Nvidia’s 800VDC Tech Unlocks AI Power & Thermal Dissipation - Data Centre Magazine

Nvidia's shift from 48V to 800VDC is not a voltage bump—it is a systemic architectural reset forced by AI cluster power densities crossing 100MW. At that threshold, copper I²R losses and cascaded AC-DC conversion inefficiencies hit a hard ceiling. This is an architecture problem, not an optimization one. The ripple effect fractures the stack in both directions. Upstream, SiC and GaN power devices migrate from automotive to data-center grade, rewriting the valuation logic for Wolfspeed and Sanan Optoelectronics. Midstream, Delta and Flex face a generational product switch with 48V inventory exposure. Downstream, operators must rebuild transformer capacity and distribution topology from scratch—power engineering becomes the binding constraint on AI infrastructure for the first time. On competition, AMD's MI350 must deliver a parity solution before Q3 or lose the newly weighted "power efficiency" criterion in enterprise RFPs. Huawei's Ascend, paradoxically, gains a window: the standard is not yet locked, leaving room for a domestic power specification. Over 12–24 months, 800VDC will make liquid cooling mandatory rather than optional, pull power semiconductors into the AI BOM, and shift data-center siting logic from "near fiber" to "near the substation." Whoever owns the voltage standard owns the pricing power of the next compute generation.