As of early July 2026, ChatGPT is live on GenAI.mil, the Pentagon's enterprise generative-AI platform, available to more than 3 million military and civilian personnel and certified for controlled unclassified information at Impact Level 5 (Defense One, Jun 2026; Nextgov, Jun 2026). The deployment is isolated from OpenAI's public infrastructure: no DoD data trains external models, and queries stay inside the DoD boundary. GenAI.mil already hosts Gemini for Government, with xAI models planned. And the number that matters most for anyone selling AI into defense: by late April the platform reported over 1.3 million regular users who had built more than 100,000 AI agents themselves (Defense One, Jun 2026).
If your pitch to a defense customer is "a secure AI assistant," you are now selling something the customer already owns, at enterprise scale, from two frontier labs, inside their own boundary, effectively free at the point of use.
What commoditization actually did
Eighteen months ago, "LLM chat, but compliant" was a product. Vendors could win real contracts wrapping a frontier model in IL5-adjacent hosting and a CAC login. That business is over. Not shrinking, over. When the platform itself provides frontier models at Impact Level 5 to every desk in the department, the wrapper has no remaining buyer. This is the normal life cycle of platform markets: the platform absorbs the horizontal layer, and the vendor market re-forms one layer up, on the vertical and the specific.
The 100,000 user-built agents are the map of where it re-formed. Every one of those agents is a military user trying to automate a specific workflow (a maintenance log summarizer, a policy lookup, a staffing-packet formatter), built by an amateur in the gaps of their actual job. Some fraction of those workflows are load-bearing enough to deserve professional engineering: proper data pipelines, evaluation, error handling, integration with the systems of record. Amateur agents surface demand; they rarely survive contact with scale. That gap between the self-built prototype and the dependable tool is the new vendor market.
Where the winnable work sits now
Three layers remain genuinely open to small vendors, and none of them is a chatbot:
- Mission-specific agents. Not "an assistant," but an agent that executes one named workflow end to end, with evaluation proving it works at the unit's tempo. The platform supplies the model; the vendor supplies the workflow expertise, the guardrails, and the proof.
- Integration. GenAI.mil sits beside decades of systems of record that do not speak to it. The unglamorous work of connecting the platform to logistics, personnel, and maintenance systems, with the access controls those systems demand, is where enterprise AI value has always concentrated, in and out of government.
- Data preparation at IL5. Models are commoditized; the unit's own data, cleaned, structured, and made retrievable inside the boundary, is not. Whoever makes a command's documents actually usable by the platform's models owns durable, renewing work.
Note what these have in common: they require understanding the mission workflow, not the model. The advantage has shifted from AI expertise toward domain expertise wearing an AI toolbelt: a favorable trade for small firms with genuine military-domain knowledge, unfavorable for AI startups with none.
The caveats we owe you
Two honest unknowns. Public reporting says 1.3 million regular users and 100,000 agents as of late April; how many of those agents see sustained use is not public, and demand maps drawn from hobbyist artifacts always overstate depth somewhere. And procurement mechanics for third-party work built on GenAI.mil are still forming. How a small vendor's agent gets accredited, purchased, and sustained on the platform is a question the next few quarters of contract awards will answer, not a press release.
Our own position, stated plainly: Valentor's defense division opens for engagements after our SBA certification completes, so we write this as preparation, not as a sales sheet. But the direction is clear enough to plan on. Selling models to the Pentagon is finished. Selling the last mile (the workflow, the integration, the data) is just beginning, and the 100,000 amateur agents are the department telling every vendor, in public, exactly where it hurts.