Why an 800-volt AI rack can still live in an AC data centre

A data centre can supply an 800-volt computing rack while keeping alternating current in the building’s upstream electrical system. In one of NVIDIA’s proposed arrangements, a separate power rack stands beside the servers, accepts the existing AC supply and converts it to high-voltage direct current. The important change is the location of that conversion.[1]

Equipment suppliers are now giving that transition physical form. Delta’s October 2026 announcements describe an OCP portfolio spanning 800-volt DC and ±400-volt DC equipment, following a Singapore showcase on 29–30 September that included a 660-kilowatt power rack. That figure is an equipment output rating, not the capacity of a completed data centre.[2]

The current problem inside a dense rack

AI accelerator systems bring large amounts of computing hardware close together so processors can exchange data rapidly. More electrical power must then reach a compact area. NVIDIA’s engineering explanation identifies a constraint in continuing to distribute that power at 54 volts: the current, and the conducting material needed to carry it, become difficult to accommodate.[3]

At the same illustrative load, raising the distribution voltage from 54 to 800 volts reduces current by a factor of about 14.8: roughly one-fifteenth as much. This calculation describes current in the DC distribution path. It does not predict a comparable reduction in a building’s electricity bill. Computing still consumes power, and actual losses depend on the conductors and conversion equipment.

The voltage must come down again

The processor does not consume power at 800 volts. Texas Instruments describes possible intermediate conversion stages at 50 volts, 12.5 volts and 6.25 volts on the way towards the chip’s much lower operating voltage. Stepping the voltage down brings high current back. The design task is to keep that demanding part of the journey short, while balancing the efficiency, size and complexity of the converters.[5]

An 800-volt input does not mean every internal connection operates at 800 volts. NVIDIA’s current white paper describes near-term equipment that converts the incoming supply to 54 volts. Native 800-volt distribution inside the rack is a further development. Input and internal bus are different design layers.[4]

The retrofit route has a physical price. NVIDIA’s initial power-rack design supports about 660 kilowatts, but the paper flags additional floor space, numerous incoming AC connections and the need to confirm the existing row’s capacity. Moving conversion equipment out of the compute rack frees room there while using room elsewhere.[4]

Beyond the voltage label

The Open Compute Project’s March 2026 framework accommodates multiple DC architectures. Their value depends on the facility, its equipment and the load at which it operates. A comparison of efficiency therefore needs a defined electrical path and operating conditions. A design that removes a conversion stage in one building may require other changes in another.[6]

DC protection is part of that work. Direct current lacks alternating current’s natural zero-crossings, so an electrical arc can persist. NVIDIA’s white paper describes interlocked connectors and coordinated fault detection and isolation. Those interfaces matter when multiple suppliers’ converters, cables and computing racks must operate together.[4]

For someone assessing an 800-volt proposal, a single-line electrical diagram is a useful place to begin. Trace the existing AC supply to the new converter, then follow the DC path into the rack. Which equipment remains? Where does the conversion move? How much row space does it occupy, and where is a fault isolated? Those answers reveal how a particular building can accommodate the new computing load.

Report an error or suggest a correction