Who owns the most NVIDIA H100 GPUs?
Microsoft owns the most NVIDIA H100 GPUs: an estimated 1,093,460 H100 and H200 chips at the end of 2025, according to Epoch AI. That is 25.7% of the 4,253,723 delivered to all owners. Meta is a distant second with 594,470, followed by Amazon (451,830), Google (423,372), the cloud provider CoreWeave (301,187), Oracle (284,484) and xAI (200,000). Another 904,920 belong to smaller clouds, governments, enterprises and labs that Epoch does not name.
These are estimates, not company disclosures. None of the big cloud companies reports its GPU count. Epoch builds its figures from NVIDIA's revenue and customer disclosures, analyst research and press reports, and publishes a range around each. Microsoft's 90% interval runs from 834,492 to 1,444,563, but even its low end is above Meta's high end of 774,476.
How many H100 GPUs does Microsoft have?
Microsoft had an estimated 1,093,460 H100 and H200 GPUs at the end of 2025, after passing 1 million in the first quarter of that year. It took the lead early: Epoch puts Microsoft at 235,326 at the end of 2023, almost twice Meta's 123,565. At a time when the chips were scarce, Omdia had the two companies tied at about 150,000 each for 2023. Microsoft rents compute to OpenAI, but Epoch credits chips to the company that owns them, so OpenAI does not appear in the ranking.
Epoch puts the rated power of Microsoft's H100 and H200 chips alone at 765,422,243 watts, or about 765 megawatts, using the 700-watt rating both chips share. A whole data center draws roughly twice its chips' rated power once cooling and servers are counted.
How many H100 GPUs does Meta have?
Meta had an estimated 594,470 H100 and H200 GPUs at the end of 2025, according to Epoch AI. In March 2024, Meta said its build-out by the end of that year would include 350,000 NVIDIA H100s within compute equivalent to nearly 600,000 H100s. Epoch's estimate for the end of 2024 is higher on both counts: 488,451 H100s and H200s, and 774,426 H100-equivalents of NVIDIA and AMD compute.
How many H100 GPUs are there in the world?
By the sum of Epoch AI's owner estimates, 4,253,723 H100 and H200 GPUs had been delivered by the end of 2025. Deliveries peaked at 777,098 in the third quarter of 2024. By the first quarter of 2025, NVIDIA was shipping more Blackwell B200s (502,385) than H100s and H200s (374,410), and by the last quarter of 2025 Hopper deliveries had fallen to 22,302, against 1,144,024 B200s and B300s.
Does Google use H100 GPUs or TPUs?
Google uses both, but mostly its own TPUs. Google owned an estimated 423,372 H100 and H200 GPUs at the end of 2025, fourth among the companies Epoch names. Counted by all AI compute, Google ranks first with 5,049,814 H100-equivalents, and 3,812,878 of them (75.5%) come from TPUs. Amazon's own Trainium chips add 1,020,727 H100-equivalents to its total in the same way.
How many GPUs does xAI have?
xAI had 200,000 H100 and H200 GPUs at its Colossus 1 data center in Memphis from early 2025, according to Epoch AI: 150,000 H100s and 50,000 H200s. Across Colossus 1 and Colossus 2, including chips newer than the H100, Epoch counts xAI's total NVIDIA compute at 553,714 H100-equivalents at the end of 2025. xAI is now part of SpaceX, and Epoch reports that Anthropic agreed in May 2026 to use all of Colossus 1's compute.
Methodology & caveats
Company figures are Epoch AI's median estimates from its AI chip owners dataset (estimates generated April 2026), read from its published CSV files. Epoch reports H100 and H200 GPUs as one category, so every "H100" count here from Epoch includes H200s. Counts are cumulative deliveries based mainly on chip vendors' recognized revenue, with nothing subtracted for retired chips; deliveries can precede installation by weeks or months. xAI's figure comes from operating capacity at its data center instead. Epoch's NVIDIA-based estimates follow NVIDIA's fiscal quarters, which end about a month after calendar quarters; Epoch flags its Q1 2026 figures as incomplete for that reason (they rest on the quarter ended Jan 25, 2026), so they are not used.
Totals, shares and compute sums (4,253,723 GPUs, 25.7%, 75.5%, each company's H100-equivalents) are derived by adding or dividing Epoch's published medians. H100-equivalents scale each chip by its peak dense 8-bit throughput relative to an H100. Ownership is attributed informally: chips rented by AI labs such as OpenAI and Anthropic count toward the cloud that owns them, and some data centers are jointly owned. Chinese buyers received export-controlled versions (H800, H20) rather than H100s; Epoch's separate estimate of smuggled chips is not included. Omdia's 2023 figures are cited as reported by Tom's Hardware because Omdia's report is not public, and they cover a different period and definition than Epoch's.
Sources
- [1]Epoch AI. AI chip owners: cumulative units by owner and chip type (ai_chip_owners_cumulative_by_chip_type.csv)https://epoch.ai/data/ai_chip_owners_cumulative_by_chip_type.csv
- [2]Epoch AI. AI chip owners: cumulative H100-equivalent compute by owner and chip designer (ai_chip_owners_cumulative_by_designer.csv)https://epoch.ai/data/ai_chip_owners_cumulative_by_designer.csv
- [3]Epoch AI. AI chip owners documentation: FAQ (definitions of ownership, H100e and the China category)https://epoch.ai/data/ai-chip-owners-documentation/faq
- [4]Epoch AI. Frontier AI data centers: Colossus 1, Memphis, Tennessee (updated Jun 16, 2026)https://epoch.ai/data/data-centers/satellite-explorer/xAIColossus1
- [5]Epoch AI. AI chip owners: units delivered per quarter by owner and chip type (ai_chip_owners_quarters_by_chip_type.csv)https://epoch.ai/data/ai_chip_owners_quarters_by_chip_type.csv
- [6]Omdia (via Tom's Hardware). Nvidia sold half a million H100 AI GPUs in Q3 thanks to Meta, Facebook (Tom's Hardware, Nov 28, 2023)https://www.tomshardware.com/tech-industry/nvidia-ai-and-hpc-gpu-sales-reportedly-approached-half-a-million-units-in-q3-thanks-to-meta-facebook
- [7]Meta. Building Meta's GenAI Infrastructure (Engineering at Meta, Mar 12, 2024)https://engineering.fb.com/2024/03/12/data-center-engineering/building-metas-genai-infrastructure/


