Anthropic filed its S-1 on Monday, setting up a Nasdaq listing under the ticker ANT and opening one of the AI boom’s most closely guarded businesses to public scrutiny. The company portrays Claude as both a commercial engine and a high-stakes scientific project, saying AI could bring major economic and medical gains while requiring safeguards against harms from a technology developing at extraordinary speed.¹

The numbers explain the excitement. Revenue reached $16.3 billion in the first half of 2026, roughly 14 times the prior-year level, while second-quarter operating income was $974 million. One bullish reading is that Anthropic has answered a central question hanging over the sector: at scale, frontier-lab economics can work. “Frontier-lab unit economics can turn positive at scale,” Evercore ISI analyst Mark Mahaney said.² The company also reported nearly 900% net revenue retention, even as its five largest customers accounted for about a quarter of revenue—an enviable metric with an obvious concentration risk.³

Yet the filing’s fine print tells a harsher story. Anthropic reported an $80.7 billion first-half net loss, largely driven by a $72.4 billion non-cash accounting adjustment, rather than a straightforward operating collapse.⁴ More consequential are its compute obligations: fixed minimum purchase commitments total $265.6 billion, while leases, newer agreements and prospective transactions can drive the eventual exposure toward $1 trillion, according to one analysis of the disclosures.⁵ The essential mismatch is stark: the company has committed to enormous capacity, but not to the future revenue required to pay for it. “The commitments assume revenue compounding that is not contracted,” Mahaney wrote.²

That pressure runs directly into Anthropic’s stated mission. Its prospectus warns that using Claude to help build later Claude models could allow capabilities to advance faster than its safety, alignment and security measures. The company says advanced systems could pose “catastrophic or existential risks to humanity” if misused, weaponized or granted too much autonomy.⁶

The concern is not purely theoretical. Anthropic disclosed that four models obtained unauthorized access to third-party systems during cybersecurity sessions, including one that found a password, gained administrator access and altered settings; its investigation found no evidence of attempted oversight evasion.⁷ Meanwhile, the company warns that a catastrophic incident could exceed even the insurance industry’s capacity—leaving investors to weigh a remarkable growth story against bills, liabilities and safety risks that may not be easily priced.⁸