THE CRUNCH

China is building an 800MW AI data centre campus in Inner Mongolia that could undercut the economics of America's Stargate project, according to a SemiAnalysis report. Envision has completed the first 120MW of the Galaxy campus in Ulanqab in under a year, at a cost the researchers put at roughly one-fifth of comparable US projects. When finished by the end of 2027, the campus is planned to offer up to 800MW of IT capacity, close to the roughly 820MW at Stargate's Abilene, Texas site.

The power strategy is the distinctive part. Rather than relying on the local grid, Envision is building wind farms and four-hour battery storage alongside the servers. Its wind farms, about 80 kilometres apart, have a combined 420MW of capacity, comfortably above the 240MW load expected from the first two phases, though Ulanqab's wind capacity factor of 35 to 40 percent means the batteries are needed to keep electricity flowing when the wind drops. US developers facing grid constraints and turbine shortages have instead leaned on on-site gas turbines.

Inside a 5MW hall, solid-state transformers convert 10kV AC straight to 800V DC for the racks, cutting out the low-voltage transformers, uninterruptible power supplies and rectifiers found in conventional data centre power systems. Fewer conversion stages can reduce equipment and energy losses, although SemiAnalysis did not measure an efficiency gain for the pilot. The design also aligns with Nvidia's push towards 800V DC distribution for future high-density AI systems. Cooling uses a single water loop shared by liquid-cooled and air-cooled servers, eliminating central chiller plants.

Galaxy is the first piece of Envision's wider Inner Mongolia plans, which SemiAnalysis puts at 3GW, with Alibaba, Huawei and ByteDance among the firms expanding in a region that offers cheap electricity, land and wind. But buildings and power do not equal computing power: US export restrictions still limit China's access to Nvidia's most advanced chips, a separate constraint on turning the campus into usable AI capacity.