EN / 中文

Envision Galaxy Campus Ulanqab: 1/5 US Construction Cost, Vertical Source-Grid-Load-Storage AI Compute Experiment

by xinzhixun·November 12, 2026

Recently, US semiconductor research firm SemiAnalysis released an on-site survey report that has placed Envision's "Galaxy Campus" in Ulanqab, Inner Mongolia, at the center of global discussions on AI infrastructure.

SemiAnalysis noted in its report that the first phase of the Galaxy Campus, with a capacity of 120MW, was built and delivered in less than one year, with unit construction costs approximately one-fifth those of similar projects in the United States. However, beneath the narrative of a "cost miracle," the true logic of this campus is not simply "fast and cheap construction," but rather a vertical integration experiment by energy companies using data centers as channels to convert electricity into computing power commodities.

Faster Construction, Costs Only 1/5 of the US

Envision Group, established in 2007, started with smart wind turbines and has evolved into a leading global green technology enterprise, with businesses covering smart wind power, energy storage systems, power batteries, green hydrogen, and intelligent IoT operating systems. Its entry into the AI data center sector around 2024 was not merely positioning itself as a "green electricity supplier," but redesigning the entire power chain from generation to computing load as an energy system integrator.

The first phase of Envision's Galaxy Campus, with a capacity of 120MW, completed construction and customer delivery in less than one year. The Phase I infrastructure was completed and put into operation on August 6 this year. After approximately three months of electromechanical commissioning and computing equipment deployment, it was delivered at the end of October. The entire campus is planned to reach a capacity of 800MW by the end of 2027, comparable to the planned scale of OpenAI's Stargate core campus in Abilene, Texas, which is about 820MW. However, the unit construction cost of the Galaxy Campus is approximately one-fifth that of similar US projects.

SemiAnalysis's survey shows that for a Chinese data center with a volume of 100–120MW, during the peak period of electromechanical installation and electrical integration, the number of on-site workers reaches 3,000 to 4,000, which is 3 to 5 times that of US construction sites with similar labor scales.

However, the cost advantage of the Galaxy Campus lies not only in having more workers to shorten the schedule, but in Envision's deep vertical integration—self-built wind farms, self-manufactured energy storage, co-designed prefabricated substations, and applying experience in energy equipment manufacturing to data center construction. For example, prefabricated substations are manufactured and tested in modules within factories, progressing in parallel with on-site civil engineering, and then transported to the site for connection. The entire process takes less than 3 months, whereas traditional indoor substations require about 6 months.

Power Supply Model: Self-Built Wind Farms and "Source-Network-Load-Storage" Integration

Unlike US AI data center developers who tend to install natural gas turbines on-site to bypass grid connection bottlenecks, Envision chose a different path.

Inner Mongolia, where the Galaxy Campus is located, possesses about 57% of China's available wind resources. Wind farms here operate at full load for an average of about 2,800 hours per year, with a capacity factor of approximately 32%. Specifically, the wind capacity coefficient in Ulanqab ranges between 35% and 40%. Relevant calculation data shows that the comprehensive electricity cost in Inner Mongolia is approximately $0.055 per kWh, making it "one of the lowest-cost regions globally."

Therefore, Envision plans to supply power to the Galaxy Campus through two dedicated self-owned wind farms located about 80 kilometers away, with a total capacity of 420MW, nearly 1.75 times the expected load of 240MW for the first two phases. A supporting 4-hour battery energy storage system is used to maintain power continuity during windless periods, aiming for over 60% of the electricity consumption in the first two phases to come from self-owned wind power.

This model is referred to by Envision as "Source-Network-Load-Storage" integration. Its underlying logic is that whoever controls the energy controls a key revenue stream in data center operations.

According to data disclosed by SemiAnalysis, data center rents in China are approximately $40 per kW per month (excluding electricity costs), equivalent to about $480,000 per MW per year. With the comprehensive electricity cost in Inner Mongolia at approximately $0.055 per kWh, calculated at full load, the annual electricity cost per MW is about $482,000, roughly equal to the rent.

For enterprises like Envision that simultaneously control wind power, energy storage, and power systems, self-built power sources not only lock in costs but also make data centers a consumption scenario for their energy equipment. SemiAnalysis pointed out in its report that data centers "create new demand for electricity and equipment already produced by Envision," and their returns "often come from elsewhere." In other words, Envision's deeper motivation for building an 800MW campus is not just to earn server room rent, but to find long-term buyers capable of absorbing massive amounts of electricity for its wind turbines and energy storage systems.

Power Distribution and Cooling: 800V DC Implementation, Central Chillers Eliminated

In terms of data center power architecture, a 5MW test machine room at the Galaxy Campus has actually implemented an 800V high-voltage DC system. Solid-state transformers directly convert the grid's 10kV AC power into 800V DC to supply server racks, eliminating multiple intermediate conversion stages such as low-voltage transformers, UPS, and rectifiers required in traditional data centers.

Although NVIDIA also promotes the adoption of 800V DC distribution for future high-density AI systems in its technology roadmap, the US industry as a whole remains in the early research stage.

The cooling design of the Galaxy Campus also breaks conventional norms. Due to Ulanqab's average annual temperature of about 4°C, natural cooling sources can be utilized for nearly 10 months a year. Therefore, the campus connects liquid-cooled and air-cooled servers to the same water circulation system, completely eliminating central chillers. This design not only reduces cooling energy consumption but also avoids the risk of overheating in computing clusters due to central chiller failures.

Cluster Effect in Ulanqab and the Computing Power Economic Landscape

The Galaxy Campus is just a slice of Ulanqab's computing power expansion. According to a report by Xinhua Daily Telegraph in October, there are 22 computing power projects under construction in the local Chahar High-Tech Zone, with nearly 30,000 construction workers. A Goldman Sachs report from August this year showed that as of the end of June, the committed data center capacity in Ulanqab (including operational, under-construction, and planned) totaled approximately 12.5GW, about ten times the actual operational capacity of 1.2GW in 2025. Companies such as Alibaba, Huawei, and ByteDance have already established bases locally, and DeepSeek is reportedly planning to build a GW-level data center there as well.

Against the backdrop of continuously tightening US chip export controls, China still faces constraints in accessing the most advanced AI accelerator chips. However, what the Galaxy Campus and its surrounding cluster demonstrate is competition in another dimension: leveraging energy cost advantages, vertically integrated supply chain capabilities, and highly parallel engineering organization to compress the unit economics of computing power to a level that cannot be ignored by the global market.

In previous reports, SemiAnalysis estimated China's total delivered data center capacity at over 24GW, compared to about 56GW in the US. Although the capacity gap remains, the delivery cycle for China's 100MW-class projects has been compressed from 18 months in the cloud era to about 12 months, with some projects even shorter, and related construction costs are also lower.

Editor: ChipInsight - Langke Jian