AIコンピュートのスケーリングに新しい電力アーキテクチャが必要な理由Why Scaling AI Compute Performance Requires a New Power Architecture

The path electricity takes from a utility substation to an individual graphics processing unit is quietly becoming one of the defining engineering challenges of AI infrastructure. In a recent blog post, NVIDIA argues that conventional power distribution is struggling to keep up with the density of modern AI compute, and it makes the case for moving data centers toward an 800-volt direct current (DC) architecture that reduces conversion losses and enables more efficient scaling of GPU-heavy "AI factories."

NVIDIA's starting point is that every new generation of accelerated computing demands more from the infrastructure underneath it — more compute performance, higher rack density, and more efficient, scalable power distribution. Crucially, the company frames the limiting factor as more than a matter of total wattage. The bottleneck, it says, is how power gets from the grid to the GPU, and the many conversion and distribution steps along the way.

To understand the argument, it helps to look at how power typically flows today. Electricity usually arrives from the grid as high-voltage alternating current (AC), is stepped down and converted several times, distributed to racks as AC, and then converted again to DC inside servers before being stepped down to the low voltages that chips require. Each of these stages introduces losses, largely as waste heat, and each requires equipment that occupies space and adds cost. As racks pack in more accelerators, the current flowing through this chain rises sharply, and resistive losses in copper conductors grow with it.

An 800V DC approach aims to reduce both the number of conversion steps and the losses associated with high current. Because power equals voltage multiplied by current, delivering the same power at a higher voltage means lower current — which in turn allows thinner conductors, less copper, and reduced resistive (I²R) losses. NVIDIA's position is that pushing higher-voltage DC deeper into the data center, closer to the racks, can cut the cumulative inefficiencies that accumulate across today's multi-stage AC systems.

The timing reflects a steep rise in rack-level power. Dense AI systems that group many GPUs together can draw far more power per rack than the general-purpose servers that data centers were originally designed around, and the trajectory appears to be climbing toward and beyond 100 kilowatts per rack for the most demanding configurations. At those levels, the practical limits of moving large currents through busbars and cabling become a serious design constraint, which is likely why power architecture is receiving attention alongside the chips themselves.

The concept is not without precedent in adjacent industries. The electric vehicle sector has already moved from roughly 400-volt to 800-volt systems to enable faster charging and reduce losses, demonstrating that higher-voltage DC can be handled safely and efficiently at scale. Data center power standards efforts, including work associated with the Open Compute Project, have similarly been examining higher-voltage and DC distribution as rack densities climb. NVIDIA's proposal appears to fit within this broader industry direction rather than standing entirely on its own.

Realizing an 800V DC ecosystem, however, would require coordination across a large supply chain. Power supplies, busbars, connectors, protection and safety systems, and rack designs would all need to align around common specifications, and vendors of transformers, power distribution units, and cooling would need to adapt. Because a single company cannot deliver this alone, the practical rollout is likely to depend on partners and standardization efforts, and on retrofitting or purpose-building facilities to accommodate the new topology.

For operators, the potential payoff is meaningful. Reducing conversion and distribution losses translates directly into lower energy waste, and freeing up space