Envision brings 2GW-planned AI compute base online in China
The Ulanqab Xinghe base entered production with a 120,000-square-metre hall, million-card parallel design and direct green-power supply, phase one of a 5GW plan.
What launched
Envision Group's Xinghe base in Ulanqab, Inner Mongolia, entered production on Thursday. The company describes it as a single AI computing hall of about 120,000 square metres — roughly twenty football pitches — engineered for million-accelerator parallel operation and a compute scale it puts at one million PFLOPS. Total planned electrical capacity for the site is 2GW, supplied largely through direct connection to Envision's own wind and solar generation rather than through the public grid.
Ulanqab is one of the eight national hubs designated under China's "East Data, West Computing" programme, chosen for cool climate, cheap land and proximity to renewable generation; roughly two-thirds of the region's electricity already comes from wind and solar. Envision, better known as a wind-turbine and battery manufacturer, is positioning the site as infrastructure for domestic AI labs and cloud providers that need large blocks of power-dense capacity inside China.
The base is the first phase of "Mission Gobi," announced in June, under which Envision says it intends to build 5GW of green AI compute capacity across desert regions worldwide by 2030.
Context
The launch lands as power, not silicon, becomes the binding constraint on AI buildouts. In the United States, Texas has frozen new data-centre grid connections pending an audit and Nashville moved this week to block a project by eminent domain; Google's $15bn Indian campus is facing scrutiny over water use. A developer that owns its own generation sidesteps the interconnection queue entirely — the structural advantage Envision is selling.
Why it matters
China's compute expansion has been discussed mostly through the lens of chip export controls, but the practical bottleneck is increasingly megawatts and the years it takes to procure them. Pairing renewable generation directly with an AI hall compresses that timeline and insulates operators from grid politics. If the claimed density holds, the site materially raises the ceiling on how much domestic training and inference Chinese labs can run at home — and offers a template that generation-owning industrial firms elsewhere are likely to copy.