The Pentagon picks a lab

The sequencing here is the point, so let me get it down in order.

On Thursday, the Trump administration issued guidance directing federal agencies to stop procuring or deploying Anthropic's models, citing the company's refusal to permit unrestricted military end-use under the terms of its Usage Policy and Responsible Scaling commitments. By Thursday evening, Sam Altman had announced, on his own account, which is now the preferred channel for this sort of thing, a classified partnership between OpenAI and the Department of Defense. The gap between the ouster and the replacement was, depending on how you count, somewhere between four and six hours.¹

By Thursday evening, Sam Altman had announced, on his own account, which is now the preferred channel for this sort of thing, a classified partnership between OpenAI and the Department of Defense.

I want to be careful about what is and isn't in the primary record here. The federal guidance itself is a short memorandum; I've read it. It directs agencies to "cease onboarding and wind down existing deployments" of Anthropic models where those deployments "cannot be reconciled with mission requirements," which is the bureaucratic construction for: the vendor said no and we are not going to litigate it. The OpenAI announcement is thinner, a press note, a tweet, and the phrase "classified partnership," which is doing a great deal of work. No contract value has been disclosed. No ceiling on end-use has been disclosed. Nothing about OpenAI's Usage Policy carve-outs for this engagement has been disclosed, which is itself a disclosure of a sort.

What Anthropic actually refused

The research brief frames this as Anthropic refusing "unrestricted military use," and that phrase is roughly right but worth sharpening. Anthropic's Usage Policy, as updated in its most recent revision, permits a range of defence and intelligence applications, back-office, analytic, cyber-defensive, while prohibiting weapons development, lethal autonomy, and certain surveillance categories. The company's existing arrangements with Palantir and AWS for government deployments sit within those lines. What the administration reportedly wanted was a waiver broad enough to cover end-uses Anthropic's RSP does not permit without additional safety evaluations the company was not prepared to fast-track.

This is, in the language of the frame, AI safety as market position doing exactly what the frame says it will do. Anthropic's safety posture is not a marketing overlay; it is a set of hard constraints written into policy documents the company has staked its enterprise narrative on. When the constraint binds against a customer, even a customer the size of the US federal government, the company loses the customer. That is what the frame predicted would eventually happen. It has now happened.

The interesting question is whether losing the customer was priced in.

The OpenAI move

OpenAI's Usage Policy permits military use with more flexibility than Anthropic's, a shift the company made in early 2024 when it quietly removed the blanket prohibition on "military and warfare" from its terms. That change, at the time, looked like housekeeping. In retrospect it looks like the necessary predicate for Thursday's announcement. The company has been building toward this: the Anduril partnership, the Pentagon's Chief Digital and AI Office contracts, the quieter work through Microsoft's Azure Government tenancy. The "classified partnership" framing is new; the direction is not.

What Altman announced is, structurally, a positioning trade. OpenAI is taking the contract, and, more importantly, the category, that Anthropic has now visibly declined. The revenue matters less than the category capture. Federal AI procurement is a flywheel: once a lab is inside the classified perimeter, displacement costs rise sharply, clearances accumulate, and the vendor gets first look at follow-on work. Anthropic is now outside that perimeter by order of the executive branch, which is a more durable form of outside than losing a bid.

Whether this is good business for OpenAI depends on what the AI performativity frame picks up. The company is burning capital at a rate that requires it to be plausibly the leading AI vendor across every surface, consumer, enterprise, developer, government. A public exclusion of its main frontier competitor from federal work is worth something on its own, separate from contract revenue, because it shores up the positioning story that justifies the capex. I'd note that the announcement came from Altman's personal account rather than through a DoD press release or an OpenAI newsroom post, which is the tell. This is a narrative move as much as a procurement move.

What the frame predicts next

Three things, with varying confidence.

First, Anthropic's enterprise narrative now has to do more work. The company's pitch to Fortune 500 buyers, particularly in regulated industries, has leaned on safety posture as a feature. That pitch is unchanged, and arguably strengthened, by the federal ouster: "we are the lab that said no" is a coherent message to a general counsel worried about model behaviour. But it is a narrower market than "we are the lab that said no and also the lab the government uses," which was the pre-Thursday position. Watch for Anthropic's next enterprise-revenue disclosure and, more pointedly, for how Dario Amodei frames the refusal in the next investor-facing appearance. If it's framed as a principled cost, the positioning holds. If it's framed as a misunderstanding being worked through, the positioning is softer than the company has suggested.

Second, OpenAI's Usage Policy is now the most important document at the company that nobody is reading carefully. Whatever end-uses the classified partnership covers, those end-uses are permitted under current OpenAI policy, or the policy will be quietly updated to permit them. Either case is worth tracking. The lab that wins the classified work is the lab whose policy bends to accommodate it; that is a structural fact about the market, not a criticism.

Third, the other frontier labs have just been handed a positioning choice. Google DeepMind, xAI, and Meta all have defence-adjacent capability. Each will now decide whether to pursue the Anthropic-shaped gap or the OpenAI-shaped one. My read is that xAI will move toward the OpenAI posture aggressively and that DeepMind, constrained by Alphabet's own AI principles, will find itself closer to the Anthropic position than it would prefer. Meta is the wildcard; its open-weights strategy makes the "classified partnership" construction harder to execute cleanly, because the models are already out.

What to watch

• Anthropic's Q2 federal-revenue disclosure, and whether the company quantifies the hit.
• Whether OpenAI's Usage Policy is revised in the next sixty days, and in which direction.
• Any procurement guidance from allied governments, UK, Australia, Canada, that either mirrors or rejects the US posture.
• The contract vehicle for the OpenAI deal. A sole-source award would tell a different story than a competitive one. As of this writing, no SAM.gov notice has been posted.²

This is a slightly strange week and I think it is going to produce slightly strange outcomes.

Footnotes

¶