Largest GPU clusters in the world
87 named GPU clusters: 61 running or partly running, 22 announced or under construction. Largest installed count on record: Colossus 1 with 220,000 GPUs, then Aurora (63,744) and El Capitan (46,080). Biggest build on paper: Meta Prometheus, targeting about 1,000,000 GPUs. Every number links to its source.
Last verified
| Colossus 1 | 220,0003 | — | xAI (SpaceXAI) | NVIDIA H100, H200 and GB200 | operational | Memphis, Tennessee, United States | 3003 | 2024 | 2026-11-11 |
|---|---|---|---|---|---|---|---|---|---|
| Aurora | 63,7441 | — | Argonne Leadership Computing Facility | Intel Data Center GPU Max Series | operational | Lemont, Illinois, United States | — | 2025-01 | 2026-11-11 |
| El Capitan | 46,0801 | — | Lawrence Livermore National Laboratory | AMD Instinct MI300A APU | operational | Livermore, California, United States | — | — | 2026-11-11 |
| Frontier | 39,4241 | — | Oak Ridge Leadership Computing Facility | AMD Instinct MI250X | operational | Oak Ridge, Tennessee, United States | — | — | 2026-11-11 |
| Meta GenAI InfiniBand cluster | 24,5761 | — | Meta | NVIDIA H100 | operational | — | — | — | 2026-11-11 |
| Meta GenAI RoCE cluster | 24,5761 | — | Meta | NVIDIA H100 | operational | — | — | — | 2026-11-11 |
| Project Ceiba | 20,7361 | — | AWS | NVIDIA GB200 (Blackwell) | operational | — | — | — | 2026-11-11 |
| Meta AI Research SuperCluster (RSC) | 16,0001 | — | Meta | NVIDIA A100 | operational | — | — | — | 2026-11-11 |
| Microsoft Azure Eagle | 14,4001 | — | Microsoft | NVIDIA H100 | operational | — | — | 2023-11 | 2026-11-11 |
| Eni HPC7 | 13,9201 | — | Eni | AMD Instinct MI300A APU | operational | Ferrera Erbognone, Italy | — | 2026-06 | 2026-11-11 |
| Eni HPC6 | 13,8881 | — | Eni | AMD Instinct MI250X | operational | Ferrera Erbognone, Italy | — | 2024-11 | 2026-11-11 |
| Leonardo Booster | 13,8241 | — | CINECA | NVIDIA A100-64 custom Ampere, 64 GiB | operational | Bologna, Italy | — | 2023-07 | 2026-11-11 |
| ByteDance MegaScale Ampere cluster | 12,2881 | — | ByteDance | NVIDIA Ampere | operational | — | — | — | 2026-11-11 |
| LUMI-G | 11,9122 | — | LUMI consortium | AMD Instinct MI250X | operational | Kajaani, Finland | — | — | 2026-11-11 |
| Alps Grace Hopper partition | 10,7521 | — | CSCS | NVIDIA GH200 (H100 GPU) | operational | Lugano and Lausanne, Switzerland | — | 2024-09 | 2026-11-11 |
| NVIDIA Eos (10,752-GPU system) | 10,7521 | — | NVIDIA | NVIDIA H100 | operational | — | — | — | 2026-11-11 |
| CoreWeave Barcelona H200 supercomputer | 10,2241 | — | CoreWeave | NVIDIA H200 | announced | Barcelona, Spain | — | — | 2026-11-11 |
| Yotta Shakti Cloud NM1 | 9,2161 | — | Yotta Data Services | NVIDIA H100 and L40S | operational | Navi Mumbai, India | — | — | 2026-11-11 |
| TensorWave MI325X training cluster | 8,1921 | — | TensorWave | AMD Instinct MI325X | operational | North America (city not published) | — | 2025-07 | 2026-11-11 |
| Perlmutter GPU partition | 7,1681 | — | NERSC | NVIDIA A100 40GB and 80GB | operational | Berkeley, California, United States | — | — | 2026-11-11 |
| ABCI 3.0 | 6,1282 | — | AIST Solutions (ABCI) | NVIDIA H200 SXM5, 141 GB HBM3e | operational | Kashiwa, Japan | — | — | 2026-11-11 |
| Isambard-AI | 5,4481 | — | University of Bristol | NVIDIA GH200 Grace Hopper | operational | Bristol, United Kingdom | — | 2025-07 | 2026-11-11 |
| NVIDIA Eos (4,608-GPU system) | 4,6081 | — | NVIDIA | NVIDIA H100 | operational | — | — | — | 2026-11-11 |
| Tuolumne | 4,6081 | — | Lawrence Livermore National Laboratory | AMD Instinct MI300A APU | operational | Livermore, California, United States | — | 2024 | 2026-11-11 |
| MareNostrum 5 ACC | 4,4801 | — | Barcelona Supercomputing Center | NVIDIA Hopper, 64 GB HBM | operational | Barcelona, Spain | — | — | 2026-11-11 |
| Horizon (TACC) | 4,0001 | — | Texas Advanced Computing Center (UT Austin, NSF) | NVIDIA GB200 (NVL4 nodes) | partly operational | Round Rock, Texas, United States | — | — | 2026-11-11 |
| Nebius Modiin Blackwell cluster | 4,0001 | — | Nebius | NVIDIA HGX B200 | operational | Modiin, Israel | — | 2025-10 | 2026-11-11 |
| JUWELS Booster | 3,7442 | — | Jülich Supercomputing Centre | NVIDIA A100 40 GB | operational | Jülich, Germany | — | — | 2026-11-11 |
| Shaheen III GPU partition | 2,8001 | — | KAUST | NVIDIA GH200 Grace Hopper | operational | Thuwal, Saudi Arabia | — | — | 2026-11-11 |
| Venado | 2,5601 | — | Los Alamos National Laboratory | NVIDIA GH200 Grace Hopper | operational | Los Alamos, New Mexico, United States | — | — | 2026-11-11 |
| NAVER Sejong | 2,2401 | — | NAVER | NVIDIA A100 | operational | Sejong, South Korea | — | — | 2026-11-11 |
| Polaris | 2,2401 | — | Argonne National Laboratory | NVIDIA A100 | operational | Lemont, Illinois, United States | — | — | 2026-11-11 |
| DeepSeek-V3 H800 training cluster | 2,0481 | — | DeepSeek | NVIDIA H800 | — | — | — | — | 2026-11-11 |
| Sakura Ishikari H100 deployment | 2,0161 | — | Sakura Internet | NVIDIA H100 | operational | Ishikari, Japan | — | 2024-08 | 2026-11-11 |
| Yandex Chervonenkis | 1,5921 | — | Yandex | NVIDIA A100 80GB | — | Sasovo, Ryazan Oblast, Russia | — | — | 2026-11-11 |
| Gefion | 1,5281 | — | Danish Centre for AI Innovation (DCAI) | NVIDIA H100 | operational | — | — | 2024-10 | 2026-11-11 |
| Jean Zay H100 extension | 1,4561 | — | IDRIS / CNRS (GENCI) | NVIDIA H100 80 GB | operational | Saclay, France | — | — | 2026-11-11 |
| Adastra GPU partition | 1,3521 | — | CINES | AMD Instinct MI250X | operational | Montpellier, France | — | — | 2026-11-11 |
| MareNostrum 5 AI training partition | 1,3441 | — | Barcelona Supercomputing Center | NVIDIA GB200 | operational | Barcelona, Spain | — | — | 2026-11-11 |
| Yandex Lyapunov | 1,0961 | — | Yandex | NVIDIA A100 40GB | — | Sasovo, Ryazan Oblast, Russia | — | — | 2026-11-11 |
| Yandex Galushkin | 1,0881 | — | Yandex | NVIDIA A100 80GB | — | Vladimir, Vladimir Oblast, Russia | — | — | 2026-11-11 |
| Chan Zuckerberg Initiative biological AI cluster | 1,0241 | — | Chan Zuckerberg Initiative | NVIDIA H100 | operational | — | — | — | 2026-11-11 |
| Scaleway Nabuchodonosor 2023 | 1,0161 | — | Scaleway | NVIDIA H100 | operational | Saint-Ouen-l'Aumône (Scaleway DC5, Paris region), France | — | 2023-10 | 2026-11-11 |
| TSUBAME4.0 | 9601 | — | Institute of Science Tokyo | NVIDIA H100 SXM5 | operational | Tokyo, Japan | — | — | 2026-11-11 |
| MeluXina accelerator module | 8002 | — | LuxProvide | NVIDIA A100 40GB | operational | Bissen (LuxConnect DC2), Luxembourg | — | — | 2026-11-11 |
| Setonix AMD GPU partition | 7681 | — | Pawsey Supercomputing Centre | AMD Instinct MI250X | operational | Kensington, Australia | — | — | 2026-11-11 |
| Berzelius Ampere | 7521 | — | NAISS | NVIDIA A100, 40GB or 80GB | operational | Linköping, Sweden | — | — | 2026-11-11 |
| MareNostrum 5 AI inference partition | 5761 | — | Barcelona Supercomputing Center | NVIDIA GB300 | operational | Barcelona, Spain | — | — | 2026-11-11 |
| Recursion BioHive-2 | 5041 | — | Recursion | NVIDIA H100 | operational | Salt Lake City, Utah, United States | — | 2024-05 | 2026-11-11 |
| Alice Recoque | — | — | Jules Verne consortium (GENCI, CEA, SURF, GRNET) for EuroHPC JU | AMD Instinct MI430X, 432 GB HBM4 | under construction | TGCC, CEA DAM Île-de-France, Bruyères-le-Châtel, France | — | — | 2026-11-11 |
| Argonne Equinox | — | 10,0001 | Argonne National Laboratory (DOE) | NVIDIA Blackwell | announced | Lemont, Illinois, United States | — | — | 2026-11-11 |
| Argonne Solstice | — | 100,0001 | Argonne National Laboratory (DOE) | NVIDIA Blackwell | announced | Lemont, Illinois, United States | — | — | 2026-11-11 |
| Colossus 2 | — | 440,0003 | xAI (SpaceXAI) | NVIDIA GB200 and GB300 | operational | Memphis, Tennessee, United States | 9463 | 2025-10 | 2026-11-11 |
| DeepL Mercury | — | — | DeepL SE | NVIDIA H100 | operational | Falun (EcoDataCenter 1), Dalarna, Sweden | — | 2023 | 2026-11-11 |
| Deutsche Telekom Industrial AI Cloud (Munich) | — | 10,0001 | Deutsche Telekom | NVIDIA B200 (DGX B200) and RTX PRO Blackwell | operational | Tucherpark, Munich, Bavaria, Germany | — | 2026 | 2026-11-11 |
| Doudna (NERSC-10) | — | — | NERSC, Lawrence Berkeley National Laboratory (DOE) | NVIDIA Vera Rubin | announced | Berkeley, California, United States | — | — | 2026-11-11 |
| ExxonMobil Discovery 6 | — | — | ExxonMobil | NVIDIA GH200 Grace Hopper | operational | — | — | 2025 | 2026-11-11 |
| Foxconn Taiwan AI Factory (Big Innovation Company) | — | 10,0001 | Big Innovation Company (Foxconn) | NVIDIA Blackwell Ultra (GB300 NVL72) | announced | Southern Taiwan | — | — | 2026-11-11 |
| GMI Cloud Taiwan AI Factory | — | 7,0001 | GMI Cloud | NVIDIA Blackwell | announced | Taoyuan, Taiwan | — | — | 2026-11-11 |
| HUMAIN GB300 AI factory (phase 1) | — | 18,0001 | HUMAIN (Public Investment Fund) | NVIDIA GB300 | announced | Saudi Arabia | — | — | 2026-11-11 |
| IBM Blue Vela | — | — | IBM | NVIDIA H100 | operational | — | — | — | 2026-11-11 |
| IREN Horizon 1 | — | — | IREN | NVIDIA GB300 NVL72 | operational | Childress, Texas, United States | 501 | 2026-08 | 2026-11-11 |
| JUPITER Booster | — | — | Jülich Supercomputing Centre (EuroHPC JU) | NVIDIA GH200 | operational | Jülich, Germany | — | 2025 | 2026-11-11 |
| KISTI National Supercomputer No. 6 | — | 8,4961 | KISTI (Korea Institute of Science and Technology Information) | NVIDIA GH200 | under construction | — | — | — | 2026-11-11 |
| Lambda Kansas City AI Factory | — | — | Lambda | NVIDIA Blackwell Ultra | announced | Kansas City, Missouri, United States | — | — | 2026-11-11 |
| Lux (ORNL) | — | 4,0001 | Oak Ridge National Laboratory (DOE) | AMD Instinct MI355X, 288 GB HBM3E | under construction | Oak Ridge, Tennessee, United States | — | — | 2026-11-11 |
| Magic G4 on Google Cloud | — | — | Google Cloud (for Magic) | NVIDIA H100 | announced | — | — | — | 2026-11-11 |
| Magic G5 on Google Cloud | — | — | Google Cloud (for Magic) | NVIDIA Grace Blackwell | announced | — | — | — | 2026-11-11 |
| Meta Hyperion | — | — | Meta | — | under construction | Richland Parish, Louisiana, United States | — | 2028 | 2026-11-11 |
| Meta Prometheus | — | 1,000,0001 | Meta | NVIDIA GB200 | partly operational | New Albany, Ohio, United States | 4712 | 2025-10 | 2026-11-11 |
| Microsoft Fairwater Atlanta | — | — | Microsoft | NVIDIA GB200 NVL72 | operational | Atlanta area, Georgia, United States | 6362 | 2025-10 | 2026-11-11 |
| Microsoft Fairwater Wisconsin | — | — | Microsoft | NVIDIA GPUs | partly operational | Mount Pleasant, Wisconsin, United States | 3692 | 2026-04 | 2026-11-11 |
| Mistral AI Bruyères-le-Châtel cluster | — | 13,8001 | Mistral AI | NVIDIA GB300 | announced | Bruyères-le-Châtel (Eclairion campus), Île-de-France, France | — | — | 2026-11-11 |
| Nebius Kansas City GPU cluster | — | — | Nebius | NVIDIA Hopper | announced | Kansas City, Missouri, United States | — | — | 2026-11-11 |
| Nebius Vineland GPU infrastructure | — | — | Nebius (tenant at DataOne) | — | partly operational | Vineland, New Jersey, United States | 503 | 2026 | 2026-11-11 |
| Nscale Loughton AI campus | — | 23,0401 | Nscale | NVIDIA GB300 | announced | Loughton, Essex, United Kingdom | — | — | 2026-11-11 |
| Nscale Narvik (Stargate Norway) | — | 52,0002 | Nscale | NVIDIA GB300 | under construction | Narvik, Norway | — | 2027 | 2026-11-11 |
| Nscale Sines AI campus deployment | — | — | Nscale | NVIDIA Blackwell Ultra | operational | Sines, Portugal | 332 | 2026 | 2026-11-11 |
| Nscale Texas Microsoft cluster | — | 104,0001 | Nscale | NVIDIA GB300 | announced | Barstow (Cedarvale site), Ward County, Texas, United States | — | — | 2026-11-11 |
| NVIDIA Israel-1 | — | 2,0481 | NVIDIA | NVIDIA H100 80GB | partly operational | Israel | — | 2023-11 | 2026-11-11 |
| SoftBank Blackwell DGX SuperPOD | — | — | SoftBank Corp. | NVIDIA B200 | operational | — | — | 2025-07 | 2026-11-11 |
| Stargate Abilene (Crusoe / Oracle) | — | — | Oracle (built by Crusoe, for OpenAI) | NVIDIA GB200 | partly operational | Abilene, Texas, United States | 6183 | 2025-09 | 2026-11-11 |
| Stargate UK (Nscale for OpenAI) | — | 8,0001 | Nscale (OpenAI offtake) | — | announced | Cobalt Park, North Tyneside, United Kingdom (one of several UK sites) | — | — | 2026-11-11 |
| Tesla Cortex 1 | — | — | Tesla | — | operational | Gigafactory Texas, Austin, Texas, United States | 1051 | — | 2026-11-11 |
| Tesla Cortex 2 | — | — | Tesla | — | operational | Gigafactory Texas, Austin, Texas, United States | 1201 | — | 2026-11-11 |
| Together AI–Hypertec Blackwell cluster | — | 36,0001 | Together AI and Hypertec Cloud | NVIDIA GB200 NVL72 | announced | — | — | — | 2026-11-11 |
| Yotta Greater Noida D2 Blackwell deployment | — | 30,0001 | Yotta | NVIDIA Blackwell B300 Ultra | announced | Greater Noida, India | — | — | 2026-11-11 |
Browse by kind
All listsHow this table is built
Full methodologyEvery value comes from a public source: the owner’s engineering post, user documentation, filing or press release, or a named report. The small number next to a figure opens that source. An empty cell means the owner has not published the figure. Installed GPUs and planned GPUs sit in separate columns, so an announcement never inflates a count.
Each row carries the date we last checked it, and the newest check was 11 Nov 2026. Spot a stale number? Email hello@gpuclusterlist.com with the source and we update the row.
Questions buyers ask
What is the largest GPU cluster in the world?
By installed GPUs that a company has put on record, xAI’s Colossus 1 in Memphis leads with over 220,000 NVIDIA GPUs on more than 300 MW (Anthropic, 6 May 2026). SpaceX’s S-1 describes 440,000 GB200 and GB300 GPUs across Colossus 2’s clusters and next expansion, and Microsoft says its Fairwater Wisconsin campus holds hundreds of thousands of NVIDIA GPUs. By plan, Meta’s Prometheus in New Albany, Ohio targets about 1 million GPUs, with hundreds of thousands running today.
Which supercomputer has the most GPUs?
Argonne’s Aurora has the most of any public supercomputer in this list, with 63,744 Intel Data Center GPU Max GPUs. Lawrence Livermore’s El Capitan follows with 46,080 AMD MI300A accelerators and Oak Ridge’s Frontier with 39,424 AMD MI250X packages. JUPITER at Jülich runs around 24,000 NVIDIA GH200 superchips.
How many GPUs does xAI’s Colossus have?
Colossus 1 has over 220,000 NVIDIA GPUs and more than 300 MW, all of which Anthropic agreed in May 2026 to use. Tom’s Hardware reports a mix of 150,000 H100, 50,000 H200 and 30,000 GB200 for it. For Colossus 2, SpaceX’s S-1 describes about 110,000 GB200 and 110,000 GB300 in its first two clusters plus at least 220,000 more GB300, 440,000 in all, which the table carries as planned. Press reports of 550,000 are unconfirmed, and Epoch AI estimates 946 MW of IT load in service.
How do you tell an announced cluster from one that is running?
Every row carries one of four statuses. Operational means the operator or a primary source says it is running. Partly operational means a phase is live and more is coming, as at Stargate Abilene (6 of 8 buildings delivered). Under construction and announced cover clusters that are contracted or built but not yet confirmed live. Planned GPU counts sit in their own column and never add to the installed count.
How are GPUs counted, and why do some clusters show no count?
The installed count is physical GPUs: nodes times GPUs per node when the operator publishes both. MI250X boards count once, not as two dies, and an MI300A APU counts as one GPU. H100-equivalents and design limits are never counted as GPUs. When an owner publishes no installed total, as with Fairwater, Tesla Cortex and DeepL Mercury, the cell stays empty and the note column carries what the source does say.