800 VDC — NVIDIA 800 VDC architecture

NVIDIA 800 VDC architecture

A megawatt rack cannot be fed the way racks are fed today. Convert the grid once, at the edge of the site, and carry 800 volts of direct current all the way to the rack.

Scroll

01 — Today

Stepped down, again and again

Utility power arrives as medium-voltage AC. It is transformed to 415 V, reconditioned by a UPS, spread through PDUs and busways, turned into 54 V DC by power supplies inside the rack, then converted once more on every compute tray. Every handoff turns some of it into heat.

  • < 90%End-to-end efficiency, grid to GPUNVIDIA · Oct 2025

Height is voltage, on a log scale. Thickness is current. Cyan is AC, lime is 800 V DC, copper is low-voltage DC.

02 — 800 VDC

Convert once, at the edge

Solid-state transformers and industrial rectifiers at the perimeter turn 13.8 kV AC straight into 800 V DC and send it down the hall. The AC power supplies leave the rack, and their fans leave the power chain.

  • Up to 5%Better end-to-end efficiencyNVIDIA · May 2025
  • Up to 70%Lower maintenance costNVIDIA · May 2025

03 — Current

P = V × I

At one megawatt, 54 V means 18,519 amps. At 800 V it is 1,250: about 15 times less current, and in the same conductor about 1/219 of the resistive loss.

  • 200 kgCopper busbar in one 1 MW rack at 54 VNVIDIA · May 2025
  • 45%Less copper than 415 V AC distributionNVIDIA · May 2025
  • 1/219I²R loss in the same conductor, (54 / 800)²Worked out here

04 — Swing

Idle to flat out in milliseconds

Thousands of GPUs start and stop in step. A rack can swing from about 30% to 100% of its power and back within milliseconds. The grid should not feel it, so storage sits in between, on two timescales.

  • ms – sCapacitors · supercapacitorsNVIDIA · Oct 2025
  • s – minFacility battery storageNVIDIA · Oct 2025

Illustrative trace. The swing from about 30% to 100% is NVIDIA's figure; the rhythm is not.

  • Capacitors · supercapacitors: Milliseconds to seconds
  • Facility battery storage: Seconds to minutes

05 — Rack

576 GPUs, one feed

800 V enters the rack on two conductors and a protective earth. On each compute tray a 64:1 LLC converter takes it straight to 12 V, in 26% less area than a multi-stage converter. Kyber, the rack designed around it, holds 576 Rubin Ultra GPUs.

  • 1 MWPer rack, from 2027NVIDIA · May 2025
  • 64:1800 V to 12 V in one stepNVIDIA · Oct 2025
  • 576Rubin Ultra GPUs in one Kyber rackNVIDIA · May 2025

06 — Rack, row, building

Then the row, then the building

The same 800 V backbone carries racks from 100 kW to beyond a megawatt without being rebuilt, and it is being standardized in the open, through the Open Compute Project.

  • 80+Manufacturers building to the specificationNVIDIA · Aug 2026
← sections

07 — The road

On the calendar

  1. May 2025

    800 VDC architecture announced

  2. Oct 2025

    Partners across silicon, power components and data center power, at the OCP Global Summit

  3. Mar 2026

    Joint OCP white paper with Google and Microsoft

  4. Jul 2026

    OCP LVDC solid-state transformer specification, v0.3

  5. H2 2026

    MGX-compatible 800 VDC power rack

  6. 2027

    Row power center, up to 2 MW a row

  7. 2027

    800 VDC data centers in full production, with Kyber

  8. Next

    DC power block, at the scale of the building

Up to

  • 5%Better end-to-end power efficiency
  • 45%Less copper than 415 V AC distribution
  • 70%Lower maintenance cost, from fewer power supplies failing
  • 30%Lower total cost of ownership

Each figure is NVIDIA's own “up to”. Amps and the loss ratio are worked out here from P = V × I.

Sources

An independent study of NVIDIA's published design. Not affiliated with NVIDIA.

← sections
四宫格