The Department of Defense signed agreements with eight technology companies to deploy artificial intelligence on its most classified networks. OpenAI, Google, Microsoft, Amazon Web Services, NVIDIA, SpaceX, Oracle, and Reflection AI (an NVIDIA-backed startup that rounds out the roster nicely if you are trying to demonstrate you are not dependent on anyone in particular). The models go into Impact Level 6 and Impact Level 7 environments, the security tiers that handle secret and top-secret data.
The urgency is not subtle. GenAI.mil, the Pentagon's central AI platform, launched five months ago with Google Gemini. It now has 1.3 million users, tens of millions of prompts, and hundreds of thousands of deployed AI agents. That is a remarkably fast adoption curve for an organization that until recently needed 18 months to approve a laptop. The Department of Defense has decided it wants AI on every network it operates, including the ones where the actually sensitive work happens, and it wants it now.
Dispensable by Design
The second thing these deals communicate is that the Pentagon does not want to need any of these companies individually. The official language is about preventing "AI vendor lock." The translation: if one vendor gets difficult about terms, there are seven others. This is how the Defense Department buys jet engines, two suppliers, competed against each other perpetually, neither one allowed to become essential. That playbook is being applied to frontier AI with eight vendors instead of two, because the Pentagon just spent several months learning what happens when you depend on a lab that has opinions about how its technology should be used.
Whether substitution actually works at the frontier is a separate question. For commodity tasks like summarizing intelligence reports, any model will do, and 1.3 million Pentagon users are generating those prompts all day. For specialized capabilities, the NSA is already using Anthropic's Mythos despite the blacklist, because nothing on the eight-vendor list does the same job. The architecture makes every lab dispensable for the bulk of workloads. The remainder is where the leverage lives.
Every deal was signed for "lawful operational use," a phrase that covers whatever a military lawyer says it covers. That is, not coincidentally, the formulation Anthropic would not accept without guardrails on autonomous weapons and surveillance. Eight companies decided it was fine. They have outsourced the ethical question to judge advocates, which is a perfectly rational move if you sell software and would prefer not to also be in the business of defining the rules of engagement.
The Pentagon now has its multi-vendor AI architecture, a platform growing faster than most consumer apps, and a legal standard that amounts to "whatever is lawful." The vendors are dispensable by design. The AI, apparently, is not.
Sources: Department of War, TechCrunch, Breaking Defense
–
By the Control Plane Editorial Team