The Pentagon’s AI network now demands more than the strongest model: it demands trust

The Pentagon's official May 1, 2026 release says the department entered classified-network AI agreements with eight companies: SpaceX, OpenAI, Google, NVIDIA, Reflection, Microsoft, Amazon Web Services, and Oracle. That is an important correction to early readings that reduced the list to seven names or placed xAI inside it.

Those partners will be able to deploy AI capabilities in Impact Level 6 and Impact Level 7 environments, the most sensitive layers of defense computing infrastructure. The Pentagon says the move accelerates its shift toward an "AI-first" military force, with data synthesis, situational understanding, and decision support as the stated goals.

Anthropic is not on the list. The Verge reports that the company previously had a $200 million deal to handle classified materials, but has now been left out after being labeled a supply-chain risk. That absence is the real political center of the story.

Anthropic's problem does not look like an ordinary procurement loss. The company is known for its safety positioning and rules around sensitive uses, including limits tied to mass surveillance and fully autonomous weapons. According to related reporting, those restrictions helped trigger a clash with the federal government and a temporary court injunction.

The stranger part is that officials still acknowledge the value of Anthropic's security models. Emil Michael, the Defense Department's chief technology officer, described Claude Mythos to CNBC as a separate national-security moment, while still tying Anthropic to supply-chain risk. That is state logic: capability matters, but it is not sufficient if a vendor fails political, legal, and operational vetting.

This is therefore not just a market contest between OpenAI, Google, and Microsoft. It is the formation of a military AI regime in which models, clouds, usage rules, and ownership structures are evaluated as one system. A vendor has to be powerful enough, integrated enough, and acceptable enough to enter a classified network.

That is why the category should move from a pure AI story to a society and governance story. Technical models sit at the center of the infrastructure, but the conflict is about the boundaries of state use, trust in vendors, and who gets to become part of the military computing backbone.