Which AI lab is growing the fastest?







Every funding round, every investor, and what the labs promised in return, in one document. Have a look ↓
Summary
Moonshot AI is the fastest-growing AI lab among those with recent public figures: its ARR rose more than 66.7x in 8 months, from an estimated $15M at the end of 2025 to more than $1B in August 2026, just ahead of DeepSeek (83.3x in 9 months) and Z.ai (46.0x in 8 months).
The top three multiples rest on small, estimated starting points and windows of eight or nine months. A better December 2025 figure could change the order among Moonshot AI, DeepSeek and Z.ai, but not their lead over every other lab.
DeepSeek's multiple is actually bigger than Moonshot's. It took one month longer to get there, which is why Moonshot comes first once both are put on a yearly footing.
In dollars, the ranking flips. By our calculation, OpenAI added about $30B of run rate in nine months and Anthropic about $15.4B of quarterly revenue in a year, while the three fastest growers added roughly $3.5B together.
Anthropic is the fastest grower at large scale by a wide margin: quarterly revenue went from $1.38B in Q3 2025 to $16.8B in Q3 2026, 12.2x in 12 months. Its quarter-on-quarter growth did slow, from 144% in Q2 2026 to 45% in Q3.
OpenAI's growth has barely slowed. It went from $2B to $6B to more than $20B between 2023 and 2025, and to about $50B in September 2026; the widely quoted $70B counted cloud-partner sales on a gross basis.
Part of the Chinese jump is pricing. DeepSeek raised its model prices 2.3 to 4.5 times, and Z.ai's average API price rose about 101% while its token volume grew more than 40-fold.
Chinese labs won usage before they won revenue. Their models took about 61% of the tokens used by OpenRouter's top 10 models in February and March 2026, at a small fraction of the price of Claude Opus 4.6.
The fastest growers rebuilt their business in a year. API sales went from 15.2% to 86.5% of Z.ai's revenue, and business customers went from about 30% to about 80% of MiniMax's ARR.
Fast growth is not profit. Z.ai lost CNY 2.07B in the first half of 2026 on CNY 954M of revenue and MiniMax lost $358M, while Anthropic says it reached adjusted profitability in Q3 2026.
Several closely watched labs, including Thinking Machines Lab, Reflection AI, Sakana AI and StepFun, cannot be ranked because none has published two comparable revenue figures.
Our answer: in percentages, the race goes to Moonshot AI; in dollars, it still goes to OpenAI and Anthropic.
Written by the New Market Pitch teamWe track new markets so founders and investors can move faster. About us
70+pages
This is just a preview page. Don’t worry if a number looks old: the report gets refreshed, and you’ll always get the latest version.
The fastest-growing startups in the frontier AI labs market
Below is a table ranking the startups of this market by annualized growth, measured between two public figures, with the most recent measurements first. You can find our methodology at the end of this page.
If you want a deeper understanding of the market and its current dynamics, get our report covering the Frontier AI Labs Market.
| Ranking | Company | Annualized Growth | Growth Multiple | Latest Figure | Earlier Figure | Metric | Latest Period | Earlier Period | Freshness | Source Quality | Confidence | Segment | Why This Ranking | Country |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Moonshot AI | >+54,330% | >66.7x in 8 months | >$1B | $15M | ARR | As of Aug 2026 | As of Dec 2025 | Very Fresh · 2mo | Third-Party Estimate | Low | General-purpose language and reasoning models | ARR rose from an estimated $15M at the end of 2025 to more than $1B in August: above 66.7x in eight months, an annualized pace above 54,000%. The extraordinary rank depends heavily on Sacra's unverified small starting figure and a short period; management's $2B year-end goal is not achieved revenue. | cn |
| 2 | DeepSeek | +36,300% | 83.3x in 9 months | $1B | $12M | Revenue Run-Rate | As of Sep 2026 | As of Dec 2025 | Very Fresh · 1mo | Third-Party Estimate | Low | Open-weight general-purpose language and reasoning models | The annualized run rate reached $1B versus Sacra's estimated $12M at the end of 2025: 83.3x in nine months, or about +36,300% annualized. The small estimated starting base and later API price increases magnify the percentage; this is not recognized annual revenue. | cn |
| 3 | Z.ai | +31,080% | 46.0x in 8 months | $1.6B | ~$34.8M | ARR | As of Aug 2026 | As of Dec 2025 | Very Fresh · 2mo | Derived / Implied | Low | GLM language models and API platform | August ARR of $1.6B was reported as 46x December 2025, implying a starting point of roughly $34.8M and an annualized pace of about +31,080% over eight months. The starting dollar figure is derived from the reported multiple, not separately disclosed; its monthly-annualized API sales should not be confused with H1 2026 recognized revenue of CNY 954M. | cn |
| 4 | MiniMax | >+2,160% | >8.0x in 8 months | >$800M | $100M | ARR | As of Aug 2026 | As of Dec 2025 | Very Fresh · 2mo | Third-Party Estimate | Low | Multimodal language, audio and video models | The August run rate exceeded $800M against an estimated $100M at the end of 2025: more than 8.0x in eight months, or above +2,160% annualized. The prior estimate is not audited; the company separately reported $116.6M of recognized revenue in H1 2026, which is not comparable to ARR. | cn |
| 5 | Mistral AI | +1,900% | 20.0x in 12 months | $400M | $20M | ARR | As of Jan 2026 | As of Jan 2025 | Fresh · 9mo | Third-Party Estimate | Low | Open-weight general-purpose language models | Estimated ARR increased roughly 20x between January 2025 and January 2026, translating to about +1,900%. The January endpoint is nine months old and both points rely on third-party estimates; later acquisitions and infrastructure projects do not provide a newer comparable sales figure. | fr |
| 6 | Anthropic | +1,120% | 12.2x in 12 months | $16.8B | $1.38B | Quarterly Revenue | Q3 2026 | Q3 2025 | Very Fresh · 1mo | Credible Reported | Medium | General-purpose language, reasoning and coding models | Quarterly recognized revenue rose from $1.38B in Q3 2025 to a press-reported $16.8B in Q3 2026, about 12.2x or +1,120% year over year. The earlier quarter appears in the October SEC filing, but Q3 2026 has not yet appeared in filed accounts. Anthropic ranks below smaller labs because growth from a multi-billion-dollar base is measured over a full year, not a short burst. | us |
| 7 | xAI | +400% | 5.0x in 12 months | $500M | $100M | Revenue Run-Rate | As of Dec 2025 | As of Dec 2024 | Aging · 10mo | Third-Party Estimate | Low | Grok general-purpose language and reasoning models | Estimated standalone AI revenue rose from $100M at the end of 2024 to around $500M at the end of 2025, or +400%. It excludes the X advertising business acquired in March 2025; SpaceX acquired xAI in February 2026, and later SpaceX filings mix Grok, X subscriptions, advertising and AI infrastructure, so those newer group figures cannot be compared directly. | us |
| 8 | Cohere | +290% | 3.9x in 12 months | $240M | $62M | ARR | As of Dec 2025 | As of Dec 2024 | Aging · 10mo | Third-Party Estimate | Low | Enterprise general-purpose language models | ARR rose from an estimated $62M at the end of 2024 to a reported $240M in 2025, about 3.9x or +290% in twelve months. The planned combination with Aleph Alpha, signed in September 2026 but awaiting approval, could change the reporting perimeter. | ca |
| 9 | OpenAI | +240% | 2.5x in 9 months | ~$50B | ~$20B | Revenue Run-Rate | As of Sep 2026 | As of Dec 2025 | Very Fresh · 1mo | Credible Reported | Medium | General-purpose language and multimodal models | Annualized revenue advanced from about $20B at the end of 2025 to nearly $50B in September 2026, 2.5x in nine months or about +240% annualized. That lower percentage reflects a much larger starting base, not weak dollar growth: about $30B of added annualized revenue in nine months. The widely circulated $70B September figure grossed up cloud-partner sales and is not comparable. | us |
| 10 | SenseTime | +51% | 1.5x in 12 months | CNY 3.63B | CNY 2.40B | Annual Revenue | FY2025 | FY2024 | Aging · 10mo | Filed / Audited | High | Generative AI models and infrastructure segment | The Generative AI business grew 51% to CNY 3.63B in FY2025 from CNY 2.40B in FY2024, a reported segment rather than total group revenue. H1 2026 Generative AI revenue then rose 28.2% to CNY 2.33B versus H1 2025, indicating slower growth; the segment also includes AI infrastructure and applications. | cn |
| NR | StepFun | Not measurable | CNY 500M | Annual Revenue | FY2025 | Aging · 10mo | Credible Reported | Low | General-purpose multimodal language models | Reported 2025 revenue was about CNY 500M, but no comparable earlier achieved revenue figure was verified. The widely discussed CNY 1.2B for 2026 is a forecast, not a measured endpoint. | cn | |||
| NR | AI21 Labs | Not measurable | $50M | Annual Revenue | FY2025 | Aging · 10mo | Credible Reported | Low | Enterprise general-purpose language models | Third-party estimates put its 2025 revenue at about $50M, but no reliable same-basis earlier annual figure makes a growth calculation defensible. Reports about the company's fundraising have also been disputed. | il | |||
| NR | Sakana AI | Not measurable | Undated | Company Disclosed | Medium | Frontier research and Japanese language models | Sakana AI has major enterprise and public-sector relationships but has not published two dated comparable commercial-scale measurements. Funding and research milestones cannot substitute for customer or revenue growth. | jp | ||||||
| NR | Thinking Machines Lab | Not measurable | Undated | Company Disclosed | Medium | Customizable general-purpose models and training platforms | The lab has introduced model-training infrastructure and public model research but no two directly comparable achieved revenue or paying-customer figures. Its announced gigawatt-scale NVIDIA collaboration measures planned compute, not sales. | us | ||||||
| NR | Reflection AI | Not measurable | Undated | Company Disclosed | Medium | Open-weight general-purpose models and agents | The lab has public funding and model announcements but no reliable pair of dated revenue, ARR or paying-customer figures supporting a calculated growth rate. | us | ||||||
| NR | Baichuan AI | Not measurable | Undated | Company Disclosed | Medium | Chinese-language foundation models and healthcare AI | Business revenue around CNY 100M was reported for 2024, without a dated comparable achieved figure for 2025 or 2026. Its increasing healthcare focus also makes its position as a broad frontier-model laboratory less clear. | cn | ||||||
| NR | Aleph Alpha | Not measurable | Undated | Company Disclosed | Medium | Sovereign language models and enterprise AI | The proposed Cohere combination remains subject to approval. Aleph Alpha has no two verified comparable public revenue points, and its more recent business has shifted toward sovereign AI software and integration. | de |
Which AI lab is growing the fastest right now?
Moonshot AI, the Beijing company behind the Kimi models, is the fastest-growing AI lab among those with recent public figures, ahead of DeepSeek and Z.ai.
Moonshot's annual recurring revenue (ARR) went from an estimated $15M at the end of 2025 to more than $1B in August 2026. That is more than 66.7x in 8 months, an annualized pace above +54,330%. DeepSeek comes second with a revenue run rate that went from $12M in December 2025 to $1B in September 2026, 83.3x in 9 months (+36,300% annualized pace). Z.ai, the lab formerly known as Zhipu, is third: its ARR reached $1.6B in August 2026, 46.0x the roughly $34.8M of December 2025 (+31,080% annualized pace).
DeepSeek's multiple is actually bigger than Moonshot's, so why is Moonshot first? Because we rank on growth per unit of time. DeepSeek needed nine months for its 83.3x, Moonshot needed eight for more than 66.7x, and once both are put on a yearly footing Moonshot comes out ahead. The gap is not large, and both rest on small starting estimates, a point we test in detail below.
The rest of the ranking looks very different. MiniMax is fourth at more than 8.0x in 8 months, Mistral AI fifth at 20.0x in 12 months, and Anthropic sixth with quarterly revenue up 12.2x in 12 months. xAI, Cohere, OpenAI and SenseTime close the list. All of the top four are Chinese, and the two labs that dominate the conversation in the West, Anthropic and OpenAI, sit at #6 and #9 on percentages while adding far more dollars than anyone else.
| Rank | AI lab | Observed growth | Annualized | Latest figure |
|---|---|---|---|---|
| 1 | Moonshot AI | >66.7x in 8 months | >+54,330% | >$1B ARR (Aug 2026) |
| 2 | DeepSeek | 83.3x in 9 months | +36,300% | $1B run rate (Sep 2026) |
| 3 | Z.ai | 46.0x in 8 months | +31,080% | $1.6B ARR (Aug 2026) |
| 4 | MiniMax | >8.0x in 8 months | >+2,160% | >$800M ARR (Aug 2026) |
| 5 | Mistral AI | 20.0x in 12 months | +1,900% | $400M ARR (Jan 2026) |
| 6 | Anthropic | 12.2x in 12 months | +1,120% | $16.8B quarterly revenue (Q3 2026) |
| 7 | xAI | 5.0x in 12 months | +400% | $500M run rate (Dec 2025) |
| 8 | Cohere | 3.9x in 12 months | +290% | $240M ARR (Dec 2025) |
| 9 | OpenAI | 2.5x in 9 months | +240% | ~$50B run rate (Sep 2026) |
| 10 | SenseTime | 1.5x in 12 months | +51% | CNY 3.63B annual revenue (FY2025) |
Why is AI lab revenue exploding this year?
AI lab revenue is exploding because companies stopped testing chatbots and started paying for models that do long, token-hungry work, mostly coding and agents.
The jump shows up in every company that publishes numbers. Anthropic's quarterly revenue went from $376M in Q1 2025 to $11.6B in Q2 2026, according to its own SEC filing, which is about 31x in five quarters. Z.ai said its token volume rose more than 40-fold since the start of 2026, and MiniMax's open platform and enterprise revenue grew more than seven times year on year in the first half of 2026. These are three different businesses, in two countries, with very different customers, all showing the same shape of curve at the same time.
What changed is the unit of consumption. A person chatting with an assistant burns a few thousand tokens. A coding agent that reads a whole codebase, plans, runs tests and rewrites files can burn millions in a session, and companies now run thousands of these sessions a day. Z.ai launched a Coding Plan subscription for coding-agent access during its growth spurt, and Anthropic's number of customers paying more than $100K a year went from over 950 in September 2025 to over 11,900 in September 2026. Both point to the same buyer: developers and enterprises paying per token, at scale.
70+pages
Just a preview. If a number here looks old, no worries: your report comes with the latest data.
Are the growth rates of Moonshot AI, DeepSeek and Z.ai too good to be true?
Partly, yes: the direction is real, but the size of the multiples for Moonshot AI, DeepSeek and Z.ai depends on small starting figures that nobody audited.
The latest figures are solid enough. Moonshot's August ARR was shared by the company with investors and reported by Bloomberg in September 2026, DeepSeek's $1B run rate came from CEO Liang Wenfeng at an investor meeting as reported by The Information, and Z.ai's $1.6B ARR came from its own management. The weak part is the starting point. Moonshot's $15M and DeepSeek's $12M for December 2025 are estimates by the research firm Sacra, and Z.ai's ~$34.8M is not a disclosed number at all: it is what you get when you divide $1.6B by the "46x December" multiple that JPMorgan analysts drew from management's comments.
Small starting numbers move percentages enormously. If Moonshot's true December 2025 ARR had been $30M instead of $15M, its multiple would drop from more than 66.7x to more than 33x, and the annualized pace would fall by about two thirds. The ranking among the top three is therefore fragile. Their lead over the rest is not: even with a starting point twice as high, each of the three would still have multiplied its revenue far faster than MiniMax, Mistral or Anthropic.
The windows are also short. Eight or nine months of growth, annualized, gives a pace, not an observed year. Moonshot's own path shows how bumpy that can be: ARR was above $200M in April 2026, $300M in June, then above $1B in August after the July launch of Kimi K3. More than tripling in two months is what one model launch can do, and nobody should assume it repeats every quarter. That is why we lead with the multiples and treat the annualized rates as a way to compare, not as a forecast.
How fast is Moonshot AI growing with Kimi?
Moonshot AI grew its ARR more than 66.7x in 8 months, from an estimated $15M at the end of 2025 to more than $1B in August 2026, the fastest pace of any AI lab with recent figures.
The intermediate numbers reported by Bloomberg tell a story of acceleration rather than steady growth. ARR passed $200M in April 2026 and reached $300M in June, which means Moonshot added about $100M in two months. Then Kimi K3 came out in July, a mixture-of-experts model with 2.8 trillion parameters and a context window of just over one million tokens, and ARR went past $1B by August. In other words, more than $700M of ARR appeared in roughly eight weeks, more than everything Moonshot had built in its first three years.
The revenue comes from two places: subscriptions to the Kimi chatbot and sales of model access to developers and businesses. Bloomberg did not give a split. A detail that hints at where the money comes from: Moonshot negotiated revenue-sharing deals with cloud providers to host K3, and on OpenRouter 16 different providers serve the model. Moonshot is distributing its model through other people's clouds, which lets usage grow faster than its own servers could.
Management told investors it aims for $2B of annualized revenue by the end of 2026. That is a goal, not achieved revenue, so it plays no part in the ranking.
70+pages
This is a preview page. Spotted an outdated number? That’s normal here. The report you get is always the freshest version.
How did DeepSeek get to a $1 billion revenue run rate so fast?
DeepSeek reached a $1B annualized revenue run rate in September 2026, 83.3x the estimated $12M of December 2025, and a good part of that jump came from charging more, not only from selling more.
For most of its life, DeepSeek was famous for being cheap. Its models were released openly and its API was priced far below American rivals. That changed in 2026: DeepSeek raised its model prices by 2.3 to 4.5 times, according to The Information's September report, and the run rate more than doubled within a few months. When a lab multiplies its prices by three or four and keeps its customers, revenue can rise much faster than usage. A share of DeepSeek's 83.3x is therefore a pricing reset rather than new demand.
DeepSeek also stands out for how little of its computing power goes to serving customers. The same report says more than 70% of its compute is used for training and less than 30% for inference, the part that actually produces revenue. DeepSeek is still run like a research lab that happens to sell, which makes its $1B run rate more surprising and also suggests room to grow if it shifts capacity toward paying users.
This is a run rate, not recognized annual revenue.
How fast is Z.ai (Zhipu) growing since its Hong Kong IPO?
Z.ai is growing at an extraordinary pace since its listing in Hong Kong: its ARR reached $1.6B in August 2026, 46.0x its level of December 2025, and it is the only one of the top three whose growth also shows up in published accounts.
Because Z.ai is listed, we can check its run rate against real revenue. In the first half of 2026, recognized revenue was CNY 954M, up almost 400% year on year. That is far smaller than the ARR suggests, because ARR here is the August month of API sales multiplied by twelve, at a time when sales were climbing every month. Management had reported a run rate of $1B in early July, so the following weeks alone added about 60%.
What makes Z.ai interesting is how completely the company changed. In 2025, local deployment, meaning installing models on the servers of governments and large companies, made up 73.7% of revenue. In the first half of 2026, API and open-platform sales were 86.5% of revenue, against 15.2% a year earlier, while local deployment shrank 20.5%. Z.ai went from a project-based IT supplier to a pay-per-token platform in about twelve months. Its token volume grew more than 40-fold since the start of the year and its average API price rose about 101%, so both usage and price pushed in the same direction.
The cost of that switch is visible too. Company gross margin fell from 50% to 26.4% because cloud sales earn less per yuan than deployment contracts, and Z.ai still lost CNY 2.07B in the half.
Is MiniMax growing as fast as the other Chinese AI labs?
No, MiniMax is growing fast but in a different league from the top three: its ARR rose more than 8.0x in 8 months, from an estimated $100M at the end of 2025 to more than $800M in August 2026.
The reason is mostly arithmetic. MiniMax started 2026 from a much larger base than Moonshot or DeepSeek, so the same kind of dollar growth gives a smaller multiple. In dollars, MiniMax added more than $700M of ARR over the period, close to what Moonshot and DeepSeek each added. On percentages it ranks #4, on dollars it plays in the same range as the labs above it.
The listed accounts confirm the shift. MiniMax reported $116.6M of recognized revenue in the first half of 2026, up 283% year on year, against $79M for the whole of 2025. Business customers made up about 80% of ARR in the first half of 2026, against about 30% a year earlier, and 61% of revenue came from outside China. MiniMax used to be best known for consumer apps such as Talkie and the Hailuo video generator. Today it is mostly a model supplier to businesses, and a largely international one.
Why are Chinese AI labs growing faster than American ones?
Chinese AI labs grow faster in percentages because they started 2026 with almost no revenue, while American labs started with billions, and because they priced their way into the same coding and agent demand.
Five of the ten ranked labs are Chinese, and they hold #1, #2, #3, #4 and #10. The three American labs sit at #6 (Anthropic), #7 (xAI) and #9 (OpenAI), with France's Mistral AI at #5 and Canada's Cohere at #8. The top four Chinese labs went, between them, from a combined starting point of roughly $160M of ARR or run rate to more than $4.4B in eight or nine months (our calculation from the four pairs of figures). That is a big step from close to zero, and still a small amount next to the American leaders, as we show further down.
Volume was the way in. On OpenRouter, a platform developers use to switch between models, Chinese models accounted for about 61% of the tokens used by the top 10 models in February and March 2026, with MiniMax M2.5, Kimi K2.5 and GLM-5 among the most used. But a token of MiniMax M2.5 cost about $0.30 to $1.10 per million on that platform, against $5 to $25 for Claude Opus 4.6. Chinese labs won usage by being cheap, and their 2026 revenue jump is partly the moment that usage started to be paid for.
So the percentage gap says more about starting points than about who is winning. Chinese labs are turning huge free or near-free usage into revenue, while American labs are growing an already large paid business.
| Country | Ranked AI labs (rank) |
|---|---|
| China | Moonshot AI (1), DeepSeek (2), Z.ai (3), MiniMax (4), SenseTime (10) |
| France | Mistral AI (5) |
| United States | Anthropic (6), xAI (7), OpenAI (9) |
| Canada | Cohere (8) |
70+pages
Preview only. Some numbers may have changed since we made this page, but your report will be up to date.
How fast is Anthropic growing right now?
Anthropic is growing faster than any AI lab of its size: quarterly revenue went from $1.38B in Q3 2025 to $16.8B in Q3 2026, 12.2x in 12 months (+1,120%).
The figures come from two places. The Q3 2025 number is in the S-1 that Anthropic recently filed with the SEC for its IPO, and the $16.8B for Q3 2026 was shared with investors and reported by the Financial Times and Reuters, but has not yet appeared in filed accounts. Put next to each other, the quarters show the curve clearly: $376M in Q1 2025, $787M in Q2, $1.38B in Q3, $2.04B in Q4, $4.73B in Q1 2026, $11.56B in Q2 and $16.8B in Q3. The third quarter of 2026 alone brought in more than the whole first half of the year ($16.3B), and more than three and a half times all of 2025 (about $4.6B).
The pace is not constant, though. Quarter on quarter, growth was 132% in Q1 2026 and 144% in Q2, then 45% in Q3. That is a clear slowdown, even if 45% in a single quarter would still be an exceptional year for almost any company. The customer base grew in step: companies paying Anthropic more than $100K a year went from over 950 to over 11,900 between September 2025 and September 2026, about 12.5x, roughly the same multiple as revenue. Growth came from many more large customers, and not only from a few giant contracts.
Why does a lab with $16.8B of quarterly revenue rank only #6? Because we compare percentages, and smaller labs measured over eight or nine months get a mechanical advantage. Anthropic's growth is measured over a full year from a base that was already above $1B per quarter.
Is Anthropic growing faster than OpenAI?
Yes, Anthropic is growing faster than OpenAI in percentage terms: 12.2x in 12 months for Anthropic's quarterly revenue, against 2.5x in 9 months for OpenAI's revenue run rate.
On a yearly footing, that is about +1,120% for Anthropic against an annualized pace of about +240% for OpenAI. The two companies are measured on different metrics, recognized quarterly revenue for Anthropic and an annualized run rate for OpenAI, but the gap is far too large for that difference to reverse it.
The dollar comparison is closer than many people think. OpenAI's run rate went from about $20B at the end of 2025 to about $50B in September 2026. Multiply Anthropic's Q3 2026 revenue by four and you get about $67B a year (our calculation, not a figure Anthropic publishes). The comparison is rough because a quarter times four and a monthly run rate are different measures, but on these numbers Anthropic is selling at a pace at least comparable to OpenAI's.
There is one accounting trap. In September 2026, several outlets reported that OpenAI was close to $70B of annualized revenue. That figure counted sales through cloud partners on a gross basis. When investors asked for numbers they could compare with Anthropic, the figure came back at nearly $50B, as reported a few weeks later by the Financial Times and Reuters. Using the $70B would have made OpenAI look 40% bigger than it is on a comparable basis.
Is OpenAI's growth slowing down?
Not really: OpenAI's run rate grew 2.5x in 9 months, from about $20B at the end of 2025 to about $50B in September 2026, which is close to the tripling it managed in each of the previous two years.
OpenAI's CFO Sarah Friar wrote in January 2026 that annualized revenue went from $2B in 2023 to $6B in 2024 and more than $20B in 2025. That is 3x, then more than 3x. An annualized pace of about +240% means roughly 3.4x a year, so the rhythm has held. The impression of a slowdown came from comparing the $50B with the inflated $70B figure, not from the trend itself.
The same letter shows how closely revenue follows computing power: OpenAI went from 0.2 gigawatts of compute in 2023 to 0.6 in 2024 and about 1.9 in 2025, again roughly 3x a year. For OpenAI, revenue growth has so far moved in step with how much compute it could bring online. That makes data-center build-out, not demand, the most likely limit on how fast it can keep growing.
OpenAI ranks #9 only because it starts from the largest base of all. About $30B of added run rate in nine months is far more than any lab ranked above it added, Anthropic aside.
Which AI lab is adding the most revenue in dollars?
OpenAI and Anthropic add far more revenue in dollars than everyone else combined, even though they rank #9 and #6 on growth rates.
Our calculation, using the two figures of each lab, gives about $30B of added annualized run rate for OpenAI in nine months, and about $15.4B of added quarterly revenue for Anthropic in a year, which would be about $62B on an annual basis. The top three on percentages, Moonshot AI, DeepSeek and Z.ai, added roughly $3.5B of ARR or run rate between them. Put differently, OpenAI alone added around eight times what the three fastest-growing labs added together.
This is why both views are needed. Percentages show who is breaking out, and in 2026 that is clearly the Chinese labs. Dollars show who is capturing the bulk of the new spending, and that is still the two American leaders. All eight other ranked labs together added about $5.4B (our calculation, with SenseTime converted from yuan), less than a fifth of OpenAI's increase.
| Rank | AI lab | From | To | Added (our calculation) |
|---|---|---|---|---|
| 1 | Moonshot AI | $15M ARR (Dec 2025) | >$1B ARR (Aug 2026) | more than ~$985M |
| 2 | DeepSeek | $12M run rate (Dec 2025) | $1B run rate (Sep 2026) | ~$988M |
| 3 | Z.ai | ~$34.8M ARR (Dec 2025) | $1.6B ARR (Aug 2026) | ~$1.57B |
| 4 | MiniMax | $100M ARR (Dec 2025) | >$800M ARR (Aug 2026) | more than ~$700M |
| 5 | Mistral AI | $20M ARR (Jan 2025) | $400M ARR (Jan 2026) | ~$380M |
| 6 | Anthropic | $1.38B quarterly revenue (Q3 2025) | $16.8B quarterly revenue (Q3 2026) | ~$15.4B per quarter |
| 7 | xAI | $100M run rate (Dec 2024) | $500M run rate (Dec 2025) | ~$400M |
| 8 | Cohere | $62M ARR (Dec 2024) | $240M ARR (Dec 2025) | ~$178M |
| 9 | OpenAI | ~$20B run rate (Dec 2025) | ~$50B run rate (Sep 2026) | ~$30B |
| 10 | SenseTime | CNY 2.40B (FY2024) | CNY 3.63B (FY2025) | ~CNY 1.23B |
70+pages
This is just a preview page. Don’t worry if a number looks old: the report gets refreshed, and you’ll always get the latest version.
How fast is Mistral AI growing?
Mistral AI, Europe's best-known AI lab, multiplied its ARR by 20.0x in 12 months, from an estimated $20M in January 2025 to $400M in January 2026, which makes it the fastest-growing lab outside China in our ranking.
Both figures are estimates by Sacra, and the latest one is nine months old, so this is the growth Mistral showed through 2025 rather than a reading of today. The Paris company does not publish detailed accounts. Its co-founder Arthur Mensch told Bloomberg at Davos in January 2026 that Mistral expects to pass $1B of revenue by the end of 2026, but that is a target, and no newer comparable figure has been published since.
Mistral sells itself as the sovereign option for European governments, banks and industrial groups that want models they can run under European control, and the chipmaker ASML became a shareholder in 2025. That pitch helps explain why a lab with far fewer users than ChatGPT could multiply its revenue twenty times in a year. It also explains the ceiling: Mistral's $400M is less than 1% of OpenAI's current run rate.
Is xAI's Grok actually growing?
Yes, xAI's Grok business grew, by 5.0x in 12 months on Sacra's estimates, from a $100M revenue run rate at the end of 2024 to about $500M at the end of 2025, but those figures are about ten months old and nobody publishes a clean newer one.
The problem is the corporate structure. xAI bought X, the former Twitter, in March 2025, and SpaceX then acquired xAI in February 2026. Our figures cover the standalone AI business, without X advertising. SpaceX's own filing reported $3.2B of revenue in 2025 for its AI segment, but that number mixes Grok with X subscriptions, advertising and AI infrastructure, and Light Reading's reading of the filing puts Grok itself well under $2B. The two numbers measure different things, which is why the larger one cannot be used to compute Grok's growth.
So xAI's growth through 2025 was real and quick, faster than Cohere and OpenAI in percentage terms over the periods measured. Whether Grok is still growing at that pace now is something its owners have not shown in a comparable form.
How fast is Cohere growing before the Aleph Alpha merger?
Cohere grew its ARR 3.9x in 12 months, from an estimated $62M at the end of 2024 to $240M in 2025, a solid pace for an enterprise-only lab but well behind the leaders.
Cohere never built a consumer chatbot and sells models mostly to large companies and governments that want to run them privately. That choice makes growth slower, since each contract takes months, and steadier. In September 2026, Cohere signed an agreement to combine with Germany's Aleph Alpha, in a deal reported to value the combined company at about $20B, and the Schwarz Group, owner of Lidl and an Aleph Alpha backer, announced €500M of financing in April 2026.
The merger still needs regulatory approval. Once it closes, Cohere's figures will include Aleph Alpha's business, and any growth measured across that date will partly come from the acquisition rather than from Cohere's own sales.
Is SenseTime still growing in generative AI?
SenseTime is still growing in generative AI, but it is the slowest of the ranked labs: the segment grew 1.5x in 12 months, from CNY 2.40B in FY2024 to CNY 3.63B in FY2025 (+51%).
Unlike most labs on this list, SenseTime's figures come from audited annual reports, so they carry far less uncertainty. They also cover more than models: the generative AI segment includes AI infrastructure and applications, which grow at a calmer, more traditional pace than token sales.
The latest half-year shows growth slowing further. Generative AI revenue rose 28.2% to CNY 2.33B in the first half of 2026, and the segment is now 79.9% of group revenue. SenseTime even posted its first net profit since listing, CNY 617M, though that came largely from fair-value gains on companies it invested in; on an adjusted basis it still lost CNY 385.9M. SenseTime has become a generative AI company, but a slow-growing one compared with its Chinese peers.
70+pages
Just a preview. If a number here looks old, no worries: your report comes with the latest data.
Are AI price increases behind the growth?
In China, partly yes: DeepSeek and Z.ai both raised prices in 2026, and those increases explain a meaningful share of their revenue jump.
The evidence is unusually direct. DeepSeek raised its model prices 2.3 to 4.5 times, according to The Information, and its run rate more than doubled in a few months. Z.ai said its average API price rose about 101% while, as mentioned earlier, its token volume grew more than 40-fold since the start of 2026. For Z.ai, volume was clearly the bigger driver, since tokens multiplied by 40 while price roughly doubled, but price alone doubled the revenue each token brought in.
This is a change of strategy. In 2024 and 2025, Chinese labs competed by cutting prices, sometimes to near zero, to win developers. In 2026, the ones with popular models started charging more and found that customers stayed. Z.ai's API business went from a gross margin of -0.4% to 24.6%, the clearest sign that pricing power is real. It also means part of the top three's growth is a one-time reset: a lab can triple its price once, but it cannot do it every year.
Which AI labs grow without publishing their revenue?
Several well-known labs, including Thinking Machines Lab, Reflection AI, Sakana AI, StepFun, Baichuan AI, AI21 Labs and Aleph Alpha, cannot be ranked because they have not published two comparable revenue figures.
StepFun is the closest to being measurable. Caijing reported that it made about CNY 500M of revenue in 2025, but no verified earlier figure exists, and the CNY 1.2B talked about for 2026 is a forecast. AI21 Labs is estimated at about $50M of revenue in 2025 with no reliable earlier figure on the same basis. Baichuan AI reported business revenue of about CNY 100M for 2024 and nothing comparable since, while it has moved toward healthcare.
The American newcomers are a different case. Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has launched its Tinker training platform and announced a gigawatt-scale collaboration with NVIDIA, but that measures planned compute, not sales. Reflection AI has announced funding and models but no revenue. Sakana AI in Japan has large corporate and government partners but no dated revenue figures. Aleph Alpha's future figures will merge into Cohere's if the deal is approved.
| AI lab | Country | Latest public figure |
|---|---|---|
| StepFun | China | CNY 500M annual revenue (FY2025) |
| AI21 Labs | Israel | $50M annual revenue (FY2025) |
| Sakana AI | Japan | No public revenue figure |
| Thinking Machines Lab | United States | No public revenue figure |
| Reflection AI | United States | No public revenue figure |
| Baichuan AI | China | No comparable recent figure |
| Aleph Alpha | Germany | No public revenue figure |
70+pages
This is a preview page. Spotted an outdated number? That’s normal here. The report you get is always the freshest version.
Are the fastest-growing AI labs making money?
Mostly not: Anthropic is the only ranked lab that says its core business turned profitable, and only on an adjusted basis.
Anthropic told investors it was profitable on an adjusted basis and cash-flow positive in Q3 2026, after an operating loss of more than $8B in 2025. The Chinese listed labs are still deep in the red. Z.ai lost CNY 2.07B in the first half of 2026, more than twice its revenue, and spent CNY 2.13B on research. MiniMax lost $358M in the same half, about three times its $116.6M of revenue. Both losses narrowed by 11% to 12% year on year, which is progress but far from break-even. SenseTime, the slowest grower, did post a net profit in the first half of 2026, but mostly thanks to gains on its investments, as we saw earlier.
So the speed of growth and profitability are moving in opposite directions for now. The fastest growers are buying their growth with compute and research spending, and the one lab that reached adjusted profit is the one growing more slowly in percentage terms but on a base of tens of billions.
Can Moonshot AI and DeepSeek keep growing this fast?
No, not at these multiples: Moonshot AI and DeepSeek cannot multiply revenue 60 or 80 times again, but they can keep growing very fast in dollars.
Even Moonshot's own target shows the deceleration. Going from more than $1B of ARR in August to $2B by the end of 2026 would mean doubling in about four months, which is still extremely fast but far slower than the more than 66.7x of the previous eight months. Anthropic's path offers a preview: 144% quarter-on-quarter growth in Q2 2026 fell to 45% in Q3 once revenue reached the tens of billions.
The one-time factors will also fade. As seen above, the 2026 price increases and Z.ai's switch from deployment projects to API sales boosted 2026 once; they will not repeat at the same scale. What will decide the next year is whether their models stay among the ones developers pick. On OpenRouter, Kimi K2.5's weekly token volume fell 20% in one March 2026 week while MiniMax M2.5's tripled, a reminder that developers switch models within days.
So which AI lab is really growing the fastest?
Moonshot AI is the fastest-growing AI lab right now among those with recent public figures, and that answer holds, even if the size of its lead is fragile.
Moonshot's ARR rose more than 66.7x in 8 months, ahead of DeepSeek (83.3x in 9 months) and Z.ai (46.0x in 8 months). As seen above, these three rates rest on small, estimated starting points over short windows, so the exact order among them could change with a better December 2025 figure. Their lead over everyone else would not.
The claim that Chinese labs are "outgrowing" OpenAI and Anthropic is exaggerated if it means they are winning the market. As we saw, OpenAI added about $30B of run rate in nine months and Anthropic about $15.4B of quarterly revenue in a year, while the top three Chinese labs added roughly $3.5B together. In percentages, the race goes to Moonshot. In dollars, it still goes to OpenAI and Anthropic, with Anthropic the fastest grower at that scale by a wide margin.
70+pages
Preview only. Some numbers may have changed since we made this page, but your report will be up to date.
Our methodology
Our objective was to rank frontier AI labs by measured growth, not by reputation, funding or headlines. We measure growth only between two public figures of the same metric for the same company: ARR to ARR, run rate to run rate, quarterly revenue to quarterly revenue, annual revenue to annual revenue. We never compare two different metrics, and we never use a forecast or a target, such as Moonshot AI's $2B year-end goal, as the latest point.
The two figures must be 6 to 36 months apart. Because windows differ, we annualize the growth so that a lab measured over eight months can be compared with one measured over twelve. Next to each annualized rate we show the simple multiple and its window, for example 2.5x in 9 months, and when the window is under 12 months we call the annualized rate a pace, since no full year has been observed. Labs whose latest figure dates from the last 12 months are ranked first; labs measured on older figures are ranked after them, even when their growth was faster, because the question is who is growing fastest now. In this market, all ten ranked labs have a latest figure from the last 12 months. When a newer figure exists on a different basis, as with SpaceX's combined AI segment after it acquired xAI, we mention it without using it. When part of the growth comes from an acquisition, we say so.
We only rank companies whose latest figure is at least $5M, because growth from a tiny base says little. When a company says revenue tripled without giving a base, we can quote the claim but we do not rank it. The ranking covers startups only: companies founded in 2000 or later, never listed or first listed within the last five years, and independent rather than divisions of larger groups, with companies acquired within the last three years, such as xAI, kept in. These companies still behave like young growth businesses, and we apply the same definition in every market we cover. This is why the AI units of Google, Meta, Alibaba or ByteDance are left out.
Each row carries a confidence level. It is high when both figures come from filed or audited accounts, medium when both come from company statements or credible reporting, and low when one point is a third-party estimate or derived from a reported multiple, with short windows weighing toward the lower level. We searched company statements, investor letters, filings, earnings releases, funding announcements and reputable financial and technology media over the last ten years to find both points.
We considered and rejected several alternatives. One common 12-month window for every lab would have left only a handful of names, since few labs publish two figures exactly a year apart, so annualization is the fairest comparison and the observed multiple sits next to it. Ranking by dollars added would simply reward the two largest labs, so we give that view in the text instead. Including growth claims without a base amount would have meant ranking numbers no one can check. For the three leaders, the starting points are small and estimated: we keep them because they are the best available figures, and we show the multiples first so that readers can judge the scale.
One limit cannot be removed: only labs that publish figures can be measured. Thinking Machines Lab, Reflection AI, Sakana AI, StepFun, Baichuan AI, AI21 Labs and Aleph Alpha are all active, but none has published two comparable figures. Our principle throughout is that two checkable figures are more useful than a growth claim without a base.
Key sources used for this analysis include: Bloomberg on Moonshot AI's ARR, Sacra's estimates for Moonshot AI, The Information on DeepSeek's run rate, Caixin on Z.ai's first-half results, Yicai Global on MiniMax's results, Sacra on Mistral AI's ARR, Anthropic's S-1 filing, Reuters on Anthropic's Q3 2026 revenue, OpenAI's CFO letter on revenue and compute, Reuters on OpenAI's net-basis run rate, Light Reading on SpaceX's AI segment, Reuters on Cohere and Aleph Alpha, SenseTime's 2025 annual report, and OpenRouter's usage rankings.
Written by the New Market Pitch teamWe track new markets so founders and investors can move faster. About us









