Pentagon slams anthropic’s bid to dodge ‘supply-chain risk’ tag, calls it a threat inside the wire
A federal judge in San Francisco will decide within weeks whether anthropic can force the U.S. military to keep buying its Claude models after the Pentagon black-listed the start-up as a supply-chain hazard. Government lawyers on Tuesday told the court the label sticks: anthropic balked at contract language that lets troops deploy its AI for “any lawful military use,” then watched its access to federal coffers evaporate.
The contract clause that detonated the fight
According to the sealed procurement file, Defense officials wanted a blanket license to fine-tune Claude for tasks ranging from logistics triage to target recognition. anthropic’s board refused, inserting hard bans on autonomous weapons design and bulk surveillance of U.S. citizens. The Pentagon read that carve-out as a potential kill-switch lodged in private hands. Translation: during a surge, the company—not the commander—could throttle code.
Defense attorneys frame the refusal as textbook “vendor risk.” In their 42-page opposition they warn that anthropic’s cloud-hosted weights require constant retraining; a silent update pushed from San Francisco could “subvert design and/or functionality” while a carrier group steams toward contested waters. The filing paints Claude not as off-the-shelf software but as a living dependency that “bleeds into operational tempo.”
Billions and constitutional smoke
anthropic counters that the freeze is retaliation dressed up as risk management. Sources close to the litigation estimate $4.7 billion in forecast federal revenue is now frozen across 11 open solicitations. The company’s emergency motion for a preliminary injunction claims the move tramples First-Amendment-protected editorial discretion and due-process protections against arbitrary blacklisting.
Justice Department litigators swat that argument aside. “This is a procurement choice, not a speech debate,” they write, noting that agencies are free to shun vendors whose terms “inject hesitation into kill-chain decisions.” They also ridicule Anthropic’s projected losses as “speculative math,” arguing that if any cash gap materializes, the proper remedy is a contract claim—not a court order compelling the Pentagon to shop somewhere it doesn’t want to.
What hangs in the balance on 24 march
The hearing will test how far a start-up can go in writing its own ethics rules before the sovereign buyer walks. A ruling for Anthropic could let every cloud vendor dictate mission scope; a ruling for the Pentagon would entrench “any lawful use” as the non-negotiable fine print in future AI tenders. Either way, the outcome will ripple through every boardroom betting on defense dollars while promising safety guardrails.
Amodei’s team has one week to file its reply brief. Meanwhile, procurement officers across the intelligence community are quietly rewriting RFPs with stiffer algorithmic-compliance annexes. The message is already echoing: if you want the Pentagon’s money, leave the kill switch at home.