Located approximately 350 kilometers northwest of Beijing, the grassland municipality of Ulanqab in Inner Mongolia has transformed into China's primary hub for artificial intelligence compute infrastructure. Historically recognized for agriculture and mineral extraction, the city now hosts nearly 100 enterprise data centers operating or under active construction, with technology firms pledging an aggregate capacity of 12.5 gigawatts (GW).
According to a research note published by Goldman Sachs, more than 70 percent of Ulanqab's total power and capacity commitments have been announced over the past twelve months. The rapid concentration of facilities reflects an industry-wide pivot: leading Chinese AI laboratories and internet giants are shifting away from third-party cloud rental models toward owned, vertically integrated physical infrastructure to support frontier model pre-training and high-throughput inference workloads.
The Shift Toward Dedicated AI Infrastructure
For several years, Chinese artificial intelligence developers maintained significantly lower capital expenditure on physical facilities compared to American hyperscalers, relying primarily on shared cloud compute allocations. That operational model is rapidly shifting as parameter counts and distributed training requirements expand.
Frontier AI startup DeepSeek is currently constructing a dedicated 1-gigawatt AI data center campus in Ulanqab. DeepSeek's direct infrastructure push is matched by major technology enterprises, including ByteDance, Alibaba, Kuaishou, Xiaohongshu, and Huawei, all of which operate or are expanding massive server facilities across the municipality's Chahar High-Tech Zone.
Geographic and Structural Advantages
The concentration of artificial intelligence infrastructure in Ulanqab is underpinned by four core structural factors:
- Abundant Low-Cost Power: Inner Mongolia is one of China's most productive renewable energy corridors, featuring vast wind turbine installations and utility-scale photovoltaic arrays supplemented by local thermal baseloads. In the first seven months of 2026, data center electricity consumption in Ulanqab reached 4.0 billion kilowatt-hours (kWh)—an 86 percent increase year-over-year—representing 7.3 percent of the entire municipality's power usage.
- Natural Thermal Dissipation: With an average annual temperature near freezing and low relative humidity, the regional climate enables year-round free-air economization, significantly reducing the cooling energy overhead and power usage effectiveness (PUE) ratios of dense accelerator clusters.
- Sub-5ms Network Latency to Beijing: Unlike remote western computing zones that suffer from high round-trip transit delays, Ulanqab is linked to Beijing via two dedicated direct optical fiber trunks constructed in 2017 and 2019. These lines maintain average latency below 5 milliseconds (<5ms), making the facilities viable not only for asynchronous offline batch training but also for latency-sensitive real-time inference routing.
- Policy Alignment with National Compute Corridors: The rapid buildout integrates directly into China's broader "Eastern Data, Western Computing" initiative, designed to channel resource-intensive compute loads from coastal economic centers to energy-rich northern and western regions.
As hardware constraints and compute demands continue to shape foundation model development, Ulanqab's emergence as a 12.5GW compute corridor underscores how geographic proximity, specialized power generation, and dedicated fiber backbones are dictating the physical geography of global AI infrastructure.
