A grasslands city suddenly at the center of AI
Ulanqab does not sound like the place where the future gets wired together. It sits in Inner Mongolia, about a couple of hours west of Beijing by train, in a stretch of grasslands dotted with cinder cones, sheep farms, and the leftovers of old coal country. It’s the kind of place where the horizon feels wider than the road network. The local population is around 1.5 million, which is a decent-sized Chinese city, but not the sort of number that usually comes with a giant cluster of server halls humming away in the background.
Yet that is exactly what has happened. Since the mid-2010s, close to a hundred data centers have been finished or started in Ulanqab, turning the city into one of the country’s busiest places for compute. What began as an odd-looking bet on an inland city has grown into a buildout large enough to matter far beyond Inner Mongolia. The city is no longer just a dot on the map west of Beijing. It has become a place where Chinese AI companies want their machines to live.
Ulanqab’s real surprise is not that it has data centers. It’s how fast the count stopped sounding like a local quirk and started reading like industrial policy with a power bill.
The numbers get even stranger. Across the city, announced projects add up to roughly 12 gigawatts of planned capacity. That is a huge amount of electricity to pin to one inland city, and most of those commitments arrived in the past year rather than over some slow, sleepy decade. In other words, this wasn’t a gradual drift of servers into a cheap district. It was a rush. Companies, builders, and local officials all seemed to decide at once that Ulanqab should absorb a serious share of China’s AI infrastructure.
For a bit of comparison, the much-discussed U.S. buildout tied to OpenAI is expected to reach about 10 gigawatts when it is finished, after spending that has been described in the hundreds of billions of dollars. Ulanqab is not doing that on the same balance sheet or with the same global fanfare, but the scale is uncomfortably close for a city most people outside China could not have picked out of a map six months ago. That’s the funny part. A place known for grassland winds and old mining country is now being measured against one of the biggest AI infrastructure bets on the planet.
The result is a story that cuts across tech news, ai policy, digital culture, and plain old power math. Ulanqab has become a national compute hub in the making, and the weirdness of that fact is part of the point. If you’re looking for the center of gravity in China’s AI race, you might not expect to find it here, amid sheep pastures and cooling towers. But that’s where the racks are going up.
Why Ulanqab beats the usual suspects
If you were drawing up a shortlist of places to park a giant AI server farm, a grasslands city in Inner Mongolia would probably not be the first name scribbled on the napkin. And yet Ulanqab keeps winning the argument for a few very unromantic reasons: cold air, cheap power, and a network that no longer behaves like a punishment for choosing the inland route.
The weather does a lot of the heavy lifting. Ulanqab sits at a fairly high elevation, and its winters run long and cold. That matters because servers dump heat like a badly insulated apartment in July. In a place where the thermometer spends a lot of time below freezing, operators can lean on outside air for cooling instead of paying more to chill rows of racks all year. The result is lower operating cost before anyone has even talked about GPUs, training runs, or inference requests.
In Ulanqab, winter is not just weather. It’s part of the cooling system.
Distance from Beijing is the next piece of the puzzle, and here Ulanqab gets to be the odd inland case that still feels close enough to matter. The city is far from the coastal megacities, sure, but it is not so far from the capital that data has to crawl across the country and lose its nerve on the way. For China’s biggest user markets, that proximity helps keep delays down to a level where real-time services still feel usable. A recommendation feed can load without dragging its feet. A chatbot can answer without sounding like it fell asleep.
Power is cheaper there too, which is never a bad thing when you’re feeding a room full of servers around the clock. Inner Mongolia has a mix that most regions would envy: strong wind, plenty of solar development, and a deep coal base that has long kept the grid supplied. That combination pushes electricity prices down. The power and politics of it are plain enough. Beijing wants more compute built where energy is abundant and cheaper; Inner Mongolia happens to have the resource mix to make that possible. A recent central government policy release on data infrastructure fits that logic neatly, even if the local execution is what really decides whether a project lives or dies.
The other old objection to western China data centers was network lag. That used to be enough to kill the idea for anything that needed quick responses. Storage is one thing. Fast AI services are another. Two dedicated fiber routes, laid in the late 2010s, changed the math. They pulled average latency down to a few milliseconds, which is a small number with a very large personality when you’re trying to serve prompts, ranking systems, or lifestyle tech apps that users expect to feel instant. The difference between “close enough for storage” and “good enough for inference” is not academic. It’s the difference between a back room and a product.
That is why Ulanqab AI data centers stopped sounding like a curiosity and started sounding like a plan. The old assumption was simple: western China could store bits, but it couldn’t move them fast enough for serious interactive work. The fiber buildout made that objection much less convincing. Local filings and project paperwork, including a 2024 provincial service notice, show how this kind of buildout gets turned from theory into concrete, cable, and server halls.
So the appeal is not some grand mystery. Ulanqab offers cold air, cheaper electricity, and a network that can keep up with nearby users in Beijing. That combination turns an apparently awkward inland city into a place where fast inference makes sense, not just warehousing. In the current round of tech news, that is enough to send a lot of capital in one direction.
The companies building their own compute
Once the conversation moves past Ulanqab’s weather and geography, the real story turns into a corporate spending decision. Chinese AI firms are no longer happy to rent whatever cloud capacity happens to be available and call it a day. They want racks they can control, chips they can schedule, and power bills they can actually predict. In other words, they want their own compute.
The roster tied to new projects in Ulanqab reads like a compact tour of China’s internet and AI sector: DeepSeek, ByteDance, Alibaba, and Xiaohongshu. That’s a very different posture from the one Chinese AI companies took for much of the last decade, when physical infrastructure spending lagged well behind what U.S. peers were pouring into land, buildings, and server halls. American giants were buying up acreage and ordering GPUs by the truckload. Many Chinese firms, by contrast, stayed lighter on capex and leaned on rented cloud capacity, which kept them agile but also left them at the mercy of someone else’s timetable.
That habit is getting harder to defend. Training bigger models takes serious muscle, and so does serving a flood of paying users who expect answers fast enough that the chatbot doesn’t feel like it’s taking a tea break. The economics of inference are especially unforgiving. If thousands or millions of users are sending prompts all day, every day, paying for distant compute can get old in a hurry. Nearby capacity cuts latency, trims some networking pain, and makes a better product for the person waiting on the other end of the screen. Nobody likes staring at a spinning cursor while a model thinks about life.
If your model makes money all day, waiting in somebody else’s server line starts to feel expensive very quickly.
That is why the Ulanqab pipeline matters for startups as much as for the big platforms. DeepSeek has become the poster child for lean but ambitious Chinese model building, yet even a frugal outfit needs serious hardware once usage moves from curious dabblers to real customers. Moonshot AI and Zhipu AI are finding the same thing. Domestic demand is no longer hypothetical. It has users, bills, and very short patience. Compute close to those users is easier to justify when the product is live, the prompts are rolling in, and every extra millisecond is one more chance for a complaint.
The state’s Eastern Data Western Compute program gives that logic a name, but the business case is more basic than the slogan suggests. Put the servers where electricity is cheaper and land is available, keep the data path short enough to make inference practical, and stop pretending that the cloud can absorb every workload forever. Ulanqab fits that calculation unusually well, which is why the city’s data-center pipeline now includes both internet names and firms that are usually better known for sales growth than for concrete pads and backup generators.
One recent National Energy Administration notice points to how closely this buildout is being tracked inside the energy bureaucracy, while a separate provincial project filing shows how quickly these plans are moving into the paperwork stage. That’s where the market stops being a concept and starts becoming a construction schedule.
A major Chinese wind-turbine maker has joined the scramble too, announcing a large AI data center in the city and tying it directly to its own clean-power supply. That kind of vertical setup makes obvious sense on paper. If you already produce the electricity and already know how to sell energy gear, building compute next to your power assets saves a pile of middlemen. It also says something about where the domestic market has landed: AI infrastructure is no longer just a cloud provider’s problem. It’s becoming a line item for everyone from model labs to hardware firms to the companies building the pipes under the whole thing.
A green promise with a coal problem
The sales pitch for Ulanqab’s server boom sounds tidy on paper. Put the chips where power is cheap, put the power where the wind blows hard, and let the computers chew through work far from China’s crowded coastal cities. In Beijing, that logic has real appeal. Policymakers have spent years trying to move more digital infrastructure westward, so inland provinces can absorb electricity that might otherwise be curtailed when grids get overloaded or demand stays weak. Data centers fit neatly into that plan because they can be placed near generation instead of forcing every watt to travel east first.
That idea sits inside a broader state project to pair western compute capacity with eastern demand. The rough shape is simple enough: build the servers inland, keep the users and app traffic in the east, and use fast fiber to stitch the two together. It’s the same reason Inner Mongolia keeps showing up in China’s AI maps. The region has room, land is cheaper than in the big coastal metros, and the power mix looks cleaner than a pure coal province at first glance. The problem is that the picture gets murkier fast once you look past the brochure version.
Inner Mongolia still gets a hefty slice of its electricity from coal. Wind farms and solar fields have grown quickly, and the region’s open land makes that easier than it would be in denser provinces. Yet coal remains part of the everyday energy stack, so every new AI compute cluster in the area is tied, at least partly, to a grid that still leans on fossil fuel. That matters because the green story around Chinese AI is not really about zero-carbon computing. It’s about reducing pressure on the coastal grid and buying time while renewables expand. Those are different things.
A server farm can be placed next to a wind farm and still run on a grid that’s only partly green.
Water makes the whole equation even messier. Ulanqab is dry enough that the comparison to a high-desert city does not feel like hyperbole. Annual rainfall is low, the air is crisp in a way that sounds nice until you remember it also means the region does not have much slack in its water system, and cooling a room full of hot silicon is not a hobby that tolerates shortages. The local utility has already had to ration water supply at night to meet peak demand, which is a fairly blunt sign that infrastructure is chasing the buildout rather than leading it.
That pressure would be awkward in any city. In a place trying to become a national AI hub, it gets louder. New centers can use less cooling water during winter, when the air is cold enough to do some of the work for free. That helps for part of the year, and nobody building in Ulanqab is pretending otherwise. But the seasonal relief does not erase the bigger issue. Once more AI compute clusters come online, the region will need more water, more grid flexibility, and a better answer for how to scale without turning every dry month into a local headache.
The irony is hard to miss. A city being sold as a home for renewable energy China projects is also sitting on a coal-heavy grid and a dry climate that keeps utility planners awake. The promise is real, but so is the constraint. For now, the boom can still be justified as a way to make use of power that would otherwise sit idle, and perhaps to keep some of the country’s most demanding digital work away from expensive coastal land. Yet the larger the buildout grows, the less forgiving the system becomes. A DeepSeek data center, or any of the other large facilities moving in, may be able to sip a little less water in January. Come summer, the arithmetic gets less charming.
What Ulanqab says about China’s AI race
Ulanqab has turned into a useful stress test for China’s AI buildout. The city is not famous for glossy product launches or chatbot demos. It is famous now because companies want somewhere cold, close enough to Beijing, and cheap enough to run very hungry servers all day and all night. That tells you a lot about where the real race is headed. Models get the headlines. Compute gets the contracts.
The bigger shift is that this isn’t being driven by planners alone. State policy helped open the door, but commercial demand pushed it wider. DeepSeek, ByteDance, Alibaba, Xiaohongshu, and a string of others have been tied to projects in and around the city because they need domestic capacity that can handle training, inference, and live user traffic without a long wait. In other words, firms are treating infrastructure as part of the product, not as a background utility they can ignore until the bills arrive. If a model answers slowly, users notice. If it answers fast, nobody throws a parade, which is usually how infrastructure is supposed to work.
China’s AI bottleneck may end up looking less like a software problem and more like a question of pipes, wires, and cold weather.
That leaves Ulanqab in a strange but useful position. The city sits on top of a cheap-power mix that still leans on coal, yet it also has wind and solar resources that make the cleaner pitch plausible enough to keep expanding. The catch is timing. Servers do not politely wait for the grid to sort itself out. They need steady electricity every hour, not just when the wind is behaving and the sun has a good mood. So the long-run winner will probably be the place that can keep renewables growing faster than coal, while still keeping latency low enough for real-time AI work.
Water is the more immediate headache. A data center water shortage sounds like the sort of thing a planner would put in a footnote and a utility manager would put on a wall. In Ulanqab, it is the sort of constraint that can decide how many racks get switched on next year. Cooling demand rises with scale, and even if winter helps, summer does not send a polite RSVP before showing up. That makes water, not just electricity, the limiter that could slow the boom if the buildout outruns local supply.
So Ulanqab may not be the obvious poster child for China’s AI ambitions. It is dusty, cold, and about as far from a shiny innovation district as you can get without leaving the country. Still, that may be the point. If China is going to keep stacking up AI capacity at home, the answer may come from places that can handle the boring stuff first. Power. Water. Latency. The glamorous part is the model. The messy part is where the future gets built.
