GPU clusters
Our database of over 500 GPU clusters and supercomputers tracks large hardware facilities, including those used for AI training and inference.
Charts and analysis of these data →
gpu_clusters.csv
474 rows · 57 columns · Updated 2026-11-09
| Name | Status | Certainty | Single cluster? | Max OP/s (log) | H100 equivalents | Chip type (primary) | Chip quantity (primary) | Country | Owner | First Operational Date | Note | Sector | Power Capacity (MW) | Hardware Cost | Location | Users | Hardware note | Quote | First Operational Date Note | Certainty Note | Energy Efficiency (log) | Builds Upon | Superseded by | Possible Duplicate | Possible Duplicate Of | Chip type (secondary) | Chip quantity (secondary) | Total number of AI chips | GPU Supplier (primary) | GPU supplier (secondary) | Include in Standard Analysis | Exclude | Rank when first operational | 16-bit OP/s (log) | Max OP/s | 8-bit OP/s | 16-bit OP/s | 32-bit OP/s | Calculated Power Capacity (MW) | Reported Power Capacity (MW) | Energy Efficiency | Calculated Cost | Reported Cost | Reported Cost (Inflation adjusted) | Cost Quote | Noteworthy | Decommissioned Date (if applicable) | Largest existing cluster when first operational | % of largest cluster when first operational | Source 1 | Source 2 | Source 3 | Source 4 | Source 5 | latitude | longitude |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| xAI Colossus Memphis Phase 3 | Existing | Confirmed | Yes | 20.737033531333875 | 275795.85649723833 | NVIDIA H100 SXM5 80GB | 200000 | United States of America | xAI | 2025-07-22 | we know from earlier that it has 150k H100s and 50k H200s, which makes the total 150k H100s 50k H200s 30k GB200s | Private | 352.352 | 3231 Riverport Rd, Memphis, TN 38109 | xAI | 150k H100s, 50k H200s, 30k GB200s,H100,GB200,H200 | "230k GPUs, including 30k GB200s, are operational for training Grok @xAI in a single supercluster called Colossus 1" | 2025-07-22 | 11.889026794712445 | xAI Colossus Memphis Phase 2 | NVIDIA GB200 | 30000 | 230000 | NVIDIA | NVIDIA | True | 20.436003535669897 | 545800000000000000000 | 545800000000000000000 | 272900000000000000000 | 136440000000000000000 | 352.352 | 774509581327.7632 | True | https://archive.ph/oHQSA | 35.069154 | -90.118494 | |||||||||||||||||||
| xAI Colossus Memphis Phase 2 | Existing | Likely | Yes | 20.597475789870376 | 200000.000002945 | NVIDIA H100 SXM5 80GB | 150000 | United States of America | xAI | 2025-02-18 | GPU types is from a tweet. Unclear exactly when it was first operational. It was confirmed to be operational Feb 17, 2025, but some evidence suggests that it was operational by December 4th, 2024 | Private | 280.28 | 7120201208.681584 | 3231 Riverport Rd, Memphis, TN 38109 | xAI | H100,H200 | This weekend, the @xAI team brought our Colossus 100k H100 training cluster online.... Moreover, it will double in size to 200k (50k H200s) in a few months. | 2025-02-18 | 11.848853685384858 | xAI Colossus Memphis Phase 1 | xAI Colossus Memphis Phase 3 | NVIDIA H200 SXM | 50000 | 200000 | NVIDIA | NVIDIA | True | 1 | 20.29644579420639 | 395800000000000000000 | 395800000000000000000 | 197900000000000000000 | 98930000000000000000 | 280.28 | 706079634651.0632 | 7120201208.681584 | xAI Colossus Memphis Phase 2 | 1 | https://archive.ph/z39uN | https://web.archive.org/web/20241217184541/https://memphischamber.com/economic-development/xai/ | https://web.archive.org/web/20241217184541/https://memphischamber.com/economic-development/xai/ | https://www.youtube.com/watch?v=AUAJ82H12qs&t=260s | 35.069154 | -90.118494 | |||||||||||
| xAI Colossus Memphis Phase 1 | Existing | Confirmed | Yes | 20.29644579420639 | 100000.00000147166 | NVIDIA H100 SXM5 80GB | 100000 | United States of America | xAI | 2024-09-02 | Private | 150 | 3802927639.101294 | 3231 Riverport Rd, Memphis, TN 38109 | xAI | H100 | "With 100k liquid-cooled H100s on a single RDMA fabric" "This weekend, the @xAI team brought our Colossus 100k H100 training cluster online.... Moreover, it will double in size to 200k (50k H200s) in a few months" | 2024-09-02 | Reported to be online by Musk | 11.819324539486733 | xAI Colossus Memphis Phase 2 | 100000 | NVIDIA | True | 1 | 19.995415798542414 | 197900000000000000000 | 197900000000000000000 | 98950000000000000000 | 49470000000000000000 | 142.688 | 150 | 659666666666.6666 | 3802927639.101294 | xAI Colossus Memphis Phase 1 | 1 | https://archive.ph/Bh9Tq | https://web.archive.org/web/20240727065656/https://www.commercialappeal.com/story/news/government/city/2024/07/09/elon-musk-xai-memphis-electricity-water-mlgw/74331544007/ | https://archive.ph/z39uN | 35.069154 | -90.118494 | |||||||||||||||
| Meta 100k | Existing | Likely | Yes | 20.29644579420639 | 100000.00000147166 | NVIDIA H100 SXM5 80GB | 100000 | United States of America | Meta AI | 2024-10-30 | Reported as operational in October 2024 earnings call | Private | 142.688 | 3810062229.221386 | United States of America | Meta | H100 | Meta Platforms is putting the final touches on one such cluster, which will be a little bigger than 100,000 H100s | 2024-10-30 | Reported by The Information | 11.8410283478729 | 100000 | NVIDIA | True | 1 | 19.995415798542414 | 197900000000000000000 | 197900000000000000000 | 98950000000000000000 | 49470000000000000000 | 142.688 | 693471069746.58 | 3810062229.221386 | True | xAI Colossus Memphis Phase 1 | 1 | https://archive.ph/85KB1 | https://web.archive.org/web/20240901171617/https://www.semianalysis.com/p/100000-h100-clusters-power-network | https://web.archive.org/web/2/https://www.tomshardware.com/tech-industry/artificial-intelligence/meta-is-using-more-than-100-000-nvidia-h100-ai-gpus-to-train-llama-4-mark-zuckerberg-says-that-llama-4-is-being-trained-on-a-cluster-bigger-than-anything-that-ive-seen | https://web.archive.org/web/20241228090751/https://www.tomshardware.com/tech-industry/artificial-intelligence/meta-is-using-more-than-100-000-nvidia-h100-ai-gpus-to-train-llama-4-mark-zuckerberg-says-that-llama-4-is-being-trained-on-a-cluster-bigger-than-anything-that-ive-seen | 39.381266 | -97.922211 | ||||||||||||||
| OpenAI/Microsoft Goodyear Arizona | Existing | Likely | Yes | 20.29644579420639 | 100000.00000147166 | NVIDIA H100 SXM5 80GB | 100000 | United States of America | Microsoft,OpenAI | 2024-10-02 | It seems likely that this became operational in phases. It seems like 75k were likely available sometime mid 2024, and ~160k will be available in the future. There is very little official public data about this cluster, it's almost all infrences/rumors Microsoft appears to own it and rent it to OpenAI The Semianalysis lecture estimates that there are 32k H100s (48MW) per building. Satellite data as of Dec 2024 seems to show that 5 building are now complete (though this doesn't imply that the clusters inside have been built or are operational). This would imply that the cluster is growing to ~160k GPUs. It's also unclear if the newer buildings would still have Hoppers or would have Blackwell chips I personally suspect that this cluster was one of the primary clusters for training GPT4.5 | Private | 142.688 | 3802450830.801299 | 33°24'28.7"N 112°21'54.3"W Goodyear, Arizona, USA | OpenAI | H100 | "You already have 100K clusters. OpenAI in Arizona" "Our data shows one of OpenAI’s next training supercomputers in Arizona was going to have more than 75,000+ GPUs in a singular site by the middle of next year" | 2024-10-02 | This estimate is from the SemiAnalysis number of GPUs/building, and satellite imagery showing that they've built 5 buildings | 11.8410283478729 | 100000 | NVIDIA | True | 1 | 19.995415798542414 | 197900000000000000000 | 197900000000000000000 | 98950000000000000000 | 49470000000000000000 | 142.688 | 693471069746.58 | 3802450830.801299 | xAI Colossus Memphis Phase 1 | 1 | https://web.archive.org/web/20241006181731/https://www.dwarkeshpatel.com/p/dylan-jon | https://web.archive.org/web/20240609051243/https://www.semianalysis.com/p/microsoft-swallows-openais-core-team | https://www.youtube.com/watch?v=hobvps-H38o | 33.190283 | -111.966202 | ||||||||||||||||
| Oracle OCI Supercluster H200s | Existing | Likely | Yes | 20.11292572483009 | 65536.0000009646 | NVIDIA H200 SXM | 65536 | United States of America | Oracle | 2024-11-21 | Unclear exactly when this first became operational, if it was ever built. The announcement November 2024 states that the largest supercomputer in the cloud is now generally available, implying that this cluster has been built and is operational. However, it's possible that they're just advertising their capacity to scale H200 clusters to be the largest supercomputer in the cloud. It seems very likely that this cluster has been built, but it's not fully certain. I estimate the first operational date to be Q1 2025 to be slightly conservative on the timing | Private | 93.51200768000001 | 2345609100.860053 | Cloud | H200 | We’re excited to announce the general availability of Oracle Cloud Infrastructure (OCI) Supercluster with NVIDIA H200 Tensor Core GPUs. The largest AI supercomputer available in the cloud*, our latest Supercluster scales up to an industry-leading 65,536 GPUs. | 2024-11-21 | 11.8410283478729 | 65536 | NVIDIA | True | 4 | 19.811895729166114 | 129695744000000000000 | 129695744000000000000 | 64847872000000000000 | 32407552000000000000 | 93.51200768000001 | 693471069746.58 | 2345609100.860053 | xAI Colossus Memphis Phase 1 | 0.6553600000000014 | https://archive.ph/B3HJm#selection-1045.0-1045.83 | https://web.archive.org/web/20241214043143/https://www.datacenterdynamics.com/en/news/oracles-65000-gpu-supercluster-now-generally-available/ | 39.381266 | -97.922211 | |||||||||||||||||||
| Tesla Cortex Phase 1 | Existing | Confirmed | Yes | 19.995415798542414 | 50000.00000073623 | NVIDIA H100 SXM5 80GB | 50000 | United States of America | Tesla | 2024-11-15 | Private | 71.344 | 1905031114.610693 | 1 Tesla Rd, Austin, TX 78725 | Tesla | H100 | Tesla completed the deployment of its 50,000 Nvidia H100 GPU training cluster during Q4 2024. | Q4 2024 | 11.8410283478729 | 50000 | NVIDIA | True | 4 | 19.694385802878433 | 98950000000000000000 | 98950000000000000000 | 49475000000000000000 | 24735000000000000000 | 71.344 | 693471069746.58 | 1905031114.610693 | xAI Colossus Memphis Phase 1 | 0.500000000000004 | https://web.archive.org/web/20250217061801/https://www.datacenterdynamics.com/en/news/teslas-50000-gpu-cortex-supercomputer-went-live-in-q4-2024/ | https://web.archive.org/web/20240822222555/https://www.fool.com/earnings/call-transcripts/2024/07/24/tesla-tsla-q2-2024-earnings-call-transcript/ | https://web.archive.org/web/20240904184237/https://www.tomshardware.com/desktops/servers/elon-musk-shows-off-cortex-ai-supercluster-first-look-at-teslas-50000-nvidia-h100s | 30.222695 | -97.618788 | ||||||||||||||||||
| Lawrence Livermore NL El Capitan Phase 2 | Existing | Confirmed | Yes | 19.941311098168587 | 44143.351592362786 | AMD Instinct MI300A | 44544 | United States of America | US Department of Energy | 2024-11-18 | Public | 35 | 607324736.6965789 | Lawrence Livermore National Laboratory 7000 East Ave, Livermore, CA 94550 | US Government,Lawrence Livermore NL,Scientific Research | AMD MI300A | the system has 11,136 nodes packed with 44,544 of AMD's MI300A APUs | 2024-11-18 | 12.096213058154332 | Lawrence Livermore NL El Capitan Phase 1 | 44544 | AMD | True | 5 | 19.640281102504606 | 87359692800000000000 | 87359692800000000000 | 43679846400000000000 | 21839923200000000000 | 69.00685209600002 | 35 | 1247995611428.5715 | 1826641892.232672 | 600000000 | 607324736.6965789 | $600M was quoted as the max cost | True | xAI Colossus Memphis Phase 1 | 0.4414335159171314 | https://web.archive.org/web/20241122140255/https://www.tomshardware.com/pc-components/cpus/amd-powered-el-capitan-is-now-the-worlds-fastest-supercomputer-with-1-7-exaflops-of-performance-fastest-intel-machine-falls-to-third-place-on-top500-list | 37.682018 | -121.768374 | ||||||||||||||
| CoreWeave H200s | Existing | Likely | Yes | 19.919695084604296 | 42000.00000061847 | NVIDIA H200 SXM | 42000 | United States of America | CoreWeave | 2024-08-28 | Private | 59.92896000000001 | 1499797837.3467495 | CoreWeave | H200 | These instances are deployed in clusters with up to 42,000 GPUs | 2024-08-28 | 11.8410283478729 | 42000 | NVIDIA | True | 1 | 19.618665088940315 | 83118000000000000000 | 83118000000000000000 | 41559000000000000000 | 20769000000000000000 | 59.92896000000001 | 693471069746.5798 | 1499797837.3467495 | CoreWeave H200s | 1 | https://web.archive.org/web/20250107015942/https://www.coreweave.com/blog/coreweave-is-the-first-cloud-provider-to-deploy-nvidia-h200-tensor-core-gpus | 39.381266 | -97.922211 | |||||||||||||||||||||
| Meta GenAI 2024b | Existing | Confirmed | Yes | 19.686956992557814 | 24576.000000361928 | NVIDIA H100 SXM5 80GB | 24576 | United States of America | Meta AI | 2024-03-12 | At least one of these was used to train Llama 3 | Private | 35.067002880000004 | 880071195.8491452 | Meta | H100 | Today, we’re sharing details on two versions of our 24,576-GPU data center scale cluster at Meta | 2024-03-12 | From official Meta announcement. At least one of these was used to train Llama 3.1 | 11.8410283478729 | 24576 | NVIDIA | True | 1 | 19.385926996893833 | 48635904000000000000 | 48635904000000000000 | 24317952000000000000 | 12157747200000000000 | 35.067002880000004 | 693471069746.58 | 880071195.8491452 | Meta GenAI 2024a | 1 | http://web.archive.org/web/20240828225612/https://engineering.fb.com/2024/03/12/data-center-engineering/building-metas-genai-infrastructure/ | https://web.archive.org/web/20240827223115/https://www.semianalysis.com/p/gpt-4-architecture-infrastructure | 39.381266 | -97.922211 | ||||||||||||||||||
| Meta GenAI 2024a | Existing | Confirmed | Yes | 19.686956992557814 | 24576.000000361928 | NVIDIA H100 SXM5 80GB | 24576 | United States of America | Meta AI | 2024-03-12 | At least one of these was used to train Llama 3 | Private | 35.067002880000004 | 880071195.8491452 | Meta | H100 | Today, we’re sharing details on two versions of our 24,576-GPU data center scale cluster at Meta | 2024-03-12 | From official Meta announcement. At least one of these was used to train Llama 3.1 | 11.8410283478729 | 24576 | NVIDIA | True | 1 | 19.385926996893833 | 48635904000000000000 | 48635904000000000000 | 24317952000000000000 | 12157747200000000000 | 35.067002880000004 | 693471069746.58 | 880071195.8491452 | True | Meta GenAI 2024a | 1 | http://web.archive.org/web/20240828225612/https://engineering.fb.com/2024/03/12/data-center-engineering/building-metas-genai-infrastructure/ | https://web.archive.org/web/20240827223115/https://www.semianalysis.com/p/gpt-4-architecture-infrastructure | 39.381266 | -97.922211 | |||||||||||||||||
| "Jupiter, Jülich" | Existing | Confirmed | Yes | 19.66817844958985 | 23536.00000034651 | NVIDIA GH200 | 23536 | Germany | EuroHPC JU,Julich Supercomputing Center | 2025-06-10 | Public | 18 | 290000685.87105626 | Jülich Supercomputing Centre (JSC) 52428 Jülich, Germany | GH200 | "JUPITER marks the debut of a quad NVIDIA GH200 Grace Hopper Superchip node configuration, based on Eviden’s BullSequana XH3000 liquid-cooled architecture, with a booster module comprising close to 24,000 NVIDIA GH200 Superchips interconnected with the NVIDIA Quantum-2 InfiniBand networking platform." "If you look at the Top500 spreadsheet, it tells you the total number of cores used in the machine, plus the number of cores used by accelerators. (These cores are synonymous with a streaming multiprocessor in a GPU architecture, not a count of CUDA cores or tensor cores.) The Jupiter GPU Booster used for the HPL test that have it the number four ranking had 4,801,344 cores and of these, there are 3,106,752 cores allocated to the GPUs, which leaves 1,694,592 cores in the CPU hosts. At 72 cores per Grace, that is 23,536 Grace CPUs and therefore 23,536 Grace-Hopper units and therefore 23,536 Hopper H200 GPUs." | 2025-06-10 | 12.111875948822563 | 23536 | NVIDIA | True | 19.367148453925868 | 46577744000000000000 | 46577744000000000000 | 23288872000000000000 | 11638552000000000000 | 32.9833504 | 18 | 1293826222222.2222 | 1089491617.5764012 | 287000000 | 290000685.87105626 | EuroHPC says in a statement that the cost of building, delivering, installing, and maintaining the Jupiter machine is €273 million ($287.3 million), and presumably the remaining part of that €500 million is to build or retrofit a datacenter for Jupiter and pay for power and cooling for the machine. | https://web.archive.org/web/20240825120701/https://nvidianews.nvidia.com/news/nvidia-grace-hopper-superchip-powers-jupiter-defining-a-new-class-of-supercomputers-to-propel-ai-for-scientific-discovery | https://archive.is/oyvz9 | https://archive.ph/wip/HgCqo | https://www.anandtech.com/show/21136/nvidia-at-sc23-h200-accelerator-with-hbm3e-and-jupiter-supercomputer-for-2024 | https://top500.org/system/180357/ | 50.921734 | 6.358342 |