800-Volt Power for AI Data Centers: Why 54 Volts Can't Feed a Megawatt Rack
Why AI data centers are moving to 800-volt DC power: a 1 MW rack at 54 volts needs up to 200 kg of copper and about 18,500 amps. How Nvidia's 800 VDC design works, what it saves, who builds it, and what the grid still has to solve.
Nvidia is preparing data center power for AI racks that draw a full megawatt each. Built the way today's racks are, one of those racks would need up to 200 kilograms of copper busbar just to carry its electricity (see Sources, [1]).
That single number explains why Nvidia is pushing the industry from 54-volt power inside the rack to 800-volt direct current across the whole building. This article walks through the physics, the claimed savings, the companies building the hardware, and the parts that are still unsolved.
From kilowatts to a megawatt
Rack power has climbed fast. Air cooling was optimized for racks of 8 to 12 kilowatts, while Nvidia's GB200 NVL72 rack pulls 120 to 130 kilowatts on its own (see Sources, [3]).
Nvidia's next step is far bigger. It says it is leading the transition to 800 VDC power infrastructure to support 1 MW IT racks and beyond, starting in 2027, and that the same infrastructure should cover racks from 100 kilowatts to more than a megawatt (see Sources, [1]).
Our AI data center power and cost calculator shows what that kind of load means for electricity use and cost; for why the same density forced the move to liquid cooling, see Liquid Cooling for AI Data Centers.
The 54-volt problem
Today's AI racks use 54-volt DC distribution inside the rack: thick copper busbars carry power from rack-mounted power shelves down to the compute trays. Nvidia says this design, built for kilowatt-scale racks, starts to hit physical limits once racks pass 200 kilowatts (see Sources, [1]).
It names three limits:
| Limit | What Nvidia says (see Sources, [1]) |
|---|---|
| Space | A GB200 NVL72 or GB300 NVL72 rack has up to eight power shelves. At megawatt scale with 54 VDC, power shelves would take up to 64 U of rack space, leaving no room for compute. |
| Copper | One 1 MW rack at 54 VDC needs up to 200 kg of copper busbar. The rack busbars alone in a 1 GW data center could need up to 200,000 kg. |
| Conversions | Repeated AC/DC transformations across the power chain waste energy and add failure points. |
Why higher voltage fixes it
Electrical power is voltage multiplied by current. Hold the power constant and raise the voltage, and the current falls in proportion.
Lower current means thinner conductors can carry the same load. Nvidia says switching row-level distribution from 415 VAC to 800 VDC busways lets 85% more power flow through the same conductor size and cuts copper requirements by 45%. It adds that DC distribution avoids AC-specific losses such as skin effect and reactive power (see Sources, [1]).
Fewer conversions, less in the rack
In Nvidia's design, 13.8 kV AC grid power is converted directly to 800 VDC at the data center perimeter, removing several AC/DC steps and many of the fan-cooled power supply units inside the racks. Inside the rack, 800 V comes in and is converted to 12 V for the GPUs. Taking rack-level AC/DC conversion out frees space for compute and improves cooling efficiency (see Sources, [1]).
The hardware is already on show: at GTC 2025, Nvidia exhibited an 800 V sidecar to power 576 Rubin Ultra GPUs in a single Kyber rack (see Sources, [1]).
Why now, if the idea is old?
Higher-voltage DC distribution has been piloted in data centers before, but adoption stalled on technical and deployment challenges. Nvidia argues three things have changed: AI-driven rack densities, advances in power conversion, and the industrial base built around electric-vehicle charging standards (see Sources, [1]).
Who builds it
Nvidia lists partners across the electrical chain (see Sources, [1]):
- Silicon: Analog Devices, Infineon, Innoscience, MPS, Navitas, onsemi, Renesas, ROHM, STMicroelectronics, Texas Instruments
- Power-system components: Delta, Flex Power, Lead Wealth, LiteOn, Megmeet
- Data center power systems: Eaton, Schneider Electric, Vertiv
What Nvidia says it saves
Nvidia claims up to 5% better end-to-end power efficiency, up to 70% lower maintenance costs from fewer power-supply failures, and up to 30% lower total cost of ownership (see Sources, [1]).
The catch
Nvidia itself says facility-wide 800 VDC brings new challenges in safety, standards and workforce training, and that it is studying both traditional transformer-based and solid-state transformer approaches. Its architecture also includes energy storage to absorb load spikes and sub-second GPU power swings (see Sources, [1]).
The grid side is behind. IEEE Spectrum reports that Nvidia and hyperscale developers are moving on 800 VDC far ahead of the power industry. In Texas, the grid operator ERCOT's Large Load Working Group is evaluating how to accommodate several 800 VDC facility designs, partly because AI loads can swing from almost nothing to hundreds of megawatts and back several times per second (see Sources, [2]).
Nvidia plans full-scale production of 800 VDC data centers alongside its Kyber rack-scale systems in 2027 (see Sources, [1]).
What it means
The bottleneck in AI infrastructure keeps moving: from chips, to memory, to cooling, and now to the copper and the voltage that carry power to the racks. Whether 800 VDC arrives on schedule depends less on Nvidia's hardware than on safety standards, trained crews, and grid operators who are still working out how to connect it.
This article is for educational and informational purposes only. It is not engineering, financial or investment advice.
Sources & References
- [1]NVIDIA 800 VDC Architecture Will Power the Next Generation of AI Factories — NVIDIA Technical Blog
- [2]How 800 VDC for Data Centers Tames AI Power Swings — IEEE Spectrum
- [3]Why AI rack densities make liquid cooling nonnegotiable — Network World