Anthropic Said No to the Pentagon and Faces the Consequences
On March 4, 2026, the U.S. Department of Defense labeled Anthropic as a "risk to the national security supply chain." The following day, Dario Amodei confirmed that the company would take the case to court. By March 10, the lawsuit was filed in the Northern District Court of California, Microsoft had submitted a supportive brief, and over a thousand employees from OpenAI and Google had signed a joint statement rejecting the sanctions.
The speed of escalation reveals something that press releases do not convey: this is not a mere contractual dispute. It is the first open test of strength between the state and the private AI sector over who has the ultimate authority to decide what an AI model can do when commanded by a military uniform.
The Contract That Sparks the Fire
In July 2025, Anthropic signed a $200 million contract with the Department of Defense. The agreement allowed the use of Claude in classified operations, including tracking immigrants for the Immigration and Customs Enforcement and high-value target capture operations. However, the contract had two explicit restrictions: Claude could not make lethal decisions autonomously without human approval, and it could not conduct mass and indiscriminate surveillance on U.S. citizens.
These restrictions were neither fine print nor a last-minute concession. They were the core of the value proposition that Anthropic has built since its founding in 2021, when Dario Amodei and Daniela Amodei left OpenAI to pursue AI development that prioritized safety over commercial speed. The Pentagon was aware of this when it signed the contract. It accepted these terms.
What changed was the administration. In January 2026, Defense Secretary Pete Hegseth issued a memorandum under the strategy termed "AI-First Combat Power," calling for all Defense Department AI contracts to include language of "any legal use" within 180 days. This memorandum directly collided with what Anthropic had agreed to. On February 24, Hegseth personally met with Amodei and warned him that if he did not lift the restrictions by the following Friday, he would invoke Title I of the Defense Production Act to enforce compliance or demand the model be retrained.
Amodei responded with three words that often prove costly in the corporate world: "We cannot comply."
What the Defense Production Act Cannot Buy
The threat to invoke Title I of the Defense Production Act deserves specific analysis, as it reveals the fragility of the government argument. This law was designed to ensure the state’s prioritized access to productive resources during crises: steel, semiconductors, medical equipment. Its historical application has centered on physical goods with traceable production chains.
Applying it to the design principles of a language model is a legal leap of considerable proportions. As noted in the Lawfare analysis referenced in the case coverage, Title I grants priority access but would complicate significantly the legal enforcement of removing safeguards embedded in the model or retraining it. The analogy to a steel mill does not hold weight. A government can order a plant to produce steel for torpedoes instead of construction beams. Ordering a company to modify the ethical values encoded in an AI model enters a legal territory with no clear precedents.
This ambiguity is, paradoxically, the most effective tool for the Department of Defense. As the Lawfare analysis suggests: the threat of invoking a law whose implications no one fully understands may be sufficient to force capitulation without ever using it. Uncertainty generates pressure. And pressure, in companies with investors and at-risk contracts, often works better than decrees.
Anthropic chose the only path that eliminates that ambiguity: the courts.
The Landscape of Loyalties and What It Reveals About the Sector
The most significant aspect of this episode is not the lawsuit, but who aligned with each party and how quickly.
Microsoft submitted a supportive brief on the same day Anthropic filed the lawsuit, arguing that AI should not be used for large-scale domestic surveillance or to initiate wars without human control. Jeff Dean and over forty industry figures signed a declaration warning that sanctioning a U.S. AI company could severely damage the country’s scientific and industrial competitiveness. The Information Technology Industry Council, representing Nvidia, Amazon, Apple, and OpenAI, cautioned that the sanctions could weaken the government’s access to the best products and services.
On the other side, companies like Elon Musk's xAI accepted the Pentagon's terms for classified work, consolidating a visible fracture in the sector between those who calculate that an ethical stance has an unacceptable opportunity cost and those who believe that capitulating is the start of a slippery slope.
This fracture matters because it exposes the underlying logic of each business model. A company that critically depends on government contracts to support its growth has little margin to maintain positions that irritate the market's largest customer. Conversely, a company that has built its differentiation precisely on the restrictions the government wants to remove faces the inverse problem: capitulating would destroy the asset that makes it valuable.
Anthropic is not defending principles in the abstract. It is defending the only competitive advantage that separates it from its competitors in a market where technical models converge each quarter.
The Precedent That No One Wants to Name
If the Department of Defense wins this case, the effect on the industry will not be immediate or dramatic. It will be gradual and systemic. Every company negotiating a federal AI contract will know that any usage restriction can be declared a national security risk at a time of political tension. That certainty changes design incentives even before any agreement is signed.
The Information Technology Industry Council articulated this succinctly: the sanctions could limit the government’s access to the best available products. The paradox is that the state, by pressing to eliminate restrictions, may end up with access to more obedient but less capable models developed by companies that have learned to build without limits because limits are politically inconvenient.
The Northern District Court of California will have to resolve a question that Congress has yet to answer: whether a private company can contractually establish usage restrictions on its technology when that usage involves lethal decisions or whether national security doctrine can dissolve any private agreement deemed inconvenient.
Legal experts in the case coverage suggest that the structural solution lies not in the courts but in Congress, which so far has not established specific legislative frameworks for the military use of AI systems. While that void persists, the Pentagon fills the space with memos, and companies respond with lawsuits.
What Anthropic put on the table on March 10, 2026, is not just its $200 million contract. It raises the question of whether an organization can uphold a foundational purpose when the world’s most powerful customer decides that purpose is an obstacle. The culture of any organization is the natural result of pursuing that purpose with conviction or, alternatively, the inevitable symptom of times when leadership chose the contract over conviction.
