The 800V Power Architecture

The conventional data center power architecture runs on 480V AC (or 415V in Europe). Power flows from the utility transformer → building switchgear → UPS → rack PDUs → server power supplies → voltage regulator modules → finally 12V or lower to the chip. Each conversion stage loses 3–7% of total energy as heat, adds physical equipment, and consumes floor space. In a 200 kW rack, conversion losses alone can total 15–20 kW.

NVIDIA announced at GTC 2025 that Rubin NVL72 and the Rubin Ultra Kyber rack would move to 800V HVDC (high-voltage direct current) distribution. In this architecture, power is converted once at a high level — from the utility's medium-voltage AC to 800V DC — and then distributed through busbars directly to the rack, where a single DC-DC stage steps it down for the GPUs. This eliminates multiple conversion stages, cuts conversion losses roughly in half, reduces copper cross-section requirements for a given power level (P = VI, so doubling voltage halves current for the same power), and frees up floor space.

Implementation requires new equipment across the entire power chain. Vertiv's 800V product line is scheduled for commercial availability in H2 2026. Eaton debuted its 800V reference architecture at OCP Summit 2025 and co-presented with NVIDIA at GTC 2026. Schneider Electric, ABB, Infineon, STMicro, and TI are all developing 800V components.

The enabling silicon is wide-bandgap semiconductors — silicon carbide (SiC) and gallium nitride (GaN). Traditional silicon MOSFETs cannot efficiently switch at 800V at the frequencies required; they either lose too much energy to switching losses or cannot handle the voltage at all. SiC and GaN can do both. SiC has roughly 3.3× the thermal conductivity of silicon and can operate above 200°C. GaN can switch at MHz frequencies, enabling smaller passive components and higher power density. The wide-bandgap semiconductor market is projected to exceed $5.3 billion in 2026. Analysts predict "all-GaN data centers" by 2028, where every conversion stage from grid to chip uses wide-bandgap silicon. Reported efficiency gains are up to 40% versus silicon incumbents.

Power in a typical data center takes an absurdly convoluted path before it reaches a chip. The utility delivers high-voltage alternating current. A transformer drops that to 480 volts. Switchgear distributes it around the building. A UPS converts it to direct current, stores some in batteries, then converts it back to alternating current. That AC goes into rack power distribution units. Inside each server, a power supply converts AC back to DC again. Then a voltage regulator chip on the motherboard drops the voltage once more to feed the GPU. That's five or six conversions, and each one wastes energy as heat.

In a small data center, the waste is annoying but tolerable. In a 600-kilowatt Kyber rack running state-of-the-art AI, the waste is unacceptable — you're throwing away tens of kilowatts per rack purely in the power conversion chain. Worse, all those conversion boxes take up physical space that could be holding chips instead.

NVIDIA's solution, which it is now mandating for its next-generation racks, is to skip most of those conversions. Power comes in at the building level, gets converted once to 800 volts DC, and flows straight through copper busbars to the rack, where one final step-down feeds the GPUs. Fewer conversions, less waste, less equipment, more floor space for computing. There's also a bonus from basic physics: when you double the voltage, you halve the current required for the same power, which means the copper cables can be thinner and cheaper.

But you can't just crank up the voltage on the old equipment. Ordinary silicon power transistors can't handle 800 volts efficiently — they either lose too much energy as heat or burn out. You need a new class of power chip made from silicon carbide (SiC) or gallium nitride (GaN). These "wide-bandgap" semiconductors can switch at much higher voltages and frequencies than silicon, they run cooler, and they are up to 40% more efficient in power conversion. The entire power infrastructure of the data center is being rebuilt around this new silicon.

So the 800V transition has two distinct sets of winners. First, the big power systems companies — Vertiv, Eaton, Hubbell, Powell — who make the transformers, busbars, switchgear, and rack power distribution units in the new architecture. Second, a smaller and more specialized group of wide-bandgap semiconductor makers — Wolfspeed, Navitas, ON Semiconductor — who make the SiC and GaN chips inside those systems. Both sets of winners benefit from the same physical fact: AI chips need more power, delivered more efficiently, than anything that came before.