Hardware shortages force Pentagon to keep relying on Anthropic despite security concerns

Anthropic’s dispute with the Pentagon has turned into a case study in how hard it is for the U.S. government to replace a frontier AI vendor once that vendor is embedded in sensitive workflows. In March, the Defense Department designated Anthropic a “supply-chain risk to national security” and moved to phase out its products after a clash over military use of Claude. But by late May, reporting from The New York Times and Bloomberg indicated U.S. agencies were still constrained by shortages of advanced chips and a lack of ready substitutes on classified systems, leaving Anthropic difficult to dislodge. The contradiction is rooted in earlier adoption. In July 2025, the Pentagon awarded contracts worth up to $200 million each to Anthropic, Google, OpenAI and xAI to speed military use of advanced AI, according to a Congressional Research Service brief. CRS said Anthropic later described Claude as the department’s most widely deployed frontier model and the only frontier model on classified networks. (nbcnews.com) ### How did Anthropic end up both blacklisted and indispensable? March 5 was the formal break point. NBC News and Politico reported that the Pentagon told Anthropic it had been designated a supply-chain risk, a status that the department said would require contractors to stop using Anthropic’s AI services for defense work. The Pentagon said the issue was Anthropic’s refusal to permit “all lawful purposes” for military use. (congress.gov) Dario Amodei, Anthropic’s chief executive, said the company would challenge the action in court. NBC reported Amodei said Anthropic would not agree to use of its systems for mass domestic surveillance or fully autonomous weapons, while still arguing that the company and the department shared an interest in U.S. national security. (nbcnews.com) ### Why couldn’t the government just switch to other models? The classified-network bottleneck appears to be one reason. CRS said Claude had already become the only frontier AI model operating on the Pentagon’s classified systems, with uses including intelligence analysis, modeling and simulation, operational planning and cyber operations. Replacing that kind of deployment is not the same as signing a new vendor. (nbcnews.com) May 22 reporting by The New York Times said the CIA and NSA could not fully deploy the latest models on classified systems because of a shortage of cutting-edge chips. Bloomberg reported on May 21 that the Pentagon was testing rival AI models with 25 departmental “power users” as it raced to find alternatives to Claude. Those two reports, taken together, suggest the government had new suppliers available in principle but not yet a simple operational substitute in practice. (congress.gov) ### What did the Pentagon do to reduce that dependence? May 1 brought a diversification push. The Pentagon announced agreements with eight companies — SpaceX, OpenAI, Google, Nvidia, Reflection, Microsoft, Amazon Web Services and Oracle — to deploy AI on classified networks for what it called lawful operational use. Anthropic was excluded. DefenseScoop, Nextgov and Federal News Network each described the move as an expansion of classified-network AI access to additional vendors. (nytimes.com) That widens the field, but it does not establish that every workflow already running on Claude can be moved immediately or without hardware constraints. That is an inference from the timing of the Pentagon’s vendor expansion, the reported chip shortage and the continued testing of alternatives. (war.gov) ### Where does the legal fight stand? March 26 brought Anthropic a temporary court win. Tech Policy Press reported that U.S. District Judge Rita Lin blocked the government from enforcing the ban and the supply-chain-risk designation, saying the Pentagon’s action appeared insufficiently tailored to its stated concerns. April 8 narrowed that relief. The same Tech Policy Press timeline said a three-judge panel of the U.S. (defensescoop.com) Court of Appeals for the D.C. Circuit denied Anthropic’s request for a stay, allowing the government to maintain the designation while litigation proceeds. ### What is the practical takeaway from this episode? The immediate fact is vendor dependence. (techpolicy.press) Anthropic was important enough to be deeply deployed, controversial enough to be blacklisted, and hard enough to replace that the Pentagon and intelligence agencies still had to work through alternatives under chip and infrastructure constraints. The next visible milestones are already on the calendar. Bloomberg reported the Pentagon’s replacement effort is underway through model testing with departmental users, and Anthropic’s court challenge to the designation remains active in federal court. (bloomberg.com) (nbcnews.com)