Here's an awkward breakup in the making: Defense Secretary Pete Hegseth is reportedly weighing whether to end the Department of War's partnership with Anthropic, the AI company that makes Claude. And according to Pentagon insiders, they're not particularly happy about being put in this position.

When AI Safety Meets Military Needs

Chief Pentagon spokesman Sean Parnell confirmed to Axios on Monday that the relationship with Anthropic is under scrutiny. But it was another senior Pentagon official who really let the frustration show, telling the publication that Anthropic could soon be slapped with a "supply chain risk" label. This official also noted it would be "an enormous pain" to unwind the arrangement and promised they'd make sure the company "pays a price" for forcing their hand.

That's not idle bureaucratic grumbling. If Anthropic gets designated as a supply chain risk, every company working with the Pentagon would need to certify they're not using Claude AI. That's a genuine headache because Claude is currently the only AI model integrated into the military's classified systems, and it's widely used across various business applications. Axios previously reported that Claude even played a role in the capture of former Venezuelan President Nicolás Maduro back in January.

So what's the actual problem? Anthropic is willing to loosen some of its terms of use, but the company is drawing a hard line on two things: it won't let its tools be used for mass surveillance of Americans or for developing autonomous weapons. The Pentagon considers these restrictions excessively limiting.

It's worth noting that the Trump administration has previously faced accusations of using Palantir Technologies Inc. (PLTR) to spy on U.S. citizens, claims the company has repeatedly denied. Neither the Department of War nor Anthropic responded to requests for comment.

Amodei's Outspoken Warnings

This Pentagon tension arrives on the heels of some pretty pointed public statements from Anthropic CEO Dario Amodei. He recently criticized NVIDIA Corp (NVDA) and the Trump administration over chip sales to China, warning the issue extends beyond the AI race into national security territory. Amodei went so far as to compare selling advanced chips to China with "selling nuclear weapons to North Korea."

Meanwhile, Anthropic just pledged $20 million to Public First Action, a bipartisan organization focused on AI governance and public education. The group aims to promote transparent safeguards, strengthen federal oversight, and enforce export controls to maintain U.S. leadership in AI.

Amodei has also been sounding cautionary notes about AI investment enthusiasm. He's questioned whether trillions in revenue will materialize as quickly as some investors expect, estimating it could take anywhere from one to five years. His warning: investing massive amounts of capital too early could be financially ruinous if the projections don't pan out.

For context, Anthropic's Pentagon contract is valued at up to $200 million, which represents a fraction of the company's $14 billion in annual revenue, according to the publication. So while losing the Pentagon as a customer wouldn't be existential for Anthropic, the broader implications of being labeled a supply chain risk could ripple through the defense contractor ecosystem in ways that matter quite a bit.