Adopt 800V DC in AI data centers

Power delivery has quietly become one of the biggest bottlenecks in AI data centers. GPUs keep getting denser, rack power keeps climbing, and the old way of feeding servers electricity is starting to look physically absurd. That is why 800V DC has moved from a niche engineering idea to an actual roadmap item. NVIDIA put that shift on the table in May 2025, and since then suppliers like Vertiv, Texas Instruments, and Schneider Electric have lined up around it for the next wave of AI infrastructure. ### Why are data centers even talking about 800V DC? Because AI racks are no longer normal server racks. NVIDIA says the industry is heading toward 1 MW IT racks and beyond, starting in 2027, while Schneider says many AI racks are already pushing past 400 kW. Traditional in-rack 48V or 54V DC distribution was built for kilowatt-scale equipment, not something that starts to resemble industrial power gear. (developer.nvidia.com) ### What breaks in the old setup? Low-voltage power means huge current. Huge current means thick copper, bulky busbars, more heat, and more conversion hardware stuffed into the rack. NVIDIA says a 1 MW rack on 54V would need power shelves taking up as much as 64U — basically the rack gets eaten by power gear instead of compute. TI puts the copper problem just as plainly: a 1 MW rack on today’s 48V distribution would need almost 450 pounds of copper. (developer.nvidia.com) ### So what does 800V DC change? Basically, it moves a lot of the ugly power conversion out of the compute rack and distributes power at a much higher voltage. That cuts current for the same power level, which shrinks copper needs and makes the system easier to route and cool. NVIDIA’s design centers on 800V high-voltage DC delivered to the rack, then stepped down near the chips. Schneider describes a similar model with separate “power racks” beside the IT racks. (developer.nvidia.com) ### Why not just stay with AC? Because every extra conversion step costs space, energy, and cooling. The appeal of 800V DC is not that DC is magically better in every context — it is that megawatt-class AI racks punish inefficiency much more than older data center designs did. NVIDIA says the new architecture could improve end-to-end efficiency by up to 5% and cut total cost of ownership by up to 30%. (developer.nvidia.com) Those are big numbers when a facility is scaling toward gigawatts. ### Is this real, or still a concept slide? It is early, but it is real enough that the ecosystem is organizing around dates and parts. NVIDIA tied 800V DC to its post-Blackwell roadmap and said deployment starts in 2027. Vertiv said in October 2025 that its 800V DC portfolio was moving from concept to engineering readiness, with product release planned for the second half of 2026 to support NVIDIA Rubin Ultra systems in 2027. (developer.nvidia.com) TI also said it is developing the sensing and power-management pieces for these systems. ### Who benefits if this happens? The winners are not just GPU vendors. Power-electronics companies, rectifier makers, busway suppliers, DC-DC converter vendors, and facility integrators all move closer to the center of AI capex. NVIDIA’s partner list already spans chip suppliers like Infineon, Navitas, STMicroelectronics, and TI, plus infrastructure names like Eaton, Schneider Electric, and Vertiv. (developer.nvidia.com) That tells you this is becoming a full-stack facilities buildout story, not just a server design tweak. ### What is the catch? Standards, safety, and deployment inertia. Data centers know AC deeply, and operators do not swap power architectures lightly. Schneider’s framing is telling here — it pitches adjacent “power racks” and modular integration precisely because nobody wants to rip apart existing sites overnight. So this is less a sudden flip than a migration path for new AI factories first. (developer.nvidia.com) ### Bottom line This is really a physics story disguised as an infrastructure trend. AI racks are getting so power-hungry that the old 48V and 54V playbook stops penciling out. 800V DC looks like the industry’s leading answer — not because it is fashionable, but because megawatt-scale compute is forcing the change. (developer.nvidia.com) (se.com)