Washington Is Taking Control of AI at Three Levels at Once — and the Labs Can't Refuse

In June, the Pentagon ended a $200 million contract with Anthropic, told other contractors they could no longer work with the company, and stamped it a supply-chain risk ^[1]. The offense was not a security breach. It was a refusal: Anthropic had barred its Claude model from powering autonomous weapons and mass surveillance, and its chief executive had said so publicly. Defense Secretary Pete Hegseth made the administration's answer plain.

We, as humans, have to have the confidence that ... it's going to deliver violence only where we intend it to be delivered. — Frank M. Bradley

That fight is the most visible edge of something larger. Washington is not just arguing with one lab. It is tightening its hold on AI at three levels at once — the land the machines run on, the models themselves, and the companies that build them. None of the three moves is new on its own. What is new is that they are all happening now, and each constrains a different point of leverage. The first level is the ground. The Defense Department is soliciting private developers to build AI data centers on unused military land — Fort Hood, Fort Bliss, Dugway Proving Ground — treating them as classified infrastructure for warfighting networks ^[2]. The Energy Department is doing the same on four federal sites, from Oak Ridge to the Savannah River Site, under lease terms where developers handle construction and decommissioning ^[3]. The appeal is not subtle. On civilian land, a data center faces zoning boards, environmental reviews, and lawsuits — the kind of fight now playing out in Ohio, where a group called the Independence Guard Alliance is suing to block a Nebius facility over secret rezoning and utility discounts ^[4]. On a military base, that fight is dramatically narrowed. The Supreme Court ruled 8-0 last year that agencies need not evaluate the indirect effects of related projects, and the administration's AI Action Plan orders the streamlining of the three statutes — the Clean Air Act, Clean Water Act, and NEPA — that make up the legal framework for environmental challenge ^[5][6]. The review that remains is a fraction of what a civilian zoning fight would allow. The second level is the models. In June, Anthropic's Mythos model penetrated nearly every classified U.S. system within hours during a security test, and the administration responded by ordering a 30-day pre-release review for frontier models ^[7]. Every major lab — OpenAI, Google DeepMind, Microsoft, xAI, Anthropic — agreed to provide early access for national security evaluation before a model ships.

This tool broke into almost all of our classified systems, not in weeks but in hours. — Mark Warner

A model no longer reaches the public until the government has looked at it first. The third level is the companies. The same month it cut Anthropic loose, the Pentagon signed deals with seven firms — Google, Microsoft, AWS, Nvidia, OpenAI, Reflection, SpaceX — to integrate AI into classified military networks ^[1]. Seven in, one out. The reason this sorting bites is that the labs have never needed the government's money more. OpenAI and Anthropic are cutting model prices by 50 to 80 percent to hold off Chinese competitors like Alibaba's Qwen and DeepSeek ^[8]. Morgan Stanley now describes low-priced model access as a "loss leader" meant to pull in more profitable cloud spending on storage and security ^[9]. The model itself — the thing these companies were built to sell — is becoming the cheapest part of the stack. Both labs have responded by launching consulting ventures, embedding engineers inside enterprises rather than selling raw intelligence ^[10]. The logical consequence is that government contracts are among the few revenue streams that still pay premium rates for frontier capability — which is precisely the leverage the administration is using to sort the compliant from the punished. The paradox sits inside the pattern. The government is building the infrastructure and, at the same moment, pushing the civilian legal venues through which its military ambitions could be publicly contested out of reach — for communities and advocates, the zoning board and the environmental lawsuit are being narrowed to a sliver. And separately, the labs are losing the commercial leverage to refuse: with model prices collapsing, the government is one of the few customers left that pays for the expensive end of the business. The one lab that drew a red line just found out what refusal costs.