Show me the gigawatts: the interactive views
Who uses, owns and is building the most AI compute — with every range kept on screen. Hover, tap or tab to any bar for its value, 90% range, date, source and confidence.
- Estimate / deployed
- Forecast / estimate
- Uncertainty range (90% CI or sourced floor/ceiling)
- Announced, not yet installed
- Earlier estimate
- Installed today / last dated value
View 1
Compute actually used by each lab
Show data table
| Lab | Median | 90% CI low | 90% CI high | As of | Source |
|---|---|---|---|---|---|
| OpenAI | 1,743,000 | 1,251,000 | 2,185,000 | 2025-12-31 | Epoch AI |
| Google DeepMind | 1,583,000 | 1,008,000 | 2,549,000 | 2025-12-31 | Epoch AI |
| Anthropic | 1,190,000 | 842,000 | 1,717,000 | 2025-12-31 | Epoch AI |
| Meta Superintelligence Labs | 996,000 | 606,000 | 1,638,000 | 2025-12-31 | Epoch AI |
| SpaceXAI (xAI) | 615,000 | 551,000 | 700,000 | 2025-12-31 | Epoch AI |
Sources: Epoch AI, “Introducing the AI Chip Users Explorer” (2026-09-09); Epoch AI, “Data on AI chip users” (2026-09-09)
View 2
Compute owned by each parent company
Note: Epoch’s page says Q1 2026 data was “not yet finalized”; Amazon and Meta’s Q1 2026 numbers are lower than Q4 2025. Treat Q1 2026 as preliminary. D2
Show data table
| Company | Q1 2026 (prelim.) | Share Q1 2026 | Q4 2025 | Share Q4 2025 | Interval |
|---|---|---|---|---|---|
| Alphabet / Google | 5,500,000 | 31% | 5,000,000 | 24.8% | none published |
| Microsoft | 3,300,000 | 18.9% | 3,400,000 | 16.8% | none published |
| Meta | 2,000,000 | 11.1% | 2,300,000 | 11.3% | none published |
| Amazon | 1,500,000 | 8.6% | 2,500,000 | 12.1% | none published |
| Oracle | 1,100,000 | 6.2% | 1,100,000 | 5.6% | none published |
Sources: Epoch AI, “Five hyperscalers now own over two-thirds of global AI compute” (2026-04-14)
Animated view
Who owns the world’s AI compute
Note: Q1 2026 shares are preliminary (Epoch: “not yet finalized”). At end-2025 the top five held 70.7% (Epoch headline “71%”). D2 · D3
Show data table
| Company | Share | H100e | As of | Note |
|---|---|---|---|---|
| Alphabet / Google | 31% | 5,500,000 | 2026-03-31 | point estimate (no interval published); Q1 2026 preliminary |
| Microsoft | 18.9% | 3,300,000 | 2026-03-31 | point estimate (no interval published); Q1 2026 preliminary |
| Meta | 11.1% | 2,000,000 | 2026-03-31 | point estimate (no interval published); Q1 2026 preliminary |
| Amazon | 8.6% | 1,500,000 | 2026-03-31 | point estimate (no interval published); Q1 2026 preliminary |
| Oracle | 6.2% | 1,100,000 | 2026-03-31 | point estimate (no interval published); Q1 2026 preliminary |
| Everyone else | 24.2% | — | 2026-03-31 | remainder: 100% − the five listed shares (75.8%), our arithmetic |
Sources: Epoch AI, “Five hyperscalers now own over two-thirds of global AI compute” (2026-04-14)
View 3
Largest single clusters
Note: The source doc converts Colossus 2’s 550K GPUs to ~1.4M H100e; Epoch describes ~1.4M H100e as xAI’s expansion target. We chart GPU counts only. D1
Show data table
| Cluster | Deployed | Announced / target | As of | Source |
|---|---|---|---|---|
| Colossus 2 (SpaceXAI) | 550,000 GB200/GB300 | +440,000 announced | 2026-10-01 | State of AI |
| Prometheus (Meta) | “hundreds of thousands” (exact count not disclosed) | ~1,000,000 GPUs | 2026-10-19 | Meta Engineering |
| Colossus 1 (SpaceXAI, rented to Anthropic) | >220,000 NVIDIA GPUs (>300 MW) | — | 2026-05-06 | Anthropic |
Sources: State of AI Report Compute Index (Nathan Benaich, Air Street Capital), “Who has the largest GPU cluster in this index?” (2026-10-01); Meta Engineering, “Inside Prometheus: Meta's Journey to a Gigawatt-Scale Supercluster” (2026-10-19); Anthropic, “Higher usage limits for Claude and a compute deal with SpaceX” (2026-05-06)
Animated view
Cluster scale: one square = 10,000 GPUs
Show data table
| Cluster | Reported | As of | Confidence | Source |
|---|---|---|---|---|
| Colossus 2 | 550,000 GPUs deployed + 440,000 announced | 2026-10-01 | operator-reported count (not an independent audit) | State of AI Report Compute Index (Nathan Benaich, Air Street Capital) — “Who has the largest GPU cluster in this index?” (2026-10-01) |
| Prometheus | “hundreds of thousands” deployed (no exact count); ~1,000,000 target | 2026-10-19 | operator-reported: 'hundreds of thousands' deployed (no exact count); ~1M GPU target | Meta Engineering — “Inside Prometheus: Meta's Journey to a Gigawatt-Scale Supercluster” (2026-10-19) |
| Colossus 1 | over 220,000 GPUs (floor) | 2026-05-06 | company-stated floor ('over 220,000 NVIDIA GPUs', >300 MW) | Anthropic — “Higher usage limits for Claude and a compute deal with SpaceX” (2026-05-06) |
Sources: State of AI Report Compute Index (Nathan Benaich, Air Street Capital), “Who has the largest GPU cluster in this index?” (2026-10-01); Meta Engineering, “Inside Prometheus: Meta's Journey to a Gigawatt-Scale Supercluster” (2026-10-19); Anthropic, “Higher usage limits for Claude and a compute deal with SpaceX” (2026-05-06)
View 4
Announced future capacity
Note: The doc says “30 GW by 2030 through Stargate.” Altman’s ~30 GW covers all partners (Stargate itself is 10 GW), and we found no 2030 date — so it is shown as undated. D4
Show data table
| Company | 2023 | 2024 | 2025 | 2026 | 2027 | 2028 | 2029 | 2030 | Announced, undated |
|---|---|---|---|---|---|---|---|---|---|
| OpenAI | 0.2 (actual) | 0.6 (actual) | 1.9 (actual) | 5–6 (forecast) | ≥10 (forecast) | — | — | — | ~30 GW total commitments (≈$1.4T), undated |
| Anthropic | — | — | 1.4 (estimate) | 5–6 (forecast) | — | — | — | — | ≥14.8 GW signed Oct 2025–Sep 2026 (+1–2 GW earlier), delivery over several years |
| Meta (Prometheus + Hyperion only) | — | — | — | 1 (forecast) | — | — | — | 3 (forecast) | Prometheus ~1 GW + Hyperion up to 5 GW 'over several years' |
Sources: Epoch AI, “Introducing the AI Chip Users Explorer” (2026-09-09); Epoch AI (Gradient Updates, opinion), “Frontier labs don't use most AI compute (yet)” (2026-05-20); Axios, “OpenAI wants to get to $1 trillion a year in infrastructure spend, Altman says” (2025-10-28); DatacenterDynamics, “Anthropic signed $517bn in compute agreements in past 11 months” (2026-09-07); Reuters (via Investing.com), “Anthropic's $518 billion AI buildout hinges largely on deals that cannot be canceled, filing shows” (2026-09-29); Meta Careers blog, “Why networking infrastructure is critical for AI innovation” (2025); TechCrunch, “Mark Zuckerberg says Meta is building a 5GW AI data center” (2025-07-14)
Animated view
Follow the watts: capacity over time
Show data table
| Company | 2023 | 2024 | 2025 | 2026 | 2027 | 2028 | 2029 | 2030 | Announced, undated |
|---|---|---|---|---|---|---|---|---|---|
| OpenAI | 0.2 (actual) | 0.6 (actual) | 1.9 (actual) | 5–6 (forecast) | ≥10 (forecast) | — | — | — | ~30 GW total commitments (≈$1.4T), undated |
| Anthropic | — | — | 1.4 (estimate) | 5–6 (forecast) | — | — | — | — | ≥14.8 GW signed Oct 2025–Sep 2026 (+1–2 GW earlier), delivery over several years |
| Meta (Prometheus + Hyperion only) | — | — | — | 1 (forecast) | — | — | — | 3 (forecast) | Prometheus ~1 GW + Hyperion up to 5 GW 'over several years' |
Sources: Epoch AI, “Introducing the AI Chip Users Explorer” (2026-09-09); Epoch AI (Gradient Updates, opinion), “Frontier labs don't use most AI compute (yet)” (2026-05-20); Meta Careers blog, “Why networking infrastructure is critical for AI innovation” (2025); TechCrunch, “Mark Zuckerberg says Meta is building a 5GW AI data center” (2025-07-14); Axios, “OpenAI wants to get to $1 trillion a year in infrastructure spend, Altman says” (2025-10-28); DatacenterDynamics, “Anthropic signed $517bn in compute agreements in past 11 months” (2026-09-07)
Unit explainer
What is an “H100-equivalent”?
H100e = H100-equivalents: a chip's peak dense 8-bit (FP8/INT8) throughput divided by an Nvidia H100's, multiplied by chip count (Epoch AI definition). For example, a Google TPUv7 does about 2.3× an H100’s 8-bit operations, so one TPUv7 counts as 2.3 H100e (Epoch AI).
Show data table
| Lab | 90% floor (squares) | Median (squares) | 90% ceiling (squares) |
|---|---|---|---|
| OpenAI | 125.1 | 174.3 | 218.5 |
| Google DeepMind | 100.8 | 158.3 | 254.9 |
| Anthropic | 84.2 | 119.0 | 171.7 |
| Meta Superintelligence Labs | 60.6 | 99.6 | 163.8 |
| SpaceXAI (xAI) | 55.1 | 61.5 | 70.0 |
Scorecard: what can and can’t be ranked
These dimensions measure different things. Don’t add them together.
| Dimension | Leader / position | Best available measure | Caveat |
|---|---|---|---|
| Compute actually used by a lab | OpenAI and Google DeepMind, effectively tied | End-2025 Epoch medians: OpenAI ~1.74M H100e; DeepMind ~1.58M. 90% intervals overlap. | OpenAI leads on the point estimate, not conclusively. |
| Compute owned by the parent | Alphabet / Google | ~5.5M H100e (prelim. Q1 2026; ~5.0M end-2025); Microsoft ~3.3M; Meta ~2.0M. | Serves Cloud customers and many internal products; not reserved for DeepMind. |
| Largest deployed single cluster | SpaceXAI, Colossus 2 | ~550,000 GB200/GB300 GPUs deployed (operator-reported); +440,000 announced. | A site is not a lab-wide fleet. Deployed ≠ announced. |
| Announced / contracted future capacity | OpenAI headline; Anthropic broadest multi-provider pipeline | OpenAI ~30 GW commitments; Anthropic ≥14.8 GW deals; Meta ~1 GW + up to 5 GW. | Different kinds of commitments and dates; not installed capacity. |
| User-serving inference capacity | Not publicly rankable | No public league table for tokens/sec, concurrent agent jobs, latency or quotas. | For users, availability and time-to-completion can matter more than cluster size. |
| Compute assigned to a named product | Not disclosed comparably | No audited accelerator-hours per Muse / Dots / Grok Bot / Claude / Spark user. | Any product-level ranking would be guesswork. |
| Usable throughput / efficiency | No single winner from H100e alone | H100e normalizes peak FP8 throughput. | Ignores utilization, interconnect, memory bandwidth, software, workload. |
Model it yourself
The Compute Brief tools add a scenario modeler, a lab-vs-lab comparator, an embed kit and a briefing deck — all on this same data. Or grab the free sample CSV.