Behind the Curtain: How Anthropic's Pentagon deal could get revived

Mike Allen; Jim VandeHei

Here are two hard truths the Pentagon and Anthropic won't state bluntly about their feud over unfettered AI use in warfare:

1. Anthropic's AI is vastly better for warfare than any other AI on the market. It could take ChatGPT, Gemini or Grok months to come close, insiders tell us.
2. Anthropic will take a massive, long-term financial hit if it remains blacklisted by the government as a "national security supply chain risk." We're talking tens of billions of dollars in direct and indirect contracts in the coming years, the insiders say.

Why it matters: Anthropic is suing the Trump administration for nixing use of Claude, the company's large language model, after the company refused to allow its AI to be used for fully autonomous warfare or mass surveillance of Americans (which the Pentagon says is already illegal).

So a compromise seems undoable. Defense Secretary Pete Hegseth is pretty dug in. A source familiar with his thinking told us: "Unless they come back to the government and say, 'We're going to agree to any lawful use,' there isn't anything to talk about. ... The secretary's bottom line is 'all lawful uses.'"

• But there is a path, sources on both sides tell us.

Behind the scenes: In private, some leaders inside the federal government want Anthropic AI, both for warfighting and cyber defense. These officials believe Anthropic is a big reason the U.S. is probably 6-12 months ahead of China in leveraging AI for national defense. Claude is also being used extensively for the Iran war.

• Also in private, lots of people at Anthropic or advising its CEO, Dario Amodei, believe they were within inches of a deal that would have and should have satisfied both sides. They're pushing Amodei and the Pentagon to quietly revive the talks.

An Anthropic court filing included a one-paragraph email from Emil Michael, the Pentagon official negotiating with the company, saying "we are very close here" to an agreement on language. The note was dated March 4, five days after Hegseth declared the company a "Supply-Chain Risk to National Security."

• The day after the letter, Michael tweeted: "I want to end all speculation: there is no active @DeptofWar negotiation with @AnthropicAI." Earlier, Michael said on X that Amodei "is a liar and has a God-complex."

One possibility: Anthropic agrees to parameters on ensuring AI is used lawfully, and arranges to donate to Trump Accounts, and/or back other AI policies both sides support.

• Brad Gerstner could be a middleman. Gerstner's firm, Altimeter Capital, is an Anthropicinvestor, and he's an Amodei adviser. He's also an architect of Trump Accounts, which provide $1,000 to every newborn whose parents enroll, and allow parents and others to contribute until the child turns 18. Older kids can get Trump Accounts through philanthropic efforts like last year's massive gift by Michael and Susan Dell.

Amodei and top Anthropic officials have said publicly they want the country, not just themselves, to be enriched by their fast-growing company. So they could conceivably fund Trump Accounts.

• It's also possible they strike a deal by compromising on language governing how the Pentagon will use Anthropic's Claude. There are also several Trump policies Amodei supports to help grease a deal.
• It's doubtful the exact language OpenAI agreed to will suffice, which remains a substantial and perhaps insurmountable obstacle.

Either way, soon-to-be-released models are likely to stir government urgency to strike some kind of deal. Anthropic, in private discussions with government officials, is warning that the next big advancement will supercharge offensive and defensive cyber capabilities.

• That means the chances of something big, bad and very public hitting U.S. infrastructure or institutions will rise substantially. If Anthropic continues to outperform other models, the government's cyberwarriors will push to keep access to Claude.

The bottom line: Any deal probably requires marriage-counselor-level mediation between Hegseth and Amodei. This is truly the territory of: Conservative defense secretaries are from Venus, and liberal Silicon Valley CEOs are from Mars. They couldn't think, talk or act more differently.

• In fact, take the two men out of it, and almost everyone else involved in the talks tells us we wouldn't be writing this column, because Anthropic AI would be the default AI of the U.S. military — until something superior comes along.

Axios' Maria Curi contributed reporting.