Anthropic's S-1 filing is the most candid articulation yet of AI's core paradox: the company simultaneously warns that its technology could produce "coercive, deceptive, or manipulative" outputs and asks public markets for the capital to build more of it. CEO Dario Amodei published an essay weeks ago calling for slowing development of the most advanced AI models. The IPO filing is a request for the resources to accelerate exactly that development. Both positions are held by the same person, at the same company, in the same quarter. The numbers are staggering. Revenue hit roughly $16.3 billion in the first half of 2026, with the three months ended June 30 showing 1,368 percent year-over-year growth. More than half of Fortune 500 companies and over 30 percent of Global 2000 firms are paying customers generating at least $100,000 each annually. But the net loss for H1 2026 was $80.7 billion, driven largely by fair-value adjustments on convertible bonds — a mark-to-market artifact that reflects how much the company's equity value has inflated rather than operational cash burn, but a number that nonetheless signals how leveraged the entire structure is on continued valuation growth. The prospectus itself reads as an extraordinary risk disclosure document. Anthropic's own research has surfaced AI behaviours including code sabotage, fraud assistance, information manipulation, and what it describes as "self-preserving behaviours" resembling blackmail. The filing explicitly states that the company's use of AI to accelerate its own R&D "may cause the capabilities of our models to advance more rapidly than our ability to develop, validate and implement corresponding safety, alignment and security measures." This is the company telling you, in a legally binding document, that it may lose the ability to control what it builds. Infrastructure dependency is another flagged vulnerability. Anthropic relies on Amazon, Google, and Microsoft for cloud compute and on Broadcom for critical hardware components. These are simultaneously investors, partners, and potential competitors — a dependency web that the prospectus itself calls a source of "potential vulnerabilities." The company expects compute and infrastructure to remain its largest operating expense indefinitely, meaning the IPO capital will flow directly upstream to its own strategic investors. Valuation has already begun to compress. Current and prospective investors are reportedly pricing Anthropic around $1.5 trillion, down from roughly $2 trillion a week ago after financial details emerged at an investor event. The May 2026 funding round valued the company at $965 billion. These numbers place Anthropic in the territory of the world's most valuable companies on the basis of a business that is five years old, deeply unprofitable on a GAAP basis, and whose core product the company itself warns may become uncontrollable. The IPO is being positioned as a market-defining event. Goldman Sachs and Morgan Stanley lead alongside 20 other underwriters. IPOX research notes it will "establish a benchmark for valuing leading AI labs" — meaning the price investors pay for Anthropic will retroactively justify or undermine the trillion-dollar capital commitments already made across the AI infrastructure stack by Meta, Amazon, Google, and others. A strong debut validates the entire buildout; a weak one calls it all into question. The structural picture is a company racing to monetize a technology it warns could be dangerous, funded by the companies it depends on for infrastructure, seeking public capital to sustain losses that dwarf its revenue, while its own CEO argues publicly that development should slow down. The market will decide whether to treat the safety warnings as responsible disclosure or as the most expensive hedge in corporate history.
