NVIDIA has published the architecture. OCP and the supplier alliance have published the spec and the timeline. The colocation operators have published, so far, very little.
NVIDIA, the Open Compute Project, and roughly thirty power-electronics vendors have spent the last twelve months publishing a coordinated 800 VDC architecture for supercomputing racks scaling from 100 kW to over 1 MW. What public-facing colocation operators have said about adopting it is a separate question, and the gap between what NVIDIA and its partners have published, what OCP and the supplier ecosystem are standardizing, and what can be inferred from the operator silence is the structure of this story.
This is the second-order companion to the trillion-dollar NVIDIA backlog hitting the grid capacity wall. The grid story is about generation, transmission, and the substation. The 800 VDC story is about the last fifty feet... the busway, the rack, and the conversion stages between them.
NVIDIA has published two architecture blogs that bracket the design. The May 20, 2025 post introduced 800 V high-voltage DC distribution and stated that "using 800 V busways and switching from 415 VAC to 800 VDC in electrical distribution enables 85% more power to be transmitted through the same conductor size." It claimed up to 5% end-to-end efficiency improvement over current 54 V systems, said the architecture supports "racks ranging from 100 kW to over 1 MW," and tied full-scale production to "NVIDIA Kyber rack-scale systems in 2027." The October 13, 2025 follow-up extended the conductor-density framing to "157% more power than 415 VAC" at the same wire gauge, and observed that traditional multi-stage AC-to-DC conversion produces end-to-end efficiency "less than 90%" in some configurations.
The 600 kW figure that has anchored most outside coverage of Kyber traces back to Jensen Huang's GTC 2025 keynote rather than a single NVIDIA spec slide. NVIDIA's Vera Rubin POD blog walks through the NVL72, NVL144, and NVL576 Kyber configurations that sit inside the published 100 kW to 1 MW band. The NVL576, the configuration most often associated with the 600 kW number, is the high end of the published architecture and the one driving most of the facility-level redesign conversation. The architecture pulls AC-to-DC rectification out of the server PSU and into a facility-level sidecar that distributes 800 VDC across the row. Inside the rack, the 48 V intra-rack bus of Open Rack v3 survives because the form factor survives. The change is upstream of the rack, not inside it.
The silicon and systems vendors most exposed to the architecture have published companion pieces into the same envelope. Texas Instruments announced a complete 800 VDC power architecture for AI data centers with NVIDIA on March 16, 2026, describing reference designs from the medium-voltage AC input through to the point-of-load at the GPU. Vertiv's May 2025 blog named the NVIDIA Rubin Ultra rollouts explicitly. On Vertiv's Q1 2026 earnings call, Morgan Stanley analyst Chris Snyder asked CEO Giordano Albertazzi about 800 V architecture timing directly. Albertazzi confirmed the portfolio launches in the second half of 2026, with shipping in 2027. That detail is not in the 8-K filing; it sits in the call transcript.
The two hyperscaler-published architecture posts sit on a different cadence. Microsoft's Azure infrastructure blog framed the disaggregated power sidecar as "fueling the next wave of AI platforms." Google's cloud blog described the same architecture as the enabling spec for 1 MW IT racks and liquid cooling at the OCP EMEA Summit. Both are explicit about co-authoring the OCP reference spec the rest of the ecosystem now bids against. Neither attaches a public site-by-site deployment schedule.
The reference architecture sits in the OCP spec library. OCP's Diablo 400 specification, v0.5.2, was published on May 30, 2025. It defines a disaggregated sidecar power rack that takes three-phase AC in and delivers ±400 VDC out — 800 V rail-to-rail, bipolar. The spec specifies droop and active current sharing, a voltage-drop budget of 0.1% at five meters of HVDC output cable, and a scaling path from 100 kW to 1 MW per IT rack on a single architecture. Authoring credit sits with Microsoft, Meta, and Google. The component layer leans deliberately on the 400 V automotive EV supply chain, so the supercomputing buildout can ride that cost curve rather than fund a separate one.
The chairs of the OCP Power Distribution sub-project are Paolo Catapane of ABB and Joshua Buzzell of Eaton, both listed on the OCP project page. That those are the chairs is itself a signal. The two industrial vendors with the deepest medium-voltage switchgear lineage are writing the rules under which their own products will be specified into the next supercomputing buildout. The 800 V HVDC Supplier Alliance was announced by NVIDIA at COMPUTEX 2025, with roughly thirty power-electronics vendors committed.
The supplier order book gives the standardization a financial floor, with attribution that needs reading carefully. Eaton's Q1 2026 report landed at $7.45 billion in total-company revenue, up 17% year-over-year. The widely cited data-center metrics (data-center orders up 240% year-over-year, data-center revenue up 50% year-over-year) sit inside that company total, not on top of it. Eaton's reported $22.8 billion backlog is a total-company figure across electrical, aerospace, vehicle, and eMobility segments; the data-center share is not separately broken out in the Q1 release. Vertiv reported $2.65 billion in revenue, up 30%, with Americas organic growth at 44% and operating profit up 51%. Schneider Electric attributed Q1 growth to AI-driven data-center demand without naming 800 VDC by voltage, and its disclosed portfolio spans broader electrification and grid-infrastructure segments well beyond the supercomputing power-delivery line. The directional signal is unambiguous; the precise data-center share is not what those headline numbers represent.
The semiconductor layer beneath the system integrators has been publishing into the same envelope for over a year. Navitas Semiconductor's October 2025 white paper laid out the GaN-and-SiC device roadmap. Flex described the ecosystem transition. STMicroelectronics' 800 V HVDC data-center blog is the clearest single statement that the supercomputing power-delivery transition rides the automotive 800 V EV cost curve for SiC and GaN. The pattern matches a structure Supercomputing News has documented before: HBM allocation, not HBM supply, is the 2026 AI infrastructure story, and the same allocation logic now applies to SiC. Capacity exists. Allocation is contracted ahead of it. Hyperscalers with Mt. Diablo authorship leverage are first in line.
Open Rack v3 is not displaced by any of the above. ORv3 still defines the mechanical rack and the 48 V intra-rack bus. What the OCP spec library reflects is a deliberate division of labor: the rack mechanicals are stable; the upstream conversion path is what the ecosystem has agreed to rewrite. The disruption clusters at the sidecar, the row backbone, and the medium-voltage interface, not inside the rack envelope.
The third layer of the story is inference, and it requires explicit framing. Each claim in this section is either a bounded search (we looked at this corpus, we did not find that) or an analytical extension (we applied the published architecture to public site specs and observed the gap), not a directly sourced operator commitment.
The operator silence is a bounded search. A review of Q1 2026 earnings releases, 8-K filings, and prepared remarks from Equinix, Digital Realty, CoreWeave, Crusoe, Iron Mountain, QTS, Aligned, Switch, Vantage, NTT, STACK, plus AWS, Azure, Google Cloud, Meta, and Oracle did not surface a named 800 VDC commitment as of May 2026. Equinix highlighted accelerating liquid-cooling deployments without naming voltage. Digital Realty announced liquid-to-chip cooling support across its 170-data-center global footprint without naming voltage. The absence of mention in this specific corpus is what is reportable; the absence of internal planning is not. Operators have at least three plausible reasons not to name an architecture commitment on a public call. An 800 VDC commitment signals architectural lock-in to competitors. Parts of the standards stack remain unratified, which makes forward-looking voltage language a securities-disclosure hazard. And a 2026 statement of "Rubin Ultra ready" is a hostage to a 2027 delivery date the operator does not fully control. Those reasons sit alongside the capex pressure that Rubin throws into 2027 data-center budgets, which is its own incentive against premature commitment.
The sovereignty observation needs the same boundedness. Among publicly documented sovereign programs, per-rack density is not uniformly disclosed. The UK's AIRR programme (Isambard-AI and Dawn) sits in a documented 80–100 kW per rack envelope. The EuroHPC Joint Undertaking's 19 AI Factory sites are individually specified at varying density bands; some published procurement documents land in the 30–60 kW air-cooled or 80–120 kW liquid-cooled range. Many AI Factory and sovereign-program sites were designed below the envelope Rubin Ultra will require; public per-rack density is not uniformly disclosed across programs. The defensible statement is narrower than a blanket one: some programs are demonstrably below the architecture's design point, some are not publicly characterized at all, and the ones architected concurrently with Rubin, including the European Commission's emerging AI Gigafactory program, have not yet committed to a public per-rack figure.
Retrofit economics is the most explicitly analytical claim in the piece. A range of roughly $1–2 million per delivered megawatt for deep electrical retrofit appears across vendor TCO modeling from Schneider, Vertiv, and Eaton, but it is not anchored to a canonical public citation. Treat the band as illustrative, not authoritative. The structural point underneath the dollar figure holds regardless: a hall sized for 80 kW per rack was sized for a different substation, transformer chain, switchgear envelope, and cooling plant. Moving that hall to a Rubin-class density typically requires re-engineering more of the facility than the rack itself.
The Tier-III, arc-flash, and labor floor is standards-stack analysis with primary sourcing where available. Uptime Institute's Tier criteria define outcomes (concurrent maintainability for Tier III, fault tolerance for Tier IV) rather than mandate a power-distribution architecture. Certification on 800 VDC is not precluded by the published criteria, but it requires procedure-level re-engineering across switching, lockout/tagout, qualification, and the operations playbook. On the arc-flash side, NETA World's column on the NFPA 70E 2027 cycle confirms that the 2027 edition was released digitally on NFPA LiNK in May 2026, and identifies Bobby Gray as chair of the DC Tables task group. Beyond that primary source, claims about the magnitude of expanded DC content are best held as ecosystem expectation, not commitment.
Several failure modes shift in character at 800 VDC. Arcs do not self-extinguish at a zero crossing, so once ionized they sustain. Most 800 VDC systems are ungrounded or high-resistance grounded; a first ground fault produces minimal current and may not trip protection, while a second fault produces a hard short. Sustained DC potential also drives electrolysis in cable sheaths, raceways, and liquid-cooling loops; equipotential bonding and insulation monitoring across the cooling distribution unit become facility-design requirements rather than facility-design afterthoughts. None of these are exotic. All of them are different enough from the AC equivalents that the practitioner workforce qualified on 480 VAC switchgear is not automatically qualified on 800 VDC. The overlap with the EV-charging-station electrician population is uncomfortably high, and both buildouts are bidding for the same labor pool. A residual air-cooling load remaining after direct-to-chip capture, in the single-digit percentage range typical of liquid-cooled supercomputing halls, is what rear-door heat exchangers and in-row coolers absorb; the specific per-rack kilowattage at Rubin Ultra density is not publicly fixed and scales with the configuration point the operator picks.
Three signals will tell the operator side of the story when it finally surfaces. Watch the Vertiv H2 2026 800 VDC portfolio launch, specifically the wording Vertiv uses when it names the SKUs, because that is the moment "available" replaces "in development" in the supplier vocabulary. Watch the European Commission's AI Gigafactory first call, expected in the back half of 2026, for whether the call documents specify 800 VDC at the row boundary or leave the question open. And watch the Q2 and Q3 2026 hyperscaler earnings cycles: the first public-facing colocation operator to say "800 VDC" on an earnings call is the data point that confirms the architectural transition has crossed from the supplier order book into the operator press release. Until then, the asymmetry holds. Vendors are publishing. Operators are not.
