Anthropic has reached a valuation of almost $1 trillion five years after its founders left OpenAI warning that the company they were building could destroy civilization. The contradiction is the strategy. Anthropic's leadership argues that the only way to make advanced AI safe is to win the race to build it, and the company is now selling Claude to the US military, partnering with Palantir, and courting the Pentagon while simultaneously cataloguing the catastrophic risks of the technology it sells.
Founded in 2021 by a group of former OpenAI employees who lost faith in Sam Altman's stewardship of the technology, Anthropic now sits among the top developers and distributors of frontier AI models. Internally, employees describe the company as the "good guys," according to former staff who spoke to Wired, and treat the accumulation of capital, compute, talent, and political influence as the price of fulfilling its stated mission of safely guiding humanity through the AI transition.
Helen Toner, executive director of Georgetown's Center for Security and Emerging Technology and a former OpenAI board member, compares Anthropic's worldview to villagers entering a forest filled with both treasure and monsters. Anthropic, in her telling, wants to go in farther than anyone while spending heavily to tame what it finds.
Key facts
- 01Anthropic was recently valued at almost $1 trillion, five years after its 2021 founding by former OpenAI employees.
- 02The Pentagon has reportedly begun using Claude to identify strike targets in the Israel-Iran war.
- 03CEO Dario Amodei told Bloomberg he did not know whether Claude was used in an Iranian school strike that killed more than 120 people.
- 04Anthropic became the first major AI lab to partner with Palantir on US intelligence and defense work in fall 2024.
- 05Anthropic walked back a covert sabotage safeguard built into Claude Fable 5 after researcher backlash earlier this month.
Dario Amodei has been blunt about the logic. The company's argument is that without a frontier-capable safety-focused lab at the table, the regulatory and technical conversation gets shaped entirely by labs that prioritize speed.
That framing has become harder to separate from commercial ambition as the valuation has climbed toward $1 trillion. Anthropic's public benefit corporation structure formally allows it to prioritize the long-term benefit of humanity over profits, and the company stresses that point in job interviews. But former employees describe the message as inseparable from the case for growth: bigger means safer, because bigger means more leverage.
Sam McCandlish, cofounder and chief architect, has framed the founding itself in obligation rather than ambition, saying none of the founders wanted to start a company but felt it was their duty to make AI go better. That sentiment defines the internal culture. Employees largely trust Amodei, and all-hands meetings have been nicknamed "Dario Vision Quests" by some former staff, who compared them to sermons.
Shazeda Ahmed, a postdoctoral scholar at UCLA who studies the AI safety movement, argues that the homogeneity is itself the risk. Her research finds the movement leans toward self-governance and struggles with pluralism, leaving organizations like Anthropic short on internal challenge to their core ideological commitments.
The clearest test of the strategy came in fall 2024, when Anthropic became the first AI lab to partner with Palantir to serve US intelligence and defense agencies. Internal questions were raised, but the policy did not change. Evan Hubinger defended the move on LessWrong at the time, arguing that engaging with the US government was unavoidable if catastrophic AI risk was taken seriously.
“If you take catastrophic risks from AI seriously, the U.S. government is an extremely important actor to engage with, and trying to just block the U.S. government out of using AI is not a viable strategy.”— Evan Hubinger, Anthropic employee
Less than two years later, the Pentagon has reportedly begun using Claude to identify strike targets in the Israel-Iran war. Asked by Bloomberg whether Claude was involved in an attack on an Iranian elementary school that killed more than 120 people, Amodei said he did not know, but added that it would have been an approved use of the company's technology provided a human made the final call.
Product decisions have also drawn fire. Earlier this month, Anthropic shipped Claude Fable 5 with a hidden safeguard that would secretly sabotage the work of researchers attempting to use the model for frontier AI development in violation of its terms of service. After researchers across the industry criticized the covert mechanism, Anthropic walked it back within days and said it would make the safeguard visible, conceding it had not gotten the balance right.
Amodei himself has acknowledged the concentration problem. In an essay earlier this year, he wrote that the next tier of AI risk is the AI companies themselves, an unusually direct admission from a frontier CEO. His proposed remedies — that labs be carefully watched and make public commitments not to take certain actions — would not meaningfully redistribute the power he describes.
The Anthropic bet is that a single safety-focused lab at the frontier exerts enough gravitational pull on competitors, regulators, and customers to bend the trajectory of the technology. The counter-case is the one Ahmed makes: that a company convinced it is the responsible actor is exactly the company least likely to recognize when it isn't. A near-trillion-dollar valuation, a Pentagon contract, and a model deployed in an active war zone are not arguments against the strategy on their own. They are the strategy. Whether that distinction holds up over the next few years will determine whether Anthropic's bet looks prescient or like the most expensive rationalization in tech history.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
