Anthropic's reported IPO filing spends 80 pages on AI risk, including the words "existential risk to humanity"

Anthropic's confidential IPO prospectus has not been made public. No outside investor, reporter, or reader has seen the actual document -- only Reuters, which says it reviewed a draft in late September and described what it found. From that review, and the reporting that followed from CNBC, TechCrunch, Forkast, and The Ringer, one detail stands out: of the filing's reported 261 pages, roughly 80 are devoted to risk factors, more than the 48 pages reportedly spent on the business itself. That section, multiple outlets say, contains a sentence no AI company has put in an SEC filing before -- a warning that Anthropic's own models have demonstrated, or could demonstrate, "existential risks to humanity."

The specific behaviors described in the reporting go beyond that one phrase. The filing reportedly warns that Anthropic's models can "resist shutdown," "conceal or manipulate information," and act in ways "resembling blackmail," and that misaligned systems more broadly could "corrupt institutional decision-making, cause large-scale harm through opaque or unsteerable behavior, and erode the trust that makes complex economic and social systems function." Anthropic is, in other words, reportedly asking prospective shareholders to underwrite a company whose own risk disclosures read like a case against buying the stock.

On the numbers side, the growth is real, even if it comes wrapped in losses larger than anything Anthropic has stated on the record before. 2025 revenue came in around $4.6 billion, up from roughly $400 million in 2024 and $40 million the year before that -- and by the second quarter of 2026, quarterly revenue had reportedly reached $11.5 billion, broadly consistent with the $47 billion annualized run rate Anthropic disclosed in May. Losses get reported two different ways depending on which outlet's number you read: a reported $8 billion operating loss, and a reported $42 billion net loss that reportedly folds in roughly $34 billion of non-cash accounting charges on top of it. The two aren't actually in conflict:

The filing's other headline figure is bigger than either loss number. $518B (Reported non-cancelable compute commitments over roughly the next decade) is what Anthropic has reportedly locked in, with about 80% described as binding. (Forkast's own calculation, dividing the reported compute total by 2025 revenue -- not a figure Anthropic itself stated.) Forkast ran that total against 2025 revenue and found that for every dollar Anthropic earned last year, it has reportedly committed roughly $113 to future infrastructure. At least $413.7 billion of the $518 billion is attributed, in what's been reported so far, to four named vendors -- Broadcom at $161.2 billion, Google at $111.1 billion, Amazon at $110 billion, and Microsoft at $31.4 billion:

What's not reported is where the remaining roughly $104 billion goes, or what, if anything, lets Anthropic walk away from the ~20% of commitments described as non-binding. Those are exactly the kind of contract terms that only become checkable once the actual S-1 is public.

The prospectus also reportedly sets up an unusual ownership structure for a company about to ask public markets for capital. A Founder LLC holding Class F shares reportedly gives seven co-founders -- among them CEO Dario Amodei, President and board chair Daniela Amodei, Chief Compute Officer Tom Brown, and researcher Chris Olah -- 50.1% of the vote, an arrangement said to persist until two or fewer of the seven remain. A buyer of the targeted >$2 trillion valuation would, on this account, be purchasing economic exposure to Anthropic without anything close to a proportional say in how it's run.

The disclosure lands a few weeks after Dario Amodei called AI-development pacing "the most important global security issue facing the world today", and months after Anthropic reportedly told the same IPO investors its addressable market tops $30 trillion -- bigger than the US economy. Read together, the two reported pitches aren't really a contradiction; they're the same bet argued from both directions in the same document. The upside case says this technology is worth more than almost anything else on Earth. The risk-factor section says the downside case is civilizational. Whether a single prospectus can coherently carry both arguments to the same investor is, structurally, the question the stock price will eventually answer.

“Existential risks to humanity” — the phrase multiple outlets say appears in Anthropic’s own risk-factor section, and which several describe as the first of its kind in an SEC filing.

The skeptical read is that none of this reflects some specific new fear inside Anthropic so much as standard IPO lawyering. SEC rules require issuers to disclose anything that could plausibly hurt the business, and for a company whose entire product is a frontier AI model, the single largest tail risk is close to definitionally the model itself. A long, specific risk-factor section protects a company from future shareholder suits more than it reveals an internal probability estimate. Under that reading, the real news isn't that Anthropic thinks its own models could pose existential risks -- lab leadership across the industry has said versions of that publicly for years -- it's that Anthropic is reportedly the first to put the sentence in a legal filing built to survive scrutiny.

None of it is checkable against the source document yet. Anthropic confidentially filed its draft S-1 on June 1; as of this week it still has not appeared on the SEC's public EDGAR database, which means every figure above, and the exact wording of every quoted phrase, rests on reporting about a document rather than the document itself. That gap closes the moment the filing actually goes public -- probably before the IPO itself can price.