AI rack densities are approaching MW scale by 2027. Explore three 800 VDC distribution architectures (rack, pod, and facility-level) for next-generation data centers.
As graphics processing unit (GPU), tensor processing unit (TPU), and other accelerator densities climb, rack densities will move toward megawatt-scale racks. At these densities, the physical limits of conventional 240/415 VAC delivery become unavoidable. More circuits, more copper, more cable trays, and less room for the compute hardware that justifies the facility in the first place.
The engineering math is straightforward and builds on proven and mature components from the electric vehicle (EV) industry. At constant current, doubling the voltage doubles the power a conductor can carry. Moving to 800 VDC cuts the number of power feeds and the volume of copper required to reach each rack, turning a cabling bottleneck into a solvable design problem.
Three models, one design question
The decision between rack-level, pod-level, and facility-level conversion depends on variables like rack density targets, facility layout constraints, and the acceptable trade-off between distributed resilience and centralized efficiency. A companion white paper, “800 VDC distribution for AI data centers: Architecture selection criteria and design elements,” explores these selection criteria and design elements in details, providing a framework for matching architecture to operational requirements.
- Rack-level conversion places rectifiers in a dedicated power rack adjacent to the IT rack, converting 400 or 480 VAC to 800 VDC. Battery backup can be integrated as a centralized UPS upstream or as battery backup units within the power rack.
- Pod or row-level conversion moves higher-capacity rectifiers to the end of a row or pod, distributing 800 VDC via power whips or overhead busway. This model scales naturally with pre-engineered pod deployments and pairs well with direct-to-chip liquid cooling.
- Facility or room-level conversion pushes rectification into gray space with multi-megawatt capacities. An alternative at this scale uses a solid-state transformer that combines MV isolation and AC-to-800 VDC rectification in a single liquid-cooled unit, replacing the conventional iron-core transformer entirely.
A common voltage cascade
Regardless of the architecture, the IT rack power train follows the same path: 800 VDC is stepped down to 54 VDC, then to 12 VDC, and finally to logic-level voltages (around 1 VDC for GPUs and CPUs). Voltage stability during battery discharge adds complexity. Without active regulation, the DC bus voltage drops as batteries deplete, creating instability at the IT load. Integrated DC-to-DC converters within the battery system maintain a steady bus voltage across all three models.
The shift to 800 VDC is an engineering response to a physical constraint that incremental AC optimization cannot resolve. Many of the components exist, some of the standards are established with others in development, and the architecture options are well-defined. The rack-level architecture will lead the way towards more centralized architectures.
