Why 800 Volts Direct Current

At the Computex trade show in Taipei in May 2025, NVIDIA said it would lead a move to 800-volt direct current power distribution in data centers, with deployment targeted for 2027 alongside its Kyber rack design for future Rubin Ultra processors. In a technical blog accompanying the announcement, the company described the problem it was trying to solve: delivering the power for a one-megawatt rack at the roughly 54 volts used inside today's racks would require up to about 200 kilograms of copper busbar in a single rack. The number is striking, but the physics behind it is simple.

Electrical power equals voltage multiplied by current. To deliver a fixed amount of power at low voltage, you need a large current, and large currents require thick conductors, which are heavy, expensive and take up space. The heat lost in a conductor also rises with the square of the current, so doubling current quadruples the loss. Raising the voltage lets the same power flow with a fraction of the current. Moving from about 54 volts to 800 volts cuts the current for a given power by roughly a factor of fifteen.

For most of the industry's history, racks drew somewhere between a few kilowatts and a few tens of kilowatts, and alternating current distribution with a final conversion inside each server worked well enough. Artificial intelligence systems have changed the arithmetic. The Open Compute Project, or OCP, an industry group founded by Facebook, now Meta, that publishes open hardware designs, has hosted a specification called Mount Diablo, written by Google, Meta and Microsoft, for racks drawing from about 100 kilowatts up to one megawatt. At those levels, the power equipment inside the rack starts to crowd out the computing equipment it serves.

Mount Diablo's answer is to disaggregate. Power conversion moves out of the server rack into a separate sidecar power rack, which feeds the computing rack at high voltage direct current, either plus or minus 400 volts or 800 volts. Microsoft has said this frees space for more accelerators in each computing rack. The choice of 400 volts was partly practical: electric vehicles already use components at that level, so a supply chain exists.

NVIDIA's proposal goes further, toward the whole building. In its design, medium-voltage alternating current from the utility, typically around 13.8 kilovolts, is converted to 800 volts direct current near the edge of the facility, possibly using a solid-state transformer, which is a power-electronics device that replaces the traditional iron-and-copper transformer. The 800-volt supply then runs down the rows to the racks, where it is stepped down to the low voltages chips require. The claim is fewer conversion steps, less copper and lower losses.

The idea has gathered partners. At the OCP Global Summit in October 2025, NVIDIA named data center power system suppliers including ABB, Eaton, GE Vernova, Hitachi Energy, Mitsubishi Electric, Schneider Electric, Siemens and Vertiv, and semiconductor makers including Infineon, Navitas, STMicroelectronics and Texas Instruments announced work on components. Several of those companies previewed equipment at the event.

Some caution is warranted. Direct current at 800 volts raises its own safety questions. Alternating current, or AC, passes through zero volts many times a second, which helps circuit breakers extinguish the arc that forms when a circuit opens; direct current, or DC, does not, so breaking a DC fault is harder and needs different equipment. Electrical codes, protective devices, training and maintenance practices built over a century for AC will need to adapt. Solid-state transformers at data center scale are still early in commercial deployment. And the efficiency gains NVIDIA and its partners describe are their own estimates, not yet the product of years of field operation.

Faraday's read is that the move toward high-voltage direct current inside data centers is well founded in physics and has unusually broad industry backing, but its timing depends on codes, safety equipment and supply chains that do not move at the speed of chip roadmaps. The final lesson draws the course together into a way of judging any power strategy.