Fears of AI apocalypse crashed into Anthropic's $2T IPO

Eighteen months ago Anthropic's own forecast had it booking $12bn of revenue in 2027. Some investors now pencil in a $320bn run-rate by the end of next year, and that arithmetic is what props up the $2tn IPO the FT has been stress-testing.

At a glance

• Anthropic was poised to file this month for an initial public offering that could have valued it at $2tn, until fears of an AI apocalypse crashed their way into the mainstream.
• On the $320bn run-rate some investors forecast for the end of next year, $2tn works out at roughly seven times 2028 sales, a shade under Microsoft's multiple according to LSEG.
• The bear case is commodity cognition: if one model looks much like the next, the price is set by the cheapest seller and the labs building the engines keep only crumbs.

If you missed the listing chatter, the doom debate landed at an awkward moment. Whether these models could one day wipe out the species has pulled the spotlight off the duller question of what the shares are worth. OpenAI, planning a float of its own, has pushed it to next year, with Sam Altman saying that listing earlier would be "ill-advised".

Seven times sales gets you to $2tn, sixteen times gets you to $5tn

The dull way to value a company works fine when the company makes regular things in predictable ways: forecast revenue a year or two out, then apply a multiple borrowed from listed peers. Anthropic breaks the forecast. Founded in 2021, it booked nothing until 2023, and by August it was running at around $65bn annualised, an unofficial number that deserves kid gloves.

The growth seems to have surprised Dario Amodei too, given that $12bn estimate for 2027. The investors now talking about $320bn own a piece of the outcome, but if they are right, $2tn is about seven times 2028 sales, a little less than Microsoft according to LSEG. SpaceX trades at 16 times that year's revenue; on that benchmark Anthropic would be $5tn.

Pricing a company whose future is this hard to map off near-term estimates sounds flimsy, and it is exactly what Wall Street does. SpaceX's underwriters Goldman Sachs and JPMorgan were among those who set target prices by taking a punt on profit a year or two out and adding a peer-group multiple.

SpaceX put enterprise apps at $22.7tn in its May listing documents

The second method starts from the total addressable market: pick the pool of money a product could theoretically address, assume a slice of it, put the revenue on a multiple and discount it back. It prices a shop by the footfall on the street rather than by the till.

It travels badly. Uber's 2019 listing advertised a $12.3tn market and its annual revenue today is under $60bn. SpaceX's May documents put enterprise apps at $22.7tn, a little more than the entire annual revenue of the S&P 500, and Microsoft's revenue is less than 2 per cent of that.

Give Anthropic 3 per cent of that pool and it makes nearly $700bn; ten times that, discounted over three years, is $4.5tn. Anthropic's filings may tell investors it sees a $30tn market, the Wall Street Journal reported, and Morgan Stanley, likely to be named an underwriter, last week put generative AI's opportunity at twice that.

Take 10 per cent of that $60tn market for knowledge work and digitisable consumer spending, which would involve replacing about 100mn workers, hand Anthropic a third of it, and you get $2tn of annual revenue and, over five years, a $10tn valuation today.

Alex Karp's bear case is that model value slides toward zero

Even if the TAMs are real, it is not obvious that the labs pocket the money. Anthropic and OpenAI compete with Alphabet, Meta, Nvidia, Chinese players and everyone else along the chain. Palantir's Alex Karp calls it commodity cognition: if models look fungible, the cheapest one sets the price, and value accrues to whoever owns the data or builds the app.

Moonshot AI's Kimi and DeepSeek undercut US prices at slightly lower quality, and Meta and Alphabet have advertising businesses big enough to give models away cheaply. Google and OpenAI have already cut prices on some less advanced offerings, customers including some of Anthropic's investors and underwriters run routing software that sends each task to the most cost-effective engine, and Nvidia, an investor in both labs, bought Hugging Face, a major distribution platform for open-weight models.

Anthropic is already moving up the stack. Its plug-in for lawyers, launched in January, set off the so-called SaaSpocalypse, in which Salesforce, ServiceNow and their peers shed a collective $300bn in a couple of days; OpenAI launched a research and writing tool called Astra for Law on Thursday. Marc Benioff calls the SaaSpocalypse "crazy nonsense", yet Salesforce and Anthropic are building Claudeforce together.

Cyber stocks rose double digits after Amodei's call to pace the frontier

Any valuation model here has to carry an unusual risk factor: a potential apocalypse. It is not a new idea. Altman, Amodei and hundreds of other researchers and experts signed the one-sentence Statement on AI Extinction Risk in 2023, and Amodei's call this month for a collective slowdown, plus warnings from former Anthropic researcher Jacob Coxon, cranked the volume up.

If AI really does make humans extinct, every investment goes to zero. Lesser destruction is more tractable: Amodei's recent essay frets about hundreds of billions of dollars in damage, and liability for a mishap on that scale would leave a hole in Anthropic's valuation, to say nothing of its reputation. In the days after the slowdown call, cyber security and virus-scan stocks rose double digits.

Founders argue the best model always carries a premium for the hard 5 per cent of problems, and the labs are racing towards recursive self-improvement, the point at which models upgrade themselves. Anthropic's research arm has posited more than $10tn of extra US GDP by 2030 in its extreme scenario, which simplistic maths turns into $100tn of equity value.

Every route to $2tn runs through numbers nobody outside can check: the $65bn is unofficial, the $320bn comes from people who own a piece of the company, and each share-of-market assumption is picked rather than measured. In our view the weakest joint is the ten-times multiple, applied to revenue that does not exist yet and attached to no named peer; halve it and the $10tn case halves with it.

What the prospectus has to show

Whether Anthropic still files this month is not stated, and Altman has moved OpenAI's float to next year. When a document does appear, three things are worth checking: the size of the market Anthropic claims, against the $30tn the Wall Street Journal reported; whether Morgan Stanley is named as an underwriter; and how the risk factors word the scenario in which every investment, including this one, goes to zero.