China’s New AI Hub Has 12.5 Gigawatts Planned—and a Water Problem

Ulanqab, an arid city in Inner Mongolia, is becoming one of Asia’s fastest-growing AI compute clusters. Chinese companies have announced an estimated 12.5 gigawatts of data-center capacity there, but the expansion is landing in a place where water demand has already forced nightly shutdowns at several waterworks and coal still supplies a significant share of electricity.

Distance stopped being a disqualifier

The city’s data-center buildout predates the current AI surge. Huawei opened its first Ulanqab facility in 2016, Apple followed in 2019, and the area became a main hub in China’s Eastern Data, Western Compute program in 2021. Those western facilities initially fit backup storage better than interactive services because they were far from China’s populous eastern coast and faced higher latency.

AI created two ways to use the location

• Model training can run for months and requires less real-time interaction, making latency less consequential than it is for live applications.
• Two dedicated fiber-optic cables built in 2017 and 2019 brought average latency below five milliseconds, a level the article says can support real-time exchanges such as AI inference.

Model builders are moving closer to the metal

The new wave is notable not only for its size, but for its owners. DeepSeek, ByteDance, Alibaba and Xiaohongshu are reportedly building or planning major AI data centers in Ulanqab. Chinese AI companies have historically spent less on physical infrastructure than their American peers; the projects point to a move toward dedicated capacity rather than renting compute solely from cloud providers.

That shift follows a more commercial logic than the region’s earlier government-led buildout, according to Singapore Management University professor Andrew Stokols. As Chinese AI startups draw more paying users, facilities close enough to Beijing and other major cities can serve inference requests without excessive delay, while retaining the cost advantages of Inner Mongolia.

Low-cost power is not automatically clean power

Ulanqab offers long, cold winters at high elevation, reducing the energy required to cool servers. Electricity is also relatively cheap, supported by wind and solar growth as well as abundant coal. That combination explains the site’s appeal, but it complicates a simple renewable-energy narrative: research cited in the article finds that about 37 percent of Ulanqab’s electricity still comes from coal.

The region is also being positioned as a way to absorb unused renewable generation. Envision, a wind-turbine manufacturer, announced this month that it will build a 2-gigawatt AI data center in Ulanqab connected directly to its own clean-power supply. But data centers operate around the clock, and the pace and extent of Inner Mongolia’s transition away from coal remain uncertain.

The harder limit may be water

Ulanqab receives roughly 14 inches of rain a year. Last month, its local water company shut off several waterworks for seven hours each night to curb peak demand, before many planned data-center projects have begun operating. Local weather data indicates the facilities need additional cooling water during only two months of the year, yet the planned scale still poses an environmental challenge in a city already struggling to meet residents’ demand.

Ulanqab has solved the network problem that once constrained western Chinese data centers, and AI has given those facilities a higher-value use. The unresolved test is physical: whether a cluster built on low costs can add the promised capacity without deepening pressure on a dry local water system or locking more AI demand into coal-backed power.