Model tracks over 1,000 Chinese datacenters – capacity estimates diverge by a factor of 15
The analysis firm SemiAnalysis published a building-level model of more than 1,000 Chinese datacenters on September 25, claiming that the installed fleet of over 24 gigawatts is far larger than the "empty China" narrative suggests. But every key figure comes from a single paid modeling effort that cannot be independently verified – not from public data.
The analysis firm SemiAnalysis published its so-called China Datacenter Model on September 25, 2026, a tracking tool that, according to the company, covers more than 1,000 datacenter facilities spread across more than 60 operators in China. The headline finding: China's installed datacenter capacity exceeds 24 gigawatts – larger than EMEA (~14 GW) and the rest of Asia excluding China (~15 GW) individually, but still under half of the United States' 56 gigawatts at the end of 2026.
According to the company, the tracking shows that China alone has a fleet of over 24 GW – larger than EMEA and larger than the rest of Asia. The figures are SemiAnalysis' own. The underlying data is behind a paywall, and none of the key figures about the Chinese market can be checked against independent sources from the available documentation. According to the company itself, this is the first building-level measurement of a market where published capacity estimates diverge by a factor of 15 – delivered by a single company with its own model.
Why the market has been flying blind
SemiAnalysis' critique of existing coverage is aimed at what the company calls two "lazy assumptions": that China is big, and that China is empty. The problem, according to the company, is that most primary sources on Chinese datacenters are in Chinese, and that Western analyses have therefore cited one another rather than the facilities themselves. Reports have repeatedly quoted high vacancy percentages.
The company suggests that the high vacancy figures in circulation are wrong, but the company's own vacancy measurements are not included in the publicly available portion of the report. The question of how full the Chinese facilities actually are remains open.
The size map: China versus the rest of the world
According to SemiAnalysis' global model, which tracks over 5,000 facilities across four regions, the capacity picture at the end of 2026 looks like this:
- United States: 56 GW – still the world leader according to the model
- China: over 24 GW
- APAC excluding China: ~15 GW
- EMEA: ~14 GW
- Latin America: ~2 GW
China's fleet is thus, according to the model, larger than EMEA and larger than the rest of Asia excluding China – each region taken individually – but less than half of the US's. That is a different picture than the "empty China" narrative, but also different from a China that has caught up with the US.
In addition to the installed base, SemiAnalysis counts roughly 20 GW in a "dated" project pipeline – that is, planned construction that SemiAnalysis itself treats as uncertain or outdated – and a further roughly 30 GW in announced projects. The installed figure of 24 GW excludes both.
The demand side: the BAT record and ByteDance's dominance
What SemiAnalysis describes as the largest capex increase in the sector's history happened in the second quarter of 2026: combined investment by Alibaba, Tencent and Baidu reached 20 billion dollars, more than doubling year over year. According to the company, all three posted negative free cash flow at the same time for the first time ever – a sign, according to the analysis, that the companies are investing faster than they earn.
The single most important customer for the Chinese colocation market is, according to the model, ByteDance. The company is said to occupy roughly a fifth of delivered datacenter capacity in China on its own, and to rent nearly all of it. That is a striking share for a company that is not listed and files no financial statements – the figure is SemiAnalysis' estimate, not a company-reported number. According to the analysis, this makes ByteDance the most important single customer for all wholesale colocation operators in the country, pointing to significant customer concentration in the market.
The state layer: pillars and the power grid
Beneath the commercial operators sits a state layer. The telecom carriers that held the majority of the market in the 2010s still own a third of national capacity, according to SemiAnalysis. And investment in the power grid has taken off: the grid operators' combined capex accelerated from 2024 and ended the 14th Five-Year Plan 24 percent above the original budget framework. The 15th Five-Year Plan (2026–2030), according to the company, adds a further 40 percent on top of this, to over 746 billion dollars (around 5,000 billion yuan).
This is partly public Chinese plans that SemiAnalysis builds upon, and partly the company's own interpretation – but it points to something that distinguishes the Chinese buildout from the American one: capacity expansion is tied directly to state investment frameworks for power supply.
The international dimension: Chinese capacity moves outward
SemiAnalysis predicts that Chinese hyperscalers' overseas leasing will double from 2026 to 2029, to nearly 4 GW of leased capacity. That is a forecast, not a measured figure – but the simultaneous news from Alibaba points in the same direction.
On Tuesday, September 22, three days before the SemiAnalysis report, Alibaba launched the Zhenwu V900 accelerator. The company calls it China's most powerful AI chip and claims it triples performance compared with its predecessor, and that it can be linked into clusters of up to 500,000 units for training frontier models. These are Alibaba's own claims, made by CEO Eddie Wu and reported by Bloomberg – not independently verified measurements.
On Wednesday, September 23, the company additionally announced, through executive Li Feifei at the Apsara conference, an accelerated expansion in Europe and the Middle East as part of a 20-gigawatt global infrastructure plan. The first new cloud regions in Turkey, Finland and the Netherlands are to open within 12 months, according to Bloomberg.
Together with SemiAnalysis' leasing forecast, this sketches a picture of a Chinese cloud and AI ecosystem that is beginning to buy capacity outside its own borders – not only build within them.
What can be verified, and what cannot
There is a clear dividing line in this story. Alibaba's announcements about European regions and the chip launch are documented through Bloomberg's coverage, even though the performance claims are the company's own. Brookings Papers on Economic Activity published research on September 23 estimating US AI infrastructure investment at 10,300 billion dollars from 2025 to 2032 – an average of 3.63 percent of US GDP per year – and warns that risk is being shifted into opaque financing off the balance sheet. As context: Investor's Business Daily wrote on September 24 that the US AI infrastructure buildout depends on Chinese components and materials, including rare-earth coatings for chip equipment and gas turbines, and optical transceivers.
But every single quantitative claim about the Chinese market – the 24 GW, ByteDance's fifth, BAT's 20 billion, the 746 billion dollars in grid capex – rests exclusively on SemiAnalysis' proprietary model behind a paywall. There are also entirely different ways of measuring the market – facility counts rather than capacity, for example – which cannot be compared directly with SemiAnalysis' gigawatt figures, but which show how different the existing baselines for the Chinese market are.
What remains to be answered
Three questions will determine whether the 24 GW figure holds: What are the actual vacancy rates and utilization levels in the Chinese facilities – SemiAnalysis challenges the high vacancy figures, but its own measurements are not in the published material. Whether any of the established datasets – at analysis firms or in Chinese public figures – can confirm the order of magnitude. And how robust ByteDance's dominant role as a customer is if the capex wave flattens out. Until more sources can corroborate SemiAnalysis' model, all the figures should be read as one company's well-reasoned estimates – not as a new, established truth about the Chinese datacenter market.