Anthropic has told investors it booked more than $11.5 billion in revenue in the second quarter and turned positive adjusted operating income — the first time a frontier AI lab has put a profitable quarter in front of the people writing the checks. The figures come from internal documents viewed by Bloomberg News and reported Aug. 14, are labeled preliminary, and are not audited. They landed in the middle of a pre-IPO campaign in which Anthropic's chief financial officer is meeting early investors, which is the first reason to read them carefully and the second reason they exist at all.
The number has moved twice. In May, the Wall Street Journal reported that Anthropic's revenue was "set to more than double to $10.9 billion in the second quarter, an explosive rate of growth that will help it turn an operating profit for the first time," and put that profit at $559 million for the June quarter against $4.8 billion of first-quarter sales. CNBC reported the same $10.9 billion projection the same day. Three months later, the actual figure Anthropic is circulating is above $11.5 billion — up more than 14-fold from $787 million a year earlier — with the first quarter restated slightly to $4.73 billion. Anthropic has confirmed none of this publicly. The company has not filed an S-1 and is under no obligation to publish anything.
That distinction matters more than the headline. The Journal noted in May that "it is unclear what accounting methods Anthropic has used to book revenue and costs, as the company isn't yet required to follow the financial-reporting requirements of a public company." The metric in the August documents is adjusted operating income, a non-GAAP figure that typically excludes stock-based compensation. No outside auditor has signed anything.
The cost story, and the argument against it
The bullish reading is that inference got cheap enough, fast enough, to outrun the compute bill — that Anthropic's per-token serving costs fell while enterprise demand for Claude Code and agentic workloads compounded. There is partial support for the demand half: Anthropic reported $14 billion in annualized recurring revenue in February alongside a $30 billion Series G at a $380 billion post-money valuation, and claimed roughly $30 billion ARR by early April.
The cost half is contested. Writer Ed Zitron, who has tracked the labs' unit economics for two years, argued in May that the timing was engineered. Anthropic's compute agreement with SpaceX for the Colossus-2 cluster commits it to $1.25 billion a month — about $15 billion a year — but at a reduced ramp-up rate covering May and June, precisely the months in which Anthropic projected its profit, with the full rate starting in July. SpaceX's own S-1 describes the arrangement.
"While I wouldn't say this is cooking the books, it's definitely a shiatsu-grade massaging of the numbers," Zitron wrote. "That operating profit is a result of accountancy rather than any improvements to its business model."
He also flagged an unresolved reconciliation problem: on March 9, CFO Krishna Rao declared under oath in a federal court filing that Anthropic had brought in revenues "exceeding $5 billion to date" — hard to square with leaked run-rate figures implying the company earned most of its lifetime revenue in a single quarter. Separately, The Information reported in January that Anthropic had missed its own gross-margin projections, with inference costs running 23% higher than anticipated. Neither point disproves the quarter. Both argue against treating it as settled.
Anthropic's own CEO has been more hedged than his boosters. "We could be profitable in 2026 if the revenue grows fast enough," Dario Amodei told Dwarkesh Patel, framing the question inside what he calls a cone of uncertainty created by data centers built on one-to-two-year lags. In a separate conversation with Stripe co-founder John Collison, Amodei made the accounting argument directly: train a model for $100 million in one year, deploy it the next for $200 million in revenue, and "if you consider each model to be a company, the model that was trained in 2023 was profitable." Company-level losses, in that framing, are the cost of the next model, not evidence the current one loses money.
Why It Matters
The two leading labs are now running visibly different financial theories, and both are about to be tested by public markets.
Anthropic's is enterprise-first and margin-forward: sell coding and agentic capacity to companies at prices that clear the compute bill, show a profitable quarter before the roadshow, and let the multiple follow the P&L. Rao has been holding early investor meetings — CNBC reported Aug. 13 that he has deliberately not discussed valuation, which is standard before underwriters are picked — with a listing targeted for the fall.
OpenAI's is scale-first and loss-tolerant: annualized revenue reached roughly $25 billion in February, up from about $9 billion at the end of 2025, against a projected 2026 loss near $14 billion and reported cumulative losses running toward $115 billion by 2029. It was valued at $852 billion privately in March. Sam Altman has called any IPO valuation under $1 trillion a "nonstarter," and the New York Times has reported the listing could slip to 2027 as a result.
Neither model is obviously correct. Anthropic's profitable quarter may be a two-month discount window; OpenAI's losses may be the correct price of owning the consumer surface. What has changed is that investors now have two priced alternatives instead of one narrative.
What to Watch
The third quarter is the real test, because it is the first full quarter at Anthropic's undiscounted SpaceX rate — the company itself has signaled it may not stay profitable for the full year given planned compute spending. Watch for an S-1, which would replace leaked adjusted figures with audited GAAP ones and settle the Rao reconciliation question. Watch whether OpenAI accepts a sub-trillion listing or waits. And treat every number above as reported rather than proven until one of these companies files.
“We could be profitable in 2026 if the revenue grows fast enough.”— Dario Amodei, CEO, Anthropic