Ulanqab: The Inner Mongolia City Powering China’s AI Data Center Surge

A mid-sized city in Inner Mongolia, two hours west of Beijing by train, has quietly become the most important piece of AI infrastructure in China. Ulanqab, population 1.5 million, now has nearly 100 data centers open or under construction since 2016. Chinese companies have pledged a combined 12.5 gigawatts of capacity there, a figure that exceeds the planned ceiling of OpenAI's $500 billion Stargate Project, and more than 70 percent of those pledges came in just the last year, according to a Goldman Sachs research note.

What happened

Data point	Detail
Data centers open or under construction since 2016	Nearly 100
Total capacity pledged by Chinese companies	12.5 gigawatts
Share of pledges made in the last year	Over 70%
OpenAI Stargate Project planned total capacity	10 gigawatts
Average annual rainfall in Ulanqab	~14 inches (similar to Denver)
Share of local electricity from coal	~37%
Average latency to eastern China via fiber	Under 5 milliseconds
Envision’s planned clean-power data center	2 gigawatts, announced this month

Ulanqab sits on the Inner Mongolian Plateau at high elevation. Its long, cold winters cut the energy needed to cool servers. Electricity prices are among the lowest in China, pushed down by surplus wind and solar power layered on top of abundant coal. And two dedicated fiber optic cables, built in 2017 and 2019, now deliver latency below five milliseconds to Beijing, fast enough for real-time AI inference tasks.

Huawei built its first data center there in 2016. Apple followed in 2019. By 2021 the area was formally designated a hub under a national program called “Eastern Data, Western Compute,” which encourages building compute infrastructure in China’s western interior. AI gave that program a second wind. Because model training runs can last months and require very little real-time interaction, the region’s historical latency disadvantage stopped mattering.

Why Chinese AI firms are building, not renting

The more consequential shift is who is now building in Ulanqab. According to reporting by Wired, DeepSeek, ByteDance, Alibaba, and Xiaohongshu are all developing data centers there. That is a notable change: Chinese AI companies have historically rented compute from cloud providers rather than owning their own hardware. The Ulanqab buildout signals that firms now see enough sustained demand, particularly from paying inference customers, to justify the capital expense.

Andrew Stokols, a professor at Singapore Management University who studies China’s compute infrastructure, told Wired that growth in Ulanqab appears driven more by commercial demand than government-led investment. Startups including DeepSeek, Moonshot AI, and Zhipu AI have been attracting paying users inside China and need inference capacity close enough to serve them without excessive lag.

This mirrors a pattern playing out in the US, where massive capital commitments for AI infrastructure are coming from both tech giants and specialized AI labs simultaneously. The scale in Ulanqab is striking: at 12.5 GW pledged, it already exceeds what OpenAI’s $500 billion Stargate Project is expected to reach when fully built.

Is the renewable energy story real?

The Chinese government frames data center growth in Inner Mongolia as a clean energy win. The region has surplus wind and solar capacity, and data centers provide a large, steady buyer for that power. Envision, one of China’s biggest wind turbine manufacturers, announced this month it will build a 2 GW data center in Ulanqab tied directly to its own clean power supply.

Damien Ma, director of Carnegie China, says there is a genuine positive correlation between regions with excess renewable capacity and those seeing the most data center construction. But experts are quick to add nuance. Stokols found in his research that roughly 37 percent of electricity in Ulanqab still comes from coal. Data centers need to operate around the clock, and operators have historically preferred fossil fuels for their reliability. Calling AI and renewables a perfect match in this context is, as Stokols puts it, too simple a story.

The water problem nobody is solving yet

The most concrete near-term constraint is water. Ulanqab receives only about 14 inches of rain per year, on par with Denver. The local government is already struggling to meet residential water demand before most planned data centers are even operational. Last month, the city’s water company had to shut down several waterworks for seven hours each night to manage peak demand.

Local government weather data shows the data centers will need extra water for cooling during only two months of the year, but the cumulative impact of dozens of facilities running simultaneously could still create a significant environmental problem for a region that was already water-stressed before this buildout began.

Our take

The Ulanqab story is worth watching closely, even if your business has nothing to do with China. The numbers tell you something real about where AI compute is heading globally: firms that were happy to rent GPU capacity are now writing checks for physical infrastructure, because the inference demand is consistent enough to justify it. That is a maturity signal for the AI market, not just a China story.

For businesses planning AI integration projects, the subtext here is that inference costs are not a fixed variable. As more self-owned capacity comes online, pricing dynamics for AI APIs can shift. The firms building now are betting that owning the hardware beats paying cloud markups at scale.

The water constraint is the kind of physical bottleneck that rarely shows up in AI investment narratives but tends to become the binding constraint eventually. Denver-levels of rainfall supporting 12.5 GW of compute is a tension that cannot be resolved by software.

If you want to track how this infrastructure buildout affects AI model availability and pricing, our AI news coverage keeps tabs on the developments that actually move costs for operators.

What to watch next

1. Monitor whether DeepSeek and ByteDance confirm or publish specs for their Ulanqab facilities, which would give clearer signals on inference capacity timelines.
2. Track Envision’s 2 GW clean-power project as a test case for whether AI data centers in China can genuinely run on renewables around the clock.
3. Watch Ulanqab’s water policy responses. Any regulatory cap on water use could slow construction faster than grid constraints would.
4. Compare API pricing from Chinese AI providers over the next 12 months as self-owned infrastructure scales up.

The practical takeaway: 12.5 GW of pledged AI compute in one inland Chinese city is a real market signal, not a vanity metric, and the water risk is the most likely thing to slow it down.

Source: WIRED · AI