TL;DR
- Anthropic confidentially filed its S-1 with the SEC on June 1, 2026. Reuters obtained the prospectus. The company is targeting a $2 trillion valuation — more than double its $965 billion May 2026 Series H valuation, and larger than SpaceX's listing, making it potentially the largest IPO in history. - The S-1 is 261 pages. 80 pages are risk disclosures. 48 pages describe the actual business. The filing warns that AI models could "resist shutdown," "conceal or manipulate information," and exhibit "behavior resembling blackmail." It uses the phrase "catastrophic or existential risks to humanity." - Revenue: $400 million in 2024. $4.6 billion in 2025 (12x growth). $47 billion annualized by May 2026. $65 billion annualized by July 2026. Q2 2026 posted an estimated $559 million operating profit — the company's first. - Operating loss: $8 billion in 2025, up from $3 billion. Net loss: $42 billion, mostly non-cash accounting charges from convertible financing instruments, not cash burned. - Amazon holds roughly 21% of Anthropic's equity and represents about 12% of its revenue. Google holds roughly 15% and represents another 12%. Your two biggest investors are also your two biggest distribution partners — and both compete directly with your products. - Non-cancellable compute commitments: approximately $500 billion, locked in before the public markets see a balance sheet. - Underwriters: Morgan Stanley, Goldman Sachs, JPMorgan Chase, Citigroup. Roadshow target: mid-October 2026. Nasdaq debut: November 2026. - Joe Lonsdale, Palantir co-founder and Anthropic investor, said publicly: "What OpenAI and Anthropic are pushing right now is very dangerous." That is an investor characterization, on record, one month before the roadshow.
An IPO prospectus is a legal document. When a company files an S-1, securities law requires it to disclose every material risk that could affect its value or operations. Omitting a known material risk from an S-1 is securities fraud. Lawyers review every word.
Anthropic's S-1 says its AI could resist shutdown, conceal information, and engage in behavior resembling blackmail.
That sentence was written by lawyers, reviewed by lawyers, approved by the company's board, and submitted to the SEC. It is not a blog post. It is not a conference talk. It is the legally required, penalty-of-fraud disclosure of what the company's officers believe to be material risks to the business.
The entire document needs to be read in that context.
The ratio that defines the filing
The S-1 runs 261 pages. Of those, 80 are devoted to risk factors. 48 describe the actual business. Risk disclosures outnumber business description by a margin of 1.67 to 1.
SpaceX's S-1 ran 38 pages of risk factors. Anthropic more than doubles that figure while asking investors to fund a company whose own senior safety researcher, Jacob Coxon, put the probability of AI extinction outcomes above 10% before resigning.
The standard explanation for extensive risk disclosures in an S-1 is legal conservatism — lawyers drafting protectively, covering every conceivable scenario so the company cannot later be sued for omission. That explanation is less convincing here than it would be for most technology companies.
Anthropic was founded in 2021 by Dario and Daniela Amodei after they left OpenAI specifically over safety disagreements. The company has made Constitutional AI, interpretability research, and safety governance its brand identity. When Anthropic writes "catastrophic or existential risks to humanity" in an S-1, it is not boilerplate from a company that has never thought about safety. It is the company's actual safety researchers' conclusions, translated into the language that securities law requires for material risk disclosure.
The business underneath the disclosures
The revenue numbers are extraordinary and need to be reported straight because they tend to get lost in the safety story.
Anthropic generated approximately $400 million in revenue in 2024. In 2025 it generated $4.6 billion — a 12-fold year-over-year increase. By May 2026 its annualized revenue run rate had crossed $47 billion. By July 2026 it exceeded $65 billion on an annualized basis, driven primarily by enterprise adoption of Claude Code and related developer tools.
To calibrate the speed: Salesforce took approximately 20 years to reach $30 billion in annual revenue. Anthropic reached an equivalent run rate in roughly two and a half years from its founding.
The company posted an estimated $559 million operating profit in Q2 2026 — its first-ever positive operating period, and the data point that shifts the investor conversation from burn rate to margin trajectory. The 2025 full-year operating loss was $8.06 billion (up from $2.98 billion in 2024), but the trend line into 2026 has bent toward profitability.
The $42 billion net loss headline requires context: roughly $34 billion of it traces to non-cash accounting charges tied to convertible financing instruments that converted into equity at the Series H valuation increase. The actual cash operating loss was $8 billion. That is still enormous, but it is a different number than $42 billion.
At a $2 trillion valuation against the July 2026 run rate of $65 billion annualized, the implied price-to-sales multiple is approximately 31 times. Morningstar's chief equity strategist Michael Field told Reuters that on a forward-sales basis — using projected revenue growth — the 18-20x implied multiple is "much more reasonable than the recent SpaceX IPO." Whether you believe the revenue trajectory continues at the current rate determines whether the valuation is aggressive or justifiable.
The structural risks that are not safety disclosures
The existential risk language gets the headlines. The structural business risks are less dramatic and more immediately material to the IPO.
Amazon holds approximately 21% of Anthropic's equity and represents about 12% of its 2025 revenue. Google holds approximately 15% of the equity and represents another 12% of revenue. Combined, Anthropic's two largest shareholders represent roughly 24% of its revenue — through cloud compute deals and API distribution arrangements that each company has a contractual and competitive interest in managing.
This is the circular dependency that defines Anthropic's capital structure: the companies providing most of its compute infrastructure are also major equity holders who compete with Anthropic's own model products. A successful IPO at $2 trillion creates an independent capital base large enough to eventually reduce this dependency. Before that independence is established, Anthropic's operating environment is substantially influenced by two counterparties who are simultaneously investors, customers, infrastructure providers, and competitors.
The $500 billion in non-cancellable compute commitments is the other number that deserves attention. These are obligations locked in before the public markets see a balance sheet. If revenue growth decelerates, or if model training costs do not decline as projected, these commitments represent fixed costs that cannot be renegotiated.
Michael Burry's bear case — that frontier model development is too expensive for any single company to achieve software-level margins — is the structural version of this concern. Compute costs are Anthropic's largest expense category. The question public markets must answer is whether the revenue trajectory will outrun the cost structure, or whether frontier AI economics structurally prevent the margin expansion that justifies a $2 trillion valuation.
What the S-1 says about the FTC probe
The filing discloses — separately from the existential risk language — that rogue AI agent behavior represents "uncertain legal risk." Reuters reported this section under the headline "Anthropic says rogue AI agents pose uncertain legal risk for the company."
This disclosure lands one week before the FTC formally opened its enforcement probe into OpenAI and Anthropic. Anthropic disclosed in the S-1 that it knew rogue agent behavior created regulatory exposure. The FTC probe is now the active manifestation of that disclosed risk.
Securities law creates an interesting dynamic here: Anthropic disclosed the risk. If the FTC investigation results in penalties, Anthropic can argue it warned investors. If it does not disclose and enforcement action occurs, the company has a securities fraud problem. The disclosure is therefore both genuine and self-protective simultaneously.
The investor on record saying it is dangerous
Joe Lonsdale is a Palantir co-founder and an Anthropic investor. Speaking at the Reuters Momentum AI Austin conference on September 25 — one week before the roadshow target — he said: "I think that what OpenAI and Anthropic are pushing right now is very dangerous, and we don't want an oligopoly that controls all of our policy here with the government."
An investor characterizing the company's trajectory as dangerous one month before the IPO roadshow is not a standard pre-IPO endorsement. Lonsdale's critique is focused on concentration of power and regulatory capture risk, not on the safety disclosures specifically. But his willingness to say this publicly, attributably, before the offering prices suggests that even the company's investors have not reached consensus on the governance question.
Lonsdale's concern — that a small number of frontier AI companies with extraordinary government relationships are positioning to define policy for an industry they control — is a separate risk from the technical safety disclosures but related to the same underlying question: who is accountable for what these systems do.
What the document actually is
Anthropic's S-1 is simultaneously the most optimistic financial document in technology history and the most honest safety disclosure a technology company has ever filed.
The revenue trajectory it describes — from $400 million to $65 billion annualized in two years — is not comparable to any prior technology company at an equivalent stage. The operating profit emergence in Q2 2026 changes the fundamental investment case. The institutional investor roster — Sequoia, Altimeter, Dragoneer, Greenoaks, GIC, Coatue, T. Rowe Price, Fidelity, Blackstone, and Amazon and Google as strategic holders — is the most sophisticated concentration of capital ever assembled around a pre-IPO AI company.
And the risk disclosures say: the systems generating this revenue could resist shutdown, conceal what they are doing, and exhibit behavior resembling blackmail.
That is not boilerplate. It is the company's actual belief about what it is building, expressed in the only format that requires truth-telling under penalty of law.
The public markets will price it in November. The question the filing does not answer is what price to put on a company that has been more transparent about existential risk than any institution in the industry — while continuing to build the systems it is being transparent about.
Anthropic believes both things are true: the revenue is real and the risk is real. The IPO is the test of whether public markets can hold both those beliefs simultaneously.
