Infrastructure research· Evidence checked 2026-10-22
Compare sourced estimates of compute used by AI labs, with uncertainty ranges and dated deployment evidence.
Historical benchmark ·
Research, training and inference capacity used or rented by each lab. Historical estimates, not today’s fleets; newer disclosures do not establish complete current totals.
H100-equivalents are a source-defined capacity index, not a physical GPU count. Read the unit definition.
5 of 5 labs with an estimate on this date
Historical estimates ·
90% interval: 1.25M–2.19M
90% interval: 1.01M–2.55M
90% interval: 842k–1.72M
90% interval: 606k–1.64M
90% interval: 587k–787k
H100-equivalents · linear scale from zero
Ranges overlap substantially. Sorting by median does not establish a reliable ranking. Missing lab-years are unrecorded, not zero; no dates between observations are filled in.
AI Compute Tracker (aicomputetracker.com), CC BY 4.0; includes data from Epoch AI (epoch.ai), CC BY 4.0. Source: Epoch AI’s frozen year-end dataset. 2026-09-09
Read each claim on its own date and scope. A contract, a site’s power, and a chip count measure different things and are not added to the benchmark.
Reported operating
Meta reported a completed first phase and hundreds of thousands of GPUs working across connected buildings at Prometheus. It did not publish an exact current GPU count or the lab's entire allocated fleet.
Engineering at Meta · 2026-10-19 →Planned capacity
Meta described approximately one million GPUs as Prometheus's full-scale goal. The target must remain separate from the smaller first phase reported as complete.
Engineering at Meta · 2026-10-19 →Contracted capacity
Anthropic agreed to use all Colossus 1 capacity, describing more than 300 MW and over 220,000 NVIDIA GPUs becoming accessible within May. The announcement is not a later confirmation of delivery or a whole-lab total.
Anthropic · 2026-05-06 →Contracted capacity
SpaceX announced an agreement giving Anthropic access to Colossus 1, a site with over 220,000 NVIDIA GPUs. SpaceX's owned infrastructure cannot therefore be assigned in full to its internal AI lab.
SpaceX · 2026-05-06 →Planned capacity
OpenAI said it had exceeded its goal of securing 10 GW of US AI infrastructure. Secured projects include future buildout; this does not establish that 10 GW was operating.
OpenAI · 2026-04-29 →Reported operating
OpenAI said GPT-5.5 was trained at Abilene on Oracle Cloud Infrastructure using NVIDIA GB200 systems. This confirms use of one site, without disclosing a whole-lab fleet total.
OpenAI · 2026-04-29 →Reported operating
Anthropic reported using over one million AWS Trainium2 chips for Claude training and serving. This is a chip count for one hardware family, not Anthropic's total compute or an H100 count.
Anthropic · 2026-04-20 →Contracted capacity
The agreement covers up to 5 GW of capacity over time. Anthropic expected nearly 1 GW of additional Trainium2 and Trainium3 capacity by the end of 2026; that target is not an operating-capacity confirmation.
Anthropic · 2026-04-20 →Contracted capacity
Anthropic announced a multi-gigawatt TPU agreement with Google and Broadcom, with capacity expected to start arriving in 2027. It is future contracted capacity, not an addition to the historical estimate.
Anthropic · 2026-04-06 →Historical disclosure
OpenAI reported roughly 1.9 GW of available compute in 2025. This power-capacity disclosure has its own scope; it is not a GPU count or an updated 2026 fleet total.
OpenAI · 2026-01-18 →A missing figure means the record is incomplete. It does not mean a lab has no compute. A parent company’s hardware fleet is not automatically assigned to its AI lab.
1.74M H100-equivalents · 2025-12-31 · estimate
Research, model training and inference attributed to OpenAI.
Mostly rented capacity. Epoch's scope excludes third-party Microsoft-hosted inference under its accounting treatment; this is not Microsoft's fleet.
90% interval: 1.25M–2.19M. Dataset source.
1.58M H100-equivalents · 2025-12-31 · estimate
DeepMind research and model inference within Alphabet.
Allocation from Google's fleet is modeled. External cloud rentals and other business uses are excluded; the boundary remains uncertain.
90% interval: 1.01M–2.55M. Dataset source.
1.19M H100-equivalents · 2025-12-31 · estimate
Claude research, training and inference, including partner-hosted inference paid for by Anthropic.
Rented capacity across providers. A cloud provider's total fleet is not an Anthropic allocation.
90% interval: 842k–1.72M. Dataset source.
996k H100-equivalents · 2025-12-31 · estimate
Frontier AI and basic AI research at Meta; the 2024 estimate covers predecessor teams.
The lab's share of Meta's fleet is modeled. Recommendation and advertising systems are outside the estimate's intended scope.
90% interval: 606k–1.64M. Dataset source.
635k H100-equivalents · 2025-12-31 · estimate
AI research, training and inference attributed to xAI, now SpaceXAI.
The benchmark predates 2026 rentals to external customers. Later SpaceX-owned capacity is not automatically capacity available to its own lab.
90% interval: 587k–787k. Dataset source.
Comparable whole-lab estimate not recorded
Whole-lab research, training and inference capacity.
No comparable whole-lab estimate is recorded in this reviewed benchmark. Unknown does not mean zero; a single training run or a parent company's fleet is a different scope.
Comparable whole-lab estimate not recorded
Whole-lab research, training and inference capacity.
No comparable whole-lab estimate is recorded in this reviewed benchmark. Unknown does not mean zero; a single training run or a parent company's fleet is a different scope.
Comparable whole-lab estimate not recorded
Whole-lab research, training and inference capacity.
No comparable whole-lab estimate is recorded in this reviewed benchmark. Unknown does not mean zero; a single training run or a parent company's fleet is a different scope.
Comparable whole-lab estimate not recorded
Whole-lab research, training and inference capacity.
No comparable whole-lab estimate is recorded in this reviewed benchmark. Unknown does not mean zero; a single training run or a parent company's fleet is a different scope.
Comparable whole-lab estimate not recorded
Whole-lab research, training and inference capacity.
No comparable whole-lab estimate is recorded in this reviewed benchmark. Unknown does not mean zero; a single training run or a parent company's fleet is a different scope.
Comparable whole-lab estimate not recorded
Whole-lab research, training and inference capacity.
No comparable whole-lab estimate is recorded in this reviewed benchmark. Unknown does not mean zero; a single training run or a parent company's fleet is a different scope.
Epoch normalizes a fleet's on-paper dense 8-bit performance (FP8 or INT8) to an NVIDIA H100. For chips without 8-bit support, it uses 16-bit performance. We reproduce that source-defined index; we do not convert it into the site's hardware FLOPS comparisons.
Source precision is retained below and in CSV downloads. Decimal places describe the source model’s output, not certainty about hardware. No interpolation or extrapolation is applied.
| Lab / estimate date | Median | 5th percentile | 95th percentile |
|---|---|---|---|
| OpenAI | 1743439.8493869188 | 1251477.0828997404 | 2185142.3455992746 |
| Google DeepMind | 1583429.9538668268 | 1008220.4725750112 | 2548755.702786412 |
| Anthropic | 1189573.8857074357 | 842241.6260268212 | 1716534.7037994058 |
| Meta Superintelligence Labs | 996228.0602346028 | 606301.475799089 | 1637712.6527012342 |
| SpaceXAI (xAI) | 634792.6058833487 | 587268.4845382306 | 786632.5928349933 |
| Google DeepMind | 420334.47667409095 | 214143.6362592907 | 775030.4133149517 |
| OpenAI | 386218.7101238488 | 276938.3521681225 | 486056.6393925586 |
| Meta Superintelligence Labs | 247706.7361251794 | 123667.98345628983 | 480076.0743148279 |
| Anthropic | 208559.98299904328 | 145189.55328493097 | 298183.1495345837 |
| SpaceXAI (xAI) | 203366.31755523162 | 171066.07146658262 | 237382.2215257479 |
| OpenAI | 102534.71096180622 | 70131.45706056355 | 136201.42880654646 |
The downloadable CSV is the numerical source of record for this page. Its SpaceXAI row differs from Epoch's introductory article; the two releases are not combined. The source contains no 2026 snapshot.
AI Compute Tracker (aicomputetracker.com), CC BY 4.0; includes data from Epoch AI (epoch.ai), CC BY 4.0. Figures reproduced from the source dataset; labels and presentation adapted. Upstream licence.
Source dataset SHA-256: a42673451e91a9896ed285ec74f66442473c07f9fedc89695a75ef3e4701169c
Inspired by Binkg’s lab-compute comparison. Figures from that post are not used as observations here. Updates require a source review; this page is not a live fleet census.
Explore the infrastructure ledger and compute contracts for individual sites and announced capacity.