Chinese AI companies have pledged 12.5 gigawatts of data center capacity in Ulanqab, an Inner Mongolian city of 1.5 million people, according to a Goldman Sachs research note published last week. More than 70% of those commitments were announced in the past year, and nearly 100 data centers have opened or begun construction there since 2016. For scale, OpenAI's $500 billion Stargate Project targets 10 gigawatts of total capacity when complete.
The buildout marks a shift in how Chinese AI firms fund their infrastructure. DeepSeek is reportedly constructing a large facility in Ulanqab, alongside projects from ByteDance, Alibaba and Xiaohongshu. Until recently, Chinese labs leaned on rented cloud capacity while spending far less on physical infrastructure than their US counterparts, despite shipping models with competitive capabilities.
Ulanqab's appeal comes down to geography and cost. The city sits at high elevation on the Inner Mongolian Plateau with long, cold winters that cut cooling loads. It's two hours west of Beijing by train, close enough to serve populous eastern regions. And electricity is cheaper there than almost anywhere else in China, supported by wind, solar and an abundant supply of coal.
Key facts
- 01Chinese firms have pledged 12.5 gigawatts of AI data center capacity in Ulanqab, with 70% of those commitments made in the last year.
- 02Nearly 100 data centers have opened or broken ground in the city of 1.5 million since 2016.
- 03OpenAI's $500 billion Stargate Project targets only 10 gigawatts of total capacity when complete.
- 04Envision announced a 2GW AI data center in Ulanqab this month, powered directly by its own wind supply.
- 05Ulanqab still draws roughly 37% of its electricity from coal, and gets just 14 inches of rain per year.
Inner Mongolia was already a data center corridor before generative AI arrived. Huawei built its first facility in Ulanqab in 2016, Apple followed in 2019, and in 2021 the region was designated a hub in Beijing's "Eastern Data, Western Compute" program. The early sites mostly handled backup storage because latency to eastern users was too high for real-time workloads.
“With the rise of AI in 2022, there was the realization that actually, those remote data centers could be well-utilized for model training.”— Andrew Stokols, Professor at Singapore Management University
Two dedicated fiber optic cables built in 2017 and 2019 cut average latency to less than five milliseconds, fast enough to support inference at scale. Model training also plays to the region's strengths: runs stretch over months and tolerate distance from users. Startups including DeepSeek, Moonshot AI and Zhiput AI now have paying users to serve, giving the remote data centers a commercial reason to exist.
The growth is running ahead of local resources, particularly water. Ulanqab is roughly as dry as Denver, receiving only about 14 inches of rain a year. Last month, the local water company shut off several waterworks for seven hours each night to manage peak demand — and that was before most of the newly announced projects come online.
Operators note that Ulanqab data centers require additional water for cooling in only two months of the year, per weather data from the local government. Even so, the combined draw from dozens of new facilities in an already water-stressed region is a real constraint. It's the same tension playing out at US hubs, only at a colder latitude with a thinner rainfall baseline.
“Inner Mongolia has long been the West Virginia of China. It's a coal country.”— Damien Ma, Director of Carnegie China
The energy story is more favorable, at least on paper. Damien Ma, director of Carnegie China, notes a positive correlation between Chinese regions with the most unused renewable capacity and those attracting the most data center investment. Envision, one of China's largest wind turbine makers, said this month it will build a 2-gigawatt AI data center in Ulanqab tied directly to its own clean power supply.
The reality is messier. Research from Andrew Stokols at Singapore Management University found that roughly 37% of Ulanqab's electricity still comes from coal. Because data centers run around the clock, operators have historically favored fossil generation for its reliability. Ma expects that mix to shift — "in three years, maybe it will be completely powered by renewables" — but the timeline is unclear.
The other open question is demand. The pledged 12.5 gigawatts assumes Chinese labs can absorb capacity on a scale comparable to Stargate, and much of that hinges on inference growth from DeepSeek, ByteDance and Alibaba retaining paying customers at current trajectories. Announced capacity is not the same as built capacity, and China's data center corridors have seen overbuilds before.
Ulanqab is the clearest signal yet that Chinese AI firms are done treating compute as something to rent. Owning the stack — from turbines to fiber to inference servers — is the same vertical integration play Microsoft, Google and Amazon executed years ago, and it changes the competitive picture for anyone selling model access into the Chinese market. If even half the pledged capacity comes online with renewables as advertised, the cost floor for Chinese inference drops in a way US labs will feel on price.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
