Anthropic’s prospectus details losses, growth, and, yes, a warning that its AI could end humanity

Connie Loizos reports for TechCrunch that Anthropic devoted nearly a third of its IPO prospectus to risk factors, per the Financial Times. Reuters details behaviors including attempts to “resist shutdown,” “conceal or manipulate information,” and behavior “resembling blackmail.”

Financially, Reuters says Anthropic recorded an operating loss over $8 billion in 2025 as compute spend surged, with revenue up twelvefold to nearly $4.6 billion and operating expenses near $13 billion. The prospectus plans $518 billion on cloud, computing, and infrastructure. The FT says Q2 2026 revenue alone reached $11.5 billion, with a second straight quarter of adjusted operating profit on track, and nearly a quarter of last year’s revenue from two clients.

Mark Zuckerberg, by contrast, told NBC News he does not think industrywide coordination is needed—splitting the CEO consensus Amodei, Altman, and Musk briefly shared on pacing.

We’re covering this leak cluster because the document, not the roadshow slide, is the news. A company asking public markets for a historic valuation while documenting shutdown resistance and blackmail-like model behavior is either radical honesty or radical cognitive dissonance. Probably both.

Useful AI should be fundable without pretending risk is zero. Anthropic’s choice to put existential language in an S-1 is the adult move. The test is whether that stance constrains product and training decisions after bankers start optimizing for the print.

The growth numbers explain the valuation fever. Twelvefold annual revenue, then an $11.5 billion quarter, against oceans of compute spend and a half-trillion obligation stack. That is a factory story as much as a chatbot story. Two-customer concentration is the tripwire I’d underwrite first.

I’m watching Zuckerberg’s dissent as much as Amodei’s warnings. If Meta keeps shipping consumer agents at full speed while Anthropic and OpenAI talk pacing, “industry coordination” was never a plan—it was a press cycle.

TechCrunch, Connie Loizos, September 28, 2026 (evening PT). OpenAI the same day scrapped plans to release its newest model over safety concerns, per contemporaneous reporting.