A SemiAnalysis report describes Envision Technology's Xinghe computing campus in Ulanqab, Inner Mongolia, which is powered by two self-built 420MW wind farms on dedicated transmission lines backed by four-hour energy storage. Critically for AI data center power design, the campus uses 800V DC power distribution and eliminates central chillers, in contrast to the roughly 97% of US data centers that still rely on AC architectures. The report highlights first-phase delivery in under a year at about one-fifth the cost of comparable US projects.
Overview: Two 800-MW Campuses, Two Very Different Timelines
A report from semiconductor and AI research firm SemiAnalysis has drawn an unusually sharp contrast between two AI computing campuses of comparable scale. In September, analysts Everlyn Tan and Jeremie Eliahou Ontiveros toured Envision Technology Group's Xinghe campus in the Bayin district of the Chahar High-tech Development Zone in Ulanqab, Inner Mongolia. Their findings, published Nov. 7, describe a facility with a planned IT load of 800 MW — the same order of magnitude as OpenAI's flagship Stargate site in Abilene, Texas — whose first 120 MW phase was delivered in August 2026, less than a year after construction began, at roughly one-fifth the cost of a comparable U.S. project. The analysts characterized the campus as China's West Texas, but noted that the speed, scale, low cost, and highly vertically integrated model stand in stark contrast to the U.S. trajectory, where anti-data-center sentiment has become a political force.
The Abilene Backlash: Noise, Rent, and Permitting
The Stargate campus in Abilene — a city of about 131,600 people roughly 180 miles from the Dallas–Fort Worth metroplex — is the most mature of the program's seven sites, with eight data halls running Nvidia Blackwell chips. It carries about 1.019 million equivalent H100 units of compute, 824 MW of IT power, and $31.2 billion in capital investment, built by Crusoe for Oracle-owned chips used by OpenAI. Residents describe a different ledger. Arlene Mendler, who has lived in Abilene for nearly 35 years, said the site operates 24 hours a day, seven days a week, with the night sky glowing white. A professor named Rusty Trowell said his home now sits about 200 feet from the data center and requires blackout blinds to sleep. Omaira Garcia's ranch abuts a large gas plant whose turbine exhaust stacks are visible from her kitchen window. The city council granted an 85% property tax abatement to retain the investment, while Zillow data show average monthly rent near $2,600, up nearly 50% year over year and now above Austin's. Residents also complain that developers filed a simple permit — which requires no environmental study, public notice, or comment period — and only a year later submitted a formal permit, with plans to more than quadruple the number of gas turbines.
Power Architecture: Self-Built Wind and Dedicated Transmission
The most consequential divergence is at the power layer. Abilene relies primarily on the public grid plus on-site gas turbines. Ulanqab takes the opposite approach: Envision built two wind farms outside the campus with a combined 420 MW of capacity, connected by dedicated transmission lines, backed by a four-hour energy storage system to reduce single-point dependence on the public grid. The target is for more than 60% of consumption across the first two phases to come from the company's own wind assets. This is a structural choice, not a marginal optimization. Research by Maurizio Porfiri's team at NYU Tandon, covering 4,283 U.S. data centers, found that 97.5% sit in metropolitan or micropolitan areas because operators prefer to attach to existing power infrastructure rather than build generation.
800V DC and the Solid-State Transformer Pilot
SemiAnalysis reports that roughly 97% of U.S. data centers still run AC distribution architectures. The Xinghe campus has piloted a solid-state transformer in a 5 MW data hall that converts 10 kV AC directly into 800V DC for server racks. Zhu Keli, founding dean of the Guoyan New Economy Research Institute and a chief expert on the intelligent economy, told National Business Daily that conventional designs pass through multiple AC/DC conversion stages, each incurring losses; a direct 800V DC output cuts distribution-chain losses with meaningful long-run energy savings, while higher equipment integration eliminates large volumes of AC switchgear and frees floor space. SemiAnalysis notes U.S. operators remain in early exploration of such architectures.
Cooling and Cost: No Central Chillers
Xinghe dispenses with central chillers entirely, reducing both energy consumption and construction cost — a departure from most U.S. facilities, which retain chiller plants. Zhu notes that for large AI clusters, electricity typically accounts for 60% or more of operating cost, making Ulanqab's abundant wind and solar resources and ample land the decisive siting factors. Combined with a sub-one-year first-phase build and one-fifth the U.S. cost benchmark, the campus compresses both capital and operating expense.
Market Context: The Politics of Siting
The contrast lands in a changed political environment. In the 2026 U.S. midterms, Democrat Gina Hinojosa, who campaigned on pausing new data center approvals, defeated incumbent Republican Governor Greg Abbott in Texas, the fastest-growing data center market in the country. Survey data provided to National Business Daily by Sean Patterson II of the University of Pennsylvania's Annenberg Public Policy Center, drawn from 1,320 U.S. adults between mid-June and mid-July, show local opposition to new data centers rising from 49% to 61% versus the spring, with majorities among Democrats (69%), Republicans (54%), and independents (53%), and 70% among adults under 30. Pew Research Center data cited in the article indicate 67% of planned new U.S. data centers are sited in rural areas.
Why the Model Does Not Travel Easily
Zhu attributes the U.S. backlash to a mismatch between supply, demand, and governance tempo: capital floods in, projects cluster near residential areas, grid expansion lags load growth, and water, noise, and land externalities hit host communities without early participation or benefit-sharing mechanisms. He argues the Ulanqab model is hard to replicate in the United States because the power system is highly fragmented across many private utilities, making cross-entity coordination difficult; wind, storage, and compute projects sit on separate approval chains; and community resistance is high. Capital can buy rural land for a data center, but rarely assemble a synchronized renewable base and unified storage dispatch into a closed loop.
Implications for AI Data Centers
The deeper lesson is that Ulanqab is not merely cheap power procurement. It requires coordinated land-use planning, energy planning, and compute-network planning, unified grid dispatch capable of absorbing variable wind across time windows, and supporting industrial policy. For U.S. operators, the transferable elements are narrower but real: DC distribution at the rack, solid-state transformation, and chiller-free cooling are technical choices independent of grid structure, and each attacks the operating-cost line that dominates AI economics.