Nvidia, Google and Microsoft are standardising an open 800 volt direct current power system for AI data centres through the Open Compute Project, Nvidia said on Tuesday, with more than 80 equipment manufacturers building to the specifications.
The 800 VDC architecture replaces the chain of alternating current conversions that feed today’s 54 volt racks, cutting the number of steps between the grid and the chips. Nvidia estimates the design improves end-to-end efficiency by as much as 5 per cent and uses 45 per cent less copper. The company also puts maintenance savings at up to 70 per cent, mostly from fewer power supply failures. The three companies released version 0.3 of a solid-state transformer specification through the OCP in July, following a joint white paper in March.
“800 VDC unlocks the compute performance and power density required for AI at scale,” said Vladimir Troy, Nvidia’s vice president of data centre infrastructure. The rollout comes in three stages: an MGX-compatible 800 VDC power rack in the second half of 2026 that can run alongside existing AC equipment, a row power centre supporting up to 2 megawatts per row in 2027, and a facility-scale DC power block that converts grid power directly.
Google frames the open specification as a way to stop the industry splintering into incompatible designs. “Common 800 VDC interfaces can help the industry scale AI infrastructure while protecting the flexibility operators need in real deployments,” said Tom Garvens, Google’s vice president of data centre technology and systems. “By working through OCP, we can align requirements early, reduce fragmentation and give suppliers a clearer path to build interoperable products.”
The OCP work covers low-voltage DC designs up to 1,500 volts and gives operators two deployment paths: sidecar power racks feeding existing halls during retrofits, and direct conversion from medium-voltage AC for new sites. Component makers are already publishing supporting designs, including a Texas Instruments gallium nitride converter with 97.6 per cent peak efficiency.
The standard answers a power problem that is compounding every year. Average rack density has reached 27 kilowatts in 2026, up from 7 kilowatts in 2021, according to a Thornburg Investment Management report published this month, and Nvidia’s current GB200 systems draw 100 to 137 kilowatts per rack. The report projects 200 to 300 kilowatts for the coming Vera Rubin platform and more than 600 kilowatts for Rubin Ultra, which arrives on Kyber racks in 2027.
Thornburg equity research analyst Baadal Chaudhary calls the mismatch “Watts and Wafers”: chips now iterate on a yearly cycle while the infrastructure that powers them takes far longer to build. The same report puts transformer lead times at two to five years and switchgear at up to three. The grid interconnection queue in Northern Virginia runs to seven years.
Demand is showing up in supplier order books. Eaton reported data centre orders up 240 per cent in the Americas and a backlog 31 per cent higher year on year. Nvidia chief executive Jensen Huang has called energy “the first principle of AI infrastructure and the binding constraint on how much intelligence the system can produce.”
The announcement came a day after Nvidia said it would mobilise more than US$500 billion for AI chips, power and data centres with six Wall Street firms.
Australian operators will meet the new hardware under mandatory national standards due before parliament in early 2027, which require large AI data centres to put as much energy into the grid as they take out.
The first MGX-compatible power racks are due from Nvidia’s partners in the second half of this year. The 2 megawatt row power centre follows in 2027, when Rubin Ultra hardware is scheduled to reach data centre floors.
