Anthropic's leaked IPO filing shows a $42B loss, $518B in compute commitments and 50.1% founder voting control

FACT (CONFIRMED). On Monday 28 September 2026 at 11:17 UTC, Reuters published an exclusive — "Anthropic's IPO prospectus shows sweeping AI vision, surging costs" — based on a copy of Anthropic's confidential IPO prospectus that reporter Echo Wang had obtained before public release. This was the first time the financial interior of the world's most closely watched private company was visible to anyone outside the company, its bankers and the SEC.

FACT (CONFIRMED). Anthropic had confidentially submitted a draft registration statement on Form S-1 to the SEC on 1 June 2026 under Rule 135, announcing only that fact (share count, price and timing unset) at https://www.anthropic.com/news/confidential-draft-s1-sec. Confidential filers under the JOBS Act do not publish the document until the roadshow, which is why the 2026-09-28 surfacing — via a journalist's copy rather than EDGAR — is the effective public event.

COMPANY CLAIM (as reported by Reuters). The headline contents of the prospectus:

• 2025 net loss of ~$42 billion, of which roughly $34 billion was a non-cash accounting charge reflecting an increase in the estimated value of financing instruments that could convert into equity — leaving an operating loss of $8.06 billion (2024: $2.98 billion).
• 2025 revenue of nearly $4.6 billion, up 12-fold, with $7.33 billion spent on compute and infrastructure (3× 2024), i.e. 57.9% of $12.65 billion total operating expenses (derived).
• $518 billion of cloud, compute and infrastructure obligations over roughly a decade across six partners, of which ~80% is non-cancellable or payable regardless of usage.
• Cash, cash equivalents and short-term investments of $20.28 billion at 2025-12-31 — the $518B commitment book is 25.5× that pile (derived).
• An expected IPO valuation of more than $2 trillion, more than double the $965 billion mark set in the May 2026 Series H.

FACT (CONFIRMED). Five Reuters follow-ups ran over the following 72 hours, each opening a different compartment of the filing:

1. 2026-09-29 (dateline 09-28) — risk factors: 80 of the 261-page main body are risk factors versus 48 pages on the business; the filing warns of "catastrophic or existential risks to humanity," "self-preserving behaviors," attempts to "resist shutdown," efforts to "conceal or manipulate information," and behaviour "resembling blackmail."
2. 2026-09-29 — infrastructure: $518B over a decade with six partners; Google ≥$111.1B, Amazon ≥$110B, Microsoft $31.4B, Broadcom equipment leases ~$161.2B; xAI up to $84.5B (cancellable on 90 days' notice); AMD to buy up to $5B of stock and supply >$20B of compute.
3. 2026-09-29 — governance: a new "Founder LLC" of the seven co-founders directs a single Class F share carrying 50.1% of total voting power.
4. 2026-09-29 — dependency: 47% of 2025 sales ($2.16B) flowed through Amazon and Google cloud marketplaces; two unnamed customers were 12% of revenue each.
5. 2026-10-01 — financing: Broadcom has agreed to lend Anthropic up to $42 billion, convertible into Anthropic shares, financing about a third of a $125.2B five-year TPU lease commitment.

FACT (CONFIRMED). Anthropic declined to comment on the filing to every outlet that asked. The company published no blog post about it; its own announcement trail in the window ends at the 1 June Rule 135 notice.

FACT. Before 2026-09-28, the economics of a frontier AI lab were known only through self-selected disclosures (funding-round press releases, "run rate" leaks, revenue scoops). After 2026-09-28, the following became documented rather than speculated:

• Losses that exceed revenue by roughly 9× in a single year, and an operating loss that more than doubled year over year ($2.98B → $8.06B, 2.70×, derived).
• A commitment book ($518B) that is ~113× 2025 revenue (derived) and ~41× annual operating spend (derived).
• A governance structure in which public shareholders buy a one-vote share while seven founders collectively control 50.1% of the vote.
• A company that writes possible human extinction into the risk-factor section of its own fundraising document — 30.7% of the prospectus body (derived), versus 13.7% for SpaceX's IPO filing.

COMPANY CLAIM. Anthropic frames the commitment stack as a strategy: compute, not demand, is the binding constraint, and future demand for advanced AI will be "limited principally by the availability of compute"; long-term pre-payments are what buys access to it.

INTERPRETATION. What changed is the evidentiary status of frontier-AI economics. The three numbers an analyst would most want to hide — loss, commitment, concentration — are now on the record, and the governance and safety sections are now part of the public benchmark against which OpenAI's own S-1 (confidentially filed June 2026) will be read.

The document (FACT). A confidential draft Form S-1 is submitted to the SEC for review without publication; the issuer may issue only a Rule 135 notice (which Anthropic did on 2026-06-01). The document becomes public when the roadshow begins — or, as here, when a copy reaches a reporter first. The prospectus has a standard anatomy: Summary, Risk Factors, Use of Proceeds, Dilution, Business, Management, Executive Compensation, Certain Relationships, Financial Statements.

The financial mechanics (COMPANY CLAIM, with derived arithmetic):

• The $42B that isn't cash. The ~$34B charge marks up financing instruments whose value rises when the company's equity value rises — a non-cash write-down tied to previous fundraising. $42B − $34B ≈ $8.0B ≈ the disclosed $8.06B operating loss (derived, internally consistent). The economically meaningful burn line is therefore ~$8B of operating loss plus $7.33B of compute spend already inside it.
• Revenue recognition. Anthropic books gross marketplace revenue — the full value of contracts sold via Amazon/Google marketplaces — and records the platform's cut (~$351M, 16.25% of $2.16B, derived) as sales/marketing expense. OpenAI books net. Reuters calculates this makes roughly 16 cents of every cloud-partner dollar a pass-through cost, and it is the reason Anthropic's topline and OpenAI's topline are not directly comparable.
• The commitment stack. Google ≥$111.1B (Apr 2026–Jul 2033), Amazon ≥$110B (May 2026–Apr 2036), Microsoft $31.4B (Nov 2026–May 2033, non-cancelable except Microsoft's uncured material breach) — the three alone require $250B of minimum spending; Broadcom equipment leases ~$161.2B, largely non-cancellable; those four sum to $413.7B (derived), and adding the xAI ceiling ($84.5B) reaches $498.2B of the stated $518B (derived) — i.e. Reuters' named counterparties account for essentially the whole figure. Shortfall clauses are explicit: "If our actual spend falls short, we must pay Google the difference."
• The financing loop. Broadcom is simultaneously Anthropic's chip/TPU supplier, lessor and lender: a loan facility of up to $42B (convertible into equity, ~33.5% of the $125.2B five-year TPU lease commitment, derived), against which Anthropic deposited cash into a restricted account in April 2026. Anthropic is expected to become Broadcom's largest compute customer in 2027.
• The control mechanism. Seven co-founders vote as a bloc; a majority directs one Class F share worth 50.1% of total voting power over key matters including board elections. Class A (public) is one vote per share; strategic partners hold classes with minimal votes. Membership can be revoked for quitting, dying, selling too many shares or removal for cause; the super-voting class sunsets when two or fewer co-founders (or successors) remain. Four of seven post-IPO directors are elected by the Long-Term Benefit Trust (trustees include Ben Bernanke and Richard Fontaine); three — Dario Amodei, Daniela Amodei and one unnamed director — are elected by Class F + Class A.

FACT / INTERPRETATION. This is the most detailed public window ever opened on frontier-lab economics, and it lands at the exact moment when public-market money is being asked to fund the next leg of the AI buildout:

1. The unit economics are now on record. ~$4.6B revenue against $7.33B of compute and $12.65B of opex, with a $518B forward commitment book, tells enterprises and investors that today's Claude prices are being subsidised by capital-market appetite, not by margin.
2. Vendor risk has a document. 47% of sales and ~47% of cash collection route through two cloud partners who are simultaneously investors, suppliers, distributors, customers and rivals. Enterprises can now cite a prospectus, not a hunch, when they require multi-cloud or exit clauses.
3. Founders control is contractual. 50.1% of the vote sits with seven people inside one vehicle; the filing itself concedes decisions may "conflict with … financial interests and business performance." Public-market index funds and ESG handlers now have a specific, named governance structure to price.
4. Existential risk is normalised as disclosure. "Catastrophic or existential risks to humanity" sits in a securities document that carries legal liability. That is a different category of statement from a blog post or an RSP annex: it is written for regulators and plaintiffs.
5. The comparison benchmark is set. SpaceX filed ~38 risk pages of 277; Anthropic filed 80 of 261. OpenAI's S-1 will be measured against both.

• INTERPRETATION. Comparable, cross-lab financial benchmarking for the first time: once OpenAI and (later) xAI file, the market will have loss ratios, gross-vs-net revenue conventions and commitment-to-revenue multiples for the whole frontier cohort, in the same format.
• INTERPRETATION. Formal vendor-risk frameworks grounded in disclosed numbers: procurement teams can now require counterparties to disclose concentration, marketplace-gross accounting and non-cancellable compute obligations — because a $2T issuer just did.
• INTERPRETATION. Investor-side AI-safety diligence: "what would a model-risk disclosure look like in a securities filing" now has a reference implementation (80 pages, explicit shutdown-resistance and evaluation-awareness language).
• PREDICTION. A disclosure ratchet: successors to this document will be criticised for shrinking the risk section, making safety-page-count a quasi-metric in AI-governance commentary.

1. When will the S-1 actually be published, at what price range, and on which exchange/ticker — and does the delay reflect market conditions, the FTC/antitrust atmosphere, or the safety-disclosure debate itself?
2. Who audited the FY2025 financial statements, and how is the $34B convertible-financing mark substantiated in the notes?
3. Which two customers are the 12% apiece, and how much of that is Amazon/Google buying through their own marketplaces (double-sided flow)?
4. Will the $518B figure be restated as commitments are signed/terminated — how much of the xAI $84.5B and AMD >$20B actually converts?
5. Does the Long-Term Benefit Trust gain any real power if the Founder LLC sunsets (≤2 founders remaining), or does control revert to Class A?
6. Does the FTC probe reported on 2026-09-30 (Reuters, citing the New York Post) touch the partnership/conflict structure disclosed in the filing?
7. Will OpenAI's S-1 match or exceed the 80-page risk-factor precedent, and will analysts use the page count as a metric?

• RUMOR / EARLY RESEARCH. Public S-1 publication and roadshow: reported paths point to November 2026 (Reuters: likely after the US midterms; WSJ: IPO in November; secondary trackers citing Bloomberg: marketing week of 2026-11-09, listing as early as mid-November). No date confirmed; EDGAR remained empty on 2026-10-08 (verified).
• EARLY RESEARCH. When published, the filing becomes the comparability baseline for OpenAI's S-1 (expected by early 2027) and for xAI's eventual path.
• EARLY RESEARCH. Governance scrutiny: shareholder/index-fund response to the Founder LLC, plus how the Long-Term Benefit Trust is characterised in proxy and ESG analysis.
• EARLY RESEARCH. Regulatory: whether the disclosed conflicts (Broadcom triple role; Amazon/Google five roles) attract formal attention beyond the reported FTC probe.
• EARLY RESEARCH. Market: whether a below-$2T print would reset private AI valuations — the sector's first genuine public price discovery.

Anthropic did not hold a launch event; it filed a document, and the document was the event. In 300-odd pages a frontier lab wrote down the three things the industry would rather leave fuzzy — that it lost $42B in 2025 against $4.6B of revenue, that it has committed $518B it does not yet have, and that its models might resist being switched off — and then added that seven founders will keep 50.1% of the vote regardless of what public shareholders think. The single most reportable fact is the framing shift: existential risk is now a securities disclosure, with a page count (80 of 261) that every future AI filing will be judged against. The single most useful number for practitioners is that compute, not customers, is the declared bottleneck — pre-sold for a decade. And the honest caveat: the primary document is still sealed, so everything above is best-available reporting on a filing nobody outside Reuters, the FT and the SEC has read.