Anthropic wants to go public on a number you're not allowed to see yet

On 1 June 2026 Anthropic announced — on its own blog, naturally — that it had confidentially filed a draft S-1 with the SEC. 'Confidential' is the operative word. The company starts the IPO clock, generates the valuation headlines, and doesn't have to show the public a single audited line of its financials until at least 15 days before any roadshow. That gap between 'announcement' and 'receipts' is not an accident of process — it's the whole point of confidential filings for companies that know their story plays better in the dark.

Let's run the tape. A $65B Series H at a $965B valuation, with co-leads who are not idiots: Altimeter, Dragoneer, Greenoaks, Sequoia, Capital Group, Coatue, D1. These are serious investors who saw the actual books. The valuation is real in that sense — people with full information set it. Then there is the ~$47B revenue run-rate, which you will see cited everywhere, and which, read carefully, is per reporting — not from the filing, not disclosed, and not auditable by anyone outside the SEC reviewers right now. The number may be right. It may be directionally right. It may be an annualisation of a particularly strong quarter. You cannot yet tell the difference, and the lack of distinguishability is structural:

The top three rows are facts. The bottom two are signals. The distinction evaporates in most headlines, which is exactly why we're making it here.

A listing reportedly could come as early as October 2026 — and the rush has a logic that is worth naming. File first, set the narrative, frame the AI-IPO race before rivals get their numbers out. Bloomberg reported that OpenAI filed its own confidential S-1 the following week. If both companies are in the public market by late 2026, the race to establish which lab commands the premium multiple is on — and whoever controls the story in the 15 days between public S-1 and roadshow sets the anchor. Anthropic moved first.

Here is what the IPO hype cannot answer and the prospectus will have to: what is the gross margin on a frontier model at scale? Training runs are staggeringly expensive. Inference at the volume Anthropic's reported run-rate implies is not cheap. If the gross margin is 50%, the $965B number requires a story about improving unit economics. If it's 70%+, the story is easier to tell. If it's below 40%, the near-trillion valuation will face serious public-market scrutiny. That single line — which the confidential S-1 is hiding — is the only number that actually matters right now. Everything else is marketing.

Watch also for customer concentration. An AI company with a handful of very large enterprise customers and strong API-volume growth looks different from one dependent on a few mega-contracts. The S-1 will have to disclose both, and the composition of that $47B (if it's confirmed) will determine whether $965B is priced on a SaaS multiple or a platform multiple. The difference, at that valuation, is in the hundreds of billions. Until the prospectus lands, you're being asked to take the management team's word for it — and the management team has every incentive to make the word sound good.