12.5 Gigawatts on Sheep Pasture and Coal Country — China's Real AI Hub Isn't Beijing

One more number puts the scale in perspective. According to a Goldman Sachs research note published last week, Chinese companies have pledged a combined 12.5 gigawatts of data center capacity in Ulanqab, more than 70% of it announced within just the past year. That makes it one of the fastest-growing computing clusters in Asia. For comparison, OpenAI's $500 billion Stargate project is targeting 10 gigawatts at completion. If you want to see where the center of gravity for China's AI infrastructure is shifting, you don't look at Shanghai or Shenzhen — you look at this plateau city.

Here's the plain version: the warehouses of computers that run AI models consume enormous amounts of electricity and generate enormous amounts of heat. That makes cheap power and cold weather a huge advantage — but places with both are usually far from major cities. Ulanqab is a rare spot that manages to be cheap, cold, and still close to Beijing.

Why here — the three-part calculation

The first factor is cooling. Ulanqab sits at high elevation on the Inner Mongolian plateau, with long, cold winters. A large chunk of a data center's electricity bill goes toward cooling servers, and cold outside air slashes that cost dramatically. According to local government weather data, Ulanqab needs supplemental cooling water for only two months out of the year.

The second factor is distance. Build a data center too far out in the western interior and power gets cheap, but the round-trip signal time to users stretches out. Unlike other western hubs, Ulanqab sits much closer to Beijing and the major metro areas, and two dedicated fiber-optic lines laid in 2017 and 2019 have cut average latency below 5 milliseconds. That's fast enough not just for training models, but for running inference services that answer user queries in real time.

The third factor is the clincher: electricity prices. Inner Mongolia has rapidly expanded wind and solar while still sitting on abundant coal, which keeps power prices among the cheapest anywhere in China. A data center's operating cost ultimately comes down to its electricity bill, so where you put the same servers can decide whether the whole operation turns a profit. Cold plateau, 5-millisecond latency, rock-bottom rates — spots where all three overlap are rare on China's map.

From backup storage to training ground — rediscovering "East Data, West Computing"

Ulanqab's rise as a data center city predates the current AI boom by years. Huawei built its first data center there in 2016, with Apple following three years later. In 2021, Beijing designated the region as a key hub in a nationwide initiative called "East Data, West Computing" (东数西算) — the idea being that data generated in the east gets processed in the west.

The problem, at first, was that these centers weren't especially useful. Being far from the population-dense eastern coast meant high latency, so they mostly served as backup storage. AI changed that picture. Andrew Stokols, who studies China's computing infrastructure at Singapore Management University, explains that as AI took off in 2022, it became clear that "those remote data centers could be plenty useful for training models." A single training run takes months and rarely needs real-time tweaking, so latency barely matters. Buildings designed as warehouses suddenly became factories.

Stokols also points to another distinctive feature of Ulanqab's growth: unlike other computing hubs, it's been driven more by commercial demand than by government investment. As startups like DeepSeek, Moonshot AI, and Zhipu AI grew their domestic paying user bases, it became commercially attractive to build inference-focused data centers close enough to serve those users without lag. Policy laid the groundwork, and the market started running on top of it.

Building with their own money, for the first time

The most notable thing about this boom is who's behind it. China's AI companies have shipped plenty of strong models while spending far less than their American rivals on physical infrastructure, typically renting compute from cloud providers instead. Now, for the first time, they're pouring real money into infrastructure of their own. According to a Bloomberg report, DeepSeek is building a massive AI data center in Ulanqab, and ByteDance, Alibaba, and Xiaohongshu are all building there too.

Wind turbine manufacturers have joined in as well. Envision, one of China's largest wind turbine makers, announced this month that it will build a 2-gigawatt AI data center in Ulanqab connected directly to its own clean-power generation. In other words, the company making the electricity is also building the facility that consumes it. For the government, that's a two-for-one: it helps China keep pace with the speed of U.S. data center construction while creating demand for surplus renewable energy. Damien Ma, who heads the Singapore-based research center MacroPolo, formerly Carnegie China, has found a positive correlation in China between regions with the most unused renewable capacity and regions with the most data center construction.

Water and coal — the two line items missing from the math

When a picture looks this clean, it's worth being suspicious. Ulanqab's first problem is water. The city is about as dry as Denver, Colorado, getting just 14 inches — roughly 360 millimeters — of rain a year. Even before most of the planned data centers are up and running, local government is already struggling to supply enough water for residents. Last month, Ulanqab's water utility had to shut down several water facilities for seven hours a night to reduce peak demand. It's a bit of luck that data centers only need supplemental cooling water for two months a year — but those two months happen to fall in summer, exactly when water is scarcest.

The second problem is the color of the power. "AI data centers running on renewables" is a story Beijing likes to tell, but Stokols's research shows that about 37% of Ulanqab's electricity still comes from coal. Because data centers need to run around the clock without interruption, operators have traditionally favored fossil fuels over wind and solar for their reliability. Ma's phrase for Inner Mongolia is "China's West Virginia" — coal country, for a long time. The region is now moving fast to shift from coal to wind and solar, but nobody can say for certain how fast or how far that shift will go. Ma thinks it's plausible the grid could run entirely on renewables within three years, but until then, the data centers on this plateau will keep burning at least some coal to stay online.

What this means at home

Ulanqab's math applies just as directly elsewhere. Where a data center ends up is ultimately determined by the overlap of cheapest power, tolerable distance, and climate you can cool with — and as domestic debate shifts from data centers clustered around the capital region toward regional dispersal, this city offers a concrete case study. In particular, the way dedicated fiber lines manufactured a "far but not too far" location by shaving latency down to 5 milliseconds shows that solving the distance problem with cable investment can actually work.

At the same time, Ulanqab is a warning. In the competition to attract data centers, water and power sources are usually the fine print at the bottom of the ledger — but in practice, they show up first, in the form of resident water shutoffs and coal dependence. Read the signal that Chinese AI companies are finally building their own infrastructure, but read the conditions of the ground it's standing on just as carefully.

If there's one sentence that matters most in this story, it's this: Chinese AI companies have started spending real money on their own infrastructure for the first time. For years, the going explanation was that "China builds great models but invests a fraction of what the U.S. does in infrastructure," and that gap was seen as the ceiling on Chinese AI. Ulanqab's 12.5 gigawatts is the first tangible evidence that ceiling is lifting. And the fact that evidence is standing not on a campus in Beijing but on grassland in a coal-mining town tells you something essential about the Chinese calculus: electricity prices matter as much as model performance.

But two line items are still missing from that calculation — water and coal. Just as America's data center debate spilled over into grid capacity and local politics, China's data center boom will eventually run into the same reckoning over water rationing and carbon accounting. Ulanqab's water shutting off for seven hours every night is where that reckoning begins. If, three years from now, this city really is running entirely on renewables, China will have solved a problem the U.S. hasn't managed to solve. If not, those 12.5 gigawatts become one of the largest coal-power demands anywhere in the world. Either way, the next round of competition in Chinese AI won't be decided by models — it'll be decided on this plateau.