Pentagon signs seven AI deals

The Pentagon just pushed frontier AI deeper into the military’s most sensitive systems. On May 1, the Defense Department said seven companies — OpenAI, Google, Microsoft, Amazon Web Services, Nvidia, SpaceX, and Reflection — can deploy AI capabilities on classified networks. That is the real news. Not another pilot, not a lab demo — actual access to Impact Level 6 and Impact Level 7 environments, where classified and top-secret work lives. (nextgov.com) ### What did the Pentagon actually sign? These are agreements to bring advanced AI tools into the Defense Department’s classified computing stack for “lawful operational use.” The Pentagon framed the point pretty plainly: faster data synthesis, better situational awareness, and help for warfight(nextgov.com)at sit inside secure systems rather than out on unclassified sandboxes. (defensescoop.com) ### Which companies made the cut? The core list repeated across coverage is seven firms: OpenAI, Google, Microsoft, Amazon Web Services, Nvidia, SpaceX, and Reflection. One outlet reported Oracle as an eighth participant, which suggests either a broader partner set or a difference between the Pentagon’s initial an(defensescoop.com)nt that drove the news cycle. (nextgov.com) ### Why do IL6 and IL7 matter? Because those labels are not marketing fluff. IL6 is the Defense Department’s standard for classified cloud workloads. IL7 is stricter still — built for top-secret and highly sensitive national-security information. So this is not just “the military using AI.” It is the military preparing to run frontier AI where the consequences of errors, leaks, or misuse are much higher. (defensescoop.com) ### Why is Anthropic missing? That is the sharp edge of the story. The Pentagon’s rollout comes after a dispute with Anthropic over limits on how its models could be used, especially around autonomous weapons and surveillance. The department later labeled Anthropic a supply-chain risk, and the fight spilled into b(defensescoop.com)t on once military customers want operational freedom. (nextgov.com) ### Why does vendor diversity matter here? The Pentagon says it wants to avoid vendor lock-in. That sounds bureaucratic, but the logic is simple — if one model family underperforms, gets restricted, or becomes politically toxic, the department does not want its whole classified AI stack trapped (nextgov.com) access. (nextgov.com) ### What changes now? The biggest shift is practical. These models are being routed through GenAI.mil, the Pentagon’s central AI platform, instead of living as isolated experiments. Google had already rolled out Gemini 3.1 Pro there in late April, which shows this was not a theoretical roadmap item waiting years to start. The plumbing is already being built. (nextgov.com) ### What is the real risk? Once frontier models sit inside classified workflows, safety becomes less about chatbot weirdness and more about control. Who can query the model? What tools can the model touch? Can outputs be audited later? Can the system prove what data it saw and what actions it tr(nextgov.com)oved from flirting with frontier AI to operationalizing it in classified space. The fight now is not whether the military will use these systems. It is who gets to supply them — and under what rules.