The AI Data-Centre Map: Who Is Building What, and How Much Is Real

Data as of 2026-10-22 · updated monthly · every row sourced

The answer

The AI build-out is mostly paper megawatts. Across 31 named projects in our coverage, 24 carry a stated capacity; by our arithmetic they add up to about 59.0 GW as stated (mixing operating, contracted and planned capacity, so the figure is an upper bound on what exists). Of those, the sources describe only 3 as operational or in production (about 0.5 GW, our arithmetic). The rest are planned, contracted, optioned or in talks. The biggest named sites are Meta's Hyperion (toward 5 GW) and Prometheus (~1 GW), SpaceX/xAI's Memphis footprint (over 2 GW as reported) and Microsoft's Indiana campus (~1.1 GW of added supply). Power is the binding constraint: Nscale's Essex site has no grid connection expected until the 2030s, and utilities are now building generation for single customers (Black Hills 564 MW for Google; NiSource ~1,100 MW for Microsoft).

Named AI Data-Centre Projects by Stated Capacity (MW, as sourced)

The tracker: first 5 of 31 rows

Project	Capacity	Status	Money	Source
Meta: Prometheus, New Albany, Ohio	roughly 1 gigawatt (2026) company claims	first phase complete (per Meta); size of the first phase not stated	—	Engineering at Meta
Meta: Hyperion, Richland Parish, Louisiana	scaling toward 5 gigawatts company claims	planned (per Meta)	—	Engineering at Meta
Amazon: AWS, 16 US states	44 gigawatts reported	under development (per source)	—	Business Insider, citing Cleanview Research
SpaceX/xAI: Memphis-area sites (Colossus)	compute footprint exceeding 2 GW (as reported) reported	two built, two under construction (per source)	—	Melvin (@MelvinInvests)
Tesla: Cortex 1 and Cortex 2, Texas	more than 105 MW + more than 120 MW; plans 400 MW by year end company claims	both in production (per Tesla)	—	Tesla Q3 2026 Update