Ulanqab's 12.5 GW Bet Puts China's AI Compute Ahead of Stargate

Ulanqab, a city of roughly 1.5 million people on the Inner Mongolian Plateau, has accumulated commitments for approximately 12.5 gigawatts of combined data center capacity, with over 70 percent of those pledges announced within the past year alone, according to a Goldman Sachs research note. That figure puts Ulanqab ahead of OpenAI's Stargate Project, which targets 10 gigawatts of total capacity at completion. For engineers thinking about where China's AI training and inference workloads will physically execute, this geography now demands attention — it shapes latency budgets, energy mix assumptions, and the competitive economics of Chinese AI deployment.

Why Ulanqab

Three structural factors converge here that are difficult to replicate elsewhere in China. The city's elevation and long, cold winters reduce mechanical cooling loads. Two dedicated fiber optic cables, installed in 2017 and 2019, brought average latency to under five milliseconds, crossing the threshold for real-time AI inference workloads. And electricity prices in Inner Mongolia rank among the lowest in China, driven by wind and solar penetration alongside substantial coal reserves.

The latency story matters historically. Before these fiber links existed, Ulanqab-area data centers were used primarily for backup storage because round-trip times to China's eastern population centers were prohibitive for interactive workloads. Andrew Stokols, a professor at Singapore Management University who studies China's compute infrastructure, notes that the arrival of large-scale model training around 2022 changed the calculus: training runs lasting months are largely insensitive to latency, so remote, cheap-power locations became viable. Sub-five-millisecond connectivity now makes inference viable there too — meaning Ulanqab can serve the full AI compute stack rather than just offline workloads.

Who Is Building

The composition of builders marks a structural shift in how Chinese AI companies treat infrastructure. DeepSeek, ByteDance, Alibaba, and Xiaohongshu are all reportedly building or planning data centers in Ulanqab — companies investing in their own compute rather than renting capacity from cloud providers. This mirrors the broader compute sovereignty logic playing out across the AI frontier: controlling physical infrastructure confers training flexibility, cost predictability, and the ability to run experiments that shared-cloud economics make prohibitively expensive.

Stokols characterises Ulanqab's growth as more commercially driven than state-directed, distinguishing it from earlier projects under China's "Eastern Data, Western Compute" initiative, in which Ulanqab was designated as a hub in 2021. The current wave appears to be responding to genuine paying-user demand from Chinese AI startups rather than policy incentives alone.

Energy Mix and Water

The Chinese government frames Ulanqab as a solution to the country's surplus renewable energy — regions with the most stranded renewable capacity correlate strongly with data center construction activity, according to Damien Ma of Carnegie China. Envision, one of China's largest wind turbine manufacturers, announced this month it will build a 2-gigawatt AI data center in Ulanqab connected directly to its own clean power supply, making the renewable pitch concrete. But Stokols's research places coal at roughly 37 percent of Ulanqab's current electricity mix, and because data centers require 24/7 uptime, operators have historically preferred the dispatch reliability of fossil fuels over intermittent renewables. Ma describes Inner Mongolia as historically analogous to West Virginia — a coal-producing region now attempting an energy transition whose pace and completeness remain uncertain.

The operational risk that receives least attention in the capacity announcements is water. Ulanqab receives approximately 14 inches of precipitation annually, and the local water utility was already forced to shut down several waterworks for seven hours each night last month to manage peak residential demand — before the majority of planned data center projects are operational. Local government weather data indicates that additional water for cooling is only required during two months of the year given the cold climate, which partially mitigates the risk, but the cumulative draw from nearly 100 facilities at gigawatt-scale still represents a significant environmental exposure.

As pipeline architecture increasingly determines where AI performance gains come from, the physical location of training and inference clusters — their energy mix, water access, and network topology — becomes a first-order engineering variable. Ulanqab's rise demonstrates that compute geography can shift rapidly when cost and connectivity thresholds align, and the more than 70 percent of capacity pledged within a single year suggests the window for that alignment to attract capital is compressing fast.