Three 800 VDC architecture models for MW-scale AI racks. Learn why the location where the conversion boundary sits shapes conductor routing, fault domains, and maintenance.
There is a point at which adding more AC circuits to a rack stops being an engineering trade-off and becomes a physical impossibility. AI accelerator roadmaps are approaching that point. With chip vendors projecting megawatt-scale rack densities by 2027, the number of 240/415 VAC feeds required to power a single rack exceeds what the cable tray, floor space, and connector density can accommodate.
Raising the distribution voltage to 800 VDC directly addresses the volume problem. At constant current, doubling the voltage doubles the power carried per conductor. Fewer feeds, less copper, and smaller cable pathways per rack.
But voltage alone does not make an architecture. The critical question is where the AC-to-DC conversion happens, and that decision shapes everything downstream: conductor routing, DC circuit breaker selectivity, space allocation, and fault domain boundaries.
A component ecosystem that already exists
One reason 800 VDC is viable now is the electric vehicle (EV) industry. EV platforms have operated at 400 to 800 VDC for years, driving volume production of connectors, contactors, circuit breakers, and safety devices at this voltage class. Regulations and standards have been established: IEC 62196-3 for connectors, IEC 60947-4-1 for contactors, IEC 60947-2 for protective devices. Technologies such as high-voltage interlock loops and pyrofuses have been field-validated on millions of vehicles. Data center power infrastructure can adopt these components after certification for data center use. This saves a lot of time in developing new devices.
Three conversion boundaries, three different facilities
At the rack level, a dedicated power rack (sidecar) sits adjacent to the IT rack and converts 400 or 480 VAC into 800 VDC. The latest configurations use 110 kW three-phase rectifier shelves and can supply over 600 kW per power rack. On the IT side, 800 VDC power shelves occupy 1U at 90 kW. Battery backup sits either upstream as a centralized UPS or inside the power rack as battery backup units, though the latter consumes rectifier shelf positions.
At the row or pod level, higher-capacity rectifier systems serve one or two rows of racks, with 800 VDC distributed through power whips or overhead busway. Pre-engineered pod configurations integrate rectification at the end of the row with direct-to-chip liquid-cooling infrastructure. Battery backup is either internal to the conversion system or housed in a separate adjacent cabinet. This model reduces the number of conversion points per rack while keeping the power electronics inside the IT space.
At the facility or room level, rectification moves into the gray space (the data hall perimeter or maintenance corridors) at multi-megawatt scale. A notable alternative here is the solid-state transformer, which replaces both the conventional iron-core MV isolation transformer and the rectifier system. Using power electronics and compact high-frequency transformers with self-contained liquid cooling, the SST performs isolation and AC-to-800 VDC rectification in a single unit. It consolidates two major pieces of infrastructure into one, with potential footprint and efficiency gains.
What stays the same across all three
No matter where the AC-to-DC boundary sits, the voltage cascade inside the IT rack is identical. DC/DC converters step 800 VDC down to approximately 54 VDC at the rack busbar. Intermediate bus converters on the server board reduce that to 12 VDC. Voltage regulator modules on the motherboard deliver the final logic-level voltages: around 1 VDC for GPUs and CPUs, with 3.3 V and 5 V rails for NICs, memory, and supporting components. Each stage introduces conversion losses, making stage count and converter efficiency a shared design constraint.
Voltage stability during battery discharge is another concern. As batteries deplete, bus voltage drops continuously without active regulation, creating instability at the IT load. Integrated DC-to-DC converters within the battery system hold the bus voltage steady. This applies equally whether batteries sit in a sidecar or in a multi-megawatt gray-space installation.
A key design variable
The choice between rack, row, and facility-level conversion is ultimately about complexity. Distributed conversion keeps fault domains small but multiplies the number of conversion points. Centralized conversion consolidates infrastructure but concentrates risk. Neither approach is inherently superior. The right boundary depends on rack density targets, cooling topology, facility layout, and how the operations team intends to manage failures. The physics of 800 VDC remain constant; the architecture is the variable.
