Anthropic planned a $2T IPO until doom fears landed

Eighteen months ago Anthropic's own forecast put its 2027 revenue at $12bn; by August it was running at around $65bn a year. This month it was poised to file for an IPO that could value it at $2tn, until fears of an AI apocalypse crashed into the mainstream, as the FT lays out.

At a glance

• OpenAI has been planning a listing too, and Sam Altman has pressed pause on it until next year, saying that floating earlier would be "ill-advised".
• Some investors put Anthropic's run-rate at $320bn by the end of next year, which would make $2tn roughly seven times 2028 sales, a little less than Microsoft, according to LSEG.
• The catch is who keeps the money: Palantir's Alex Karp calls it commodity cognition, and price cuts, routing software and cheap Chinese models are already pushing model prices down.

If you have not been following the arc: Anthropic was founded in 2021 and had not produced a dollar of revenue until 2023, the same year Amodei and Altman signed the one-sentence Statement on AI Extinction Risk along with hundreds of other researchers and experts. Theories of AI doom have been percolating in online forums for decades. This month Amodei called for a collective slowdown in model development, and stark warnings from former Anthropic researcher Jacob Coxon made those fears loud again.

Seven times sales, if the $320bn arrives

The ordinary way to price a company is to forecast earnings a year or two out, or revenue if there are no earnings, and put that number on a multiple similar to listed peers. Anthropic breaks that habit by changing too fast even by Silicon Valley standards. The $65bn annualised figure for August is unofficial, and the FT says such numbers deserve kid gloves.

Some investors now predict a run-rate of $320bn by the end of next year, and the FT notes they are not impartial. Take the figure at face value and $2tn is about seven times 2028 sales, a little less than Microsoft, according to LSEG. SpaceX trades at 16 times that year's revenue; use that benchmark and Anthropic lands near $5tn.

The method is not exotic. On SpaceX, underwriters Goldman Sachs and JPMorgan were among the many who set generous target prices by taking a punt on profit a year or two out and slapping on a peer-group multiple.

Where $4.5tn and $10tn come from

Everything above that comes from total addressable market, always run through the same recipe: size a market, assume a slice of it, put a multiple on the revenue that slice would produce, and discount back for the years it would take.

SpaceX's May listing 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, nearly $700bn, apply a multiple of 10 and discount three years: $4.5tn.

Morgan Stanley, likely to be named an underwriter, sized the opportunity at $60tn of knowledge work and digitisable consumer spending. Taking 10 per cent of it would involve replacing about 100mn workers; a third of that for Anthropic is $2tn of annual revenue, which at the same 10 times, five years out, is a $10tn valuation today.

Uber touted a $12.3tn TAM at its 2019 listing and makes less than $60bn a year today. Anthropic's research arm has posited more than $10tn of extra US GDP by 2030 in its extreme adoption scenario, which simplistic maths turns into $100tn of equity value.

Alex Karp's bet is that the models themselves go to zero

Anthropic and OpenAI are dab hands at making models, but so are Alphabet, Meta, Nvidia and Chinese players, and everyone along the chain, from customers to chip and data centre providers, will try to claw back some of AI's economic benefit. The danger for the labs is fungibility: if one model is much like the next, the cheapest one sets the price.

Karp calls this commodity cognition and argues the value accrues to whoever owns the data or builds the application the model runs in, leaving the engine makers able to charge enough to cover costs and not much more.

The signs are there. Moonshot AI's Kimi and DeepSeek undercut US prices at slightly lower quality, Meta and Alphabet have advertising businesses large enough to give models away, and Google and OpenAI have cut prices on less advanced offerings. Customers, including some of Anthropic's own investors and underwriters, run routing software that sends each task to the cheapest engine.

Nvidia, an investor in both labs, has acquired Hugging Face, a distribution platform for free-to-use open-weight models. AI founders counter that the best models keep a premium for the hard 5 per cent of problems.

A plug-in for lawyers triggered a $300bn sell-off

If models slide toward zero, the way out is to sell software-like apps, and Anthropic has been moving up the stack. Its plug-in for lawyers, launched in January, triggered the SaaSpocalypse: shares in companies such as Salesforce and ServiceNow shed a collective $300bn in a couple of days. OpenAI launched a research and writing tool called Astra for Law on Thursday.

A $2tn price tag is, after all, less than a third of the combined market capitalisation of software and services companies in the S&P 500. Marc Benioff calls the SaaSpocalypse "crazy nonsense", though Salesforce has since tied up with Anthropic on Claudeforce, which fuses Anthropic's models with Salesforce's customer service wares.

Then there is the risk factor most prospectuses never carry. Amodei's recent essay frets about "hundreds of billions of dollars in damage", and a lab held liable for that would have a hole in its valuation as well as its reputation. In the days after his call to "pace the frontier", cyber security and virus-scan stocks rose in double digits.

Two numbers carry most of the weight here, and both are soft: the $65bn annualised run-rate is unofficial, and the $320bn forecast comes from investors the FT flags as anything but impartial. In our view the weak link is the multiple rather than the market, since the 10 times revenue used in both the $4.5tn and the $10tn case does most of the work, and nothing in the exercise explains why the number is 10.

When the $320bn gets tested

The FT describes the filing as poised rather than made, and names no new date for it; Altman has said OpenAI will not list before next year. The nearest hard checkpoint is the run-rate investors have staked out for the end of next year, $320bn, against roughly $65bn annualised in August. Worth watching too is whether the prospectus really claims a $30tn market, as the Wall Street Journal reported.