Envision has built the first 120 megawatts of the Galaxy AI campus in Ulanqab, Inner Mongolia, in less than a year, according to SemiAnalysis. The researchers put construction cost for that phase at one-fifth of comparable US projects. The completed Galaxy AI campus is planned to provide up to 800 megawatts of IT capacity by the end of 2027, placing its planned scale close to the roughly 820 megawatts of critical IT capacity at Stargate Abilene in Texas. The comparison matters because the price of AI compute increasingly depends on the cost of the powered building, not only the price of the accelerator.
Galaxy AI campus: Power and construction economics
SemiAnalysis visited the site in September and published its report on November 7, 2026. The report rests on that visit and Envision’s own materials, so the cost comparison has not been independently verified. Agora Intelligence noted that Galaxy’s costs are claims made by the builder and the analyst, and that SemiAnalysis sells a paid model covering these data centers.
Still, the figures point to a sharp regional gap. For 2026, SemiAnalysis puts Chinese capital expenditure below $3 million per megawatt. With Galaxy close to the Chinese average, the implied US benchmark sits below $15 million per megawatt. Chinese rents are about $40 per kilowatt per month, excluding power, while power in Inner Mongolia costs about $0.055 per kilowatt hour all-in. Agora Intelligence argued that the cost of building a powered hall, rather than the public price per token, will weigh more on AI compute over the next three years.
The scale comparison holds because the campuses are close in size. Crusoe began building Abilene in mid-2024 and leases it to Oracle, which runs it for OpenAI. The Texas site has about 820 megawatts of critical IT capacity on 1.2 gigawatts of power. Envision’s Galaxy AI campus is a different company from the Galaxy that leases the Helios campus in Texas to CoreWeave, which is important for readers tracking names across the market.
Renewables, batteries and the rack
Envision is not waiting for the local grid to carry the entire load. The company has installed more than 100 gigawatts of wind turbines worldwide and wants its wind farms to provide over 60 percent of the electricity needed by the first two phases. Those wind farms sit about 80 kilometers apart and have a combined capacity of 420 megawatts, well above the expected 240 megawatt load from those phases. A four-hour battery storage system smooths the supply when wind generation falls.
Inside the buildings, SemiAnalysis visited a 5 megawatt hall where solid-state transformers take 10 kilovolt AC electricity and convert it directly into 800 volt DC for the server racks. The 800 volt DC design removes low-voltage transformers, uninterruptible power supplies and rectifiers from the power chain. Fewer conversion stages can reduce equipment requirements and energy losses, although SemiAnalysis did not provide a measured efficiency improvement for the pilot.
Envision says its controller coordinates the transformer and batteries within milliseconds. The company describes the deployment as the first commercial use of this configuration, although independent performance figures have not been disclosed publicly. TechRadar said the claimed operational advantages await detailed independent evidence. The facility uses what is described as the world’s first commercial integrated 800 volt DC power system, and Envision plans a larger version of the pilot in later construction phases.
Cooling and factory-built infrastructure
The campus runs liquid-cooled and air-cooled servers on a single water loop, eliminating the need for conventional central chiller plants. The approach can simplify infrastructure, though it depends on local climate conditions and the facility’s ability to reject heat through its water loop. The prefabricated substation arrives factory-tested and is installed in under 3 months, against about 6 for a conventional substation. Envision also factory-tests batteries from its affiliate AESC, according to Agora Intelligence, while single-component suppliers leave commissioning to the site.
These design choices align with a broader shift in power distribution. Nvidia’s Kyber racks draw about 600 kilowatts each. SemiAnalysis says about 97 percent of US data centers still use AC uninterruptible power supply systems, while about a quarter of the halls built in China over the past decade run on DC. That difference makes Galaxy’s pilot a commercial test of an architecture that US operators have largely avoided at scale.
Regional buildout and the accelerator question
Galaxy is the first major project in Envision’s Inner Mongolia datacenter plans, which add up to 3 gigawatts. The surrounding Chahar High-tech Zone has nearly 30,000 construction workers spread across 22 data center projects. The Inner Mongolia region has 25 gigawatts of announced datacenter projects, and Alibaba, Huawei and ByteDance are among the firms expanding infrastructure in the region. SemiAnalysis compares the area to West Texas for its open land, cheap energy and few nearby residents to object, but says Texas is seeing its highest ever level of anti-datacenter sentiment.
The main Galaxy site covers 120,000 square meters and is designed to hold up to one million AI accelerators. Its target computing capacity is one million PFLOPS. Roughly 120 megawatts of capacity has already been commissioned, including capacity tied to agreements with major Chinese tech firms, and a 200 megawatt phase was expected to be handed over by late October 2026. Envision’s long-term target for the site is 3 gigawatts of capacity by the end of 2028.
More than 80 percent of the facility’s electricity is meant to come from wind and solar, and Envision has secured offtake agreements running through 2029 to 2030 with customers in fields such as autonomous driving and content generation. The campus is the flagship of Envision’s Mission Gobi initiative, which aims to build 5 gigawatts of green AI datacenter capacity worldwide by 2030. If Galaxy alone reaches its 3 gigawatt target, one campus would account for most of that global goal.
Access to power and buildings does not automatically create usable AI computing power. China’s access to advanced accelerators remains a separate constraint, particularly as US export restrictions limit its ability to acquire Nvidia’s most powerful chips. Even so, the Galaxy AI campus shows that the physical hall can now be built faster and more cheaply in Inner Mongolia than in Texas. For capacity buyers, the practical risk sits in the price per megawatt built into a lease, not in the accelerator list price. As Agora Intelligence put it, margin goes to whoever builds the megawatt cheaply more than to whoever owns the gigawatts.