The Pentagon's Frontier-Model Surge: 1.7M Users, 500K Power Users — and Grok and ChatGPT Heading to Classified Networks
The Department of Defense's enterprise AI platform GenAI.mil has become one of the largest frontier-model deployments in the world — and it's expanding into classified systems. At the DefenseTalks conference (reported by DefenseScoop, September 23, 2026), James Mazol, deputy undersecretary of defense for research and engineering, gave the numbers: "1.7 million of our 3 million [personnel] have used GenAI.mil and the frontier models that we have on that," with about half a million power users "using generative AI pretty much every day to do their jobs." When the second Trump administration began, only ~80,000 DoD personnel used generative AI.
The platform — launched December 2025, adopted by every service branch, with Grok and ChatGPT added August 31 alongside Gemini — is now going classified. Mazol: "And by the way, we're going to proliferate frontier models, not just on our controlled unclassified network, but also SIPR, which is secret, and then JWICS for top secret, and into [Special Access Programs] as well." Earlier this year the workforce had used the platform to create 100,000 agents.
The constraints are infrastructural, not doctrinal. Pentagon chief digital and AI officer Cameron Stanley: "AI is getting put into everything. That scares me, mainly because of capacity. We don't have the capacity. It's not just us; it's industry writ large. Capacity for compute is going to be difficult for us to overcome." He wants models that are "the right size and scope and shape for what we're using them for, so we can reduce the inference side and the inference cost."
What it means. Two threads of this topic converge here. First, adoption: while the pacing debate worries about frontier-model risk, the US military is moving the same frontier models (Grok, ChatGPT, Gemini) from unclassified back-office use toward top-secret and SAP networks — the largest single institutional bet that these systems are controllable. Second, compute: DoD's top digital official says the binding constraint is the same one squeezing everyone — inference capacity — which is why the Pentagon cares about model right-sizing, and why the data-center power fights (see Congress's First Data-Center Cost Bill: House Passes Ratepayer Protection Act 417-3, but Senate Bid Blocked by Heinrich) now have a national-security customer attached.