NVIDIA Adds Chinese Partner to 800V Power Ecosystem

On August 1, NVIDIA updated its official list of collaborators supporting its cutting-edge 800V high-voltage direct current (HVDC) power architecture. China's GaN technology leader, Innoscience, has officially joined the ecosystem and will provide a full-stack GaN-based power solution for NVIDIA's Kyber rack system—becoming the only Chinese company included in this collaboration. The announcement sent Innoscience's stock soaring nearly 64% on the Hong Kong exchange.

Back in May during Computex Taipei, NVIDIA unveiled a new alliance aimed at building an 800V HVDC power infrastructure for data centers—a critical step toward enabling 1-megawatt AI server racks by 2027. The alliance includes major names in semiconductors and power solutions, such as Infineon, MPS, Navitas, Rohm, STMicroelectronics, Texas Instruments, Delta, Flex Power, LiteOn, Megmeet, Eaton, Schneider Electric, and Vertiv.

On the same day, Navitas Semiconductor, a key power IC innovator, also confirmed its collaboration with NVIDIA for the Kyber system. The next-gen rack-level architecture will rely on Navitas' GaNFast™ and GeneSiC™ power technologies to supply GPUs like the Rubin Ultra. Navitas' stock surged up to 164% in U.S. trading following the announcement.

The shift to 800V architecture comes as AI data centers face mounting power demands and system complexity due to frequent voltage conversions. Traditional 54V DC racks are already nearing their 200kW power ceiling. NVIDIA's next-gen NVL72 racks may contain up to eight dedicated power bays just to meet 1MW demands—taking up significant space using legacy 54V setups.

To resolve this, the NVIDIA 800V HVDC architecture uses industrial-grade rectifiers to convert 13.8kV AC directly into 800V DC at the facility's perimeter. This streamlines power conversion, boosts end-to-end efficiency by 5%, and minimizes the number of internal conversion stages, reducing energy losses.

Beyond efficiency, the new architecture cuts the number of fan-cooled power units, resulting in lower thermal output, improved system reliability, and up to 70% less maintenance costs. With cooling accounting for a large portion of energy use in AI data centers, reducing thermal load is a major breakthrough.

By moving from 415V AC to 800V DC and utilizing 800V busbars for distribution, the architecture delivers 85% more power through the same conductor size. This allows thinner wires to handle the same load, cutting copper usage by 45% and avoiding typical AC inefficiencies like the skin effect and reactive power loss.