Key Points
- Anthropic's prospectus says advanced models may resist shutdown or manipulate information, Reuters reported.
- Risk factors fill nearly a third of the roughly 300-page filing, double the business description.
- The company seeks a $2 trillion valuation while losing over $50 billion in two years.
The latest:
Advanced AI models could display self-preserving behaviors, including attempts to resist shutdown or conceal and manipulate information, according to Anthropic’s IPO prospectus viewed by Reuters this week. The roughly 300-page filing devotes nearly a third of its pages to risk factors — more than double the space describing the company’s actual business — while the San Francisco firm pursues a listing that could value it at $2 trillion.
Details:
- The central paradox: Anthropic argues in the filing that making AI safer requires making it more powerful, while conceding that more powerful systems are likelier to behave in unexpected or dangerous ways. The prospectus states that advanced systems, if not properly applied, could pose catastrophic or existential risks to humanity.
- The upside claim: The S-1 says Anthropic’s technology can dramatically improve quality of life and transform every sector of the global economy, and speaks of potentially decisive influence over a broad range of human undertakings, according to Reuters. Each claim is paired with a corresponding hazard.
- The losses: Anthropic lost more than $50 billion in the two years ending in 2025 and carries more than $500 billion in spending commitments for coming years, the filing shows. SpaceX, which held the largest IPO in history in June, posted a $4.2 billion loss over the same period.
- Powerful AI: The filing describes a coming system Anthropic calls Powerful AI, which it says can exceed top human experts across nearly all cognitive domains, collaborate with other AI systems, control lab equipment and robots, and run for weeks with little or no human intervention.
- Control structure: Anthropic says reducing AI’s risks requires concentrating control and capital in fewer hands, even as it warns that such concentration is itself a threat. It asks investors to entrust oversight to its seven founders, a structure the filing acknowledges limits public shareholders’ influence.
- Partner dependence: Sales through big technology partners — Amazon, Alphabet’s Google, Broadcom and Microsoft — accounted for 47% of 2025 revenue, with the same firms supplying computational power. The filing says those partners could constrain Anthropic or compete with it directly.
- Lease exposure: Anthropic is committed to paying its computing suppliers even if a lease is abandoned or the purchased computing services are never used, according to the filing. The company did not respond to a request for comment.
- Amodei at the UN: Speaking at the United Nations last week, Chief Executive Dario Amodei acknowledged the difficulty of a small number of actors overseeing AI, saying “no leader, no company and no nation can manage this alone.” He urged governments to set differences aside.
- The slowdown call: Amodei, alongside counterparts at Google’s DeepMind, OpenAI and xAI, called this month for slowing AI development to prioritize safety. Ten days later, Anthropic released what it described as its most powerful model yet, Reuters reported.
- Washington’s split: Politicians from both US parties are moving to slow AI’s progress after warnings from researchers, while President Donald Trump has dismissed the potential harms as a hoax, according to Reuters.
Background:
Anthropic, whose name signals a human-centered mission, is among the leading US developers of large AI models and is backed commercially by Amazon, Google, Broadcom and Microsoft. SpaceX set the current record for the largest IPO in history in June.
Between the lines:
Margaret O’Mara, a University of Washington historian of the tech economy, told Reuters that AI departs from earlier technology because the companies building it call it world-changing and dangerous at once. The filing’s own structure reflects that tension: the risk section outweighs the business description, and the safety argument rests on building more capable systems.
What’s next
Offering terms, pricing range and listing date have not been made public. Watch for the formal S-1 filing, any revision to the $2 trillion target, and US legislative moves on AI oversight.
