An $80.7 Billion Loss, a $16.8 Billion Quarter: How to Read Anthropic’s IPO Filing

中文版 Chinese version

Figures below are from reporting on Anthropic’s draft S-1 by the Wall Street Journal, Reuters, and Yahoo Finance (October 21–22, 2026); the filing itself was not yet public at publication time. Anthropic has not commented. This is analysis, not investment advice.

The filing, and the race around it

Anthropic could make its IPO prospectus public as early as Friday, October 23, according to Yahoo Finance reporting citing three people familiar with the plans. The company filed confidentially with the SEC in June; a public filing starts a 15-day clock before a roadshow, which is reportedly targeted for the week of November 9, on Nasdaq, led by Morgan Stanley, Goldman Sachs, and JPMorgan.

The context matters as much as the document. Rival OpenAI has pushed its own IPO to 2027 and is instead considering another private round — a reversal from a year in which both labs were racing to list first. Anthropic is going anyway, into a market that just watched a South Korean data-center firm fall 32% on debut and an Nvidia-backed peer shelve its listing. Yahoo Finance calls the deal “a crucial test of investor appetite for the artificial intelligence boom.” That is not hype; at a discussed valuation of $1.5 trillion (down from earlier talk of $1.8–2 trillion) and a raise that could reach $100 billion, this would exceed SpaceX’s June IPO as the largest debut ever attempted.

The loss that isn’t — and the growth that is

The headline number is the loss: $80.7 billion net loss in the first half of 2026, against $13 billion a year earlier. Read one layer down and it dissolves:

• About $72 billion of it is a non-cash fair-value charge on convertible notes and similar instruments. When Anthropic’s valuation rises, those instruments are worth more, and accounting rules book the increase as a loss. No cash left the building; the “loss” is, perversely, a symptom of the valuation going up. The same pattern produced 2025’s roughly $42 billion net loss.
• The underlying operating loss is about $8.1 billion for H1 2026 — roughly the same as the operating loss for all of 2025.

Now the other side of the ledger, which is genuinely extraordinary:

• Revenue: $1.1B (H1 2025) → $16.3B (H1 2026) — about fifteen-fold in a year. Quarterly: $4.7B in Q1, $11.5B in Q2, and — per what the company told investors on October 14 — $16.8B in Q3 alone, up 45% sequentially. Q3 revenue exceeded the entire first half’s quarterly average pace by a wide margin.
• Anthropic says it has now been profitable on an adjusted basis for two consecutive quarters, with positive free cash flow, generating $1.70 of revenue per $1 of compute in Q3 — a compute margin above 40%.

So the bull case writes itself: operating losses flat while revenue grows fifteen-fold is what operating leverage is supposed to look like, and the scariest number in the filing is an accounting artifact. If that were the whole filing, the IPO would price itself.

The number that should actually price this deal

It is not the whole filing. The same draft contains the figures that explain why some investors are marking the deal down from $2 trillion toward $1.5 trillion:

• Compute spend of nearly $9 billion in H1 2026 alone — already more than the $7.3 billion spent in all of 2025.
• Future commitments of at least $518 billion contracted over the coming decade (about $450 billion of it cloud agreements, most non-cancelable), with total obligations reported above $600 billion including deals under negotiation. The capacity is rented from Google, Amazon — and, in a detail that captures 2026 perfectly, GPU capacity at SpaceX’s Colossus 1 data center.
• Customer concentration: two customers account for about 20% of revenue, and the filing expects a small number of large customers to remain a significant share.
• And the structural observation analysts keep circling: Anthropic’s compute costs have grown nearly in lockstep with its revenue. For a normal software company, scale brings margin expansion almost automatically. For a frontier lab, every new dollar of revenue arrives with roughly its proportional share of GPU attached. The Q3 figure — $1.70 back per $1 of compute — is the first hard evidence that ratio can improve. One quarter is not a trend.

My take

The debate about this IPO will be conducted as if it were about the $80.7 billion loss, because that number fits in a headline. It should be about a different question: is a frontier lab a software company with unusual costs, or an infrastructure utility with unusually good software?

If it is the former, today’s lockstep cost growth is a phase — inference gets cheaper, margins widen, and a $1.5–2 trillion valuation is a growth multiple like any other. If it is the latter, the correct comparables are not Microsoft and Google but power utilities and cloud landlords: businesses valued on returns on enormous, continuous, non-cancelable capital commitments — the $518 billion kind. Utilities do not trade at software multiples, and they are not supposed to.

Anthropic’s own filing argues both sides at once, which is why Friday’s public S-1 is worth reading rather than summarizing: fifteen-fold growth and an improving compute ratio on one page; half a trillion in locked-in spending and a two-customer fifth of revenue on the next. OpenAI looked at the same market and chose to stay private another year. Anthropic is betting the window — post-SpaceX, pre-whatever-comes-next — is worth more than the wait. By the week of November 9, the roadshow will tell us which reading the market bought.

Sources: Yahoo Finance (Oct 21, 2026) on filing timing, roadshow, and syndicate; Wall Street Journal draft-S-1 reporting via Reuters, Quartz, PYMNTS, and TradingView (Oct 21–22, 2026) for H1 revenue, the $80.7B loss and its $72B non-cash component, compute spend, and commitments; Reuters reporting of the October 14 investor day for Q3 revenue and adjusted profitability; Motley Fool (Oct 22, 2026) for the OpenAI IPO delay and Q3 compute-margin figures.