Google, Microsoft and Nvidia are lining up behind 800-volt direct current as the standard way to power AI data centers, and they want the rest of the industry to follow the same blueprint.
The three companies are working through the Open Compute Project, the hardware consortium that gave the world open server and rack designs, to turn 800VDC into a common spec for power distribution. They are aligning on shared requirements for power conversion, power quality, system interfaces and safety, according to a Network World report published this week.
The pitch is simple: today's AI accelerators draw so much power that the old alternating-current distribution system no longer scales. Nvidia's technical blog lays out the numbers. When the company moved from its Hopper architecture to Blackwell, the power draw of a single GPU climbed about 75 percent. But because the NVLink domain grew to 72-GPU systems, rack power density jumped 3.4 times. Racks that once ran at tens of kilowatts are now pushing past 100 kilowatts, with the megawatt-per-rack mark on the horizon.
Delivering that much electricity at traditional voltages is physically awkward. The enormous current needed at low voltage produces resistive losses and demands huge amounts of copper. Nvidia says the same wire gauge can carry 157 percent more power at 800VDC than it can at 415VAC. A DC setup also needs three wires instead of four, which cuts conductor and connector counts further.
The savings add up fast. The Open Compute Project says an all-DC facility can cut copper usage by 50 to 80 percent and shave 8 to 12 percent off annual energy-related operating costs through lower conversion and distribution losses. For a 10-megawatt AI build, skipping the upstream AC stage can save $4 million to $8 million in capital costs. At gigawatt scale, that means millions of pounds of copper that never gets pulled. Those numbers matter to operators who are already watching power budgets swallow a growing share of their build costs.
Why AC is becoming the bottleneck
Traditional data centers step utility power down through a chain of transformers and conversion stages, ending at 415VAC, then each rack's power supplies convert that down to 54VDC for the compute trays. Every conversion wastes a little energy as heat, and end-to-end efficiency in these older designs can land below 90 percent.
The 800VDC approach collapses that chain. Medium-voltage AC gets converted straight to 800VDC at the facility level by large power conversion systems, then that DC backbone runs through the data hall to the racks. That removes layers of switchgear, transformers and rack-level power distribution units, and it frees floor space for compute.
There is a second problem that voltage alone does not fix, and it may be the harder one. AI training runs as a single synchronous workload. Thousands of GPUs compute in lockstep, then pause to exchange data, in near-perfect unison. That creates facility-wide power swings that are massive and abrupt — a rack can jump from about 30 percent utilization to 100 percent and back in milliseconds. Across a whole data hall, that is hundreds of megawatts ramping up and down in seconds, which joint research from Nvidia, Microsoft and OpenAI warns can cause grid-scale oscillations.
The proposed answer is multi-timescale energy storage treated as part of the power architecture, not as a backup. Short-duration storage — high-power capacitors and supercapacitors near the racks — absorbs the millisecond spikes. Facility-level battery systems at the utility interconnection handle the slower, larger shifts in the seconds-to-minutes range and ride through generator transfers. The 800VDC backbone makes it easier to place storage where it does the most good, because the batteries connect directly to the DC bus instead of being bolted onto an AC feed.
The industry is already converging
High-voltage DC is not a new idea outside the data center. Electric vehicles and utility-scale solar farms have run on 800VDC or higher for years, which means components and engineering practices already exist and can be adapted.
Nvidia's next-generation Kyber rack architecture is designed around 800VDC. Power flows at high voltage to each compute node, where a late-stage converter steps it down to 12 volts right next to the GPU. Nvidia says that single conversion stage occupies 26 percent less area than traditional multi-stage approaches, leaving more room near the processor.
The company has also lined up a broad partner list across the electrical ecosystem. Silicon suppliers include onsemi, Texas Instruments, Infineon, Analog Devices, STMicroelectronics and Renesas. Power system names include ABB, Eaton, GE Vernova, Hitachi Energy, Schneider Electric, Siemens and Vertiv. That breadth matters, because an open standard only works if equipment makers can build to one spec instead of designing separate power systems for each hyperscaler.
One notable absence: AWS
The company missing from the table is Amazon Web Services. AWS has made no public commitment to 800VDC and has not announced its own power architecture plans. The Network World report notes that Amazon is nonetheless redesigning its AI data centers for much higher density, with an internal initiative reportedly focused on next-generation infrastructure for Nvidia's GB200-class systems and future Vera Rubin hardware.
That silence is worth watching. AWS runs the largest share of public cloud capacity in the world, and its choices tend to set the floor for what equipment vendors build. If Amazon eventually adopts the same 800VDC spec, the standard gets an enormous installed base almost overnight. If it pursues a separate high-voltage design, suppliers face the cost of supporting two incompatible power ecosystems, and the industry loses the economies of scale that open standards are supposed to create.
There is also a timing question. Power delivery hardware has long lead times, and data centers being designed today will be built out over the next two to three years. A decision delayed is effectively a decision made, because the equipment orders for those facilities get placed months before the first rack arrives. The three companies pushing the standard know this, which is why they are publishing specs and lining up partners now rather than waiting for consensus to emerge on its own.
The three companies started presenting their 800VDC work through the Open Compute Project in 2025. The effort has moved past demonstrating feasibility and into writing specifications the supply chain can adopt. An OCP workstream is focused on converting medium-voltage AC directly to 800VDC, with aligned requirements for power quality, power smoothing and end-to-end interfaces.
Nvidia has published a technical whitepaper on the architecture and says the transition will happen in phases, giving the component ecosystem time to mature. The end state it describes is a native-DC data center where the utility connection, the energy storage and the compute racks all speak the same electrical language.
For Cloud & Edge Computing operators, the stakes are practical. Power is the constraint that decides where AI capacity gets built, and the industry has been watching grid interconnection queues stretch for years. A common 800VDC standard would not add a single megawatt to the grid, but it would stretch every megawatt further — less copper, fewer conversion losses, cooler halls and storage that can smooth the load instead of just backing it up.
The open question is how quickly the rest of the supply chain follows. OCP has a track record of turning consortium specs into shipped hardware, and the component vendors are already on board. The bigger test may be AWS, which is big enough to go its own way and has been known to. If the largest cloud operator eventually signs on, 800VDC stops being a club of three and becomes the default. If it does not, the industry gets two standards to juggle, which is precisely the fragmentation the Open Compute Project was created to avoid.
Either way, the direction of travel is clear. AI data centers are becoming power plants with compute attached, and the companies building them have decided that volts, not just chips, are the thing to standardize. The Network World report details the full effort, and Nvidia's blog lays out the technical case for the 800VDC ecosystem. For anyone planning the next wave of AI capacity, the practical question is no longer whether the industry moves to high-voltage DC, but how quickly the rest of the supply chain can build to the new spec.