GPU 集群地图：每一台已知的 AI 超级计算机，标注在地图上

This map plots 459 GPU clusters with a known position, holding about 1.5M H100e between them across roughly 2.6M AI chips. The ones with a published power figure draw 2.58 GW. A cluster is the machine itself — a coherent pool of accelerators on one interconnect — rather than the building around it, which is why the same GPUs also appear on the data center map.

The United States holds 81.4% of all located cluster compute across 119 clusters, ahead of China (109.4K H100e across 188 clusters). 36 countries have at least one located cluster.

The largest single cluster is xAI Colossus Memphis Phase 3, operated by xAI, at 275.8K H100e on 230,000 chips. Ownership is as concentrated as geography: xAI, Meta AI, Oracle, Microsoft + OpenAI and US Department of Energy together hold 66.7% of the located total. Circle area is proportional to compute rather than radius, so a cluster ten times larger looks ten times larger by area instead of a hundred. Coordinates are published by Epoch AI and used as given; the compute and chip counts are modelled estimates reconstructed from chip orders, reported spending and public statements, not disclosures.