China's "Stargate" in Inner Mongolia: 120 MW in a year and 800 V DC

While everyone debates which accelerators China can access, SemiAnalysis suggests looking at another part of the equation: where those accelerators will be housed and what will power them. The team visited the so-called "Chinese Stargate" in Inner Mongolia and published a report. Its post on X provides only a brief summary, which deserves a closer look.

What they saw

According to SemiAnalysis, this is an 800 MW campus powered by wind. Of that capacity, 120 MW was built in less than a year. For perspective, 120 MW is already the scale of a large hyperscale data center, rather than a demonstration site.

There are also two engineering details. First, a 5 MW pilot built around solid-state transformers (SST) and 800 V direct-current power distribution. Second, a cooling system without central chillers.

A caveat up front: we have only a short announcement, and we have not read the full report. We therefore do not know exactly which accelerators are installed at the campus, their utilization rates, how much the project costs, or how the pilot performs under real workloads.

Why these details matter

A conventional data center power system works like this: a medium-voltage grid connection, a bulky transformer, uninterruptible power supplies, alternating current distributed along the rows, and a power supply in each server converting it back to direct current. Each conversion consumes a few percent of the energy and adds equipment that takes up space and can fail.

An SST replaces a heavy conventional transformer with power electronics that can supply direct current directly. An 800 V DC bus eliminates some intermediate conversions and allows the same power to be transmitted through thinner cables. Similar thinking is already visible in industry plans for next-generation racks: power density is rising so quickly that the old architecture is running up against copper and space constraints.

The cooling side is simpler. Central chillers are large compressor-based systems that cool water for an entire building. Doing without them implies relying on warmer fluid in the cooling loop and free cooling. The region's cold climate probably helps, but that is our assumption, not a statement from the report.

Why combine all of this? Increasingly, the constraint on AI infrastructure is how quickly inexpensive power can be delivered and heat removed, rather than chip availability.

Limitations and practical implications

The numbers need to be read soberly. Of the planned 800 MW, 120 MW has been built, or about 15%. The 5 MW pilot represents only about 0.6% of the announced capacity. This is an experiment, not a proven standard. Wind generation is also intermittent, and the summary does not explain how the campus covers nighttime and windless periods.

Practical takeaways for people who do not build data centers themselves:

— If your product depends on the price per token or GPU-hour, watch the power sector as well as chips. It determines the underlying cost of inference over the long term. — Do not base forecasts on the assumption that export restrictions alone limit the pace of China's AI infrastructure development.

The speed at which capacity comes online is a separate factor, and the claim here is 120 MW in a year. — When renting capacity from cloud or colocation providers, ask about the power supply architecture, power density per rack, and cooling design. In a couple of years, these will affect prices and delivery times as much as the processor model.

— For infrastructure startup founders: 800 V DC, SST, and liquid cooling represent an emerging market for equipment and services.

The post points to an interesting development, but the details need to be checked against the SemiAnalysis report itself. The direction, however, is clear. The AI race is also being fought in transformers, power buses, and cooling towers, and the winner is whoever connects megawatts faster and more cheaply.