Pentagon Clears 8 AI Firms for Classified Networks, Excludes Anthropic

The Pentagon on Friday cleared eight technology companies to deploy their AI models and hardware on its most secure classified networks — and pointedly left Anthropic, the maker of Claude, off the list. The deals give Amazon Web Services, Google, Microsoft, OpenAI, SpaceX, Nvidia, Reflection, and Oracle access to the Department of Defense's Impact Level 6 and Impact Level 7 environments, the two highest tiers of the military's cloud security architecture.

It's the broadest single AI procurement action the Defense Department has ever announced, and it solidifies a divide that's been brewing for months: every major American AI vendor is now welcome on classified DoD systems except the one that wouldn't agree to "any lawful purpose" use.

What the deals cover

IL6 covers data classified up to the Secret level. IL7 covers compartmented intelligence and the Pentagon's most sensitive operational systems — the kind of environments where targeting decisions, signals intelligence, and command-and-control workflows live. Until this week, only a narrow set of vendors could touch those networks, and only for tightly scoped use cases.

The new agreements change that. Each of the eight companies can now deploy frontier AI models, accelerators, and managed services into IL6 and IL7 enclaves. The Pentagon's announcement framed the goal as "data synthesis," "situational understanding," and "augmenting warfighter decision-making in complex operational environments." In plainer terms: the DoD wants commercial frontier AI inside its most classified workflows, fast.

The list itself is a who's-who of American AI. AWS, Google, and Microsoft bring the hyperscale cloud and their respective frontier models. OpenAI brings GPT-5.5. Nvidia brings the chips that all of them run on. SpaceX brings Starlink and the orbital networking layer the military increasingly depends on. Reflection — the well-funded open-source frontier lab — represents the bet that domestic open models will matter for sovereign deployments. Oracle, added hours after the initial announcement of seven companies, brings classified cloud capacity it has been building out for years.

Why Anthropic is missing

The exclusion of Anthropic is not subtle. Anthropic's Claude is already running on classified networks today, embedded in Palantir's Maven targeting toolkit. The company has spent the past year arguing — in court, in op-eds, and in private negotiations — that its models can support defense work but should not be used for fully autonomous lethal weapons or for domestic mass surveillance of U.S. citizens.

The Trump administration disagreed. The DoD wants vendor commitments to support "any lawful purpose" without carve-outs, and Anthropic refused to sign that language. In response, the department designated Anthropic a national supply chain risk, effectively blacklisting it from new defense contracts. Anthropic sued. A California federal judge blocked part of the blacklisting. A D.C. appeals court let the supply chain designation stand. The litigation is ongoing.

Pentagon AI chief Radha Plumb confirmed the position again Friday, telling reporters that Anthropic's blacklist status remains unchanged and is separate from any technical evaluation of Claude Mythos, the company's flagship model. In other words: the issue is policy, not capability.

Why this matters

For the Pentagon, the deals are a victory of speed over scrutiny. The DoD has been racing to operationalize generative AI for intelligence analysis, planning, logistics, and eventually autonomous systems. Vendor agreements that require case-by-case review for each classified deployment slow that down. The "any lawful purpose" framing removes those gates and lets the department move at the pace it wants.

For the eight winning companies, this is real revenue and real strategic position. Classified workloads are some of the stickiest contracts in tech: once a model is integrated into a Combatant Command's workflow, it's enormously expensive to swap out. Google's Gemini, OpenAI's GPT-5.5, Microsoft's Azure OpenAI services, and Nvidia's accelerators now have a defined path into the country's most sensitive systems.

For Anthropic, the cost is harder to measure but easier to feel. The company's annualized revenue hit $30 billion this year and it's reportedly weighing a $50 billion funding round at a $900 billion valuation. Defense contracts are a small slice of that. But the symbolism — that the most safety-focused frontier lab is being locked out of national security work specifically because of its safety commitments — is a story that will follow the company through every enterprise sales cycle.

What's next

Implementation will be the test. IL6 and IL7 enclaves are physically and logically separated from the public internet, with strict requirements on hardware supply chains, personnel clearances, and software updates. Standing up frontier AI in those environments is non-trivial. Expect the early deployments to focus on retrieval over classified document corpora and on assistive tools for analysts, with autonomous decision-making coming later — and only after a lot of evaluation work.

Anthropic's legal fight is also not done. The company is still pursuing its claim that the supply chain designation was retaliatory, and additional rulings are expected this summer. If those go in Anthropic's favor, the picture could shift again.

The bigger question hanging over Friday's news is whether "any lawful purpose" will hold up in practice. Several of the eight winning companies have their own internal policies against autonomous lethal use and mass surveillance — Google in particular faces an ongoing employee revolt over exactly this — and how those policies interact with a contract that allows everything lawful is the next chapter to watch.

The bottom line

The Pentagon now has eight AI vendors cleared for its most sensitive networks, and the one company that asked for the strongest ethical guardrails is on the outside. Whatever you think of that outcome, it tells you something important about how the U.S. government is actually choosing AI partners in 2026: capability matters, but compliance with the customer's terms matters more.