Anthropic's IPO filing puts a price on the AI arms race

Anthropic, the maker of the Claude chatbot, has put a figure on its appetite for computing power. It is a very large one.

In an amended prospectus filed with the American regulator on Monday, the company disclosed purchase commitments and lease agreements with an aggregate committed spend of about $895.4bn, substantially all of it for compute and infrastructure. "Compute" is simply the processing capacity, chiefly in specialised chips and the data centres that house them, needed to train and run AI models. About $447.5bn of the total is expected to fall due before the end of 2030, with the remainder running between 2031 and 2045, and roughly $69.8bn is due within a year of 30 September. The details are in the amended S-1 on the SEC's website.

A bill that dwarfs the balance sheet

Put in context, the number is startling. Yahoo Finance's analysis notes that the full bill is almost seven times the roughly $132bn of cash and short-term investments Anthropic held at the end of September. The same report records that the company's first operating profit lasted a single quarter. Third-quarter revenue was $16.96bn, yet compute and infrastructure expense reached $9.9bn, including $5.0bn bought under short-term commitments that cost materially more than multiyear deals. In plain terms, demand for Claude ran ahead of the capacity Anthropic had booked, and it paid a premium to make up the gap.

The flexibility, or lack of it, is the crux. According to gHacks' reading of the filing, only about $213bn of the $895bn in contracts would let Anthropic avoid payments, and deals with Microsoft, Google, Amazon and Broadcom, totalling $437bn, are non-cancellable, with some requiring payment whether or not the capacity is used. The company's own risk language is candid on the mismatch: its revenue is primarily consumption-based and largely not locked in by long-term contracts, while its infrastructure commitments stretch over many years.

The filing also shows how entangled the industry has become. Yahoo's chart of the largest commitments lists Broadcom equipment leases of $161.2bn, Google at $121.3bn, SpaceX's AI unit at up to $120.5bn and Amazon at $110bn. One estimate puts the real exposure higher still: gHacks reports that further estimated payments of $150bn to $175bn to secure adequate compute would take the total above $1 trillion. A bet this size is a bet that demand will keep compounding.

Is the demand there?

There is evidence it is, at least among consumers. A new study of anonymised UK bank and card data, reported by The Next Web, found that Claude took 29.4% of UK consumer spending on AI subscriptions in July, up from 7.4% in December 2025, while ChatGPT's share fell from 73.5% to 52.5%. Separately, reporting on the same research says UK adults now spend over £1bn a year on paid AI plans, and 5.1% subscribe. Britons, in other words, are paying real money.

The physical side is less certain. Early indications from across the Atlantic are that the politics of data centres are hardening: one news aggregator's roundup records that Ohio voters passed AI data-centre bans or restrictions in at least 14 municipalities. If building capacity becomes slower and dearer, firms that have pre-bought it hold an advantage, but those that have over-committed carry the risk.

What it means for UK businesses

For a British manager, the lesson is not about Anthropic's share price. It is that the companies supplying your AI tools are carrying fixed costs on a scale that must be recovered from customers. That is our judgement rather than anything stated in the filing, but it points one way: expect pricing, usage limits and plan structures to keep shifting, and expect suppliers to push towards larger, stickier contracts.

The sensible response is unglamorous. Avoid building a critical process around a single model provider, and keep the work portable between them. Prefer contract terms that can be revisited, and track what AI spending is actually delivering, because the suppliers will certainly be tracking it. The cost of intelligence has fallen sharply in recent years, but this filing is a reminder that someone has to pay for the machinery underneath. Over the next few years, that will increasingly be you.