800 VDC Power Architecture for AI Data Centers: Standards, Vendors, and Lead Times in 2026

800 VDC data center power moves the AC-to-DC conversion out of the IT rack and distributes 800 volts DC to racks that step it down to 50 V, 12 V, or 6 V close to the GPU. NVIDIA states that full-scale production of 800 VDC data centers will coincide with its Kyber rack-scale systems in 2027, while today’s GB200 and GB300 NVL72 racks remain on a 54 V DC busbar fed by 415/480 VAC power shelves ([1]). Specifically, this guide covers the architecture, the OCP specifications, the vendor ecosystem with dates, the US standards position, and the hot-swap and leak-detection practices an electrical team needs before energizing an 800 VDC rack.

What 800 VDC data center power replaces, and why NVIDIA is moving to it

The conventional AI hall power chain runs medium-voltage AC (roughly 35 kV) to a transformer, then 415 VAC through an AC UPS and PDUs or busway to in-rack power shelves, which rectify 415 VAC to 54 VDC for the busbar, followed by tray-level DC/DC conversion to the GPU ([2]). Because that chain involves so many conversion stages, NVIDIA puts its end-to-end efficiency at less than 90% and counts four conductors per AC feed against three for DC (positive, return, protective earth) ([2]).

The constraint is, specifically, copper and shelf space. NVIDIA calculates that a 1 MW rack on a 54 VDC busbar would need up to 200 kg of copper busbar, and that a Kyber rack at 54 V would consume up to 64 U of power shelves; as a result, the limitations of 54 V begin above roughly 200 kW per rack ([1]). Moving from 415 VAC to 800 VDC therefore delivers 85% more power through the same conductor and reduces copper requirements by 45% in NVIDIA’s May 2025 analysis; its October 2025 note, using a same-wire-gauge basis, instead states 157% more power than 415 VAC ([1]; [2]).

Consequently, NVIDIA’s stated targets for the native architecture are up to 5% end-to-end efficiency gain versus 54 V, up to 70% lower maintenance cost from fewer PSUs, and up to 30% lower TCO ([1]). The Kyber rack, which carries 576 Rubin Ultra GPUs, is scheduled for 2027 ([3]). In addition, conversion stages drop: NVIDIA’s reference uses a single 64:1 LLC stage from 800 VDC to 12 VDC near the GPU with 26% less area than the current path ([2]).

The GB200 and GB300 NVL72 power train today: 54 V busbar, AC power shelves

A GB200 NVL72 rack draws about 120 kW and is fed by power shelves that convert AC input to a nominal 50-51 V DC on the rack busbar; each shelf holds six 5.5 kW PSUs (33 kW per shelf), and a full rack carries eight shelves in an N+N arrangement fed by power whips from a remote power panel ([4]). Similarly, HPE’s QuickSpecs for the same platform list six shelves standard (eight maximum), a 50 V DC busbar, IEC 60309 560P6W five-wire whips, and a 132 kW nominal TDP with an EDPp of roughly 192 kW that busways should be provisioned to carry ([5]).

This is, importantly, the baseline every 800 VDC decision is measured against. NVIDIA states that its rack-adjacent 800 VDC power rack “slots into” existing AC distribution without building electrical changes, and that the same MGX footprint supports GB300 NVL72 and Vera Rubin NVL72 ([6]; [3]). For more on the AC chain as built today, see Power Delivery for AI.

OCP Open Rack v3, Mt. Diablo, and the LVDC standardization push

Google announced at OCP EMEA 2025 that it was moving from 48 VDC to ±400 VDC to support racks from 100 kW to 1 MW, expecting more than 500 kW per rack before 2030, and released the Mt. Diablo 0.5 draft with Meta and Microsoft in May 2025; specifically, the design is an AC-to-DC sidecar power rack with a claimed efficiency gain of about 3% ([7]). The Mt. Diablo specification, in particular, supports either ±400 VDC bipolar or 0-800 VDC output ([8]).

In August 2026, Google, Microsoft, and NVIDIA published a joint OCP position. Namely, the OCP Power Distribution project’s LVDC work covers up to 1,500 VDC, keeps the ±400 V Mt. Diablo sidecar spec, plans a Mt. Diablo 2.0 for native 800 V, released a Solid-State Transformer specification v0.3, and is working with UL Solutions, NFPA, IEEE, and IEC ([9]). Notably, NVIDIA states that more than 80 companies are building to the specification and that the joint white paper was published in March 2026 ([6]). However, an open spec is not the same thing as a listed product, a distinction covered in The 800 VDC Standard That Doesn’t Exist.

AI factory power train architecture: the three deployment options and their dates

NVIDIA’s August 2026 roadmap and the accompanying white paper define three stages, and the dates matter more than the diagrams.

The white paper’s safety requirements for all three stages, notably, include high-resistance grounding, continuous insulation monitoring, active arc-fault detection, and mechanical connector interlocks that prevent disconnection under load ([10]). In contrast, Schneider Electric’s field view is narrower than NVIDIA’s roadmap: its executives told The Next Platform that initial sidecars would supply 600 kW to 1 MW across one to three racks, centralized converters would come in 2-5 MW blocks, and no more than about 10% of new AI nodes would require 800 VDC by 2030 ([11]).

AI power delivery trends: densities, swings, and energy storage

Rack density is, first and foremost, the driver. NVIDIA reports a 75% GPU TDP increase from Hopper to Blackwell and a 3.4x increase in rack density, with rack loads swinging from 30% to 100% in milliseconds; consequently, the reference architecture answers the swing with supercapacitors near the rack for millisecond-to-second events and facility batteries for second-to-minute events ([2]). By contrast, Uptime Institute data cited in August 2026 puts the industry’s average rack density at just over 11 kW ([12]).

Which companies have actually deployed UL-listed 800 VDC power systems in production US data centers?

None of the vendor announcements reviewed for this article claims a UL listing for an 800 VDC power shelf, sidecar, or power rack, and none names a production US data center running one. Instead, the closest public deployments are outside the US or still in design: Foxconn’s 40 MW Kaohsiung-1 facility in Taiwan is being built for 800 VDC, and CoreWeave, Lambda, Nebius, Oracle Cloud Infrastructure, and Together AI are “designing for” 800 V ([3]). For example, Lite-On has a 1.5 MW-class 800 VDC rack in a named hyperscaler program with validation in Q3 2026 and production in Q4 2026 ([13]), and Wolfspeed reports its silicon carbide MOSFETs qualified for Lite-On’s 800 VDC sidecar platforms as of August 2026 ([14]).

UL Solutions’ position is that the applicable product standards do not yet exist in final form. Specifically, in January 2026 it stated that “many local codes and inspectors are not yet prepared” and that dozens of standards for switchgear, panelboards, busbars, cables, and overcurrent devices need updating; UL leads the Codes and Standards workstream inside the OCP Power Distribution sub-project with ABB and Eaton as co-leads ([15]). Therefore, the practical reading is that 800 VDC racks entering US sites in 2026 will be field-evaluated or certified against component standards, not “UL Listed” as a system category.

The 800 VDC vendor ecosystem: announced products, dates, and status

No primary source was found for Rolls-Royce/mtu or Bloom Energy 800 VDC products, so they are omitted. Instead, GE Vernova’s assessment is that the transformer-plus-rectifier side power rack “can be immediately deployed at scale,” while medium-voltage SSTs are “coming years” away ([31]).

Our facility electrical team is nervous about 800 VDC: what’s the current state of US standards and product availability?

The nervousness is, in fact, justified on standards and premature on products. The installation code, the workplace safety standard, and the arc-flash calculation method all exist for DC, but the product safety standards that would let an inspector accept an 800 VDC shelf as a listed category mostly stop at 600 V.

Three points follow. First, the NEC and NFPA 70E already govern 800 VDC work: NEC 110.16 labeling applies to DC, and NFPA 70E Table 130.4(E)(b) sets a limited approach boundary of 3 ft 6 in (fixed part) and a restricted approach boundary of 1 ft for the 301 V-1 kV DC range ([37]; [47]). Second, and importantly, the incident-energy method is the weak link: IEEE 1584 is AC-only, NFPA 70E’s DC methods are conservative, and the NFPA Fire Protection Research Foundation has opened a project, sponsored by Siemens and Schneider Electric, whose Phase 1 tests an 800 Vdc rectified-converter case ([37]). However, Schneider’s own study of two 800 VDC architectures found incident energy below 1.2 cal/cm² and argues that converter-limited fault current, capacitor discharge in the first milliseconds, and reverse-current blocking govern the result more than voltage does ([48]).

Third, and perhaps most consequential, the shock hazard is the training gap. Specifically, at a 1,000 ohm body-resistance baseline, contact with 800 V yields 800 mA, and technicians qualified only to 600 VAC need requalification ([49]). Similarly, Europe has the same product-standard gap: ETSI’s DC interface standard stops at 400 VDC and IEC 62368-1 at 600 V ([46]; [41]).

Which 800 VDC rack power systems are designed to allow safe hot-swapping without full de-energization?

Hot-swap at 800 VDC is being engineered at two levels: the power shelf or module, and the server-tray input. First, at the module level, Schneider Electric describes its 800 VDC sidecar conversion shelves as “Live Swap capable” ([19]), and Heron Power lists field-replaceable modules on its SST and DC panel ([33]). Second, at the tray level, Infineon’s 800 VDC hot-swap controller with a CoolSiC JFET is designed so operators “can exchange server boards in an 800 VDC architecture while other servers continue to operate in the same rack” ([26]), and Texas Instruments demonstrated an 800 V hot-swap controller in March 2026 ([27]).

NVIDIA’s white paper requires mechanically interlocked connectors that prevent disconnect under load ([10]). However, Vertiv notes that protection at this voltage and current is currently fuse-and-disconnect, with resettable DC breakers “still maturing” ([50]).

What vendors offer combined 800 VDC power and liquid cooling systems with integrated leak detection and safety?

Delta is the only vendor in the reviewed set that ties the two together in one announcement: a 2.4 MW liquid-to-liquid CDU with N+1 pumps running on 800 VDC for 800 VDC AI factories, alongside its 660 kW in-row power racks and a 4RU 140 kW in-rack CDU for GB300 NVL72 ([21]). By contrast, Vertiv and Eaton position integrated power-and-cooling platforms (the Vertiv 800 VDC MGX reference architecture and Eaton’s Beam Rubin DSX), and NVIDIA’s Vera Rubin NVL72 introduces a liquid-cooled busbar with 45°C liquid cooling ([3]; [18]).

Leak detection itself is, however, standardized only at the guidance level. Vertiv’s cold-plate white paper states that conductive coolant leaks risk equipment failure and electrocution if accumulated coolant becomes energized, recommends placing high-voltage equipment higher in the system, and describes two-wire rope sensors (leak present) and four-wire sensors (leak location) ([51]). Meanwhile, OCP’s guidelines target an annual cold-plate and coolant-loop failure rate of 0.3% or lower ([52]). No reviewed document specifies a leak-to-trip interlock for 800 VDC circuits; that logic instead remains an owner specification. Cooling design is covered in Liquid Cooling for AI.

800 VDC power shelf lead times 2026: what is published and what is not

No vendor publishes an order-to-ship lead time for an 800 VDC power shelf. What exists instead are production windows: Lite-On’s 110 kW MGX shelves entered mass production in Q2 2026 and its 800 VDC rack is scheduled for Q4 2026 production ([25]; [13]); similarly, NVIDIA’s MGX power rack and Vertiv’s portfolio are both due in H2 2026 ([6]; [16]). Heron Power’s 24-week SST lead time is the only published figure, and importantly, it is a vendor claim for a pre-production product ([33]).

Upstream AC equipment is instead the binding schedule constraint, not the shelf.

We’re designing power distribution for GB200 NVL72 racks: should we go 800 VDC from the bus bar or stick with 54 V?

Stay on the 54 V busbar with AC power shelves for GB200 and GB300 NVL72, and build the electrical room so an 800 VDC power rack can be added later. Because the busbar voltage is fixed by the platform, GB200/GB300 NVL72 is a Gen 1, roughly 145 kW, AC-fed system, and native 800 VDC input arrives with Vera Rubin NVL72 (Gen 2) and Kyber ([10]; [1]). Similarly, Schneider Electric’s retrofit guidance is consistent: facilities at up to 200 kW per rack can add a DC sidecar later, while above 400 kW per rack the infrastructure should be purpose-built ([56]).

Future-proofing steps that follow from the published designs:

• First, size busway and whips to the EDPp, not the TDP; HPE specifies 192 kW provisioning for a 132 kW nominal GB200 NVL72 rack ([5]).
• Second, reserve a rack position and 415/480 VAC feed adjacent to each compute pod for a Gen 2 power rack, which takes AC in and delivers 800 VDC by DC whips ([10]).
• Third, specify busway that is rated or upgradable to 800 VDC; Eaton’s Pow-R-Way III and PowerWave 2 are rated to 600 VDC today with a roadmap to 800 and 1,500 VDC ([17]).
• Additionally, plan high-resistance grounding with insulation monitoring, which the NVIDIA specification mandates for 800 VDC ([10]).
• Finally, order long-lead AC gear now; a 15 kV switchgear lineup runs 52-80 weeks ([55]).

How do I write a safety plan for hot-swapping components in liquid-cooled 800 VDC racks?

The outline below maps each plan section to the governing standard. Since it is an engineering checklist, not legal advice, confirm cited editions with the AHJ.

1. Scope and system description. Define the DC system: voltage (800 VDC or ±400 VDC), grounding method (high-resistance, per the NVIDIA specification), and every energized boundary a technician can reach ([10]).
2. Hazard identification. Classify DC shock thresholds under NFPA 70E Article 310 and set approach boundaries from Table 130.4(E)(b): 3 ft 6 in limited (fixed part) and 1 ft restricted for 301 V-1 kV DC ([39]; [47]).
3. Incident-energy analysis. State the method and its limits: because IEEE 1584 does not cover DC, use an NFPA 70E DC method or a vendor topology study, and record that the FPRF 800 Vdc model is pending ([37]; [48]).
4. Labeling. Apply NEC 110.16 labels with nominal voltage, arc-flash boundary, incident energy or PPE level, and assessment date; notably, 110.16 applies to DC ([35]; [37]).
5. Hot-swap authorization. Permit hot-swap only for modules the manufacturer has documented for insertion and removal under load with interlocked connectors; require the test report and, where available, third-party live-swap certification ([26]; [45]). Otherwise, treat the work as energized work requiring an energized electrical work permit and, under NFPA 70E 130.2(A)(2), an additional person trained in contact release per 110.4(C)(1) ([39]).
6. Worker qualification. Since technicians qualified only to 600 VAC are not qualified at 800 V, requalify them for DC above 600 V, with rated insulated tools and gloves ([49]).
7. Liquid-cooling interface. Require rack-level leak detection with automatic response, route conductive coolant away from DC connections, and define the 800 VDC trip logic on a detected leak; this is an owner specification because no reviewed standard defines it ([51]).
8. Stored energy. Address supercapacitor and BBU shelves under NFPA 70E Article 370 and the DC shock and arc-flash thresholds that apply to them ([39]; [49]).
9. Verification and revision. Finally, set a re-analysis trigger for every rack generation change and for publication of the FPRF DC model and NFPA 70E updates.

How SAVRN builds for this

SAVRN’s factory-built compute blocks are designed so the power train can change without rebuilding the hall. Specifically, the Atom compute block is liquid-cooled and manufactured off site, and the Electron family supplies on-site power, which places the AC-to-DC conversion point inside a block boundary that SAVRN controls rather than in a utility-fed electrical room with 52-80 week switchgear lead times ([55]). As a result, that boundary is where a Gen 1 54 V busbar pod and a Gen 2 800 VDC power rack can coexist on the same footprint, which is the configuration NVIDIA’s own roadmap expects through 2027 ([6]).

SAVRN treats the standards position described above as a design input: no listed 800 VDC shelf category, DC arc-flash methods still under research, and leak-to-trip logic undefined by any standard. Its published work on this subject, in particular, sets out why an open specification is not a UL listing in The 800 VDC Standard That Doesn’t Exist, how today’s AC chain is built in Power Delivery for AI, and how cooling and power interact at the rack in Liquid Cooling for AI. Additionally, the campus model that lets blocks be added by generation is described in Modular AI Campus. None of this requires an operator to buy from SAVRN; instead, the standards, dates, and lead times above apply to any 800 VDC project.

Chad Harris, Founder, SAVRN