← Back to feed

Envision Group has commissioned the Ulanqab Galaxy Base, a 120,000-square-meter (roughly 29.7-acre) AI computing facility in Inner Mongolia that it claims is the world’s largest single AI data center building. The facility went live on August 6.

The headline number is the building size. The more consequential detail is the power architecture.

Bypassing the Queue

Western AI infrastructure is running into a structural constraint that has nothing to do with chip supply or capital. Grid interconnection queues in Northern Virginia, Phoenix, and Dallas now run four to seven years. Goldman Sachs Research estimates the United States faces a 9.3-gigawatt power shortfall in 2026, widening to 45 gigawatts by 2028. Companies can secure land, sign GPU contracts, and raise capital in months. Getting permission to draw power from the grid takes a decade.

Envision’s approach at Ulanqab is different: build the energy generation assets next to the servers and skip the grid entirely. The facility uses what Envision calls its AI Power System — a proprietary platform that coordinates on-site renewable generation (solar and wind), battery energy storage, and dedicated transmission infrastructure to deliver power directly to the compute floor. No interconnection queue, no utility negotiation.

The Facility

• Footprint: 120,000 sqm — equivalent to about 20 standard football fields
• Location: Ulanqab, Inner Mongolia, in the Gobi Desert
• Planned capacity: 2+ gigawatts across the full campus buildout
• Compute density: Million-card parallel computing capability at current commissioning; million-P scale targeted
• Power: Ultra-high proportion of direct renewable connection, no primary grid dependency

The campus is the flagship project under Envision’s “Gobi Mission,” which targets locations where renewable generation assets already exist and where the economics of building dedicated power infrastructure are more favorable than waiting for grid connection in urban markets.

Why Inner Mongolia

The Gobi Desert offers the same logic that drew early cloud computing to the American Southwest: cheap land, cold air for natural cooling, and proximity to renewable generation capacity. Inner Mongolia has extensive wind resources and receives strong solar irradiance. Building generation assets on the same site as the compute — rather than connecting to a grid built for residential and industrial load in a different era — eliminates the interconnection bottleneck as a variable.

Envision’s model also sidesteps the political friction that has dogged Western AI data center proposals. Community opposition has become a meaningful constraint: $130 billion in AI data center capacity was blocked in Q1 2026 alone, with residents citing noise, water use, and electricity competition. A Gobi Desert site far from residential areas removes that variable.

The Token Factory Framing

Envision’s internal framing for the campus is “token factory” — a facility optimized not just for raw compute capacity but for token output per watt and per dollar. The million-P scale reference aligns with current Chinese compute density standards using domestic accelerators, primarily Huawei and domestic AI chip suppliers rather than Nvidia hardware.

China’s AI compute buildout is following a parallel path to the US: massive capital deployment, but on fundamentally different infrastructure assumptions. Where US hyperscalers are retrofitting grid-connected campuses, Chinese AI infrastructure is being built from scratch in locations where renewable power is abundant and the regulatory environment moves faster. Envision’s Galaxy Base is the most visible example yet of what that looks like at scale.