Anthropic's IPO Draft Filing: 80 Risk Pages, $42 Billion in Losses, and a $2 Trillion Target

Anthropic is working toward a stock market listing that could value the company at $2 trillion, and the confidential prospectus seen by Reuters devotes nearly twice as much space to risk as to the business itself — with warnings about…

Illustration: Two stacks of paper on a table, a huge dark stack looming over a thin white one with a small red weight on the thin stack, evoking risk pages dominating Anthropic's IPO filing.
Illustration
Gift article

Anthropic's IPO Draft Filing: 80 Risk Pages, $42 Billion in Losses, and a $2 Trillion Target

Anthropic is working toward a stock market listing that could value the company at $2 trillion, and the confidential prospectus seen by Reuters devotes nearly twice as much space to risk as to the business itself — with warnings about models that "resist shutdown" and $518 billion in commitments.

The listing paperwork is not publicly available, but in late September 2026 Reuters obtained a draft of the S-1 document, and several outlets — CNN, Benzinga, Forkast/Yahoo Finance, TNW and The Coin Republic — reported the details the same week. The listing is expected to be delayed until after the US midterm elections in November, according to Reuters sources cited by Benzinga. Until public documentation appears on the SEC's EDGAR database, this is everything we know — and everything below is built on secondary reporting.

The prospectus that warns against its own product

The most-reported detail is the proportions: of a prospectus of roughly 300 pages, nearly a third — around 80 pages — is devoted to risk factors, against just 48 pages describing the business itself, according to Reuters as cited by CNN Business (cd7f6395). Forkast's analysis of the document puts the total at 261 pages — so the page counts differ between sources, as expected for a draft that can be revised.

The risk warnings are unusually direct. Reuters writes that the prospectus warns that advanced models may exhibit "self-preserving behavior," including attempts to "resist shutdown" or "conceal or manipulate information," or produce output that could be interpreted as "coercive, deceptive or manipulative" (e31b9ca7). It is worth underscoring: these are the company's own risk descriptions in a legal document, not an external assessment that the models actually do this.

At the same time, the prospectus is a sales document. It describes what the company calls "Powerful AI" — a system that, according to the company itself, "can surpass top human experts at nearly all cognitive domains," designed to collaborate with other AI systems, operate lab equipment, robots and other machines, and run for weeks continuously with little or no human intervention (e31b9ca7). The same document thus promises both technology that exceeds humans and pages warning about the consequences.

The financials: $4.6 billion in revenue, $518 billion in commitments

The numbers behind the vision are heavy. Revenue in 2025 was around $4.6 billion ($4.59 billion according to Forkast), but the company had, according to CNN citing Reuters, a net loss of $42 billion in 2025 (cd7f6395). The Reuters article simultaneously states that the company lost "over $50 billion" in the two years ending in 2025 (e31b9ca7). The figures need not contradict each other, but they are reported differently and should be kept apart until the full prospectus is available.

The loss is also partly an accounting matter: according to The Coin Republic, citing the Financial Times and the prospectus, roughly $34 billion of the $42 billion came from a non-cash accounting item. The operating loss was thus about $8.06 billion on roughly $12.65 billion in operating costs, and the company had $20.28 billion in cash at the end of the year (3826ff1a).

The cost side is dominated by compute: $7.33 billion in computing and infrastructure spending in 2025, a tripling from 2024, and more than half of total operating costs (57265302). And this is the past. The prospectus describes, according to Forkast, roughly $518 billion in planned cloud, computing and infrastructure commitments, distributed as follows: around $161.2 billion to Broadcom, $111.1 billion to Google, $110 billion to Amazon and $31.4 billion to Microsoft. About 80% of the commitments are said to be binding and non-cancellable (a180fecd).

It is against this picture that the $2 trillion valuation target should be read: a listing of that magnitude could, according to Reuters, be the largest in history, and the company's previous private valuation was $965 billion in May 2026 (57265302). But $2 trillion is the prospectus's target, not a confirmed transaction.

Seven founders with half the votes

Anyone buying shares in a company with $518 billion in binding supplier commitments is buying a company where ordinary shareholders have limited power. Forkast describes a Founder LLC structure that gives seven co-founders 50.1% of voting rights through a Class F share — among them CEO Dario Amodei, president and chair Daniela Amodei, Chief Compute Officer Tom Brown and Chris Olah. The founders are also said to have pledged 80% of their own equity to charity (a180fecd).

Customer concentration is another governance risk: two customers each accounted for around 12% of 2025 revenue, according to the prospectus, and many large enterprise customers lack long-term contractual commitments (3826ff1a).

The conflict line to Washington

The prospectus describes, according to TNW's account of Reuters, several clashes with the US government over the past year: the president ordered federal agencies to stop using Anthropic's models in February; the Pentagon labeled the company a supply chain risk, which a judge temporarily blocked in March (bdf9169b). Public contracts account for less than 1% of revenue, but the conflict is best illustrated by the export controls: the Commerce Department imposed controls on Anthropic's top models Fable 5 and Mythos 5 in June. The company disabled both models for all customers on June 12 and restored access on July 1, after the controls were lifted (bdf9169b).

What we don't know

All of this rests on a confidential draft, known through secondary reporting in which most details can be traced back to Reuters' viewing of the document. The prospectus can be revised before any public listing, and investor Byron Deeter characterized, according to the reporting, market reactions as responses to precisely "a draft." The figures differ somewhat between sources — on losses, page counts and totals — and some disputed details, such as a reported Akamai deal, a possible NVIDIA investment and an alleged meeting between Amodei and the president, are not confirmed across sources and are therefore omitted here. The first time the public can read the document itself is when it is posted publicly on EDGAR — presumably close to a listing after the November election.

AIMag.no
AIMag.no
The AIMag.no editorial team covers artificial intelligence, tools, research, and regulation.

Get the best of AI MAG in your inbox

News, analysis, and ideas at the intersection of AI and society.