800 volts DC: rewiring the rack for the megawatt era

Here is a sentence that should give any data-center operator pause: the way we deliver power inside a server rack is about to hit a wall of physics. As racks climb past a few hundred kilowatts toward a full megawatt, the humble copper busbar simply cannot carry the current. NVIDIA's answer is to rewire the rack at 800 volts. It sounds like an electrical footnote. It is actually a redesign of the building, from the grid edge all the way to the chip.

The wall, in numbers

Today's racks distribute power internally at around 54 volts, an approach that runs out of road somewhere above 200 kW. Push past that and the current gets so high that the copper busbars become impractically thick, losses mount, and efficiency falls. The physics, not the budget, sets the ceiling (NVIDIA Technical Blog).

NVIDIA's fix is 800 V HVDC. At GTC 2025 it showed an 800-volt design able to feed 576 Rubin Ultra GPUs in a single Kyber rack. The architecture converts grid power (around 13.8 kV AC) straight to 800 V DC at the building perimeter via solid-state transformers, cutting conversion steps and shrinking the copper needed by up to 45%. By GTC 2026 a roster of power-component makers, including Navitas, STMicroelectronics and Texas Instruments, had lined up behind the standard (Embedded).

Why a voltage change is a building change

It's tempting to file this under "electrical engineering, not my problem." That would be a mistake. Moving to 800 V DC means converting grid power differently, at the perimeter rather than in stages, which changes the transformers, the distribution, the rack design and the safety regime. You cannot bolt an 800-volt rack into a hall plumbed for yesterday's architecture. The voltage decision propagates outward into the concrete.

This is the same pattern liquid cooling already taught us. A new generation of silicon didn't just demand a better chip; it demanded a different room. Now power delivery is doing the same. The megawatt rack is not a bigger version of today's rack. It's a different electrical animal, and the facilities that can host it are the ones designed for where power is going, not where it has been.

Power delivery is the new density frontier

Step back and the trend is unmistakable. The story of AI infrastructure is the story of cramming more power into less space, and every layer of the stack is being re-engineered to allow it: liquid to move the heat, 800 volts to move the current, denser racks to shrink the footprint. The constraint that began as "how many kilowatt-hours can you buy" has migrated inward to "can your building physically deliver them to the chip."

That makes power delivery, not just power supply, a competitive variable. The operator whose facility is built for high density and ready for the HVDC future can host the next generation of racks. The one whose hall maxes out at a few hundred kilowatts per rack on legacy distribution is holding a depreciating asset, no matter how cheap the electricity coming in.