The market as of early October 2026: revenue, users, enterprise adoption, token volume and the capital behind them. Every figure is dated and tagged by how firm it is.
The market as of early October 2026: revenue, users, enterprise adoption, token volume and the capital behind them. Every figure is dated and tagged by how firm it is.
Two years ago one company led generative AI on almost every measure. Today the leader depends on the measure, because the measures count different things. OpenAI leads on consumer reach. Anthropic leads on enterprise adoption and, on the latest press reports, on revenue. Google leads on distribution. Open-weight models, many of them Chinese, carry most of the token volume on OpenRouter, a widely tracked developer marketplace. The four largest US platforms lead on capital, with about $748B of capex planned for 2026.
Amazon, Alphabet, Microsoft, Meta · company guidance
Five things to take from this briefing
Revenue has moved to enterprise. Anthropic booked $11.5B in Q2 2026 against OpenAI's $6.7B67. OpenAI's own business revenue more than doubled in Q3, and its run rate was nearing $70B by late September1.
Consumers still mean ChatGPT. It reports 1.2 billion weekly users and holds 55.5% of visits to gen-AI chatbot websites, down from 86.7% in January 202589.
Businesses are buying Claude. 43.5% of US businesses on Ramp's platform paid Anthropic in July, against 39.7% for OpenAI10. Menlo Ventures puts Anthropic at 40% of enterprise LLM API use11.
Volume is not revenue. US-origin models fell from about 70% to 30% of tokens on OpenRouter in a year as cheaper open-weight models took bulk work12. The dollars stayed mostly with closed frontier models.
The buildout dwarfs the revenue. Planned 2026 capex at Amazon, Alphabet, Microsoft and Meta is about 4.4× the combined run-rate revenue of OpenAI and Anthropic.
Most figures for private companies come from press reports of investor materials, not audited statements. Each figure below is marked company-statedpress reportsingle reportestimate or filing. Anthropic's public S-1, expected this month, would be the first audited look at either lab's finances.
Seven scoreboards
Each row answers a different question. Read across a row to see what the metric captures, who leads on it, and what it leaves out.
Google: Gemini in Android and Search; Amazon: largest 2026 capex ($220B)
Microsoft, Meta
Capex serves many businesses, not only AI
Revenue: the fastest-growing numbers are the least audited
Both leading labs disclose run-rate revenue: recent revenue extrapolated to a full year. The Information reported that both multiply four weeks of revenue by 13, with Anthropic adding monthly subscriptions multiplied by 1213. A run rate is a speedometer reading, not an odometer reading. It is not audited and can move quickly in either direction.
On those figures, Anthropic went from about $9B at the end of 2025 to above $65B at the end of July, a 7.2× increase142. OpenAI doubled over the same stretch, from $20B to above $40B1516. Anthropic passed OpenAI on reported run rate in April at about $30B. In late September Axios reported OpenAI nearing $70B, with business revenue more than doubling since July1. A separate Axios report, citing the New York Times, put Anthropic's pace above $100B3. That figure rests on a single report and is drawn as a dashed segment below.
OpenAI and Anthropic revenue, three ways
US$ billions. Switch views with the tabs.
OpenAI
Anthropic
company-stated
press report
single report
OpenAI
Anthropic
Full-year 2025: OpenAI $13.1B17; Anthropic about $10B6. Anthropic Q2 per Bloomberg (preliminary); the Wall Street Journal reported $11.6B7.
Compare at
Axios sourceOpenAI memo
How to read it. The vertical line marks April 2026, when Anthropic's reported run rate passed OpenAI's. Filled dots are company statements; hollow dots are press reports citing people familiar with the figures. Dates for the two September points are 11 days apart. The faint bar in the gross-vs-net view is Anthropic's figure as reported; the solid bar is after the adjustment you choose. OpenAI's figure is already booked net for Azure sales.Data table
Anthropic's Q2 2026 revenue was $11.5B, 2.4× its Q1 figure; OpenAI's was $6.7B, up 18%67. OpenAI was still ahead in Q1 ($5.7B to $4.7B)24. Anthropic reported its first quarter of positive adjusted operating income in Q2, about $560M, and told investors it does not expect to stay profitable in the quarters that follow as compute spending rises25. OpenAI spent $3.7B of cash in Q1, 65% of its revenue, according to The Information24.
The two companies book cloud sales differently
When a customer buys Claude through Amazon Bedrock or Google Cloud, Anthropic records the full sale as revenue and the cloud provider's cut as an expense. When a customer buys OpenAI models through Microsoft Azure, OpenAI records only its share2627. Both treatments are allowed under US accounting rules. They change reported revenue and margins, not operating profit. In April, OpenAI's chief revenue officer told staff the difference inflated Anthropic's then-$30B figure by about $8B, roughly 27%28. In September, a person familiar with Anthropic's finances told Axios that a net restatement would cut its revenue by 6–10%27. The third tab above applies either adjustment. Even at the competitor's 27%, Anthropic's reported September pace stays ahead of OpenAI's, by about $3B. The margin is thin, it compares figures from different weeks, and Ramp's data show OpenAI growing faster among business customers so far in Q310.
Everyone else is smaller, or does not say
xAI's audited 2025 revenue was $3.2B, including X's advertising, against a $6.4B operating loss; its Q1 2026 AI-segment revenue was $818M29. Mistral, Europe's largest model developer, reported about $400M of annualized revenue in January and targets $1B by year end30. Google does not report Gemini revenue31. Microsoft reported an AI business run rate above $37B in the March quarter, up 123%32, and Amazon an AWS AI run rate above $25B33. Those cloud figures include what the labs themselves spend on compute, so adding them to lab revenue counts the same dollars twice.
Consumer reach: OpenAI's lead is large and no longer uncontested
ChatGPT is the largest consumer AI product. OpenAI told developers on September 29 that it reaches 1.2 billion weekly users, up from 800 million a year earlier4; in February it reported 50 million paying subscribers34. Google's Gemini app passed 1 billion monthly users on August 11, up from 750 million in February531. Meta reported 1 billion monthly users for Meta AI in May 2025, a count that includes use inside WhatsApp, Instagram and Facebook35. Anthropic does not publish consumer user counts; Sensor Tower estimated about 245 million monthly users for Claude in May5.
Reported users of AI assistants
Millions of users. Labels say whether a count is weekly or monthly; * marks third-party estimates.
OpenAI
Google
Anthropic
Others
Not like-for-like. A weekly count is a stricter bar than a monthly count, so ChatGPT's figure understates its lead on a monthly basis. Meta AI's figure is from May 2025 and counts assistant use inside Meta's apps. Chinese figures are QuestMobile estimates for May 202636; Grok's is from SpaceX's IPO filing29.Data table
On the web, ChatGPT's share of worldwide visits to gen-AI chatbot sites fell from 86.7% in January 2025 to a low of 52.7% in May 2026, then recovered to 55.5% in August9378. Gemini's share grew 4.5× to 25.6%. Claude's grew from 1.5% to 9.3%, most of it in the first half of 2026. DeepSeek (3.4%), Grok (2.4%), Microsoft Copilot (1.6%) and Perplexity (0.9%) trail8.
Share of worldwide visits to gen-AI chatbot websites
Percent of visits, Similarweb panel. Dots mark published snapshots.
ChatGPT
Gemini
Claude
All others
What web share leaves out. Mobile apps, assistants built into Android and iOS, and API calls from software do not register as website visits. Google can route users to Gemini from Android and Search; OpenAI and Anthropic reach many paying users through desktop apps, coding tools and APIs, none of which count as web visits. Snapshots come from three Similarweb releases38; a separate report of the same May data gives Gemini 27.8% rather than 27.3%8.Data table
The two challengers grew by different routes. Gemini's growth follows Google's distribution: default placement on Android and in Search, where AI Overviews reach more than 2.5 billion people a month and AI Mode more than 1 billion31. Claude grew without a platform default, and Similarweb finds its users return daily at a slightly higher rate than Gemini's39. ChatGPT is also becoming an advertising business: Similarweb found ads in 26% of US desktop ChatGPT conversations in June 202640.
Enterprise: where the money is, Anthropic leads
Two independent sources point the same way. Menlo Ventures' survey of about 150 technical leaders put Anthropic at 40% of enterprise use of model APIs at the end of 2025, OpenAI at 27% and Google at 21%. At the end of 2023 OpenAI held 50% and Anthropic 12%4111. Menlo is an Anthropic investor. Ramp's AI Index, built from corporate card and bill-pay transactions rather than surveys, shows Anthropic passing OpenAI in April 2026 (34.4% of Ramp businesses paying, against 32.3%) and leading 43.5% to 39.7% in July, when 55.7% of Ramp businesses paid for at least one AI tool, up from 50.6% in April424310.
Enterprise share, two ways
OpenAI
Anthropic
Google
Other
Share of enterprise LLM API use
Menlo Ventures survey, percent
US businesses paying each vendor
Ramp AI Index, percent of Ramp customers (labels rounded)
Caveats. Menlo's shares are modeled from a survey of enterprise API use, and Menlo invests in Anthropic. Ramp counts whether a business pays a vendor, not how much; its customers skew toward startups and technology firms; and its sample grew from about 50,000 to more than 70,000 businesses during 2026, so early and late readings are not strictly comparable. May 2025 Ramp values are approximate (OpenAI ≈32%, Anthropic ≈8–9%).Data table
Ramp's lead economist also reported that OpenAI was outgrowing Anthropic among Ramp's customers in Q3 to date, crediting OpenAI's newest flagship model10. The enterprise lead is real on current data, and it is being contested.
Coding is the largest and most contested use case
Claude Code, Anthropic's coding agent, reached a $2.5B run rate by February 202644. Cursor, a leading independent AI code editor, reportedly crossed about $4B in early June45; days after its IPO, SpaceX agreed to buy it for $60B in stock and pair it with xAI's compute46. OpenAI's Codex had 1.6 million weekly users in February47. Seat-based products are scaling too: Microsoft 365 Copilot passed 30 million paid seats in July48, Google had sold more than 8 million Gemini Enterprise seats by February49, and OpenAI reports more than 2 million business customers50.
Token volume: open-weight models carry the traffic, closed models keep the dollars
OpenRouter, a marketplace that routes developers' requests across many models, publishes usage data that gives one of the few public views of where tokens go. In a year, US-origin models' share of its tokens fell from about 70% to about 30%, open-weight models came to carry most of the volume, and DeepSeek became the largest single provider12. Anthropic held roughly 12% of volume51.
70% → 30%
US-origin models' share of OpenRouter tokens, June 2025 to June 2026 (approximate)12
60–69%
Open-weight models' share of OpenRouter tokens, mid-2026, depending on the tracker12
~12%
Anthropic's share of OpenRouter volume, against about 40% of enterprise API use in Menlo's survey51
Volume and revenue diverge because frontier models cost many times more per token, often ten times or more1252. Bulk, repetitive work moves to the cheapest adequate model; complex coding and agent work stays with frontier closed models. For a manager, this is a pricing story: open-weight models cap what any lab can charge for routine work, and OpenAI cut prices on several models during 202653.
China is a separate market. QuestMobile counted 499 million monthly users of AI-native apps in China in May 2026, up 85.4% from a year earlier, led by ByteDance's Doubao (382 million), Alibaba's Qwen (167 million) and DeepSeek (130 million)36. US services are largely unavailable there, while Chinese open-weight models are widely used by developers elsewhere.
Capital: the buildout is several times the revenue
Amazon, Alphabet, Microsoft and Meta guide to between $735B and $760B of capital spending in 2026, most of it for AI data centers and chips33545556. At the midpoint, about $748B, that is 4.4× the combined late-September run rate of OpenAI and Anthropic (about $170B), or 7.1× using the firmer July figures. Both Amazon and Alphabet raised their guidance after the second quarter.
2026 capital expenditure guidance
US$ billions. Lighter extensions show the top of a guided range.
Different things on one axis. The bars are planned spending; the vertical line is two labs' annualized revenue. They share a unit, not a purpose: the platforms' capex also serves search, advertising, retail and cloud customers. Microsoft's figure is its 2026 guidance from April and may differ from its fiscal-year number.Data table
The demand side is contracted, at least on paper. At June 30, Microsoft reported $678B of commercial remaining performance obligations, Google Cloud a $514B backlog and AWS $496B, about $1.69T in total485457. A large part of that comes from a few AI customers; The Information reported that Anthropic's Google Cloud commitment alone is $200B over five years58. Cloud growth has accelerated: Google Cloud revenue rose 82% in Q2 to $24.8B at about a 36% operating margin, AWS rose 37% to $42.2B, and Azure grew 43% in Microsoft's June quarter543348. The cost shows up in cash: Alphabet's free cash flow turned negative in Q2, and Amazon's has also swung negative5957.
Private valuations imply multiples that have fallen as revenue grew. Dividing each round's post-money valuation by the run rate at the time:
Round
Date
Valuation
Run rate then
Multiple
Anthropic · Series G
2026-02-12
$380B
$14B
27.1×
OpenAI · March 2026 round
2026-03
$852B
$25B
34.1×
Anthropic · Series H
2026-05-28
$965B
$47B
20.5×
OpenAI at latest reported run rate
Sep 2026
$852B
$70B
12.2×
Anthropic at latest reported run rate
Jul / Sep 2026
$965B
$65B / $100B
14.8× / 9.7×
Sources: 62672318. Multiples computed here; run rates are press-reported or company-stated as noted above.
Regions: adoption tracks policy and infrastructure more than where models are built
Microsoft's AI Economy Institute estimates the share of working-age people who used a generative AI tool, from Windows and Microsoft product telemetry. In Q1 2026 that share was 17.8% worldwide. The UAE led at 70.1%, 3.9× the world rate; the United States ranked 21st at 31.3%6869. The gap between the Global North (27.5%) and South (15.4%) widened. In Q2 the world figure rose to 18.8% and the UAE to 73.3%70.
Share of working-age people using generative AI, Q1 2026
Percent of population aged 15–64. Italic rows are aggregates.
Measurement limits. The estimate relies mainly on Windows and Microsoft products, captures Apple devices only partly, and has no consolidated data for China, Russia or Iran69.Data table
Economy
Q1 2026
UAE
70.1%
Singapore
63.4%
United States
31.3%
Global North
27.5%
World
17.8%
Global South
15.4%
United States and Canada
31.3%of US working-age people (Q1). Home to the three leading labs and most enterprise spending; 55.7% of Ramp's US business customers pay for an AI tool10.
Europe
~50%in Ireland (49.9%) and France (49.6%), Q2. EU AI Act: model-provider duties since August 2025 and transparency duties since August 2, 2026; high-risk obligations moved to December 2027 and August 2028 by the Digital Omnibus7172. Mistral, based in Paris, is the region's best-funded model developer.
Gulf
73.3%in the UAE (Q2), first in each Microsoft release since 2025, backed by national AI policy and government targets for AI-agent services70.
China
499Mmonthly users of AI-native apps (May), on domestic products: Doubao, Qwen, DeepSeek. Chinese open-weight models also carry much of the world's developer token volume3612.
Asia-Pacific
64.3%in Singapore (Q2), second worldwide. South Korea, Thailand and Japan posted the largest gains in Q1, which Microsoft links to better performance in Asian languages68.
Global South
15.4%of working-age people (Q1), against 27.5% in the Global North, a gap of 12.1 points that widened during the quarter. Connectivity, electricity, skills and language are the binding constraints69.
Events and policy in 2026
FebruarySpaceX merges with xAI65. Anthropic raises $30B at $380B, with a $14B run rate2362. OpenAI raises $110B from Amazon, Nvidia and SoftBank34.
Early 2026The US Department of Defense blacklists Anthropic after talks over military use of its models break down (reported)73.
May 28Anthropic raises $65B at a $965B valuation62.
June 1 and June 8Anthropic and OpenAI confidentially submit draft S-1s6374.
June 12 – July 1A Commerce Department export-control directive leads Anthropic to suspend access to Claude Fable 5 and Mythos 5; the controls are lifted June 30 and access returns July 175.
JuneSpaceX lists on Nasdaq; on June 16 it agrees to buy Cursor for $60B in stock46.
July 27 and August 2The EU Digital Omnibus enters into force, deferring high-risk rules; AI Act transparency duties apply from August 271.
September 12 and 18Altman says OpenAI will not go public in 2026. Investor materials show $278B of projected negative free cash flow for 2026–3060.
September 29OpenAI DevDay: ChatGPT at 1.2 billion weekly users4; run rate reported near $70B1.
Mid-October (expected)Earliest start of Anthropic's IPO marketing, per Reuters; the public S-1 must be filed at least 15 days before64.
What to watch next
Anthropic's public S-1. The first audited revenue for a frontier lab, its gross-vs-net policy, margins, and how much revenue depends on Amazon and Google.
OpenAI's Q3 numbers. Whether booked revenue tracks the ~$70B run rate, and whether cash burn stays near the $25B forecast for 2026.
Gemini monetization. Google reports users, not Gemini revenue or paying subscribers.
Price pressure. Open-weight models' share of tokens, and how far frontier prices fall.
Backlog conversion. About $1.7T of contracted cloud revenue against roughly $748B of capex this year.
Policy. Export controls, defense procurement, and EU transparency enforcement.
Questions for discussion
Choose the scoreboard you would use to value OpenAI and the one you would use to value Anthropic. Are they the same? What does each choice assume about where profits will come from?
Use the gross-vs-net slider. What haircut would put OpenAI ahead in each period? What else would you need to know before deciding which company is larger?
ChatGPT has the largest audience; Anthropic books more revenue. Sketch revenue per user and per business customer for each. Which model holds up better against open-weight price pressure?
The four platforms plan to spend about four times the labs' combined run-rate revenue this year. If demand slows, who carries the loss: the labs, the clouds, or the chipmakers?
Reading the numbers
Run-rate revenue
Recent revenue scaled to a year (four weeks × 13, or a month × 12). A measure of current pace, not revenue earned.
Recognized revenue
Revenue booked in a period under accounting rules. Lags a fast-growing business.
Gross vs net
Booking a partner-channel sale at full value with the partner's share as an expense (gross), or at the seller's share only (net). Both are permitted; gross booking raises revenue and expenses by the same amount, leaving operating profit unchanged.
Weekly vs monthly users
A weekly active user returned in the last seven days; a monthly active user in the last thirty. The same product has more monthly than weekly users.
Web traffic share
Share of visits to chatbot websites in Similarweb's panel. Excludes apps, assistants built into operating systems, and API traffic.
Business adoption (Ramp)
Share of Ramp's customers that paid a vendor in a month. Says nothing about how much they spent.
API share (Menlo)
Modeled share of enterprise use of model APIs, from a survey of technical leaders.
Token share (OpenRouter)
Share of tokens processed on one marketplace. A cheap token counts the same as an expensive one.
Data notes
Every number in this briefing was entered once, with its source, date and basis, in a Python script that writes a single data file; the page and charts are generated from that file, and the build checks internal consistency (shares sum to at most 100, derived ratios recomputed). Where outlets disagree, the text names the alternative: Anthropic's Q2 revenue is $11.5B in Bloomberg's preliminary figure and $11.6B in the Wall Street Journal; Similarweb's May 2026 Gemini share appears as 27.3% and 27.8% in two reports. Menlo's mid-2025 report described its shares as usage; some later summaries describe the December figures as spend, so this briefing calls them API use. Approximate dates are marked with ~ in the tables. Country-level user counts and US-only web-share splits that circulate in secondary summaries were left out because no primary source could be found.
Research and drafting assistance came from Claude, a model made by Anthropic, one of the companies covered here. All figures were taken from the press reports, filings and company statements listed below.
The Decoder. OpenAI and Anthropic before the IPO: different balance sheets make comparison difficult (citing The Information). 2026-03. https://the-decoder.com/?p=33666