China's datacentre fleet, counted by building

SemiAnalysis has extended its building-level datacentre model across China's border. Tracking more than 1,000 facilities from over 60 players, it puts China's fleet above 24GW, more than EMEA at about 14GW and APAC excluding China at about 15GW, though well behind the US at 56GW as of the end of 2026. The China figure leaves out roughly 20GW of dated pipeline and about 30GW of announced projects.

The case for counting from the ground up is that published estimates of China's capacity differ by 15 times. The largest tenant files no 10-K, several of the biggest landlords have never listed, and most primary sources are in Chinese.

Who is building

Spending has reached an inflection point. In the second quarter of 2026, the combined capex of Alibaba, Tencent and Baidu reached $20B, more than double a year earlier, and all three posted negative free cash flow for the first time on record. ByteDance, which is private and absent from that total, occupies roughly a fifth of delivered capacity in China and rents nearly all of it. GDS and VNET, the only Chinese landlords listed in the US, signed 1.3GW of wholesale orders in the first half of 2026 but captured barely a third of ByteDance and Alibaba orders from 2024 to date.

The state remains a large builder. The three carriers still own a third of national capacity. Grid companies' combined capex ended the 14th Five-Year Plan 24% over the original blueprint, and the 15th plan, for 2026 to 2030, adds another 40% on top, to more than $746B. SemiAnalysis says 100MW facilities are routinely delivered in under 12 months, and that overseas leasing by Chinese hyperscalers is set to double between 2026 and 2029, approaching about 4GW.

The report does not wave away the weak spots: vacancy rates are high, developers compete hard on price and chip supply is constrained by export restrictions. Its argument is that neither has stopped AI capacity from being built and filled quickly.