> ## Content Index
> Fetch the complete content index at: https://adikumar.co/llms.txt
> Use this file to discover other available public pages before exploring further.

# The architecture map (where every stage lives, who builds it, what's standardized) (3/13)
- URL: https://adikumar.co/the-architecture-map/
- Published: 2026-05-08T13:16:12.000Z
- Updated: 2026-10-07T05:39:14.000Z
- Description: The AI data centre power train is six stacks, not one. Where each stage lives, who builds it, what is standardised. Reference map for architects speccing new AI campuses.
- Author: Adi Kumar
- Tags: #Essay, The DC-DC Transition, AI, Energy, Standards, Nvidia, AI Infrastructure, 800V DC

[The DC-DC Transition series](https://adikumar.co/dc-dc-transition/) · Part 3 of 13 

Most operators making AI infrastructure decisions in 2026 are getting them wrong. The reason moves past what most commentary highlights.

The mistake is treating the architecture as monolithic; separable choices at each stage carry more of the signal. Picking the wrong voltage class or backing the wrong supplier is secondary; the monolithic view is what compounds. “We’re going 800 V” or “we’re committing to Mt. Diablo” reads as a single decision. In practice it is six decisions taken stage by stage. Each stage sets its own commercial logic; each has its own vendor set; each moves on its own timeline.

The smart operators choose *per stage*. Their Stage 2 decision (perimeter conversion) does not dictate their Stage 4 decision (rack-level conversion). Their Stage 3 decision (building distribution) is separable from their Stage 5 decision (in-rack busbar). They make commitments where commitments earn them something, and preserve optionality everywhere else.

The operators getting it wrong treat the headline voltage choice as if it determined every stage below it, when it does not. The cost of that mistake, measured in stranded capex, vendor lock-in, and capacity that cannot accept the next generation of accelerator, compounds over the 15-25 year operational life of the facility.

This essay is the framework. The essays and supplements that follow (on solid-state transformers, sidecar power racks, BESS coupling, connectors, and the value chain) each go deep on one of these stages. Before we get there, you need the map.

## A note before we start

This essay runs structural, not argumentative. The essay maps the stack so that when you’re reading any of the next ten essays, you know which stage the argument is happening at, and what the stages above and below are doing.

If the first section reads like reference material, that is intentional. Treat it that way. Bookmark it and come back to it.

## 1\. The six stages

FIGURE 3.1 · THE SIX-STAGE POWER STACK adikumar.co · essay 3/13 From the utility meter to the accelerator package. Six sequential conversion stages, each with its own voltage, vendor pool, and standards regime. STAGE CONVERSION VOLTAGE CONCENTRATION 01 · UTILITY Utility interface The data centre meets the grid at the substation. MV AC service 13.8 / 22 / 33 kV depending on region 13.8 kV utility side of the boundary REGULATED GE / Siemens / Hitachi 02 · PERIMETER Perimeter conversion Transformer or SST steps grid AC into the facility bus. MV AC → LV AC / DC the SST or classic transformer choice 480 VAC or 800 VDC the OCP LVDC standardisation lands here CONCENTRATED Vertiv / Schneider / ABB 03 · BUILDING Building distribution Bus or switchgear moves power to the compute halls. bus / cable / switchgear the argument about hall backbone ±400 VDC or 480 VAC bipolar vs unipolar decision made here COMPETITIVE Eaton / Legrand / Starline 04 · RACK-LEVEL Rack-level conversion Sidecar power rack or in-rack shelf lands power at 48 V class. AC or DC → 48 V class Mt. Diablo v0.7.0 reference lives here 48 V nominal OCP ORv3 HPR busbar interface CONTESTED 80+ OCP ecosystem partners 05 · IN-RACK In-rack distribution Busbar and connectors route 48 V to each blade slot. busbar + connectors high-current interface engineering 48 V, 600-1200 A connector wars sit at this interface COMMODITY TE / Molex / Amphenol 06 · POINT-OF-LOAD Point-of-load conversion Multi-phase VRM feeds the GPU core at sub-volt levels. 48 V → 0.8-1.2 V multi-phase step-down at the package 0.8 V, 1500+ A chosen by the silicon designer SILICON-LOCKED Vicor / Infineon / TI / ADI Six stages. Six commercial regimes. One asset. SOURCES · OCP LVDC standardisation (Aug 2026) · Mt. Diablo v0.7.0 · Nvidia developer blog on 800 VDC · adikumar.co · essay 3/13 

Figure 3.1\. The six-stage power stack. Each stage has its own vendors, standards, and competitive structure.

From the utility connection to the GPU pin, every AI data centre passes power through six distinct stages. The voltage transitions, the conversion stages, and the protection regimes are different at each.

**Stage 1 · Utility interface.** Where the data centre meets the grid. Medium-voltage AC delivery, with the exact voltage set by the utility and site (13.8 kV, 22 kV, and 33 kV are common in different markets). Ownership and operating responsibility typically change at the metering or interconnection boundary, but the exact demarcation is site- and utility-specific. This is also a common point for BESS interconnection in modern designs, although BESS can be coupled downstream depending on topology.

**Stage 2 · Perimeter conversion.** The first equipment inside the operator’s domain. In a sidecar architecture, this is a conventional step-down transformer to 480 VAC. In a native architecture, this is an MVAC-to-DC conversion block, implemented as a transformer-rectifier or, in some designs, a solid-state transformer (SST), producing the facility DC voltage. The 800 V class is the current OCP and Nvidia focus; higher-voltage designs sit within a separate voltage-ceiling question. The choice here sets the AC/DC conversion boundary and constrains the vendor and interface set downstream.

**Stage 3 · Building distribution.** The AC or DC distribution that moves power from perimeter conversion to the compute halls. In a native 800 VDC architecture, this is an 800 VDC hall or row backbone, and Stage 3 also brings a wiring-topology choice: 2-wire 800 V (positive and return) is the simpler interface; 3-wire bipolar (±400 V + midpoint) is the topology used by the current OCP Diablo 400 specification; a 2-wire 800 V option is also defined. The choice changes insulation-to-ground exposure, grounding, protection, and the rack-input interface. The 2-wire vs 3-wire decision at Stage 3 propagates directly into the Stage 4 rack-input specification and the Stage 5 protection scheme. In a sidecar architecture, Stage 3 is conventional 480 VAC distribution feeding each sidecar power rack, which handles the AC/DC step at Stage 4\. In legacy AC, this is the LV switchgear and busway. Topology matters enormously here for fault management, redundancy and BESS coupling.

FIGURE 3.2 · 2-WIRE VS 3-WIRE ARCHITECTURE adikumar.co · essay 3/13 Two 800 VDC building topologies, both defined in OCP Diablo 400 v0.7.0. Upstream source equipment, rack input, protection and grounding all change with the topology. 2-WIRE UNIPOLAR · 0 to +800 V single-polarity bus with a return conductor Utility service grid interconnection, voltage in utility AC Transformer + rectifier or SST skid galvanic isolation + AC-to-DC conversion 2-wire 800 VDC + PE output · Diablo v0.7.0 option 2-wire 800 VDC DC BREAKER +800V 0V ret RACK SHELF 2-wire input 800 → 48 V ORv3 HPR RACK SHELF 2-wire input 800 → 48 V ORv3 HPR RACK SHELF 2-wire input 800 → 48 V ORv3 HPR BOND BAR SOURCE SST or TR feeding 2-wire 800 VDC + PE RACK INPUT Single-polarity shelf, 2-conductor + PE PROTECTION DC breaker at source and each rack drop GROUNDING System-negative bonded to earth via bond bar INSULATION-TO-GROUND 800 V full-bus on the positive conductor 3-WIRE BIPOLAR · ±400 V split bus around a midpoint conductor Utility service grid interconnection, voltage in utility AC Transformer + rectifier or SST skid galvanic isolation + AC-to-DC conversion, bipolar output stage ±400 VDC + midpoint + PE output · Diablo v0.7.0 default 3-wire ±400 VDC BIPOLAR BREAKER +400V 0V mid \-400V RACK SHELF 3-wire input ±400 → 48 V ORv3 HPR RACK SHELF 3-wire input ±400 → 48 V ORv3 HPR RACK SHELF 3-wire input ±400 → 48 V ORv3 HPR IMD + Z-BOND SOURCE SST or TR feeding ±400 VDC + midpoint + PE RACK INPUT Bipolar ±400 V shelf, 3-conductor + PE PROTECTION Bipolar DC breaker at source and each rack drop GROUNDING Midpoint bonded, isolated, or impedance-grounded INSULATION-TO-GROUND 400 V per conductor to midpoint reference Same nominal voltage. Different source output, rack input and grounding equipment. The choice at Stage 3 propagates into the Stage 4 rack input specification and Stage 5 protection. SOURCES · OCP Diablo 400 specification v0.7.0 (March 2026) · IEC 61140 (grounding classes) · IEC 61557 (IMD / insulation monitoring) · adikumar.co 

Figure 3.2\. 2-wire vs 3-wire building topologies as system diagrams. Upstream utility and transformer/rectifier stages shown; source, rack input, protection and grounding equipment all change with the topology.

**Stage 4 · Rack-level conversion.** The step from building-distribution voltage to what the rack uses internally. In a native architecture with an 800 VDC backbone, Stage 4 is a DC/DC step closer to the rack (or, in newer reference designs, delivered as 800 VDC into an in-rack power shelf). In a sidecar architecture, Stage 4 is where the AC/DC conversion happens: a sidecar Mt. Diablo shelf takes 480 VAC and produces 800 VDC or ±400 V for the row. In legacy 48 V architectures, this conversion sat inside the IT rack via the power shelf. The disaggregation of this stage from the IT rack is one of the bigger structural shifts in the transition.

**Stage 5 · In-rack distribution.** The busbar that runs vertically through the IT rack, distributing power to each blade slot. The 50-54 V class remains the dominant intermediate rack distribution voltage in today's OCP architectures, even as 800 VDC moves upstream toward the rack. The 48 VDC busbar architecture established by OCP’s ORv3 spec has remained remarkably stable even as higher-voltage distribution moves upstream.

**Stage 6 · Point-of-load.** The final power-conversion stages between rack-level distribution and the accelerator’s operating rails. Modern accelerators can require hundreds of amps at sub-1 V core rails, and delivery involves multiple rails and intermediate conversion close to the package. This is where Vicor, Infineon, Texas Instruments and a handful of others fight a quiet but consequential battle. The decisions made here affect efficiency, cooling and silicon longevity.

The layers aren’t independent. A choice at Stage 2 (SST topology) constrains what’s possible at Stage 3 (which busbar architectures work) which constrains Stage 4 (which sidecar designs are viable). But the stages are *separable* in a way that the marketing doesn’t always make clear. An operator could combine an Nvidia-reference Stage 4 architecture with a Mt. Diablo-style Stage 3, provided the interfaces are engineered accordingly; whether that makes economic or operational sense is site-specific.

## 2\. The standards field (and why it’s fragmented)

FIGURE 3.3 · STANDARDS GOVERNANCE BY STAGE adikumar.co · essay 3/13 No single body governs the full stack. Five standards bodies split the six stages between them. Overlap is at the interfaces. 01 · UTILITYUtility interface02 · PERIMETERPerimeter conversion03 · BUILDINGBuilding distribution04 · RACK-LEVELRack-level conversion05 · IN-RACKIn-rack distribution06 · P.O.L.Point-of-load IECInternational Electrotechnical · DC SSTCommission voltage classes, distribution,PRIMARYPRIMARYPRIMARYSECONDARYn/an/aIEEEInstitute of Electrical and ElectronicsEngineers · bus, protection, DC UPSSECONDARYPRIMARYPRIMARYSECONDARYn/an/aULUnderwriters Laboratories · safety DCcertification, breakers, arc-flashn/aPRIMARYPRIMARYPRIMARYPRIMARYn/aOCPOpen Compute Project · rack, busbar,Mt. Diablo shelf, LVDCn/aSECONDARYSECONDARYPRIMARYPRIMARYn/aNEC / NFPANational Electrical Code · cabling,grounding, wet-locationSECONDARYPRIMARYPRIMARYSECONDARYSECONDARYn/a Primary jurisdiction (the body that publishes the operative standard) Secondary (touches the stage through a related spec) Not applicable at this stage SOURCES · IEC 62933 · IEEE 946 / 1584 · UL 2436 · OCP LVDC v0.5 (Aug 2026) · NEC Article 690/705 · adikumar.co 

Figure 3.3\. Standards governance by stage. No single body governs the full stack. In practice, the least mature required layer can set the deployment schedule.

FIGURE 3.4 · OCP LVDC · TWO DEPLOYMENT PATHS adikumar.co · essay 3/13 OCP Aug 2026 documents two deployment paths for 800 VDC. Path A retains existing 480 VAC building infrastructure and localises the AC/DC conversion. Path B moves the conversion to the facility perimeter. PATH A · EXISTING AC + SIDE POWER RACK 480 VAC building distribution retained; sidecar handles AC/DC A1 · UTILITY MVUtility MV servicegrid interconnectionA2 · MV→LV TRANSFORMERStep-down transformerMV to 480 VACA3 · 480 VAC DISTRIBUTION480 VAC building busexisting infrastructure preservedA4 · SIDE POWER RACKSidecar AC/DC conversionORv3 HPR spec, ±400 V or 2-wire 800 VA5 · 800 VDC TO COMPUTE800 VDC to compute rowsserved into ORv3 racksA6 · COMPUTE RACKCompute rackaccelerator-agnostic PATH B · DIRECT MVAC → 800 VDC FACILITY AC/DC conversion at the perimeter; 800 VDC through the hall B1 · UTILITY MVUtility MV servicegrid interconnectionB2 · TR / SST SKIDTransformer-rectifier or SSTMV AC directly to 800 VDC (facility-scale block)B3 · 800 VDC HALL BACKBONE800 VDC hall or row backbonefacility-wide DC distributionB4 · IN-RACK POWER SHELFIn-rack shelf800 VDC → 48 V step-downB5 · 48 V BUSBAR48 V rack busbarORv3 HPR interface into blade slotsB6 · COMPUTE RACKCompute rackaccelerator-agnostic SOURCES · OCP LVDC blog Aug 2026 (opencompute.org) · Diablo 400 v0.7.0 · Nvidia developer blog on 800 VDC · adikumar.co · essay 3/13 PATH A retains legacy 480 VAC. PATH B replaces AC building distribution with 800 VDC. 

Figure 3.4\. OCP LVDC deployment paths. Path A adds a side power rack to existing AC infrastructure; Path B converts at the perimeter and distributes DC.

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 stages. The standards field is fragmented along organizational lines that have nothing to do with the operator’s view of the system.

Five standards, certification and governance layers are particularly relevant through 2030:

**IEC** (International Electrotechnical Commission). Develops the international standards that influence voltage classification and equipment design. IEC 60038 defines the voltage classes (LVDC, MVDC, HVDC). Essay 1 uses this framework for the LVDC/MVDC/HVDC 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.** Publishes engineering practices and recommended standards for DC distribution. 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 covers DC switches and disconnects within its applicable scope. The standard is often referenced in 800 V discussions, though the specifics of scope, voltage range, and certification history matter. 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/NRTL certification is one of several inputs (alongside NEC and AHJ requirements, equipment listings, utility rules, and insurer criteria) that decide whether 800 V data centre topologies move from “engineerable” to “deployable at scale.”

**OCP** (Open Compute Project). Industry consortium that publishes open specifications and reference designs. OCP specs are adopted through voluntary interoperability commitments; legal certification stays absent. 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). The principal US electrical installation code framework. 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 may need to satisfy IEC-derived equipment requirements, applicable UL/NRTL certification requirements, NEC/AHJ installation requirements and OCP interoperability specifications, depending on jurisdiction and deployment model. Each body moves at its own pace. In practice, the least mature required layer can set the deployment schedule.

This is also why standards-extension work matters: potential extension of UL 98B beyond its current scope, IEC SC 8A working drafts and future revisions of IEEE 1709 are relevant to how higher-voltage DC systems move from engineering validation toward routine deployment. It’s not that the technology isn’t ready. Multiple bodies need to agree on what “ready” means before an installation can move from engineering validation to routine deployment.

## 3\. The vendor map by stage

The vendor field is heterogeneous. Each stage has its own competitive structure.

FIGURE 3.5 · THE VENDOR MAP BY STAGE adikumar.co · essay 3/13 Concentration runs U-shaped: silicon and utility ends captive, middle stages are procurable. Named vendor pools per stage, with a concentration marker. STAGE PRIMARY VENDOR POOL CONCENTRATION 01 · STAGEUtility interfaceGE Vernova · Siemens Energy · Hitachi Energy · ABBHIGH90%02 · STAGEPerimeter conversionVertiv · Schneider Electric · ABB · Eaton (SST: Delta, Bloom, StarPower)MEDIUM-HIGH75%03 · STAGEBuilding distributionLegrand · Starline · Vertiv · Eaton · Anord MardixMEDIUM55%04 · STAGERack-level conversion80+ OCP partners · Delta · Great Wall · Flex · Compuware · VertivLOW-MEDIUM45%05 · STAGEIn-rack distributionTE Connectivity · Molex · Amphenol · PositronicMEDIUM-HIGH70%06 · STAGEPoint-of-loadVicor · Infineon · TI · Analog Devices (post Maxim + Empower)HIGH90% SOURCES · OCP LVDC ecosystem census (Aug 2026) · issuer disclosures · adikumar.co supplier tracker CONCENTRATION SCALE · low < 40% share held by top three · medium 40-70% · high > 70% 

Figure 3.5\. The vendor map by stage. Stages near silicon and utility are concentrated; the middle stages hold the procurement leverage.

**Stage 1 (Utility interface).** Operator’s own engineering team plus the utility. The core relationship is with the utility; a conventional equipment supplier market sits elsewhere; competition appears mainly around interconnection equipment and BESS. 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.

**Stage 2 (Perimeter conversion / SST).** This is where concentrated supplier power emerges as the transition deepens. SSTs are technically demanding, capital-intensive and subject to significant qualification and reliability requirements. The Western supplier field includes Eaton, Schneider Electric, ABB, Hitachi Energy, Siemens, GE Vernova, Mitsubishi Electric and Vertiv. Their exposure differs, though: participation in the 800 VDC ecosystem should not be read as equivalent SST maturity. Supplement D covers the actual vendor-by-vendor SST classification. On the Chinese side, CRRC, BYD and TBEA have relevant power-electronics and transformer capabilities; Supplement D examines their specific positioning.

(*Author note for transparency:* I work at Eaton. I’ll say more below about how I’m handling that disclosure across this series.)

**Stage 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.

**Stage 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 stage where the OCP spec ecosystem creates direct vendor competition. The ORv3 HPR specification creates an open interface intended to allow multiple vendors to build compatible equipment.

**Stage 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 48 VDC busbar is mature technology; differentiation shifts toward mechanical integration, thermal performance, manufacturability and interface engineering.

**Stage 6 (Point-of-load).** A concentrated group of specialised power-management IC suppliers. Vicor, Infineon, Texas Instruments and Analog Devices, which completed its acquisition of Empower Semiconductor in July 2026\. This is a concentrated stage and the one most invisible to the operator buying conversation. The silicon and platform designers (Nvidia, Google, Amazon, Meta, Microsoft) determine much of the point-of-load architecture when they design the accelerator package; the operator generally inherits those requirements through the server or rack system it procures.

The pattern is worth noticing: the stages nearest the silicon (5 and 6) are concentrated and least visible to the operator. Stage 1 is constrained by the utility and interconnection process; a conventional supplier market plays a smaller role; Stage 2 is supplier-concentrated but visible. The middle stages (3, 4) are more directly contestable through procurement.

When the rest of this series talks about “where the value goes” in the DC transition, Stages 2, 3, and 4 are the main areas to examine for incremental value created by the DC transition. Stage 6 already has established semiconductor economics; Stage 5 is comparatively mature.

## 4\. The hybrid reference architecture (sketched)

Large AI-oriented campuses are likely to contain more than one power architecture over their build-out life. Sidecar and native run hall-by-hall; campus-wide commitment framing overstates them; they are conversion-boundary choices that can be taken hall-by-hall, or even row-by-row, given workload mix and tenant base.

A realistic 2027 hybrid scenario for a major colocation operator could include SSTs or transformer-rectifiers in the AI halls and traditional step-down transformers in the legacy ones; DC bus to the AI halls and AC distribution to the legacy ones; sidecar power racks for one tenant class and Nvidia-reference rack architectures for another; 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.

FIGURE 3.6 · TWO CONVERSION-BOUNDARY ARCHITECTURES adikumar.co · essay 3/13 Native and sidecar rendered end-to-end through the six-stage frame. The AC to DC conversion boundary lands at Stage 2 in native, and at Stage 4 in sidecar. Everything else follows. NATIVE ARCHITECTURE AC→DC conversion at Stage 2 · 800 VDC backbone through the hall 01 · UTILITYUtility MV service13.8 kV class02 · PERIMETERTransformer-rectifier or SSTMVAC → 800 VDC (conversion boundary)◀ CONVERSION BOUNDARY03 · BUILDING800 VDC hall or row backboneunipolar or ±400 V bipolar04 · RACKIn-rack power shelf800 VDC → 48 V step-down05 · IN-RACK48 V busbarORv3 HPR interface06 · POLMulti-phase VRMto accelerator operating rails SIDECAR ARCHITECTURE AC→DC conversion at Stage 4 · legacy 480 VAC preserved upstream 01 · UTILITYUtility MV service13.8 kV class02 · PERIMETERStep-down transformerMV → 480 VAC03 · BUILDING480 VAC building distributionexisting switchgear + busway04 · RACKSidecar Mt. Diablo shelf480 VAC → 800 VDC (conversion boundary)◀ CONVERSION BOUNDARY05 · IN-RACK48 V busbarORv3 HPR interface after in-rack step-down06 · POLMulti-phase VRMto accelerator operating rails Same six stages. The AC to DC conversion boundary moves between Stage 2 and Stage 4. SOURCES · OCP LVDC Aug 2026 · Diablo 400 v0.7.0 · Nvidia developer blog on 800 VDC · adikumar.co · essay 3/13 

Figure 3.6\. Native and sidecar rendered through the six-stage frame. The AC/DC conversion boundary moves between Stage 2 (native) and Stage 4 (sidecar).

Hybrid is 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 (sidecar at one $/kW/month, native 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 hyperscalers run heterogeneous platform estates: Microsoft is deploying Maia alongside Nvidia Vera Rubin and AMD systems, and Google, Amazon and Meta each mix their own accelerators with Nvidia-reference platforms. A heterogeneous compute estate makes a single campus-wide power architecture less obviously sufficient, even inside a self-operated fleet.

The operators most exposed to this complexity are those still treating the architecture as monolithic. They make commitments at Stage 2 that can constrain later choices at Stage 4, or they procure Stage 3 equipment that limits what is deployable at Stage 5\. The architecture map is how operators avoid that outcome.

Essay 6 walks through the operator economics, the case-study deployment patterns, and the procurement implications of running hybrid as a product. This essay establishes that hybrid is the practical operating model for campuses that contain multiple generations, workloads, or power architectures. Supplement B provides the quantitative scenario case.

The DC-DC Transition · Navigate the series

- **← PREVIOUS** · [Essay 2 · Two architectures wearing the same name](https://adikumar.co/two-architectures-wearing-the-same-name/)
- **→ NEXT** · Essay 4 forthcoming
- **FRAMEWORK MAP** · [The AI Power Chain framework](https://adikumar.co/framework-ai-power-chain/)
- **DEEP DIVE** · [Supplement A · The Transition Tax](https://adikumar.co/dc-dc-supplement-a-transition-tax/)
- **DEEP DIVE** · [Supplement B · Hybrid forever](https://adikumar.co/dc-dc-supplement-b-hybrid-forever/)

## 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.
- **LVDC / MVDC** · Low-Voltage / Medium-Voltage Direct Current. The IEC 60038 voltage classes referenced throughout this series.
- **OCP** · Open Compute Project. Hyperscaler-led industry consortium publishing open specifications and reference designs for data centre hardware.
- **SST** · Solid-State Transformer. A power-electronics-based transformer that provides electrical isolation and voltage conversion; in the AI data-centre architectures discussed here, it is being developed to convert medium-voltage AC to 800 VDC.
- **VDC** · Volts Direct Current. Used in 48 VDC / ±400 VDC / 800 VDC references to rack-level and data-centre power distribution.

For the full corpus glossary of acronyms used across all essays, see [adikumar.co/glossary](https://adikumar.co/glossary/).

Series expansion · six new supplements

The DC-DC Transition series now includes six deep-dive supplements addressing specific questions that the main 13-essay series flagged but did not fully develop. Consulting-grade visuals throughout. Vendor concentration matrices, procurement flow diagrams and capex decomposition views:

1. [Supplement A. The Transition Tax](https://adikumar.co/dc-dc-supplement-a-transition-tax/): capex decomposition across the six domains. Where the 800V DC premium accumulates in retrofit versus greenfield paths.
2. [Supplement B. Hybrid Forever](https://adikumar.co/dc-dc-supplement-b-hybrid-forever/): why most facilities will run mixed AC/DC through 2035+. Hyperscaler zone strategies compared.
3. [Supplement C. The Voltage Ceiling](https://adikumar.co/dc-dc-supplement-c-voltage-ceiling/): 1500V DC and the MVDC question. Standards + vendor readiness by voltage class.
4. [Supplement D. The Solid-State Transformer](https://adikumar.co/dc-dc-supplement-d-solid-state-transformer/): facility-distribution supplier dynamics. Western vs Chinese vendor field.
5. [Supplement E. The Sidecar Power Rack](https://adikumar.co/dc-dc-supplement-e-sidecar-power-rack/): the rack-shelf architecture. OCP Mt. Diablo v0.7.0 reference. 80+ ecosystem partners.
6. [Supplement F. The Connector Wars](https://adikumar.co/dc-dc-supplement-f-connector-wars/): rack and on-package interface engineering. High-current connector vendor concentration + failure modes.

Full series index: [/tag/dc-transition/](https://adikumar.co/tag/dc-transition/)All supplements now live.

In brief

AI data centre power passes through six stages between grid meter and the accelerator package: utility interface, perimeter conversion, building distribution, rack-level conversion, in-rack distribution, and point-of-load. Each stage has different vendor concentration, standards maturity, and lead-time exposure. A 100 MW campus accumulates conversion and distribution losses across the chain; the magnitude depends on topology, equipment efficiency, loading and redundancy. Sidecar and native are the two principal 800 VDC conversion-boundary architectures this essay maps.

The AI Power Chain: series companions

- [The AI Power Chain: Vendor Screen](https://adikumar.co/ai-power-chain-vendor-screen/). The vendor map across every layer
- [Pricing Under Scarcity](https://adikumar.co/pricing-under-scarcity/). Where premium capture is compounding
- [The Services Inversion](https://adikumar.co/services-inversion/). Why services now command product-like multiples

## Frequently asked

What does a 50 MW AI campus power chain typically involve?

There is no single standard topology. A representative 50 MW AI campus takes medium-voltage utility power (13.8 kV or 34.5 kV), transforms to 480 VAC and distributes AC to sidecar power racks that convert to 800 VDC for high-density AI rows. A purpose-built AI hall may instead use MVAC → 800 VDC conversion at the perimeter with DC facility distribution. The appropriate topology depends on rack density, existing infrastructure, redundancy requirements, workload mix and site economics.

What are the six stages of the AI data centre power architecture?

The six stages, from grid meter to the accelerator package, are: 1) Utility interface (medium-voltage substation, BESS interconnection); 2) Perimeter conversion (transformer or SST to 800V DC / ±400V / legacy 480 VAC); 3) Building distribution (bus or switchgear to compute halls); 4) Rack-level conversion (sidecar power rack or in-rack power shelf); 5) In-rack distribution (48V busbar to blade slots); 6) Point-of-load (VRM at 0.8V, hundreds of amps to GPU core). Each stage has different vendor concentration, standards maturity 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. A 1 GW AI factory would generally require a high-voltage utility interconnection and one or more site distribution substations, with downstream medium-voltage distribution feeding either conventional AC infrastructure with local 800 VDC sidecar conversion or direct MVAC → 800 VDC conversion for purpose-built AI halls. The choice changes where conversion, protection, energy storage and distribution sit in the chain, and each stage has its own vendor concentration.