Corpus written 2026-09-06
Global Compute Infrastructure Map
Facilities drawn where a source gives coordinates; the rest are on record but unplaced.
Facilities drawn where a source gives coordinates; the rest are on record but unplaced.
Plate I The map, live interactive · screen only
Plate I is interactive and does not print. Plate II below prints the same corpus in an equal-area projection, with the records Plate I cannot show. The live plate is at scrutica.com/map.
Two per cent of the sites on this map report a capacity figure
3,639 sites, one mark each. 2,961 name a site organisation or hardware owner, 72 state a capacity and 20 state an accelerator count.
Available information
Filter facilities by the information recorded about them.
Every record that has coordinates.
A site organisation or AI-hardware owner is named. These are separate fields; a hardware owner need not operate the facility.
A power-capacity value is recorded in megawatts.
An accelerator count is recorded.
Coordinates are classified as site-level in the record.
No site organisation, hardware owner, power capacity or accelerator count is recorded.
All 3,639 geolocated records selected. Then filter by facility class, status or capacity.
The first three categories account for 4,226 public records. Licence-withheld records sit outside that total.
Each dot represents one facility record, with no grouping or capacity weighting. All dots have the same size. The Equal Earth projection preserves relative areas across latitudes; grid lines at 30° intervals replace geographic boundaries. Records without coordinates, records excluded by map filters and licence-withheld records are listed separately beneath the map.
Of 4,226 public facility records, 586 have no coordinates. Map filters exclude 1 more record: duplicate pins on identified campuses, and OpenStreetMap records above 70° north with no site organisation, power figure, accelerator count or description. In total, 587 public records are absent from the map.
A further 316 records are withheld under source-licensing restrictions. They are excluded from the public total of 4,226.
The map contains 3,639 facility records. A country’s share of these records is not its share of computing capacity.
Sparse coverage may reflect gaps in the sources as well as fewer facilities. The number of map points alone cannot establish how much compute a country has.
General registries account for 90% of mapped records, from 4 sources. Compute-focused catalogues account for 80 of 3,639 mapped records. The source comparison below gives each source’s record count and geographic coverage.
Plate II The reach of the sources volume against reach
| Source | Records placed0–1,492 | Countries reached0–93 |
|---|---|---|
| PeeringDBGeneral registry | Records placed:1,492 | Countries reached:76 |
| OpenStreetMap — data-centre extractGeneral registry | Records placed:1,170 | Countries reached:93 |
| IM3 (PNNL) — US data-centre bulk corpusGeneral registry | Records placed:389 | Countries reached:1 |
| OpenStreetMapGeneral registry | Records placed:231 | Countries reached:33 |
| Epoch AI — frontier data centresCompute catalogue | Records placed:63 | Countries reached:4 |
| Microsoft Azure region docsDisclosures and reporting | Records placed:56 | Countries reached:32 |
| Oracle Cloud region docsDisclosures and reporting | Records placed:45 | Countries reached:27 |
| Google Cloud region docsDisclosures and reporting | Records placed:43 | Countries reached:29 |
| AWS region docsDisclosures and reporting | Records placed:34 | Countries reached:27 |
| Company press releasesDisclosures and reporting | Records placed:25 | Countries reached:11 |
| OVHcloud — Universal Registration Document (2024)Disclosures and reporting | Records placed:20 | Countries reached:8 |
| Source | Records placed0–1,492 | Countries reached0–93 |
|---|---|---|
| Equinix — SEC 10-K (2024)Disclosures and reporting | Records placed:15 | Countries reached:8 |
| EuroHPC Joint UndertakingCompute catalogue | Records placed:12 | Countries reached:12 |
| GDS Holdings — SEC 20-F (2024)Disclosures and reporting | Records placed:10 | Countries reached:1 |
| DayOne — corporate siteDisclosures and reporting | Records placed:9 | Countries reached:3 |
| Iron Mountain — SEC 10-K (2024)Disclosures and reporting | Records placed:8 | Countries reached:3 |
| Epoch AI — GPU clustersCompute catalogue | Records placed:5 | Countries reached:5 |
| GDS Holdings — Q4 2024 resultsDisclosures and reporting | Records placed:4 | Countries reached:2 |
| Scaleway — corporate siteDisclosures and reporting | Records placed:3 | Countries reached:1 |
| Press reportingDisclosures and reporting | Records placed:2 | Countries reached:1 |
| Scrutica curated reference recordsUnclassified | Records placed:2 | Countries reached:2 |
| European Commission decisionDisclosures and reporting | Records placed:1 | Countries reached:1 |
| Mistral Compute — announcementDisclosures and reporting | Records placed:1 | Countries reached:1 |
General registry General-purpose listings and maps, including data-centre extracts and grid-connection records.
Compute catalogue Catalogues of frontier data centres, accelerator clusters and publicly funded HPC sites.
Disclosures and reporting Company disclosures, government records and press reports.
The countries holding the most records the map cannot draw, ordered by how many, and every country that places nothing, regardless of volume. The rest are named in full beneath the table.
| Country | Placed | Held | Unplaced |
|---|---|---|---|
| United States | 1,051 | 1,277 | 226 |
| Japan | 67 | 111 | 44 |
| No country on the record | 85 | 112 | 27 |
| India | 106 | 132 | 26 |
| France | 280 | 301 | 21 |
| China | 35 | 56 | 21 |
| South Korea | 19 | 36 | 17 |
| United Kingdom | 181 | 197 | 16 |
| Brazil | 93 | 109 | 16 |
| Germany | 198 | 211 | 13 |
| Saudi Arabia | 6 | 19 | 13 |
| Russia | 50 | 61 | 11 |
| Italy | 68 | 78 | 10 |
| Thailand | 15 | 24 | 9 |
| Poland | 39 | 47 | 8 |
| Sweden | 30 | 38 | 8 |
| United Arab Emirates | 10 | 18 | 8 |
| Singapore | 53 | 60 | 7 |
| Netherlands | 132 | 138 | 6 |
| Canada | 98 | 104 | 6 |
| Malaysia | 16 | 22 | 6 |
| Taiwan | 7 | 13 | 6 |
| Djibouti | 0 | 1 | 1 |
| Guatemala | 0 | 1 | 1 |
| Oman | 0 | 1 | 1 |
24 further countries hold 59 unplaced records between them, none more than 5 each: Australia, Finland, Indonesia, Switzerland, New Zealand, Norway, Ukraine, Argentina, Turkey, Israel, Mexico, Hong Kong, Iceland, Philippines, Vietnam, Tanzania, Spain, Ireland, Denmark, Greece, Kenya, Ghana, Qatar, Slovenia. 3 countries hold records and place none: Djibouti, Guatemala, Oman. Each is in the table above whatever it holds.
| Facility | Country | Operator | Type | Status | Power (MW) | Accelerators |
|---|---|---|---|---|---|---|
| Colossus 2 | United States | — | AI Training | Operational | 946 | 530,000 |
| Meta | United States | Meta Platforms, Inc. | Hyperscale DC | Operational | — | 202,224 |
| Microsoft Goodyear | United States | — | AI Training | Operational | 202 | 100,000 |
| ORNL Frontier (Exascale HPC) | United States | — | HPC Center | Operational | 22.7 | 37,888 |
| JUPITER (EuroHPC) | Germany | EuroHPC JU | HPC Center | Operational | — | 23,752 |
| Jupiter, Jülich | Germany | — | AI Training | Operational | 18 | 23,536 |
| Leonardo (EuroHPC) | Italy | EuroHPC JU | HPC Center | Operational | — | 13,824 |
| Eclairion Bruyères-le-Châtel (Mistral Compute) | France | Eclairion | AI Training | Operational | 44 | 13,800 |
| LUMI (EuroHPC) | Finland | EuroHPC JU | HPC Center | Operational | — | 11,912 |
| LUMI Supercomputer (CSC Finland) | Finland | — | HPC Center | Operational | 7.1 | 10,240 |
| MareNostrum 5 (EuroHPC) | Spain | EuroHPC JU | HPC Center | Operational | — | 4,480 |
| Microsoft | United States | Microsoft Corporation | Hyperscale DC | Operational | — | 4,000 |
| Cineca Marconi-100 | Italy | — | AI Training | Operational | 2.644 | 3,952 |
| GENCI Adastra | France | — | AI Training | Operational | 1.609 | 1,352 |
| Hut 8 | Canada | Hut 8 | Hyperscale DC | Operational | — | 1,000 |
| EuroHPC MeluXina | Luxembourg | — | AI Training | Operational | 0.687 | 800 |
| MeluXina (EuroHPC) | Luxembourg | EuroHPC JU | HPC Center | Operational | — | 800 |
| EuroHPC Karolina | Czech Republic | — | AI Training | Operational | 0.548 | 576 |
| Karolina (EuroHPC) | Czech Republic | EuroHPC JU | HPC Center | Operational | — | 576 |
| Vega (EuroHPC) | Slovenia | EuroHPC JU | HPC Center | Operational | — | 240 |
| Stargate UAE (OpenAI/G42/Oracle) | United Arab Emirates | Oracle Corporation | AI Training | Under construction | 5,000 | — |
| AWS Project Rainier (New Carlisle, IN) | United States | Amazon.com, Inc. | AI Training | Operational | 2,200 | — |
| Meta Richland Parish (Hyperion) | United States | Meta Platforms, Inc. | AI Training | Under construction | 2,000 | — |
| Stargate Abilene (OpenAI/Oracle/SoftBank) | United States | Oracle Corporation | AI Training | Operational | 1,200 | — |
| Microsoft Fairwater Atlanta | United States | — | AI Training | Operational | 636 | — |
| Meta Prometheus | United States | — | AI Training | Operational | 631 | — |
| TeraWulf Lake Mariner Campus | United States | — | AI Training | Operational | 510 | — |
| Khazna–Eni Ferrera Erbognone AI Campus | Italy | — | Hyperscale DC | Announced | 500 | — |
| Microsoft Fairwater Wisconsin | United States | — | AI Training | Operational | 369 | — |
| Colossus 1 | United States | — | AI Training | Operational | 340 | — |
| Google New Albany | United States | — | AI Training | Operational | 339 | — |
| Google Columbus | United States | — | AI Training | Operational | 303 | — |
| CoreWeave Lancaster PA | United States | CoreWeave | AI Training | Announced | 300 | — |
| Amazon Madison Mega Site | United States | — | AI Training | Operational | 284 | — |
| CoreWeave Denton TX | United States | — | AI Training | Operational | 262 | — |
| CoreWeave Kenilworth NJ (NEST Campus) | United States | CoreWeave | AI Training | Under construction | 250 | — |
| Huawei Horinger | China | — | AI Training | Operational | 241.8 | — |
| DayOne Nusajaya Tech Park | Malaysia | DayOne Data Centers Limited | Colocation | Operational | 240 | — |
| QTS Richmond 1 | United States | — | AI Training | Operational | 238 | — |
| Google Council Bluffs (East) | United States | — | AI Training | Operational | 237 | — |
| Google Omaha | United States | — | AI Training | Operational | 237 | — |
| Google Papillion | United States | — | AI Training | Operational | 237 | — |
| Amazon Ridgeland | United States | — | AI Training | Operational | 228 | — |
| VNET Bayin Ulanqab | China | — | AI Training | Operational | 221 | — |
| Microsoft Project Osmium | United States | — | AI Training | Operational | 190 | — |
| QTS Richmond 2 | United States | — | AI Training | Operational | 180 | — |
| Meta Jeffersonville | United States | — | AI Training | Operational | 178 | — |
| Meta Rosemount | United States | — | AI Training | Operational | 178 | — |
| Alibaba Zhangbei | China | — | AI Training | Operational | 169 | — |
| Google Storey County | United States | — | AI Training | Operational | 161 | — |
| Google The Dalles | United States | — | AI Training | Operational | 154 | — |
| Meta Montgomery | United States | — | AI Training | Operational | 153 | — |
| Meta Kuna | United States | — | AI Training | Operational | 152 | — |
| Meta Temple | United States | — | AI Training | Operational | 152 | — |
| QTS Richmond 3 | United States | — | AI Training | Operational | 144 | — |
| Google Lincoln | United States | — | AI Training | Operational | 141 | — |
| Google Lancaster | United States | — | AI Training | Operational | 137 | — |
| Coreweave Helios | United States | — | AI Training | Operational | 132 | — |
| Google Midlothian | United States | — | AI Training | Operational | 103 | — |
| Google Mesa | United States | — | AI Training | Operational | 90 | — |
| Google Waltham Cross | United Kingdom | — | AI Training | Operational | 88 | — |
| Meta Gallatin | United States | — | AI Training | Operational | 87 | — |
| Meta Los Lunas | United States | — | AI Training | Operational | 87 | — |
| STACK Infrastructure NVA02 | United States | — | AI Training | Operational | 85 | — |
| CoreWeave Chester VA | United States | — | AI Training | Operational | 82 | — |
| Google Arcola | United States | — | AI Training | Operational | 77 | — |
| Google Red Oak | United States | — | AI Training | Operational | 77 | — |
| Microsoft SAT40 | United States | — | AI Training | Operational | 75 | — |
| CoreWeave Ellendale ND | United States | — | AI Training | Operational | 68 | — |
| CoreWeave Marble NC | United States | — | AI Training | Operational | 65 | — |
| Stream Phoenix | United States | — | AI Training | Operational | 64 | — |
| Google Pryor (North) | United States | — | AI Training | Operational | 53 | — |
| Microsoft SAT14 | United States | — | AI Training | Operational | 47 | — |
| Start Campus Sines Data Campus | Portugal | — | AI Training | Operational | 33 | — |
| Vantage TX1 | United States | — | AI Training | Operational | 32 | — |
| CoreWeave Dalton 1 & 2 | United States | — | AI Training | Operational | 28 | — |
| DayOne SG1 Singapore | Singapore | DayOne Data Centers Limited | Colocation | Under construction | 20 | — |
| Equinix SG3 | Singapore | Equinix, Inc. | Colocation | Operational | 20 | — |
| Scaleway DC5 (PAR2) | France | Iliad Sa | Hyperscale DC | Operational | 20 | — |
| Equinix LD5 (London) | United Kingdom | Equinix, Inc. | Colocation | Operational | 12 | — |
| Scaleway DC3 (PAR1) | France | Iliad Sa | Colocation | Operational | 6.9 | — |
| Scaleway DC2 (PAR1) | France | Iliad Sa | Colocation | Operational | 3.8 | — |
| ASML Veldhoven (EUV Manufacturing HQ) | Netherlands | ASML | Packaging | Operational | — | — |
| AWS US-East-1 (Northern Virginia) | United States | Amazon.com, Inc. | Hyperscale DC | Operational | — | — |
| AWS Wavelength Zone (Chicago) | United States | Amazon.com, Inc. | Edge | Operational | — | — |
| AWS Africa (Cape Town) | South Africa | Amazon.com, Inc. | Hyperscale DC | Operational | — | — |
| AWS Asia Pacific (Hong Kong) | Hong Kong | Amazon.com, Inc. | Hyperscale DC | Operational | — | — |
| AWS Asia Pacific (Taipei) | Taiwan | Amazon.com, Inc. | Hyperscale DC | Operational | — | — |
| AWS Asia Pacific (Tokyo) | Japan | Amazon.com, Inc. | Hyperscale DC | Operational | — | — |
| AWS Asia Pacific (Seoul) | South Korea | Amazon.com, Inc. | Hyperscale DC | Operational | — | — |
| AWS Asia Pacific (Osaka) | Japan | Amazon.com, Inc. | Hyperscale DC | Operational | — | — |
| AWS Asia Pacific (Mumbai) | India | Amazon.com, Inc. | Hyperscale DC | Operational | — | — |
| AWS Asia Pacific (Hyderabad) | India | Amazon.com, Inc. | Hyperscale DC | Operational | — | — |
| AWS Asia Pacific (Singapore) | Singapore | Amazon.com, Inc. | Hyperscale DC | Operational | — | — |
| AWS Asia Pacific (Sydney) | Australia | Amazon.com, Inc. | Hyperscale DC | Operational | — | — |
| AWS Asia Pacific (Jakarta) | Indonesia | Amazon.com, Inc. | Hyperscale DC | Operational | — | — |
| AWS Asia Pacific (Melbourne) | Australia | Amazon.com, Inc. | Hyperscale DC | Operational | — | — |
| AWS Asia Pacific (Malaysia) | Malaysia | Amazon.com, Inc. | Hyperscale DC | Operational | — | — |
| AWS Asia Pacific (New Zealand) | New Zealand | Amazon.com, Inc. | Hyperscale DC | Operational | — | — |
| AWS Asia Pacific (Thailand) | Thailand | Amazon.com, Inc. | Hyperscale DC | Operational | — | — |
Showing 100 of 3,639. View the full keyboard-accessible table →