Nine Tenths of Anthropic’s $80.7B Loss Is an Accounting Line

Anthropic filed a Form S-1 with the Securities and Exchange Commission on 26 October 2026, applying to list its Class A common stock on the Nasdaq Global Select Market under the ticker ANT. An amended filing followed. Everything below is read from the filing itself, with reporting clearly marked where we relied on it.

What the filing actually reports

For the comparable half-year, revenue was $1,163 million in the first six months of 2025 against $16,286 million in the first six months of 2026 — a fourteen-fold increase. The first half of 2026 on its own is more than three and a half times the whole of 2025.

The loss figure is not what most coverage made of it

Headlines put the number at $80.7 billion. The filing shows that $72,388 million of it — just under 90% — sits on a single line: the change in fair value of convertible notes, derivative liabilities and SAFEs. The footnote attributes it primarily to an increase in the fair value of the outstanding convertible promissory notes.

That line is not money spent. It is a remeasurement. When a private company issues convertible notes and the company then becomes worth far more, accounting rules require the liability to be written up, and the write-up runs through the income statement as a loss. The more the company is worth, the larger the reported loss. It is close to the opposite of the signal a reader instinctively takes from the word loss.

None of this makes the company profitable, and we are not claiming it does. The point is narrower: the $80.7 billion figure cannot be read as cash consumed, and a comparison against any normal operating loss is a category error. Reporting of the filing put second-quarter operating income at $974 million, against a $1.7 billion operating loss in the first quarter — we have not located those two figures in the filing text ourselves, so treat them as reported rather than confirmed.

The number that genuinely changed: compute

Compute and infrastructure cost $2,522 million in 2024 against $386 million of revenue. That is 653% of revenue — six and a half dollars of compute for every dollar earned. In 2025 it was $7,326 million against $4,585 million, or 160%. In the first half of 2026 it was $8,945 million against $16,286 million, or 55%.

That inversion is the single most important trend in the document. A business that spent six times its revenue on inference two years ago now spends roughly half. Compute spending did not fall in absolute terms — the first six months of 2026 cost $8,945 million, more than the whole of 2025 at $7,326 million. Revenue simply grew faster, by a wide margin.

For anyone buying AI capacity, that ratio is also the clearest available evidence on where model serving costs are heading. It is one of the few places a frontier lab has been required to show the arithmetic.

Five share classes, and only one of them is being sold

The filing sets out an unusual capital structure. Public investors buy Class A. What Class A carries, relative to the others, is worth reading before anything else in the document.

Read those two middle rows together. Class F holds a permanent voting majority on shareholder votes, and Class T elects most of the board. A Class A holder owns economics and a vote that cannot, by construction, decide anything a Class F holder disagrees with.

Multi-class structures are ordinary in technology listings. Five classes, with voting majority and board control split across two of them that the public cannot buy, is not ordinary. Whatever one makes of it, it is the part of the filing least represented in the coverage so far.

The commitments, and what is still blank

The filing describes commitments with an aggregate committed spend whose figure is literally left blank in the version we read, noted as substantially all related to compute and infrastructure, with part expected before 31 December 2030 and substantially all of the remainder between 2031 and 2045. Reporting of the filing has put a SpaceXAI arrangement at up to $120.5 billion over terms extending into 2029, at roughly $3.1 billion a month once capacity is delivered. We could not confirm that figure in the filing text, so it is reporting.

Still blank in the document

The price range per share, the number of shares offered, the total committed spend and the use of proceeds are all unfilled. That is normal for an initial filing — those numbers arrive in a later amendment, close to pricing. Any valuation figure circulating now is therefore reporting on private conversations, not a disclosed number. Press reports have described a range between $1.5 trillion and $2 trillion under discussion, against roughly $965 billion in May.

Twenty-two underwriters are named, led by Goldman Sachs and Morgan Stanley, with J.P. Morgan, Citigroup, Barclays, Wells Fargo Securities, BofA Securities, Deutsche Bank Securities and others following.

What this changes if you build on Claude

Very little immediately, and two things over time.

A listed company files quarterly. Pricing changes, usage limits, model deprecations and enterprise contract terms start being discussed on earnings calls and in filings, which means anyone building on Claude Code or the API gets a steadier flow of information about the economics behind the product than they have had.

The second is the compute ratio. A company whose inference cost has gone from 653% of revenue to 55% now has a public obligation to keep showing that line. If it holds or improves, price pressure on API tokens is more likely to go downward than upward. If it reverses, you will see it two or three quarters before it reaches your invoice.

Frequently asked questions