Anthropic's IPO Numbers: What They Mean for Claude Buyers

On September 28, 2026, Reuters reported the financial figures in Anthropic’s IPO prospectus, a document it had seen but which is not public. Revenue grew 12-fold in 2025 to nearly $4.6 billion. The company spent $7.33 billion on compute and infrastructure and faces $518 billion in cloud and infrastructure obligations in the years ahead. For a business buying Claude, the useful question is what those numbers say about price, capacity and contract terms.

Every company figure on this page is as reported by Reuters from the prospectus; Anthropic declined to comment, and the filing is not on the SEC’s public EDGAR system as of October 3. Our June explainer on Anthropic’s IPO filing covers the listing itself; this page covers only the new figures.

The Reuters report gives four figures that matter to buyers. The net loss it also reports, about $42 billion, is mostly a roughly $34 billion accounting charge for the rising value of financing that could turn into shares, rather than money spent running the business, so it says little about Claude’s economics.

Two more details round out the picture. The operating loss widened to $8.06 billion from $2.98 billion in 2024, and the company held $20.28 billion in cash, equivalents and short-term investments at the end of 2025. Reuters, citing sources, has also reported that the listing is likely to come after the November US midterm elections.

Figures circulating since the report go further than Reuters does. Some sites give a precise revenue figure to two decimal places, a split of the $518 billion between named cloud providers, and a share of the obligations that cannot be cancelled. None of those appear in the Reuters report we read, and the prospectus is not public, so we leave them out until the filing can be checked.

The rest of this section is our reading of the reported figures, not a forecast of Anthropic’s prices.

The first row has a recent example. When Anthropic launched Claude Fable 5.1 on September 1, input and output prices stayed at $10 and $50 per million tokens and only the cache-read price fell, by 75%, which lowers bills where cached context dominates. The contract side connects to our note on applying AI spend commitments through cloud marketplaces, which covers the terms worth checking in any commitment.

A company can submit its registration statement to the SEC for confidential review, which is why this one is not public and Reuters reported from a copy it had seen. The SEC’s procedure requires it to publish the registration statement and the earlier confidential drafts at least 15 days before any roadshow. If Reuters’ sources are right about a listing after the November 3 midterms, the public document must appear at least 15 days before that roadshow. No roadshow date has been reported, so it could appear at any time before then.

That filing is the one to read. Look at the risk factors on capacity and customer contracts, any discussion of pricing, and how the company describes its largest customer agreements.

None of this needs a change of model or vendor. The figures describe a company spending heavily to add capacity and keen to lock in customers, which gives buyers with predictable usage some leverage. The resilience side is unchanged from our Claude outage playbook, and our AI transformation work includes model-cost reviews and fallback design.

The routing above tells a team facing a board or finance question to cite only reported figures and revisit after the public filing. That is easy to say and awkward to do when the question is “is our main AI vendor financially sound?” and the honest answer is “here is what one outlet says a document we cannot see contains.” The brief has to carry that honesty without sounding evasive.

A workable structure has three short parts. First, the source and its status in one sentence: Reuters reported these figures on September 28 from a prospectus it had seen; Anthropic declined to comment; the filing is not public. Second, the figures themselves, each with Reuters’ own qualifier attached, so “nearly $4.6 billion” stays “nearly” and the $42 billion net loss is presented beside the roughly $34 billion accounting charge that makes up most of it. Third, what your team is doing about it: the commitment conversation, the tested fallback, and the date the public filing is expected to be readable.

What the brief should not contain is any figure that is not in the Reuters report, however widely it is circulating, and any forecast of Anthropic’s prices. The table in the buyers section is explicitly our reading; a board brief should say so if it borrows from it.

Date the brief and plan to replace it. Everything in it is true as of the Reuters report and will be superseded by the public filing, which under SEC procedure must appear at least 15 days before any roadshow. A brief that says “valid until the public filing; to be rewritten from the filing within a week of its appearance” has told the board exactly how much weight to put on it.

Each illustrative misuse below takes a figure Reuters reported and turns it into one it did not. They are the reasons the “what is not confirmed” section exists.

The precise revenue: “nearly $4.6 billion” becomes a figure to two decimal places in a slide, borrowed from a site whose figure does not appear in the Reuters report. The allocated obligations: the $518 billion is split between named cloud providers with percentages, when the Reuters report gives neither the split nor the share that cannot be cancelled. The cash-burn reading: the roughly $42 billion net loss is presented as money spent running the business, when Reuters attributes most of it to an accounting charge for financing that could turn into shares. The priority assumption: the fact that two customers are nearly a quarter of revenue is read as meaning smaller buyers will be served first if capacity tightens, when the buyers table draws the opposite inference.

All four are avoided by the same rule the brief follows: quote the outlet, keep its qualifiers, and stop where its report stops. The guide on reading an AI valuation leak, linked above, is the longer version of that rule.