Live
Abstract illustration: rising revenue blocks on thin pillars beneath an oversized slab
Business & Startups

Anthropic’s Leaked S-1: Skip the $42B Loss. Watch the $518B Bill

Two numbers from Anthropic’s draft IPO prospectus look like they cannot both be true. Fortune’s snapshot of the income statement lists a 2025 net loss of US$42 billion. The Decoder reports an operating loss of $8.06 billion for the same year. A $34 billion gap is not a rounding error, so it is worth settling before anything else.

Both are right. They measure different things. According to Reuters, which saw the prospectus, about $34 billion of the net loss is an accounting charge reflecting the higher estimated value of Anthropic’s convertible financing, the instruments that turn into shares later. When a company’s valuation climbs as fast as Anthropic’s has, what it owes holders of those instruments climbs on paper too, and that increase runs through the income statement. No cash leaves the building. The operating loss, which widened from $2.98 billion in 2024 to $8.06 billion, is the number that describes the business.

That matters because the $42 billion figure will travel. It is the scariest number in the document and the least informative one.

What this document is, and isn’t

This S-1 is not on EDGAR. Anthropic confidentially filed a draft registration with the SEC in June. The version now in circulation was, per the Financial Times as cited by The Decoder, sent to a small group of partners in the days before Sept. 29, and Reuters reviewed it. Fortune’s headline calls it “leaked.” Anthropic declined to comment to Reuters on the timeline or the findings.

So every figure here is secondhand, a draft, and subject to change before a public filing. Drafts get revised. Risk factors get lawyered. Treat this as a preview, not the record.

The revenue curve

Anthropic booked nearly $4.6 billion in revenue in 2025, up twelvefold, per Reuters. Then the curve went vertical. First-quarter 2026 revenue was $4.73 billion, more or less a full 2025 in a single quarter, and the second quarter came in at $11.5 billion, a figure first reported in August that Fortune now includes in its snapshot of the filing. Fortune says the company will be profitable on an operating basis for a second straight quarter. The Decoder adds the qualifier the bulls will skip: only on an adjusted basis.

The company is reportedly aiming for a valuation above $2 trillion, more than double the $965 billion it was valued at in May, with the listing expected in November, after the U.S. midterms.

Two customers, $518 billion

Here is our read on what actually deserves the attention. It sits in two lines that won’t make anyone’s headline.

First, concentration. Two customers made up nearly a quarter of 2025 revenue, and Reuters reports the prospectus warns that many of Anthropic’s largest clients are not locked into long-term contracts and could cut spending. The customers are unnamed. On 2025’s $4.6 billion, a quarter is roughly $1.15 billion riding on two relationships that can, by Anthropic’s own account, walk.

The 2026 numbers may dilute that. A quarter of $4.6 billion is small next to an $11.5 billion quarter. But we don’t know the 2026 concentration figure from the reporting, and a customer base that grew that fast tends to be lumpy in ways a prospectus describes only after the fact.

Second, commitments. The prospectus discloses $518 billion in future cloud, computing and infrastructure obligations, against $20.28 billion in cash at the end of 2025. Compute and infrastructure spending in 2025 alone tripled to $7.33 billion. Annualize the second quarter and Anthropic is running at about $46 billion a year in revenue, our arithmetic. The obligations are more than eleven times that.

None of the coverage we read gives a schedule for when the $518 billion comes due, and that schedule is the single most important number not yet public. Spread over many years against a revenue line growing at this pace, it is a bet on demand. Front-loaded, it is a liability that has to be refinanced by investors who are buying the stock in November. For a company where two customers can move a quarter of revenue, the gap between those two readings is the whole investment case.

The risk language

Then there is the section everyone is quoting. The Decoder reports that risk factors take up nearly a third of the lengthy document, and that Anthropic warns increasingly advanced models “could manipulate, blackmail, or otherwise behave in unpredictable ways,” and of technology posing “existential risks to humanity.” Reuters adds that Anthropic’s own research found models in controlled tests sabotaging code, assisting fraud and manipulating information.

It reads as alarming. Our read: it is also a company restating its own research in the language securities lawyers prefer, where you disclose everything so that nothing later counts as a surprise. The behaviours Reuters lists come from Anthropic’s own tests. Moving them into a prospectus changes who can sue over them, not what is known.

When the public version does land on EDGAR, three things will tell investors more than any risk factor. The payment schedule behind the $518 billion. A 2026 customer-concentration figure to set against 2025’s. And the bridge from the adjusted operating profit reported for 2026 to the GAAP operating line, which for 2025 showed an $8.06 billion loss. The public filing is where the footnotes have to be complete, and where a draft’s numbers can still move.

The sentence that should keep an underwriter up at night is less dramatic. It is the one about customers who aren’t locked in, sitting a few pages from a $518 billion bill.

Sources

// Business Editor
Mira Okonkwo

Mira Okonkwo covers the business of technology for prompt/power: venture capital, startups, IPOs and earnings. She treats a valuation as a mood rather than a number until the S-1 says otherwise.

Latest from prompt/power

  1. Gemini’s Free Tier Shrinks Oct. 9: What You Keep and What Costs ExtraOct 5
  2. How to Read an AI Company’s S-1: The 7 Numbers That MatterOct 5
  3. OpenAI’s Safety Lead Quit Over Culture. California’s AG Was Already InOct 5
  4. When an AI Agent Breaks In, Who Answers for It?Oct 5
  5. The New AI Models Don’t Talk. They Decide.Oct 5

Leave a Reply

Your email address will not be published. Required fields are marked *