China's AI Boom Has a Water Problem

Two hours by train from Beijing, on the high cold plateau of Inner Mongolia, the city of Ulanqab is doing something no other city in the world is doing at this scale: hosting the buildout of nearly 100 AI data centers on roughly 14 inches of rain a year. Last month, before most of those projects are even running, the local water utility was forced to shut off service.

The buildout is colliding with a constraint underneath it: water. China's biggest AI labs are not building in the coastal cities where the engineers, the venture capital, and the headline data centers have lived for a decade. They are building in Ulanqab because the plateau is cold, dry, and cheap. The cheapest electricity in China comes through here, drawn from a grid stacked with wind, solar, and the coal seams of Inner Mongolia. That same weather that powers the cluster also starves it of water.

Chinese companies have now pledged 12.5 gigawatts of data center capacity in Ulanqab, according to a Goldman Sachs note cited by Wired. More than 70 percent of those commitments were announced in just the last year. For a sense of scale, OpenAI's $500 billion Stargate project is set to reach only 10 gigawatts of total capacity when it is complete. The U.S. benchmark is useful here. The subject is Ulanqab.

DeepSeek, ByteDance, Alibaba, and Xiaohongshu (RedNote) are each building their own facilities in the city, a break from the older pattern of renting compute from China's hyperscale cloud providers. The Japan Times describes DeepSeek targeting roughly one gigawatt of capacity in Inner Mongolia. DataCenterDynamics reports RedNote is weighing a 600 megawatt campus of its own. These are not cloud-region leases. They are owner-operated builds, the kind that signal a long-term bet on a single geography.

Ulanqab sits at high elevation on the edge of the Gobi, and gets about as much precipitation as Denver. Most data center cooling systems use water in some form, even when the marketing copy says "air-cooled." Hyperscale operators have spent a decade building dry-cooling loops and water-recycling plants to keep new builds off municipal supplies. In Ulanqab, those mitigations are still on the drawing board for most of the 12.5 gigawatts that has been announced. The utility's shutdown last month was for residents, not data centers. The constraint is binding on the population before it binds on the compute.

The economic logic that pulled the cluster to Ulanqab is not going away. Coal-fired power in Inner Mongolia remains among the cheapest in the country. Cold air cuts cooling load in winter. Proximity to Beijing keeps latency low enough for training and inference alike. The same geography that makes the buildout cheap is what makes it fragile. The U.S. side of the race has its own version of this story: Stargate's 10 gigawatts, the five-site expansion OpenAI announced, all of it has to be sited, powered, cooled, and permitted. Ulanqab just makes the physical inputs visible earlier, in a place where the data is already public.

The next test is whether the new data centers in Ulanqab come online with closed-loop cooling and on-site water reuse at hyperscale norms, or whether they draw from the same municipal supply that was already turned off last month. If the former, the resource-wall framing ages out. If the latter, the constraint becomes a permitting and political fight, not an engineering one. The Goldman Sachs note, the 70 percent spike in announcements, and the operator-specific builds from DeepSeek, ByteDance, Alibaba, and Xiaohongshu give the next twelve months enough public milestones to track. The water meter in Ulanqab is the cleanest one.