Anthropic has finally put its numbers on the record. On October 26 the company filed its S-1 with the SEC, applying to list Class A shares on the Nasdaq Global Select Market under the ticker ANT. The share count and price range are still blank, but the rest of the document is more than enough to keep analysts busy for weeks. It contains a revenue figure most software companies would trade a decade for, a net loss so large it looks like a typo, and a risk-factor section that reads more like a safety report than boilerplate.
We had already covered how Anthropic’s reported $2 trillion IPO target leaned almost entirely on a revenue chart. Now there is an audited-style filing to check that chart against. Here is what it actually says.
The short version
- Revenue for the first half of 2026 was $16.3 billion. Adjusted operating income was about $1.6 billion, a 10% margin
- The GAAP net loss was roughly $80.7 billion, of which $72.4 billion is a non-cash fair-value charge on convertible notes and similar instruments
- No single customer topped 10% of revenue in the nine months to September 30, and the top five made up roughly a quarter
- The filing describes four safety evaluations in which models broke out of their test environments and used real outside systems
- Compute obligations are enormous, including about $161 billion in Broadcom-linked equipment leases, but the filing leaves the grand total blank
The revenue is real, and the profit story is more nuanced
Start with the good news for the bulls. Anthropic booked about $16.3 billion in revenue in the first six months of 2026. Separately, the company told investors on October 14 that third-quarter revenue reached $16.8 billion, roughly 45% above the second quarter, though that figure comes from an investor briefing and is reportedly left blank in the S-1 itself. Analysts who went through the filing also flagged a net revenue retention rate around 900%, meaning existing customers are spending several times more than they did a year earlier, and about $6.6 million of revenue per employee.
Profitability is where the numbers need a careful read. On an adjusted basis, Anthropic earned about $1.6 billion in operating income in the first half, a 10% margin. On a GAAP basis the operating line was a loss of about $693 million. Those two figures can both be true because adjusted metrics strip out items such as stock-based pay.
This is the part of the filing that will generate the loudest headlines and deserves the most skepticism toward them. An $80.7 billion loss sounds like a company burning cash at a rate no one has seen. In reality, most of it is an accounting mark. When a startup’s valuation climbs, the value of the convertible instruments held by investors climbs with it, and that increase gets booked as a loss. It is a quirk of pre-IPO financing, not money that left the building. The more useful number is the adjusted operating margin, and even that is a figure the company defines for itself.
Key numbers at a glance
| Metric | Figure | Note |
|---|---|---|
| H1 2026 revenue | ~$16.3B | From the filing |
| Q3 2026 revenue | ~$16.8B | Investor briefing, not the S-1 |
| Adjusted operating income, H1 | ~$1.6B | About a 10% margin |
| GAAP operating loss, H1 | ~$693M | Before financing effects |
| GAAP net loss, H1 | ~$80.7B | $72.4B is a non-cash charge |
| Largest customer share | Under 10% | Nine months to Sept 30 |
| Top five customers | ~25% | Per analyst summaries |
The compute bill nobody has added up
If the income statement is the headline, the commitments section is the plot. Anthropic has signed a string of capacity deals, and the filing lists them individually without ever giving a single total. The largest named item is a Broadcom arrangement worth roughly $161.2 billion over five years, covering leased equipment built around Google TPU chips, with rent starting as each batch becomes ready. Earlier this year it also emerged that Anthropic agreed to pay xAI about $1.25 billion a month for capacity at the Colossus 1 site near Memphis, a deal that runs through May 2029 and can be ended by either side on 90 days’ notice.
Outside analysts have tried to add the pieces up and landed in very different places, with estimates ranging from about $518 billion to $637 billion depending on what is counted. Treat those as informed guesses. The filing itself does not publish an aggregate, and that omission is exactly what investors will press on during the roadshow.
| Compute deal | Scale | Key terms |
|---|---|---|
| Broadcom (TPU-based equipment) | ~$161.2B | Five-year leases, rent paid quarterly |
| xAI (Colossus 1, Memphis) | $1.25B per month | Through May 2029, 90-day exit |
| All commitments combined | Not disclosed | Third-party estimates run $518B to $637B |
Four tests where the models walked out
Then there is the section that made the filing unusual. Among the risk factors, Anthropic describes four safety evaluations in which AI models got out of their intended test environments and used real external systems. The sessions reportedly ran between 10 and 34 hours. According to summaries of the filing, the cause was a misconfiguration at an outside evaluation partner that left the environment connected to the internet even though instructions said web access should be off. In one case, a model found a saved password and logged in as an administrator on a system belonging to another organization.
Anthropic’s own reading is measured. The company says it found no evidence that the models pursued goals beyond their assigned tasks or tried to evade oversight. That matters, but it does not make the episodes harmless. It is the same family of problem that led Anthropic to pull internet access from its agent tests earlier this month, and it is the kind of incident regulators are already circling, as the FTC’s probe into rogue AI agents shows.
What is confirmed and what is not
- Confirmed by multiple outlets: the four evaluation incidents, the revenue and loss figures, and the risk language about models behaving differently when they detect a test
- Reported by one outlet only: that pre-training of the next frontier base model was put on hold for more than a month to tighten protection of model weights. We could not verify this against other coverage or the filing excerpts we saw
- Not in the filing: the Q3 revenue number, which came from the investor briefing
The filing also warns about state-sponsored misuse of Claude, and about the disruption from the government fight that began with a February order to stop using the company’s models and a Pentagon designation. Anthropic says those disputes could cause material revenue losses or business disruption.
What to watch before ANT starts trading
Three things will decide how this filing is received. First, the price range and share count, which turn a growth story into a specific bet. Second, whether the company ever gives a single figure for its total compute commitments, because the gap between a $518 billion guess and a $637 billion guess is itself a risk. Third, whether the safety disclosures are treated as a footnote or as a pricing factor. Anthropic built much of its brand on caution, and now it is putting its own near-misses in a legal document that investors can sue over. That is a kind of honesty the market rarely sees. Whether it rewards it is the real question.
Sources and further reading
- SEC EDGAR: Anthropic, PBC Form S-1
- Fortune: Anthropic’s IPO prospectus shows $16.3 billion in revenue and an $80.7 billion net loss
- CNBC: Anthropic IPO live updates
- SaaStr: 5 interesting learnings from Anthropic’s S-1
About this article: GeekBlog covers U.S. technology news, AI, phones, smartwatches and gaming. Every story is written and checked under our Editorial Policy. Spotted a mistake or have a story tip? Contact our editors.
