Journey simply two hours west of Beijing by practice, and also you’ll end up surrounded by the rolling grasslands and historical cinder cones of Internal Mongolia. This huge, arid land has lengthy been China’s capital of sheep farming and coal mining, however over the previous couple of years, it has turn into the most well liked place within the nation to construct an AI data center.
In Ulanqab, a metropolis in Internal Mongolia residence to about 1.5 million individuals, practically 100 information facilities have been opened or begun development since 2016. Chinese language firms have pledged to construct initiatives with a mixed estimated capability of 12.5 gigawatts within the metropolis, and over 70 % of the whole commitments have been introduced in simply the final 12 months, making it one of many quickest rising compute clusters in Asia, in accordance with a analysis notice revealed by Goldman Sachs final week. For comparability, OpenAI’s $500 billion Stargate Challenge is about to achieve solely 10 gigawatts of complete capability when it’s full.
Chinese language firms are flocking to Ulanqab for a lot of causes. The town sits at excessive elevation on the Internal Mongolian Plateau and has lengthy, chilly winters, which suggests information facilities there don’t want to make use of as a lot power to remain cool. It’s additionally comparatively near Beijing, so information may be transmitted to China’s populous areas with minimal latency. However essentially the most engaging issue has to do with prices. Electrical energy is cheaper in Internal Mongolia than virtually wherever else in China, pushed by each the sturdy development of wind and photo voltaic power and an ample provide of coal.
What’s additionally fascinating is who is constructing these information facilities. For the primary time, Chinese language AI firms are making large investments in their very own infrastructure, relatively than renting compute from cloud firms. DeepSeek is reportedly constructing an enormous AI information heart in Ulanqab, as are ByteDance, Alibaba, and Xiaohongshu. For years, Chinese language AI firms have spent far much less on constructing bodily infrastructure than their American friends, regardless of creating a lot of standard AI fashions with spectacular capabilities. The Ulanqab information heart increase alerts that now they’re lastly beginning to catch up.
There’s only one downside: discovering sufficient water. Ulanqab is about as dry as Denver, getting solely roughly 14 inches of rain annually. The native authorities is already struggling to offer sufficient water to satisfy resident demand—earlier than most of the deliberate information heart initiatives are even up and operating. Final month, the native water firm in Ulanqab was compelled to turn off a number of waterworks for seven hours every evening to mitigate peak demand. The information facilities being constructed within the metropolis will want much less water within the winter—climate information from the native authorities of Ulanqab exhibits they solely require further water for cooling throughout two months out of the 12 months—however the entire new infrastructure might nonetheless pose a major environmental problem for the area.
The Boonies
Internal Mongolia has been a knowledge heart scorching spot for no less than a decade, lengthy earlier than the present AI increase. Huawei constructed its first one in Ulanqab in 2016, and Apple adopted swimsuit three years later. In 2021, the realm was designated as one of many fundamental hubs of a country-wide authorities mission dubbed “Eastern Data, Western Compute,” which goals to construct information facilities within the Western hinterlands of China.
There was one main downside, although. As a result of they’re situated removed from China’s populous jap coast, these information facilities initially confronted excessive latency charges when transferring information to nearly all of customers. Because of this, they had been initially largely relegated to backup storage—till AI gave them a brand new goal. “With the rise of AI in 2022, there was the belief that truly, these distant information facilities might be well-utilized for mannequin coaching,” says Andrew Stokols, a professor at Singapore Administration College who research China’s compute infrastructure. A coaching run for an AI mannequin can take months and doesn’t require a lot real-time tinkering, so latency is much less of a problem.
Comparatively talking, Ulanqab can also be probably not that distant. Internal Mongolia is far nearer to Beijing and different main metropolitan areas in China than any Western information heart hub is. And it’s now related by two devoted fiber optics cables inbuilt 2017 and 2019 that lowered common latency speeds to lower than 5 milliseconds, quick sufficient to help real-time information exchanges like AI inference.
