AI campuses and GPU halls
Stargate, Colossus, the CoreWeave and Crusoe sites: halls built around tens of thousands of GPUs. 72 data centres on EveryDatacenter so far.
An AI campus is a data centre built around very large clusters of GPUs or other accelerators for training and running AI models. Racks in these halls draw 100 kW or more each, several times a normal rack, so they are cooled by liquid and fed by power lines and substations of hundreds of megawatts. Stargate at Abilene, Colossus at Memphis and Meta's Hyperion in Louisiana are the best known.
Why AI changes the building
Training a large model needs tens of thousands of chips that talk to each other at high speed, so they must sit close together on one site. A single Nvidia GB200 NVL72 rack draws about 120 kW, against 5 to 15 kW for a typical air-cooled rack. That density forces direct liquid cooling, heavier floors, new power distribution and much bigger grid connections. Many AI campuses are planned at a gigawatt or more, the output of a large power station.
Who builds them
Hyperscalers build their own, but much of the new capacity comes from specialist "neoclouds" such as CoreWeave, Nebius and Nscale, from crypto miners turning their sites over to AI, and from joint ventures such as Stargate, formed by OpenAI, SoftBank, Oracle and MGX in January 2025. Governments are building national AI capacity too, from Saudi Arabia's Humain to India's IndiaAI Mission, which buys GPU time from Indian data centre firms.
Power and local debate
Because grids cannot always supply such loads quickly, some AI campuses run their own gas turbines. Colossus in Memphis drew complaints from residents and environmental groups over turbines run before permits were granted, while xAI said the turbines were temporary and later received a permit for some of them. Supporters stress jobs, investment and national competitiveness; critics question air quality, water use and who pays for new power plants.