Anthropic’s Reported Prospectus Sets $518 Billion of Commitments Against $4.59 Billion of Revenue. The $42 Billion Loss Is Mostly a Non-Cash Charge.

On 28 and 29 September 2026, Reuters reported figures from Anthropic's confidential draft IPO filing: 2025 revenue of nearly $4.6 billion, an operating loss of $8.06 billion, a net loss of $42 billion, and at least $518 billion of compute and infrastructure commitments. Two different losses and one very large commitment number invite the wrong conclusions. This piece explains what each figure measures. It does not revisit Anthropic's valuation, which we covered in our earlier piece on the pre-IPO multiple (anthropic-ipo-2-trillion-target-multiple-pre-ipo-market-disagreement-2026).

Two losses, one company

The headline net loss of about $42 billion and the operating loss of $8.06 billion measure different things. The operating loss is the business burning money: revenue of roughly $4.59 billion minus operating expenses of $12.65 billion gives the $8.06 billion figure (our arithmetic, which matches the reported number). It widened from $2.98 billion in 2024. Compute and infrastructure costs were $7.33 billion, about three times the 2024 figure, according to Reuters. Revenue grew about twelvefold, which is why the operating loss, though large, grew by a smaller multiple, about 2.7 times.

The rest of the gap to $42 billion is an accounting charge. Reuters reports that $34 billion of the net loss reflects a higher estimated value of financing that could later convert into Anthropic shares. It is not cash spent running the business. The mechanism is the same one that makes a company's reported loss jump when the value of a liability tied to its own share price rises: as the estimated value of the convertible financing goes up, the company books the increase as a loss. By our arithmetic, $34 billion is about 81% of $42 billion, and the remaining $8 billion is almost exactly the operating loss. That is a useful check on the reading, though it is our observation, not a reconciliation the filing is reported to provide. Cash and short-term investments stood at $20.28 billion at the end of 2025, per Reuters.

The commitment stack

The $518 billion figure is described as at least that much in cloud, compute and infrastructure commitments over roughly a decade with six partners. Reuters' breakdown: about $111.1 billion to Google, $110 billion to Amazon and $31.4 billion to Microsoft, on terms of seven to ten years, plus about $161.2 billion in Broadcom equipment leases. Separately, agreements with xAI cover up to $84.5 billion of Nvidia-based capacity through 2029, most of which Reuters says can be cancelled on 90 days' notice. About 80% of the obligations are reported as non-cancellable or payable regardless of usage; the filing is quoted as saying that if actual spend falls short, Anthropic must pay Google the difference.

Summing the named items gives $413.7 billion for Google, Amazon, Microsoft and Broadcom, and $498.2 billion with the xAI figure included. That leaves roughly $20 billion of the $518 billion that the reporting we read does not itemise. Eighty percent of $518 billion is $414.4 billion, which is close to the four-name total. We would not read that as a match, since the reporting does not say the non-cancellable share and the four named items are the same set. It is a reminder that the breakdown is partial until the document is public.

The $113-per-dollar figure, tested

Forkast, relaying Reuters, calculates about $113 committed for every $1 of 2025 revenue. The arithmetic is right: 518 divided by 4.59 is about 113. But it compares a roughly ten-year stack of obligations with a single year of revenue from a company growing at a multiple of itself. That makes it a striking number and a poor measure of burden. Two alternative framings, both ours, change the picture. CNBC reported in August that second-quarter 2026 revenue reached $11.5 billion; annualised, that is about $46 billion, and $518 billion is then about 11 times annualised revenue. And if the commitments were spread evenly across ten years, which the sources do not say they are, the average would be about $52 billion a year, roughly the size of that annualised run rate.

The Q2 2026 figure comes from CNBC, not from the prospectus reporting, and we have not seen it confirmed in the filing. CNBC also reported that Anthropic expects a second straight quarter of adjusted operating profit, which is a different measure from the GAAP operating loss above. None of this makes the commitments small. A fixed stack that is 80% non-cancellable is a real constraint if revenue falters, and the filing warns that many large customers have no long-term contracts and could cut or stop spending. It does mean the single ratio overstates how out-of-scale the commitments are against current revenue and understates how sensitive the picture is to growth continuing.

The revenue side

Reuters reports that two customers, not named, accounted for nearly a quarter of 2025 revenue. The reporting does not support linking the concentration to any particular company. The filing is also reported to devote roughly 80 of its 261 pages to risk factors, including a warning that advanced AI could pose catastrophic or existential risks to humanity and a statement that demand for AI is limited principally by the availability of compute. That second statement is the company's explanation for the commitments: it is buying capacity ahead of demand because capacity is the constraint. It is also the claim that the commitments are meant to be tested against.

What the numbers cannot settle

• The document is not public. The figures are secondhand from Reuters' reading of a confidential draft, which may change before any public filing.
• Conflicting headline losses are not a conflict. The $8 billion and $42 billion figures measure the operating business and the total including a non-cash charge. Coverage that quotes one without the other is incomplete.
• The commitment schedule is unknown. Without year-by-year payments, no ratio of commitments to revenue can describe the burden accurately.
• 2025 is already stale. The year-end figures are for a company whose later-quarter revenue is reported to be several times higher.
• Valuation is out of scope here. Reuters reports a target above $2 trillion and a listing likely after the US midterm elections in November; neither is confirmed by the company.

The practical takeaway is to read each figure for what it measures. The operating loss shows the cost of running a business growing at an extraordinary rate. The net loss is dominated by an accounting item that depends on how the company's own financing is valued. And the commitment stack is a large, mostly fixed obligation whose weight depends on a schedule and a growth rate that the reporting has not yet shown. When the filing becomes public, those three details are where to look first.