Unique Research · Global AI Company Revenue Ranking · October 23, 2026
Application-layer companies are the most numerous — yet the money is flowing first to the two ends closest to models and compute.
Unique Research · Global AI Company Revenue Ranking
$148.7 billion: the top two take 79%; No. 87 has only just crossed $100 million. Eight Chinese companies make the list, totaling $3.32 billion — 2.2% of the ranking.
Every figure below comes from our newly re-verified Global AI Company Revenue Ranking:
$148.7 billion: estimated combined revenue of the Top 100.
79%: Anthropic and OpenAI alone take $117 billion; the other 98 companies split $31.7 billion.
87 companies have crossed $100 million in revenue. A few years ago, $100 million ARR was a star-line milestone. On this ranking, it is only No. 87.
First, the caliber — so nobody argues later. This ranking covers global unlisted AI-native companies. Revenue combines annual revenue, ARR, and annualized run rates, drawn from company disclosures, media reports, and our estimates. Figures are converted to USD billions and ranked on unrounded amounts; “over / at least” is shown at the lower bound. Disclosure periods and definitions are not uniform across companies — this is not same-fiscal-year actual revenue, and it is not market share. Cutoff date: October 22, 2026.
This edition’s No. 1 is Anthropic, estimated at $67 billion, on a quarterly-annualized basis: Q3 revenue of $16.8 billion, up nearly 50% from Q2’s $11.6 billion, after Q1’s $4.7 billion. Two consecutive quarters of adjusted profitability, adjusted gross margin of 85%, and free cash flow turning positive.
No. 2 is OpenAI, at roughly $50 billion annualized. When it briefed investors in early October, that figure was about $20 billion below the nearly $70 billion that had been circulating. One caliber adjustment can become a finance headline for a simple reason: the industry’s valuations sit on “annualized revenue.” Caliber is everything.
That is why we keep disclosure definition and date as two columns on the ranking. Anthropic’s $67 billion is quarter × 4; OpenAI’s $50 billion is a late-September annualized run rate. The denominators are not identical.
More important than rank is capital-market action. Anthropic may file its S-1 publicly with the SEC as early as October 23, chasing what could be the largest IPO on record, with roadshows as early as November. Wall Street expects a raise that could reach $100 billion, with valuation expectations cut from $2 trillion toward the $1.5 trillion range.
OpenAI is discussing roughly $30 billion in new financing, with a valuation target cut from $1.4 trillion to $1.25 trillion, and listing pushed to 2027.
On the product side, Anthropic shipped the proactive agent Claude Duet on October 21 — colliding head-on with OpenAI’s Dots and Meta’s Muse in the personal-agent lane.
If Anthropic lists on schedule, this will be its last appearance on an unlisted ranking. Strip out the top slot and No. 1 becomes Anduril at $2.2 billion — a 97% collapse in the headline number.
Incidentally, since October alone, 15 companies on the ranking have updated revenue figures — even Mira Murati’s Thinking Machines disclosed in October that its annualized run rate had crossed $100 million. With financing and IPO windows open, revenue numbers are the best roadshow material.
Take the two mountains off the table and the remaining 98 companies split $31.7 billion. Where that $31.7 billion goes is this edition’s most valuable finding.
Companies that supply training data and expert labor to models — 6 firms totaling $5.68 billion: Mercor over $2 billion, Surge AI $1.2 billion, Scale AI listed at a $1 billion ceiling, Micro1 $1 billion, Snorkel AI $375 million, AfterQuery over $100 million.
Mercor is the clearest sample. Its CEO disclosed in July that annualized run rate had broken $2 billion. Unpack that: the run rate does not deduct contractor and similar costs — it is not net revenue — and demand is highly concentrated on a few frontier labs. If those labs build data capacity in-house, the orders can reverse.
Scale AI needs its own note: the circulating $2 billion was its April 2025 expectation for that year’s revenue. After re-verification we switched to actuals — 2025 actual revenue was close to but under $1 billion, so we list it at the upper bound as No. 14.
The other line is inference cloud and compute platforms — 10 firms totaling $5.53 billion: Modal, Fireworks AI, and Together AI at roughly $1 billion each; Baseten $600 million; Fal about $400 million; Runpod $325 million; Verda $165 million; OpenRouter $140 million; DeepInfra and Prime Intellect at about $100 million each.
As model call volume rises, tokens are the unit of pricing. In the same press release where DeepInfra disclosed ARR crossing $100 million, token processing volume was up 3×. AMD CEO Lisa Su simply called 2026 “the year of inference,” expecting inference to account for about 60% of AI compute.
The two lines together are about $11.2 billion — 35% of combined revenue from Nos. 3 through 100. Application-layer companies are the most numerous, yet the money is flowing first to the two ends closest to models and compute.
One more caveat on stale data: Surge AI’s $1.2 billion is 2024 revenue; Fal’s $400 million and Together AI’s $1 billion are third-party annualized estimates; VAST Data’s $500 million is committed annual recurring revenue (CARR), not yet recognized on the books. Given industry growth rates, some of these present-day figures are probably higher — but the caliber has to be stated first.
Coding is the hardest application track. Cognition ($1 billion), Lovable ($600 million), Emergent ($120 million), Augment ($100 million), Factory ($60 million) — five firms totaling $1.88 billion.
Cognition’s official blog disclosed May annualized run rate of about $492 million, breaking $1 billion in September — doubling in four months. McKinsey’s late-August 2026 AI survey supplied demand-side evidence: 32% of surveyed enterprises abandoned buying at least one software product because they could build it themselves with AI coding tools.
AI coding is not only creating new revenue — it is eating traditional software procurement budgets.
Customer-service and voice agents collectively crossed $100 million. Sierra over $200 million; Decagon, Crescendo, and Cresta at about $100 million each; Retell $60 million — five firms totaling about $560 million.
Individual scale is modest, but “collective arrival” is itself a signal: customer service is one of the enterprise use cases where buyers write checks most readily.
Legal, healthcare, and finance verticals — 7 firms totaling about $1.07 billion: Harvey and OpenEvidence at about $300 million each; Legora over $200 million; Abridge about $117 million; Heidi, Hebbia, and Rogo at about $50 million each.
But a sword hangs over this track: according to The Information, as Hebbia and Rogo grow revenue, Claude’s finance-oriented products are closing in. Model vendors building vertical apps themselves is this layer’s largest uncertainty.
Content generation has the longest long tail — 11 firms totaling about $3.5 billion: Higgsfield $1 billion, ElevenLabs nearly $600 million, Kling AI $500 million, Suno and Midjourney $300 million each, Runway and HeyGen $200 million each, Synthesia $150 million, Black Forest Labs and Meshy $100 million each, Pika $85 million.
Higgsfield deserves its own unpack. Its $1 billion is a “last four weeks × 13” run rate — from zero to $1 billion in 18 months; enterprise-customer contract revenue up 10× since June; subscription share down from over 90% to just over 60% as pay-as-you-go expands; the company says cash flow has roughly broken even in recent months.
Customers are no longer buying a one-off generation — they are buying ongoing ad-creative production capacity.
Physical AI is past “just getting started.” Leading the pack is Anduril: Reuters confirmed in July that its actual 2025 revenue was $2.2 billion, putting it straight at No. 3 overall — and the only full-caliber actual annual revenue figure in the top five.
Skild, Microagi, Mecka, and WorldEngine all disclosed in September–October that annualized revenue had just hit $100 million; drone-maker Skydio’s source wording was “hundreds of millions,” so we list it conservatively at $100 million.
Skild’s sample is the most complete: about 10 months from first commercial deployment to $100 million ARR, roughly $50 million of recognized revenue, 60+ paying customers across warehousing, delivery, security, and factory scenes.
Eight Chinese companies total $3.32 billion — 2.2% of the full ranking, and 10.5% of Nos. 3–100.
Moonshot (月之暗面) at about $1 billion, No. 6. Per Bloomberg in September, its ARR rose from about $300 million in June to over $1 billion in August — more than 3× in two months — driven mainly by July’s Kimi K3, with an internal year-end target of $2 billion.
It has confidentially filed with the Hong Kong Stock Exchange, planning to raise about $3 billion, with private-market valuation talked up to $50 billion.
DeepSeek at about $1 billion, No. 10. Per The Information, its annualized run rate was still under $500 million a few months ago. After an August API price hike of 2.3× to 4.5×, Liang Wenfeng told investors there was no material customer churn; revenue is almost entirely from API; the chat app is free with no ads; API gross margin through July was about 82.9%.
The company is advancing a raise of about RMB 50 billion, at a valuation of RMB 500 billion, has selected CITIC Securities as sponsor, and is preparing a Shanghai STAR Market listing. Revenue for the first seven months was about RMB 475 million — roughly 10× last year’s full-year figure.
Against an industry-wide price war, a leading model vendor has shown clear pricing power for the first time. If that ability holds, the cost structure of the entire chain will be rewritten.
Then on October 22, the U.S. Department of Homeland Security ordered federal civilian agencies to fully remove DeepSeek; the same day, the U.S. Treasury proposed new rules restricting U.S. capital investment in Chinese AI and chip companies. Making the ranking and being banned — in the same month.
Behind the two model labs sit companies selling products to global users.
Kling AI (可灵AI) at about $500 million, No. 22. This is the unlisted entity under Kuaishou that completed a carve-out and independent financing — Kuaishou still retains control. Official May disclosure put March annualized run rate near $500 million; August disclosure put Q2 single-quarter revenue above RMB 850 million.
Note: $500 million is the March run rate. At this climb rate, the present figure is probably higher.
Yanyu Technology (演语科技) over $300 million, No. 31. LiblibAI, Lovart, LibTV, and Xingliu all sit under this company; the founder confirmed overall ARR above $300 million in a LatePost interview.
PLAUD at $180 million, No. 47. A Shenzhen team’s AI recording card — we use the realized $180 million run rate and do not count the company’s own $250 million target. On a16z’s U.S. consumer revenue ranking it sits at No. 16 — a textbook hardware-going-global sample.
Haiyi AI (海艺AI) over $100 million, No. 55. Disclosed October 22: coverage across 200+ countries and regions; SeaArt cumulative registered users over 70 million, with about 3 million original character models accumulated. New core product SeaHaze combines character interaction, story exploration, and emotional companionship, and forms a suite with MoreShort and SeaSoul.
Add Kuaizi Tech (筷子科技) (No. 50, nearly RMB 1 billion ARR) and Ruiqi Software / PictureThis (No. 85, $100 million), and of the eight Chinese companies only two are model labs — the other six are all applications and hardware.
On video/image, content production, and AI hardware, Chinese AI has seen money earlier than on foundation models.
One more caliber note: China’s highest-revenue model companies — Zhipu (August ARR about $1.6 billion) and MiniMax (about $800 million) — are already listed in Hong Kong and fall outside this ranking. Add the four together at about $4.4 billion and you still have less than 7% of Anthropic’s quarterly-annualized revenue.
MiniMax broke below issue price on October 22; market-cap P/ARR about 9×, Zhipu about 18×, while the market gives Anthropic 22–29×.
Country distribution is even harsher than revenue distribution. The United States: 75 companies totaling about $142.1 billion — 96%. Even stripping out the two mountains, the remaining 73 U.S. companies still take about 79% of Nos. 3–100 revenue.
Europe: 10 companies totaling about $2.52 billion — Sweden about $800 million via Lovable and Legora; France about $750 million via Mistral AI and Dataiku; three UK firms totaling about $650 million; two German, one Finnish.
Canada: 3 companies, led by Cohere at $240 million. Europe’s bright spots are almost entirely at the application layer; the only foundation-model name is Mistral.
Revenue power law is more extreme than people think — and will not soften soon. The top two take 79%; the other 98 combined equal only 47% of Anthropic alone. Model capability is still iterating; revenue will keep concentrating at the strongest model labs.
$100 million ARR is becoming the new entry ticket — and the time-to-arrive keeps shrinking. Eighty-seven companies have crossed the line; Microagi took only six months, Crescendo under two years. Fast starts mean denser competition. The real watershed will be the $1 billion tier — currently only 14 companies.
Money flows first to wherever is closest to compute and models. Data services plus inference cloud total about $11.2 billion — 35% of non-top revenue.
The application layer is being squeezed from both ends. Upstream, model vendors are starting to raise prices — DeepSeek is the first sample. Downstream, model vendors are building vertical products themselves — Claude’s finance direction is already knocking on Hebbia and Rogo’s door. Application companies without data, workflow, or customer-relationship moats will feel cost and competition pressure at once. That is our judgment.
Growth is yielding to growth quality. OpenAI’s caliber episode shows the market has begun asking how annualized revenue is calculated, whether it renews, and what the gross margin looks like. The metrics worth watching next: retention, gross margin, customer concentration — and whether each company’s caliber is comparable.
Finally, the boundaries have to be stated clearly.
Scope first. This edition applies a strict AI-native definition: the company’s founding starting point and core delivery must be AI models, applications, or AI-specific technology. Under that definition, ByteDance’s large-model business (estimated $4 billion), Canva ($4 billion), Handshake ($1 billion), Notion ($600 million), and similar non-native AI businesses or transformed companies are out — otherwise No. 3 would change hands.
Kling AI is the only special case: included as the carved-out, independently financed unlisted entity, with Kuaishou’s control relationship retained in the notes. Cursor, having been acquired by SpaceX and set to list with it, no longer meets the unlisted criterion.
The ranking’s floor is not low either: No. 100 is Heidi at $50 million.
Then revenue caliber. The ranking mixes several methods: actual annual revenue, ARR, monthly × 12, quarterly × 4, last-four-weeks × 13, CARR, and third-party estimates.
Mercor’s $2 billion is an annualized run rate that does not deduct contractor costs; Higgsfield’s $1 billion is a four-week conversion; Kling’s $500 million is a March run rate; VAST’s $500 million is a committed figure.
Scale is listed at a $1 billion ceiling; Skydio at a conservative lower bound; Anduril’s $2.2 billion is the rare full-caliber actual annual revenue; Tipsy Chat has broken $100 million on an ARR basis while first-half cash collected was $15.2 million.
When you compare horizontally, keep that ruler in mind.
One hundred companies. $148.7 billion. This ranking proves one thing: AI is no longer an industry that “will make money someday.” It is making money now — and for the first time, companies have started raising prices.
Cutoff date is October 22. We will keep updating. The full ranking for this edition follows in the original Chinese release.
Source: Unique Research WeChat essay, October 23, 2026. Mid 2247523130 · idx 1.