The word institutional buyers used for a $1.5 trillion valuation was “barely”. A month earlier the number in circulation had been $2 trillion.
Investors preparing to take part in Anthropic’s initial public offering have cut their expectations this week to a range of $1.2 trillion to $1.3 trillion, with some looking at $1 trillion, according to CNBC. The company starts its formal roadshow next week.
The compression is the story. Between the $2 trillion discussed a month ago and the $1.2 trillion now on the table sits a 40% markdown, arrived at before a single page of the prospectus has been marketed. That is what price discovery looks like in reverse: the buyers moved first, and the company has to decide whether to meet them or delay.
Bankers do this work in private before the public marketing starts, calling the anchor investors whose orders set the shape of the book. Those conversations are where a range is made, and a range that lands 20% below the number buyers were already calling marginal is an admission that the earlier figure was aspirational.
The comparison with OpenAI is unavoidable. OpenAI is raising privately at a valuation of $1.25 trillion, and Anthropic’s new range puts the two labs within a rounding error of each other. Two companies being priced at almost the same amount, in the same month, through entirely different mechanisms, is a coincidence that helps the underwriters and complicates the argument that either business is worth much more than the other.
What the valuation actually asserts is a claim about enterprise revenue. Anthropic sells mostly to businesses, through cloud partners and direct contracts, which makes its revenue more predictable than a consumer app’s and makes its growth dependent on how quickly companies move from pilots to production. A trillion-dollar price assumes that migration happens on schedule and that the company keeps its share of it.
The demand data published the same day argues for patience rather than urgency. TechCrunch reported, citing the payments company Ramp, that enterprise AI spending has flattened as prices fall and finance departments tighten their grip on costs. If corporate budgets for AI are leveling off, the growth that justifies the price has to come from more seats at the same price, or from workloads that were experiments last year becoming line items this year.
There is a counterargument, and it is a fair one. Ramp measures card and expense data from a sample of companies, which captures what firms pay each month rather than what they have contracted for. Falling prices can make revenue look flat while usage climbs sharply, and a model that gets cheaper per token can be a worse revenue line and a better product at the same time. The spending data and the usage data can point in opposite directions without either being wrong.
The structure of the offering matters as much as the range. Underwriters prefer to price a deal below where they expect it to trade, leaving the first buyers with a gain and the syndicate with a cushion if sentiment turns. A $1.2 trillion range does not mean the company is worth $1.2 trillion. It means the people selling the deal want the first day to go well, and they have learned from a year of listings that opened strong and faded.
The window is the risk nobody prices until it closes. A listing of this size needs buyers who are willing to hold a volatile asset through a period when the cost of debt for AI infrastructure is rising. SoftBank’s credit default swaps widened to a four-year high this week and its shares fell 11% in Tokyo. CoreWeave’s bonds yield above 16%. Those are the financing conditions a growth company’s investors read before they commit, and they are not friendly.
Anthropic has an advantage that the lenders do not: it does not need to borrow to build. Its compute is bought largely from cloud partners that carry the capital cost, which keeps its own balance sheet light and its obligations contractual. That is a stronger position than a company that owns its data centers, until the contracts come up for renewal and the partner decides the price was set too low.
The roadshow will test something the private rounds never have. Anthropic has to answer questions in public, on a schedule, with numbers attached and lawyers present, and the answers become the record that every future quarter is measured against. A private valuation can be renegotiated quietly. A public one is repriced in public, every day, by anyone with an account.
That is the trade the company is making. It gives up the ability to manage its own mark in exchange for capital that does not need a syndicate, and it accepts that the first number it prints will be compared with OpenAI’s private mark for as long as both companies exist.
Whether $1.2 trillion is the right answer will not be settled next week. What will be settled is whether the buyers who cut their expectations this week are the same people who show up with orders, and whether the company is willing to accept their price to get the listing done in this window rather than the next one.
