The Pentagon is dropping the one AI vendor that tried to hold a line, and it’s doing so days after a federal judge told the Pentagon the line it drew instead was illegal. The Department of War announced Monday that it’s adding OpenAI’s ChatGPT Mil and Starshield AI’s Grok for Government to GenAI.mil, the platform that’s already reached 1.7 million of the department’s 3 million-plus personnel since its Google Gemini launch in December. In the same news cycle, Pentagon CTO Emil Michael reiterated that the department will be “completely removing Anthropic’s platforms by the end of September.” A court had just struck down the Pentagon’s earlier designation of Anthropic as a “national security supply chain risk,” calling it illegal and baseless. The department is walking away from Anthropic anyway.

Here’s the detail that makes this more than a routine vendor swap. That “risk” designation didn’t come out of nowhere. It followed Anthropic asking for safeguards against its models being used for autonomous weapons targeting and mass surveillance, the kind of guardrail a company can build into its usage policy but can’t force a government customer to respect once the contract is signed. I’ve written before about how June’s Fable 5 and Mythos 5 shutdown showed how much leverage the US government holds over any AI vendor’s ability to keep serving customers at all, export controls in that case, contract terms in this one. What’s different this time is who’s on which side of the lever. Washington isn’t using its leverage against a foreign buyer. It’s Washington using it against the vendor that asked for limits on how its own product gets used to kill people, and a court has already said the government’s stated reason for pushing that vendor out doesn’t hold up.

I keep turning this over because it inverts the sovereignty story I’ve spent the last month writing about. Denmark built a control layer specifically so a foreign government’s export decision can’t take its AI access down in a day. AXA walked its own sovereignty architecture back toward Microsoft’s ambient indexing anyway, because deployment depth beat the principle every time procurement actually decided. Mistral’s whole pitch depends on European buyers valuing the switch they control over the convenience Microsoft offers. Every one of those stories is about a customer weighing safety and control against convenience and deciding, more often than not, that convenience wins. The Pentagon just ran that same calculation and picked the vendors with fewer objections to the customer’s use case, over the one that tried to negotiate what its own product would be used for.

None of this makes OpenAI or xAI the villain of the story. Onboarding a second and third model onto GenAI.mil is a reasonable vendor-diversification move on its own, and the Pentagon has said plainly it wants multiple frontier models rather than a single point of failure. What I can’t get past is the timing: a court just ruled the government’s justification for excluding Anthropic doesn’t survive scrutiny, and the response wasn’t to reconsider; it was to restate the September deadline in public. That’s not a department responding to a legal setback. That’s a department that had already decided the outcome and is treating the court ruling as background noise.

I don’t have a clean read on where this leaves Anthropic’s federal ambitions, and I’m not going to pretend I do. What I do think is worth sitting with is how quietly this reframes the safety-versus-deployment argument that’s been running across every AI lab all year. Anthropic built its whole brand around being the safety-conscious one, the lab willing to pause, disclose, and ask for guardrails before something goes wrong instead of after. For one customer, those guardrails collided directly with stated demand, and being safety-conscious cost it the account.

Sources