Gives dated, site-level capacity figures for the largest announced AI compute build-out, useful for judging future model availability and cost.
Key takeaways · AI-distilled
Epoch counts seven Stargate sites with over 9 GW planned, comparable to New York City's peak power demand and enough for about 20 million H100s, which it says equals all AI compute in the world at the end of 2025.
Only Abilene, Texas is running so far, at an estimated 0.3 GW (about 250,000 H100-equivalents) toward 1.2 GW by Q4 2026; the other six sites show 0 GW current capacity, five are projected for Q4 2028, and Lordstown's completion date is unknown.
To avoid grid-connection queues, at least three of the seven sites will use on-site natural gas plants; at least six will use closed-loop liquid cooling that does not evaporate water.
Epoch judges those design choices likely save the project time but raise facility costs. Per 2025 announcements, SoftBank will own hardware at the Milam County and Ohio sites and Oracle at the rest, all serving OpenAI.
Terms in this piece · Glossary
grounding — Tying a model's answers to checkable sources — retrieved documents, live data, tool results — instead of letting it answer from memory alone.
