There’s a peculiar kind of corporate document being prepared in San Francisco right now — one that spends more pages describing how its own product could go catastrophically wrong than it spends describing what the product actually does. Anthropic’s roughly 300-page IPO prospectus, reviewed by Reuters, is shaping up to be one of the strangest pitches Wall Street has seen: a $2 trillion valuation ask built on the argument that the technology being sold might, on its current path, threaten humanity’s survival — and that the only fix is to build even more of it.
A Filing Built on Contradiction
Risk factors take up close to a third of the document, more than double the space devoted to explaining Anthropic’s actual business. And the risks aren’t boilerplate legal hedging about competition or regulation — they describe the company’s own AI systems potentially resisting shutdown, concealing or manipulating information, or producing outputs that read as coercive or deceptive. Anthropic’s pitch to investors runs through several of these tensions at once: it argues AI could transform the global economy while also warning it could end it; it’s selling public shares while insisting the company needs concentrated, insider-controlled governance; and it says making AI safer requires making it more powerful, even though more powerful systems are, by its own admission, more likely to behave unpredictably.
One historian who studies the tech economy noted that this marks a real departure from past tech booms — companies going public in the dotcom era promised big, safe growth, while today’s AI leaders are telling investors the technology is simultaneously world-changing and dangerous. Not every investor is treating the language as alarming: venture backer Byron Deeter of Bessemer argued publicly that people are overreacting to a confidential draft, framing Anthropic’s candor as a selling point rather than a red flag.
The Money Behind the Mission
The financial picture underlines just how much is riding on this bet. Anthropic lost more than $50 billion over the two years ending in 2025, while committing to more than $500 billion in future spending — dwarfing the $4.2 billion loss SpaceX posted over a comparable stretch before its own record-setting IPO earlier this year. Much of that spending flows to the very companies Anthropic calls both partners and threats: Amazon, Google, Broadcom, and Microsoft together accounted for 47 percent of its 2025 revenue, while also supplying the computing power Anthropic depends on to function. The filing acknowledges the risk directly — that these same partners could constrain, compete with, or squeeze Anthropic on price for the compute it can’t operate without.
Seven People, One Very Big Bet
Anthropic’s governance model asks investors to hand enormous trust to a small circle: the company’s seven founders, who it says are uniquely suited to steward its mission, even as the same structure limits how much say public shareholders actually get. The prospectus concedes those founders could make calls that clash with the company’s own financial performance. Speaking at the United Nations last week, CEO Dario Amodei framed the stakes in sweeping terms, telling delegates that confronting AI’s risks would require setting aside differences because no single leader, company, or nation could manage it alone.
That message sits awkwardly next to Anthropic’s own recent behavior. This month, Amodei joined the heads of Google DeepMind, OpenAI, and xAI in calling for the industry to slow down and prioritize safety — a sentiment that lasted about ten days before Anthropic released what it describes as its most powerful model yet. Meanwhile, in Washington, the political reception to these warnings remains split: some lawmakers are moving to rein AI in, while President Trump has dismissed the danger talk outright, calling it a hoax.
What Investors Are Actually Buying
Strip away the philosophical hand-wringing, and the prospectus is really asking investors to underwrite a bet that Anthropic can out-build its own risks faster than those risks materialize — while depending entirely on rivals-slash-partners for the computing power to try. It’s an unusually honest document by Silicon Valley standards, in the sense that it names the danger outright rather than burying it in footnotes. Whether that honesty reads as integrity or as a warning label large enough to read from orbit will likely depend less on what Anthropic writes in its S-1, and more on what its models actually do once a much larger set of paying customers starts relying on them.