Anthropic just passed OpenAI in revenue. $30 billion to $25 billion. Four months ago the gap was $9B to $24B · a 3.3x jump for the challenger in a single quarter. 1,000+ enterprise customers now pay a million dollars a year on Claude · doubled in under two months. Eight of the Fortune 10 are Claude customers. For three years we watched the wrong scoreboard · daily active users told us one story, weekly enterprise invoices told the real one.
On April 6, Anthropic posted what looked like a routine partnership announcement · an expanded TPU compute deal with Google and Broadcom, 3.5 gigawatts of capacity coming online from 2027. Inside the same post, buried below the infrastructure news, sat the revenue number. $30 billion annualized. Up from $9 billion at the end of 2025.
Do that math once. Do it again. They tripled in four months.
Run-rate revenue is the best recent month annualized, not trailing twelve-month collections. Both companies report on that basis. Neither is profitable. Both have caveats. We will get to all of it. But the trajectory is not a rounding error. Here is Anthropic's curve, straight from CEO Dario Amodei at Davos and the Series G disclosures:
$87 million annualized in January 2024. $1 billion by December 2024. $5 billion by August 2025. $9 billion by the end of 2025. $14 billion by February, $19 billion by March (Amodei confirmed this at a Morgan Stanley conference), $30 billion by April. Roughly 10x per year, three years running. At the end of 2023, OpenAI's revenue was 15 times Anthropic's. Today Anthropic has passed it.
The scoreboard we kept watching
Consumer usage says OpenAI won. ChatGPT has roughly 900 million weekly active users. 50 million paying subscribers. 60–65% share of the consumer AI market. The product is a verb now · "let me ChatGPT it." That is the kind of brand equity that should be unassailable. Meanwhile ChatGPT's share of AI-product traffic has dropped from 77% to 57% in twelve months. Gemini quadrupled to around 25%. Claude nearly tripled in three months. Those are the numbers people quote on TV. They are true. They are also the wrong place to look.
Menlo Ventures' 2025 State of Generative AI report nailed the shift before most people clocked it. Anthropic's share of enterprise LLM spend went from 12% in 2023 · to 24% in 2024 · to 40% in 2025. Over the same window, OpenAI's share fell from 50% to 27%. The lines crossed in mid-2025. The revenue numbers are now catching up to the share numbers.
Inside that, coding is the real story. Enterprise AI spending on coding hit $4 billion in 2025 · 55% of departmental AI spend, the single largest category across the entire application layer. In that category, Anthropic commands 54% market share. OpenAI has 21%. More than double the competition in the highest-value enterprise AI use case.
Why it actually happened
Here is the line that changes everything. Dario Amodei on stage at Davos in January: "One of the good choices we made early was to be a company that is focused on enterprise rather than consumer. It is very hard to fight your own business incentives. It is easier to choose a business model where there is less need to fight your own business incentives."
That is not a product strategy. That is an organizational design decision made five years ago that is now paying out. Consumer AI has a specific gravity · optimize for engagement, retention, virality, ads, time-on-app. Enterprise has different gravity · a Fortune 500 CTO signing a three-year contract wants reliability, trust, auditability, predictable capability curves. Those requirements pull the model in the same direction as the safety research Anthropic was already doing. The business model and the mission were, in Amodei's phrase, synergistic.
You can see this synergy play out in the customer list. Eight of the Fortune 10 are Claude customers. Uber deployed Claude across software engineering, data science, finance, and trust and safety · wall to wall. Salesforce rolled it out to their entire global engineering org. Stripe is an enterprise Claude customer with no ChatGPT deployment · and notably has over 100 former Stripe employees now on Anthropic's staff. Netflix, Spotify, Snowflake, Rakuten, Novo Nordisk, Ramp, Instacart, Honeycomb, Accenture (tens of thousands of developers) · all in production.
What the scoreboard you were watching says · and what the invoice scoreboard says.
Three signals inside the same data
Past the pilot phase.
An Andreessen Horowitz survey this January captured the consolidation clearly: 75% of Anthropic's enterprise customers have Claude running in production, versus 46% for OpenAI. Anthropic's customers are not experimenting. They are running their business on it. The revenue number is catching up to a deployment reality that has been building for 18 months.
The killer use case.
Coding is 55% of enterprise AI spend · the single largest category across the entire application layer. Anthropic has 54% of that category. OpenAI has 21%. Claude Code's run-rate alone crossed $2.5 billion and more than doubled since the start of 2026. One product drove the bulk of the $9B to $30B climb in six months. Developer habits are sticky in ways consumer apps are not.
The accounting counter-punch.
OpenAI's CRO Denise Dresser told staff that Anthropic's run-rate is inflated by roughly $8 billion via gross revenue accounting on cloud-reseller sales. Under her stricter convention Anthropic drops from $30B to $22B · still ahead of OpenAI's net-reported $24B only on the margin. The accounting dispute changes the scoreboard. It does not change the 3.3x trajectory from $9B.
The compute bet that backs it up
You do not sign a 3.5 gigawatt TPU commitment through 2031 because you are bluffing about demand. Anthropic's expanded deal with Google and Broadcom gives them access to roughly 3.5 gigawatts of next-generation TPU capacity starting in 2027. A single gigawatt is enough to power about 750,000 homes. Three and a half for one customer is industrial-scale infrastructure. Broadcom's regulatory filing flagged it as a business risk factor contingent on Anthropic hitting commercial milestones. Translation · the compute partners are betting on the revenue, and the revenue is betting on the compute.
The weirdest part is the efficiency ratio. Anthropic trains at one-quarter the cost of OpenAI. Anthropic's 2025 training budget was approximately $5 billion. OpenAI's was approximately $20 billion. Four times the efficiency, leading the enterprise market. WSJ obtained confidential IPO-prep financials. OpenAI projects spending $121 billion on compute in 2028 alone, with $85 billion in losses that year. Breakeven slips past 2030. Anthropic's training costs peak around $30B in the same window and positive free cash flow by 2027.
The lesson is older than AI
This is not the first time enterprise-first beat consumer-first, and it will not be the last. Look at the 2000s cloud wars. AWS launched S3 in 2006 with no marketing spend, no celebrity endorsement, no billboards. For years Silicon Valley snickered at "just S3" while Google and Microsoft raced to dominate consumer cloud storage. Ten years later AWS was generating more operating income than Amazon's entire retail business. The quiet B2B flywheel compounded while everyone was watching the loud B2C contest.
Or go further back. In 1985, Oracle had a fraction of IBM's mindshare. Oracle sold a weird database to bank back offices. Fifteen years later Oracle ran the enterprise software world and IBM was pivoting to consulting.
The consistent pattern · the loud brand wins the moment. The quiet builder wins the decade. It happens in cloud, it happens in databases, it happens in payments infrastructure, and now in AI. Consumer revenue is fragile · it churns on vibes, ad-blockers, policy changes, competitor launches. Enterprise revenue is durable · it churns on procurement cycles measured in years, not weeks.
Before you buy, renew, or expand any AI contract · run every vendor through three questions.
- Who else pays them a million dollars a year · and why? Ask for three reference calls with customers at your revenue tier. Ask what they had to do to get the vendor's support team to respond at 2 AM. The answers predict your vendor's next five years better than any benchmark chart.
- What is the production deployment rate, not the pilot rate? Usage numbers are vanity. Production deployments are revenue. A vendor where 75% of enterprise customers are in production has a different risk profile from one at 46%. Get both numbers before you sign.
- What is our migration plan if this vendor has a six-month quality regression? Most AI stacks are one bad quarter away from a fire drill that no one has scripted. Treat your model layer like your database layer · best tool per job, swap when the math changes, never let a single vendor hold the whole stack hostage.
The quiet builder just ate the loud leader. Stop watching the demos. Start watching the invoices.
For three years we watched the wrong scoreboard. Daily active users told us one story. Weekly enterprise invoices told the real one. Anthropic won by choosing a business model that aligned with its technical mission, serving the customer whose procurement timeline compounds, and letting 18 months of coding-benchmark dominance turn into 1,000+ million-dollar contracts. The loud brand wins the moment. The quiet builder wins the decade. The question is not who wins the next snapshot · it is whether you are paying attention to the snapshot that matters.