SemiAnalysis Maps China's AI Build-Out: 1,000+ Data Centers and 24 GW Delivered

For years, anyone trying to size China's AI infrastructure was working half-blind. The biggest tenant doesn't publish a 10-K, several of the largest landlords have never listed, and most of the primary documents are in Chinese. On Thursday, SemiAnalysis published a building-by-building count. Its new China Datacenter Model tracks more than 1,000 facilities run by over 60 operators and puts the country's delivered fleet at more than 24 gigawatts. By the firm's numbers, that makes China larger than all of EMEA (about 14 GW) and larger than the rest of Asia-Pacific combined (about 15 GW). Only the United States is bigger, at a projected 56 GW by the end of 2026.

The headline figure also leaves out a lot. SemiAnalysis excludes roughly 20 GW of dated pipeline and another roughly 30 GW of announced projects. The report, written by Everlyn, Dylan Patel and Patrick Schaabi, starts by going after what it calls two lazy assumptions the market had settled on: "China is big, and China is empty." It notes that published estimates of China's capacity have differed by as much as 15x.

One tenant, a fifth of the country

The model's most notable finding is how much of the market depends on ByteDance. SemiAnalysis estimates the TikTok parent occupies about one-fifth of all delivered data center capacity in China and rents nearly all of it. That makes ByteDance the most important customer for every wholesale colocation provider in the country. Even without ByteDance, which is private, the listed players are spending at a record pace. Alibaba, Tencent and Baidu together spent $20 billion on capex in the second quarter of 2026, more than double the year before. SemiAnalysis says it is the first time on record that all three posted negative free cash flow in the same quarter.

Adding ByteDance, the firm puts combined capex for the four companies at roughly $35 billion in 2024 and more than $50 billion in 2025, on a path to about $100 billion in 2026. TrendForce reports a similar trend. In an August 3 forecast, it projected that combined 2026 capex for ByteDance, Tencent, Alibaba and Baidu would rise more than 80% year over year, with ByteDance growing the most. TrendForce says that money is going mainly into large AI data centers, in-house ASIC development and GPU clusters.

The two US-listed Chinese landlords, GDS and VNET, signed about 1.3 GW of wholesale orders in the first half of 2026 and less than 10 MW of retail. Even so, SemiAnalysis says the pair won only about a third of ByteDance and Alibaba orders between 2024 and 2026 year to date. The rest went to unlisted developers, state carriers and hyperscaler self-builds. The report sums up the shift this way: "The AI era is wholesale. The legacy retail stock is not participating."

Built retail-first, then flipped

SemiAnalysis says China's market grew up very differently from America's hyperscale-first model. For its first two decades, it was run by China Mobile, China Telecom and China Unicom, which together held 60-70% market share and rented low-density racks, about 90% of them under 2 kW, to websites, game studios and content delivery networks. A cloud price war from 2015 to 2021 was followed by a sharp slowdown in 2022 and 2023. Tencent's management described that period as a move "from pursuing revenue growth to healthy growth," and Tencent cut its capex budget by nearly half in mid-2022.

That history explains the vacancy numbers critics point to. SemiAnalysis says the national utilization rate of about 50% is "largely a monument" to older retail stock that can't be retrofitted for H20 or Ascend servers. Meanwhile, wholesale AI buildings are filling quickly. At VNET, wholesale utilization is back above 70% while legacy retail racks sit near 60%. This is a different picture from 2025 reporting by MIT Technology Review, later picked up by Tom's Hardware, which described facilities built in 2023 and 2024 that sat underused as GPU rental prices collapsed. Monthly rent for an eight-GPU H100 server fell from ¥180,000 to ¥75,000.

The building is also moving west. Under the Eastern Data, Western Compute program the NDRC launched in February 2022, energy-consumption quotas effectively stopped new construction outside eight national hubs. Inner Mongolia, where power costs roughly half the Tier-1 city rate, has become what SemiAnalysis calls "China's Johor." Speed is the other advantage. The standard delivery time for a 100 MW facility has dropped from about 18 months to about 12, and permitting takes three to six months. Alibaba's modular CUBE 5.0 design goes from finished shell to server arrival in 100 days.

Why It Matters

The report reframes the US-China infrastructure race. According to SemiAnalysis, "The US buildout is power-gated" by interconnection queues and transformer lead times, while "China's is chip-gated." China has abundant power, state utilities that welcome data center load, and plentiful construction labor, so it can put up buildings faster and more cheaply than the West. In that framing, the real constraint on Chinese AI capacity is access to chips under US export controls, not buildings or power. That makes export policy the key lever, and it makes the domestic accelerator ramp at Huawei and others the number to watch. SemiAnalysis also expects Chinese hyperscalers' overseas leasing to double between 2026 and 2029 to about 4 GW. That figure doesn't count the hundreds of thousands of GPUs they rent from Western clouds, so export controls on hardware shipped into China cover only part of their compute.

What to Watch

The best indicator will be ByteDance's tender calendar, which SemiAnalysis says effectively sets the order book for every wholesale operator in China. It is also worth watching whether GDS's estimate of about 18 more months to absorb the pricing reset holds up. Two Ningxia entities ByteDance registered in July point to a possible fourth self-build cluster. Beyond that, the 20 GW dated pipeline will show how many projects actually get delivered and whether Chinese AI chip supply grows fast enough to fill the new buildings.