From 80,000 to 1.7 million
When the current administration took office, about 80,000 people in the department used generative AI. Last week officials said GenAI.mil has about 1.7 million users, roughly 500,000 of them heavy users, out of a force of about 3 million. Users have built about 100,000 agents. Several frontier models are available on the controlled unclassified network, and James Mazol, deputy undersecretary for research and engineering, said they will extend to SIPR, JWICS, and special access networks.
That is fast adoption by any standard. It shows that people across the department want AI in their daily work and will use it when it is put in front of them.
The quote that matters
Cameron Stanley, the chief digital and AI officer, said something at the same event that deserves more attention than the user count. "AI is getting put into everything. That scares me, mainly because of capacity." He added that compute capacity is going to be difficult to overcome.
Every one of those 1.7 million users depends on a data center and a network path to reach it. That works at a desk in the Pentagon. It does not work for a platoon under jamming, a ship in emissions control, or a team at a listening post with a satellite terminal turned off.
Two architectures, both needed
The enterprise platform and the edge are complements. GenAI.mil is the right answer for staff work, policy drafting, and analysis on connected networks. The edge needs the same kind of help, answering questions about the order, transcribing radio traffic, summarizing a shift, reading a captured document, delivered a different way.
Special operators said as much at SOF Week in May. They asked for smaller models that need less compute and still understand intent, and for processing close to where the data is collected, because cloud tools do not reach remote areas.
How EdgeLance runs AI at the edge
EdgeLance runs mission AI on the hardware the unit already carries: phones, laptops, and forward servers. A unit can use any model its program has approved, local or API, and the inference gateway routes each request to the best place it can run right now. When the link is up and policy allows, a request can go to an enterprise model. When the link is down, it runs on the device or on a peer across the mesh.
Answers come with citations to the unit's own documents, and the models a team carries are staged in mission packs and checked before step-off. Stanley's capacity problem also gets smaller this way. Every inference that runs on a device the unit already owns is one the data center does not have to serve. The AI capability page covers the rest.
Count the users who can't connect
The department has proved demand for AI at enterprise scale. The next number worth tracking is how many of those users can still get an answer when the network is gone. For the people closest to the fight that number is close to zero today, and it is the one that matters most in a war with a capable jammer. We made the cost case for this in renting versus owning intelligence last November.