Access HPS Whitepaper: Powering the AI Factory: The Transition from AC to 800 VDC Architecture

Abstract

AI rack power has climbed from tens of kilowatts to several hundred within two GPU generations, with roadmaps converging on one-megawatt racks before the decade ends. The conventional train — medium-voltage AC stepped to 415/480VAC, double-conversion UPS, PDUs, to in-rack 48VDC — fails on loss, copper, footprint, component count, and storage integration. The answer, per NVIDIA and OCP white paper, is a single conversion from medium-voltage AC to a facility-level 800VDC backbone. This paper argues that the line-frequency transformer (LFT) plus solid-state rectifier (Transformer Rectifier Unit or TRU) is the right way to build it now. A TRU is estimated to deliver ≈98.5% MV-to-800VDC efficiency — matching today's solid-state transformer (SST) demonstrators — with a 1:50–100 component-count advantage, MTBF above 250,000 hours, 40–50 year transformer life, better than 3:1 CAPEX advantage, and decades of gigawatt-class field history in both passive multi-pulse (AEG’s DC3) and active front-end (AEG’s DC4) form. The same platform should be capable of scaling to 1500 VDC for gigawatt AI factories of the future. Central rectification, not the rack-row sidecar, is the target architecture; the sidecar leaves the AC bottleneck intact. SSTs remain promising but unproven at production scale. AI factories can be energized sooner with TRU while remaining future proof.