China's AI Tech Giants Fuel Massive 12.5 GW Data Center

What happened

Ulanqab, an arid city in Inner Mongolia located roughly two hours west of Beijing, has rapidly transformed into China's primary hub for artificial intelligence computing. 5 gigawatts of data center capacity across nearly 100 facilities, with over 70 percent of those commitments occurring in the past year alone. This collective infrastructure footprint outpaces major Western initiatives, including OpenAI's planned 10-gigawatt Stargate supercluster.

Unlike previous infrastructure waves where remote facilities served primarily as passive cold storage, this expansion is tailored directly for active AI workloads. High-speed dedicated fiber-optic connections have cut latency between Ulanqab and Beijing to under five milliseconds, enabling these centers to run both months-long model training runs and responsive real-time inference.

Notably, leading domestic AI developers such as DeepSeek, ByteDance, Alibaba, and Xiaohongshu are financing and constructing proprietary facilities rather than depending exclusively on third-party cloud compute providers.

Why it matters

This capital-intensive building spree marks a pivotal transition in China's artificial intelligence strategy. Historically, Chinese AI enterprises have prioritized efficient algorithm design and cost-effective research over massive physical cluster buildouts. The rapid escalation of self-owned compute assets indicates that domestic tech leaders are aggressively scaling their sovereign infrastructure to support growing user adoption and retain control over specialized hardware pipelines.

Geographically, Ulanqab provides an optimal mix of operational advantages and national policy alignment. The region's cold climate on the Inner Mongolian Plateau minimizes cooling overhead, while an abundance of local solar, wind, and coal resources yields some of the lowest industrial power tariffs in China.

This structure directly serves Beijing's "Eastern Data, Western Compute" initiative, allowing the state to channel surplus green energy into heavy compute loads while helping domestic AI developers lower long-term training and inference expenditures.

What to watch

Resource bottlenecks—especially severe regional water constraints—will determine the longevity of this compute buildout. Ulanqab averages just 14 inches of rainfall annually, and municipal authorities have already implemented temporary evening water shutoffs to manage baseline residential demand before most newly pledged data centers go fully operational.

While frigid winter temperatures reduce cooling requirements for ten months of the year, peak summer thermal loads could severely strain municipal water supplies. Observers should track how data center operators implement closed-loop liquid cooling architectures and navigate local resource tensions as more gigawatt-scale facilities come online.