When Anthropic filed to go public, the number at the top of the range was $2 trillion. This week, the institutional investors invited into Anthropic’s listing were talking about half a trillion dollars less, and the shift had little to do with anything the company itself disclosed.
The Wall Street Journal reported on Nov. 2 that investors preparing to take part in the initial public offering had pushed their pricing expectations to between $1.5 trillion and $1.6 trillion, after a selloff in technology shares the previous week. When Anthropic first filed to go public, the upper bound under discussion in the market was $2 trillion, a figure that would have made the listing the largest in American history on its first day of trading.
Nothing in the business deteriorated between the filing and this week. The company’s revised registration statement, filed the same day as the Journal’s report, showed third-quarter revenue of $17 billion for the three months ended Sept. 30, an increase of 1,127% from a year earlier. Claude, its family of models, accounted for essentially all of it.
The growth rate is where the argument starts. Anthropic booked $11.6 billion in the second quarter, so sequential growth slowed from 144% to 47% in a single step. A company that more than doubles in one quarter and grows by half in the next is still expanding at a pace almost no software business has sustained, but the curve bends, and public investors pay for the shape of the curve rather than the level.
Margins tell a similar story. Adjusted operating income was $2.5 billion, a 15% margin. On a standard accounting basis the company recorded an operating loss of $567 million, because operating expenses rose from $10.6 billion to $17.5 billion. Anthropic attributed part of the increase to short-term computing contracts that priced well above multi-year agreements, and part to a $2 billion non-cash charge tied to employee donation matching.
The filing also put a number on what the company has promised to spend. Purchase commitments and lease agreements totaled about $895.4 billion as of Sept. 30 and since, an obligation that runs for years and cannot be canceled the way a cloud bill can. Against that, remaining performance obligations, the contracted revenue not yet recognized, came to $13.8 billion at the end of September, with about three-quarters expected to be booked in the following twelve months.
Those two figures, read together, describe a company that sells mostly on demand rather than on long contracts, and that has locked in capacity on the assumption the demand continues. That is a reasonable bet for a lab whose models sit inside other companies’ products. It is also the kind of bet a public shareholder has to underwrite quarter by quarter, which is a different exercise from underwriting it once, in private.
At the lower end of the new range, the company would be valued at roughly 22 times its annualized third-quarter revenue. At the figure discussed when it filed, the multiple would have been close to 29 times. Neither is unusual for a software company growing this fast. Both assume the deceleration is a pause rather than the start of a normal curve.
The buyers who moved the range are the ones who will set the price. Institutional accounts that allocate to large listings compare each new name with what they already own, and in the past two weeks what they own has been repriced. A broad decline in technology shares lowers the comparable set, and a lower comparable set lowers the multiple an investor will accept, whatever the company’s own numbers say.
The cut also shows how price talk works. The range that circulates before a roadshow is not a valuation the company has agreed to. It is a set of conversations between bankers and prospective buyers, and it moves when the buyers’ other holdings move. A fifth to a quarter off the top of the range is a large revision, but it is a revision of expectations rather than of the business.
What a lower price changes is arithmetic downstream. A company raising a fixed sum at a lower valuation issues more shares, which dilutes existing holders, including employees whose options were struck against a private mark. A company that holds the raise constant and accepts the dilution is saying that the capital matters more than the price. Neither choice is free, and the filing does not say how much Anthropic intends to sell.
For the wider market, the comparison is the thing to watch. Anthropic is among the first of the large model developers to test public demand for frontier AI at a scale that requires index funds to own it. A listing at either end of that range would give every private AI company a reference point and every public investor a way to own the theme directly. A listing that prices at the bottom of the range, or breaks it, would tell the next company in line that the window is narrower than the pitch decks suggest.
The range will be fixed after the roadshow, and the roadshow has not begun. Until then, the figure buyers are discussing is a number they say they will pay, not one the company has accepted, and the gap between those two things is where every large listing is actually decided.
People familiar with the discussions said the revision reflects the market rather than the issuer, and that no final range has been set. Bankers typically begin a roadshow with a range they believe they can beat, which is why the early conversations are treated as a floor rather than a ceiling.
