Anthropic's IPO prospectus: a 42 billion dollar loss on 4.6 billion dollars of revenue

On 28 September 2026, Reuters and the Financial Times reported from Anthropic's IPO prospectus, which both newsrooms had obtained. Anthropic has been a private company whose numbers circulated only as leaks and projections. The prospectus is the first accounting under public-market rules, and it reads differently from the success stories of recent weeks, which had the company profitable for at least one quarter this year. The three numbers that matter: a net loss of 42 billion US dollars in 2025, revenue of 4.6 billion in the same year, and planned spending of 518 billion for cloud capacity, compute and infrastructure over the coming years.

Anthropic Claude mark — The IPO opens the books behind this mark for the first time. The brand graphic comes from the Commons collection, every figure in this post from the prospectus as reported by Reuters, the Financial Times and heise, state of 28 and 29 September 2026. (Quelle: Anthropic / Wikimedia)

What the numbers say

The operating loss more than doubled between 2024 and 2025, from just under 3 billion to over 8 billion US dollars (Reuters, 28.09.2026). The 42 billion net loss sits above that figure; none of the published summaries itemises what drives it. The ratio still carries the message: the net loss is close to nine times annual revenue. The planned 518 billion in spending forms the second half of the story, because those are commitments to data centres that today's revenue cannot come close to servicing. The IPO is meant to cover it: around 100 billion in proceeds at a valuation of roughly two trillion, with the listing now scheduled after the US midterm elections in November (heise, 29.09.2026). For proportion: SpaceX's record listing raised 86 billion this summer, Aramco managed 29 billion in 2019.

Neither a vendor figure nor our own measurement: numbers from the prospectus as reported by Reuters and heise. The planned 518 bn in spending would blow this scale, so it stays in the text. (Quelle: heise / Reuters)

Two customers, both owned by competitors

Nearly a quarter of the 2025 revenue came from just two customers (Reuters, 28.09.2026). The prospectus keeps them anonymous; the reporting names them, a credit that goes back to VentureBeat: the coding tool Cursor and GitHub Copilot. Copilot belongs to Microsoft, which is at the same time the largest customer and the largest investor of competitor OpenAI. Cursor repositioned itself spectacularly over the summer: SpaceX signed the merger agreement with maker Anysphere on 16 June 2026, paying 60 billion US dollars in its own shares, and the deal closed in the third quarter of 2026 (SEC filing, heise). SpaceX had already folded xAI into its corporate structure in early 2026 and runs the Colossus supercomputer, where meanwhile even Google and Anthropic rent compute.

A quarter of the revenue therefore hangs on two companies that build models of their own. Reuters supplies a second line that disappears under the shock of the loss figure: many of the largest customers hold no long-term contracts and may cut or stop their spending. For a provider that has to service multi-year compute commitments out of API revenue, that is the most fragile sentence in the entire document.

The risk chapter

According to the Financial Times, nearly a third of the prospectus is devoted to risk factors. Named there: the potential of the technology for manipulation and extortion, unpredictable behaviour, the possibility that its own AI carries existential risks for humanity, and the extreme concentration of the user base. At that breadth it is a disclosure duty to the US securities regulator, since a prospectus that conceals risks gets more expensive in court than one that spends thirty pages on them. The striking part remains the combination inside one document: the company warns in black and white about the risk of its own technology, and funds a quarter of itself through two competitors who build models too.

What this means for companies buying AI

The numbers change nothing about how useful the models are. They change how you grade the contract side. Four conclusions, and we run the first two at Senn ourselves:

Model access runs through a gateway with placeholders; our software speaks alias names, no vendor SDKs. A provider swap should cost one configuration line, never a project restart. How to set that up is written up in Anchoring AI in the enterprise.
Tasks without frontier requirements, meaning classification, retrieval-augmented generation, extraction, standard texts, run on open weights, and on own hardware small models are good enough today. The post Local AI gets smaller keeps the dimensions on record.
Today's API prices are displacement prices, cross-subsidised from venture capital. A multi-year contract at fixed prices is a bet against your supplier. A contract with a price-adjustment clause and an interchangeability clause is no bet at all.
Legally, the dependence on US corporations is set in stone by this prospectus: Anthropic raises capital through the SEC, the compute contracts sit under US law, the listing goes to a US exchange. The energy demand behind that planned infrastructure, which critics of the AI arms race point to, belongs to the same bill.

That the technology works was never the open question. What the prospectus leaves open is which price is the real one tomorrow, and whether the contract you sign today has an answer.