Anthropic vs. the Department of War

Anthropic launched Claude 3 Haiku and Claude 3 Sonnet on the AWS Marketplace for the U.S. Intelligence Community and AWS GovCloud. Alongside this rollout, Anthropic published its government acceptable use policy—contractual exceptions to its standard terms “carefully calibrated to enable beneficial uses by carefully selected government agencies.”

Permitted uses included legally authorized foreign intelligence analysis, combating human trafficking, identifying covert influence campaigns, and providing advance warning of military activities. All standard restrictions remained in place: the general Usage Policy continued to prohibit disinformation campaigns, the design or use of weapons, censorship, and malicious cyber operations. The government exceptions also did not extend to AI systems capable of domestic surveillance.

Donald Trump was inaugurated as president. CEO Dario Amodei notably did not attend, unlike Sam Altman (OpenAI), Tim Cook (Apple), and Sundar Pichai (Google). This absence was repeatedly cited by administration allies as early evidence of Anthropic's political hostility—a narrative that would intensify over the following year.

At the World Economic Forum in Davos the same week, Amodei urged the U.S. to maintain its ban on chip exports to China, warning of "1984 scenarios, or worse." He later published "On DeepSeek and Export Controls," arguing that DeepSeek's breakthrough made export controls "even more existentially important."

OpenAI received the first $200 million CDAO frontier AI contract—a prototype Other Transaction Agreement to develop agentic workflows for national security—and simultaneously launched "OpenAI for Government." The deal established the template all subsequent contracts would follow.

The White House released "Winning the Race: America's AI Action Plan"—a 28-page document with 90 federal policy positions. An accompanying executive order, "Preventing Woke AI in the Federal Government," mandated "truth-seeking" and "ideological neutrality" in AI prompts used by the federal government.

The plan was authored by OSTP Director Michael Kratsios, AI/Crypto Advisor David Sacks, and Secretary of State Marco Rubio. The order's framing of safety restrictions as "ideological bias" set the conceptual groundwork for the administration's later attacks on Anthropic's usage policies.

Semafor reported that Anthropic had declined requests from contractors working with the FBI, Secret Service, and ICE to make exceptions allowing its AI tools to be used for certain tasks, including surveillance of U.S. citizens. Two senior administration officials told Semafor they worried Anthropic was "selectively enforcing its policies based on politics."

This was the moment the confrontation shifted from political friction to operational conflict. Anthropic's refusal wasn't abstract policy; it blocked specific law enforcement use cases the administration wanted. The "two red lines"—no mass domestic surveillance, no fully autonomous weapons—were no longer theoretical.

Secretary Hegseth launched GenAI.mil, a secure generative AI platform for all ~3 million Department of War personnel, initially running Google's Gemini. Anthropic was the only one of the four $200M contract holders that did not join, due to its insistence on usage restrictions.

At his second Davos appearance, Amodei made international headlines by slamming the Trump administration's approval of Nvidia H200 chip sales to China, notable because Nvidia had committed up to $10 billion in Anthropic investment two months earlier.

Days later, Amodei published "The Adolescence of Technology," a nearly 20,000-word essay identifying five categories of AI risk. He argued that in some cases, large-scale AI surveillance should be considered a "crime against humanity," and wrote that "we are considerably closer to real danger in 2026 than we were in 2023"—while cautioning that "nothing here is intended to communicate certainty."

Hegseth summoned Amodei to what a senior Defense official previewed as "a sh*t-or-get-off-the-pot meeting." Present on the Pentagon side: Hegseth, Deputy Secretary Steve Feinberg, and Under Secretary Emil Michael. The Pentagon demanded Claude be available for "all lawful purposes" without limitation. Anthropic refused to drop its two red lines.

Hegseth said he would not let any company dictate Pentagon operational decisions.

Anthropic was given until 5:01 PM Friday, February 27 to comply. If it refused, Hegseth would: (1) terminate the $200M contract, (2) designate Anthropic a supply chain risk, and (3) invoke the Defense Production Act—a tool that had never been used to compel a company to produce a product it deems unsafe.

On February 26, Anthropic published a defiant statement: "These threats do not change our position: we cannot in good conscience accede to their request."

Claude hit #1 on Apple's U.S. top free apps chart, surpassing ChatGPT for the first time. ChatGPT uninstalls surged 295%; Anthropic's daily sign-ups tripled; paid subscribers more than doubled in 2026.

The consumer market delivered its own verdict. While the government moved to blacklist Anthropic, millions of individual users moved in the opposite direction—punishing OpenAI for what they perceived as opportunism and rewarding Anthropic for its refusal to capitulate.

Anthropic filed two simultaneous lawsuits. In the Northern District of California, a 48-page complaint named the DoD, Hegseth, ~20 federal agencies, and GSA Administrator Edward Forst as defendants, alleging five counts: APA violations, First Amendment retaliation (viewpoint discrimination against Anthropic's safety stance), constitutional retaliation for penalizing protected speech, Due Process violations, and APA violations for unilateral contract cancellations.

Separately, in the D.C. Circuit Court of Appeals, Anthropic filed a narrower petition challenging the supply chain risk designation under the Federal Acquisition Supply Chain Security Act of 2018.

That same day, 37 employees from OpenAI and Google DeepMind, including Chief Scientist Jeff Dean, filed an amicus brief supporting Anthropic.

Four major tech industry groups filed an amicus brief calling for a pause on the Pentagon's supply chain risk designation of Anthropic. The Computer & Communications Industry Association (CCIA), the Information Technology Industry Council (ITI), the Software & Information Industry Association (SIIA), and TechNet represent hundreds of companies with Pentagon contracts—including Google, OpenAI, Meta, Microsoft, Nvidia, Cloudflare, Adobe, Accenture, and Deloitte.

The groups argued the Pentagon bypassed the typical procurement and supply-chain security processes that Congress created, and that the supply chain risk designation was being misused against a domestic company over a policy disagreement.

The filing marks a striking moment of industry unity: competitors who normally fight for the same defense contracts collectively warning that the Pentagon's actions could chill innovation and reshape how the government treats all AI vendors.

Judges Henderson, Katsas, and Rao denied Anthropic's emergency motion to stay the supply-chain risk designation issued under 41 U.S.C. § 4713. While Judge Lin's preliminary injunction in the Northern District of California blocked the presidential directive and Hegseth's order, this D.C. Circuit case addresses the § 4713 designation directly—and that designation remains in effect.

The per curiam opinion acknowledged that Anthropic's petition raises "novel and difficult questions" and that the company would likely suffer "some degree of irreparable harm" without a stay. But the panel found the balance of equities favored the government, noting that a stay would force the military to "prolong its dealings with an unwanted vendor of critical AI services in the middle of a significant ongoing military conflict."

The court also noted that Anthropic may have financially benefited from the dispute, citing Amodei's internal statement that the public sees Anthropic as "the heroes" and a Digiday analysis suggesting the lost Pentagon contract "may turn out to be the best marketing spend in Silicon Valley for years."

Despite denying the stay, the panel granted expedited briefing and scheduled oral argument for May 19, 2026. The court directed both parties to address three specific questions: whether it has jurisdiction, whether specific procurement actions have been taken, and whether Anthropic can modify Claude's safeguards after delivery.