Monday, October 13th 2025

NVIDIA Moves to 800 Volt DC Power Infrastructure to Drive Gigawatt-Scale AI

NVIDIA is implementing a major change in data center power architecture by using 800-volt DC distribution as the practical foundation for next-generation AI systems. Presented at the 2025 OCP Global Summit, the company's specifications and partner commitments frame 800 VDC not as an experimental idea but as a necessary response to racks that already draw well over 100 kW and are headed toward megawatt levels, soon to reach gigawatt of power per cluster. The core problem is simple and physical: low-voltage, high-current delivery forces bulky copper conductors and complex buswork into every rack, wasting space and money while creating thermal and installation headaches. Raising the distribution voltage reshapes those tradeoffs. At 800 volts, the same conductors can carry substantially more power, shrinking copper mass and freeing rack real estate for compute instead of power hardware.

NVIDIA's approach also simplifies the conversion chain by moving the main AC-to-DC rectification to the grid edge and using compact DC-DC modules in racks, reducing the number of conversion stages that generate heat and incur losses. Across large installations that efficiency improvement compounds. Even a few percentage points saved in delivery translates to megawatts less wasted energy at the gigawatt scale. Practical deployment plans tie the electrical advances to cooling and mechanical design: water-cooled busbars and integrated liquid cooling will be central to new rack designs that pack hundreds of GPUs while keeping temperatures and copper needs manageable.
NVIDIA is contributing designs to the Open Compute Project and working with power-infrastructure and component manufacturers to standardize connectors, protection, and safety practices. The result is likely a split market: greenfield gigawatt facilities built around 800 VDC, and legacy sites that remain on lower-voltage systems until major refreshes. For AI's appetite for scale, the higher-voltage path looks like the more sustainable way forward. NVIDIA's Kyber rack-scale system is the first in the adoption path. With 800 VDC, the same copper design as previous NVL systems can deliver over 150% more power, removing the need for massive 200-kilogram busbars to supply a single rack.
Source: NVIDIA
Add your own comment

27 Comments on NVIDIA Moves to 800 Volt DC Power Infrastructure to Drive Gigawatt-Scale AI

#26
dyonoctis
Vayra86Yes, so it enters a market and that market does get regulated. Nowhere does it mean regulation equals your prices never increase ;)
Beyond the prices, it seems that it was a bad idea when it comes to the quality of the service...and a bit weird because the (partially owned for a while) state-owned company is still around (EDF), and is still the owner of the network, they merely sell access to the network to the competition...if something bad happens to the hardware of said network, you as a customer can't call the competitor to fix it, you have to call Enedis, which EDF owns. 95% of the network is being taken care of and expanded by Enedis.

So in the end, those private companies don't really do much, besides piggybacking on the network. They won't build new nuclear plants; they have to buy that energy from EDF at a very low price, but EDF still gotta handle the maintenance of those plants alone.

EDF was partially privatised when the market was opened, but the state had to buy it back. Turns out, state money is needed to handle something so critical. Even with 75% of market share, there was a noticeable drop in performance. Along with a massive debt.
Posted on Reply
#27
Wirko
Dr. DroThat has gotta be one amazing piece of circuitry, alright. As far as I'm aware, VRMs don't have perfect conversion efficiency, so getting rid of that stage at the board level and supplying them with clean power from a single source should save, at this scale, tens of thousands of watts. Very impressive.
The VRMs next to each power hungry processor remain the same. The only difference is that the input voltage doesn't have to be 12V but can be 54V (mentioned in one of the slides here) or 48V (another datacenter standard) if that makes power delivery at the rack level more economical.
Posted on Reply
Add your own comment
Nov 17th, 2026 19:21 CST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts