Ulanqab has pledged 12.5GW of AI data centers, and its water goes off at night

Goldman Sachs is counting pledges, and pledges are cheap ^[3]. The pace inside its note is the harder number: roughly 8.75 gigawatts of that capacity was announced in a single twelve-month window ^[2], in a city of about 1.5 million people ^[1]. That is around 8.3 kilowatts of promised capacity per resident ^[3]. The total sits 25 percent above the 10 gigawatts OpenAI's Stargate project expects to reach at completion ^[1], and Stargate carries a $500 billion price tag ^[6].

None of that was decided by silicon. Ulanqab sits high on the Inner Mongolian Plateau with long cold winters, so cooling draws less power ^[7], and electricity in Inner Mongolia is cheaper than almost anywhere else in China, the product of fast wind and solar growth alongside abundant coal ^[8]. Huawei built the city's first data center in 2016 and Apple followed three years later ^[16]; in 2021 the region was named a main hub of the state's "Eastern Data, Western Compute" program ^[17]. For years the distance from the populous coast left those buildings doing backup storage ^[18]. Andrew Stokols of Singapore Management University told Wired that the rise of AI in 2022 brought the realization that remote sites could be well used for model training ^[19], which runs for months and tolerates latency ^[20].

So the qualifying list is cold air, cheap electrons, fiber and land. Water is not on it. Ulanqab gets about 14 inches of rain a year, roughly as dry as Denver ^[12], and the local government already cannot supply enough water for residents, before most of the announced projects are running ^[13]. The defense of the buildout is seasonal: weather data from the local government shows the centers need additional cooling water in only two months of the year ^[15]. That is the municipality's own figure, and it is the figure holding up the pledge column.

The ownership change is what turns a water question into a corporate one. Wired frames the Ulanqab spending as the first time Chinese AI companies have put real money into infrastructure they own rather than renting compute from cloud providers ^[9], after years of spending far less on physical plant than their American peers ^[11]. Renting is exposure to a price list. Owning the buildings is exposure to a municipal water table and a grid interconnection queue, in a jurisdiction where, Wired reports, growth has been driven more by commercial demand than by government direction, unlike other Chinese hubs ^[22]. Nearly 100 facilities have opened or broken ground there since 2016 ^[2], which means the operators arriving now are inheriting a water budget already spoken for.