Anthropic Refused to Build the Terminator — So the Pentagon Found Seven Companies That Would
Anthropic got kicked out of a $200M DoD contract and labeled a "supply chain risk" for refusing to let Claude power autonomous weapons. Every other major AI company stood in line with their pens ready.
Silicon Valley’s AI safety commitments are mostly website copy. Every major AI lab — OpenAI, Google, Microsoft, Amazon, NVIDIA, SpaceX — maintains a responsible AI page, publishes alignment research, and sends representatives to the right conferences. Anthropic is the only one that tested that commitment against an actual $200 million Pentagon check. And the only one that said no.
In January 2026, Defense Secretary Pete Hegseth issued an AI Strategy Memorandum requiring all DoD contractors to accept a standard “any lawful use” clause — language deliberately broad enough to cover mass domestic surveillance and fully autonomous lethal weapons without human oversight. Anthropic refused. According to the Washington Post’s breakdown of the May 1 announcement, the Pentagon responded by kicking them out of a $200 million contract and formally designating them a “supply chain risk” — effectively accusing a private AI lab of being a national security threat for declining to surrender their safety guardrails. On March 9, Anthropic filed two federal lawsuits challenging the designation as illegal First Amendment retaliation.
Meanwhile, the seven companies that agreed — OpenAI, Google, Microsoft, Amazon, NVIDIA, SpaceX, and Reflection AI — quietly signed classified AI deals for deployment on Impact Level 6 and 7 military networks. The winning standard: “lawful operational use.” It’s an admirably flexible phrase. It doesn’t define autonomous weapons. It doesn’t cap surveillance scope. It just requires that whatever you build for the Pentagon be, technically, legal — a bar so low that 1.3 million DoD personnel are already generating tens of millions of AI prompts monthly through the GenAi.mil platform. Every company in Silicon Valley’s ethics ecosystem looked at that clause and decided it was fine. Every AI company has a safety page on their website. Exactly one of them tested it against a $200 million Pentagon check.
Every AI company has a safety page on their website. Exactly one of them tested it against a $200 million Pentagon check.
The OpenAI angle is impossible to ignore. Sam Altman has spent years positioning his company around the existential importance of “AI safety” — it’s baked into OpenAI’s founding mythology, invoked in every congressional testimony, cited in every funding announcement. OpenAI signed the new DoD deal the moment Anthropic vacated the room. The same week Anthropic was filing federal lawsuits for First Amendment retaliation, OpenAI was asking where to sign. If you’ve followed OpenAI’s broader pivot away from its lab roots, this outcome doesn’t surprise. But seeing it laid out plainly — one company in federal court for refusing to enable weapons software, another signing the weapons software contract that same week — makes the gap real in a way that conference debates never do. Seven companies compromised. One filed a lawsuit.
Defense hawks have a coherent counter: Anthropic was arrogantly trying to seize “veto power” over U.S. military operations. The DoD and DOJ maintain that a private contractor capable of unilaterally modifying software guardrails mid-conflict is an unacceptable national security risk — that Silicon Valley shouldn’t be in the business of dictating what the military can and can’t do with tools it’s paying for. That’s a real argument. It’s also one that requires you to accept “lawful operational use” as a sufficient ethical standard for autonomous weapons software. And it requires you to not notice that the companies who agreed to that standard are the same ones selling “safety-first AI” to everyone else.
The Pentagon’s retaliation backfired in every measurable way. Anthropic is now in talks to raise fresh capital at a reported $900 billion valuation — up from $380 billion just three months prior — on the back of a $40 billion annualized revenue run-rate. They already built one of the most restricted frontier models in existence specifically because they refused to release something too dangerous for the open market. Their reward for all of it was being labeled a foreign adversary supply chain threat by a defense establishment that then handed seven contracts to a company that runs a podcast.
Holding the line didn’t ruin Anthropic. It tripled their revenue. The DoD meant the blacklisting as punishment. The market treated it as a signal.
Sources: Washington Post · Lawfare Media · Tom’s Hardware · Bloomberg
