Global · AI labs
AI lab economics: who earns what, what they are worth, and what each GW brings in
Anthropic's annualised revenue reached $65bn at the end of Jul 2026 (Bloomberg, reported), passing OpenAI's ~$50bn (Financial Times, Sep 2026, reported). Neither publishes audited accounts; both figures are unchecked against company filings. The one US lab with filed numbers, xAI inside SpaceX, earned $2.56bn in Q2 2026 and lost $1.26bn from operations (10-Q, 4 Aug 2026). Europe's largest lab, Mistral, is at about $1bn (Epoch AI, 2026-09-08).
Updated 11 October 2026 · data build 2026-10-10 · how we know · corrections
Anthropic's run-rate ($65bn) has passed OpenAI's (~$50bn)
Annualised revenue as stated by each company or reported by the press, last report in each quarter. Not audited. Reported
Source: Epoch AI, Data on AI companies (CC BY 4.0) · as of 2026-10-11 · original reports: The Information, Financial Times, New York Times, Wall Street Journal, CNBC, Anthropic, Bloomberg, X post · Compiled by StackPath
The two labs may not count revenue the same way: whether sales through cloud partners (Azure, AWS Bedrock, Google Vertex) are counted gross or net can move a run-rate by tens of billions. A circulating ~$70bn figure for OpenAI is disputed; we use the ~$50bn reported by the Financial Times.
US labs: $125bn of run-rate; China's four: $4.4bn; Europe's largest: $1bn
Sum of each lab's latest reported annualised revenue. Google DeepMind and Meta disclose no model revenue and are not counted. Reported
Source: Epoch AI, Data on AI companies (CC BY 4.0) · SpaceX 10-Q (xAI) · as of 2026-10-11 · original reports: Wall Street Journal, Financial Times, Yahoo Finance, The Information, Bloomberg, CNBC, HKEX filing, South China Morning Post · Compiled by StackPath
Our reading. Mistral's ~$1bn is about 1/65 of Anthropic's run-rate, and its last valuation ($24bn) about 1/40 of Anthropic's. China's four labs together earn less than xAI's segment ($4.4bn against $10.2bn). The model-revenue race is, on these numbers, between two US companies. Europe's full lab list, with funding and compute, is on European AI labs.
Each GW of compute brings in about $7–11bn of revenue a year
Revenue divided by the compute the lab says it had. OpenAI from its CFO (reported); xAI from SpaceX's 10-Q (filed figures, annualised by us, so an estimate). Our reading
| Lab | Period | Revenue $bn | GW | $bn per GW | Source |
|---|---|---|---|---|---|
| OpenAI | 2023 | 2 | 0.2 | 10 | OpenAI (CFO Sarah Friar) Reported |
| OpenAI | 2024 | 6 | 0.6 | 10 | OpenAI (CFO Sarah Friar) Reported |
| OpenAI | 2025 | 20 | 1.9 | 10.5 | OpenAI (CFO Sarah Friar) Reported |
| xAI | 2025-Q2 | 3 | 0.4 | 7.4 | SpaceX 10-Q Estimate |
| xAI | 2026-Q2 | 10.2 | 1.4 | 7.3 | SpaceX 10-Q Estimate |
Bases. OpenAI: revenue is annualised revenue (ARR) at year end, as stated; GW is compute available at year end, as stated. OpenAI: revenue is annualised revenue (ARR) at year end, as stated ('$20bn+'); GW is compute available at year end, as stated. xAI: revenue is quarterly segment revenue × 4 (includes X advertising and compute rented to others); GW is nameplate compute draw at quarter end.
Download CSV ↓Source: OpenAI CFO, 2026-01-18 · SpaceX 10-Q, Q2 2026 · as of 2026-10-11 · Compiled by StackPath
xAI's segment revenue includes X advertising ($0.37bn in Q2 2026) and about $1.6bn of new revenue from renting its compute to other companies ("cloud services", which began in 2026). One unnamed customer of the AI segment ("Customer B") was 19.5% of all SpaceX revenue in Q2 2026. Without the rental income, xAI's own products earn less per GW than the figure above.
On press reports, OpenAI and Anthropic spent more on compute than they earned
Revenue for the calendar year against reported spending on training and inference compute. All figures press-reported, via Epoch AI. Reported
| Lab | Year | Revenue $bn | Training $bn | Inference $bn | All compute $bn | Compute ÷ revenue | Reported by |
|---|---|---|---|---|---|---|---|
| OpenAI | 2024 | 3.7 | 4 | 1.8 | 5.8 | 1.6× | Bloomberg, The Information |
| OpenAI | 2025 | 13 | 8.3 | 8 | 16.3 | 1.3× | New York Times, The Information |
| Anthropic | 2024 | 0.4 | 1.5 | – | 2.5 | 6.6× | The Information |
| Anthropic | 2025 | 4.5 | 4.1 | 2.7 | 6.8 | 1.5× | The Information |
All compute is training plus inference where both are reported, otherwise the reported total. Epoch's lower-confidence rows (analyst estimates it hides from its own charts) are left out.
Download CSV ↓Source: Epoch AI, Data on AI companies (CC BY 4.0) · as of 2026-10-11 · original reports mostly The Information, CNBC and the New York Times · Compiled by StackPath
Our reading. Microsoft recorded $24.1bn of revenue "from commercial arrangements with OpenAI, inclusive of revenue-sharing payments" in the year to June 2026 (10-K, 29 Jul 2026). Summing OpenAI's press-reported run-rates over the same months gives about $21bn of OpenAI revenue (our estimate; the same method gives $7.8bn for H2 2025 against $8.7bn reported, so it may run about 10% low). The two figures are of similar size, but they are not like for like, so this does not show that OpenAI paid Microsoft its whole revenue. Microsoft's figure is revenue it recognised, not cash received ($6bn was still owed at 30 June); the filing does not split cloud services from the revenue share; part of OpenAI's cloud spending may have been paid for with money Microsoft invested ($11.9bn funded to date of a $13bn commitment); and OpenAI's side is an estimate from unaudited, press-reported run-rates that may count Azure sales differently. Microsoft's line for its OpenAI stake moved from a $4.8bn net loss in FY2025 to a $6.5bn net gain in FY2026, which it attributes mainly to a dilution gain from OpenAI's October 2025 recapitalisation.
xAI lost $3.7bn from operations in six months; its gross margin rose to 57%
SpaceX's AI segment (xAI, Grok and X) by quarter, from its 10-Q. Q1 figures are the half-year less Q2. Measured
| Quarter | Revenue $bn | Cost of revenue $bn | Gross margin | R&D $bn | Operating loss $bn |
|---|---|---|---|---|---|
| 2025-Q1* | 0.73 | 0.45 | 38% | 0.91 | −0.94 |
| 2025-Q2 | 0.74 | 0.55 | 25% | 1.12 | −1.52 |
| 2026-Q1* | 0.82 | 0.46 | 44% | 2.38 | −2.47 |
| 2026-Q2 | 2.56 | 1.11 | 57% | 2.18 | −1.26 |
* derived: six months less the second quarter. R&D includes training Grok and data-centre costs for training.
Download CSV ↓Source: SpaceX 10-Q, quarter to 30 Jun 2026 (segment note) · as of 2026-08-04 · Compiled by StackPath
| Lab | Measure | Period | Value | Source | Kind |
|---|---|---|---|---|---|
| OpenAI | Adjusted gross margin (company-defined) | 2025 | 33%, down from 40% in 2024; target was 46% | The Information, via The Decoder, 2026-02 | Reported |
| Anthropic | Adjusted operating income (company-defined) | Q2 2026 | Positive, on revenue of more than $11.5bn | Bloomberg, 2026-08-14 | Reported |
| MiniMax | Gross margin; loss | H1 2026 | Gross margin 17.9% (12.1% a year earlier); net loss $358.0m on revenue of $116.6m | MiniMax interim results, 2026-08-26 | Reported |
| Z.ai (Zhipu) | Loss | H1 2026 | Net loss of about RMB2.0bn on revenue of RMB953.9m | The Standard (Hong Kong), 2026-08-31 | Reported |
Company-defined, unaudited measures (adjusted gross margin, adjusted operating income) are not comparable across labs, or with the xAI segment above.
Investors price the leaders at 15–17× run-rate
Post-money valuation at the last priced round (as announced by the company or reported, via Epoch AI), divided by the press- or company-reported run-rate before it. For listed labs, market value. Our reading
| Lab | Region | Latest valuation | Date | Latest run-rate | Multiple | Equity raised (Epoch) |
|---|---|---|---|---|---|---|
| OpenAI | US | $852bn | 2026-03-31 | $50bn | 17× | $183bn |
| Anthropic | US | $965bn | 2026-05-28 | $65bn | 15× | $139bn |
| xAI | US | $230bn | 2026-01-06 | $10.2bn (segment, Q2 ×4) | – | $37bn |
| Mistral AI | Europe | $24bn | 2026-09-15 | $1bn | 24× | $6.5bn |
| Cohere | Canada | – | – | $0.24bn | – | – |
| DeepSeek | China | $50bn | 2026-06-16 | $1bn | 50× | $7.5bn |
| Moonshot AI | China | – | – | $1bn | – | – |
| Z.ai (Zhipu) | China | $39.4bn | 2026-03-31 (market) | $1.6bn | 25× | $0.6bn |
| MiniMax | China | $37.2bn | 2026-03-31 (market) | $0.8bn | 47× | $0.6bn |
xAI's valuations include X and its tracked run-rates do not, so no multiple is shown. Equity raised sums Epoch's tracked rounds (it misses some early private rounds for the Chinese labs).
Source: Epoch AI, Data on AI companies (CC BY 4.0) · Amazon 10-Q (Anthropic stake) · Microsoft 10-K (OpenAI stake) · as of 2026-10-11 · Compiled by StackPath
The suppliers are also the owners. Amazon carried its Anthropic stake at $190.4bn on 30 June 2026 (preferred stock plus convertible notes), up from $60.6bn six months earlier, and committed $50bn to OpenAI's preferred stock. Microsoft holds about 25% of OpenAI and has funded $11.9bn of a $13bn commitment (filings above). Alphabet is also an Anthropic investor, but its 10-Q does not name or value the stake (it reports $124.3bn of non-marketable equity securities in total at 30 June 2026; 10-Q). See the compute deals for what the labs owe them back.
What each lab discloses
Disclosure varies: one segment in an SEC filing, two Hong Kong listings, and press reports for the rest. Reported
| Lab | Status | What is disclosed | Latest run-rate |
|---|---|---|---|
| OpenAI US | Private (public benefit corporation since Oct 2025) | No public audited accounts. Revenue figures are company statements or press reports; Microsoft's 10-K gives what OpenAI paid Microsoft. | $50bn (2026-09-30, Financial Times, press report) |
| Anthropic US | Private (public benefit corporation) | No public audited accounts. Run-rates mostly stated by the company; Amazon's 10-Q gives the value Amazon carries its stake at. | $65bn (2026-07-31, Bloomberg, press report) |
| xAI US | AI segment of SpaceX (listed June 2026) | Quarterly segment P&L in SpaceX's 10-Q (measured). The segment includes X (formerly Twitter) advertising and compute rented to other companies. | $10.2bn (segment, Q2 ×4) |
| Google DeepMind US | Part of Alphabet | No model revenue disclosed. Alphabet reports Google Cloud as a whole. | not disclosed |
| Meta (Superintelligence Labs) US | Part of Meta | No external model revenue; models are used inside Meta's own products. | not disclosed |
| Mistral AI Europe | Private | Annual recurring revenue stated by the company to press. | $1bn (2026-09-08, Wall Street Journal, company statement) |
| Cohere Canada | Private | Run-rates from press and company statements. | $0.24bn (2025-12-31, CNBC, press report) |
| DeepSeek China | Private (High-Flyer) | Run-rates from press reports only. | $1bn (2026-09-24, The Information, press report) |
| Moonshot AI China | Private | Run-rates from press reports. | $1bn (2026-08-31, Bloomberg, press report) |
| Z.ai (Zhipu) China | Listed in Hong Kong (Jan 2026) | Half-year results filed in Hong Kong (interim, unaudited); ARR stated by the company. | $1.6bn (2026-08-31, South China Morning Post, company statement) |
| MiniMax China | Listed in Hong Kong (Jan 2026) | Half-year results filed in Hong Kong (interim, unaudited); ARR stated by the company. | $0.8bn (2026-08-26, South China Morning Post, company statement) |
Source: Epoch AI, Data on AI companies (CC BY 4.0) · as of 2026-10-11 · each row links its original publisher · Compiled by StackPath
Our reading. The money is real and growing fast, but every lab that files accounts (xAI in the US, Z.ai and MiniMax in Hong Kong) shows losses, and the two leaders' figures are self-reported and defined differently. On the cost side, the one filed figure, Microsoft's $24.1bn of revenue from OpenAI in a year, is of the same order as our estimate of OpenAI's own revenue, though the two are not measured the same way (see above). What the labs have committed to spend next is on AI compute commitments; whether revenue can catch up with the build-out is on the AI revenue gap.
Run-rates are the latest month (or week) × 12, as stated by each company or reported by the press; Epoch AI marks each as company disclosure or media report, and we keep that label. None is audited. OpenAI and Anthropic may count revenue from cloud partners (Azure, Bedrock, Vertex) differently (gross or net), so their run-rates are not strictly comparable. xAI is reported as SpaceX's AI segment, which also contains X (advertising and subscriptions) and, from 2026, compute rented to other companies. Epoch's earlier xAI figures are for xAI alone and are not comparable with the segment. Valuation multiples divide a post-money valuation by the latest run-rate reported before the round. Post-money valuations are set by the last investors and are not market prices; for the Hong Kong-listed labs we use Epoch's market value.
Questions people ask
- Is Anthropic bigger than OpenAI by revenue?
- On annualised run-rate, yes, as reported: Anthropic $65bn at end-Jul 2026 (Bloomberg) against OpenAI's ~$50bn at end-Sep 2026 (Financial Times). Over calendar 2025 OpenAI still booked more (about $13bn against Anthropic's $4.5bn, both press-reported). Run-rates are company or press figures, not audited accounts, and the two may count cloud-partner sales differently.
- What is run-rate (annualised) revenue?
- The latest month's (or week's) revenue multiplied up to a year. It shows the current pace, not what was earned over a year, and it overstates the year's revenue when a company is growing fast. ARR (annual recurring revenue) counts only subscriptions and contracts that recur.
- How much revenue does an AI lab make per gigawatt of compute?
- About $7–11bn a year per GW on the numbers disclosed: OpenAI's CFO gave revenue and GW for 2023–2025 (reported), and SpaceX's 10-Q gives xAI's segment revenue and 1.4 GW of compute draw (measured; revenue annualised by us). SemiAnalysis argues API inference on new GB300 racks can earn far more per GW; whole-company figures include training compute that earns nothing directly.
- Are AI labs profitable?
- Not on the numbers filed. xAI's segment lost $3.73bn from operations in H1 2026 (SpaceX 10-Q). OpenAI and Anthropic publish no accounts; press reports say OpenAI's adjusted gross margin fell in 2025 and that Anthropic reached an adjusted operating profit in Q2 2026. Those are company-defined, unaudited measures.
- How do European AI labs compare with US and Chinese labs?
- Mistral, Europe's largest model lab, reported about $1bn of annual recurring revenue (Sep 2026), against $125bn summed for OpenAI, Anthropic and xAI and $4.4bn for four Chinese labs (DeepSeek, Moonshot, Z.ai, MiniMax). See /europe/ai-labs for Europe's labs in full.