China’s AI-infrastructure advantage may sit between the grid and the chip
Envision Energy has begun powering live computing workloads with an 800-volt direct-current system at its Galaxy data centre in Inner Mongolia, The Information reports. The commercial pilot is only 5MW, but it is operating before Nvidia and its partners deploy comparable infrastructure for their next generation of racks.
Conventional facilities convert electricity several times between alternating and direct current before it reaches a server. Envision’s solid-state transformers convert 10,000-volt AC from the grid directly to an 800V DC bus. The company says this lifts conversion efficiency from roughly 94–95% to 98.5%, while its battery system absorbs rapid changes in GPU load. Those figures are vendor claims, and a new power path must prove reliability around exceptionally expensive chips.
The timing reflects an unexpected consequence of export controls. Chinese accelerators individually perform less work than Nvidia’s best chips, so Chinese companies assemble more of them into “super racks” that can draw 400–500kW. Nvidia’s current Blackwell racks use about 140kW; its planned 2027 Kyber racks may reach 600kW. Both markets are therefore converging on high-voltage DC even though their compute architectures differ.
Envision’s advantage is integration. It builds wind turbines, batteries, transformers and control software, reducing coordination across suppliers. Its system costs 10–15% more upfront, and the company expects about two years before broad adoption. The pilot does not establish global leadership, but it shows why the AI race cannot be read from chip benchmarks alone: usable compute depends on how efficiently a country can manufacture, connect and operate the electrical system around the silicon.