Anthropic's IPO filing is a price-hike notice addressed to every enterprise customer

Every company that paid Anthropic for Claude in 2025 got a subsidised product. According to the leaked IPO prospectus reported by Reuters, Anthropic brought in nearly $4.6 billion in revenue last year and spent $7.33 billion on compute and infrastructure. Before it paid a single researcher, salesperson or landlord, it had spent more on servers than it collected from every customer combined. If you buy from Anthropic, or from anyone reselling its models inside their own product, I think you should read that filing as a letter about your next invoice.

I went in expecting to side with John Gruber, who read the prospectus at Daring Fireball as the business plan of a company that only makes sense if it is about to build a machine god. He has a point about the tone. CNBC counted around 80 of the filing's 261 pages devoted to risks, against 48 on the actual business, and the risks include models that might resist shutdown. But I read the document as a buyer, and from that seat it says something more mundane and more expensive.

Start with the headline loss, because Gruber leans on it too hard. The $42 billion net loss for 2025 includes a charge of roughly $34 billion reflecting a higher estimated value of financing that could convert into shares. That is an accounting entry, and nobody's AI budget paid for it. Strip it out and Reuters still reports an operating loss of more than $8 billion on total operating expenses of $12.65 billion. On top of that sit $518 billion of cloud, computing and infrastructure obligations in coming years. What surprised me is that the cleaned-up figures make my case stronger. A company losing that much on operations, with that much committed spend ahead, has one realistic source of improvement, and it is the people sending it money.

Those people are almost entirely you. Government contracts are under 1% of revenue, so the commercial customer base carries the whole load. Nearly a quarter of revenue came from just two customers. And among its own risk factors, Anthropic admits that plenty of its biggest accounts have signed no multi-year commitment and remain free to scale back or walk away whenever they like. I doubt any company puts that admission in a filing and then leaves the problem alone. The obvious fix is to convert month-to-month usage into multi-year committed spend, because contracted backlog is what an IPO roadshow sells to fund managers. Reuters has reported the listing is likely to slip until after the November US midterms. I expect the Anthropic sales team to spend the weeks before then chasing exactly that kind of paper.

The valuation adds pressure. Anthropic is reportedly aiming for more than $2 trillion, more than double the estimated $965 billion it was worth in May. Public shareholders at that price will want to see gross margins climb, and the cheapest lever for any software vendor is to charge existing customers more for the same thing. I have no evidence that Anthropic has raised a list price yet, and I would be surprised if it did so crudely. Tighter rate limits, new premium tiers, smaller discounts and higher minimum commitments all lift revenue without a headline.

The reasonable reply is that buyers hold the cards. Gruber argues that commodity open-source models are rapidly closing the gap, and if that is true, any vendor that squeezes customers will watch them leave. I agree the alternatives exist. But leverage belongs only to buyers who can actually switch, and most enterprise AI deployments I hear about have been welded to one model: prompts tuned for its quirks, tool-calling built around its formats, evaluation suites that only ever ran against it. A finance team that has never tested a second model has no credible threat to make at the renewal table, however good the open-weight options look on a leaderboard.

The concentration point cuts against smaller buyers too. If either of the two customers behind nearly a quarter of revenue trims its spend, the shortfall has to be recovered from everyone else, and the mid-sized accounts with the least negotiating weight are the easiest place to recover it.

If your Claude contract, or a contract with any vendor whose product runs on Claude, comes up for renewal before the listing, assume the account team will offer a fatter discount in exchange for a longer commitment. Take it only with a written cap on per-token prices for the full term and a right to exit if rate limits or tier definitions change, and spend the fortnight before the meeting running your heaviest workload on at least one rival model so the walk-away is real.

Prompted by Anthropic’s IPO Prospectus Is a Fucking Doozy, Daring Fireball.