On this page · 11 sections
- What changed in 2026: from a 2027 switch-over to a phased standard
- What 800 VDC changes, stage by stage
- Rectifiers, solid-state transformers and busways
- Safety, standards and protection
- The partners NVIDIA named
- Retrofit or new build: a decision table
- Worked example: feeding one 600 kW rack
- India grid context
- Implementation checklist for a retrofit
- How eCorpIT can help
- References
Summary. 800 VDC is NVIDIA's answer to AI racks that outgrow AC distribution, and in 2026 it became a phased option beside AC rather than a 2027 switch-over. NVIDIA says the same copper carries 157% more power at 800 VDC than at 415 VAC, and that a 1 MW rack at 54 VDC would need up to 200 kg of copper busbar. Kyber, targeted for 2027, is designed for 800 VDC input; SemiAnalysis reported in July 2026 that it had slipped to 2028, and NVIDIA said its roadmap was intact. There are three ways in: a 660 kW power rack beside the compute rack, due to enter production in the third quarter of 2026; a row power centre of up to 2 MW, as soon as the third quarter of 2027; and hall-level DC power blocks of about 4.8 MW, with a next-generation solid-state transformer (SST) expected toward 2029. In India, an 800 V bus falls in the above-650 V chapter of the CEA's safety regulations, and a battery storage chapter applies from 1 April 2027. Our verdict: use power racks only where you must host 800 VDC-input racks before 2028, design new halls for row power centres, and leave SSTs to the 2029-plus build cycle.
About this article. For data centre owners, colocation buyers and engineers planning AI halls. On 7 November 2026 we checked each figure against NVIDIA's blogs, white papers and datasheet, OCP's LVDC white paper and SST specification v0.3, Vertiv and ABB announcements, and the Gazette texts of the CEA (Measures relating to Safety and Electric Supply) Regulations, 2023 (notified 8 June 2023) and Amendment Regulations, 2026 (notified 27 March 2026, in force 1 April 2027). Efficiency, cost and timing figures are vendor claims, and the Kyber delay is a report. This is not engineering, legal or procurement advice, and no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.
What changed in 2026: from a 2027 switch-over to a phased standard
Our GTC 2027 buyer briefing covers the Kyber timing dispute and our Rubin Ultra NVL576 Kyber rack guide the rack. This piece covers the electrical design, which has changed since 2025. Milestones come from NVIDIA, OCP and Tom's Hardware's report of the SemiAnalysis claim.
| Date | Milestone |
|---|---|
| May 2025 | NVIDIA ties full-scale production of 800 VDC data centres to Kyber in 2027 |
| Oct 2025 | First NVIDIA white paper; partners now include ABB, GE Vernova, Hitachi Energy and Siemens; Foxconn details a 40 MW 800 VDC site |
| 30 Mar 2026 | OCP LVDC power distribution white paper v1.0 |
| 22 Jun 2026 | OCP SST specification v0.3 by Google, Microsoft and NVIDIA takes effect (announced in July) |
| Jul 2026 | SemiAnalysis reports a Kyber slip to 2028; NVIDIA says its roadmap is intact |
| Aug 2026 | NVIDIA: power rack in the second half of 2026, row power centre in 2027, more than 80 companies building |
| Oct 2026 | Visionbay's K-1 design phases 800 VDC in; MGX 1.2 supports racks up to 250 kW |
Three shifts stand out. First, OCP says 800 VDC does not replace AC but coexists with it, and NVIDIA's page describes a gradual move. Second, it no longer depends on Kyber: an 800 V to 54 V power shelf of six 15 kW supplies lets an MGX compute rack, including Vera Rubin NVL72, run from an 800 VDC power rack. Third, hyperscalers now co-write the requirements, so suppliers build to one set of interfaces.
What 800 VDC changes, stage by stage
A conventional AI hall steps medium voltage (for example 35 kV) down to 415 or 480 VAC, passes it through an AC UPS and PDUs, and converts it to 54 VDC inside each rack. NVIDIA says end-to-end efficiency in that chain can fall below 90%. An 800 VDC hall moves the AC-to-DC step upstream to a rectifier or SST and converts once more inside the rack; in Kyber a 64:1 converter steps it straight to 12 VDC beside the GPU.
Copper and current. NVIDIA's first white paper compares power per square millimetre of conductor at a fixed 48 A continuous rating.
| Distribution | Conductors | kW per mm² | Gain over 415 VAC |
|---|---|---|---|
| 415 VAC three-phase | 4 (three phases, earth) | 0.6 | Baseline |
| 480 VAC three-phase | 4 | 0.8 | +16% |
| 800 VDC | 3 (positive, return, earth) | 1.7 | +157% |
| 1,500 VDC | 3 | 3.1 | +382% |
NVIDIA has published three numbers: 85% through row busways of the same conductor size (May 2025), over 150% (October 2025 OCP blog) and 157% (October 2025 technical blog and white paper). The bases differ, so quote the one that matches your design. A hall already on 480 VAC has banked 16%, so its uplift is about 120% (2.57 divided by 1.16).
Claimed savings. NVIDIA claims up to 5% better end-to-end efficiency than 54 V systems, up to 70% lower maintenance cost and up to 30% lower total cost of ownership; ABB claims more than 5%. Treat these vendor projections as a ceiling; in Europe they also count toward duties such as Germany's PUE limits.
Why 800 V, single-ended. NVIDIA chose one positive and one return conductor rather than ±400 V, because bipolar systems need three-pole protection devices that are not widely available. Fuses on both converter inputs still let ±400 V sources feed the rack, such as sidecars built to OCP's Mt Diablo specification. NVIDIA treats 1,500 VDC as a possible long-term target.
Rectifiers, solid-state transformers and busways
Retrofit cost and risk sit in the equipment. NVIDIA's white papers describe four ways to make 800 VDC and two scales of distribution.
| Equipment | What it does | Maturity per NVIDIA | Timing |
|---|---|---|---|
| Power rack (sidecar) | 415 to 480 VAC in, 800 VDC out to adjacent racks; about 660 kW | Entering production | Q3 2026 |
| Facility rectifier | AC below 600 V to 800 VDC; about 1.5 MVA strings at about 99% | Mature in BESS and renewables | Available |
| Transformer-rectifier unit (TRU) | Line-frequency transformer plus rectifier; MV units up to 7.5 MVA at 98.5% or better | Proven in mining, electrolysis, rail | Practical for about 5 MW blocks |
| Solid-state transformer | High-frequency isolation; medium voltage straight to 800 VDC | Developing; most units 15 kV class | Next generation toward 2029 |
| Row DC busway | Overhead busway, 1,250 A feeders, 125 A tap boxes | In development and validation | From Q3 2027 |
| Hall DC switchboard | 6,000 A-class DC busduct and switchboard | Specified for DC power blocks | No date given |
The reference block. NVIDIA's first white paper sketches a 17.5 MW block: five 3.5 MW medium-voltage rectifiers, five-to-make-four, convert 35 kVAC into a 5,000 A DC board. Protected 1,500 A busducts feed four 1.1 MW racks with 1+1 feeds; one rectifier can carry up to 3.3 MW if its partner fails.
TRU first, SST later. NVIDIA's second white paper weighs UPS-derived IGBT converters, aggregated rack power shelves behind a transformer, and the Panama design with a phase-shifting transformer. SSTs are smaller, but most commercial units accept about 15 kV, while many hyperscale sites distribute at 34.5 kV. Vertiv runs two tracks: an MVDC UPS it calls more ready today, and SSTs as they mature. ABB markets an SST-based Infinitus portfolio, with no ship date or price on its page.
What the SST specification asks for. OCP's SST specification v0.3 defines 13.8 kV and 34.5 kV input versions (each plus or minus 10%) rated at 5 MW, with a 10 MW option at 34.5 kV. It asks for an 800 VDC unipolar output within 1% static tolerance, at least 98% power-train efficiency from 50% to 100% load, and floating or high-resistance grounding. The enclosure is 10 to 20 feet long, NEMA 3R rated for outdoor or grey-space use, from minus 30°C to 40°C, with derating above 40°C and above 1,000 m.
Energy storage moves onto the DC bus. Rack-side capacitors absorb swings as fast as 400 microseconds; a battery system at the grid connection covers slower ramps and generator transfers.
Safety, standards and protection
DC behaves differently in a fault, the main reason 800 VDC is arriving in stages.
| Topic | AC | 800 VDC | Design response |
|---|---|---|---|
| Fault current | Crosses zero every cycle | No zero crossing | Faster interruption and current limiting |
| Series arcs | Tend to self-extinguish | Can sustain | Interlocked connectors, arc detection |
| Fault energy | High from utility and machines | Often lower in converter-limited systems | Limit current at every source |
| Ground faults | Overcurrent and earth-fault relays | Insulation or residual current monitoring | Depends on grounding scheme |
| Breakers | Mechanical breakers common | Solid-state and hybrid breakers clear fastest | Faster clearing cuts incident energy |
Grounding is not settled. Four schemes are under evaluation: high-resistance midpoint, high-resistance return, floating and solid. NVIDIA credits midpoint grounding with more uniform fault detection, but system studies will decide.
Two protection zones. From the rectifier or SST to the tap boxes, only qualified staff have access and insulation monitoring raises alarms without immediate shutdown. From the tap box to the rack, where people work, residual current monitoring and fast isolation come first. Each 125 A DC whip sits behind a normally open contactor that closes only once the rack connector is locked, a design borrowed from EV charging.
Breakers. Moulded-case circuit breakers (MCCBs) are the day-one choice for 125 A branches but can take about 20 ms to clear, so designs need redundant feeds or hold-up time. Solid-state circuit breakers (SSCBs) clear in under a millisecond; NVIDIA targets 125 A air-cooled units for in-rack power shelves and 1,250 A, likely liquid-cooled, units for future native 800 VDC racks.
Certification. IEC 60364-1 and NEC Article 100 both put the low voltage line at 1,500 V DC, hence the label LVDC. NVIDIA says its first power rack can be certified now under UL 62368-1 for UL and IEC regions, because it keeps 800 VDC exposure inside the rack. For hall-level distribution, Google, Microsoft and NVIDIA are working with UL Solutions, NFPA, IEEE and IEC. The SST specification already lists NFPA 70 and 70E, UL 347A, UL 1741, IEEE 1584 for arc flash and IEEE 519 for harmonics.
The partners NVIDIA named
NVIDIA named three data centre power system companies in May 2025 (Eaton, Schneider Electric, Vertiv) and longer lists in October 2025. A name on the list is not a shipping product.
| Category | Companies NVIDIA lists |
|---|---|
| Data centre power systems | ABB, Eaton, GE Vernova, Heron Power, Hitachi Energy, Mitsubishi Electric, Schneider Electric, Siemens, Vertiv |
| Power system components | BizLink, Delta, Flex, Lead Wealth, LITEON, Megmeet (GE Vernova also appears here in one list) |
| Silicon | Analog Devices, AOS, EPC, Infineon, Innoscience, MPS, Navitas, onsemi, Power Integrations, Renesas, Richtek, ROHM, STMicroelectronics, Texas Instruments |
| Operators designing for 800 VDC | Foxconn (40 MW Kaohsiung-1), CoreWeave, Lambda, Nebius, Oracle Cloud Infrastructure, Together AI |
Vertiv said in October 2025 that its 800 VDC portfolio of rectifiers, DC busways and rack converters was planned for the second half of 2026, aligned with the 2027 Rubin Ultra rollout. It now says AC delivery breaks down between 350 and 400 kW per rack. In October 2026 NVIDIA described Visionbay's K-1 design, built with Vertiv equipment, as phasing 800 VDC in through cluster-level power centres. We found no published prices from any vendor named here.
Retrofit or new build: a decision table
The options come from NVIDIA's second white paper, the 250 kW line from MGX 1.2, and the 227 kW figure from the Vera Rubin NVL72 datasheet, which also lists 187 kW and 156 kW modes.
| Your situation | Recommended path | Why | Watch for |
|---|---|---|---|
| Existing AC hall, racks up to 250 kW | Stay on AC | MGX 1.2 supports racks up to 250 kW; Vera Rubin NVL72 peaks at 227 kW | Peak, not nominal, load |
| AC hall needing 800 VDC-input racks, in-row battery backup or N+N feeds before 2028 | Option A: power racks | No upstream change; production stated for Q3 2026 | One power rack per one or two compute racks; 12 to 24 AC whips each |
| Existing hall, selected high-density rows from 2027 | Option B: row power centre | Up to 2 MW per row; frees white space; leaves upstream AC rooms alone | DC breaker and busway certification; Q3 2027 at the earliest |
| New hall opening 2028 or later for racks near 1 MW | Option C: DC power blocks | About 4.8 MW blocks in 20 MW deployment units; TRU now, SST later | Outdoor containerised equipment, grounding, inspector approval |
| Campus planned for 2029 onward | Reserve space for direct medium-voltage conversion | Next-generation SSTs expected toward 2029, blocks near 10 MW | SST input voltage versus your supply |
Power racks are the fastest path but add a conversion stage, and NVIDIA calls them a near-term step. The further upstream DC conversion moves, the more of the building's protection and inspection regime changes.
Worked example: feeding one 600 kW rack
Take a Kyber-class rack at 600 kW, the 2025 figure Tom's Hardware reported Jensen Huang confirming. NVIDIA has published no Kyber rating, and the design has moved: it was shown with 144 GPU packages at GTC 2026, and a switch to two dies per package has been reported. Treat 600 kW as a planning number, before derating.
Current. At 54 VDC, 600,000 W divided by 54 V is about 11,100 A. At 415 VAC three-phase and unity power factor, 600,000 divided by (1.732 x 415) is about 835 A per phase. At 800 VDC it is 750 A.
Whips. A 125 A DC whip at 800 V carries 100 kW, so six feed the rack. A 100 A three-phase whip at 415 V carries about 72 kW (1.732 x 415 V x 100 A), so the rack needs nine. At N+N it is twelve against eighteen.
Conversion space. One adjacent 660 kW power rack, fed by 12 to 24 AC whips of 100 A, covers this rack. A 2 MW row power centre feeds three such racks (1.8 MW) with no power racks.
Efficiency. NVIDIA's gain of up to 5% across 10 MW of IT load is up to 500 kW less loss, or 4,380 MWh a year at constant full load (500 kW x 8,760 hours). A power rack retrofit captures less.
India grid context
415 V halls get the full comparison. For a 415 VAC hall, NVIDIA's 157% applies as published, since that is its baseline. The power rack takes true three-phase 415 to 480 VAC with no neutral, through IEC 60309 connectors.
Medium voltage and the SST specification. A 33 kV supply falls inside the 34.5 kV version's plus or minus 10% band (31.05 to 37.95 kV), but an 11 kV supply falls outside the 13.8 kV band (12.42 to 15.18 kV). The specification gives input frequency as 57 to 63 Hz, a band around 60 Hz, though the same cell notes 50/60 Hz; India's grid code sets a 50 Hz reference, so get 50 Hz ratings in writing. Check its 40°C ambient ceiling against your site's summer peak. Since 34.5 kV SSTs need more development, we read a transformer-rectifier as the near-term route for a 33 kV site; confirm with your utility and suppliers.
The 650 V line in CEA's rules. The CEA (Measures relating to Safety and Electric Supply) Regulations, 2023, notified on 8 June 2023, run to 136 regulations in 14 chapters and split safety provisions at 650 V: Chapter V up to 650 V, Chapter VI above. So 800 VDC, low voltage in IEC terms, sits in CEA's higher category. HVDC is defined as 100 kV and above, so neither the HVDC nor the EV charging chapter covers a data hall bus. Regulation 45 ties inspection to a voltage notified by the appropriate government: above it, the Electrical Inspector must approve before energising; at or below it, owners self-certify. Regulation 14 points to Indian standards, India's National Electrical Code and National Building Code, and international standards where Indian ones are missing, which is how UL and IEC work would reach you.
Battery storage rules from April 2027. CEA's listing now shows an amendment. The Amendment Regulations, 2026 take effect on 1 April 2027 and add Chapter XA for battery storage connected above 650 V, covering two-fault tolerance, battery management, fire detection and suppression, ventilation, emergency stops and an independent fire safety audit within three months. A new regulation 137 has CEA list relevant standards in the same window. Its power conversion definition assumes storage feeding the AC grid, so if batteries sit on the 800 VDC bus, ask your inspector whether Chapter XA applies.
People and grid. People working on an installation's lines and apparatus need a certificate of competency or electrical work permit, and an installation above 250 kW of connected load needs an Electrical Safety Officer. NVIDIA calls grid interconnection a primary bottleneck because of synchronised GPU load swings, and in October 2026 Google, Meta, Microsoft and NVIDIA published an OCP Grid Resilience Specification under which AI sites ride through disturbances and restore near-full load in under two seconds. Raise ramp-rate and ride-through expectations with your discom early; our Karnataka data centre policy note covers discom-side site factors.
Implementation checklist for a retrofit
- Measure peak rack power for every planned AI row; mark rows above about 250 kW.
- Audit spare AC capacity, row space and whip routes for power racks.
- Choose the conversion point: power rack now, row power centre from 2027, DC power block for a new hall.
- Agree grounding and protection philosophy with your electrical consultant before ordering DC equipment.
- Specify interlocked DC connectors, insulation or residual current monitoring, and breakers with known clearing times.
- Decide where storage sits (rack capacitors, in-row battery backup, facility BESS) and check Chapter XA.
- Check certification for every DC component and map approvals with your Electrical Inspector.
- Train electricians on DC faults, lock-out and arc flash, and plan DC acceptance testing.
- Write slip and substitution clauses into equipment and GPU contracts; Kyber and row-level DC gear carry 2027 dates.
How eCorpIT can help
eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our data centre and cloud networking team helps infrastructure buyers turn GPU and power roadmaps into dated facility requirements and compare retrofit options. Talk to us at /contact-us/.
Last updated: 7 November 2026.
