The power problem in AI is usually described at the scale of the grid: gigawatts of new load, years-long interconnection queues, gas turbines and nuclear deals. Nvidia's latest push is about the last few meters. At the Open Compute Project Global Summit this week, the company and about 20 industry partners showed equipment for 800-volt direct current data centers, a redesign of how power reaches the chips that Nvidia says is needed once a single rack draws a megawatt.
"Moving to 800VDC infrastructure from traditional 415 or 480 VAC three-phase systems offers increased scalability, improved energy efficiency, reduced materials usage, and higher capacity for performance in data centers," Nvidia said, as reported by Data Center Dynamics.
Why the old design breaks
Today's AI racks distribute power inside the rack at 54 volts DC. Copper busbars carry electricity from power shelves to compute trays. Nvidia's engineers wrote in May that this approach begins to hit physical limits once racks exceed 200 kW.
The numbers they gave are striking. The company's current GB200 NVL72 and GB300 NVL72 racks already use up to eight power shelves. Powering a megawatt-scale Kyber rack at 54 volts would require power shelves taking up to 64 units of rack space, leaving no room for compute. A single 1 MW rack would need up to 200 kg of copper busbar, and the rack busbars in a 1 GW data center could require up to 200,000 kg of copper. Repeated conversions between alternating and direct current also waste energy and add failure points.
What changes
Nvidia's design converts 13.8 kV alternating current from the grid directly to 800 volts DC at the edge of the data center using industrial rectifiers, eliminating most intermediate conversion steps. Row-level busways then carry 800 volts DC to the racks, which receive it directly, removing the AC to DC conversion inside each rack and freeing space for computing hardware.
The company claimed several gains in its May technical post. It said 800-volt DC busways can carry 85% more power through the same conductor size as 415-volt AC distribution while cutting copper requirements by 45%. It projected up to a 5% improvement in end-to-end power efficiency, up to a 70% reduction in maintenance costs and the ability to scale from 100 kW to more than 1 MW per rack on the same infrastructure. Full-scale production of 800-volt data centers is meant to coincide with Kyber systems in 2027, which Nvidia said could cut total cost of ownership by up to 30%. This week the company went further, saying more than 150% more power can be transmitted through the same copper at 800 volts DC compared with today's systems.
These are vendor claims, not independent measurements. But the direction is clear, and the electric vehicle and solar industries have already moved to 800-volt DC for similar reasons, Nvidia noted.
Kyber and Vera Rubin
The architecture is built for Kyber, the successor to Nvidia's current Oberon rack design, which will house 576 Rubin Ultra GPUs by 2027. Kyber stands compute blades vertically, "like books on a shelf," with up to 18 blades per chassis and switch blades connected through a cable-free midplane at the back.
Nvidia also outlined the Vera Rubin NVL72 rack, which uses 45°C liquid cooling, a new liquid-cooled busbar and 20 times more energy storage to keep power steady. That last feature matters for the grid as well as the chips. Large AI training jobs can swing power demand as thousands of GPUs start and stop together, and storage inside the rack helps smooth those swings before they reach the utility connection.
The design also cuts the number of power supply units with fans in the power chain. Nvidia said fewer units mean higher reliability, less heat to remove and better efficiency, and that a single conversion from AC to DC reduces electrical complexity and maintenance needs.
Who is on board
The partner list spans the electrical supply chain. Silicon suppliers include Analog Devices, Infineon, Innoscience, MPS, Navitas, onsemi, Renesas, ROHM, STMicroelectronics and Texas Instruments. Power system makers Eaton, Schneider Electric and Vertiv are involved, and Vertiv unveiled an 800-volt DC reference architecture combining power and cooling. HPE announced support for Kyber.
Several chipmakers said the shift requires new power semiconductors, particularly silicon carbide and gallium nitride, to handle the higher voltages efficiently, according to DCD.
Operators are also signing up. CoreWeave, Lambda, Nebius, Oracle Cloud Infrastructure and Together AI are designing for 800-volt data centers, Nvidia said, and Foxconn's 40 MW Kaohsiung-1 data center in Taiwan is being built for the architecture.
What it means for power demand
Efficiency gains inside the building do not shrink the AI build. A 5% improvement in power delivery efficiency is meaningful at gigawatt scale, but Nvidia's goal is to let each rack draw far more power, not less. The architecture is designed to scale from 100 kW to more than 1 MW per rack, and the 54-volt systems it replaces start to struggle above 200 kW.
What it does change is how quickly that power can be put to use. Fewer conversion stages, less copper and fewer power supplies mean less equipment to source, at a time when transformers, switchgear and other electrical gear are in short supply. Denser racks also mean a given amount of computing needs less floor space, which can make it easier to fit large clusters onto sites where power is already available.
What to watch
The milestones are hardware shipments. Kyber and Rubin Ultra are scheduled for 2027, and the first 800-volt facilities such as Kaohsiung-1 will show whether the promised efficiency and maintenance gains hold up in operation. Nvidia acknowledged in May that further innovation is needed to deliver the expected overcurrent protection and reliability gains.
For utilities and grid planners, the takeaway is that the unit of AI demand is getting bigger. When a single rack draws a megawatt, the facility around it starts to look less like a server room and more like an industrial plant, with its own medium-voltage rectifiers, DC busways and on-site storage.