Pentagon strikes classified AI deals

The Pentagon just moved commercial AI deeper into the classified side of the U.S. military. That matters because this is where the sensitive work happens — secret planning, intelligence support, and operational decision tools. The gap until now was pretty simple: lots of frontier AI lived on public or controlled enterprise systems, but not broadly inside the Defense Department’s most restricted networks. On May 1, that changed, when the department approved seven companies to bring advanced AI into those environments. (breakingdefense.com) ### Which companies got in? The list is Amazon Web Services, Google, Microsoft, OpenAI, NVIDIA, SpaceX, and Reflection. Some of those names already have deep Pentagon ties through cloud, software, launch, chips, or existing defense programs. Reflection stands out because it is much newer than the others, which tells you this was not just a rollover for the usual giant contractors. (breakingdefense.com) ### What exactly did the Pentagon approve? Not a vague partnership. The department said these companies can deploy advanced AI capabilities onto classified networks at Impact Level 6 and Impact Level 7. IL6 is the standard bucket for secret data. IL7 is the more restrictive end of the classified stack — basically the places where the Pentagon gets far more ca(breakingdefense.com)ze data, improve situational understanding, and support decisions in complex environments. (breakingdefense.com) ### Why is classified access the real story? Because this is the jump from “interesting software” to infrastructure. An AI model on an unclassified pilot system can write summaries or help with paperwork. An AI model cleared onto secret networks can sit much closer to real military workflows. That does not mean autonomous warbots tomorrow. But it does mean these models can be used where the inputs are richer, the users are more operational, and the consequences are much higher. (breakingdefense.com) ### Why seven vendors instead of one? Basically, the Pentagon does not want to bet the future of military AI on a single company. The department’s broader AI strategy has been about scaling adoption across the enterprise, not treating AI as one bespoke program. A multi-vendor setup also gives officials leverage — on pricing, resilience, security, and performa(breakingdefense.com)media.defense.gov) ### So why is Anthropic missing? That is the sharpest part of the story. Anthropic was not included even though its tools had already appeared in defense workflows through Palantir’s Maven stack, and multiple reports say the administration had labeled the company a supply-chain risk earlier this year. The dispute appears to center on military-use terms and access limits. In plain English — the Pentagon wanted broader operational freedom than Anthropic was willing to grant. (breakingdefense.com) ### Do we know the money or rollout timeline? Not yet. The Pentagon did not say when each model would go live on the classified networks, and it did not disclose financial terms. That missing detail matters because “approved” is not the same thing as “fully deployed at scale.” The real test is which commands adopt these systems first, and for what missions. (breakingdefense.com) ### What changes now? The military now has a much bigger menu of frontier AI inside the walls that matter most. That should speed experimentation and probably procurement too. But the catch is that this is also where arguments about safety limits, oversight, autonomy, and vendor trust get much more concrete. Once AI is on classified networks, the debate stops being theoretical. (breakingdefense.com) ### Bottom line? This was not just another tech-announcement Friday. It was the Pentagon saying frontier AI is moving from pilots and policy decks into the classified operating core of the department — and that access will go to companies willing to meet its terms.