On September 25, SemiAnalysis published a preview of its China datacenter model. The authors say published estimates of China's datacenter capacity differ by 15 times, while reports keep citing high vacancy, so the market kept two easy assumptions at once: China is large, and China is empty. Their building-level count is different. Tracking more than 1,000 facilities across more than 60 operators, they put delivered capacity in China above 24 GW. That figure excludes about 20 GW of dated pipeline and about 30 GW of announced projects.
They set the global model beside it. As of year-end 2026, the United States has 56 GW, Asia-Pacific outside China about 15 GW, EMEA about 14 GW, and Latin America about 2 GW. On that basis, China's delivered fleet is larger than EMEA and larger than the rest of Asia. The global model previously stopped at the Chinese border.
[1]
The authors put combined capital spending by Alibaba, Tencent, and Baidu in the second quarter of 2026 at $20 billion, more than double the year before, with all three posting negative free cash flow at the same time. They call it the largest capex step-up on record for the sector. ByteDance is private, so it is outside that $20 billion. In their China model, ByteDance alone occupies roughly a fifth of delivered datacenter capacity in the country and rents nearly all of it, which makes it the most important customer for wholesale colocation operators there.
GDS and VNET, the two Chinese datacenter landlords listed in the United States, signed 1.3 GW of wholesale orders in the first half of 2026. The model says that, from 2024 through 2026 year to date, they captured barely a third of ByteDance and Alibaba orders. State-owned carriers still own about a third of national capacity. Separately, the authors say combined capital spending by the national power-grid companies ended the 14th Five-Year Plan 24 percent above the original plan, and that the 15th plan, for 2026 through 2030, adds another 40 percent, to more than $746 billion, or 5 trillion yuan. That is grid-company spending, not datacenter-owner spending.
On delivery speed, they say China routinely delivers 100 MW facilities in under 12 months, that modular buildings are the standard method, and that Tencent deployed a third-generation modular design in 2014. Overseas leasing by Chinese hyperscalers, in the global model, is set to double from 2026 to 2029 and approach about 4 GW. That number excludes GPUs rented from Western clouds.
Vacancy is not described as gone. The authors say vacancy rates are still high, developers compete heavily on price, and export controls constrain chip supply, and that neither has stopped AI datacenters from being built and filled. Citing a VNET disclosure, they describe one leading colocation operator whose utilization fell from the mid-70 percent range to the mid-50 percent range between 2010 and 2022. From 2023, wholesale buildings filled back above 70 percent on AI demand, while legacy retail racks sat near 60 percent. The history they give is a market built first as telecom retail racks: in the 2010s the three state carriers together held about 60 to 70 percent, and about 90 percent of facilities ran below 2 kW.
[1]The 24 GW figure is an output of the SemiAnalysis model, not a government statistical release. The public page says the tenant-by-tenant detail for ByteDance, Alibaba, Tencent, Baidu, and Huawei sits behind the paywall. This article does not add those figures. The approximate marks in the source, including the roughly one-fifth share, stay approximate. They are not an audit to the last digit.
[1]要点
- The model puts delivered capacity in China above 24 GW, excluding announced projects that are not yet delivered.
- ByteDance is about a fifth of delivered capacity and rents nearly all of it.
- Vacancy is still high, while the authors say wholesale buildings have filled faster on AI demand.