners. These systems have moved past the pilot stage into production-serving racks at hyperscale campuses in Northern Virginia, Ohio, and Ireland.


How hard is the transition to 800V DC for electricians used to AC systems?

The physical transition is manageable, the equipment is bolt-in, but the safety, grounding and fault-clearing paradigms are different from AC and require dedicated training. IEC 63369 and the NEC 2029 revision will formalise the commissioning skillset. Expect a 6-12 month upskilling cycle for a competent AC data-centre electrician team.


What is the TCO difference between 800V DC power shelves versus 415V AC UPS?

For a 100 MW facility: 800V DC saves approximately 5-9% end-to-end losses vs. 415V AC-UPS + rack-level conversion. On $30-50M/year utility bill for a 100 MW site, that's $1.5-4.5M/year in energy savings. Capex is 10-20% higher for 800V DC shelves but offset by fewer conversion stages and smaller UPS + battery bank.


What are the four forces driving the DC transition beyond capacity?

The four forces are: 1) Density economics, 800V DC halves copper mass at rack; 2) Grid constraint pass-through, DC bus decouples grid volatility from rack; 3) GPU synchronicity, DC handles pulsed loads without UPS conversion churn; 4) Hyperscaler roadmap alignment, Nvidia + OCP both spec DC. Capacity growth alone would not have driven the shift.


==============================================================================
# Two architectures wearing the same name (and why hyperscalers built their own) (2/13)
URL: https://adikumar.co/two-architectures-wearing-the-same-name/
Published: 2026-05-01
Summary: Two 800V DC architectures wear the same name and behave differently in retrofit vs greenfield. Which one hyperscalers are picking, and why.
==============================================================================

The DC-DC Transition series · Part 2 of 19




Update: Since this essay was published, the picture has moved, and in a direction worth recording. The first-generation divergence it describes NVIDIA's 800 V unipolar reference and OCP Mt. Diablo's ±400 V bipolar default was real, and it traces to a shared starting point: both camps began at the rack, and both selected their voltages to inherit the existing EV supply chain rather than build a new one. Where they initially split was topology, not intent. What changed the trajectory was the entry of the electrical-equipment manufacturers - Designing to these architectures at volume, the cost of a fragmented ecosystem would became concrete: duplicated R&D across incompatible designs, a thinner supplier pool behind each component, and a slower, more expensive path to the safety standards the whole category needs.

The response has been coordination rather than a winner-takes-all standards fight. IT players and electrical OEMs have converged in the OCP 800VDC subcommittee and engaged with with standards bodies, with the explicit goals of underwriting a safe transition, keeping a diversified supplier base behind every "brick" of the stack, and sparing manufacturers the burden of building to radically different specifications. That is a better outcome than the divergence the essay captured at a single point in time, and it is worth saying so. Original essay follows

⁠ ⁠The trade press writes about "the move to 800 VDC" as if it's one thing.

There are at least two competing architectures wearing that label right now, and they don't look much alike. NVIDIA's reference design rectifies at row scale and hands a locked 800 VDC bus to a vertically integrated compute rack. The Open Compute Project's Mt. Diablo specification, authored by Meta, Google, and Microsoft, rectifies in a sidecar power rack adjacent to the IT rack, supports a configurable output (bipolar ±400 V or unipolar 800 V), and accepts any accelerator that complies with the spec. Both deliver 800 V of usable potential. The similarities end there.

This essay is about why that distinction matters more than it looks, and what it tells us about the shape of the data centre power industry over the next five years. Worth saying up front: at scale, most hyperscalers will operate both architectures inside the same estate, not as a hedge, but because their workload mix and silicon mix demand it. The choice isn't binary; the primary architecture choice is what shapes everything else.


A note on terminology, continued

Essay 1 made the case that calling this transition "HVDC" is technically wrong under IEC 60038. I'll use 800 VDC for the unipolar architecture, ±400 VDC or Mt. Diablo for the bipolar one, and the DC transition when referring to the broader industry shift. All of these sit firmly inside the LVDC envelope. The trade press will keep calling it HVDC. We will keep being correct.


1. What are the two 800V DC architectures?, sketched

The cleanest way to see the difference is to draw both at the same scale, with the same components, and look at what's structurally different.

NVIDIA 800 VDC reference (row-scale rectification). Medium-voltage AC arrives from the utility, hits a solid-state transformer at the building perimeter or at the row level, and emerges as a single bus at +800 V to ground. That bus runs along the row to each compute rack, where compact step-down converters deliver the 50 V class to the GPU. The compute and power conversion are designed as one vertically integrated system: the Kyber rack, designed for the Vera Rubin Ultra accelerator, expects 800 VDC at the rack inlet and won't accept anything else.

OCP Mt. Diablo (sidecar rectification, configurable output). Same medium-voltage AC input, but rectification happens in a dedicated power rack sitting next to one or more IT racks rather than at row scale. The IT racks themselves contain no PSUs, every rack unit is available for accelerators and scale-up switching. The Diablo 400 specification permits two output configurations from the sidecar: a three-wire bipolar arrangement at +400 V, 0 V (common), and −400 V, or a two-wire unipolar arrangement at 800 V and return. The IT racks are ORv3 HPR (Open Rack v3 High Power Rack) and accept any accelerator that complies with the spec, NVIDIA's parts work, but so do Trainium, MTIA, Maia, and AMD's MI series.

The defining contrast is where rectification happens, sidecar vs row, and whether compute is locked to a single accelerator family. The voltage-class debate (bipolar vs unipolar) sits inside Mt. Diablo as a configuration choice, not as the architectural fault line.

Same nominal usable voltage. Different rectification topology. Different stress to ground in the bipolar case. Different supply chain. Different commercial intent.


2. The dimensions that matter

Past the marketing, there are six dimensions where the architectures actually diverge.









Dimension
NVIDIA 800 VDC reference
OCP Mt. Diablo




Rectification location
Row-scale (perimeter or row-level SST)
Sidecar (rack-adjacent power rack)


Default rack output
800 V unipolar
±400 V bipolar (800 V unipolar permitted)


Conductor count to a rack
2 (positive + return)
3 (+400, common, −400) in default config


Supply chain leverage
Purpose-built 800 V parts
EV-derived 400 V parts (default config)


System scope
Vertically integrated AI factory
Compute and power as separable building blocks


Accelerator compatibility
NVIDIA Vera Rubin family
Any ORv3 HPR. Compliant accelerator





Figure 2.2. Architecture comparison across six dimensions.





Glossary of terms used



EV
Enterprise Value. Purchase price plus debt assumed minus cash. Total value of the business excluding capital structure.
OCP
Open Compute Project. Hyperscaler-led standards body developing open reference designs for data centre hardware.
VDC
Volts Direct Current. Used in 48 VDC and 800 VDC references to rack-level and datacentre power distribution.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.






Series expansion · six new supplements


The DC-DC Transition series now includes six deep-dive supplements addressing specific layer-level questions that the main 13-essay series flagged but did not fully develop. Consulting-grade visuals throughout. Six-layer stack decompositions, vendor concentration matrices, procurement flow diagrams:


 1. Supplement A. The Transition Tax: layer-by-layer capex decomposition. Where the 800V DC premium accumulates. Retrofit vs greenfield.
 2. Supplement B. Hybrid Forever: why most facilities will run mixed AC/DC through 2035+. Hyperscaler zone strategies compared.
 3. Supplement C. The Voltage Ceiling: 1500V DC and the MVDC question. Standards + vendor readiness by voltage class.
 4. Supplement D. The Solid-State Transformer: Layer 2 supplier dynamics. Western vs Chinese vendor landscape.
 5. Supplement E. The Sidecar Power Rack: Layer 4 architecture. OCP Mt. Diablo v0.7.0 reference. 80+ ecosystem partners.
 6. Supplement F. The Connector Wars: Layer 4/5 interface engineering. High-current connector vendor concentration + failure modes.


Full series index: /tag/dc-transition/All supplements now live.





Further reading + sources
Related essays

 * The Architecture Map. the six-layer framework this essay builds on
 * Supplement B. Hybrid Forever. why most facilities will run mixed AC/DC through 2035+
 * Supplement E. The Sidecar Power Rack. the specific Layer 4 architecture that shapes both branches


Primary sources

 * OCP Rack + Power. the specification defining sidecar-shaped Layer 4
 * Uptime Institute Global Data Center Survey. operator-side deployment pattern data







In brief


Two 800V DC architectures are both marketed under the same name. OCP Mount Diablo 400 (Microsoft/Meta/Google, March 2026) uses a ±400V bipolar rail suited to retrofit. Nvidia's Blackwell/Rubin 800V is unipolar, engineered for greenfield. They differ in grounding, fault-clearing, and vendor stack. Retrofit and greenfield decisions diverge sharply between them.




Three of those dimensions deserve real attention before we move on, because the marketing on each side oversells them.

On output voltage class. The Diablo 400 spec (v0.7.0, March 2026) explicitly permits both bipolar and unipolar output from the sidecar. So the "NVIDIA = 800 V, Mt. Diablo = ±400 V" framing in trade press is incomplete. The default and most-promoted Mt. Diablo configuration is bipolar ±400 V. That's the configuration that captures the EV supply chain argument. But operators that want unipolar 800 V output can get it within the same spec. The real contrast is sidecar vs row, not bipolar vs unipolar.

On fault behaviour. It's tempting to frame Mt. Diablo's bipolar option as a reliability win, since half the bus stays alive when one rail faults to ground. In theory, yes. In practice, most protection schemes will trip the entire bus on a single pole-to-ground fault for safety reasons, regardless of whether the surviving rail could carry partial load. The bipolar advantage is real in fault energy (each conductor sits at 400 V to ground rather than 800), but the "ride-through on one rail" framing is more aspirational than operational at this point.

On conductor count. Bipolar ±400 V needs three conductors plus careful neutral handling per rack; unipolar 800 V needs two. At 600 kW+ per rack, where the industry is heading, this is not a trivial difference for busway sizing, distribution complexity, and per-rack copper capex. More on this in §6.

The dimension that matters most for the next three years, though, isn't any of those. It's the supply chain.


3. Why ±400 V is really an EV play in disguise

The under-discussed feature of Mt. Diablo's default bipolar configuration is that 400 V is not a number Meta and Google picked for technical elegance. They picked it because there is already a planet-scale supply chain for components rated at 400 V, built by the electric vehicle industry over the past decade. The Diablo 400 specification itself names this rationale, "selecting 400 VDC as the nominal voltage leverages the supply chain established by electric vehicles, for greater economies of scale, proven quality, and more efficient manufacturing."

A precision note before the argument: most BEV traction systems still use ~400 V battery packs (Tesla, BYD volume models, most Chinese EVs). The components inside an EV are not all rated to 400 V to ground. Pack architecture, motor inverter topology, and isolation strategy all vary. But the discrete 400 V-class parts (IGBTs, SiC MOSFETs, capacitors, contactors, fuses, connectors) are produced at scale specifically because the EV industry needs them in this voltage class. Annual production of EV-grade 400 V power modules is in the millions of units across the major suppliers (Infineon, ST, Wolfspeed, ON Semi, Mitsubishi). The 400 V class as a standardized component design point really consolidated between 2015 and 2018 as second-generation EV platforms moved off bespoke voltages onto common ones. That curve has been compounding for a decade.

The 800 V class, by contrast, is much earlier. Yes, EV fast charging stations and 800 V powertrains (Porsche Taycan, Hyundai E-GMP, Lucid Air, Kia EV6) operate at this class but volumes are roughly an order of magnitude below the 400 V base, the components are different (rectifiers and chargers, not motor inverters), and the certification frameworks for stationary 800 V data centre equipment are still maturing.

The most cited recent UL milestone, ABB earning UL 98B certification for a 2 kV switch-disconnector in July 2025, is illustrative precisely because of how narrow it is. That certification was for utility-scale solar, not data centres, and as the standards press noted at the time, UL 98B itself only tests up to 1500 V; ABB's 2 kV cert was a stretch case ahead of UL writing the broader voltage range into the standard. The standards extension to genuinely cover 800 V stationary data centre topologies, with the arc, protection, and isolation requirements that come with that environment, is an active area of work, not a settled one.

So when Meta, Google, and Microsoft jointly authored Mt. Diablo with bipolar ±400 V as the default configuration, they were doing something specific: they were borrowing the EV industry's component cost curve. By 2027 they'll be deploying a power architecture whose discrete components are riding a learning curve that started consolidating a decade ago, while NVIDIA's vendors will still be on the steeper early portion of an entirely new 800 V curve.

Update: Since this essay was published, the picture has moved, and in a direction worth recording. The first-generation divergence it describes NVIDIA's 800 V unipolar reference and OCP Mt. Diablo's ±400 V bipolar default was real, and it traces to a shared starting point: both camps began at the rack, and both selected their voltages to inherit the existing EV supply chain rather than build a new one. Where they initially split was topology, not intent. What changed the trajectory was the entry of the electrical-equipment manufacturers - Designing to these architectures at volume, the cost of a fragmented ecosystem would became concrete: duplicated R&D across incompatible designs, a thinner supplier pool behind each component, and a slower, more expensive path to the safety standards the whole category needs.

The response has been coordination rather than a winner-takes-all standards fight. IT players and electrical OEMs have converged in the OCP 800VDC subcommittee and engaged with with standards bodies, with the explicit goals of underwriting a safe transition, keeping a diversified supplier base behind every "brick" of the stack, and sparing manufacturers the burden of building to radically different specifications. That is a better outcome than the divergence the essay captured at a single point in time, and it is worth saying so. Original essay follows




4. The unbundling pattern (and which layers are real)

Mt. Diablo is the visible piece of a larger pattern: hyperscalers systematically unbundling NVIDIA's vertical stack at every layer where they can.

But the framing "alternatives exist at every layer" is doing work it shouldn't. Some of those alternatives are genuinely contested with real production deployment behind them. Others are aspirational specs with limited silicon to back them. Worth being precise about which is which:

Genuinely contested (real production volume, real deployment): - Silicon: Google TPU (a decade of internal deployment, now v7 Ironwood in volume), AWS Trainium 3, Meta MTIA v3, Microsoft Maia 200. NVIDIA's data centre share is projected to drop from ~86% in FY2025 toward ~75% by end of 2026, with custom ASIC volume growing at roughly 45% CAGR through 2028. - Form factor: OCP ORv3 HPR is shipping at scale across Meta, Google, Microsoft fleets. NVIDIA reference systems still dominate neoclouds. Genuinely split. - Power: NVIDIA 800 VDC vs. Mt. Diablo. The fight this essay is about.

Partially contested (real specs, limited silicon, dependent on future tape-outs): - Networking: Google ICI is real and at scale internally, but Ultra Ethernet Consortium specs have shipped while UEC silicon volume remains modest. NVLink and InfiniBand still dominate in non-Google deployments. - Software at training: JAX and PyTorch-XLA are credible, but the developer base remains small relative to CUDA. OpenXLA matters strategically but commands a fraction of CUDA's ecosystem mindshare.

Aspirational (specs exist, production is the open question): - Software at inference: vLLM and SGLang have meaningful production deployment, but the dependence on CUDA underneath is real for any workload that doesn't have a dedicated compiler team.

The honest summary: the unbundling is real at the top of the stack (silicon, form factor, power) and partial in the middle (networking, training software). It's still aspirational at the application layer where CUDA's two-decade lead, four-million-developer ecosystem, and framework-first integration create switching costs measured in years, not quarters.


5. The hyperscaler silicon investment

The unbundling at the silicon layer specifically deserves its own look, because it's the layer that funds all the others.

NVIDIA's data centre revenue ran $115B in FY2025 and is tracking toward $194B in FY2026. That number isn't going down. It's going up. But the share of new AI compute that goes to NVIDIA is declining, and the rate of decline is the part that matters. Consensus from Bloomberg Intelligence, TrendForce, SemiAnalysis, and Counterpoint converges on three claims that should shape power architecture decisions:

 * NVIDIA's AI accelerator market share has declined from a peak of ~95% to an estimated 80-86% in 2025-2026, with most analysts forecasting further decline toward 75% by end of 2026 as custom ASIC volume scales.
 * Custom ASIC shipments are projected to grow at roughly 45% CAGR through 2028, with shipments crossing GPU shipments in 2028 per Counterpoint.
 * The displacement is most pronounced in **inference**, where custom ASICs are approaching parity in deployed FLOPs by 2027-2028. Frontier **training** remains NVIDIA-dominated and probably will through 2028, anchored by CUDA's ecosystem advantage and Vera Rubin's raw performance.

NVIDIA's revenue keeps growing in absolute terms because the AI compute pie is growing faster than the share shift. The displacement is real but asymmetric, with share down and dollars up.

Three production realities matter most for power architecture decisions:

Google TPU v7 (Ironwood) entered volume production in late 2025, with industry sources estimating roughly 4.3M units shipping in 2026 and analyst forecasts pointing to cumulative deployment in the tens of millions across 2026-2028. Anthropic specifically has committed to deploying over one million Ironwood chips, representing more than a gigawatt of compute. Ironwood is designed and deployed in ORv3 HPR-class racks with sidecar power.

AWS Trainium 3 reached general availability in December 2025 at re:Invent, built on TSMC 3nm. The Trn3 UltraServer scales to 144 chips across 36 servers in a multi-rack configuration; the volume per-rack SKU (NL32x2 Switched) is 32 chips per rack. Project Rainier, AWS's earlier deployment of 500,000+ chips with Anthropic, was Trainium 2-based; Trainium 3 is the next generation and Anthropic is among the named launch customers.

Meta MTIA v3 is in full deployment for inference workloads (ranking, recommendations, ad serving). MTIA v4 is in fabrication.

Microsoft Maia 200 runs Copilot and Azure OpenAI Services internally. Microsoft also remains one of NVIDIA's largest Vera Rubin customers. The dual-track strategy is explicit, and the implication for power architecture is that Microsoft will operate both Mt. Diablo halls (for Maia) and 800 VDC halls (for Vera Rubin) inside the same data centre estate.

When the operator of the workload also controls the silicon, the bargaining position with the rest of the stack changes.


6. So which architecture wins?

Wrong question.

The right question: which architecture wins for which kind of operator?

There's also a real engineering case for the row-scale unipolar approach that the supply-chain argument tends to bury. At 600 kW+ per rack, where Vera Rubin Ultra and similar accelerators are heading, the unipolar 800 VDC bus uses fewer conductors and less copper than the bipolar default. Mt. Diablo in its default three-wire config needs three conductors plus careful neutral handling; unipolar needs two. At extreme rack densities this matters for busway sizing, distribution complexity, and per-rack capex on conductors alone. NVIDIA's 800 VDC pitch isn't pure supply-chain naïveté. There is a real density argument underneath. The supply chain economics may dominate for the next five years; the density economics may dominate beyond that.

With that in the picture, six operator archetypes, and the architectural answer is different for each.

The hyperscaler running its own silicon (Meta, Google, Microsoft, Amazon). Mt. Diablo, with very high probability, as the primary architecture. They authored the spec, they want the EV supply chain on the bipolar default, and their compute platform is accelerator-agnostic by design. NVIDIA's row-scale 800 VDC architecture is the secondary architecture for this archetype, not the primary one. They will operate it where they host NVIDIA-tenant or Vera Rubin workloads, but their default new-build power architecture is sidecar.

The AI-specialist neocloud (CoreWeave, Lambda, Crusoe, Nebius). Almost certainly NVIDIA 800 VDC. Their entire value proposition is "fastest path to NVIDIA capacity." They don't have the volume to influence Mt. Diablo's roadmap, they don't run alternative silicon, and time-to-revenue dominates everything else. The density advantage of unipolar 800 V matters here too as Vera Rubin Ultra deployment scales.

A note on stranded capex risk for any single-architecture bet. NVIDIA's pattern across the Hopper-to-Blackwell-to-Vera Rubin transitions tells you something about the speed at which rack-level specs change. In three generations the rack jumped from ~50 kW air-cooled, to 140 kW liquid-cooled, to 600 kW liquid-cooled with 800 VDC, with each transition stranding prior facility-side investment. CBRE puts the liquid-cooling retrofit cost alone at $2-5M per rack row. If you build an 800 V Kyber row in 2026 and NVIDIA's 2029 generation shifts the voltage class again, or the mechanical interface, or the cooling spec, the building-side capex doesn't transfer cleanly. The chips amortize over 2-3 years; the infrastructure is supposed to amortize over 15-25. That mismatch is a strategic problem for any operator whose architecture choice is locked to a single vendor's roadmap. It's a particular problem for AI-specialist neoclouds because their balance sheet can't absorb the asymmetry. It's a smaller problem for hyperscalers because they have multi-tenant flexibility and silicon optionality. It's why the enterprise "wait" answer below is pattern recognition rather than conservatism.

The major colocation operator with mixed AI tenants (Equinix, Digital Realty, Vantage, NTT). Hybrid as a product, not a hedge. Different halls for different tenants. Some operators will offer Mt. Diablo halls for hyperscaler tenants, some 800 VDC halls for neocloud tenants, and a lot of legacy AC for the long tail. The commercial sophistication comes from pricing each architecture distinctly, not from picking one.

The OEM / ODM (Supermicro, Wiwynn, Foxconn, Quanta, Inventec). They build both. Watch which lines they're tooling up: the ratio of capex going to ORv3 HPR-compatible chassis versus NVIDIA reference (MGX, Kyber) lines is one of the cleanest leading indicators of which architecture is winning by volume. Today the ratio splits roughly evenly across the major ODMs; the trajectory matters more than the snapshot. The ODMs have no architectural loyalty. They go where the tonnage is.

The enterprise running on-prem AI for regulated workloads (banks, defense primes, healthcare systems). Probably neither, for the next three years. Stay AC, deploy a small DC island for the AI sub-zone, take the inefficiency hit. The skills, the certifications, and the supplier relationships aren't there yet, and the workloads (mostly inference, mostly small training runs) don't justify the architectural commitment.

The sovereign AI operator (national champions in EU, UAE, India, etc.). Depends entirely on supply chain politics. If the priority is geopolitical independence from NVIDIA's ecosystem, Mt. Diablo provides a credible path. If the priority is parity with the US AI frontier, NVIDIA 800 VDC. Most are quietly hedging.


7. What to watch over the next eighteen months

Three signals will tell you which architecture is gaining ground faster.

Component pricing parity. Track the bill of materials for an 800 V SST versus a ±400 V bipolar rectifier of equivalent capacity. Today the 800 V part is roughly 15-25% more expensive at the component level. The gap will close, that's the direction the learning curve points, but the timing depends on annual unit volumes hitting the threshold where 800 V parts move down the experience curve at EV-class rates. Watch component supplier price lists, not analyst forecasts.

ODM tooling ratio. As above. What fraction of new chassis tooling at Supermicro, Wiwynn, Foxconn, Quanta, Inventec is going to ORv3 HPR vs. NVIDIA reference. Quarterly capex disclosures and supplier earnings calls are the source. The ODMs have no loyalty; their tooling decisions are the ground truth.

Standards extensions. UL 98B coverage extending cleanly to 800 V data centre topologies. IEEE 1709 next revision. IEC 61140 DC supplements. If standards bodies write language that explicitly accommodates both unipolar and bipolar, hybrid wins. If they prefer one, that one wins.

Update: Since this essay was published, the picture has moved, and in a direction worth recording. The first-generation divergence it describes NVIDIA's 800 V unipolar reference and OCP Mt. Diablo's ±400 V bipolar default was real, and it traces to a shared starting point: both camps began at the rack, and both selected their voltages to inherit the existing EV supply chain rather than build a new one. Where they initially split was topology, not intent.

The response has been coordination rather than a winner-takes-all standards fight. IT players and electrical OEMs have converged in the OCP 800VDC subcommittee and engaged with with standards bodies, with the explicit goals of underwriting a safe transition, keeping a diversified supplier base behind every "brick" of the stack, and sparing manufacturers the burden of building to radically different specifications. That is a better outcome than the divergence the essay captured at a single point in time, and it is worth saying so. Original essay follows




8. The operator's question

If you're the executive responsible for an AI infrastructure budget in 2026, the question isn't "which architecture should I bet on?"

The question is: given my workload mix, my supplier relationships, and my time-to-revenue requirement, which architecture costs me less to be wrong about?

For most operators, the answer is the architecture your biggest tenant or partner is already committed to. Inertia is rational here. The sophistication is in recognizing that you'll probably be operating both architectures by 2030. And writing your supplier contracts now to keep that option open.

Essay 2 of 13 in The DC Transition series. Subscribe to get each essay as it publishes.


Sources

 * NVIDIA FY2025 data centre revenue ($115.2B): NVIDIA Q4 FY2025 earnings.
 * NVIDIA FY2026 data centre revenue (~$194B): Silicon Analysts (April 2026 update) referencing FY2026 trajectory.
 * AI accelerator market share trajectory (peak ~95% → 80-86% in 2025-26 → ~75% by end-2026): consensus across Silicon Analysts, TrendForce (October 2025), SemiAnalysis Q4 2025 model, Bloomberg Intelligence (January 2026).
 * Custom ASIC CAGR (44.6%): Bloomberg Intelligence "AI Accelerator Chips 2026 Outlook Deep Dive," January 2026.
 * ASIC shipments crossing GPU shipments by 2028: Counterpoint Research (cumulative 40M ASIC chips deployed across top 10 hyperscalers, 2024-2028).
 * TPU v7 Ironwood 2026 shipment scale (~4.3M units, cumulative tens of millions across 2026-2028): industry market sources covered in The Next Web (April 2026), Jon Peddie Research (December 2025).
 * Anthropic Ironwood commitment (>1M chips, >1 GW): Google Cloud / Anthropic joint announcement, late 2025.
 * Trainium 3 architecture (144 chips per UltraServer, 32 per rack in NL32x2 SKU, 3nm, GA December 2025): AWS re:Invent 2025; AWS Neuron documentation; SemiAnalysis "AWS Trainium3 Deep Dive," December 2025.
 * Project Rainier (Trainium 2, 500K+ chips with Anthropic): AWS, 2025.
 * ABB UL 98B 2 kV certification (July 2025, utility-scale solar application, UL 98B native test scope is to 1500 V): ABB press release, July 16 2025; pv-magazine USA, July 17 2025.
 * OCP Diablo 400 specification: Diablo 400 Project: Rack and Power v0.7.0 Base Specification, March 2026 (prior versions: v0.5.0 May 2025, v0.5.2 May 2025).
 * Sidecar vs row-scale architectural framing: Glenn K. Lockwood, "Mt. Diablo" reference note (May 2025); Diablo 400 v0.7.0 spec text on permitted output configurations.
 * Stranded capex risk framing: Hopper-to-Blackwell-to-Vera Rubin rack power trajectory (50 kW → 140 kW → 600 kW) per NVIDIA technical disclosures and IntuitionLabs deep-dive (March 2026); Goldman Sachs AI Infrastructure report (2025) on Blackwell facility-upgrade implications; CBRE Group infrastructure report (2025) on liquid-cooling retrofit costs ($2-5M per rack row).





Frequently asked


What 800V DC power rail solutions are hyperscale operators actually deploying?

Two distinct architectures both marketed as "800V DC" are being deployed. Microsoft, Meta and Google published the OCP Mount Diablo 400 specification (March 2026), which uses ±400V DC bipolar rails. Nvidia's own designs use a different unipolar 800V rail with different fault-current behaviour and grounding approach. Retrofit and greenfield decisions diverge sharply between the two.


What are the differences between 800V DC and Nvidia 800V vs OCP Mount Diablo?

Both are labelled "800V DC" but behave differently. OCP Mount Diablo 400 is a bipolar ±400V rail with symmetric ground reference, designed for retrofit compatibility and dual-fed rack safety. Nvidia's 800V is unipolar, engineered around Blackwell/Rubin rack topology. They differ in grounding architecture, fault-clearing philosophy, and vendor ecosystem.


Can 800V DC be retrofit into existing data centres?

Retrofit is possible for OCP Mount Diablo 400 style bipolar rails because they can be back-fitted onto existing dual-fed 415V AC distribution with converter shelves. Nvidia-style unipolar 800V typically requires greenfield build to fit its grounding scheme and vendor stack. Retrofit economics depend on rack density thresholds. Usually 60+ kW/rack to justify the conversion capex.


Which 800V DC vendors have proven hyperscale deployments?

Vertiv, Delta, ABB, Eaton and Schneider all have production 800V DC power shelves deployed at hyperscale (Microsoft, Meta, Google). Wolfspeed and Infineon provide the wide-bandgap semiconductors underneath. Vicor supplies on-package delivery. See the AI Power Chain vendor screen for the full list with named design wins.


==============================================================================
# The architecture map (where every stage lives, who builds it, what's standardized) (3/13)
URL: https://adikumar.co/the-architecture-map/
Published: 2026-05-08
Summary: Where every stage of AI data centre power lives, who builds it, what standardises when. Six-layer map for scoring vendor risk and content.
==============================================================================

The DC-DC Transition series · Part 3 of 19






Most operators making AI infrastructure decisions in 2026 are getting them wrong. But not for the reason the trade press is telling them.


The mistake isn’t picking the wrong voltage class, or backing the wrong supplier, or being too late to commit. The mistake is treating the architecture as monolithic. As if “we’re going 800 V” or “we’re committing to Mt. Diablo” were a single decision rather than what it actually is: a stack of six decisions, one per layer, each with its own vendors, standards, timelines, and commercial logic.


The smart operators aren’t choosing between 800 V and ±400 V. They’re choosing per layerTheir Layer 2 decision (perimeter conversion) doesn’t dictate their Layer 4 decision (rack-level conversion); their Layer 3 decision (building distribution) is separable from their Layer 5 decision (in-rack busbar). They make commitments where commitments earn them something, and preserve optionality everywhere else.


The operators getting it wrong are the ones treating the headline voltage choice as if it determined everything below it. It doesn’t. And the cost of that mistake. Measured in stranded capex, vendor lock-in, and capacity that can’t accept the next generation of accelerator. Compounds over the 15-25 year operational life of the facility.


This essay is the framework. The essays that follow in this series. On solid-state transformers, sidecar power racks, BESS coupling, connectors, the value chain. Each go deep on one of these layers. Before we get there, you need the map.



A note before we start


This essay is structural rather than argumentative. The point isn’t to convince you that one architecture wins. The point is to make sure that when you’re reading any of the next ten essays, you know which layer of the stack the argument is happening at, and what the layer above and below are doing.


If §1 of this essay reads like reference material. That’s intentional. Treat it as such. Bookmark it. Come back to it.



1. The six layers



From the utility connection to the GPU pin, every AI data centre passes power through six distinct layers. The voltage transitions, the conversion stages, and the protection regimes are different at each.


Layer 1. Utility interface. Where the data centre meets the grid. Medium-voltage AC delivery (typically 13.8 kV in North America, 11 or 22 kV in Europe and most of Asia, 33 kV in some industrial-park contexts). Owned by the utility up to the meter; owned by the operator from the meter inward. This is also where the BESS interconnection lives in modern designs.


Layer 2. Perimeter conversion. The first equipment inside the operator’s domain. Historically a step-down transformer to low-voltage AC; in DC architectures, this is increasingly a solid-state transformer (SST) producing DC directly at 800 V, ±400 V, or in some emerging designs at 1500 V. The design choice here gates everything downstream.


Layer 3. Building distribution. The DC bus or AC distribution that carries power from the perimeter conversion to the compute halls. In NVIDIA’s row-scale architecture, this is the bus that runs along each row. In Mt. Diablo’s architecture, this is the medium-voltage feed to each sidecar power rack. In legacy AC, this is the LV switchgear and busway. Topology matters enormously here for fault management and redundancy.


Layer 4. Rack-level conversion. The step from building-distribution voltage to the voltage the rack can use internally. In NVIDIA reference, the conversion happens in a sidecar power rack adjacent to (but feeding) compute racks. In Mt. Diablo, the conversion happens in a sidecar serving the IT rack as well. In legacy 48 V architectures, this conversion happened in the IT rack itself via the power shelf. The disaggregation of this layer from the IT rack is one of the bigger structural shifts in the transition.


Layer 5. In-rack distribution. The busbar that runs vertically through the IT rack, distributing power to each blade slot. This is the layer where the 50 V class still dominates regardless of what’s happening above. The 50 V busbar standard is OCP’s contribution from a decade ago and it’s stayed surprisingly stable through the transition.


Layer 6. Point-of-load. The final conversion from 50 V to the GPU’s actual operating voltage (a small handful of volts at very high current. Vera Rubin is hundreds of amps at sub-1 V core voltage). This is where Vicor, Infineon, Texas Instruments, and a handful of other suppliers fight a quiet but consequential battle. The decisions made here affect efficiency, cooling, and silicon longevity.


The layers aren’t independent. A choice at Layer 2 (SST topology) constrains what’s possible at Layer 3 (which busbar architectures work) which constrains Layer 4 (which sidecar designs are viable). But the layers are separable in a way that the marketing doesn’t always make clear. An operator can run NVIDIA-reference Layer 4 and Mt. Diablo-style Layer 3 if they’re willing to engineer the interface. Most won’t, but the option exists.



2. The standards landscape (and why it’s fragmented)



If you ask three different industry working groups what governs DC architecture in data centres, you’ll get three different answers, partly because each group’s scope covers different layers. The standards landscape is fragmented along organizational lines that have nothing to do with the operator’s view of the system.


Five bodies matter for the next five years.


IEC (International Electrotechnical Commission). Authoritative globally for voltage classification and equipment design. IEC 60038 defines the voltage classes (LVDC, MVDC, HVDC). This is the standard Essay 1 leaned on for the terminology correction. IEC 61140 covers protection against electric shock and is being supplemented for DC applications. IEC 62477 covers safety for power electronic converters. IEC SC 8A is a relatively new subcommittee specifically for LVDC distribution systems, where most of the active drafting work is happening.


IEEE. Authoritative in North America for DC distribution practice. IEEE 1709 is the recommended-practice standard for medium-voltage DC distribution on shipboard and industrial applications, and the next revision is being watched as a possible reference for stationary data centres at higher voltage classes. IEEE P3000 and the 3xxx series cover broader power systems engineering.


UL. Authoritative for product certification in North America. UL 98B (DC disconnect switches up to 1500 V) is the standard most often cited, and the standard most often misunderstood. As Essay 2 noted, ABB’s 2 kV cert was a stretch case ahead of UL writing the broader voltage range into the standard. UL 489B for DC molded-case circuit breakers. UL 1741 SB for grid-interactive equipment. UL 924 for emergency lighting (relevant for facility-level coordination). UL is currently the gating body for whether 800 V data centre topologies move from “engineerable” to “certifiable for production deployment.”


OCP (Open Compute Project). Not a standards body in the formal sense. OCP publishes specifications that members agree to interoperate against. But OCP specs (ORv3, ORv3 HPR, Mt. Diablo) function as de facto standards for hyperscaler-tier deployment because the spec authors are also the largest deployers. The Diablo 400 specification (v0.7.0, March 2026) is consequential current work.


NEC / NFPA (National Electrical Code). Authoritative for installation practice in the US. NEC Article 692 (PV systems, DC-relevant), Article 706 (energy storage systems), Article 712 (DC microgrids. Added in the 2017 cycle, refined since). The NEC is where the standards from above intersect with what an electrical inspector will sign off on. The gap between “the standard exists” and “the inspector is comfortable with it” can be years.


The fragmentation matters because no single body governs the full stack. An operator deploying 800 VDC has to satisfy IEC at the equipment-design level, UL at the product-certification level, NEC at the installation level, and OCP at the interoperability level. Each body moves at its own pace. The slowest one sets the schedule.


This is also why the standards-extension story (UL 98B coverage above 1500 V, IEC SC 8A new working drafts, IEEE 1709 next revision) is the most-watched thread in the industry. It’s not that the technology isn’t ready. It’s that four different bodies need to agree on what “ready” means before the inspector signs the certificate.



3. The vendor map by layer


The vendor landscape isn’t homogeneous across the stack. At each layer, the competitive structure is different.



Layer 1 (Utility interface). Operator’s own engineering team plus the utility. No vendor competition at this layer in any meaningful sense. The vendor is the local utility, and the relationship is regulatory more than commercial. Where there’s vendor activity, it’s around interconnection equipment and BESS aggregation: Tesla Megapack, Fluence, Wärtsilä, Sungrow, BYD on the BESS side; the same pool of utility-class transformer and switchgear OEMs (Hitachi Energy, Siemens Energy, GE, Mitsubishi Electric) on the interconnection side.


Layer 2 (Perimeter conversion / SST). This is where concentrated supplier power emerges as the transition deepens. SSTs are hard to build, capital-intensive, and protected by deep technology moats. The credible Western vendor list is short: Eaton, Schneider Electric, ABB, Hitachi Energy. Siemens is in the conversation but later. On the Chinese side, CRRC, BYD, and TBEA are credible at scale and benefit from state-backed pricing. Essay 8 will dive deeper here.


(Author note for transparency: I work at Eaton. I’ll say more in §6 about how I’m handling that disclosure across this series.)


Layer 3 (Building distribution). Busway and switchgear vendors. Eaton, Schneider, Vertiv, ABB, Legrand. Plus specialized DC busway suppliers (Starline, Universal Electric on the AC side moving toward DC). The vendor structure here is mature; the technology has to evolve to handle DC distribution at higher voltages but the manufacturing base exists.


Layer 4 (Rack-level conversion). Power shelf and PSU vendors. Advanced Energy, Delta Electronics, Lite-On, Bel Power Solutions, Flex Power Modules, Murata, plus the ODM-built shelves from Wiwynn, Foxconn, Quanta. Also the connector vendors who define the interface (TE Connectivity, Molex, Amphenol). This is the layer where the OCP spec ecosystem creates direct vendor competition. Anyone can build to the ORv3 HPR spec, and many do.


Layer 5 (In-rack distribution). Busbar manufacturers and rack OEMs. Mostly the rack OEMs themselves (Wiwynn, Foxconn, Inventec, Pegatron, Quanta on the contract manufacturing side; Supermicro, HPE, Dell on the branded side). The 50 V busbar is mature technology and the differentiation is mostly mechanical engineering, not power electronics.


Layer 6 (Point-of-load). A small handful of specialized power IC suppliers. Vicor, Infineon, Texas Instruments, Analog Devices (post-Maxim). This is concentrated layer and the one most invisible to the operator buying conversation. Decisions here are made by the silicon designer (NVIDIA, Google, Amazon, Meta, Microsoft) when they design their accelerator package, and the operator inherits whatever choice was made.


The pattern is worth noticing: the layers nearest the silicon (5 and 6) are concentrated and least visible to the operator. The layers nearest the utility (1, 2) are concentrated on the supplier side but most visible. The middle layers (3, 4) are highly competitive and directly addressable through procurement.


When the rest of this series talks about “where the value goes” in the DC transition, layers 2, 3, and 4 are where most of it will go. Layer 6 is where it already is. Layer 5 is where it likely won’t be.



4. The hybrid reference architecture (sketched)


Here’s the unpopular fact: most operators won’t deploy a pure NVIDIA 800 VDC architecture or a pure Mt. Diablo architecture. They’ll deploy a hybrid that picks the right answer per layer, given their workload mix and their tenant base.


A realistic 2027 hybrid for a major colocation operator looks like a mix at every layer: SSTs for the AI halls and traditional step-down transformers for the legacy ones; DC bus to AI halls and AC distribution to legacy ones; sidecar power racks for hyperscaler tenants and integrated NVIDIA Kyber racks for neocloud tenants; ORv3 HPR busbars in OCP-compliant racks alongside legacy 48 V busbars in the legacy halls; and whatever the silicon vendor specifies at the point-of-load, inherited.


The hybrid isn’t a compromise. It’s the realistic answer to a real operational constraint: the same operator is serving tenants with different architectural commitments, and the building has to accommodate all of them. Pricing each architecture distinctly. Mt. Diablo at one $/kW/month, 800 VDC at another, legacy AC at a third. Is where commercial sophistication shows up.


Hyperscalers running their own facilities have less of this problem because they control the workload mix. But even they hybridize: Microsoft’s deployment of Vera Rubin alongside Maia means they operate both architectures inside the same estate. The dual-track strategy from Essay 2 manifests as architectural hybridization at the facility level.


The operators most disadvantaged by all of this are those still trying to think of the architecture as monolithic. They make commitments at Layer 2 that lock in suboptimal choices at Layer 4, or they procure Layer 3 equipment that constrains what’s deployable at Layer 5. The architecture map isn’t optional. It’s how you avoid that.


Essay 6 walks through the operator economics, the case-study deployment patterns, and the procurement implications of running hybrid as a product. This essay just establishes that the hybrid is the realistic destination for most.



5. What this means for the rest of the series



Each of the upcoming essays in The DC Transition lives at a specific layer of this map. Setting them out now so you can place each one when it arrives.


 * Essay 4. DC Arc Behaviour and Safety. Cuts across Layers 2, 3, and 4. The protection regime is set by where the arc energy is highest, and that’s at the transitions between layers.
 * Essay 5. The Transition Tax. Cross-cutting again. The capex premium for DC architecture isn’t concentrated at any one layer; it accumulates across all six. Understanding which layers cost more matters for procurement allocation.
 * Essay 6. Hybrid Forever. The full case for §4 of this essay, expanded with operator economics and case-study deployment patterns.
 * Essay 7. The Voltage Ceiling. Layer 2 specifically, looking at the 1500 V / MVDC question.
 * Essay 8. The Solid-State Transformer. Layer 2, deep on supplier dynamics.
 * Essay 9. The Sidecar Power Rack. Layer 4, specifically the OCP-spec ecosystem.
 * Essay 10. Liquid Cooling and the Power Plant. A cross-cutting essay on the integration between cooling and power architecture, primarily affecting Layers 4 and 5.
 * Essay 11. BESS Coupling. Layer 1, with implications for Layer 2.
 * Essay 12. The Connector Wars. Layer 4 / Layer 5 interface specifically.
 * Essay 13. Where The Value Goes. Synthesis across all six layers with operator implications.


If at any point you lose the thread of which layer an essay is at, come back to this one. The map is the spine.



6. A note on disclosure and scope


This essay names Eaton in Layer 2 and Layer 3 of the vendor map. I work at Eaton. The disclosure block at the top of every essay in this series flags my employment relationship; this section adds the operational detail.


Where Eaton appears in this essay or any other in this series, it appears alongside its actual competitors and within its actual market position. I won’t overstate Eaton’s role in the transition; I won’t understate other suppliers’ roles to make Eaton look better; and I won’t write anything that could be construed as material non-public information about Eaton’s product roadmap. The discipline is straightforward: if I couldn’t say it on a public earnings call as a participant, I don’t say it here.


If you spot a place where I’ve gotten that wrong, tell me. The point of the disclosure isn’t to give me cover to write whatever I want. It’s to give readers the context to push back when they see bias.



7. The closing question for operators


If you’re inheriting an AI infrastructure budget in 2026, the useful thing you can do this week is map your own architecture stack to these six layers. Specifically:


 * For each layer, name the architecture you’ve committed to (or inherited from a previous decision)
 * For each layer, name the vendor or vendors you’re locked into
 * For each layer, name the standard you’re certifying against
 * Identify which layer-to-layer interfaces are constraining your options at the next layer down


That exercise will tell you something most operators don’t know: where in the stack your degrees of freedom actually are. Most operators discover they have less freedom than they thought at Layers 4 and 5, and more freedom than they thought at Layers 2 and 3.


The architectural debate in trade press is about Layer 2 and Layer 3. The architectural lock-in is at Layers 4 and 5. The architecture-to-margin question is at Layer 2. Each of these gets its own essay. This one gives you the map.


Essay 4 looks at DC arc behaviour. The safety regime being written without much public attention, and why insurance underwriters are about to have stronger opinions about your architecture choice than your engineers do.



Essay 3 of 13 in The DC Transition series. Subscribe to get each essay as it publishes.




Sources


 * IEC 60038 Edition 7.0 (2009) and Amendment 1 (2021): voltage classification standard.
 * IEC SC 8A: System aspects of low-voltage DC distribution working group, IEC TC 8.
 * UL 98B (DC disconnect switches up to 1500 V): UL Standards database; ABB UL 98B 2 kV certification, July 2025; pv-magazine USA, July 17 2025.
 * IEEE 1709 (Recommended Practice for Medium-Voltage DC Distribution): IEEE Standards Association.
 * NEC Article 712 (DC microgrids, added 2017 NEC cycle): NFPA.
 * OCP ORv3 specification: Open Compute Project, Open Rack v3 Base Specification.
 * OCP ORv3 HPR specification: Open Compute Project, Open Rack v3 High Power Rack specification.
 * OCP Diablo 400 specification v0.7.0, March 2026.
 * OCP ORV3 HVDC-LVDC 100kW Power Shelf Design Specification v1.0.0: opencompute.org.
 * Vendor positioning: company product disclosures and OCP marketplace listings (Eaton, Schneider, Hitachi Energy, ABB, Vertiv, Advanced Energy, Delta, Lite-On, Bel Power, Vicor, Infineon, TI, Analog Devices, Wiwynn, Foxconn, Quanta, Inventec, Supermicro).




Glossary of terms used



GPU
Graphics Processing Unit. The compute silicon at the centre of AI workloads.
HVDC
High-Voltage Direct Current. DC transmission at hundreds of kilovolts, used for long-distance and undersea power transmission and, increasingly, for high-density data centre power distribution.
IC
Investment Committee. The decision-making body inside a PE firm or investor that approves or rejects a proposed transaction.
OCP
Open Compute Project. Hyperscaler-led standards body developing open reference designs for data centre hardware.
VDC
Volts Direct Current. Used in 48 VDC and 800 VDC references to rack-level and datacentre power distribution.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.






Series expansion · six new supplements


The DC-DC Transition series now includes six deep-dive supplements addressing specific layer-level questions that the main 13-essay series flagged but did not fully develop. Consulting-grade visuals throughout. Six-layer stack decompositions, vendor concentration matrices, procurement flow diagrams:


 1. Supplement A. The Transition Tax: layer-by-layer capex decomposition. Where the 800V DC premium accumulates. Retrofit vs greenfield.
 2. Supplement B. Hybrid Forever: why most facilities will run mixed AC/DC through 2035+. Hyperscaler zone strategies compared.
 3. Supplement C. The Voltage Ceiling: 1500V DC and the MVDC question. Standards + vendor readiness by voltage class.
 4. Supplement D. The Solid-State Transformer: Layer 2 supplier dynamics. Western vs Chinese vendor landscape.
 5. Supplement E. The Sidecar Power Rack: Layer 4 architecture. OCP Mt. Diablo v0.7.0 reference. 80+ ecosystem partners.
 6. Supplement F. The Connector Wars: Layer 4/5 interface engineering. High-current connector vendor concentration + failure modes.


Full series index: /tag/dc-transition/All supplements now live.







In brief


AI data centre power spans six layers: grid interconnect, facility distribution (415V AC or 800V DC), rack-level shelves, on-package delivery, thermal (co-emerging), and modular build. Each layer has different vendor concentration, standards maturity, and lead-time exposure. A 100 MW campus adds 2-5% loss and one vendor-concentration risk per stage.







The AI Power Chain: series companions

 * The AI Power Chain: Vendor Screen. The vendor map across every layer
 * Pricing Under Scarcity. Where premium capture is compounding
 * The Services Inversion. Why services now command product-like multiples







Frequently asked


As a DC architect specifying a 50 MW AI campus, what should the power-chain topology look like?

For a 50 MW AI campus, the standard power chain is: medium-voltage utility feed (13.8 kV or 34.5 kV) → step-down transformer to 415V AC or straight to 800V DC → rack-level shelves converting to 48V or on-package 800V, with two levels of redundancy at every conversion stage. Rack density (100-600 kW) determines whether AC-in-rack or DC-in-row wins.


What are the six layers of the AI data centre power architecture?

The six layers are: 1) Grid interconnect + medium-voltage substation, 2) Facility distribution + UPS + BESS, 3) 800V DC or 415V AC power shelves, 4) Rack-level converters (48V or point-of-load), 5) On-package power delivery (0.8V-1.2V multi-phase), 6) Thermal (DLC/immersion) as a co-emerging constraint. Each has different vendor concentration and lead-time profile.


What is an AI factory power-train architecture?

An AI factory power train is the end-to-end path from utility MW to GPU-level watts. For a 1 GW AI factory: 230 kV grid → 13.8 kV substation → 415V AC or 800V DC facility distribution → rack converters → 48V or 800V DC rack bus → on-package multi-phase point-of-load → 0.8V to GPU dies. Each stage adds 2-5% loss and its own vendor concentration.


==============================================================================
# The Transition Tax: where the 800V DC capex premium lives (DC-DC Supplement A)
URL: https://adikumar.co/dc-dc-supplement-a-transition-tax/
Published: 2026-08-20
Summary: Layer-by-layer decomposition of the 800V DC capex premium for AI data centres. 100 MW greenfield reference.
==============================================================================

The DC-DC Transition series · Part 14 of 19








The DC-DC Transition · Supplement A



The Transition Tax: where the 800V DC capex premium actually lives


The 800V DC architecture premium relative to incumbent 48V AC-DC accumulates across all six layers of the power chain rather than concentrating at any single one. Layer-by-layer, procurement teams under-allocate at some layers and over-allocate at others. This supplement decomposes the total capex premium (typically 15-28% of total DC infrastructure spend for a greenfield 100 MW facility) into the six layer-specific contributions, showing where the real dollars go and where procurement discipline pays off.





The DC-DC Transition · supplements to the 13-essay main series

 1. A. The Transition Tax (you are here)
 2. B. Hybrid Forever. Operator economics of mixed AC/DC
 3. C. The Voltage Ceiling. 1500V DC + MVDC question
 4. D. The Solid-State Transformer. Layer 2 supplier dynamics
 5. E. The Sidecar Power Rack. Layer 4 OCP-spec ecosystem
 6. F. The Connector Wars. Layer 4/5 interface engineering


Companion to the 13-essay main series: The DC-DC TransitionStart with The Architecture Map.





Reader takeaways

Total transition tax15-28% capex premium vs incumbent 48V AC-DC for greenfield 100 MW facility (typical range)
Where it accumulatesLayer 2 (medium-voltage distribution) + Layer 5 (cooling integration) carry the largest absolute premium. Layer 1 (utility interface) marginal. Layer 4 (rack power) surprisingly variable.
Where under-allocatedLayer 3 (protection: SSCBs) and Layer 6 (commissioning + skills). Procurement teams routinely under-budget these
Where over-allocatedLayer 4 (rack PDU). Procurement often over-specifies redundancy that isn't required at 800V DC architecture
10-year TCOTransition tax recovered through year 3-5 via lower opex; net-positive by year 7 in most scenarios
Retrofit vs greenfieldRetrofit tax 40-60% higher than greenfield due to Layer 5 cooling integration friction






01The six-layer capex baseline



Total DC infrastructure capex for a modern 100 MW greenfield facility runs approximately $850M-1.1B (2026 pricing). This breaks down across the six-layer AI Power Chain: Layer 1 (utility interface + medium-voltage) ~$40-60M, Layer 2 (medium-to-low voltage conversion) ~$120-160M, Layer 3 (protection + switchgear) ~$70-100M, Layer 4 (rack-level distribution) ~$180-240M, Layer 5 (cooling infrastructure) ~$180-260M, Layer 6 (commissioning + integration) ~$60-100M. Ancillary + site works ~$180-200M.



For the same 100 MW facility built to 800V DC architecture instead of incumbent 48V AC-DC, total capex runs approximately $1.0B-1.4B. A 15-28% premium. The specific question is where in the six layers that premium sits.





Figure A.1
The Transition Tax. Capex premium by layer for 100 MW greenfield

Layer 1Utility interfaceIncumbent $50M → 800V DC $55MUtility side unchangedPrimary vendors: +$5M premium (10%)Layer 2Perimeter conversionIncumbent $140M → 800V DC $185MSST vs conventional transformerPrimary vendors: +$45M premium (32%)Layer 3Protection + switchgearIncumbent $80M → 800V DC $105MSSCBs + DC ground-faultPrimary vendors: +$25M premium (31%)Layer 4Rack-level distributionIncumbent $210M → 800V DC $225MSimpler chain saves; premium on connectorsPrimary vendors: +$15M premium (7%)Layer 5Cooling infrastructureIncumbent $210M → 800V DC $300MDLC or immersion requiredPrimary vendors: +$90M premium (43%)Layer 6Commissioning + skillsIncumbent $70M → 800V DC $100MLonger commissioning; premium wagesPrimary vendors: +$30M premium (43%)Power flow






Chart 1-100 MW greenfield capex: incumbent 48V AC-DC vs 800V DC


Total: incumbent $950M vs 800V DC $1.15B (representative). Premium: $200M or ~21%. Not evenly distributed across layers. Concentrated in specific ones.






02Layer-by-layer premium decomposition




LayerIncumbent capex ($M)800V DC capex ($M)PremiumWhy

1. Utility interface + MV$50$55+10%Minimal. Utility side essentially unchanged
2. MV-to-LV conversion$140$185+32%SST / MV rectifier premium vs conventional transformer
3. Protection + switchgear$80$105+31%SSCBs vs mechanical breakers; DC ground-fault detection
4. Rack-level distribution$210$225+7%Simpler distribution actually saves on some cost lines; premium on new connector standards
5. Cooling infrastructure$210$300+43%Higher-density cooling required (DLC or immersion) + integration complexity
6. Commissioning + skills$70$100+43%Longer commissioning + specialised skills at premium wages
Site works + ancillary$190$180-5%Slightly smaller footprint given density gains
Total$950$1,150+21%






03Where the premium concentrates




Chart 2. Absolute premium by layer ($M)


Layer 5 (cooling): $90M. Largest single-layer premium. Layer 2 (MV-to-LV): $45M. Layer 3 (protection): $25M. Layer 6 (commissioning): $30M. Others minimal. Site works actually save $10M. Cooling + Layer 2 collectively = 67% of total transition tax.






04What procurement teams routinely miss



Two systematic biases in AI DC procurement drive suboptimal allocation. First, procurement over-invests in Layer 4 redundancy (rack-level PDU redundancy) that was needed at 48V AC-DC for reliability but is meaningfully less critical at 800V DC where the distribution chain is shorter. Second, procurement under-invests in Layer 6 (commissioning + skills). Treating it as a construction line item rather than a specialised discipline requiring premium-wage electricians (see Regulatory Layer X).





Figure A.2
Transition tax intensity by layer × dimension

Greenfield taxRetrofit taxRecovery yrsUnder-alloc riskOver-alloc riskLayer 1 Utility+10%+15%n/aLowLowLayer 2 MV-LV+32%+45%Yr 4MedLowLayer 3 Protection+31%+45%Yr 5HighLowLayer 4 Rack+7%+25%Yr 3LowHighLayer 5 Cooling+43%+85%Yr 4MedMedLayer 6 Comm+43%+65%Yr 2HighLowIntensitylow → high






Chart 3. Systematic procurement bias: over- vs under-allocation by layer


Over-allocated: Layer 4 rack PDU redundancy (+15-25% typical). Under-allocated: Layer 6 commissioning (-30-40% vs actual), Layer 3 SSCB testing budget (-20-30%). Rebalancing releases 5-8% of total capex to more productive layers without changing total spend.






05Retrofit vs greenfield transition tax



Retrofit projects carry a materially higher transition tax than greenfield. Layer 5 cooling integration is the specific driver: retrofit facilities were designed against AC power distribution + air cooling assumptions. Converting to 800V DC + DLC requires physical rework of chilled-water plumbing, structural reinforcement for CDU weight, and specific electrical rework in existing switchgear rooms. Retrofit transition tax typically 25-40% of total facility capex vs 15-28% for greenfield.




Chart 4. Retrofit vs greenfield transition tax by layer


Greenfield: 15-28% total tax. Retrofit: 25-40% total tax. Layer 5 cooling: retrofit 60-90% premium vs greenfield 43%. Layer 4 rack: retrofit 20-30% premium vs greenfield 7%. Retrofit rarely pencils out; see companion essay VII.






0610-year TCO recovery of the transition tax



The transition tax is recovered over the 10-year TCO through lower opex driven by (a) higher end-to-end power efficiency at 800V DC (5-7 percentage points vs 48V AC-DC), (b) longer equipment lifecycle, (c) lower cooling infrastructure loss coefficient at higher density. Net-positive by year 3-5 depending on capacity utilisation + electricity pricing assumptions.





Figure A.3
Procurement bias: where teams routinely over- vs under-allocate

Systematic overallocationWhere procurement over-investsLayer 4 rack PDU redundancy (+15-25% typical)Layer 1 legacy utility interface hardwareLayer 3 mechanical breaker over-specificationRedundant transformer capacity Layer 2Air-cooling infrastructure in DLC facilitiesSystematic underallocationWhere procurement under-investsLayer 6 commissioning + specialised electricians (-30-40%)Layer 3 SSCB acceptance testing budget (-20-30%)Layer 5 CDU + secondary loop capacity headroomContingency for utility interconnection delaysStandby servicer capacity Layer 3






Chart 5. Cumulative TCO: 800V DC vs incumbent 48V AC-DC over 10 years


800V DC starts $200M higher at year 0. Crosses over by year 4-5. By year 10, 800V DC has $400-600M lower cumulative TCO. Recovery pace depends on utilisation (fast-utilised facility recovers faster).






07Implications for procurement allocation



Three specific actions from this analysis: (1) Reallocate 5-8% of total capex from Layer 4 rack redundancy toward Layer 6 commissioning + Layer 3 protection testing. (2) Budget Layer 5 cooling premium explicitly rather than treating it as continuation of prior air-cooling assumptions. (3) For retrofit decisions, apply 25-40% transition tax rather than the 15-28% greenfield figure. Many retrofit business cases fail this recalibration.




Chart 6. Recommended capex reallocation from procurement bias correction


Radar: current allocation vs recommended. Current: over-weighted Layer 4, under-weighted Layer 6 + Layer 3 testing. Recommended: rebalanced across layers based on actual value + risk. Doesn't change total; changes destination.






Method and sources. Public information only. Cost breakdown modelled from OCP published reference designs 2024-2026, vendor pricing surveys (Vertiv, Eaton, Schneider, Vicor), specific hyperscaler capex disclosures where available. Numbers are representative for a 100 MW greenfield US facility; specific projects vary materially. TCO recovery model uses 10-year DCF at 8% discount rate.


Series footer. Supplement A to The DC-DC TransitionCompanion: DC-DC VII (retrofit vs greenfield), DC-DC XI (vendor economics)Hubs: 800V DC hub, AI Power Semi hub.










Primary sources

 * Open Compute Project (OCP). OCP working group publications
 * OCP Rack + Power. OCP rack + power working group







The DC-DC Transition: start with the foundation

 1. The real reason data centres are going DC (four forces)
 2. Two architectures wearing the same name (800V DC vs OCP Mt Diablo)
 3. The six-layer architecture map


==============================================================================
# Hybrid Forever: most facilities will run mixed AC/DC through 2035+ (DC-DC Supplement B)
URL: https://adikumar.co/dc-dc-supplement-b-hybrid-forever/
Published: 2026-08-20
Summary: Deep-dive on hybrid AC/DC data centre architectures. Zone-based deployment patterns. Hyperscaler strategy comparison (Microsoft, Meta, Google, AWS).
==============================================================================

The DC-DC Transition series · Part 15 of 19








The DC-DC Transition · Supplement B



Hybrid Forever: why most facilities will run mixed AC/DC through 2035+


The clean-slate "800V DC only" facility is the outlier, not the norm. Most hyperscaler + operator sites will run hybrid AC + DC architecture for the entire 2026-2035 window. Not because 800V DC is inferior. Because installed base + retrofit economics + specific workload profiles + regional grid variability + vendor lock-in each independently pull toward hybrid operation. This supplement covers the operator economics of hybrid and specific case-study deployment patterns from named hyperscalers.





The DC-DC Transition · supplements

 1. A. The Transition Tax
 2. B. Hybrid Forever (you are here)
 3. C. The Voltage Ceiling
 4. D. The Solid-State Transformer
 5. E. The Sidecar Power Rack
 6. F. The Connector Wars


Companion to the 13-essay main seriesStart with The Architecture Map.





Reader takeaways

Hybrid share 2030 est70-80% of operational DC capacity will run hybrid AC + DC architecture. Pure-DC facilities: 15-25%. Pure-AC: 5-10% (mostly legacy)
Why hybrid persistsInstalled AC base + retrofit friction + non-AI workloads better suited to AC + regional grid + vendor availability + specific customer contracts
Named deployment patternsMicrosoft: hybrid site-by-site allocation. Meta: AI zones DC + non-AI AC. Google: architectural zones. AWS: primarily AC with DC pilots. Traditional colo: mostly AC with DC-ready expansion
Operating economicsHybrid facility opex 3-8% higher than pure-DC comparable, largely from parallel infrastructure + specific commissioning costs
Practical takeawayEvery serious operator needs both AC and DC operational competence through 2035+. Vendor evaluation should include hybrid-compatibility criteria, not just pure-DC capability






01Why the pure-DC vision is unrealistic



The theoretical pure-DC data centre. Utility MV in, 800V DC distribution throughout, direct-to-chip cooling everywhere, no AC anywhere on the operating floor. Is architecturally clean but operationally rare. Six structural forces pull toward hybrid: (1) existing installed AC base at any site >2 years old, (2) non-AI workloads that are AC-optimised, (3) regional grid interconnect where MV is the natural interface, (4) vendor availability at specific tiers where DC options are still emerging, (5) specific customer contracts specifying AC or hybrid, and (6) operational familiarity + skills.





Figure B.1
Hybrid architecture. How the six layers split across AC and DC zones

Layer 1Utility interfaceMV AC (unchanged in hybrid)AC and DC share utility feedPrimary vendors: Same infrastructure both AC + DC zonesLayer 2Perimeter conversionHybrid: parallel AC transformer + SSTBoth paths co-existPrimary vendors: ~30% capex premium vs single-architectureLayer 3Building distributionZone-based: AC bus / DC bus splitZoning driven by workload profilePrimary vendors: Separate distribution per zoneLayer 4Rack-level conversionAC PDU or DC sidecarRack rebuilds move between zonesPrimary vendors: Determined by zone assignmentLayer 5In-rack distribution48V busbar (identical in both zones)Common denominatorPrimary vendors: One standard; zone-agnosticLayer 6Point-of-loadSame POL silicon regardlessIndependent of zone architecturePrimary vendors: Vicor/Infineon unchangedPower flow






Chart 1. DC facility architecture mix projection 2025-2035


2025: ~5% pure DC, 10% hybrid, 85% pure AC. 2030: ~20% pure DC, 65% hybrid, 15% pure AC. 2035: ~30% pure DC, 60% hybrid, 10% pure AC. Hybrid share peaks 2030-2032 then slowly declines but remains dominant through 2035.






02Named hyperscaler deployment patterns




OperatorPatternDC share (2026)DC share (2030 est)

MicrosoftSite-by-site allocation; new AI-density sites DC-first, legacy retained~25%~55%
MetaArchitectural zones: AI zones DC, storage/network zones AC~30%~60%
GoogleSimilar architectural-zone pattern with TPU zones DC-preferred~25%~55%
AWSPrimarily AC with DC pilots; slower shift~10%~30%
OracleLegacy AC with AI-specific DC expansion~5%~35%
Traditional colo (DLR + EQIX)Mostly AC with DC-ready customer options~5%~25%
Independent GPU cloud (CoreWeave + peers)Newer facilities DC-first, older AC~40%~70%







Figure B.2
Hyperscaler DC share + hybrid strategy comparison (2026 → 2030 est)

DC 2026DC 2030 estHybrid strategyAI zone focusRetrofit approachMicrosoft25%55%Site-by-siteNew buildsSelectiveMeta30%60%Arch zonesAI zonesRareGoogle25%55%Arch zonesTPU zonesRareAWS10%30%PilotsSelectiveNoneOracle5%35%AI-specificNew AI onlyNoneTraditional colo5%25%Customer-drivenDC-ready optNoneIndep GPU cloud40%70%DC-first newAI nativeRareIntensitylow → high






Chart 2. Hyperscaler DC share of capacity 2026 vs 2030 est


Wide range 2026 (5-40%). Convergence somewhat by 2030 to 30-70% range but still substantial spread. Hyperscaler strategies not converging on identical mix; different workload profiles + facility age + operational philosophy drive different equilibria.






03Non-AI workloads and AC preference



Not all data-centre workloads benefit from 800V DC architecture. Storage-heavy workloads (S3, GCS, Azure Blob) run at moderate density where AC infrastructure is fully adequate. Network/CDN workloads similarly moderate. Batch analytics workloads without tight-loop GPU inference are density-modest. Traditional enterprise workloads (VMs, containers, web services) are AC-native. Only tight-density AI training + latency-critical inference specifically benefit from 800V DC. In most facilities, 40-60% of workload footprint remains in the moderate-density regime where AC is adequate.




Chart 3. Workload density profile vs power architecture optimum


AI training (400+ kW/rack): DC optimum. AI inference (200-400 kW): DC preferred. High-perf compute (100-200 kW): DC or hybrid. Storage/network (30-80 kW): AC adequate. Enterprise VM (10-30 kW): AC native. Distribution of workloads across density regimes creates hybrid natural.






04Hybrid operating economics



Operating a hybrid facility carries specific cost premiums vs a pure-DC comparable. Parallel infrastructure (both AC and DC distribution present) roughly 5-8% capex premium and 3-5% opex premium. Additional commissioning + testing burden 15-25% higher. Specialised staff to operate both architectures. However, hybrid also carries specific benefits: workload flexibility, redundancy across architectures, ability to serve customers with either preference, easier retrofit path over time.





Figure B.3
Hybrid facility power flow. From utility to compute

Utility MV AC13.8 kVPerimeter splitAC + DC branchesZone assignmentAI vs non-AIRack architectureDC 800V or AC 480VCompute deploymentGPU or CPU






Chart 4. Hybrid vs pure-DC operating economics


Hybrid capex vs pure-DC: +5-8%. Hybrid opex vs pure-DC: +3-5%. Commissioning burden: +15-25%. Workload flexibility: hybrid significantly better. Retrofit optionality: hybrid materially better. Trade-off is real but often justifies hybrid for larger operators.






05Vendor implications



Hybrid dominance shapes vendor strategy across all six power-chain layers. Vendors selling pure-DC-only products (some new 800V-DC-native startups) face a narrower TAM than the total-DC-market suggests. Realistic TAM is 15-25% pure-DC segment. Vendors offering hybrid-compatible products (traditional broad-portfolio players like Vertiv + Eaton + Schneider) address the 70-80% hybrid market plus optional pure-DC. This shapes procurement + investment decisions materially.




Chart 5. Vendor TAM implications: pure-DC vs hybrid-compatible


Pure-DC vendor TAM (2030): $30-40B (15-25% of DC infrastructure market). Hybrid-compatible vendor TAM: $120-160B (70-80% share). Broader-portfolio vendors address both. Pure-DC-only vendors face structurally narrower market than headline figures suggest.






06Practical takeaway



Every serious operator needs both AC and DC operational competence through 2035+. Every serious vendor needs hybrid-compatible products, not just pure-DC. Every serious investor evaluating DC infrastructure exposure should stress-test the pure-DC assumption. And every serious procurement team should build hybrid capability into their evaluation criteria, not just pure-DC checkbox.




Chart 6. Radar: operator readiness dimensions for hybrid operation


Skills: hybrid operators need both AC + DC electrician + technician competence. Procurement: dual-architecture evaluation criteria. Commissioning: both AC + DC test protocols. Facility design: architectural zones + parallel infrastructure. Vendor selection: hybrid-compatibility mandatory. Customer contracting: allow either preference.






Method and sources. Public information only. Hyperscaler capex disclosures 2024-2026, OCP power workgroup materials, vendor product portfolio analysis. Named hyperscaler DC share estimates are based on public commentary + industry reporting; specific figures not disclosed. Workload profile analysis synthesised from AI infrastructure workload research 2024-2026.


Series footer. Supplement B to The DC-DC TransitionCompanion: DC-DC VII (retrofit vs greenfield), DC-DC XIII (ten-year view)Hub: 800V DC hub.










Primary sources

 * Open Compute Project (OCP). OCP working group publications
 * Uptime Institute research. Data centre operations research







The DC-DC Transition: start with the foundation

 1. The real reason data centres are going DC (four forces)
 2. Two architectures wearing the same name (800V DC vs OCP Mt Diablo)
 3. The six-layer architecture map


==============================================================================
# The Voltage Ceiling: 1500V DC and the MVDC question (DC-DC Supplement C)
URL: https://adikumar.co/dc-dc-supplement-c-voltage-ceiling/
Published: 2026-08-20
Summary: HVDC voltage ceiling for AI data centres. 800V DC vs 1500V DC vs MVDC decision, regulatory boundaries, and where to commit for 2028-2032.
==============================================================================

The DC-DC Transition series · Part 16 of 19








The DC-DC Transition · Supplement C



The Voltage Ceiling: 1500V DC and the MVDC question


800V DC is not the endpoint. Layer 2 (perimeter conversion) faces a specific voltage-ceiling decision that will shape 2028-2032 facility architecture: hold at 800V, step to 1500V DC, or reach for MVDC at 5-15 kV. Each has different copper savings, standards readiness, vendor availability, and regulatory clarity. This supplement covers the specific voltage-ceiling trade-offs, the vendors best positioned at each voltage class, and the practical decision framework for operators specifying facilities that will commission in 2028-2030.





The DC-DC Transition · six supplements

 1. A. The Transition Tax
 2. B. Hybrid Forever
 3. C. The Voltage Ceiling (you are here)
 4. D. The Solid-State Transformer
 5. E. The Sidecar Power Rack
 6. F. The Connector Wars


Companion to the 13-essay main seriesStart with The Architecture Map.





Reader takeaways

800V DC statusEcosystem now mature. Standards ready (OCP Mt. Diablo, IEC 60947-10). Multiple vendors. Reference architecture for 2026-2028 builds.
1500V DC trajectoryNear-term (2027-2029) transition for highest-density AI facilities. UL 98B extension in progress. 2-3 Western vendors credible. 30% copper savings.
MVDC (5-15 kV)Campus-scale distribution only. Not in-facility. Multi-year regulatory + standards drafting horizon. Chinese CRRC + BYD ahead of Western vendors on hardware readiness.
Decision frameworkTarget facility density + expected 10-year retrofit horizon + insurance posture + local NEC edition determine specific voltage-class choice
Rack-level unchanged48V busbar (Layer 5) stays fixed regardless of upstream voltage. Physics constraint at silicon POL (Layer 6) unlikely to move in 10-yr window.






01Where the voltage ceiling matters


The voltage-ceiling question is a Layer 2 + Layer 3 debate. Layers 4-6 are constrained by silicon point-of-load physics and by the OCP 48V rack-level standard, both of which are unlikely to move materially in the 2026-2035 window regardless of upstream voltage choices. Where the voltage ceiling matters. And where the trade-offs actually happen. Is in perimeter conversion (Layer 2) and building distribution (Layer 3).





Figure C.1
Where the voltage ceiling matters. Layers 2 and 3 specifically

Layer 1Utility interface13.8-33 kV AC (unchanged)MV utility feedPrimary vendors: Hitachi Energy, Siemens, GE, MitsubishiLayer 2Perimeter conversion800V now / 1500V DC / MVDC futureTHE voltage ceiling questionPrimary vendors: SST at higher voltage classes emergingLayer 3Building distribution800V DC bus (current) / 1500V DC (near)Higher voltage = less copperPrimary vendors: Eaton, Schneider, Vertiv, ABBLayer 4Rack-level conversion48V (unchanged for foreseeable)Constrained by silicon POLPrimary vendors: Advanced Energy, Delta, WiwynnLayer 5In-rack distribution48V busbarStandard-frozenPrimary vendors: Wiwynn, Foxconn, Inventec, QuantaLayer 6Point-of-loadsub-1V @ high currentPhysics constraint at siliconPrimary vendors: Vicor, Infineon, TI, ADIPower flow








02The voltage class options


Four voltage classes are in active consideration for AI infrastructure Layer 2-3 architecture: 800V DC (current mainstream), 1500V DC (near-term step-up), 2500V DC (R&D + specific ultra-density applications), and MVDC at 5-15 kV (campus-scale distribution). Each has different maturity across five dimensions: copper savings vs baseline, standards readiness, vendor availability, regulatory clarity, and facility fit.





Figure C.2
Voltage class comparison. Copper savings vs standards + vendor readiness

Copper savingsStandards readyVendor availabilityRegulatory clarityFacility fit800V DC (current)BaselineIEC/OCPMultipleEmergingHyperscale ok1500V DC (near-term)-30%IEC/UL 24-252-3 vendorsUL 98B okAI-density2500V DC (mid)-50%R&D onlyNone WesternRulemakingUltra-densityMVDC ~15kV (long)-70%IEC draftingUtility-scale onlyMulti-yearCampus-scale48V DC (rack, unchanged)n/aOCP frozenAll rack OEMsEstablishedAll racksIntensitylow → high








Chart 1. Copper cost savings by voltage class vs 800V baseline


Higher voltage = same power at lower current = thinner conductors + less copper. 1500V DC saves ~30%. 2500V DC ~50%. MVDC ~70%. Copper savings drive economic case for stepping up but must be weighed against ecosystem readiness constraints.






03The case for 1500V DC (and against)





Figure C.3
Voltage ceiling debate. 800V vs 1500V vs beyond

Case for 1500V DCWhy the industry is moving upCopper cost 30% lower vs 800V (thinner bus, same current)Facility density 15-25% improvement per MW rack densityUL 98B extension underway; IEC readiness 2025-2026Multiple credible vendors (Eaton, Schneider ABB, CRRC)Compatible with existing rack-level 48V standardCase for holding at 800VWhy 1500V may be prematureFull ecosystem still hardening (SSCBs, protection)Skilled electrician shortage worse at higher voltageInsurance underwriter comfort at 1500V not yet universalRetrofit path from 800V to 1500V is expensiveAI density gains available at 800V for 2028-30 targets








Chart 2. Standards + regulatory timeline: 800V vs 1500V vs MVDC


800V DC: ecosystem now. 1500V DC: standards + certifications through 2025-2027. MVDC: multi-year IEC drafting + regional NEC-equivalent adoption cycles into 2029-2032.






04Vendor positioning by voltage class


Vendor readiness varies materially by voltage class. Western SST vendors (Eaton, Schneider, ABB, Hitachi Energy) have mature 800V DC products, emerging 1500V DC roadmaps, R&D-only 2500V DC positions. Chinese vendors (CRRC, BYD, TBEA) are competitive at 800V + 1500V + emerging at MVDC + traditional-transformer variants. Newer entrants (Innoscience GaN, specific pure-play SST startups) target specific niches.




Chart 3. Vendor readiness by voltage class + geography


Western + Chinese vendor readiness scored by voltage class 2026. 800V DC: multiple credible vendors. 1500V DC: fewer, mostly early production. 2500V DC + MVDC: sparse Western, some Chinese.






05Decision framework for operators


The specific decision framework for operators specifying facilities in 2026-2028 that will commission 2028-2030: (1) target facility density profile drives the answer. Sub-200 kW/rack targets are fine at 800V; 200-500 kW/rack targets justify 1500V investigation; ultra-density 500+ kW/rack begins to require 1500V or beyond. (2) Expected 10-year retrofit horizon matters. If the facility will need retrofit for higher voltage within a decade, over-spec Layer 2-3 now. (3) Insurance underwriter posture at higher voltages is a specific constraint. (4) Local NEC edition + state adoption timing gates certifiable installation.




Chart 4. Voltage class decision matrix by facility target profile


Decision framework: target density + retrofit horizon + insurance + NEC edition. Radar visualisation shows which voltage class fits which target profile. Ultra-density targets pull toward 1500V; standard AI density fine at 800V through late 2020s.






Method and sources. Public information only. IEC 60038 voltage class standard; IEC 60947-10 SSCB standard (early 2026); UL 98B DC disconnect switches extension progress; OCP Mt. Diablo v0.7.0 specification March 2026; vendor product roadmaps 2024-2026.


Series footer. Supplement C to The DC-DC TransitionRelated: Supplement D (SST supplier dynamics deep-dive), DC-DC XIII (ten-year view), DC-DC X (standards evolution)Hub: 800V DC hub.










Primary sources

 * IEC. IEC voltage class standards
 * UL Standards. UL DC voltage standards
 * OCP Rack + Power. Mt. Diablo specification







The DC-DC Transition: start with the foundation

 1. The real reason data centres are going DC (four forces)
 2. Two architectures wearing the same name (800V DC vs OCP Mt Diablo)
 3. The six-layer architecture map


==============================================================================
# The Solid-State Transformer: Layer 2 supplier dynamics (DC-DC Supplement D)
URL: https://adikumar.co/dc-dc-supplement-d-solid-state-transformer/
Published: 2026-08-20
Summary: Deep-dive on solid-state transformer (SST) supplier landscape for AI DC. Western vs Chinese vendor comparison.
==============================================================================

The DC-DC Transition series · Part 17 of 19








The DC-DC Transition · Supplement D



The Solid-State Transformer: Layer 2 supplier dynamics


The solid-state transformer (SST) is the consequential piece of equipment in the 800V DC architecture. And the layer where vendor power is most concentrated in the six-layer stack. Four Western vendors (Eaton, Schneider, ABB, Hitachi Energy) plus three credible Chinese entrants (CRRC, BYD, TBEA) constitute the entire universe. Silicon carbide device concentration + engineering difficulty + capital intensity keep the field narrow. This supplement covers supplier dynamics, competitive positioning at each voltage class, and the strategic implications for operator procurement.





The DC-DC Transition · six supplements

 1. A. The Transition Tax
 2. B. Hybrid Forever
 3. C. The Voltage Ceiling
 4. D. The Solid-State Transformer (you are here)
 5. E. The Sidecar Power Rack
 6. F. The Connector Wars


Companion to the 13-essay main seriesStart with The Architecture Map.





Reader takeaways

SST vendor universeWestern: Eaton, Schneider, ABB, Hitachi Energy (Siemens later). Chinese: CRRC, BYD, TBEA. Total ~7 credible global suppliers.
ConcentrationTop 3 Western vendors: ~70% of Western-served AI DC design wins. Chinese: dominant in domestic + Southeast Asian markets.
Pricing powerSST pricing 2024-2026 held firm despite volume growth. Concentration + engineering barriers protect margins vs commoditising LV transformer market.
Design win dynamicsHyperscaler-scale deals often multi-vendor for supply security; mid-cap operators single-vendor for engineering efficiency.
Lead times12-24 months for large-scale units (2-5 MW). Chinese-domestic supply 8-15 months. Export controls + CFIUS scrutiny reduce global Chinese availability.
Innovation trajectory800V products mature. 1500V products in early production 2026-2028. MVDC in R&D + specific utility deployments; Chinese vendors ahead.






01What SST does that legacy transformers do not


A conventional transformer performs one function: voltage step-down from MV AC to LV AC, at a fixed turns-ratio, with electrical isolation. It does nothing else. Downstream from a conventional transformer, additional equipment is required to rectify AC to DC (if DC is the ultimate distribution). This adds equipment stages, equipment losses, and specific reliability failure modes.





Figure D.1
Solid-state transformer power flow. From MV AC to DC bus

MV AC input13.8-33 kVSST high-voltage stageRectify + isolateDC-DC conversionRegulate to 800V or 1500V DCDC output800V or 1500V DC busBuilding distributionLayer 3 downstream







An SST integrates rectification + voltage regulation + isolation in a single equipment stage using silicon carbide power electronics. Directly converts MV AC input to specified DC output voltage. Voltage regulation is active (not fixed-ratio like conventional transformer). Fault interruption is inherent in the power-electronic stage rather than requiring downstream mechanical breakers. End-to-end efficiency 2-4 percentage points above LV transformer + rectifier stack.




02Vendor landscape and concentration





Figure D.2
SST vendor landscape: Western + Chinese by product readiness

800V product1500V roadmapMVDC R&DAI DC design winsGlobal availabilityEatonYes2027YesMultipleGlobalSchneider ElectricYes2026-27YesMultipleGlobalABBYes2027YesGrowingGlobalHitachi EnergyYes2028YesUtility-adjGlobalSiemensLater2028+YesLimitedGlobalCRRC (CN)YesAvailableYesCN domesticCN+SEABYD (CN)YesAvailableR&DCN domesticCN+SEATBEA (CN)YesAvailableYesCN domesticCNIntensitylow → high








Chart 1. Western vs Chinese SST vendor market share 2026 (est)


Western vendors dominate outside China: Eaton ~22%, Schneider ~18%, ABB ~16%, Hitachi Energy ~12%, Siemens ~5%. Chinese vendors dominate domestic + SEA: CRRC ~40%, BYD ~25%, TBEA ~15%. Global shares differ materially by geography.






03Why the field stays concentrated





Figure D.3
SST value proposition vs execution difficulty

SST advantageWhy AI DC operators want SSTDirect MV to DC in one equipment stageHigher end-to-end efficiency (2-4 pp above LV transformer stack)Voltage regulation + rectification integratedFault interruption faster than mechanical breakersEnables 1500V DC + higher without step-changeSST challengeWhat makes SST hardSilicon carbide devices at MV expensiveThermal management at high current densityInsulation coordination at MV DCLong lead times (12-24 mo for large units)Vendor concentration limits pricing pressure








Chart 2. SST manufacturing capex + engineering entry barrier


Entering credible large-SST manufacturing requires: ~$300-800M facility investment, 5-8 years engineering + design certification, specific SiC device supply relationships (Wolfspeed, ST, onsemi), test + validation infrastructure. Total 8-12 year investment horizon deters new entrants.






04Chinese vendor competitive dynamics


CRRC + BYD + TBEA benefit from three specific advantages: (1) domestic scale from Chinese AI infrastructure buildout provides amortisation base, (2) industrial-policy support keeps capital costs low, (3) SiC device supply from Chinese domestic sources (Sanan, TianKe) reduces cost + supply constraint vs Western supply chain. Global export access constrained by CFIUS scrutiny + specific export-control considerations, but Chinese domestic + Southeast Asian + Middle East + African markets substantial + growing.




Chart 3. Chinese SST vendor capacity + geographic access 2026


Chinese vendors have ~40 GW aggregate annual manufacturing capacity (2026), growing to ~80 GW by 2028. Geographic access: dominant China + growing SEA + material Middle East + emerging Africa. Restricted from most Western markets by regulatory + national-security considerations.






05Operator procurement implications


For hyperscaler-scale deployments: dual-source procurement (typically Western + Chinese where geographically permissible; two Western vendors otherwise) for supply security. For mid-cap operators: single-vendor typical for engineering efficiency + volume pricing. For AI-native GPU cloud operators: often single-vendor determined by contract-collateralised financing lender preferences (see Fin I DDTL).




Chart 4. Operator procurement patterns by scale + type


Hyperscaler: dual-vendor or triple-vendor. Mid-cap operator: single-vendor with contingency. AI-native GPU cloud: single-vendor lender-preferred. Traditional colo: single or dual depending on refresh cadence. Patterns reflect scale + engineering resource + capital-structure constraints.






Method and sources. Public information only. Vendor product datasheets + investor commentary 2024-2026. Yole Group + OMDIA power-electronics market analyses. Industry reporting on SST design wins + capacity.


Series footer. Supplement D to The DC-DC TransitionRelated: Supplement C (voltage ceiling context), DC-DC XI (vendor economics), AI Power ChainHub: AI Power Semi hub.










Primary sources

 * Eaton. Eaton SST product line
 * Schneider Electric. Schneider SST product documentation
 * ABB. ABB SST product documentation
 * Hitachi Energy. Hitachi Energy SST product line







The DC-DC Transition: start with the foundation

 1. The real reason data centres are going DC (four forces)
 2. Two architectures wearing the same name (800V DC vs OCP Mt Diablo)
 3. The six-layer architecture map


==============================================================================
# The Sidecar Power Rack: Layer 4 and the OCP-spec ecosystem (DC-DC Supplement E)
URL: https://adikumar.co/dc-dc-supplement-e-sidecar-power-rack/
Published: 2026-08-20
Summary: Deep-dive on sidecar power rack architecture for AI-density data centres. OCP Mt. Diablo v0.7.0 specification.
==============================================================================

The DC-DC Transition series · Part 18 of 19







The DC-DC Transition · Supplement E



The Sidecar Power Rack: Layer 4 and the OCP-spec ecosystem


The sidecar power rack is the specific Layer 4 architecture that OCP Mt. Diablo (v0.7.0, March 2026) established as the reference for hyperscaler AI-density deployments. Physical separation of rack-level power conversion (800V DC to 48V DC) from the IT compute rack. Enables specialised thermal design, shared power across multiple IT racks, and service without IT downtime. The OCP-spec ecosystem. Wiwynn, Foxconn, Delta, Advanced Energy, plus specific power module + connector vendors. Competes intensely on this specification. This supplement covers what the sidecar rack actually does and the vendor landscape shaping the ecosystem.



The DC-DC Transition · six supplements
 1. A. The Transition Tax
 2. B. Hybrid Forever
 3. C. The Voltage Ceiling
 4. D. The Solid-State Transformer
 5. E. The Sidecar Power Rack (you are here)
 6. F. The Connector Wars

Companion to the 13-essay main seriesStart with The Architecture Map.



Reader takeaways

Sidecar architecturePhysical rack adjacent to but separate from IT compute rack. Takes 800V DC input, outputs 48V DC to multiple IT racks. Enables shared power + specialised thermal + service without IT downtime.
OCP Mt. DiabloReference specification v0.7.0 (March 2026). 80+ ecosystem partners. Reference for hyperscaler AI-density deployments 2027 onward.
Vendor ecosystemRack OEMs: Wiwynn, Foxconn, Inventec, Quanta. Power modules: Delta, Advanced Energy, Lite-On, Bel Power, Flex, Murata. Connectors: TE Connectivity, Molex, Amphenol (see Supplement F).
Competitive dynamicsAnyone can build to OCP spec, and many do. High-competitive layer where operator procurement leverage exists.
Cost per kWSidecar architecture ~5-10% higher upfront capex vs integrated but wins on TCO through shared-power efficiency + serviceability.
Adoption trajectory2026: pilots. 2027-2028: hyperscaler default for new AI-density builds. 2029+: mainstream for mid-cap operators.






01What the sidecar power rack does


Traditional AI-DC rack architecture places 48V rectifier + PSU inside the IT compute rack itself, at the bottom. Power conversion + compute share the same physical enclosure + thermal envelope. Works fine at moderate density. At AI-density (100+ kW/rack), heat rejection from power conversion competes with heat rejection from compute for the same cooling infrastructure. Serviceability of power components requires IT downtime.





Figure E.1
The sidecar power rack. Physical separation of Layer 4 from IT rack

Layer 1Utility interfaceMV ACUnchangedPrimary vendors: Utility + interconnection OEMsLayer 2Perimeter conversionMV to 800V DCFeeds sidecarPrimary vendors: SST vendorsLayer 3Building distribution800V DC busRow-scale or centralisedPrimary vendors: Busway + switchgear vendorsLayer 4Rack-level: SIDECAR RACK800V DC input, 48V DC outputTHE sidecar power rackPrimary vendors: OCP-spec: Wiwynn, Foxconn, Delta, Advanced EnergyLayer 5IT rack (compute)48V busbar inputCompute racks receive 48VPrimary vendors: Rack OEMs: Supermicro, Dell, HPE, ODMLayer 6Point-of-loadsub-1V to GPUSilicon-integratedPrimary vendors: Vicor, Infineon, TI, ADIPower flow







Sidecar architecture solves this by physically separating power conversion. A dedicated sidecar rack (typically 42U or similar) sits adjacent to a group of IT racks (typically 2-6 IT racks per sidecar). Sidecar takes 800V DC from building distribution, converts to 48V DC (via isolated DC-DC modules), outputs 48V to a shared busbar that feeds the multiple IT racks. Sidecar has its own thermal envelope (typically air-cooled with dedicated fans + heat exchanger). IT racks are dedicated to compute + storage + networking, thermally optimised for those loads.




02Rack-level architecture comparison





Figure E.2
Rack-level power architecture comparison

AI density fitServiceabilityCost per kWOCP standard matchAdoption 2026Sidecar (OCP Mt. Diablo)HighHighBaselineYesGrowingIntegrated (48V in IT rack)MedMedLowerPartialLegacyDistributed PDULowHighHigherNoDecliningRow-scale converterMed-highLowLower at scaleEmergingNVIDIA refLegacy in-rack shelfLowLowHistoricOCP legacySunsetIntensitylow → high








Chart 1. Rack architecture capex + TCO comparison per MW


Sidecar: $180-220k/MW upfront, $18-22k/MW annual opex. Integrated: $155-190k/MW upfront, $22-26k/MW annual. Distributed: $200-240k/MW upfront, $25-30k/MW annual. Sidecar wins TCO by year 3-4 despite higher upfront.






03The OCP Mt. Diablo spec ecosystem


OCP Mt. Diablo v0.7.0 (published March 2026) specifies the sidecar rack physical dimensions, electrical interface, cooling interface, and management protocol. 80+ ecosystem partners have committed to build to spec or integrate with spec-conforming products. This creates a specifically competitive vendor layer. Any rack OEM can build a Mt. Diablo sidecar, any power module vendor can build modules to slot in, any connector vendor can meet the interface spec. Result: procurement leverage for operators buying sidecar systems is higher than at any other layer.





Figure E.3
Sidecar power rack signal flow. From building distribution to GPU

800V DC inFrom building distributionSidecar rectifier/converter stageIsolated DC-DC48V DC outputTo adjacent IT racksIT rack 48V busbarVertical distributionPOL to GPUSilicon-integrated








Chart 2. OCP Mt. Diablo ecosystem participation 2024-2026


2024: initial working group ~15 companies. 2025: v0.6 draft with 40+ committed. 2026: v0.7.0 published with 80+ ecosystem partners committed. Growing at ~50% annually with hyperscaler-scale operator demand pulling participation.






04Vendor landscape by module type


Within the sidecar ecosystem, four distinct vendor sub-segments compete. Rack OEMs (Wiwynn, Foxconn, Inventec, Quanta) build the physical rack + integration. Power module vendors (Delta, Advanced Energy, Lite-On, Bel Power, Flex, Murata) build the 800V-to-48V conversion modules that slot into the sidecar. Connector vendors (see Supplement F). Management + BMC software vendors integrate for monitoring + control.




Chart 3. Sidecar vendor market share by sub-segment (2026)


Rack OEMs: Wiwynn ~30%, Foxconn ~25%, Inventec 15%, Quanta 15%, other 15%. Power modules: Delta ~28%, Advanced Energy 22%, Lite-On 15%, others 35%. Fragmented + competitive across all sub-segments.






05Adoption trajectory 2026-2030


Sidecar adoption tracks hyperscaler AI-density deployment cadence. 2026: hyperscaler pilot deployments (Microsoft, Meta, Google). 2027: hyperscaler production deployments for new AI-density facilities. 2028: mid-cap operators begin adoption. 2029-2030: mainstream for new construction. Legacy integrated + distributed architectures continue in existing facilities but rare for new AI-density builds by 2029.




Chart 4. Sidecar rack adoption 2026-2030 (% of new AI-density DC construction)


2026: ~15% of new AI-density DC construction. 2027: ~35%. 2028: ~55%. 2029: ~72%. 2030: ~82%. Rapid adoption driven by OCP standards clarity + hyperscaler pull + vendor supply availability.






Method and sources. Public information only. OCP Mt. Diablo v0.7.0 specification March 2026. Vendor product documentation + investor commentary 2024-2026. Industry reporting on OCP ecosystem growth.


Series footer. Supplement E to The DC-DC TransitionRelated: Supplement F (connector interface deep-dive), DC-DC X (standards evolution), DC-DC XI (vendor economics)Hub: 800V DC hub.










Primary sources

 * OCP Rack + Power. Mt. Diablo v0.7.0 specification
 * Wiwynn. Wiwynn sidecar rack products
 * Foxconn. Foxconn rack integration







The DC-DC Transition: start with the foundation

 1. The real reason data centres are going DC (four forces)
 2. Two architectures wearing the same name (800V DC vs OCP Mt Diablo)
 3. The six-layer architecture map


==============================================================================
# The Connector Wars: Layer 4/5 interface engineering (DC-DC Supplement F, series close)
URL: https://adikumar.co/dc-dc-supplement-f-connector-wars/
Published: 2026-08-20
Summary: Deep-dive on high-current DC connector engineering for AI-density data centres. Layer 4/5 interface (sidecar to IT rack).
==============================================================================

The DC-DC Transition series · Part 19 of 19







The DC-DC Transition · Supplement F



The Connector Wars: Layer 4/5 interface engineering


The Layer 4/5 interface. Where the sidecar power rack meets the IT rack's 48V busbar. Is the connector-engineering choke point of the AI-density facility. 400+ ampere continuous current at 48V DC. Blindmate operation required. High cycle count expected. Failure mode is fire. Three vendors dominate (TE Connectivity, Molex, Amphenol) with a small pack of second-tier + specialty + Chinese-entrant alternatives. This supplement covers the specific engineering constraints + vendor competitive dynamics + operator procurement discipline required at this layer.



The DC-DC Transition · six supplements
 1. A. The Transition Tax
 2. B. Hybrid Forever
 3. C. The Voltage Ceiling
 4. D. The Solid-State Transformer
 5. E. The Sidecar Power Rack
 6. F. The Connector Wars (you are here)

Companion to the 13-essay main seriesStart with The Architecture Map.



Reader takeaways

Current + voltage envelope48V DC at 400-800+ ampere continuous. Peak current above 1000A during transients. Ambient temperature 30-45°C typical AI-DC deployment.
Failure consequenceContact resistance rise → arcing → thermal event → fire. Response time from onset to fire: seconds. Non-recoverable equipment damage typical.
Vendor concentrationTE Connectivity + Molex + Amphenol = ~75% of AI-DC 48V connector market. Second-tier (Delphi/Aptiv, Positronic, Anderson) ~15%. Chinese entrants ~10% + growing (mostly China + SEA).
OCP specOCP Mt. Diablo specifies connector interfaces + test procedures. All three top vendors certified against spec. Interoperability guaranteed between spec-conforming products.
Procurement disciplineTorque specifications critical. Cycling ratings must match operational profile. Regular inspection + re-tension mandatory in maintenance cadence.
Cost per rackLayer 4/5 connectors ~$3-8k per IT rack + sidecar mating. Small % of total capex but disproportionate consequence if under-specified.






01Where connectors matter (and where they don't)


Across the six-layer power chain, connector engineering matters at specific interfaces. Layer 1-2-3 primarily use bolted bus-bar connections (permanent + high current, engineered per installation). Layer 4/5 interface. Sidecar rack output to IT rack input. Is where blindmate connectors become critical. Layer 5-6 (in-rack to compute blade) uses blade-mounted connectors + backplane. Layer 6 (POL to silicon package pins) is silicon-packaging engineering, not connector-vendor territory.





Figure F.1
Where connectors matter most. Layer 4/5 interface (sidecar to IT rack)

Layer 1Utility interfaceMV AC (unchanged)Not connector-relevantPrimary vendors: Utility equipmentLayer 2Perimeter conversion800V DC outputBus-bar interfacePrimary vendors: SST vendorsLayer 3Building distribution800V DC busBolted connections primarilyPrimary vendors: Busway + switchgearLayer 4Rack-level: SIDECAR800V DC in / 48V DC outCRITICAL connector zonePrimary vendors: THE connector layer 4/5 interfaceLayer 5IT rack: 48V busbarVertical distributionBlindmate connectorsPrimary vendors: Blade connectors + backplaneLayer 6Point-of-loadSub-1V @ 1000+A to diePhysics constraintPrimary vendors: Silicon package pinsPower flow








02Vendor landscape at the critical Layer 4/5 interface





Figure F.2
High-current DC connector vendor landscape

48V high-current800V DC blindmateOCP-spec certifiedAI DC design winsGlobal supplyTE Connectivity600A+YesMultipleManyGlobalMolex600A+YesMultipleManyGlobalAmphenol600A+YesMultipleManyGlobalDelphi (Aptiv)400AEmergingSomeGrowingGlobalPositronic600A+ MILYesSelectiveDef+nicheGlobalAnderson Power400ALimitedSomeNicheUS+EUChinese entrants (est)400A+YesGrowingCN domesticCN+SEAIntensitylow → high








Chart 1. AI-DC 48V high-current connector market share 2026


TE Connectivity ~30%. Molex ~28%. Amphenol ~17%. Delphi (Aptiv) ~7%. Positronic ~5%. Anderson Power ~3%. Chinese entrants ~10% (mostly China + SEA). Top-3 vendor concentration high given engineering + certification depth required.






03Failure modes + mitigation





Figure F.3
Connector failure modes vs mitigation strategies

Connector failure modesWhat can go wrongContact resistance rise (heat, arcing, fire)Blindmate mis-alignment (mechanical damage)Cycling wear (limited insertion cycles at high current)Ingress: dust, moisture, contaminationVibration-induced looseningMitigation strategiesHow vendors + operators respondSilver-plated or gold-plated contactsPrecision mechanical alignment featuresCycling ratings certified (typically 500-10,000 cycles)Sealed connector housings + regular inspectionTorque-critical mounting + regular re-tension








Chart 2. Connector failure mode consequence + probability matrix


Bubble chart of failure modes: probability (x-axis) vs consequence severity (y-axis) vs bubble size (frequency in field). Contact resistance rise: moderate probability, extreme consequence (fire). Blindmate mis-alignment: moderate probability, high consequence. Others less severe.






04OCP Mt. Diablo connector spec + certification


OCP Mt. Diablo (v0.7.0, March 2026) specifies the physical + electrical + mechanical requirements for the Layer 4/5 interface connector. Test procedures include: (a) contact resistance under continuous current, (b) cycling endurance at rated current, (c) thermal cycling under load, (d) ingress protection (dust + moisture), (e) blindmate mis-alignment tolerance. All three top vendors have Mt. Diablo-certified products in market as of Q2 2026.




Chart 3. OCP Mt. Diablo connector spec certification status by vendor


TE Connectivity: multiple certified products. Molex: multiple certified. Amphenol: multiple certified. Delphi (Aptiv): certified. Positronic: some certified. Anderson: emerging. Chinese entrants: growing certification. Certification enables plug-compatibility across vendors.






05Operator procurement discipline


Procurement discipline at the connector layer is disproportionate to its share of capex. Under-specified connectors are the leading root cause of AI-DC facility electrical fires (per industry incident data). Best-practice procurement: (1) always specify to OCP Mt. Diablo Rev+; (2) require vendor cycling certification matching operational profile; (3) mandate torque specifications + validation on installation; (4) require regular inspection + re-tension in maintenance cadence; (5) never accept connectors that meet only voltage + current ratings without cycling + thermal certification.




Chart 4. AI-DC facility electrical fire incidents by root cause (2022-2026)


Connector-related: ~35% of incidents. Cable-related: ~20%. Battery-related: ~15%. Cooling-fluid electrical: ~10%. Other: ~20%. Connector under-specification + poor installation practice is the leading root cause. Procurement discipline is direct fire-risk mitigation.






Method and sources. Public information only. TE Connectivity + Molex + Amphenol product documentation. OCP Mt. Diablo v0.7.0 specification. Industry incident data (aggregated + anonymised) from insurance underwriter reports 2022-2026.


Series footer. Supplement F to The DC-DC TransitionSeries close. Companion: Supplement E (sidecar architecture), DC-DC IV (arc behaviour)Hub: 800V DC hubAll six supplements: /tag/dc-transition/.










Primary sources

 * TE Connectivity. TE Connectivity DC power connector line
 * Molex. Molex DC power connector line
 * Amphenol. Amphenol DC power connector line
 * OCP Rack + Power. Mt. Diablo connector specifications







The DC-DC Transition: start with the foundation

 1. The real reason data centres are going DC (four forces)
 2. Two architectures wearing the same name (800V DC vs OCP Mt Diablo)
 3. The six-layer architecture map


==============================================================================
# Equinix + CPP / atNorth $4B: what DD would flag on a Nordic data centre take-out
URL: https://adikumar.co/dealteardown-equinix-cpp-atnorth-aug-2026/
Published: 2026-08-19
Summary: Commercial DD scoping on the Equinix + CPP $4B take-out of atNorth. Nordic interconnection scarcity, PPA book, CoC exposure, joint acquisition governance.
==============================================================================

Deal tear-down · Issue 01 · commercial DD applied to a named M&A transaction



Equinix + CPP / atNorth $4B: what commercial DD would flag on a Nordic data centre platform take-out


Partners Group exits atNorth after building it into an eight-site, 1 GW+ Nordic data centre platform. CPP Investments takes 60 percent, Equinix takes 40 percent as strategic operator. The joint sponsor+strategic take-out structure is worth the DD walkthrough on its own, but the operating asset has three specific tests any acquirer at this scale needs to run.


Published 2026-08-19 · author's independent analysis, no advisory conflict on any named party





01The deal in one paragraph


On February 27, 2026, Equinix and Canada Pension Plan Investment Board announced a joint agreement to acquire Nordic data centre operator atNorth from Partners Group for an EV of US$4 billion. CPP Investments holds a 60 percent controlling stake; Equinix takes 40 percent as the strategic operating partner. Per Equinix’s disclosure, atNorth brings eight operational sites, an active development pipeline of approximately 800 MW expected over the next five years, and an additional 1 GW of secured power beyond that pipeline. The transaction is funded via a US$4.2 billion financing package and is expected to be immediately accretive to Equinix’s AFFO per share on close. Subject to regulatory approvals; expected close in second half 2026.




02Why this deal matters


This deal matters for three overlapping reasons. It is first a template for how PE-owned digital infrastructure gets transitioned during sponsor exit windows: Partners Group needed an exit, Equinix wanted the operating footprint without shareholder-dilutive capital, CPP could supply the capital duration Equinix could not, and pairing the two produced a bid Partners Group could accept. Expect the shape to recur through 2026-2027 as DigitalBridge, IPI, Blackstone, and KKR-owned platforms reach exit maturity. Beyond the structure, roughly 800 MW of active development pipeline plus an additional 1 GW of secured Nordic power across eight sites is a scarce asset in an interconnection-constrained market, and grid interconnection status per site (the workstream most acquirers under-weight) is where a lot of the strategic value in this transaction actually sits. The third reason is narrower and mostly technical: the immediate-accretion claim on Equinix’s AFFO is a specific representation that DD needs to check against a DD-normalised EBITDA walk rather than take from the CIM.



What Equinix specifically brings to the joint venture, beyond capital, is Platform Equinix and xScale: a global interconnection ecosystem with cloud on-ramps, network cross-connects, dense fabric routing, and existing hyperscaler qualification status across every named tier-one customer, plus the xScale product line built for hyperscale-scale colocation contracts (originally structured as a JV with GIC). Neither is available to a standalone Nordic operator. Plugging atNorth capacity into Platform Equinix commands premium pricing versus generic Nordic wholesale colocation, reduces the change-of-control re-qualification friction that a financial-buyer-only take-out would face, and accelerates hyperscaler pre-leasing on the 800 MW pipeline. That strategic uplift is the case for Equinix’s 40 percent stake, and it is what CPP as a pure financial owner cannot supply.




03What I would test first


If a sponsor asked me to scope commercial DD on this platform. Or if I were on the Equinix side testing the acquirer’s own thesis before signing. Here is my first-90-minute test list, ranked by risk-weighted urgency.



 1. Grid interconnection status per site (Regulatory + Manufacturing, DD Part XI + Part VI). Equinix’s disclosure separates ~800 MW of active development pipeline from an additional 1 GW of secured power beyond the pipeline. The DD test on each: for the pipeline, how much has a signed ISA versus land-only positions; for the secured 1 GW, how much is energised and drawing versus contracted-not-yet-energised versus optioned. Nordic queues (particularly in Sweden, Denmark, Norway) have tightened over the past three years, though 2026 has shown early signs of improvement in Sweden following Svenska kraftnät process reforms. An 800 MW pipeline that is 70 percent permit-ready with signed ISAs is a different asset from one that is 30 percent shovel-ready.
 2. Customer concentration and CoC exposure (Customer DD, DD Part IV). Nordic data centre customers skew hyperscaler and crypto-mining. Both segments carry specific risks. Hyperscaler contracts often carry change-of-control termination rights that can accelerate if the acquirer is unable to inherit the qualification status. Crypto customers carry price sensitivity to power cost movements and payment-cycle risk in downturns. Test: what percentage of trailing revenue and contracted future revenue is (a) hyperscaler subject to CoC review, (b) crypto or high-density compute, (c) enterprise colocation. The mix determines the risk-adjusted revenue base.
 3. PPA and power-cost structure by bidding zone (Legal + Capex, DD Part X + Part VIII). Nordic power markets are not monolithic. Northern hydro-surplus zones (SE1, SE2, NO3, NO4) and Iceland’s isolated geothermal grid clear at roughly €15-40/MWh through most of the recent cycle. Southern demand zones (SE3, SE4, NO1, DK1, DK2) saw wholesale prices spike into €80-150/MWh territory in 2022-2024. atNorth’s siting concentrates in the cheap-power zones (Iceland, northern Sweden), which is the platform’s underlying cost-of-electricity advantage. Legacy 2018-2022 PPAs at those zone prices sit at very different rates from PPAs available now. Test: PPA book by counterparty, bidding zone, tenor, and price; mark-to-market against the specific zone benchmarks; roll-off schedule; and merchant-market exposure in the specific zones each site draws from.
 4. QoE walk on Partners Group-managed EBITDA (QoE, DD Part VII). Sellers preparing for exit optimise EBITDA presentation. PE-owned targets in a competitive process routinely present CIM EBITDA that requires 8-15 percent normalisation to reach DD-defensible figures across the category. Add-back categories typical for data centre platforms: fit-out costs miscategorised as expense, capitalised interest treatment, stock-based compensation, and one-time facility commissioning costs. Test: full add-back walk. The atNorth-specific number depends on the QoE outputs; the category benchmark says a 5-10 percent adjustment range is where the walk typically lands, and the DD budget should be sized against that expectation rather than against the CIM headline.
 5. Development pipeline realism (Capex + Product, DD Part VIII + Part V). The “several developments” language covers a range from “shovel-ready with signed customer LOIs” to “land option with concept design.” The valuation likely embeds pipeline value. Test: for each development site, what is the permit status, customer commitment, capex schedule, and power interconnection status? Discount aggressively where any of the four is missing.
 6. Joint acquisition governance mechanics (Structure, DD Part XII). CPP 60 percent + Equinix 40 percent is a controlling+strategic structure that only works if the shareholder agreement is precisely drafted. Test: exit rights (tag/drag), buyout mechanics on Equinix if operating performance disappoints, capital call obligations, board composition, tie-breaker on capex authorisation. Bad drafting here creates the year-3 dispute that consumes management attention.
 7. EU regulatory approvals for a Canadian majority buyer (Regulatory, DD Part XI). CPP Investments is a Canadian pension fund taking control of a critical-infrastructure asset in multiple EU jurisdictions. EU FDI screening applies in Denmark, Sweden, Iceland, and potentially Norway depending on site locations. Timeline exposure: 3-9 months per jurisdiction, potentially requiring undertakings. Test: which jurisdictions have signed off, which are pending, and what is the fallback if any block?
 8. Financing structure and refinance exposure (Financing, DD Part XIII). The $4.2B financing package is larger than the $4B EV, implying some transaction costs and possibly a revolver capacity buffer. Test: term loan sizing vs DD-normalised EBITDA, coupon, tenor, covenant package, and refinance date. A 2029 refinance in a compressed European private credit market with tightening covenant regimes is a different exposure than a 2030 refinance at a rate that has come down.
 9. Platform Equinix strategic-commitment durability (Structure + Product, DD Part XII + Part V). If atNorth’s post-close value uplift depends on plugging into Platform Equinix and xScale. And per Section 02 it does, that is the strategic rationale for the 40 percent stake. The DD needs to understand how durable that integration commitment is under future Equinix leadership, strategy shifts, or portfolio rationalisation. Test: what is the contractual guarantee on the integration commitment (network peering, cross-connect enablement, cloud on-ramp access, xScale product roadmap alignment), what is the buyout mechanism if Equinix later exits the strategic partner role, and how does CPP protect against becoming the majority financial owner of a de-networked Nordic colocation platform if Equinix’s priorities move? This is a novel workstream this structure creates.




Chart 1. Nordic 1 GW power book: seller representation vs plausible DD decomposition


The Partners Group representation of "1 GW+ secured power" reads clean at headline level. Any plausible DD-lens decomposition against Nordic queue mechanics separates the number into three states with meaningfully different valuation implications. The blue bars show a defensible upper bound; the red bars show the pessimistic case DD needs to price.



Illustrative decomposition based on typical Nordic operator mix as of 2026. Actual decomposition requires site-by-site verification.





Chart 2. Nordic interconnection queue: application-to-energisation timeline, 2018-2026


Time from interconnection application to energisation has extended substantially across the four Nordic markets. Sweden and Denmark are now at the tightest end. Any development pipeline value that assumes 2020-vintage queue speed needs discounting to current queue speed.



Weighted average time from formal application to first energisation event. Sources: Svenska kraftnät, Energinet, Statnett, Fingrid annual reports 2018-2025.





04Questions I would want answered


 * What is the decomposition of “1 GW+ secured power” between energised, contracted-not-energised, and queue-application-stage?
 * What is the customer mix by segment (hyperscaler, crypto, enterprise colocation, other) as percent of TTM revenue and percent of contracted future revenue?
 * Which top-10 customer contracts contain change-of-control termination rights?
 * What is the PPA book by counterparty, tenor, price, and roll-off schedule?
 * What was the CIM-adjusted EBITDA and what is the DD-normalised figure after add-back scrutiny?
 * For each development site: permit status, customer LOI or contract, capex committed vs authorised, power interconnection status?
 * What are the governance mechanics between CPP and Equinix. Specifically buyout triggers, capex authorisation thresholds, and exit provisions?
 * Which EU FDI screening regimes have been notified, and what is the timeline?
 * What is the financing package structure and 2029 refinance covenant risk?
 * What is Partners Group’s post-close involvement (rollover, transition services, non-competes)?




Chart 3. AtNorth customer mix hypothesis and CoC re-qualification exposure by segment


Nordic operators typically show hyperscaler-dominant contracted revenue with a crypto tail and growing enterprise colocation share. Each segment carries a different CoC re-qualification risk profile that the DD needs to price. The radar overlays revenue concentration against re-qualification probability across five segments.



Segment shares and CoC-risk scoring are the author's DD-lens hypothesis. Actual mix requires verification against the seller-side data room.





Chart 4. Nordic power rates by bidding zone: cheap-power zones vs demand zones vs Iceland, 2018-2028E


Nordic power markets are not monolithic. Northern hydro-surplus zones (SE1/SE2, NO3/NO4) and Iceland cleared at €15-40/MWh through most of the recent cycle. Southern demand zones (SE3/SE4, NO1, DK1/DK2) spiked to €80-150/MWh in 2022-2024. atNorth’s siting concentrates in the cheap-power zones, which is the platform’s underlying cost-of-electricity advantage.



Weighted-average bidding-zone clearing prices. Sources: Nord Pool, Svenska kraftnät, Energinet, Statnett, Landsvirkjun 2018-2026.





05What I would flag


Most of the $4B EV is being paid for what atNorth has already secured on the utility side. In Nordic markets where interconnection currently runs three years or more in the most-constrained southern zones (though Svenska kraftnät recently reported roughly a 30 percent reduction in preliminary-assessment lead times), a portfolio of ~800 MW of active pipeline plus an additional 1 GW of secured power is a scarcity position that is genuinely difficult to replicate. The DD needs to decompose both figures separately. For the 800 MW pipeline: what percent has a signed interconnection service agreement, what percent is permit-in-progress, what percent is land option only. For the additional 1 GW: what percent is energised and drawing today, what percent is contracted-not-yet-energised, what percent is queue-application-stage. The state distribution at Nordic zone-specific power and colocation pricing swings the DD-defensible value of the platform by several hundred million dollars.



The immediate-AFFO accretion Equinix has guided to is worth stress-testing as a second flag. Equinix takes a 40 percent equity-method stake, and the accretion depends on how the joint venture is consolidated, what intercompany pricing applies on the colocation-services relationship between Equinix and atNorth, and how much operating uplift Equinix embeds from Platform Equinix integration. A DD-normalised version of the accretion calculation may sit below the guidance number, may match it, or may exceed it depending on how those three inputs land; the point is that the DD should build the walk from first principles rather than accept the accretion representation on face.



The third flag is governance. Joint sponsor-plus-strategic acquisitions like this one tend to strain three or four years in, and the strain is almost always capital allocation. A majority financial owner like CPP is running the platform for stable cash yield to service its beneficiaries. A strategic operator like Equinix wants growth capex to fund customer acquisition and route optionality. Both are defensible positions and both can be accommodated at signing when interests happen to align around growth. They diverge under pressure. A bad year on hyperscaler bookings, a Nordic power-price shock, a permitting delay on a major site. And the party holding the tie-breaker vote on capex authorisation determines what the platform actually does. That clause is the piece of the joint venture agreement that most matters for how the platform performs from year three onwards, and it is nearly impossible to renegotiate once the platform is under stress.




Chart 5. Joint sponsor-plus-strategic governance: illustrative friction pattern from comparable transactions


Capital allocation disagreement between a majority financial owner and a minority strategic operator is a documented pattern in comparable JV structures. The chart shows an illustrative composite pattern derived from published post-close performance reviews of similar-shape transactions; it is not a prediction about the CPP/Equinix arrangement specifically. Whether the pattern applies here depends on how the shareholder agreement is drafted and how the two owners actually work together after close.



Composite pattern from published post-close performance reviews of comparable joint sponsor-plus-strategic acquisitions 2018-2025.





06What the market is getting right, and what it is getting wrong


The market has read this deal correctly as a scarcity-asset acquisition rather than a growth-story acquisition, which explains why trading multiples on it are broadly in line with recent comparable Nordic transactions on a per-MW basis. That much is well-priced. Two parts of the coverage are missing what actually matters, though. The renewable-power branding, where most of the reporting sits, is real but secondary; the deal is more usefully read as the template for how PE-owned digital infrastructure gets transitioned through sponsor exit windows into joint sponsor-plus-strategic hands, and it will be studied inside every DigitalBridge, IPI, Stonepeak, and Blackstone platform reaching year-five maturity over the next eighteen months. Separately, change-of-control exposure on the hyperscaler contract book is being under-covered in the analyst notes I have read, though the coverage that does exist tends to over-weight the risk. Hyperscaler procurement teams do re-qualify colocation providers on change of control, but the standard DD stress case (which for a comparable financial-buyer-only take-out I would size at five to fifteen percent of revenue over the two years post-close) narrows here because Equinix already sits on the approved-vendor lists at every named tier-one hyperscaler. The Platform Equinix qualification status is one of the reasons Equinix rather than a generic operator is the strategic partner on this deal.




07Structural insight


Joint sponsor-plus-strategic acquisitions like this one are going to be a common shape for large digital infrastructure deals over the next two to three years, for the simple reason that neither side can do the deal alone. Strategic operators like Equinix, Digital Realty, GDS, and NTT cannot fund 1-to-5 GW-scale platform take-outs from their own balance sheets without shareholder dilution that public equity holders will not accept. Pure financial sponsors like Blackstone, KKR, CPP, and Brookfield have the capital duration but need operating expertise to run these assets after close. Pairing the two solves both problems, so this shape recurs. For anyone running commercial DD on a digital infrastructure transaction over the next year, the practical implication for the workplan is that joint-acquisition governance mechanics, which have historically been treated as a legal-team item late in Stage 4, should move into the commercial DD scope in Stage 2. The operating alignment between the two owners is what determines whether the platform performs after close. The fourteen-workstream DD framework I used on Eaton/Boyd applies here almost unchanged; the weightings shift toward regulatory (EU FDI screening), structure (joint-acquisition governance mechanics), and grid interconnection (asset scarcity as the primary value driver).




Chart 6. Equinix + CPP / atNorth: $4B EV composition, DD lens on where the value actually sits


Grid interconnection scarcity accounts for the majority of the strategic value. Operating platform and customer book are the DD-verifiable base. Development pipeline is the value-at-risk that DD scrutiny will discount.



Value allocation is author's DD framework view. Purchase price allocation for accounting purposes will differ.






Glossary of terms used



AFFO
Adjusted Funds From Operations. REIT-industry cash-flow metric, roughly operating cash flow minus recurring maintenance capex. The metric Equinix guides accretion against.
CIM
Confidential Information Memorandum. Marketing document seller-side advisors prepare for prospective buyers, presenting the revenue trajectory, EBITDA, growth story, and rationale for sale.
CoC
Change of Control. Contractual clauses that trigger renegotiation, consent requirement, or termination when a counterparty is acquired or its ownership changes.
CPP
Canada Pension Plan Investment Board. Canadian sovereign-scale pension fund and the majority buyer on this transaction.
DD
Due Diligence. The workstream discipline of testing a target's claims before committing to a transaction.
EU FDI
European Union Foreign Direct Investment screening. Country-specific regulatory regimes that review acquisitions of critical infrastructure by non-EU acquirers.
EV
Enterprise Value. Purchase price plus debt assumed minus cash, the total value of the business excluding capital structure.
GW
Gigawatt. One thousand megawatts. Unit of electrical power capacity.
ISA
Interconnection Service Agreement. Contract between a customer and the transmission owner authorising a specific load or generation connection at a specific bus and MW rating.
JV
Joint Venture. Business arrangement where two or more parties pool resources for a specific project or platform, with a separate governance structure and defined economic sharing.
MW
Megawatt. Unit of electrical power. A large modern data centre draws tens to hundreds of MW.
PPA
Power Purchase Agreement. Long-term contract between an electricity generator and a buyer locking in a price per MWh over a defined term, typically 10-15 years.
QoE
Quality of Earnings. DD workstream that normalises reported EBITDA by reclassifying add-backs and adjusting for one-time items to reach a DD-defensible earnings figure.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.





Method. Public information only. No advisory relationship with any named party. First-90-minute DD scoping, not comprehensive diligence.


Sources.
Equinix newsroom announcement (Feb 27, 2026) ·
Equinix investor release ·
Data Center Knowledge on the transaction ·
TipRanks summary ·
PRNewswire full release



Series footer. This tear-down applies the fourteen-workstream framework from Due Diligence for the AI BuildoutRelated reading: Part IV (Customer / pricing / competitive) on CoC exposure, Part XI (Regulatory / CFIUS / environmental / cyber) on EU FDI screening, Part XII (Deal structure / R&W / escrow) on joint acquisition governance mechanics, Supplement B on the Eaton / Boyd end-to-end worked example. Companion context: The Investment Layer Part IV on hyperscaler capital sourcing, The AI Power Chain Part VI on modular datacenter developer stack economics.


Series. Deal tear-down applies the fourteen-workstream commercial DD framework to named M&A transactions. Companion series: Deal watch, which covers named capital events broadly (financings, backstops, JVs, strategic partnerships) using the operational and diligence lens. Issue 01 covers Equinix + CPP / atNorth, published mid-August 2026.


Written in a personal capacity. No advisory conflict on any named party. Nothing here is investment advice.













In brief


The Equinix + Canada Pension Plan + atNorth transaction announced August 2026 is a reported $4B commitment for a Nordic data centre platform. Sovereign-adjacent capital pairing with a specialised colocation operator to enter a lower-cost power geography maps to a specific PE playbook covered elsewhere in the Investment Layer series.


==============================================================================
# NVIDIA / OpenAI PORTS-Pike: what DD would flag on a $105B guarantee for a phased 4.25 GW Ohio campus
URL: https://adikumar.co/dealwatch-nvidia-openai-ohio-aug-2026/
Published: 2026-08-19
Summary: Operational DD lens on the NVIDIA-OpenAI PORTS-Pike deal. $105B guarantee + $1.5B equity + $4.2B AEP Ohio grid partnership for a phased 4.25 GW Ohio.
==============================================================================

Deal watch · Issue 01 · operational and diligence lens on named capital events



NVIDIA / OpenAI PORTS-Pike: what DD would flag on a $105B guarantee for a phased 4.25 GW Ohio campus


A chip vendor guaranteeing $105B of a customer's lease and power obligations, taking a $1.5B direct equity investment in the project SPV, and standing behind a $4.2B utility grid partnership. This is a chip vendor moving beyond compute supply into infrastructure financing at scale. Here is the operational DD lens most of the coverage is missing.


Published 2026-08-19 · updated 2026-08-19 · author's independent analysis, no advisory conflict on any named party





01The deal in one paragraph


On August 17, 2026, NVIDIA, OpenAI, and SB Energy jointly announced a large-scale AI data centre campus at the PORTS-Pike Technology Campus in Pike County, Ohio (a decommissioned uranium-enrichment site south of Columbus). SB Energy will build, own and operate the facility under a 20-year lease with OpenAI as the anchor tenant, initially deploying 4.25 IT-GW of AI compute capacity coming online in phases beginning 2028, with NVIDIA holding an option to expand the campus to 8 IT-GW as demand warrants. NVIDIA is guaranteeing up to $105B of conditional lease and power payment obligations, taking a $1.5B direct equity investment in SB Energy, and standing behind SB Energy's $4.2B partnership with AEP Ohio to build regional grid infrastructure. SB Energy plans to build at least 10 GW of new power generation to support the campus. Initial reporting (Bloomberg, WSJ, DCD, late July 2026) had contemplated an up-to-$250B guarantee before the structure was revised. A parallel financing conversation for chip supply had been reported at up to $350B.




02Why this deal matters


This deal matters for three connected reasons. The financing architecture is genuinely new: a chip vendor sitting simultaneously on the guarantee leg ($105B), the equity leg ($1.5B in the project SPV), and the utility-partnership leg ($4.2B with AEP Ohio) is a chip vendor that has moved past compute supply and into infrastructure financing at scale. The revision from an initial $250B guarantee contemplation to the announced $105B commitment within three weeks of the story becoming public is a live-price signal on how much risk NVIDIA's advisors were willing to underwrite; the specific cause is not publicly disclosed, and reading which of the plausible drivers (lender constraints, NVIDIA's own risk-appetite recalibration, project scope change, negotiation dynamics) was operative shapes what the flag actually means. And the campus is contractually exclusive to NVIDIA silicon per the joint announcement, which changes the fungibility calculus if OpenAI ever needs a substitute anchor tenant.




03What I would test first


If a sponsor asked me to scope diligence on the underlying platform. Not the financing structure but the campus operating asset it collateralises. Here is my first-90-minute test list, ranked by risk-weighted urgency. Every item is a stress-scenario worth modelling; none is a prediction that the deal will fail.



 1.  Phased delivery vs financing tenor and lease payment triggers (Financing DD, DD Part XIII). The campus is phased: initial 4.25 IT-GW coming online 2028, option to expand to 8 IT-GW. The financing package sits underneath a 20-year OpenAI lease. Test: what specific capacity milestones trigger lease payments and what capacity milestones trigger debt service; how are the two calendars synchronised; what happens if phase-1 slips 12-18 months (e.g. from 2028 to 2029/2030); is there a make-whole mechanism protecting the debt if delivery lags. This is the sharper version of the "10 GW takes 5-8 years" concern that older reporting framed.
 2.  Behind-the-meter power architecture (Capex + Market DD, DD Part VIII + Investment Layer Part V). SB Energy plans to build at least 10 GW of new power generation supporting the campus, in partnership with AEP Ohio for regional grid infrastructure. The near-term BTM answer is gas turbines; SMRs are a longer-horizon option (regulatory and construction timelines make them unlikely to contribute before 2030). Test: what is the mix of BTM gas generation, AEP Ohio grid capacity, dedicated PPA, and any longer-horizon SMR option? What is the emissions profile of the initial phase versus the option-expansion phase? What is the pass-through risk on fuel prices to the lease economics?
 3.  Hyperscaler-fungibility if OpenAI defaults (Customer DD, DD Part IV). The joint announcement states the campus will exclusively host NVIDIA AI compute. If OpenAI's credit deteriorates, could the campus realistically be repurposed for Microsoft, Amazon, Meta, or Oracle? Test: is the physical infrastructure (rack density, cooling architecture, network fabric topology, power distribution) fungible across hyperscaler AI compute demands, or is it tailored to OpenAI-specific cluster topology; what tenants are contractually eligible as substitutes under the lease; how large a retrofit would fungibility require; what is the fallback capital source for a retrofit if OpenAI defaults during phase-1?
 4.  NVIDIA combined structural exposure across three legs (Investment Layer Part IV + Part VI). The guarantee, equity, and utility-partnership legs are structurally different but concentrate on a single credit. Test: how do the three legs interact under downside scenarios; can any leg be called without triggering the others; how do rating agencies (S&P, Moody's) treat the combined exposure in NVIDIA's credit profile; what is the equity-holder cost-of-capital implication if any leg gets debt-equivalent treatment.
 5.  OpenAI standalone credit posture (Legal + Financing, DD Part X + Part XIII). OpenAI is a private company with rapid revenue growth but unresolved unit economics and a governance structure that has already been subject to public restructuring. Test: what is OpenAI's standalone senior unsecured shadow rating; how does that price into the underlying lease debt structure; how do the R&W insurance markets treat this exposure; what specific credit covenants sit inside the 20-year lease.
 6.  PJM interconnection status per AEP Ohio partnership (Regulatory + Utility, DD Part XI + Investment Layer Part V). The AEP Ohio $4.2B grid partnership meaningfully changes the PJM queue exposure story that early reporting emphasized. Test: how much of the required grid infrastructure is under signed interconnection service agreement with AEP versus still queue-application-stage; what is the AEP Ohio commitment structure. Take-or-pay, capex-share, some other shape; what happens if AEP hits its own capex constraints or regulatory approval delays.
 7.  Site environmental and land-use profile (Regulatory + Environmental, DD Part XI). PORTS-Pike is a decommissioned uranium-enrichment site with a specific environmental and remediation history. Test: what Phase I/II environmental sign-offs have been completed; what remediation obligations, land-use restrictions, and easements attach to the parcel from prior nuclear use; what surface-water and groundwater monitoring is required. NRC involvement becomes central only if the plan includes on-site nuclear generation (e.g. SMRs); otherwise the workstream is standard brownfield environmental DD at higher complexity than a greenfield.
 8.  Chip supply concurrency vs lease timing (Product + Manufacturing, DD Part V + Part VI). The parallel chip financing conversation (reported up to $350B) implies NVIDIA is on the hook for both the guarantee leg and the chip revenue that services OpenAI's ability to pay the lease. Test: what chip delivery schedule is embedded in the lease phase milestones; what happens contractually if chips slip; how is the risk of concurrent chip and grid slippage handled.
 9.  Cross-default and acceleration mechanics across financing legs (Structure + Financing, DD Part XII + Part XIII). Test: is there any cross-default mechanism connecting the $105B guarantee, the $1.5B equity investment, the AEP Ohio grid partnership, and the parallel $350B chip financing? An acceleration in any one leg under a stress case should be modelled against the others.
 10. Medium-voltage equipment supply (Manufacturing + Capex, DD Part VI + Part VIII). The campus needs MV equipment on a scale that today's vendor capacity (Vertiv, Eaton, Schneider, GE Vernova, Hitachi Energy) delivers on 24-to-36 month lead times. Test: are the specific vendor contracts signed, what allocation share has been secured, what is the fallback if a vendor cannot deliver on a phase-1 milestone.




Chart 2. PJM interconnection queue reality: 2026 composition and the AEP Ohio partnership context


PJM is holding a queue that has doubled in size since 2023, driven almost entirely by data-centre load. The AEP Ohio $4.2B partnership meaningfully changes the queue exposure story on this specific project because grid capacity is being purpose-built with utility cost-sharing rather than joining the standard queue. Other 2027-2028 projects without a comparable utility partnership face the queue in the shape shown here.



PJM queue data from 2024-2026 quarterly reports. Data-centre share estimate from PJM 2026 LTLLR analysis.





Chart 3. Behind-the-meter power architecture: three plausible mixes across the phased 4.25 → 8 IT-GW campus


SB Energy plans at least 10 GW of new power generation. The near-term BTM answer is gas turbines with AEP Ohio grid partnership feeding the balance. SMRs are a longer-horizon possibility. The three scenarios below vary the mix, each carrying different emissions profile, fuel-price pass-through, and NRC involvement.



Author's illustrative scenarios. Actual composition awaits detailed SB Energy site disclosure. GW figures refer to power supply capacity, not IT compute capacity.





04Questions I would want answered


 * What specific capacity milestones trigger OpenAI lease payments, and how do those milestones synchronise with debt service on the $105B guarantee?
 * What is the BTM power mix (gas / SMR / PPA / AEP Ohio grid) at phase-1 versus at 8 IT-GW option-expansion?
 * Is the campus infrastructure fungible across other hyperscaler AI compute demands, or specifically tailored to OpenAI cluster topology?
 * What tenants are contractually eligible as substitutes under the 20-year lease if OpenAI defaults?
 * How do the $105B guarantee, $1.5B equity, and AEP Ohio partnership interact under downside scenarios; is there cross-default?
 * What is OpenAI's standalone senior unsecured shadow rating, and how does it price into the underlying lease debt?
 * How is the NVIDIA combined exposure treated by S&P and Moody's in NVIDIA's credit profile?
 * What triggered the initial $250B contemplation to be revised to $105B. Lender feedback, NVIDIA risk-appetite recalibration, project scope change, or negotiation dynamics?
 * What is the AEP Ohio commitment structure. Take-or-pay, capex-share, or some other shape?
 * Which MV equipment vendors have signed contracts, and what phase-1 milestone allocations are secured?
 * What Phase I/II environmental sign-offs and remediation obligations attach to the PORTS-Pike parcel?




Chart 4. Medium-voltage equipment lead times 2020-2026: the constraint the phased buildout has to clear


Lead times for transformers and switchgear in the MV range have extended by a factor of three since 2020. A 4.25 IT-GW phase-1 buildout (roughly 5-6 GW of MV equipment capacity when generation and distribution are included) needs equipment allocations on a scale that today's vendor capacity delivers on 24-to-36 month lead times. Equipment classes shown are LPT, MV switchgear, HVDC converters, and standard rack PDU / UPS.



Composite lead-time indices from Vertiv, Eaton, Schneider, GE Vernova, Hitachi Energy public commentary and analyst reports 2020-2026.





Chart 5. Phased delivery vs lease payment trigger: where the cash-flow gap opens under stress


The base case follows the announced schedule: initial 4.25 IT-GW online 2028, option expansion to 8 IT-GW by ~2030-2032. Under an 18-month stress-case delay, phase-1 shifts to 2029/2030 while the financing package continues to accrue costs. The gap between financing calendar and delivery calendar under stress is where the cash-flow problem lives, and it is the specific question DD needs to model.



Base case tracks the announced phased schedule. Stress case is an illustrative 18-month delay scenario; not a prediction.





05What I would flag


The largest single exposure sits at the phased-milestone level rather than at the gross build-time level. Phase-1 delivering 4.25 IT-GW by 2028 depends on the AEP Ohio grid partnership executing to schedule, on SB Energy's BTM power generation coming online in parallel, on MV equipment vendors delivering against 24-36 month lead times booked now, on chip supply concurrency, and on brownfield permitting on the PORTS-Pike parcel. Any of those slipping under stress translates to lease-payment triggers not firing on time while debt-service on the underlying financing accrues. The specific cash-flow gap size depends on how the lease and financing calendars are structured contractually; the DD scoping question is exactly that.



The correlated exposure across NVIDIA's three financing legs is being under-treated in public reporting. The $105B guarantee, the $1.5B direct equity investment in SB Energy, and the parallel chip financing conversation ($350B, reportedly) all concentrate on the same OpenAI credit. A downside case that stresses OpenAI stresses all three concurrently. NVIDIA's aggregate exposure to a single-customer credit event should be added together for underwriting purposes rather than assessed piecemeal, and rating agencies will likely take that view.



Fungibility under an OpenAI default is a sharper question than the standard fallback-tenant test. The joint announcement states the campus exclusively hosts NVIDIA AI compute. That is fine as long as OpenAI performs. If OpenAI does not, the substitute tenant needs to be another NVIDIA-silicon customer at 4.25-8 IT-GW scale, which is a small pool (Microsoft, Meta, Amazon, Oracle, arguably one or two others). And the physical infrastructure may or may not be fungible across their compute architectures without retrofit capex. This is a workstream this specific structure creates.




06What the market is getting right, and what it is getting wrong


The market has read the structural point correctly: this deal is a precedent for how chip-vendor infrastructure financing gets structured, and every hyperscaler-adjacent conversation in Q4 will reference it explicitly. The three-legged structure (guarantee + equity + utility partnership) is genuinely new and worth studying. Two parts of the coverage are missing what actually matters, though. Framing the deal as "$120B backstop for a 10 GW site" (as some initial reporting did, and my own earlier writing repeated) conflates the 10 GW of planned power generation with the 4.25 → 8 IT-GW of contracted compute; those are different quantities with different DD implications. Coverage also tends to treat the AEP Ohio grid partnership as background context when it is arguably the piece that meaningfully reduces the PJM queue risk on this specific project. While raising a new question about AEP Ohio's own execution capacity. On OpenAI credit risk, coverage treats it as the primary sensitivity when it is one of several correlated sensitivities: if OpenAI credit deteriorates 18 months from now the debt is impaired in year three, but if phase-1 delivery slips 18 months the debt is impaired in year one and the recovery mechanism is nonexistent.




07Structural insight


The NVIDIA / OpenAI PORTS-Pike structure is the first public example of a chip vendor sitting on three different financing legs of a single customer's infrastructure at this scale: guarantee, direct equity, and utility partnership. If it closes as announced it establishes a template that will be copied. Expect similar shapes within nine months from AMD or Broadcom guaranteeing customer real-estate debt, from hyperscaler equipment vendors extending vendor financing to specific site builds, and from NVIDIA itself running additional deals of this shape for other named customers. The broader implication for how the AI Power Chain vendor set gets valued is that these firms are being pulled from equipment-supplier posture into infrastructure-financier posture, which is a different business shape and needs to be underwritten as such by anyone taking vendor equity exposure. For anyone running commercial DD on a hyperscaler-adjacent transaction in the next year, the practical shift is that phased-delivery-vs-lease-payment-trigger synchronisation. Historically treated as a legal-team item late in Stage 4. Moves into the commercial DD scope in Stage 2.




Chart 1. NVIDIA/OpenAI PORTS-Pike: financing structure across three legs plus the initial contemplation


The revised guarantee is $105B. The direct equity investment in SB Energy is $1.5B. The AEP Ohio grid partnership adds $4.2B on the utility side. The initial reporting had contemplated up to $250B. A parallel chip financing conversation had been reported at up to $350B. Aggregate correlated exposure to a single-customer credit event adds rather than netting.



Analyst view of aggregate correlated exposure. Public reporting treats the financings as independent transactions.





Chart 6. NVIDIA balance-sheet exposure vs market cap: how much of the equity story sits behind the OpenAI-linked commitments


A $105B guarantee against NVIDIA is a substantial share of the equity story even at current market cap; adding the $1.5B direct equity and any pro-rata attribution of the AEP Ohio commitment nudges the number higher. If rating agencies treat the guarantee as debt-equivalent, NVIDIA's cost of capital moves across the whole business.



Illustrative. Market cap is a moving reference; use the closest recent print for actual analysis.






Glossary of terms used



AEP
American Electric Power. Investor-owned utility operating in 11 US states including Ohio. AEP Ohio is the SB Energy partner on the $4.2B grid infrastructure commitment for the campus.
BTM
Behind-the-Meter. Power generation located on the customer side of the utility meter, not going through the public grid. For a 10 GW site, BTM gas turbines are the near-term technology answer to bypass PJM queue timelines.
DD
Due Diligence. The workstream discipline of testing a target's claims before committing to a transaction.
GW
Gigawatt. One thousand megawatts of electrical power. In this article, 10 GW refers to the planned power generation supporting the campus, distinct from IT-GW below.
HVDC
High-Voltage Direct Current. DC transmission at hundreds of kilovolts, used for long-distance and undersea power transmission and, increasingly, for high-density data centre power distribution.
IT-GW
IT Gigawatt. Power delivered to IT equipment inside a data centre. Always smaller than total site power supply because cooling, distribution, and infrastructure consume some of the incoming power. The campus is 4.25 IT-GW initial with option to expand to 8 IT-GW.
MV
Medium Voltage. Typically 1 kV to 35 kV. The voltage class connecting data centre power distribution to the utility grid.
MVA
Mega-Volt Amperes. Apparent power rating unit used for transformers and switchgear.
NRC
Nuclear Regulatory Commission. US federal agency regulating civilian use of nuclear materials and facilities. Relevant to the PORTS-Pike site only if the plan includes on-site nuclear generation such as SMRs.
PJM
PJM Interconnection LLC. Regional grid operator covering 13 US states plus DC, including Ohio where the PORTS-Pike campus sits.
PPA
Power Purchase Agreement. Long-term contract locking in electricity price per MWh over a defined term, typically 10-15 years.
R&W
Representations and Warranties. Seller statements about a target's condition. R&W insurance covers buyer's post-close claims arising from breaches.
SMR
Small Modular Reactor. Nuclear reactor design with output typically below 300 MWe, envisioned for co-located power generation at data centre campuses.
SPV
Special Purpose Vehicle. A legal entity created for a specific transaction or asset, ring-fenced from the parent's other operations. SB Energy owns the project SPV that NVIDIA has invested $1.5B in.



For the full corpus glossary of acronyms used across all essays, see adikumar.co/glossary.





Method. Public information only. No advisory relationship with any named party. First-90-minute DD scoping, not comprehensive diligence. All facts on the deal structure verified against the NVIDIA newsroom announcement of August 17, 2026, and cross-checked against Axios, CNBC, Interesting Engineering, TipRanks, and w.media coverage. Analytical framing and DD flags are the author's own.


Sources.
NVIDIA newsroom announcement (Aug 17, 2026) ·
Axios coverage ·
CNBC on the $105B financing (Aug 17) ·
w.media on the 4.25 GW campus ·
Interesting Engineering on 8 IT-GW total capacity ·
Bloomberg on the initial $250B talks (Jul 26) ·
WSJ / Yahoo on the revision from $250B



Series footer. This piece applies the framework from Due Diligence for the AI Buildout (fourteen essays plus three supplements). Related reading: Part XI (Regulatory / CFIUS / environmental / cyber) on the site-specific regulatory scope, Part XIII (Financing structure, covenants, closing risk) on the phased-delivery-vs-lease-payment analysis, Part IV (Customer / pricing / competitive) on single-customer concentration and fungibility risk, Part XII (Deal structure / R&W / escrow) on the three-legged structural exposure. Companion context: The Investment Layer Part IV on hyperscaler capital sourcing and Part VI on refinancing wall, The AI Power Chain Part IV on medium-voltage interconnect equipment lead times and Part VI on modular datacenter developer stack economics.


Series. Deal watch is a monthly-ish operational and diligence-lens read on named capital events reshaping AI infrastructure. Financings, backstops, JVs, strategic partnerships, and any structural transaction that matters even if it is not an acquisition. Companion series: Deal tear-down, which applies the fourteen-workstream commercial DD framework specifically to named M&A transactions.


Update note (Aug 19, 2026). This piece was revised after publication to reflect the NVIDIA newsroom announcement of August 17, 2026, correcting the guarantee figure to $105B (from an earlier $120B reference based on WSJ intermediate reporting), separating the 10 GW of power generation from the 4.25-to-8 IT-GW of compute capacity, adding the $1.5B NVIDIA equity investment in SB Energy, and adding the SB Energy + AEP Ohio $4.2B grid infrastructure partnership. Feedback from independent readers on the initial version drove reframing of the deliverability test around phased milestones, softening of causal attribution on the guarantee revision, integration of behind-the-meter power architecture depth, and addition of the hyperscaler-fungibility test.


Written in a personal capacity. No advisory conflict on any named party. Nothing here is investment advice.













In brief


The NVIDIA/OpenAI PORTS-Pike guarantee announced August 2026 is a reported $105B commitment for a phased 4.25GW Ohio campus. The transaction structure raises DD questions across capital, interconnection, thermal supply chain, and long-dated performance covenants that the framework in this analysis would surface pre-signing.


==============================================================================
# July 2026: the month the grid became the ceiling: What Changed Issue 01
URL: https://adikumar.co/what-changed-july-2026/
Published: 2026-08-20
Summary: One-page monthly recap on AI infrastructure power. July 2026: what NVIDIA, OCP, hyperscalers, and power equipment vendors did last month, what the market…
==============================================================================

What changed · Issue 01 · AI infrastructure power



July 2026: the month the grid became the ceiling


A one-page monthly recap on AI infrastructure power. What NVIDIA, OCP, hyperscalers, and power equipment vendors did last month, what the market is getting wrong, and what I expect over the next six months.


Published 2026-08-19 · covers July 2026





01NVIDIA


Two moves shaped the month. On the demand side, NVIDIA entered talks to backstop up to $250B of OpenAI’s lease on the 10 GW SB Energy site in southern Ohio. A full project cost above $500B if it lands, and the largest single data centre programme ever announced. On the deliverability side, the 800 VDC ecosystem crossed 31 named partners (Vertiv, Delta, TI, Schneider, Eaton among them) with second-half shipping commitments. Taken together, NVIDIA is now underwriting both the compute buyer and the power distribution stack that gets electrons to the chip. This is a substantially different business shape than pure GPU supply, and the equity market has not yet repriced it.




02OCP


Mt. Diablo is now the reference implementation for rack-level DC power in the hyperscaler world. The Google-Meta-Microsoft authored specification pushes distribution from 48 VDC to ±400 VDC or 800 VDC, targeting between 100 kW and 1 MW per rack. What matters is that the open hyperscaler-driven standardisation track (OCP) and the proprietary chip-driven track (NVIDIA) have converged on the same voltage, which is unusual and is what makes 800 VDC bankable as 2027 equipment capex rather than a research trajectory. The vendor set that gets this right is the one taking orders in Q4.




03Hyperscalers


Q2 2026 earnings put combined 2026 big-four capex at roughly $725B, up 77 percent from 2025. Alphabet raised guidance to $175-205B and the stock fell seven percent on July 24. AWS is guiding to around $200B, Microsoft to around $190B, Meta raised guidance twice to $125-145B. The pattern in the sell-off is worth reading closely because investors were not pushing back on the demand thesis; they were asking whether the physical infrastructure could absorb the spend on the timeline the guidance implied. That question, not the compute-demand question, is the one that determines whether 2027 revenue actually shows up.




Big-four hyperscaler capex, 2024 → 2026


Combined capex up 77 percent year-over-year. The $315B increase over 2025 is roughly 3× the total 2019 capex of the same four firms.



Company guidance as of Q2 2026 earnings.





04Power equipment


Vertiv had a substantial Q2. Guidance raised to $13.8-14.2B revenue and EPS $5.82-5.92, the $600M buyback completed, a Bitzero AI/HPC partnership announced, and backlog above $15B. The 800 VDC portfolio ships in H2 aligned to NVIDIA, and the PurgeRite integration produced the NearZero fluid management service, which is commercial with hyperscaler case studies quoting 78 percent water-consumption reduction. Eaton and Schneider ran the same shape of Q2 (backlog compounding, pricing power holding, capacity build the operative constraint), which reads as a group that is short of factory floor space rather than short of orders.




05What the market is getting wrong


Consensus in the AI infrastructure trade prices chips first, then power, then land, in that order. What the July signals are collectively saying is that the sequence should be inverted. Four days after month-end, on August 3, Texas Governor Abbott ordered a full audit of the ERCOT interconnection queue and paused Batch Zero. ERCOT is holding 474 GW of load requests, which is around five times its all-time record peak, and roughly 90 percent of that queue is data centres. This looks like a Texas problem in the reporting but MISO, PJM, and CAISO carry queues of similar structural shape with somewhat different regulatory postures, and any $725B combined big-four capex guidance number assumes interconnection availability that current queue mechanics cannot deliver on 2027 and 2028 timelines. Reading the market by chip supply misses the operative constraint, which is copper on transmission towers plus medium-voltage switchgear on the utility side of the point of common coupling. The Vertiv, Eaton, and Schneider order books are the leading indicator for how much of the guidance actually shows up as revenue; the hyperscaler capex line is a lagging one.




06What to expect in the next six months


 * OCP Global Summit (October). 800 VDC gets formal reference-design status. A second wave of vendor commitments follows within 30 days.
 * Hyperscaler Q3 earnings (October-November). 2027 capex guidance starts landing. Watch whether the anchor firms extend or hold the 2026 growth rate; extension implies confidence in grid-queue resolution, hold implies acceptance that 2027 is capacity-constrained.
 * ERCOT audit outcome (Q4). Signals the template other ISOs adopt. If Texas culls the queue by 30-50 percent, other regions will follow. If it merely reorders it, the queue remains the operative constraint.
 * Second thermal services roll-up. Expect a PurgeRite-shape acquisition by end-Q1 2027. Candidates: independent liquid cooling service specialists with hyperscaler exposure.
 * Private credit for medium-voltage capex. Utility-side transformer and switchgear capacity will get direct-lender financing packages this window. First deals of this shape close by year-end.
 * Grid-queue reform legislation. One serious federal-level proposal by Q1 2027 to standardise large-load interconnection. Its fate signals whether the constraint becomes structural or gets policy relief.




Method. Sourced from July 2026 earnings releases, industry reporting, and regulatory dockets. Prediction section is my own view.


Sources.
Data Center Knowledge, July highlights ·
Bloomberg on NVIDIA-OpenAI Ohio ·
OCP Mt. Diablo ·
CNBC on hyperscaler capex reaction ·
Vertiv Q2 2026 ·
Utility Dive on ERCOT 474 GW queue



Series. What Changed is a monthly one-page recap. Issue 01 covers July 2026, published mid-August. Next issue: August 2026, publishing early September.


Companion reading. The AI Power Chain (six-part technical series), The Investment Layer (eight-part capital-flow series), Due Diligence for the AI Buildout (fourteen-part diligence series plus three supplements).


Written in a personal capacity. Nothing here is investment advice.













In brief


July 2026 was the month the grid became the visible binding constraint on AI infrastructure buildout. Transformer lead times crossed 200 weeks. FERC interconnection queue reform proceedings picked up. Hyperscaler nuclear PPA signings compounded (Microsoft/Constellation, Amazon/Talen, Google/Kairos). The pass-through effect on data centre economics is starting to show in colocation contract terms.