Anthropic will be the first pure AI lab to face a public bid. That's either a head start or a trap.

Anthropic is lining up its investor meetings. According to CNBC and Bloomberg, the bankers running the deal — Goldman Sachs, Morgan Stanley and JPMorgan, the three biggest on the Street — have started scheduling the conversations that come before a roadshow, on the back of a confidential filing made last month, with an initial public offering possible as soon as October. The last private round, in May, put the company at $965 billion. That would make Anthropic the most valuable company ever to go public, and it would get there ahead of OpenAI, which has pushed its own listing to 2027. Before anyone reaches for the word bubble, let me say the useful part first: Anthropic is a genuinely extraordinary company. That has never been the question. The question is what you pay, and what the number you are being asked to pay is actually a measurement of.

So let's separate the two things the way you always have to, because conflating them is how people lose money in both directions. There is the company, and there is the price. The company builds Claude, has real customers paying real money, and is growing at a rate that would have been dismissed as a typo three years ago. The price is $965 billion, and it is not a fact about the company. It is a fact about the last people who transacted in it — a small number of private investors, negotiating in a room, none of whom had any interest in marking it low. That distinction is the whole column.

A run-rate is last month wearing a calendar

Start with the growth, because it is the engine under the valuation and it is genuinely staggering. Anthropic reported hitting a $47 billion revenue run-rate around its funding round in late May, up from roughly $30 billion in April, $14 billion in February, and about $1 billion at the end of 2024. Read that trajectory and you can see exactly why investors paid what they paid. You can also see why I want to slow down on the phrase "run-rate," because it is doing more work than most readers realize.

A run-rate is not a year's revenue. It is a recent month or quarter, multiplied out as if the rest of the year already happened at the same pace. It is last month wearing a calendar. When a company is compounding, a run-rate flatters, because it annualizes the fastest month you have ever had and quietly assumes every month after it is at least as good. That assumption has been correct for Anthropic so far — spectacularly so. But it is an assumption, not a receipt, and the gap between the two is precisely where valuations get set and reset. A $965 billion mark on a $47 billion run-rate is about twenty times a number that is itself an extrapolation. Measure it against the revenue Anthropic has actually banked over the trailing year, rather than the annualized snapshot of its best recent month, and the multiple is a good deal larger than twenty. Neither number is wrong. They are just answers to different questions, and the euphoric one always gets quoted.

A run-rate is last month wearing a calendar. It annualizes your best month and assumes the rest of the year already happened.

The first pure lab to meet a public bid

Here is what makes this listing different from every AI-adjacent one that came before it, and why it matters more than the size of the number. The AI companies that have gone public so far have mostly been infrastructure — the neoclouds, the memory makers, the chip names, the picks-and-shovels businesses whose economics, whatever you think of the price, are legible. You can model a company that rents out GPUs. Anthropic is not that. It is a pure frontier lab, and it will be the first one to stand in front of a public market and let strangers set its price. Everything about a lab's economics that stayed comfortably vague while it was private has to be written down in a prospectus that, unlike a pitch deck, carries legal liability if it lies.

Think about what actually has to appear in that document. The gross margin on serving the models — the real cost of the compute behind every Claude response, which is the one number that tells you whether this is a software business or a company reselling someone else's expensive hardware at a markup. The customer concentration, and the awkward fact that Anthropic's fastest-growing product, Claude Code, lives in coding assistance, which is also the single most competitive and most rapidly commoditizing corner of the whole industry, with Microsoft, Google, OpenAI and a dozen open-weight models all pushing the price toward zero. The governance: Anthropic is a public-benefit corporation with a Long-Term Benefit Trust sitting above ordinary shareholders, an arrangement explicitly designed to let the board choose safety over returns. That is admirable. It is also, in the flat prose of a risk-factors section, a disclosure that the people you are buying shares from have reserved the right to not maximize your shares.

And then there is the dependency that a private lab can wave away and a public one cannot. Anthropic's largest investors, Amazon and Google, are also its cloud landlords and its compute suppliers. Money flows in as investment and back out as infrastructure spending, and some meaningful part of that spectacular revenue is being paid by the same companies underwriting the whole enterprise. That circularity is not a scandal — it is how this build-out has been financed across the board. But it reads very differently in an S-1, under oath, than it does on a stage.

The precedent nobody wants to remember

I have a file of first-of-their-kind listings, and the one I keep returning to this week is Netscape, in August 1995. It was the first pure play on a genuinely world-changing technology — the web browser — and its IPO is the event people point to as the moment the internet became an investable idea. The technology was completely real. The browser did change everything. And within a few years Netscape had been run over by a better-capitalized incumbent who bundled a free competitor into every computer sold, and the stock that had made the category investable was effectively worth nothing. The lesson was never that the internet was fake. The lesson was that being first and being real are not the same as being safe, and that the company which opens a category is often not the one that keeps it.

I am not predicting that for Anthropic, and I want to be honest about my own record here, because being early is a way of being wrong and I refuse to launder it into being right. I have looked at AI valuations before and thought the number couldn't hold, and the number held, and then it went higher. If you had shorted the run-rate story at any point in the last two years you would have been carried out. So take my skepticism at the correct weight: I have been early on exactly this trade, and early is wrong until it isn't. What I am describing is not a prediction of collapse. It is a description of what changes when a private mark meets a public bid for the first time.

What actually changes in October

Because that is the real event here, underneath the banks and the confidential filing. A private valuation is a negotiated number. Everyone in the room that set the $965 billion mark was already long, already believed, and had every incentive to agree on a high figure, because a high figure is good for the people who already own the thing. A public price is a discovered number. It is set, every day, by the marginal buyer and the marginal seller — including people who can short it, people who need to sell, and passive funds that don't care about the story at all. Those are different machines, and they do not always produce the same answer. We watched the gap open in real time this year: SpaceX went public in June, priced at $135, ran to the low $200s, was briefly one of the most valuable companies on earth, and gave most of it back within weeks. The company did not change in those weeks. The mechanism setting its price did.

So here is what I would actually watch, rather than the valuation headline. Watch whether the run-rate keeps compounding or simply flattens, because the entire multiple is built on the assumption that it keeps bending upward, and a single soft quarter from the hyperscalers everyone is selling to would turn "annualize the best month" from a tailwind into a trap. Watch the gross margin when it finally has to be disclosed, because that one line decides whether you are buying a software company or a compute reseller. And watch what the first real public bid says, because for two years the price of the AI labs has been set entirely by people who were already believers, and Anthropic is about to do the one thing that private markets never make you do: find out what a skeptic will pay. Anthropic is a remarkable company. I hope the ones building the actual future get properly rewarded for it. I just remember, every time I'm told a number only goes up, to ask the question my father taught me at the dinner table after 2008: compared to when — and priced by whom?