Where the world’s AI compute is
534 located sites — every tracked AI data center and GPU cluster with a known position, sized by the compute it holds.
This map plots 534 AI compute sites with a known position: 75 data centers holding about 12.5M H100e, and 459 GPU clusters holding about 1.5M H100e. The data centers draw roughly 11.86 GW between them — the scale of a mid-sized country's electricity demand. The two figures should not be added: they are separate datasets at different granularity, and a cluster usually sits inside one of the data centers, so the same GPUs appear in both.
The concentration is extreme. United States holds 93.6% of all located data-center compute across 66 sites, ahead of Malaysia (281.7K H100e across 1 site). The five largest countries hold 99.7% of it, out of 38 countries with a located site. Counted by individual clusters rather than by capacity the picture shifts: China has 188 separately tracked clusters, second only to United States, on 2.4% of its data-center compute.
The single largest site is Colossus 2, operated by SpaceXAI, at 1.1M H100e and 946 MW. The largest individually tracked cluster is xAI Colossus Memphis Phase 3 (xAI), at 275.8K H100e. Circle area is proportional to compute rather than radius, so a site ten times larger looks ten times larger by area instead of a hundred. Every figure here is a modelled estimate reconstructed from permits, satellite imagery, chip orders and reported spending — not a disclosure.
Located compute by country
| Country | Data centers | DC compute | DC power | Clusters | Cluster compute |
|---|---|---|---|---|---|
| United States | 66 | 11.7M H100e | 10.80 GW | 119 | 1.3M H100e |
| Malaysia | 1 | 281.7K H100e | 240 MW | 1 | -- |
| China | 3 | 278.5K H100e | 632 MW | 188 | 109.4K H100e |
| United Kingdom | 1 | 123.8K H100e | 88 MW | 6 | 8K H100e |
| Indonesia | 1 | 84.4K H100e | 72 MW | -- | -- |
| Portugal | 1 | 31.8K H100e | 33 MW | -- | -- |
| Germany | -- | -- | -- | 11 | 26K H100e |
| Norway | -- | -- | -- | 3 | 20.5K H100e |
| Japan | -- | -- | -- | 25 | 18.7K H100e |
| France | -- | -- | -- | 14 | 18.4K H100e |
| Switzerland | -- | -- | -- | 4 | 17.2K H100e |
| Finland | -- | -- | -- | 4 | 11.8K H100e |
| Saudi Arabia | -- | -- | -- | 3 | 8.4K H100e |
| Italy | -- | -- | -- | 9 | 8.2K H100e |
| Sweden | -- | -- | -- | 4 | 4.9K H100e |
| Spain | -- | -- | -- | 1 | 4.5K H100e |
| India | -- | -- | -- | 3 | 4.4K H100e |
| Canada | -- | -- | -- | 5 | 3.1K H100e |
| Israel | -- | -- | -- | 2 | 3.1K H100e |
| Denmark | -- | -- | -- | 1 | 3K H100e |
| South Korea | -- | -- | -- | 6 | 2.4K H100e |
| Thailand | -- | -- | -- | 5 | 2.3K H100e |
| Russia | -- | -- | -- | 8 | 1.8K H100e |
| United Arab Emirates | 1 | -- | -- | 3 | 1.7K H100e |
| Taiwan | -- | -- | -- | 3 | 1.7K H100e |
| Brazil | -- | -- | -- | 9 | 1.3K H100e |
| Poland | -- | -- | -- | 4 | 1.2K H100e |
| Singapore | -- | -- | -- | 3 | 1.2K H100e |
| Netherlands | -- | -- | -- | 3 | 947 H100e |
| Australia | 1 | -- | -- | 3 | 637 H100e |
| Hong Kong | -- | -- | -- | 2 | 400 H100e |
| Luxembourg | -- | -- | -- | 1 | 252 H100e |
| Iceland | -- | -- | -- | 1 | 248 H100e |
| Czechia | -- | -- | -- | 1 | 182 H100e |
| Slovenia | -- | -- | -- | 1 | 75.67 H100e |
| Vietnam | -- | -- | -- | 1 | 50.45 H100e |
| Philippines | -- | -- | -- | 1 | -- |
| Mexico | -- | -- | -- | 1 | -- |
A dot on this map is not always the same kind of claim. Epoch AI publishes street addresses for data centers but no coordinates, so those are geocoded here from the address and every result records how far down it resolved. Sites whose address is a construction description rather than a postal address are placed from the town named in the record, and are never claimed as better than region-level.
- Resolved to the building
- 23 sites
- Resolved to the street
- 11 sites
- Resolved to the district
- 3 sites
- Resolved to the town
- 22 sites
- Resolved to the county or region
- 16 sites
GPU cluster coordinates come from Epoch's own export and are used as published.