Anthropic IPO prospectus warns of AI catastrophe while pitching growth

There is a narrower finding worth taking seriously: In Anthropic’s controlled experiments, Reuters reported, autonomous models sabotaged code, helped with fraud and distorted information. A controlled test can expose a vulnerability without establishing how often it would arise in ordinary use. Neither treating those results as an accomplished disaster nor dismissing them because they came from tests would help people judge the risk.

The prominence of the warning is striking. Reuters counted roughly 80 pages of risk factors in the prospectus’s 261-page main body, compared with 48 pages describing the business. Page counts show how much space Anthropic devotes to disclosure; they do not show whether its safeguards would work or who could enforce them.

The distinction matters because the complete prospectus is not available in the material behind these reports. Euronews reported that the document had been shared with a small group of partners. The published accounts do not establish every safety threshold, oversight provision or remedy it may contain. They do establish that Anthropic is asking prospective investors to consider extreme harms alongside a case for continued development.

The business rewards the next release

Anthropic’s pitch is not simply that AI will sell well. Reuters reported that the prospectus compares AI’s potential economic effect with past transformations such as industrialization, electricity and the internet. It says new models increase customer use and revenue, and that overlapping releases are needed to remain at the leading edge of AI development. CEO Dario Amodei has also called for the global AI community to slow the release of new capabilities, Reuters reported.

Those positions need not be insincere to be in tension. If the next model brings in customers while a delay gives rivals time to advance, the business faces a powerful reason to keep releasing. Reuters reported that Anthropic introduced a new Opus model while Amodei was urging a slower pace across the industry. A company can believe the industry should slow down and still feel pressure not to slow down alone.

Anthropic says dependable and secure AI is a shared responsibility and believes the market will reward it. That is the strongest case for its approach: Trust could be valuable to customers, while revenue from new products could support more safety work. But a hoped-for market reward is not an independent check on a release decision. Customers and investors cannot assess that claim fully without knowing what would trigger a delay and who could insist on one.

The resource figures provide only a partial view. Anthropic told Reuters that about 6% of the computing power used for AI research went to safety work during a sample week in July. The company said safety competes with other demands for computing capacity and costly talent, and that its financial return is uncertain. Reuters reported that the prospectus does not disclose a total safety-research spending figure. The sample-week percentage is not an annual budget or a measure of all safety work.

Who stands to gain from the expansion?

The scale of the financial pitch makes the release pressure concrete. Reuters reported that Anthropic’s 2025 revenue rose twelvefold to nearly $4.6 billion, while its operating loss exceeded $8 billion. The prospectus reported a net loss of about $42 billion, but Reuters said roughly $34 billion of that was an accounting charge associated with financing that could convert into shares. Calling the full net loss cash spent on operations would obscure the company’s actual costs.

Those costs are substantial without exaggeration. Anthropic spent $7.33 billion on computing and infrastructure in 2025, Reuters reported, out of $12.65 billion in operating expenses. The company described $518 billion in planned cloud, computing and infrastructure obligations in coming years.

The obligations show how much infrastructure Anthropic expects its expansion to require. The published accounts do not establish who would ultimately benefit from that spending or its full effects on workers and communities. Without a breakdown, readers cannot tell how the gains and costs would be distributed. The size of the commitment does, however, make the company’s need for continued financing and revenue difficult to separate from its decisions about what to build and release.

There is commercial uncertainty, too. Two customers accounted for nearly a quarter of Anthropic’s 2025 revenue, Reuters reported, and many of its largest customers could reduce spending without breaking long-term contracts. Investors are being asked to weigh rapid growth against large costs and concentrated revenue. People who might be affected by the harms in Anthropic’s warning face a different question: What protection would they have if commercial pressure and safety judgment conflict?

Who gets to make the safety decision?

The proposed ownership structure sharpens that question. CTech reported that a new Founder LLC would give Anthropic’s seven co-founders 50.1% of voting power through a special class of shares, while a Long-Term Benefit Trust would select four additional directors. That is CTech’s account of a proposed arrangement, not a verified final structure. If adopted as described, it would preserve majority founder voting power even as the company sought money from public shareholders.

Founder control could, in principle, help leaders resist investors demanding faster releases. It could also make it harder for shareholders to challenge those leaders. Neither possibility answers whether workers, users or people outside the company would have a meaningful say when a system poses risks to them. The published accounts do not establish the trust’s precise authority over model releases or whether it could stop one.

The case for taking extreme AI risks seriously should not crowd out nearer-term scrutiny. Sarah Shoker, a former OpenAI geopolitics-team leader, told The Associated Press that emphasis on existential threats can draw attention away from military AI, surveillance, hacking and data-center impacts. AP also reported that AI-safety critics worry company-favored oversight could disadvantage smaller competitors. That is a concern about how rules might be designed, not evidence that Anthropic’s warnings are a tactic to suppress rivals.

Independent assessment is no simple box to check. Andrew Strait, a former British AI-safety official, told AP there are no universal standards for testing AI safety and security. Conrad Stosz, a former U.S. AI-standards official, questioned whether evaluators chosen by companies would have enough independence and access. If an IPO broadens who can invest in Anthropic without clarifying who can scrutinize or stop a dangerous release, capital will spread more widely than authority over the risk. That gap, rather than the size of the warning alone, is what the company’s public case still needs to address.