Anthropic's $42 Billion Loss: $34 Billion Is a Non-Cash Charge. Read the Rest

Anthropic's draft S-1, as reported by Reuters, shows a $42 billion net loss on $4.59 billion of revenue for 2025. About $34 billion of that loss is an accounting charge on convertible financing, not cash spent. The operating loss was $8.06 billion. The Anthropic S-1 figures that matter sit below that line: compute at 58% of operating expenses, and 2 customers at 24% of revenue.

What the leaked Anthropic prospectus reports, and what it is not

On 28 September 2026 Reuters published figures from a draft IPO prospectus it had seen. The Financial Times, Fortune and CNBC followed within hours. Anthropic declined to comment and has not published the document.

That sourcing shapes everything below. The company submitted its draft registration statement confidentially in June 2026. A confidential submission carries no disclosure obligation, so nothing is on EDGAR and nothing has been audited in public. What you are reading in every outlet, including this one, is a report of a draft.

The draft runs to 261 pages, with around 80 of them given to risk factors, per Capital Brief's summary of the Reuters report. The listing is expected after the US midterm elections in November, at a valuation above $2 trillion. Neither the date nor the valuation is confirmed.

The operating line is the one to read. Three numbers are being quoted as if they were interchangeable. Revenue of $4.59 billion. An operating loss of $8.06 billion. A net loss of $41.97 billion. The first two describe the business. The third describes what happened to the value of the company's financing during the year.

My position is that the $42 billion figure tells you almost nothing about whether Anthropic can run profitably. The $8 billion figure tells you a great deal, and the compute line underneath it tells you the most. The revenue ramp, the enterprise mix and the gross-basis accounting question were covered before the leak in the earlier piece on Anthropic's enterprise-heavy revenue, and this post does not repeat them.

From $42 billion to $8 billion: the Anthropic S-1 loss reconciliation

Start with the two reported endpoints and work out what sits between them. Reuters gives a net loss of nearly $42 billion and an operating loss of more than $8 billion. Capital Brief, working from the same Reuters report, gives the precise figures as $41.97 billion and $8.06 billion.

The difference is $33.91 billion. Reuters describes the non-operating item as a roughly $34 billion accounting charge that reflected an increase in the estimated value of financing that could eventually turn into Anthropic shares. The residual and the reported charge agree to within rounding, which is the first check worth running on any leaked income statement.

How a convertible becomes a loss on the income statement

Anthropic raised much of its capital through instruments that convert into equity later, at terms set in advance. Under US accounting rules, an instrument of that kind can be carried as a liability and remeasured at fair value every reporting period. When the company's valuation rises, the claim those instruments hold on future shares is worth more, and the increase is booked as an expense.

The mechanics produce a result that reads backwards. The better the year, the larger the loss. Anthropic's estimated valuation reached $965 billion in May 2026, and the 2025 charge is the accounting shadow of that climb. No cash left the company because of it, which is why Reuters was careful to separate it from money spent running the business.

The same instruments on Amazon's books

You do not have to take the mechanism on trust, because the other side of the trade is already in a public filing. Amazon's 10-Q for the quarter ended 31 March 2026 states that it invested $8.0 billion in Anthropic convertible notes between Q3 2023 and Q4 2025.

Those notes, net of portions already converted to preferred stock, carried an estimated fair value of $45.8 billion at 31 December 2025, with an unrealised gain of $39.5 billion. Amazon's gain and Anthropic's charge are two views of one set of instruments. They are not directly comparable, because Amazon's gain is cumulative since 2023 and covers only Amazon's notes, while Anthropic's $34 billion is one year's charge across all its convertible financing. The direction and the order of magnitude match.

That is the strongest corroboration available for the leak. A leaked draft can be wrong. A noteholder's audited balance sheet, showing a $39.5 billion gain on $8 billion of cost, cannot be reconciled with anything other than a very large remeasurement on Anthropic's side.

What happens to the charge at the IPO

Notes that convert into shares stop being remeasured. The charge for those instruments ends at conversion, and the accumulated amount stays in the historical accounts as a loss that never represented cash. Until then, each uplift in the company's estimated value produces another charge, so the 2026 figure is likely to be large as well.

I would expect the public S-1 to present an adjusted operating result alongside the GAAP figure, and I would read both. The GAAP loss is real in the sense that it is the price of the capital structure. It is not real in the sense that matters to a buyer deciding whether the vendor survives.

Sources: Reuters, 28 September 2026, and Capital Brief's summary of the same report. The 2024 compute figure is implied by the reported 190% growth rate and is Zan Digital's arithmetic, not a disclosed line. Blank cells are not reported in any coverage this post could verify.

Anthropic compute costs: $7.33 billion against $4.59 billion of revenue

Compute and infrastructure cost Anthropic $7.33 billion in 2025, up 190% on 2024 and about 58% of its $12.65 billion of operating expenses. Everything else the company does, including research, sales, marketing, partnership fees and administration, fits inside the remaining $5.32 billion.

Set that against revenue. For every $1 of 2025 revenue, Anthropic spent about $1.60 on compute. Add the other $5.32 billion and total operating spend was $2.76 for every $1 earned.

What the ratio says about scale

Here is the number the pessimists skip. Revenue grew about 12 times in 2025. The operating loss grew 2.7 times. Operating loss as a share of revenue fell from roughly 770% in 2024 to about 176% in 2025.

That is a business whose unit economics improved sharply even as its absolute losses tripled. It is consistent with the FT's report, relayed by Fortune on 29 September 2026, that Q2 2026 revenue reached $11.5 billion, more than all of 2025. The same report says the company expects to be profitable on an operating basis for a second consecutive quarter. Those 2026 figures come from reporting, not from the draft income statement, and this post treats them as directional.

Compute is not cost of revenue, and that matters. The draft, as reported, presents compute as an operating expense line rather than splitting it between cost of revenue and research. That makes gross margin unreadable from the leak. Training compute and inference compute have very different economics, and a single line hides which one is growing faster. The analysis of inference costs and AI margins explains why that split decides whether margins improve with volume, and the AI COGS guide shows how to read past a vendor's reported line to its real gross margin.

I disagree with the reading that a 58% compute share proves the model is unprofitable at the unit level. It proves nothing of the kind, because training spend is front-loaded and sits in the same line. The public S-1 will have to split it. Until then, the honest statement is that nobody outside the company knows the inference margin.

The $518 billion of commitments behind the compute line

The $7.33 billion is last year's compute bill. The prospectus also discloses what the company has already agreed to spend. Reuters reported, in a second story on the filing, that Anthropic plans at least $518 billion of infrastructure spending over a decade with 6 partners. About 80% of that sum is non-cancellable or payable regardless of usage.

Source: Reuters (Echo Wang and Sabrina Valle), via Zawya, September 2026. The named commitments do not sum to $518 billion and Reuters does not reconcile them. The sixth partner is not identified in the coverage this post could reach.

This belongs in a loss reconciliation because it is the cost structure the operating line will carry for a decade. A company with $20.28 billion of cash and $8 billion of annual operating losses has roughly 2 and a half years of runway at 2025 rates, before any growth. The commitments above are measured against a revenue line that has to keep compounding to absorb them.

The circularity is the part a buyer should notice. Amazon and Google are investors, compute suppliers, distributors and competitors at once, an arrangement mapped in the breakdown of how the AI circular deals work. The prospectus itself warns, per Reuters, that the partners' incentives may not be fully aligned with Anthropic's.

Customer concentration: 2 customers, 24% of revenue, and half the sales through 2 storefronts

The concentration disclosure is the one most likely to move a buyer's negotiating position. It comes in 2 layers.

12% each, neither named

Reuters reported that 2 unnamed customers each generated 12% of revenue in 2025. On $4.59 billion of revenue that is roughly $550 million apiece, about $1.1 billion between them. The risk factors add that many of the largest customers are not on long-term contracts and could reduce or stop spending.

That pairing is the real disclosure. Concentration on its own is normal for a 3-year-old enterprise vendor. Concentration without contract term is a different thing, because the revenue can leave inside a quarter while the compute commitments above stay for a decade.

Half the revenue arrives through 2 storefronts

Separately from the 2 largest customers, 47% of 2025 sales, about $2.16 billion, were routed through the Amazon and Google cloud marketplaces. That share was 11% in 2023 and 32% in 2024. Reuters estimates the platforms collected roughly $351 million in distribution fees, around 16 cents per dollar of marketplace sales.

The receivables tell the same story. The 2 partners were responsible for collecting 60% of the $909 million of customer bills outstanding at the end of 2025, up from 42% a year earlier. Anthropic contracts directly with those customers, so it recognises the full sale as revenue, yet the cash arrives through a counterparty that is also a competitor. The filing warns that disputes or delays in that pipeline could hurt cash flow.

One more mix figure is worth holding. Reuters reports that about $3.8 billion of 2025 revenue was consumption-based and $789 million was subscription. The business is 83% metered, which is why a customer with no long-term contract can shrink so quickly.

What the Anthropic prospectus means if you buy from the company

You are not pricing an IPO. You are deciding whether to deepen a vendor relationship with a company whose public accounts you are about to see for the first time. Three things follow from the figures above.

Find out which counterparty you actually have

If you buy through AWS or Google Cloud, you are inside the 47%. Your terms, your data processing agreement and your support path all run through the marketplace, and the marketplace collects your payment. If you buy direct, you are one of the accounts the risk factors describe as able to leave at any time. That is a negotiating asset, and it is worth more before the public filing than after it.

Price the compute pass-through. A vendor spending $1.60 on compute per $1 of revenue has one obvious lever, which is price. Expect tighter usage limits, more metering and fewer unlimited allowances as the listing approaches. The contract clauses that protect you from a mid-term pricing change are listed in the CFO's guide to AI contract clauses.

Read the public S-1 in this order.

Where this reading is weakest

A post that reconciles a leaked income statement owes the reader a list of the ways it could be wrong.

Every figure is second-hand. No outlet has published the document. Reuters saw it; everyone else, including Fortune, Capital Brief and this post, is working from Reuters. 6 restatements are 1 source. If the Reuters transcription of a single line is off, every analysis built on it is off in the same place. The public filing may also differ from the June draft.

The reconciliation assumes one non-operating item. The $33.91 billion residual between operating and net loss is attributed here entirely to the financing charge, because that is how Reuters describes it. Interest income on $20 billion of cash, tax, and other items will also sit in that gap. Reuters said "roughly $34 billion", and the match is close enough to be reassuring, not close enough to be proof.

The 2026 profitability claim is the weakest link. The operating profit for 2 consecutive quarters comes from the FT by way of Fortune, and the "adjusted" qualifier appears in some accounts and not others. An adjusted operating profit can exclude stock compensation and the remeasurement charge. Until the public S-1 shows the quarterly table, I would not repeat that claim without the qualifier.

The case that the $42 billion number is the right one to quote

There is a defensible counter-view. The charge is the price of the capital structure, and the capital structure is a choice. A company that raised $65 billion in May 2026 on convertible terms knew the accounting consequence.

One analyst, New Constructs' Kyle Guske II, deducts the full $518 billion of commitments in his reverse discounted cash flow. His low case is $144 billion of equity value, 93% below the reported target. I think his treatment of the commitments is too harsh, since most are payable over a decade against revenue that doubled in a quarter. His insistence on reading the whole capital structure, not just the operating line, is correct.

Frequently asked questions

Where to start this week

Pull your Anthropic invoices for the last 2 quarters and write 1 line at the top: direct or marketplace. That line decides which half of this post applies to you, and most teams cannot answer it from memory.

Then set an EDGAR alert for a public Form S-1 from Anthropic, PBC. When it lands, read the fair value note and the cost of revenue split before the growth chart. Those 2 pages will settle more of this debate than anything published in the 10 days since the leak.

Related reading