China builds giant AI data centers fast and at a fifth of US cost

Analysts from SemiAnalysis toured a huge AI data center site in Inner Mongolia, China, and came away stunned by the speed and low cost. The first 120MW phase of the Galaxy campus was finished within a year, at a fifth of US costs. A megawatt (MW) is a measure of electric power.

What they saw

Galaxy sits in Ulanqab and is planned for up to 800MW. That is about the size of the core campus at Stargate Abilene, the Texas site built for OpenAI that has become a symbol of America's AI buildout. All four phases of Galaxy are built to order for two internet giants. The report does not name them.

The surrounding Chahar High-tech Zone had nearly 30,000 construction workers on site across 22 data center projects, according to a local report. With that much labor, Chinese builders rarely use a fully modular, factory-made approach, yet costs keep falling.

How it stays cheap

The owner is Envision Group, which has installed more than 100GW of wind turbines worldwide. It builds its own wind farms and a dedicated power line. Its blade factory is across the road. It also designed much of the power equipment inside and made the batteries that smooth out the load.

• Power: a pilot hall uses solid-state transformers to turn grid power straight into 800V DC, skipping several middle steps.
• Cooling: liquid- and air-cooled servers share one water loop, with no central chiller plant.
• Precedent: about a quarter of Chinese data centers built in the past decade run on DC, versus about 97% of US ones using AC backup systems.

Texas versus Inner Mongolia

The report says Texas faces its highest ever anti-data-center sentiment, while Inner Mongolia is speeding up with no such backlash. The region has 25GW of announced data center projects, including several at gigawatt scale. Local policy encourages bringing new demand to where the wind power is.

What is not known

The report gives no dollar figures in the material available here, and the cost claim comes from SemiAnalysis, not an outside audit. Returns on these projects are thin, so the payoff often comes from elsewhere. For Envision, data centers create demand for power and equipment it already makes. Whether the 800V pilot scales up is also untested. SemiAnalysis plans a follow-up on limits such as power, land and water.

Why it matters

Cheaper, faster buildouts abroad could shape what AI costs to run and how hard US firms must compete. The same 800V design is still on the US roadmap, so China offers an early look at how it works in practice.