Anthropic's IPO paperwork is sealed. The $65 billion run rate everyone's quoting is a leak, not a line.

There is a difference between a number a company reports and a number a company leaks, and the entire Anthropic trade right now is being conducted in the second kind. The company has confidentially filed to go public — a draft S-1 with the SEC, a listing reported for as early as October — and the figure doing the work in every write-up is a $65 billion annualized run rate. Sit with that phrase. Not $65 billion in revenue. $65 billion annualized: a recent, strong stretch multiplied out to a year as if the year will behave. It might. But 'run rate' is what a round number wears when it wants to be taken for a fact, and I have been doing this long enough to check the collar.

Start with the word 'confidential,' because it is carrying more than people notice. A confidential filing means the disclosures — the audited revenue, the cost of revenue, the customer concentration, the compute obligations, the losses — are sitting with the SEC and not with you. What's public is what the company and its friends have chosen to make public, which is the growth number and not the bill behind it. That is not a scandal; it is completely legal and completely ordinary. It is also exactly the condition under which a market talks itself into a price. You are pricing the trailer, not the film.

So let me do the part of this job that keeps me honest, which is to say plainly what is real. Anthropic's revenue did not inch up; it multiplied. The reported run-rate ladder across 2026 climbs from roughly $9 billion at the start of the year to $14 billion in February, $19 billion in March, $30 billion in April, and into the forties by late spring, on the way to the $65 billion now being quoted. Its second quarter revenue was reported at around $11 billion, against something like $800 million in the same quarter a year earlier — a fourteen-fold jump. That is not a rounding artifact or a financing mirage. Enterprises are paying Anthropic real money for a product that does real work, and reportedly about four-fifths of the revenue comes from enterprise contracts rather than consumer whims. A great many bad trades are built on top of a real thing. This is the real thing. Remember that I said so, because the next several hundred words are going to sound like I don't believe it.

What a run rate hides

A run rate is a snapshot annualized, and the faster you're growing, the more flattering the snapshot. Take the numbers as reported: if the current quarter is running at a $65 billion pace, the actual trailing-twelve-month revenue — the sum of what came in over the past year, which is the figure a public shareholder eventually gets to see — is a great deal smaller, because most of that year was spent at $20 billion, then $30 billion, then $40 billion. The run rate isn't lying. It's just describing the exit velocity of a single quarter and inviting you to assume the next four look like it. In a business compounding this fast, the gap between 'annualized from last quarter' and 'earned over last year' can be the difference between two entirely different companies.

This is not a new trick, and I keep a file of the old ones. In 1999 and 2000 the fashion was the 'run rate' and its cousin the 'pro forma' — numbers that annualized a good stretch or excluded the inconvenient parts — and the people quoting them were not all liars. Some of those companies were real; a few are still around. The internet was not a hoax. Pets.com still went to zero. The lesson was never 'the technology is fake.' The lesson was that a great technology and a sane price are two different claims, and a bull market's favorite move is to let the first one smuggle in the second. AI is the realest thing I've covered in a decade. That tells you nothing, by itself, about what Anthropic is worth.

A bad company or a bad price

These are not the same question, and confusing them is how people lose money in both directions. Is Anthropic a bad company? On the evidence, plainly not: a product people pay for, revenue that compounds, gross margins reported around 50 percent, and enterprise customers that don't churn the way consumers do. Is $965 billion a sane price? That is a completely different sentence, and nobody quoting the run rate has shown their work on it.

The $965 billion comes from the Series H round that closed in the spring — $65 billion raised, a post-money mark that stepped neatly over OpenAI's $852 billion, because in this market the valuations don't just rise, they leapfrog. A private round sets a price the way a single auction bidder sets a price: it tells you what the most motivated buyer in the room would pay for a slice, not what the whole thing is worth in daylight. The public market is a different room, with a different memory, and it is going to ask a question the private rounds never had to. Against that price, what are the obligations?

Here is the number I'd put next to the run rate, and it is not one anyone leaked with the same enthusiasm. Anthropic has reportedly committed to something on the order of $80 billion in cloud and compute costs through 2029 — up to five gigawatts of capacity with Amazon, and another five gigawatts of next-generation TPUs with Google and Broadcom. Gigawatts. The unit of a power grid, not a software company. Revenue growing sevenfold is the line you're sold; a multi-year, multi-tens-of-billions compute obligation is the line that comes with it, and a model lab's cost of revenue is not a rounding error — it is the business. The whole bull case rests on the reported claim that gross margins march from about 50 percent today toward the high 70s by 2028. Maybe they do. But that is a forecast wearing the costume of a fact, and the compute contracts are signed in ink while the margins are still in a slide deck.

Revenue growing sevenfold is the line you're sold. Ten gigawatts of compute you've promised to pay for is the line that comes with it.

There's a strength in the numbers worth naming, because I'm not here to cry bubble, and it's the enterprise mix. Reportedly around 80 percent of revenue comes from enterprise, with more than a thousand customers spending at least $1 million a year. That's a sturdier base than a consumer app riding a fad — enterprises sign contracts, integrate deeply, and don't leave on a whim. But concentration cuts both ways. Revenue leaning on large enterprise commitments is revenue exposed to the moment those same CFOs decide the experiment has a budget after all, or that the open-weights model running in their own data center is now good enough for two-thirds of the tickets. A strong customer base and a captive one look identical right up until the renewal.

What's already priced in

The useful question is never 'is it a bubble.' It's 'what does this price already assume, and what has to keep being true for it to hold.' At $965 billion — before whatever the public listing adds — the price assumes the run rate is a floor and not a peak; that a company reportedly turning only a slim operating profit in one recent quarter converts that into the margins of a mature software business; that $80 billion of compute gets paid out of revenue that keeps compounding rather than plateauing; and that the two or three labs at the frontier don't spend the next three years cutting each other's prices to the bone. Every one of those can come true. The point is that all of them are already in the price. You are not being offered the upside. You're being asked to underwrite the assumptions.

You can see the round number doing its work in the comparisons, too. The listing is being described as one that could rival SpaceX's record-setting debut from earlier this year — a sentence that sounds like analysis and functions like a target. It anchors you to a very large figure and invites you to treat Anthropic's eventual market cap as a race to match it. But 'as big as the biggest one we just had' is not a valuation method; it's a mood. SpaceX has a launch monopoly and hard assets in orbit. Anthropic has software, contracts, and a compute bill. Pricing the second by reference to the first is exactly the category error a bull market loves, because it changes the subject from 'what is this worth' to 'why shouldn't it be worth as much as that.'

And the assumption I'd watch hardest is the one about price — the actual per-token kind. The trouble with selling intelligence by the token is that your competitor can wake up tomorrow and charge less for it, and several of them have. Anthropic made its own Sonnet pricing cut permanent earlier this year; the open-weights labs keep shipping models good enough to cap what anyone can charge for merely-very-good; and a market where the product is fungible and the buyers are enterprises with procurement departments is not one that hands you 77 percent margins without a fight. The revenue is real. Whether the pricing that produced it survives contact with three rivals and a rack of open models is the whole question, and it is not one a run rate can answer.

I should say, because it's the rule I hold myself to: I have been early on this before, and early is a way of being wrong that I refuse to launder into being right. I called the AI-capex math circular a good year before anyone else got nervous about it, and for most of that year the stocks went up and I looked like a scold. Being early cost real money to anyone who acted on it too soon. So take this in the right spirit. I am not telling you Anthropic is overvalued and the listing is a short. I'm telling you that the number the whole trade is built on is a run rate, that the figures which would let you check it are sealed inside a confidential filing, and that a $965 billion price has an $80 billion bill and three hungry competitors standing quietly behind it.

Here's what I'll be watching for. Confidential filings don't stay confidential; the S-1 goes public before the stock does, and on that day the run rate everybody's quoting gets set next to a trailing-twelve-month number, a real cost of revenue, a real line for those compute commitments, and a real accounting of losses. That is the day the trailer meets the film. Sometimes the film is better than the trailer promised — it happens, and Anthropic might be that. But the reliable pattern, the one in my file, is that 'run rate' quietly becomes 'restated,' and the people who paid the round number find out what the audited one was. The company is real. The technology is real. The revenue is real. None of that was ever the question. The question, the way it always is, is what you paid — and right now you're being asked to pay for a number the company hasn't yet had to sign its name to.