Entities
Registers of the facilities, companies and countries on record, with the ownership chains between them.
709 of 4,257 facility records have an operational measure
- Epoch AI Frontier Data Centers7064 measured91.4%
- Epoch AI GPU Clusters437396 measured90.6%
- PeeringDB facility registry1,56963 measured4.0%
- IM3 / PNNL data-center corpus3912 measured0.5%
- OpenStreetMap (tagged data centers)1,1731 measured0.1%
- OpenStreetMap2310 measured0.0%
- Microsoft Azure region documentation560 measured0.0%
- Oracle Cloud region documentation450 measured0.0%
- 43 smaller catalogues285183 measured64.2%
Coverage by catalogue. Each band is one catalogue: its length is that catalogue’s share of the largest one’s row count, and the inked run inside it is the rows with at least one operational measure.
Nothing here is schematic; both lengths are measured proportions of the counts printed beside them.
No share-of-megawatts is drawn: every megawatt-bearing row satisfies the measured predicate by construction, so a capacity share would be a tautology.
Counts cover the full individually-browsable set, 4,257 rows, read 5 September 2026.
The smallest catalogues are folded into one band whose run is hatched, because it averages catalogues of both kinds. Each named catalogue’s key opens the register below scoped to its rows.
4,257 facility records are individually browsable here, of which 709 have an operational measure: power, GPU or FLOP capacity, or a substantive description. The other 3,548 establish that a site exists and stop there. Which of the two a record is depends almost entirely on the catalogue it arrived in.
Two counts appear on this page and they are not the same quantity. The masthead stamps 4,573 tracked records: the full corpus across 110 countries, the figure used site-wide. The register and the plate count the 4,257 that are individually browsable; the 316 in between are rows a licensing term keeps from being served as their own record. They inform aggregates and they are not hidden. They cannot be opened, so a figure about what a reader can inspect must not count them. The same universe is plotted in space on the map, rolled up by jurisdiction on Countries, and the organisations that own and operate these sites are in the company register.
Epoch AI Frontier Data Centers runs 91.4% measured across 70 records; OpenStreetMap runs 0.0% across 231. No single catalogue falls in between; the hatched band does, but it is the fold of 43 smaller catalogues and reads mid-scale only because it averages the two kinds. The separation exists because the two kinds of catalogue do different jobs: one is a research dataset built by measuring sites, the other an infrastructure registry built by recording that they exist.
Shown whole, most of the 4,257 rows could not answer the question a reader arrived with. The positional records are not filler: a site that exists is a real fact, and for a governance reader tracing where capacity might appear it is often the fact that matters. But it is a different fact from a megawatt figure, and a register that prints them in one undifferentiated count makes a uniformity claim its records do not support.
The register
Showing the 709 records that have an operational measure. A further 3,548 positional records establish that a site exists and have no capacity figure. Show all 4,257.
| Facility | Country | Type | Status | Power (MW) |
|---|---|---|---|---|
| Imbue 10k ClusterEpoch AI GPU Clusters | US | AI training | Operational | 15 |
| Tesla 10k H100 ClusterEpoch AI GPU Clusters | US | AI training | Operational | 15 |
| XTX Markets ClusterEpoch AI GPU Clusters | — | AI training | Operational | 15 |
| TensorWave MI300X Cluster 1 Phase 2Epoch AI GPU Clusters | US | AI training | Announced | 15 |
| TensorWave MI300X Cluster 2Epoch AI GPU Clusters | US | AI training | Announced | 15 |
| Meta 2017 V100 ClusterEpoch AI GPU Clusters | US | AI training | Operational | 16 |
| Oak Ridge NL TitanEpoch AI GPU Clusters | US | AI training | Decommissioned | 16 |
| Sakura's B200s Phase 2Epoch AI GPU Clusters | JP | AI training | Announced | 16 |
| Jupiter, JülichEpoch AI GPU Clusters | DE | AI training | Operational | 18 |
| DayOne SG1 SingaporeDayOne corporate disclosures (May 2026) | SG | Colocation | Under construction | 20 |
| xAI Fulton GeorgiaEpoch AI GPU Clusters | US | AI training | Announced | 20 |
| Equinix SG3Company disclosure (press / IR) | SG | Colocation | Operational | 20 |
| Scaleway DC5 (PAR2)Scaleway corporate disclosures | FR | Hyperscale DC | Operational | 20 |
| Port Jeff -Mt. Sinai Energy StorageGridStatus | US | Hyperscale DC | Decommissioned | 20 |
| Gemini 1.0 Ultra training cluster AEpoch AI GPU Clusters | US | AI training | Operational | 20 |
| Microsoft Azure EagleEpoch AI GPU Clusters | US | AI training | Operational | 21 |
| Microsoft GPT-4 clusterEpoch AI GPU Clusters | US | AI training | Operational | 21 |
| ORNL Frontier (Exascale HPC)Scrutica internal fixture record | US | HPC center | Operational | 23 |
| Yotta Shakti Cloud D1Epoch AI GPU Clusters | IN | AI training | Announced | 23 |
| Yotta Shakti Cloud NM1 Phase 2Epoch AI GPU Clusters | IN | AI training | Announced | 23 |
| NexGen Cloud Hyperstack AQ Compute SupercomputerEpoch AI GPU Clusters | NO | AI training | Operational | 23 |
| Oracle OCI Supercluster H100sEpoch AI GPU Clusters | US | AI training | Operational | 23 |
| Google Oklahoma TPU v4 PodsEpoch AI GPU Clusters | US | AI training | Operational | 24 |
| Google TPUv5eEpoch AI GPU Clusters | — | AI training | Operational | 24 |
| Oracle OCI MI300xEpoch AI GPU Clusters | US | AI training | Operational | 25 |
| Sesterce PegasusEpoch AI GPU Clusters | — | AI training | Announced | 25 |
| Foxconn Big Innovation Cloud AI factoryEpoch AI GPU Clusters | TW | AI training | Announced | 26 |
| Oracle OCI Supercluster A100sEpoch AI GPU Clusters | US | AI training | Operational | 27 |
| CoreWeave Dalton 1 & 2Epoch AI Frontier Data Centers | US | AI training | Operational | 28 |
| Andreessen Horowitz OxygenEpoch AI GPU Clusters | US | AI training | Operational | 29 |
| AWS EC2 P5 UltraClustersEpoch AI GPU Clusters | US | AI training | Operational | 29 |
| G42 Microsoft 30 MW UAE Cluster AEpoch AI GPU Clusters | AE | AI training | Announced | 30 |
| G42 Microsoft 30 MW UAE Cluster BEpoch AI GPU Clusters | AE | AI training | Announced | 30 |
| NVIDIA Israel Blackwell SupercomputerEpoch AI GPU Clusters | IL | AI training | Announced | 30 |
| Yotta D1Yotta corporate disclosures | IN | Hyperscale DC | Operational | 30 |
| Inflection AI ClusterEpoch AI GPU Clusters | US | AI training | Announced | 31 |
| Vantage TX1Epoch AI Frontier Data Centers | US | AI training | Operational | 32 |
| Start Campus Sines Data CampusEpoch AI Frontier Data Centers | PT | AI training | Operational | 33 |
| Lawrence Livermore NL El Capitan Phase 2Epoch AI GPU Clusters | US | AI training | Operational | 35 |
| Meta GenAI 2024aEpoch AI GPU Clusters | US | AI training | Operational | 35 |
| Meta GenAI 2024bEpoch AI GPU Clusters | US | AI training | Operational | 35 |
| Google A3 VMsEpoch AI GPU Clusters | US | AI training | Announced | 36 |
| YTL AI JohorEpoch AI GPU Clusters | MY | AI training | Announced | 37 |
| Nebius Kansas City Phase 2Epoch AI GPU Clusters | US | AI training | Announced | 40 |
| Oak Ridge NL FrontierEpoch AI GPU Clusters | US | AI training | Operational | 40 |
| Eclairion Bruyères-le-Châtel (Mistral Compute)Mistral Compute press announcements | FR | AI training | Operational | 44 |
| NVIDIA CoreWeave Eos-DFW Rumored Phase 2Epoch AI GPU Clusters | US | AI training | Announced | 45 |
| Lambda Labs H100/H200Epoch AI GPU Clusters | US | AI training | Operational | 47 |
| Microsoft SAT14Epoch AI Frontier Data Centers | US | AI training | Operational | 47 |
| Project Ceiba Phase 2Epoch AI GPU Clusters | US | AI training | Announced | 50 |