Anthropic has told prospective investors that its revenue in the second quarter of 2026 topped $11.5 billion. That is more than fourteen times the $787 million it booked in the same quarter a year earlier, and more than double the $4.73 billion it did in the first quarter of this year. As a growth number it is almost hard to read.

But the number that actually matters for the IPO is the smaller phrase sitting next to it: positive adjusted operating income. Among the current crop of frontier LLM labs — Anthropic, OpenAI, xAI — it is the first to show investors a profit line ahead of a public listing. (Google DeepMind turned an accounting profit back in 2020, but as an internal Alphabet unit that never had to answer to public shareholders — a different thing entirely.)

Why one word is doing so much work

For two years the bear case on the frontier labs has been simple: enormous revenue, even more enormous losses. Training runs, inference, talent and compute leases have swallowed every dollar coming in and then some. "Real business or subsidised demo?" has been the open question hanging over every AI valuation.

An adjusted-operating-profit quarter is the first hard swing at that question — and Anthropic is showing it to investors now, ahead of a listing, for exactly that reason. The disclosure is not an accident of timing. With Morgan Stanley, Goldman Sachs and JPMorgan running the book and a confidential filing already in, this is the listening-tour phase: you show the market your best number and watch what price it will bear.

The caveats a buyer should read first

Before anyone reprices the sector, three words deserve scrutiny.

"Adjusted." Adjusted operating income is a non-GAAP figure: it is not net income, and it is not free cash flow. Measures like it typically strip out items such as stock-based compensation and one-offs — and Anthropic has not disclosed exactly what its version excludes. At a company that pays much of its talent in equity, stock comp is not a footnote. A positive adjusted operating line can sit comfortably above a deeply negative actual bottom line. Until the audited S-1 lands, we do not know how wide that gap is.

"Preliminary." These figures are exactly that. They can move before they are ever filed, and numbers disclosed on a listening tour are the ones a company most wants you to see.

The compute bill. An operating-level profit says almost nothing about the cash going out the door for chips and data centres. The capital intensity of frontier AI lives largely below the operating line — in the leases, the build-outs, and the multi-year compute commitments. A lab can be "operating-profitable" on an adjusted basis and still be burning cash at a spectacular rate to fund the next model.

What is actually new here

Strip the caveats away and something real remains. The latest reported run-rate revenue — a separate, more recent figure than the quarter itself — has moved past $65 billion heading into the IPO, on an enterprise-heavy mix. Anthropic's business is overwhelmingly companies paying for Claude through the API, not consumers on a free tier — the rough inverse of OpenAI's consumer-heavy shape. Revenue at this scale, growing at this rate, with any credible path to profitability shown to investors, is a materially different pitch from the one AI companies were making a year ago.

That is the story the IPO is built on: not "we will make money eventually," but "look, here is a quarter where we did." Whether that line survives contact with an audited prospectus — with stock comp, depreciation and the true compute bill all on the page — is the question that will actually set the price. For now, Anthropic has handed the bulls their first real data point, and handed everyone else a number to interrogate.

The listing is targeted for the autumn. The figures above are what the market is being asked to believe. The S-1, when it comes, is where they get tested.