Feature

China races to build data centres in bid for AI supremacy

Part one of a three-part FT series on China’s data-centre push, reported from Ulanqab in Inner Mongolia. Per SemiAnalysis, China has 24GW of operational data-centre capacity — more than the rest of Asia combined but less than half the US’s 56GW — with a further 50GW under construction or announced. Ulanqab has 89 data centres built or planned and about 15GW of committed capacity, making it Asia’s fastest-growing hub according to an operator executive. The FT saw at least a dozen companies across four campuses, including Alibaba, Huawei (three data centres), ZData (converting an abandoned high school) and VNET; ByteDance is China’s largest renter of compute, and people briefed on the matter say DeepSeek and Z.ai are building there (Z.ai did not respond). Huawei’s Eric Xu chose Ulanqab for cheap power and its two-hour high-speed rail link to Beijing. Power is the draw: Inner Mongolia had 117GW of wind capacity by June (nearly four times the UK’s total) plus about 130GW of mostly coal-fired capacity, possibly the largest local electricity oversupply of any region (The Lantau Group’s David Fishman); data centres pay about US$0.05/kWh versus more than US$0.08 in many cities and about US$0.11 in Beijing, on the independent Mengxi grid, some with directly connected plants. Goldman Sachs puts construction costs about 20% below larger cities, and prefabricated builds take 12–18 months versus 18–24 in the US. People familiar with the policies say data centres using domestic processors instead of Nvidia get better tax benefits and power and water discounts. Caveats: chips are the binding constraint given US and Chinese restrictions on Nvidia purchases and limited Chinese fab capacity (“The question is whether the data centres can get enough chips” — Jefferies’ Edison Lee, who nonetheless sees “no risk of overbuilding”); water is scarce in the region; the piece gives no capex total.

Why it matters

Shows China competing on the AI-compute variables it controls — cheap power, fast build times and subsidies tilted towards domestic chips — while the chip constraint stays unresolved; a benchmark for Europe’s own data-centre and sovereign-compute ambitions, which face the opposite problem of expensive power and slow permitting.

Read the full article: Financial Times

More top stories