China has 24 GW of AI data centers running, and 50 GW more on the way
A new SemiAnalysis model counts more than 24 gigawatts of live data center capacity in China, more than Europe, the Middle East and Africa combined, against roughly 56 gigawatts in the United States. The bigger story is the pipeline: about 50 gigawatts more is under construction or announced, much of it in Inner Mongolia, and ByteDance alone rents about a fifth of everything that is already running.
When people talk about the AI race between the United States and China, they usually talk about models and chips. Who has the best frontier model this month, who can buy which Nvidia part, who is copying whom. Much less attention goes to the thing that actually turns chips into answers: buildings full of servers, cooling and power lines. Data centers are the factories of the AI era, and their size is measured not in square meters but in gigawatts of electricity.
A new estimate from the research firm SemiAnalysis puts real numbers on that part of the race for the first time at this level of detail. According to its new China Datacenter Model, China has more than 24 gigawatts of data center capacity that is already delivered and running. That is more than the rest of Asia combined, and clearly more than the roughly 14 gigawatts across Europe, the Middle East and Africa. The United States, by the same kind of count, sits at about 56 gigawatts. The Financial Times reported the figures this week, after The Next Web covered the first part of the report at the end of September.
So the US is still more than twice as large. But the headline number hides the part that should make people in Washington and Silicon Valley sit up: China has around 50 gigawatts more under construction or announced. If that pipeline gets built, China would roughly triple its live base.
What a gigawatt actually means
A gigawatt is a unit of power, not of computing. Data centers are compared this way because electricity is the hard limit. Every GPU, every network switch and every cooling fan needs power, and the grid connection is usually the slowest thing to get. One gigawatt is roughly the output of a large nuclear reactor, and it can run on the order of several hundred thousand modern AI accelerators once cooling and overhead are included.
So when SemiAnalysis says China has 24 gigawatts live, it is saying China already runs the equivalent of about two dozen big power plants worth of servers. Not all of that is AI. A lot of the older capacity serves cloud storage, video, payments and the everyday internet. But the new capacity that is being added is overwhelmingly built for AI training and inference.
It is also worth being clear about what kind of number this is. SemiAnalysis is a private analyst firm. Its model tracks more than 1,000 facilities run by over 60 operators, based on satellite images, permits, filings, supply chain checks and its own industry contacts. That is a serious effort, but it is still an estimate. Chinese companies do not publish this data in a standard way, and as the report itself puts it, "the biggest tenant files no 10-K. Several of the largest landlords have never listed." The numbers here should be read as a good outside view, not an official census.
The pipeline is the real story
The most striking figure is not the 24 gigawatts. It is the roughly 50 gigawatts that SemiAnalysis sees under construction or announced in China. Announced projects do not always get finished, and some will be delayed or shrunk. But even a large haircut would leave China adding capacity at a speed that very few countries can match.
Where is all that going? Much of it is moving west and north, away from the expensive coastal megacities. Since 2022 China has followed a national plan called "Eastern Data, Western Compute". The idea is simple: the demand sits in Beijing, Shanghai and Shenzhen, but land and electricity are cheaper in the interior, so the heavy computing should happen there and the results travel over fiber.
The clearest winner is Inner Mongolia. According to the FT summary of the report, the region around Ulanqab is the fastest growing data center hub in Asia, with 89 data centers and roughly 15 gigawatts of capacity planned or running. SemiAnalysis says power there costs about half of what it costs in China's biggest cities. For an industry where electricity can be the largest running cost, that matters a lot. Inner Mongolia has become the default place to put a large new AI campus.
Why China can build so fast
The second big theme of the report is speed. SemiAnalysis says China routinely delivers a 100 megawatt facility in under 12 months, and the FT summary mentions 12 to 18 months for typical new campuses thanks to cheap power and modular construction. Prefabricated power and cooling blocks are produced in factories and then put together on site, a bit like building with very large Lego bricks.
The bigger difference sits before anyone pours concrete. Permits in China take about three to six months, according to the report. In the United States the comparable step takes about 12 to 13 months. That is not a small gap. It means a Chinese project can be finished and running in roughly the time a US project is still waiting for approval.
This is the same argument that American AI leaders have been making for over a year. OpenAI, Microsoft, Google and others keep telling lawmakers that the limiting factor for American AI is no longer chips but power and permits. Grid connections, environmental reviews and local opposition can stretch projects across years. The SemiAnalysis numbers give that argument fresh ammunition: China may be behind in chips, but it is not waiting on paperwork.
ByteDance, the quiet giant
One of the most surprising details concerns a single company. SemiAnalysis estimates that ByteDance, the owner of TikTok and its Chinese twin Douyin, occupies roughly one fifth of China's delivered data center capacity. And it rents almost all of it rather than owning the buildings. That makes ByteDance the key customer for nearly every wholesale data center operator in the country.
ByteDance has good reason to need so much. Its AI assistant Doubao serves about 345 million monthly users, according to the report, which makes it one of the most used chatbots in the world. On top of that come recommendation systems for its video apps, which are among the most compute hungry consumer services anywhere. ByteDance also runs its own campuses in Datong, Wuhu and Inner Mongolia, and it registered two new entities in Ningxia, another western region, in July.
The company is also a big tenant abroad. The FT has reported that ByteDance brought in nearly 75 percent of the 2025 revenue of Nscale, a European AI cloud company. SemiAnalysis expects the overseas capacity leased by Chinese hyperscalers to roughly double between 2026 and 2029, to nearly 4 gigawatts.
The money behind it
All of this costs a lot. SemiAnalysis says Alibaba, Tencent and Baidu together spent about 20 billion dollars on capital expenditure in the second quarter of 2026, more than double the year before. For the first time on record, all three had negative free cash flow in the same quarter. In other words, they are spending more on buildings and hardware than their businesses bring in.
Including ByteDance, which does not publish figures, the four companies spent roughly 35 billion dollars in 2024 and more than 50 billion dollars in 2025. SemiAnalysis says they are on course for about 100 billion dollars this year.
That is still smaller than the American giants. Microsoft, Alphabet, Amazon and Meta are each spending at a similar scale on their own, and together they are far ahead. But the direction is the same on both sides of the Pacific: a doubling of spending in roughly a year, financed more and more by debt and leases.
Not every rack is useful for AI
There is an important caveat in the report. Not all of China's capacity is equal. Older retail data centers near the big cities, where many companies rent a few racks each, run at only 30 to 50 percent use, and many are not suitable for modern AI servers. Those servers need far more power per rack and often liquid cooling, which old buildings were never designed for.
SemiAnalysis sums it up in one line: "The AI era is wholesale. The legacy retail stock is not participating." The real growth is in large wholesale campuses that rent whole halls or buildings to one big tenant at a time. Of the two Chinese landlords listed in the US, GDS and VNET, the report says they signed about 1.3 gigawatts of wholesale orders in the first half of 2026, yet won only around a third of the big ByteDance and Alibaba orders between 2024 and 2026. Much of the rest went to private operators that rarely appear in Western news.
The chip question
Buildings are only half the equation. A data center without top chips is just a very expensive warehouse. US export controls have limited China's access to Nvidia's best accelerators for years, and Chinese firms have filled the gap with domestic chips from Huawei and others, plus older or restricted Nvidia parts. The FT and The Next Web both note reports that Beijing may let ByteDance and Alibaba buy new Nvidia chips, which would change how quickly the new space fills up.
That is why gigawatts alone do not decide the race. A gigawatt filled with the newest American accelerators produces far more AI work than a gigawatt filled with older or less efficient hardware. The US lead in chips and in efficiency per watt is real. But the SemiAnalysis numbers suggest that if the chip supply loosens, China will not be short of places to plug them in.
What this means
For anyone following AI, a few points stand out.
- The US is still ahead, with about 56 gigawatts live versus about 24 in China, by these estimates.
- China's pipeline is enormous, with around 50 gigawatts under construction or announced, much of it in Inner Mongolia.
- Speed is a Chinese advantage. Short permit times and modular builds mean new capacity can arrive in about a year.
- ByteDance matters far more than its Western profile suggests, renting about a fifth of China's live capacity.
- Spending is doubling, toward about 100 billion dollars this year for the four biggest Chinese players, with real pressure on cash flow.
- The numbers are estimates from a private research firm, not official data.
The AI race is often framed as a battle of clever models. This report is a reminder that it is also a very physical contest of concrete, transformers, cooling towers and power contracts. On that front, China is building at a pace that the rest of the world will have trouble matching, even if the United States still holds the bigger lead today.
Sources
Source: ft.com