The AI Bill Has Fixed Due Dates. The Revenue Does Not

On Monday 26 October Meta announced an equity raise of up to $95 billion and Anthropic made its listing document public. In its Financial Times analysis, columnist Craig Coben sets out the context: AI stocks have fallen in recent weeks, some listings of compute companies have been postponed, and investors reportedly asked Anthropic to lower its expected valuation. On 16 October, in its worst week since June according to CBS, the Nasdaq lost 2.1% and CoreWeave 24%. The ten-year Treasury yield stood at 5.1%, and the Federal Reserve, which had already raised rates in September, had a 74% probability of another hike in December.

The scepticism is not only the market’s. In its 7 July stability report, the Bank of England wrote that AI company valuations have become more stretched and that the sector increasingly funds itself with credit, simulating a downturn that could cut British GDP by up to 2.2 points. In this climate Anthropic is among the first American model labs to show its accounts to a public market: the Chinese labs Zhipu and MiniMax debuted in Hong Kong before it, and OpenAI is not yet listed.

Meta comes from a bill that did not add up. In the third quarter free cash flow was negative by $2.12 billion, the first in more than ten years, according to Reuters, with annual investment planned at $135–145 billion and future commitments of about $825 billion. Those who bought back shares for years, Coben observes, are now issuing them. Reuters adds that Nvidia, AMD and Broadcom have expressed interest in up to 5 billion shares: the suppliers that sell the chips also appear among those buying the stock, the pattern I described when writing about circular demand.

Anthropic has a balance sheet that appears to slide the other way. According to reports on Anthropic’s filing, first-half revenue is $16.3 billion, against $1.2 billion a year earlier, and the second quarter shows its first operating profit. The half-year as a whole, though, closes with an operating loss of $693 million, according to the filed prospectus. Anthropic has raised $141.1 billion so far, and its actual debt at 30 June was $6.2 billion. The net loss of the half, $80.7 billion, is largely accounting: $72.4 billion is the change in value of convertible securities, with no cash going out. The debt, however, is coming, and it comes from suppliers. According to the filing, as reported on convertible notes, Broadcom can lend up to $42 billion, Google up to $20 billion and Amazon up to $15 billion, not yet drawn at 30 June. Broadcom is at once the chip supplier, the equipment lessor and the lender: its $42 billion covers about a third of a five-year lease of $125.2 billion, according to Yahoo Finance. Between January 2023 and June 2026, Forbes reports, Anthropic paid 19.2 billion for compute and infrastructure and took in 21.3. So far, then, the books balance.

The problem lies in the due dates. The spending commitments declared in the prospectus relate almost entirely to compute and infrastructure: part must be met by 31 December 2030, nearly all the rest between 2031 and 2045, with minimum purchases and take-or-pay clauses, meaning payable whatever the usage. The total changes with what is counted: signed or under negotiation, cancellable or not. Fortune reports compute purchases already committed of at least $475 billion, up to $750 billion counting leased chips; Forbes reaches a trillion through 2045; InvestingLive puts it above $600 billion and breaks it down: about $450 billion of cloud over seven to ten years, mostly non-cancellable (roughly $111 billion with Google, $110 billion with Amazon, $31 billion with Microsoft), up to $86 billion for Nvidia chips through Nscale and Lambda, more than $160 billion for Google chips leased through Broadcom, and up to $150 billion of deals still under negotiation. On the other side, about 80% of first-half revenue comes from pay-as-you-go usage, and customers can cut back or leave on short notice: only $10.8 billion sits in non-cancellable multi-year contracts, about three quarters of which is recognised within a year. At the end of 2025 customer spending fell by about 20% in two weeks before recovering. Anthropic itself writes that it expects revenue growth to slow and fluctuate, and that it may have periods of unused compute with cash going out and no revenue coming in.

The analysis firm Dolphin Research estimates that to sustain the payments at their peak, revenue would have to grow four to five times. That is its own estimate, not the filing’s, but it gives the order of magnitude of the bet. Those in the middle are exposed: according to Forbes data, the compute company Nscale gets 43% of its contracts from Anthropic.

This is not proof of a bubble. Revenue has grown about fourteen times in a year, and at 30 June declared liquidity exceeded $130 billion. The experts are divided: Dan Ives calls the listing a decisive moment for AI, while Roger Montgomery reads it as an attempt to get ahead of the burst. I suspect the risk is a mismatch of timing: commitments that run for twenty years, revenue that moves in weeks, and a capital market that now costs more.

Over all of this weighs an asymmetry that concerns the writer too. The alarm makes bigger news than the promise, and the promise is what actually lets us look at the machine’s development with hope: drugs, diagnoses, expertise within reach of poor countries, a less sluggish State. It is what the investor is buying. A piece about due dates, like this one, necessarily tells the side that frightens, and that is worth saying.

There is also something the prospectus does that is rarely seen in stock-market documents: it devotes a great deal of space to how its own technology could go wrong. The summary of risk factors warns that the models may behave in unintended, misaligned or autonomous ways. In a sector that spent years selling the promise, showing concern has become a requirement of respectability: whoever is not afraid seems not to have understood. I suspect this self-flagellation does two jobs at once, reassuring the public and protecting the issuer, because what has been written as a risk is hard to contest as an omission. It is a suspicion, and it explains why the alarm and the promise come out of the same document with such different weights.

Anthropic makes Claude, and that is worth keeping in mind while reading. The dream has a due date too, and the prospectus writes it down: 2030 for part of the bill, 2045 for the rest. Before calling it a bubble or a crash, the useful question is which of the two curves has the nearer deadline.