# Capacity — Training evidence ## Training claims for 203 companies Checked against site power, accelerator inventory, financial disclosure and export-control exposure. ### The register 0companies where all four signals are present 9companies with 3 of the four 141companies with 2 of the four 53companies with 1 of the four Order bySignals presentFacility powerGPU inventoryName Cite CSVJSONJSON+Prov Company Facility power GPUs Financials Export controls Power (MW) GPUs on record Meta Platforms, Inc. 4 countries 37 facilities 904,448 GPUs, estimated financial figure from a primary filing no export-control designation matched 6,127 904,448 Alphabet 30 countries 95 facilities 179,800 GPUs, estimated financial figure from a primary filing no export-control designation matched 3,002 179,800 Amazon.com, Inc. 27 countries 92 facilities 94,544 GPUs, estimated financial figure from a primary filing no export-control designation matched 2,811 94,544 Microsoft Corporation 32 countries shared site 120 facilities 648,752 GPUs, estimated financial figure from a primary filing no export-control designation matched 1,592 648,752 Tesla 1 country 12 facilities 306,904 GPUs, estimated financial figure from a public aggregator no export-control designation matched 218 306,904 NVIDIA 4 countries 20 facilities 75,376 GPUs, estimated financial figure from a primary filing no export-control designation matched 98 75,376 Samsung 2 countries 5 facilities 4,400 GPUs, estimated financial figure from a primary filing no export-control designation matched 5 4,400 Intel 2 countries 2 facilities 4,000 GPUs, estimated financial figure from a primary filing no export-control designation matched 5 4,000 IBM 3 countries 6 facilities 4,200 GPUs, estimated financial figure from a public aggregator no export-control designation matched 4 4,200 Oracle Corporation 27 countries shared site 60 facilities 262,656 GPUs, estimated no published financial figure no export-control designation matched 8,032 262,656 IREN Limited 2 countries 4 facilities 816 GPUs, estimated no published financial figure no export-control designation matched 2,751 816 OpenAI 1 country shared site 3 facilities 1,150,000 GPUs, estimated no published financial figure no export-control designation matched 2,500 1,150,000 CoreWeave 2 countries 13 facilities 86,000 GPUs, estimated no published financial figure no export-control designation matched 1,353 86,000 xAI 1 country 3 facilities 352,448 GPUs, estimated no published financial figure no export-control designation matched 1,306 352,448 Reliance Industries 1 country 2 facilities 450,000 GPUs, estimated no published financial figure no export-control designation matched 1,168 450,000 Sesterce 1 country 10 facilities 490,976 GPUs, estimated no published financial figure no export-control designation matched 995 490,976 Micron 1 country 1 facilities no GPU inventory on record financial figure from a primary filing no export-control designation matched 576 — Humain 1 country 4 facilities 18,000 GPUs, estimated no published financial figure no export-control designation matched 500 18,000 Nebius 3 countries 7 facilities 254,512 GPUs, estimated no published financial figure no export-control designation matched 433 254,512 Applied Digital 1 country 4 facilities 335,000 GPUs, estimated no published financial figure no export-control designation matched 400 335,000 Yotta Data Services 1 country shared site 9 facilities 36,864 GPUs, estimated no published financial figure no export-control designation matched 327 36,864 Huawei Technologies 1 country 4 facilities no GPU inventory on record no published financial figure 4 designation matches, all by name similarity 242 — U.S. Department of Energy 1 country 17 facilities 235,265 GPUs, estimated no published financial figure no export-control designation matched 198 235,265 Alibaba 16 countries 24 facilities no GPU inventory on record financial figure from a public aggregator no export-control designation matched 169 — G42 2 countries shared site 15 facilities 192,400 GPUs, estimated no published financial figure no export-control designation matched 163 192,400 SK Telecom 1 country 2 facilities 61,040 GPUs, estimated no published financial figure no export-control designation matched 104 61,040 Together AI 1 country 2 facilities 40,424 GPUs, estimated no published financial figure no export-control designation matched 93 40,424 Government of Telangana 1 country 2 facilities 29,000 GPUs, estimated no published financial figure no export-control designation matched 50 29,000 Lambda Labs 1 country 1 facilities 32,000 GPUs, estimated no published financial figure no export-control designation matched 46 32,000 EuroHPC JU 12 countries 21 facilities 107,732 GPUs, estimated no published financial figure no export-control designation matched 44 107,732 Eclairion 1 country 1 facilities 13,800 GPUs, estimated no published financial figure no export-control designation matched 44 13,800 Foxconn 1 country 2 facilities 14,608 GPUs, estimated no published financial figure no export-control designation matched 37 14,608 YTL Power 1 country 1 facilities 15,428 GPUs, estimated no published financial figure no export-control designation matched 37 15,428 Voltage Park 1 country 7 facilities 25,528 GPUs, estimated no published financial figure no export-control designation matched 34 25,528 Iliad Sa 1 country 5 facilities 1,016 GPUs, estimated no published financial figure no export-control designation matched 32 1,016 Inflection AI 1 country 1 facilities 22,000 GPUs, estimated no published financial figure no export-control designation matched 31 22,000 TensorWave 1 country 3 facilities 21,000 GPUs, estimated no published financial figure no export-control designation matched 30 21,000 Andreessen Horowitz 1 country 1 facilities 20,000 GPUs, estimated no published financial figure no export-control designation matched 29 20,000 Nexgen Cloud 1 country 1 facilities 16,384 GPUs, estimated no published financial figure no export-control designation matched 23 16,384 Eni 1 country 5 facilities 30,338 GPUs, estimated no published financial figure no export-control designation matched 22 30,338 Sakura Internet 1 country 3 facilities 12,784 GPUs, estimated no published financial figure no export-control designation matched 17 12,784 Eth Domain 1 country 4 facilities 25,896 GPUs, estimated no published financial figure no export-control designation matched 16 25,896 SoftBank Group Corp. 2 countries 7 facilities 10,128 GPUs, estimated no published financial figure no export-control designation matched 16 10,128 Northern Data Group 2 countries 5 facilities 10,240 GPUs, estimated no published financial figure no export-control designation matched 15 10,240 Imbue 1 country 1 facilities 10,000 GPUs, estimated no published financial figure no export-control designation matched 15 10,000 Xtx Markets 1 facilities 20,000 GPUs, estimated no published financial figure no export-control designation matched 15 20,000 Saudi Aramco 1 country 3 facilities 28,053 GPUs, estimated no published financial figure no export-control designation matched 14 28,053 Poolside 1 facilities 10,000 GPUs, estimated no published financial figure no export-control designation matched 14 10,000 Igenius 1 country 1 facilities 5,760 GPUs, estimated no published financial figure no export-control designation matched 14 5,760 Ministry of Science and ICT 1 country 2 facilities 8,800 GPUs, estimated no published financial figure no export-control designation matched 12 8,800 Saudi Data And Artificial Intelligence Authority 1 country 1 facilities 5,000 GPUs, estimated no published financial figure no export-control designation matched 12 5,000 National Institute of Advanced Industrial Science and Technology 1 country 4 facilities 17,840 GPUs, estimated no published financial figure no export-control designation matched 11 17,840 FPT Corporation 2 countries 2 facilities 6,000 GPUs, estimated no published financial figure no export-control designation matched 8 6,000 UK Research and Innovation 1 country 2 facilities 5,448 GPUs, estimated no published financial figure no export-control designation matched 8 5,448 Exxonmobil 1 country 2 facilities 6,032 GPUs, estimated no published financial figure no export-control designation matched 7 6,032 Smc Sustainable Metal Cloud 1 country 2 facilities 6,024 GPUs, estimated no published financial figure no export-control designation matched 7 6,024 Petrobras 1 country 6 facilities 7,280 GPUs, estimated no published financial figure no export-control designation matched 6 7,280 Barcelona Supercomputing Center 1 country 1 facilities 4,480 GPUs, estimated no published financial figure no export-control designation matched 6 4,480 French Alternative Energies and Atomic Energy Commission 1 country 3 facilities 4,788 GPUs, estimated no published financial figure no export-control designation matched 6 4,788 Nat Friedman And Daniel Gross 1 country 2 facilities 6,912 GPUs, estimated no published financial figure no export-control designation matched 6 6,912 United States Government 1 country 2 facilities 8,320 GPUs, estimated no published financial figure no export-control designation matched 6 8,320 Panaai 1 country 1 facilities 4,088 GPUs, estimated no published financial figure no export-control designation matched 6 4,088 Oneasia 1 country 1 facilities 4,000 GPUs, estimated no published financial figure no export-control designation matched 6 4,000 Tokyo Institute of Technology 1 country 3 facilities 7,340 GPUs, estimated no published financial figure no export-control designation matched 6 7,340 King Abdullah University of Science and Technology 1 country 1 facilities 2,816 GPUs, estimated no published financial figure no export-control designation matched 5 2,816 KDDI 1 country 2 facilities 2,000 GPUs, estimated no published financial figure no export-control designation matched 5 2,000 Vultr 1 country 1 facilities 3,000 GPUs, estimated no published financial figure no export-control designation matched 5 3,000 GENCI 1 country 5 facilities 11,908 GPUs, estimated no published financial figure no export-control designation matched 4 11,908 Jülich Supercomputing Centre 1 country 2 facilities 4,512 GPUs, estimated no published financial figure no export-control designation matched 4 4,512 Yandex 2 countries 4 facilities 3,776 GPUs, estimated no published financial figure no export-control designation matched 3 3,776 Novo Nordisk Foundation 1 country 1 facilities 1,528 GPUs, estimated no published financial figure no export-control designation matched 3 1,528 Horizon Compute 1 country 2 facilities 3,072 GPUs, estimated no published financial figure no export-control designation matched 3 3,072 Naver 1 country 2 facilities 3,360 GPUs, estimated no published financial figure no export-control designation matched 3 3,360 U.S. Department of Defense 1 country 2 facilities 1,176 GPUs, estimated no published financial figure no export-control designation matched 3 1,176 CINECA 1 country 3 facilities 3,952 GPUs, estimated no published financial figure no export-control designation matched 3 3,952 Japan Agency for Marine-Earth Science and Technology 1 country 2 facilities 11,600 GPUs, estimated no published financial figure no export-control designation matched 3 11,600 TotalEnergies 1 country 1 facilities 3,384 GPUs, estimated no published financial figure no export-control designation matched 2 3,384 KT 1 country 3 facilities 1,992 GPUs, estimated no published financial figure no export-control designation matched 2 1,992 Gcore 1 country 2 facilities 1,320 GPUs, estimated no published financial figure no export-control designation matched 2 1,320 Ubilink Ai 1 country 1 facilities 1,024 GPUs, estimated no published financial figure no export-control designation matched 2 1,024 Preferred Networks Inc 1 country 2 facilities 2,560 GPUs, estimated no published financial figure no export-control designation matched 2 2,560 Joint Center for Advanced High Performance Computing 1 country 1 facilities 1,120 GPUs, estimated no published financial figure no export-control designation matched 2 1,120 Compute Canada 1 country 3 facilities 2,092 GPUs, estimated no published financial figure no export-control designation matched 2 2,092 National Applied Research Laboratories Taiwan 1 country 1 facilities 2,016 GPUs, estimated no published financial figure no export-control designation matched 1 2,016 Sbercloud 1 country 2 facilities 1,992 GPUs, estimated no published financial figure no export-control designation matched 1 1,992 Denvr Dataworks 1 country 1 facilities 1,024 GPUs, estimated no published financial figure no export-control designation matched 1 1,024 Lepton Ai 1 country 1 facilities 1,024 GPUs, estimated no published financial figure no export-control designation matched 1 1,024 Neevcloud 1 country 2 facilities 41,024 GPUs, estimated no published financial figure no export-control designation matched 1 41,024 Ori Industries 1 facilities 1,024 GPUs, estimated no published financial figure no export-control designation matched 1 1,024 SIAM AI 1 country 1 facilities 1,024 GPUs, estimated no published financial figure no export-control designation matched 1 1,024 Vng Corporation 1 country 1 facilities 1,024 GPUs, estimated no published financial figure no export-control designation matched 1 1,024 Nhn Corporation 1 country 1 facilities 1,000 GPUs, estimated no published financial figure no export-control designation matched 1 1,000 Hut 8 2 countries 2 facilities 2,000 GPUs, estimated no published financial figure no export-control designation matched 1 2,000 Rensselaer Polytechnic Institute 1 country 2 facilities 2,120 GPUs, estimated no published financial figure no export-control designation matched 1 2,120 University of Florida 1 country 2 facilities 1,120 GPUs, estimated no published financial figure no export-control designation matched 1 1,120 Recursion Pharmaceuticals 1 country 2 facilities 824 GPUs, estimated no published financial figure no export-control designation matched 1 824 The University of Texas at Austin 1 country 4 facilities 1,508 GPUs, estimated no published financial figure no export-control designation matched 0.9 1,508 Moscow State Univeristy 1 country 2 facilities 1,572 GPUs, estimated no published financial figure no export-control designation matched 0.9 1,572 AGH University of Krakow 1 country 2 facilities 848 GPUs, estimated no published financial figure no export-control designation matched 0.9 848 Pawsey Supercomputing Centre 1 country 1 facilities 768 GPUs, estimated no published financial figure no export-control designation matched 0.8 768 DeepL 1 country 1 facilities 544 GPUs, estimated no published financial figure no export-control designation matched 0.8 544 Nec Corporation 1 country 1 facilities 928 GPUs, estimated no published financial figure no export-control designation matched 0.8 928 SURF 1 country 3 facilities 1,568 GPUs, estimated no published financial figure no export-control designation matched 0.7 1,568 Mosaicml 2 facilities 696 GPUs, estimated no published financial figure no export-control designation matched 0.7 696 Ahrefs 1 country 1 facilities 504 GPUs, estimated no published financial figure no export-control designation matched 0.7 504 Bny Mellon 1 facilities 500 GPUs, estimated no published financial figure no export-control designation matched 0.7 500 Japan Atomic Energy Agency 1 country 1 facilities 1,088 GPUs, estimated no published financial figure no export-control designation matched 0.7 1,088 Princeton University 1 country 2 facilities 928 GPUs, estimated no published financial figure no export-control designation matched 0.7 928 University of Illinois Urbana-Champaign 1 country 1 facilities 880 GPUs, estimated no published financial figure no export-control designation matched 0.7 880 Max Planck Society 1 country 1 facilities 768 GPUs, estimated no published financial figure no export-control designation matched 0.7 768 Csiro 1 country 1 facilities 448 GPUs, estimated no published financial figure no export-control designation matched 0.6 448 Karlsruhe Institute of Technology 1 country 1 facilities 740 GPUs, estimated no published financial figure no export-control designation matched 0.6 740 Linköping University 1 country 2 facilities 1,232 GPUs, estimated no published financial figure no export-control designation matched 0.6 1,232 Nagoya University 1 country 1 facilities 884 GPUs, estimated no published financial figure no export-control designation matched 0.6 884 Centre for Development of Advanced Computing 1 country 2 facilities 992 GPUs, estimated no published financial figure no export-control designation matched 0.5 992 National Science and Technology Development Agency 1 country 1 facilities 704 GPUs, estimated no published financial figure no export-control designation matched 0.5 704 Indiana University 1 country 2 facilities 616 GPUs, estimated no published financial figure no export-control designation matched 0.5 616 Hessian Ai 1 country 1 facilities 632 GPUs, estimated no published financial figure no export-control designation matched 0.5 632 Corvex 1 country 1 facilities 256 GPUs, estimated no published financial figure no export-control designation matched 0.5 256 Polish Academy of Sciences 1 country 1 facilities 348 GPUs, estimated no published financial figure no export-control designation matched 0.5 348 Aleph Alpha 1 country 1 facilities 512 GPUs, estimated no published financial figure no export-control designation matched 0.4 512 Wrocław University of Science and Technology 1 country 1 facilities 304 GPUs, estimated no published financial figure no export-control designation matched 0.4 304 Distributed Research Using Advanced Computing Dirac 1 country 1 facilities 448 GPUs, estimated no published financial figure no export-control designation matched 0.4 448 National Computational Infrastructure Australia 1 country 1 facilities 640 GPUs, estimated no published financial figure no export-control designation matched 0.4 640 The MITRE Corporation 1 country 1 facilities 256 GPUs, estimated no published financial figure no export-control designation matched 0.4 256 Fastweb 1 country 1 facilities 248 GPUs, estimated no published financial figure no export-control designation matched 0.4 248 Opera 1 country 1 facilities 248 GPUs, estimated no published financial figure no export-control designation matched 0.4 248 Goethe University Frankfurt 1 country 1 facilities 880 GPUs, estimated no published financial figure no export-control designation matched 0.3 880 The University of Tokyo 1 country 1 facilities 360 GPUs, estimated no published financial figure no export-control designation matched 0.3 360 Massachusetts Institute of Technology 1 country 1 facilities 488 GPUs, estimated no published financial figure no export-control designation matched 0.3 488 National Supercomputing Center Singapore 1 country 2 facilities 608 GPUs, estimated no published financial figure no export-control designation matched 0.3 608 Continental 1 country 1 facilities 448 GPUs, estimated no published financial figure no export-control designation matched 0.3 448 Osaka University 1 country 1 facilities 336 GPUs, estimated no published financial figure no export-control designation matched 0.3 336 Technische Universität Dresden 1 country 1 facilities 272 GPUs, estimated no published financial figure no export-control designation matched 0.3 272 Leonardo Spa 1 country 1 facilities 320 GPUs, estimated no published financial figure no export-control designation matched 0.3 320 University of Cambridge 1 country 2 facilities 320 GPUs, estimated no published financial figure no export-control designation matched 0.3 320 Laboratório Nacional de Computação Científica 1 country 1 facilities 368 GPUs, estimated no published financial figure no export-control designation matched 0.2 368 University of Minnesota 1 country 1 facilities 264 GPUs, estimated no published financial figure no export-control designation matched 0.2 264 CSC – IT Center for Science 1 country 1 facilities 320 GPUs, estimated no published financial figure no export-control designation matched 0.2 320 University of Tsukuba 1 country 1 facilities 320 GPUs, estimated no published financial figure no export-control designation matched 0.2 320 Immunitybio 1 country 1 facilities 320 GPUs, estimated no published financial figure no export-control designation matched 0.2 320 Senai Cimatec 1 country 1 facilities 312 GPUs, estimated no published financial figure no export-control designation matched 0.2 312 University of Oxford 1 country 2 facilities 504 GPUs, estimated no published financial figure no export-control designation matched 0.2 504 Riken 1 country 1 facilities 416 GPUs, estimated no published financial figure no export-control designation matched 0.2 416 Sidi 1 country 1 facilities 200 GPUs, estimated no published financial figure no export-control designation matched 0.2 200 Texas A&M University 1 country 1 facilities 200 GPUs, estimated no published financial figure no export-control designation matched 0.2 200 Dell Technologies 1 country 1 facilities 192 GPUs, estimated no published financial figure no export-control designation matched 0.2 192 University of Stuttgart 1 country 1 facilities 192 GPUs, estimated no published financial figure no export-control designation matched 0.2 192 Mts 1 country 2 facilities 160 GPUs, estimated no published financial figure no export-control designation matched 0.1 160 Vingroup 1 country 1 facilities 160 GPUs, estimated no published financial figure no export-control designation matched 0.1 160 4paradigm no facility power on record no GPU inventory on record no published financial figure 1 designation matches, all by name similarity — — Academic Computer Centre Cyfronet AGH 1 country no facility power on record 464 GPUs, estimated no published financial figure no export-control designation matched — 464 AMD no facility power on record no GPU inventory on record financial figure from a primary filing no export-control designation matched — — Apple 3 countries no facility power on record no GPU inventory on record financial figure from a primary filing no export-control designation matched — — Applied Materials no facility power on record no GPU inventory on record financial figure from a primary filing no export-control designation matched — — Beijing E-Town Semiconductor Technology Co., Ltd. no facility power on record no GPU inventory on record no published financial figure 1 designation matches, all by name similarity — — Beijing Lingxin Intelligent Technology Co no facility power on record no GPU inventory on record no published financial figure 1 designation matches, all by name similarity — — Beijing University of Posts and Telecommunications no facility power on record no GPU inventory on record no published financial figure 1 designation matches, all by name similarity — — Biren Technology no facility power on record no GPU inventory on record no published financial figure 2 designation matches, all by name similarity — — Broadcom Inc. no facility power on record no GPU inventory on record financial figure from a primary filing no export-control designation matched — — Cambricon no facility power on record no GPU inventory on record no published financial figure 2 designation matches, at least one an exact name identity — — China Mobile 2 countries no facility power on record 38,000 GPUs, estimated no published financial figure no export-control designation matched — 38,000 China Unicom 2 countries no facility power on record 14,000 GPUs, estimated no published financial figure no export-control designation matched — 14,000 CloudWalk Technology no facility power on record no GPU inventory on record no published financial figure 1 designation matches, all by name similarity — — Dahua Technology no facility power on record no GPU inventory on record no published financial figure 1 designation matches, at least one an exact name identity — — Dongfang Jingyuan Electron Ltd. no facility power on record no GPU inventory on record no published financial figure 1 designation matches, all by name similarity — — GuoSheng no facility power on record no GPU inventory on record no published financial figure 1 designation matches, all by name similarity — — H3C Technologies Co., Ltd. no facility power on record no GPU inventory on record no published financial figure 1 designation matches, all by name similarity — — Hangzhou Hikvision Digital Technology Co., Ltd. no facility power on record no GPU inventory on record no published financial figure 1 designation matches, at least one an exact name identity — — Harbin Institute of Technology no facility power on record no GPU inventory on record no published financial figure 1 designation matches, at least one an exact name identity — — Hewlett Packard Enterprise 2 countries no facility power on record 1,000 GPUs, estimated no published financial figure no export-control designation matched — 1,000 HiSilicon no facility power on record no GPU inventory on record no published financial figure 57 designation matches, all by name similarity — — Huada Empyrean no facility power on record no GPU inventory on record no published financial figure 1 designation matches, at least one an exact name identity — — Huawei Investment & Holding Co., Ltd no facility power on record no GPU inventory on record no published financial figure 3 designation matches, all by name similarity — — iFLYTEK Co., Ltd. no facility power on record no GPU inventory on record no published financial figure 1 designation matches, at least one an exact name identity — — iFLYTEK Research no facility power on record no GPU inventory on record no published financial figure 1 designation matches, all by name similarity — — Inspur no facility power on record no GPU inventory on record no published financial figure 2 designation matches, all by name similarity — — Intellifusion no facility power on record no GPU inventory on record no published financial figure 1 designation matches, at least one an exact name identity — — Lam Research no facility power on record no GPU inventory on record financial figure from a primary filing no export-control designation matched — — Loongson no facility power on record no GPU inventory on record no published financial figure 1 designation matches, all by name similarity — — Marvell Technology no facility power on record no GPU inventory on record financial figure from a primary filing no export-control designation matched — — Megvii Inc no facility power on record no GPU inventory on record no published financial figure 1 designation matches, at least one an exact name identity — — Moscow Institute of Physics and Technology no facility power on record no GPU inventory on record no published financial figure 1 designation matches, at least one an exact name identity — — NAURA no facility power on record no GPU inventory on record no published financial figure 2 designation matches, at least one an exact name identity — — Northwestern Polytechnical University no facility power on record no GPU inventory on record no published financial figure 1 designation matches, at least one an exact name identity — — Peng Cheng Laboratory no facility power on record no GPU inventory on record no published financial figure 1 designation matches, at least one an exact name identity — — Piotech no facility power on record no GPU inventory on record no published financial figure 3 designation matches, at least one an exact name identity — — Qualcomm Inc no facility power on record no GPU inventory on record financial figure from a primary filing no export-control designation matched — — QuantumCTek Co., Ltd. no facility power on record no GPU inventory on record no published financial figure 2 designation matches, all by name similarity — — SCHNEIDER ELECTRIC S E label artefact no facility power on record no GPU inventory on record no published financial figure 1 designation matches, all by name similarity — — SenseTime Group B label artefact no facility power on record no GPU inventory on record no published financial figure 1 designation matches, all by name similarity — — Sichuan University no facility power on record no GPU inventory on record no published financial figure 1 designation matches, at least one an exact name identity — — Sino Ic Technology Co Ltd no facility power on record no GPU inventory on record no published financial figure 1 designation matches, at least one an exact name identity — — Skolkovo Institute of Science and Technology no facility power on record no GPU inventory on record no published financial figure 1 designation matches, at least one an exact name identity — — SMIC no facility power on record no GPU inventory on record no published financial figure 6 designation matches, all by name similarity — — SMIC Shanghai no facility power on record no GPU inventory on record no published financial figure 1 designation matches, at least one an exact name identity — — SMIC-Beijing no facility power on record no GPU inventory on record no published financial figure 2 designation matches, at least one an exact name identity — — Stargate OpenAI 1 country shared site no facility power on record 35,000 GPUs, estimated no published financial figure no export-control designation matched — 35,000 Texas Instruments no facility power on record no GPU inventory on record financial figure from a primary filing no export-control designation matched — — Tianjin University no facility power on record no GPU inventory on record no published financial figure 1 designation matches, at least one an exact name identity — — University of Electronic Science and Technology of China no facility power on record no GPU inventory on record no published financial figure 1 designation matches, at least one an exact name identity — — Yangtze Memory Technologies Co., Ltd. no facility power on record no GPU inventory on record no published financial figure 1 designation matches, at least one an exact name identity — — Zhipu AI no facility power on record no GPU inventory on record no published financial figure 1 designation matches, at least one an exact name identity — — Reading The four signals do not accumulate on the same companies — they partition. **41** companies here have an export-control designation and **19** have a public financial disclosure; **none** has both. A company that files with the SEC is generally not a company the BIS lists, so the corroboration available for a training claim depends on which side of that line the claimant sits — and the two sides are auditable in entirely different ways. That is why no row shows four marks, and it is a fact about the disclosure regimes rather than about the companies. Reading the table 203 companies for which Scrutica holds at least one of the four signals. A filled square is a signal present from a primary or research source; an open square is present but with estimated values; a short horizontal rule means the signal is absent. 3 sites are attributed to more than one company, so the Power and GPU columns overlap between those rows and do not sum down the register; the rows say which, and the detail panel gives every published reading. The instrument A training run has four physical preconditions that leave marks in other people's records: a site drawing serious power, accelerators bought from someone, capital a public company has to book, and — where the buyer sits on the wrong side of an export control — a paper trail at the border. An announcement leaves none of them. Absence The four marks are not a score. A missing signal usually records a company's legal status: a private company files nothing, and a US firm has no export-control designation by construction. Not published here Licensed financial values never leave the server. Where a company's only financial figure comes from a licensed database, this page shows no financial signal at all. ### Evidence for Meta Platforms, Inc. 3 of 4 signals presentBack to register Facility power6,127 MW Across **37** facilities. Nameplate capacity, not measured draw. GPU inventory904,448 Accelerators disclosed across its sites. 9 of 37 facilities have a per-model inventory rather than a total alone. The mark beside this figure is open because at least one contributing count is estimated or has no per-model breakdown, which is weaker evidence than an itemised count from a primary source. Financial disclosure$69.7B Annual capital expenditure, FY 2025. Tier 1 · a primary filing (data.sec.gov). [Source](https://data.sec.gov/api/xbrl/companyfacts/CIK0001326801.json) Export-control exposure No designation matched. This company holds sites both inside and outside the US (DK, IE, SE, US), so the lists reach part of its footprint and an empty result is informative about that part. Sites attributed to Meta Platforms, Inc.. Facility Country Power (MW) GPUs Basis [Meta Louisiana Datacenter](/facilities/fac-epoch-gc-meta-louisiana-datacenter) US 2,260 — estimated [Meta Richland Parish (Hyperion)](/facilities/fac-meta-richland-parish) US 2,000 — estimated [Meta Prometheus New Albany](/facilities/fac-epoch-meta-prometheus) US 631 500,000 estimated [Meta Jeffersonville](/facilities/fac-epoch-meta-jeffersonville) US 178 — estimated [Meta Rosemount](/facilities/fac-epoch-meta-rosemount) US 178 — estimated [Meta Montgomery](/facilities/fac-epoch-meta-montgomery) US 153 — estimated [Meta Kuna](/facilities/fac-epoch-meta-kuna) US 152 — estimated [Meta Temple](/facilities/fac-epoch-meta-temple) US 152 — estimated [Meta 100k](/facilities/fac-epoch-gc-meta-100k) US 143 100,000 estimated [Meta Gallatin](/facilities/fac-epoch-meta-gallatin) US 87 — estimated [Meta Los Lunas](/facilities/fac-epoch-meta-los-lunas) US 87 — estimated [Meta GenAI 2024a](/facilities/fac-epoch-gc-meta-genai-2024a) US 35 24,576 disclosed [Meta GenAI 2024b](/facilities/fac-epoch-gc-meta-genai-2024b) US 35 24,576 disclosed [Meta 2017 V100 Cluster](/facilities/fac-epoch-gc-meta-2017-v100-cluster) US 15 22,000 disclosed [Meta Research SuperCluster (RSC-1) Phase 2](/facilities/fac-epoch-gc-meta-research-supercluster-rsc-1-phase-2) US 13 16,000 disclosed [Meta Research SuperCluster 2 (RSC-2)](/facilities/fac-epoch-gc-meta-research-supercluster-2-rsc-2) US 7 8,000 disclosed [Meta 2017 P100 Cluster](/facilities/fac-epoch-gc-meta-2017-p100-cluster) US 0.7 992 disclosed [Facebook Altoona Data Center 1](/facilities/fac-im3-facebook-altoona-data-center-1-ia) US — — disclosed [Facebook Fort Worth Datacenter](/facilities/fac-im3-facebook-fort-worth-datacenter-tx) US — — disclosed [Meta](/facilities/fac-im3-meta-nm) US — 202,224 disclosed [Meta Aiken Data Center](/facilities/fac-im3-meta-aiken-data-center-sc) US — — disclosed [Meta Beaver Dam Data Center](/facilities/fac-im3-meta-beaver-dam-data-center-wi) US — — disclosed [Meta Cheyenne Data Center](/facilities/fac-im3-meta-cheyenne-data-center-wy) US — — disclosed [Meta Clonee 1 Data Center](/facilities/fac-osm2-w518701317) IE — — disclosed [Meta Data Center](/facilities/fac-im3-meta-data-center-mo) US — — disclosed [Meta Eagle Mountain Data Center](/facilities/fac-im3-meta-eagle-mountain-data-center-ut) US — — disclosed [Meta Forest City datacenter](/facilities/fac-im3-meta-forest-city-datacenter-nc) US — — disclosed [Meta Henrico Data Center](/facilities/fac-im3-meta-henrico-data-center-va) US — — disclosed [Meta Huntsville Data center](/facilities/fac-im3-meta-huntsville-data-center-al) US — — disclosed [Meta Odense Data center](/facilities/fac-osm2-w648731568) DK — — disclosed [Meta Prineville Data Center](/facilities/fac-meta-prineville) US — — disclosed [Meta Research SuperCluster (RSC-1) Phase 1](/facilities/fac-epoch-gc-meta-research-supercluster-rsc-1-phase-1) US — 6,080 disclosed [Meta Sarpy](/facilities/fac-epoch-meta-sarpy) US — — estimated [Meta Sarpy Data center](/facilities/fac-im3-meta-sarpy-data-center-ne) US — — disclosed [Meta Stanton Springs Data Center](/facilities/fac-osm2-w1054387470) US — — disclosed [Meta Temple Texas](/facilities/fac-epoch-gc-meta-temple-texas) US — — estimated [Metas datacenter i Luleå](/facilities/fac-osm-260883600) SE — — disclosed ### The industry-wide upper bound on chip supply The cumulative chip series bounds the field at industry level. It has no per-company decomposition, so it appears as context rather than a fifth matrix column. Median estimates 20,446,391 cumulative H100-equivalents through 2025-12-31, summed across the 25 accelerator types that have a cumulative series — the most recent date on which every one of them reports, so the figure is a single vintage rather than a mixture. A further 1,204,109 has been reported since, by the 13 of those series that run through 2026-03-31; the remaining 12 have not reported that quarter, so the 21,650,500 those two figures make together is a mixed-vintage sum and is not comparable to any quarter-stamped total. 25 of the 28 accelerator types the chip-sales dataset covers have a cumulative series; H200, TPU v3, Trainium3 have none and are shown above with the columns they cannot fill left empty. These are Epoch AI's median estimates with published 5th/95th percentile bounds; the medians alone are shown because summing percentile bounds across chip types would imply a joint distribution nobody has computed. Accelerator Designer Cumulative units H100-equivalents Through B200 Nvidia 2,040,240 5,154,724 2026-03-31 B300 Nvidia 1,714,074 4,330,657 2026-03-31 H100 Nvidia 4,317,291 4,317,291 2026-03-31 TPU v6e Google 2,563,261 2,378,043 2026-03-31 TPU v7 Google 410,393 956,825 2026-03-31 Trainium2 Amazon 1,400,000 918,948 2025-12-31 Instinct MI300X AMD 526,066 695,104 2025-12-31 Ascend 910C Huawei 650,000 499,184 2025-12-31 TPU v5e Google 2,391,837 474,983 2026-03-31 TPU v5p Google 672,187 311,807 2026-03-31 A100 Nvidia 961,830 303,275 2026-03-31 H20 Nvidia 1,489,899 222,844 2026-03-31 Ascend 910B Huawei 600,000 208,784 2025-12-31 Instinct MI350X AMD 82,457 192,247 2025-12-31 Instinct MI355X AMD 48,765 124,025 2025-12-31 H800 Nvidia 115,430 115,430 2026-03-31 Instinct MI300A AMD 114,901 113,868 2025-12-31 Instinct MI325X AMD 79,239 104,103 2025-12-31 Trainium1 Amazon 445,000 94,441 2025-12-31 A800 Nvidia 121,421 38,285 2026-03-31 Siyuan 590 Cambricon 110,000 38,280 2025-12-31 TPU v4 Google 261,998 36,407 2026-03-31 Instinct MI308X AMD 30,646 10,123 2025-12-31 Instinct MI250X AMD 45,102 8,728 2025-12-31 TPU v4i Google 30,031 2,094 2026-03-31 H200 Nvidia — — no series published TPU v3 Google — — no series published Trainium3 Amazon — — no series published ### How to read a row #### Facility power Nameplate capacity of sites attributed to the company, summed. Nameplate is not measured draw and a site is rarely dedicated to training. Treat it as an upper bound on what the company could run, not an estimate of what it does. #### GPU inventory Accelerators disclosed at those sites. Where Epoch's GPU-cluster dataset has a per-cluster inventory it takes precedence over a facility total, because a per-model breakdown is stronger evidence than a headline count. #### Financial disclosure Capital expenditure from an SEC filing where one exists, revenue from a public annual report otherwise. The tier grades where the figure was read from, not how good the figure is: tier 1 is a primary filing (a regulator's host or the issuer's own) and tier 3 a public aggregator. Each figure names the host it resolves to. #### Export-control exposure Designations from the BIS lists matched to the company, read from the designation tables on each regeneration of this page rather than from a snapshot captured once at build. Every match has a confidence score and the method that produced it, both shown beside the designation. Matches below 0.85 confidence are excluded from this page — from the column, from the company count, and from the export — because a fuzzy name match printed the same way as an exact one asserts a designation the evidence does not support. [The register](/export-controls) holds the per-match chain.