The AI Power Chain framework
The six engineering layers between grid interconnection and the accelerator die. What the framework decomposes, why it exists, how each layer maps to vendor concentration, and where the framework has been applied across the corpus.
The framework in one paragraph
The AI Power Chain runs from the utility service territory boundary to the point-of-load conversion at the GPU or accelerator die. The six intermediate layers cover grid interconnection, facility distribution, rack-shelf rectification and wide-bandgap conversion, on-package power delivery, thermal management, and modular build. Vendor concentration, supply-chain lead times, standards trajectory, and qualification cycles all vary layer by layer, none of them predictable from the others. Sponsors, operators, and hyperscaler strategy teams that treat the six as one commodity input systematically under-price recovery risk in specific layers and over-invest in others.
Why the framework exists
AI infrastructure discussion tends to collapse around three visible actors: chip companies, hyperscalers, and data centre operators. The physical infrastructure between the grid meter and the accelerator package sits invisibly under those three, treated as commodity input.
That treatment held when rack power sat below 30 kW and facility distribution standardised on 415V AC. It has stopped holding as rack power crossed 100 kW during the 2024-2025 GB200 NVL72 deployments, and as facility distribution began to fragment between legacy 415V AC, retrofit-oriented 800V DC per OCP Mount Diablo v0.5, and greenfield 800V DC per the Nvidia Blackwell and Rubin reference platforms. At that point the invisible infrastructure becomes a set of six distinct engineering problems, each with its own vendor set, its own qualification cycle, and its own lead-time exposure. The vendor consolidation on Layer 5 (Eaton acquiring Boyd Thermal in March 2026 for approximately $9.5B per company disclosures, plus Ecolab acquiring CoolIT from KKR for approximately $4.75B per Ecolab disclosures) is being priced by strategic buyers who already read the market this way. Independent sponsors and asset-level equity holders are largely still pricing the layers as commodity.
The framework sits in the same analytical tradition as classical value-chain decomposition in industrial-tech investing, updated for the specific supply-chain, thermal and standards constraints of accelerator-dense compute above 100 kW per rack. It was first developed for internal use during commercial + technical DD workstreams in 2024-2025 and codified in the architecture map essay in early 2026.
The six layers
Layer 1. Grid interconnection
Starts at the utility service territory boundary. Ends at the 34.5 kV or 132 kV medium-voltage bus inside the site substation. Scope covers transformer sizing, switchgear topology, MV cable ampacity, interconnection queue navigation and utility contracting.
Engineering constraint
Transformer lead times crossed 200 weeks in mid-2026 per industry disclosures. FERC-regulated ISO interconnection queues run multi-year in most territories. Substation build-out schedules are the binding constraint on gigawatt-scale AI infrastructure siting. Hyperscaler behind-the-meter power procurement (gas turbine, fuel cell, small modular reactor) has moved past the one-off-deal phase into a systematic strategy for side-stepping utility interconnection queues.
Vendor concentration
High and consolidating. Eaton closed Fibrebond in 2024 for approximately $1.4B, Ultra PCS in 2025 for approximately $1.55B, and Boyd Thermal in March 2026 for approximately $9.5B, all per company disclosures and all inside a 24-month window. ABB, Hitachi Energy, Siemens and Schneider hold the balance of the medium-voltage switchgear and transformer stack. Prysmian and Nexans dominate underground MV cable. Fibrebond and Compass ship pre-integrated substation skids.
Standards
IEEE and IEC-mature at the equipment level. AI-specific interconnection patterns (behind-the-meter generation, multi-tenant hyperscaler campuses, hyperscaler-nuclear PPA structures) are still being negotiated case-by-case with no standardised template yet emerging.
Framework calibrated on live deals
Case studies applying the physical scorecard alongside the Investment Layer verdict on 2024-2026 transactions.
- AirTrunk / Blackstone + CPP (Sep 2024) — per-layer scores put physical at ~22% (LOW-MEDIUM band). The clean-asset calibration anchor.
- Crusoe / Stargate Abilene — physical HIGH on Layers 1 and 3 (behind-the-meter gas; single-accelerator SiC concentration). The framework's Layer 1 grid-and-generation scenario, materialised.
- CoreWeave DDTL 3.0 — portfolio-averaged physical MEDIUM. Compute delivered from colocation footprint rather than greenfield; a debt-facility calibration.
Related work
Layer 2. Facility distribution
Starts at the MV bus inside the site substation. Ends at the rack input. Scope covers the low-voltage distribution architecture (415V AC, ±400V DC bipolar, 800V DC unipolar), the associated switchgear and protection, and the facility-wide backup and ride-through design.
Engineering constraint
This is the layer currently in architectural flux. The 415V AC-UPS architecture that carried the industry through the pre-AI era does not efficiently distribute the 100-500 kW/rack power density that GB200 NVL72 and forward-generation platforms require. Two 800V DC architectures are competing for greenfield and retrofit deployments, with different grounding schemes, fault-clearing behaviour and vendor stacks. Sponsors underwriting hyperscale campuses in 2026 are choosing between locking in a bipolar retrofit path (OCP Mount Diablo v0.5), a unipolar greenfield path (Nvidia platform), or committing to 415V AC for the remaining life of the incumbent equipment.
Vendor concentration
Concentrating around the 800V DC transition. Delta Electronics, Vicor, Vertiv and Schneider hold design-in wins on the OCP Mount Diablo v0.5 draft. Nvidia-platform 800V DC design wins are similarly concentrated. Legacy 415V AC-UPS vendors (Eaton, Vertiv, Schneider, ABB) retain the retrofit market for the pre-AI installed base.
Standards
In active development. OCP Mount Diablo v0.5 draft published March 2026. Nvidia 800V DC production specification live for Blackwell and Rubin platforms. IEC and IEEE working groups are running catch-up on unified standards, with practical convergence unlikely before 2028-2029.
Framework calibrated on live deals
Case studies applying the physical scorecard alongside the Investment Layer verdict on 2024-2026 transactions.
- AirTrunk / Blackstone + CPP (Sep 2024) — per-layer scores put physical at ~22% (LOW-MEDIUM band). The clean-asset calibration anchor.
- Crusoe / Stargate Abilene — physical HIGH on Layers 1 and 3 (behind-the-meter gas; single-accelerator SiC concentration). The framework's Layer 1 grid-and-generation scenario, materialised.
- CoreWeave DDTL 3.0 — portfolio-averaged physical MEDIUM. Compute delivered from colocation footprint rather than greenfield; a debt-facility calibration.
Related work
Layer 3. Rack shelf and wide-bandgap conversion
Starts at the rack input. Ends at the 12V or 48V rack bus that feeds the accelerator boards. Scope covers the rectification stage, the bulk capacitance for pulsed load ride-through, and the wide-bandgap semiconductor content that makes efficient high-voltage-to-low-voltage conversion possible at rack density.
Engineering constraint
The 800V DC facility transition compounds silicon-carbide MOSFET demand at the 1200V voltage class because rectification stages need higher blocking voltage than the incumbent 650V silicon. GaN HEMTs address the higher-frequency point-of-load conversion segments below 100V. The rack shelf itself is where the wide-bandgap adoption curve is currently steepest, and where the qualification cycles for hyperscaler platforms are longest.
Vendor concentration
Five vendors hold roughly 90% of merchant SiC revenue: STMicroelectronics, onsemi, Infineon, Wolfspeed and ROHM. Wolfspeed emerged from 2025 restructuring with expanded DC-oriented SiC capacity coming online at Mohawk Valley and JP1, repositioning it as a scaling data-centre SiC vendor. Infineon's $830M acquisition of GaN Systems (October 2023) consolidated the merchant GaN HEMT stack around Infineon, Navitas, EPC and Innoscience. Power Integrations remains merchant on the boundary between silicon and wide-bandgap.
Standards
Silicon-carbide qualification is mature at the device level under JEDEC and AEC-Q101 for automotive-adjacent applications. Data-centre-specific qualification is being negotiated bilaterally between hyperscalers and merchant vendors on multi-year cycles.
Framework calibrated on live deals
Case studies applying the physical scorecard alongside the Investment Layer verdict on 2024-2026 transactions.
- AirTrunk / Blackstone + CPP (Sep 2024) — per-layer scores put physical at ~22% (LOW-MEDIUM band). The clean-asset calibration anchor.
- Crusoe / Stargate Abilene — physical HIGH on Layers 1 and 3 (behind-the-meter gas; single-accelerator SiC concentration). The framework's Layer 1 grid-and-generation scenario, materialised.
- CoreWeave DDTL 3.0 — portfolio-averaged physical MEDIUM. Compute delivered from colocation footprint rather than greenfield; a debt-facility calibration.
Related work
Layer 4. On-package power delivery
Starts at the 12V or 48V rack bus. Ends at the 0.8V rail feeding the accelerator die. Scope covers the multi-phase point-of-load converter topology, the on-package (or vertical) placement of the converters, the on-die decoupling capacitance, and the backside power delivery network topology when present.
Engineering constraint
An 800W GPU package at 0.8V draws 1,000 amps. A 2,250W package draws 2,800 amps. Copper interconnect physics constrain how much of that current can flow through package substrate before ohmic loss becomes the dominant efficiency term. Vertical power delivery and backside power delivery network are the two architectural responses. Both require converter topology and packaging co-design that only a small number of merchant vendors can execute at hyperscaler qualification volumes.
Vendor concentration
Concentrated. Monolithic Power Systems holds the largest merchant share of hyperscaler design-in wins for point-of-load converters. Vicor holds the vertical-power position with its factorised architecture and lateral-topology variants. Empower Semiconductor holds boutique positions on specific accelerator platforms. Analog Devices and Texas Instruments retain merchant share on the boundary with board-level multi-phase.
Standards
Little formal standards activity. The interfaces are proprietary to accelerator platforms. Merchant vendors compete on qualified reference designs and per-platform integration engineering rather than on common standards.
Framework calibrated on live deals
Case studies applying the physical scorecard alongside the Investment Layer verdict on 2024-2026 transactions.
- AirTrunk / Blackstone + CPP (Sep 2024) — per-layer scores put physical at ~22% (LOW-MEDIUM band). The clean-asset calibration anchor.
- Crusoe / Stargate Abilene — physical HIGH on Layers 1 and 3 (behind-the-meter gas; single-accelerator SiC concentration). The framework's Layer 1 grid-and-generation scenario, materialised.
- CoreWeave DDTL 3.0 — portfolio-averaged physical MEDIUM. Compute delivered from colocation footprint rather than greenfield; a debt-facility calibration.
Related work
Layer 5. Thermal management
Starts at the accelerator die-attach layer where heat is generated. Ends at the facility rejection layer where heat leaves the site. Scope covers cold-plate design, coolant chemistry, coolant distribution unit (CDU) architecture, secondary-loop separation, and the choice between direct-to-chip liquid cooling and full-rack immersion.
Engineering constraint
Air cooling stops being physically viable above roughly 100 kW per rack at any reasonable facility PUE. Direct-to-chip liquid cooling with single-phase coolant becomes the default for AI training racks in the 100 to 400 kW/rack range. Two-phase immersion extends the density curve past 500 kW/rack for Rubin Ultra and custom silicon workloads. Coolant chemistry sourcing, CDU architecture qualification, and multi-year approval cycles for hyperscaler-qualified quick-disconnects are the current supply-side chokepoints.
Vendor concentration
Reshaping through 2024-2026 M&A. CoolIT held the largest merchant share of hyperscaler DLC design-in wins pre-acquisition; Ecolab acquired CoolIT out of KKR for approximately $4.75B in March 2026. Motivair took a majority position with Schneider Electric in 2024. Boyd Thermal was acquired by Eaton for approximately $9.5B (closed March 2026). Vertiv built cooling capability through capability tuck-ins including Strategic Thermal Labs and PurgeRite rather than through a headline DLC acquisition.
Standards
OCP Sustainability Project drove quick-disconnect and cold-plate interface standardisation across 2024-2026. Coolant chemistry (particularly PFAS-based dielectrics used in two-phase immersion) is subject to active regulatory pressure in the EU and California.
Framework calibrated on live deals
Case studies applying the physical scorecard alongside the Investment Layer verdict on 2024-2026 transactions.
- AirTrunk / Blackstone + CPP (Sep 2024) — per-layer scores put physical at ~22% (LOW-MEDIUM band). The clean-asset calibration anchor.
- Crusoe / Stargate Abilene — physical HIGH on Layers 1 and 3 (behind-the-meter gas; single-accelerator SiC concentration). The framework's Layer 1 grid-and-generation scenario, materialised.
- CoreWeave DDTL 3.0 — portfolio-averaged physical MEDIUM. Compute delivered from colocation footprint rather than greenfield; a debt-facility calibration.
Related work
Layer 6. Modular datacenter
Starts at the facility building envelope. Ends at the site perimeter fence. Scope covers prefabricated power skids, cooling modules, structural enclosures, and the on-site assembly process that turns factory-built components into a commissioned data centre.
Engineering constraint
The dominant constraint on hyperscale AI capacity today comes from the qualified craft-labour base of medium-voltage electricians, industrial mechanical contractors, and switchgear commissioning specialists available to build gigawatt-scale sites on a two-year cycle. That craft-labour bottleneck now binds harder than chip supply for most 2027-2028 hyperscale build schedules. US Bureau of Labor Statistics numbers on qualified MV electricians suggest the skilled labour base is undersupplied by a factor of two to three against announced 2028 hyperscale build pipeline. Modular prefab collapses the on-site labour requirement per gigawatt of installed capacity by two to three times, which is the primary economic case for the layer.
Vendor concentration
Growing. Fibrebond (Eaton) ships pre-integrated power skids at scale. Compass Datacenters and Vantage ship prefabricated buildings and cooling modules. Vertiv, Delta, Schneider and specialist integrators ship modular power and cooling components. The hyperscaler-captive stack (Microsoft's Ballard-adjacent capacity, Meta's internal modular teams) is a growing share of the total buildout.
Standards
Limited standardisation across vendors. Interface specifications for water, power and control connections at the skid boundary are the leading engineering constraint on multi-vendor commissioning and drive the qualification cycles.
Framework calibrated on live deals
Case studies applying the physical scorecard alongside the Investment Layer verdict on 2024-2026 transactions.
- AirTrunk / Blackstone + CPP (Sep 2024) — per-layer scores put physical at ~22% (LOW-MEDIUM band). The clean-asset calibration anchor.
- Crusoe / Stargate Abilene — physical HIGH on Layers 1 and 3 (behind-the-meter gas; single-accelerator SiC concentration). The framework's Layer 1 grid-and-generation scenario, materialised.
- CoreWeave DDTL 3.0 — portfolio-averaged physical MEDIUM. Compute delivered from colocation footprint rather than greenfield; a debt-facility calibration.
Related work
What the framework tells you
The framework produces three analytical claims that individual essays develop in depth. Vendor concentration is highest at Layer 1 (grid interconnection), Layer 3 (wide-bandgap) and Layer 5 (thermal), and lowest at Layer 6 (modular build). Value capture concentrates unevenly across the six layers, which the operator-level narrative that treats them as commodity fails to price.
Lead-time exposure is highest at Layer 1 (200+ week transformer lead times) and Layer 5 (multi-year qualification cycles for coolant chemistry and quick-disconnects). Layers 3 and 4 have supply-side capacity growing on a two-to-four year investment cycle. Layer 2 lead-time exposure is currently short but concentrating around the 800V DC transition. Layer 6 lead-time exposure is driven by skilled-labour availability rather than equipment supply.
Standards fragmentation is highest at Layer 2 (three competing distribution architectures), Layer 4 (proprietary per-accelerator interfaces), and Layer 6 (limited multi-vendor interface standards). Sponsors and operators making architectural decisions in fragmented layers commit to specific vendor stacks for the useful life of the equipment, and the cost of choosing wrong compounds over the full asset life.
The AI Power Chain Vendor Screen scores 58 named vendors across the six layers against moat, execution risk and chokepoint concentration. The scoring is the operational output of the framework applied to specific investible names.
What the framework does not do
The framework decomposes the physical infrastructure into engineering layers. It surfaces where value concentrates by layer and where lead-time and standards risks compound. Asset-level, layer-specific financial modelling requires separate machinery and sits in the Investment Layer essays.
The framework is a snapshot of the physical infrastructure architecture as of August 2026. It does not encode the software layer above the accelerator, the model-training economics that drive accelerator demand, or the customer-side capacity procurement decisions that shape the demand curve. Those sit in adjacent frameworks that were developed and applied separately.
The framework treats the six layers as sequential in the physical power path but not necessarily sequential in the operator or sponsor analytical process. Real DD workstreams cover the six layers in parallel with heavy interaction between Layer 2 (facility distribution architecture) and Layer 5 (thermal), and between Layer 1 (grid interconnection) and Layer 6 (modular build). The layer numbering is a physical-scope ordering, not an analytical priority order.
Where the framework has been applied
The framework structures the analytical content across the AI Power Chain series, the Investment Layer series, the Due Diligence for the AI Buildout series, and the DC-DC Transition series. The three cross-cutting essays inside the AI Power Chain series explicitly work across all six layers:
- The AI Power Chain Vendor Screen, 58 named vendors scored across all six layers
- Pricing Under Scarcity, where premium capture is compounding by layer
- The Services Inversion, where services multiples are converging with product multiples by layer
The Investment Layer series applies the framework to capital allocation decisions:
- The AI infrastructure M&A map, 2023-2026 transaction pattern mapped to the six layers
- The PE playbook, layer by layer, sponsor economics by layer
- Underwriting AI infrastructure debt, recovery risk by layer
The Due Diligence for the AI Buildout series operationalises the framework across sixteen DD workstreams, with several workstreams (product architecture and technology DD, manufacturing, IP and supply chain DD) explicitly organised by layer.
Framework evolution
Three open questions currently shape how the framework will evolve through 2027-2028. Each will be revisited as the underlying market data resolves.
The Layer 2 architecture question resolves as 800V DC deployment either converges on OCP Mount Diablo v0.5 bipolar, Nvidia unipolar, or splits durably along retrofit vs greenfield lines. Layer 3 and Layer 4 vendor concentration follow the Layer 2 outcome with a two-to-three year lag.
The Layer 5 versus Layer 6 boundary is dissolving as thermal capacity is increasingly integrated into modular prefab skids at the vendor level (Boyd/Eaton, Motivair/Schneider). Whether the two remain analytically separable at the framework level is an open question and will be revisited if the operational integration compounds.
The Layer 1 versus Layer 2 boundary is also under pressure as behind-the-meter power procurement (nuclear PPAs, on-site generation, hyperscaler-utility JVs) transitions from one-off transaction into repeatable strategy. The framework may absorb some of the Layer 1 scope into a new "generation" layer if the pattern compounds.