Ulanqab Pledged More AI Compute Than OpenAI's U.S. Stargate. Most of It Isn't Built Yet.

Ulanqab sits in central Inner Mongolia, four hours west of Beijing by rail, and until this year it was best known for potatoes. As of June 2026 it has 12.5 gigawatts of AI data-center capacity committed, a number that, per a Goldman Sachs research note summarized by WIRED, puts it past OpenAI's initial 10-gigawatt U.S. Stargate pledge. Ulanqab, not Shenzhen, not Hangzhou. The buildout is being selected by power, land, and cold air, not by where the labs are.

Goldman counts the 12.5 GW as the sum of operating plus planned project commitments, and UBS pegs today's live capacity at roughly 1.2 GW. That gap is the story. The pledge is real, the dirt is not yet turned. QbitAI (量子位) sent a reporter to Ulanqab in August; the field report found that most new campuses are still in pre-construction, with permits pulled, land scraped, and no power online. The earliest operational date the reporter could verify is the end of 2027, after the spring thaw.

Ulanqab sits on a high plateau where hyperscalers pick sites for free-air cooling, which keeps the power budget on the GPU rack instead of the chiller. Electricity in Inner Mongolia is also cheap because it is coal-fired and often curtailed, a problem that turns into a subsidy when you build a load that absorbs it. Land is abundant, and the city government will pre-grade a site. Per the Goldman note, via WIRED, more than 70% of the 12.5 GW was announced in the past twelve months.

The tenant list is the second tell. GW-scale projects are underway from three of China's top four data-center operators, GDS, VNET, and Chindata, plus Envision, ZDATA, and Centrin. Kuaishou, UCloud, ByteDance, and Z.AI are building their own. Envision opened the first phase of its Galaxy Campus on August 6: 120,000 square meters, 2 GW designed capacity, room for a million GPUs. The U.S. equivalent of a single anchor tenant taking that much power in a county of 1.7 million people would be a federal permitting event. In Ulanqab it is a ribbon cutting.

Two other signals from the same week of Chinese-language reporting make the buildout feel less inevitable. National Day, China's October 1 holiday, was the first mass-market test of AI travel assistants, the kind that plan an eight-day route across a country the size of the United States. The wire called it a triumph. The actual product reviews, summarized in the ChinAI #376 roundup, called it a letdown. Trip planners hallucinated closed restaurants, ignored booking windows, and recommended routes that ignore the Golden Week rail reservation system. Travelers went back to Xiaohongshu.

The second signal is quieter. The platform companies at the top of the Chinese AI stack — Baidu and Alibaba — are reportedly redoing pre-training datasets, the trillion-token corpora that go into a base model, because the previous round picked up too much synthetic and duplicated web text. A pre-training redo is not a launch. It is the unglamorous admission that the last model was trained partly on its own exhaust, and that the next one will not be. Capability will keep arriving, but the iteration cost is the data itself.

Taken together, the three signals say the same thing. Where AI gets built next is a function of megawatts and land deeds, and how fast it gets better is a function of how clean the next training corpus is. Both are slower than the demo cycle suggests. Ulanqab is the rare case where you can see the difference, because the GPUs are not yet on, the rooftops are not yet covered, and the next twelve months will be spent turning a pledge into a building.