Inteldefense
The Pentagon's Seven AI Projects Are a Deployment Test
The January 2026 strategy names seven pace-setting projects. The hard rule is model parity within 30 days, which collides with accreditation.
The most important sentence in the Department of War's AI strategy is not about a model. It is about a clock.
The strategy, issued with two companion memoranda on 9 January 2026 and rolled out on 12 January, names seven pace-setting projects. Each has a single accountable leader, aggressive timelines, and monthly reporting to the Deputy Secretary and the Chief Technology Officer. Components must also identify at least three fast-follow projects within 30 days.
That structure is the story. The seven projects are not a budget line. They are a management test, and the test is whether the department can field a capability on a schedule it publishes.
The seven projects, in plain terms
The list spans warfighting, intelligence, and enterprise work. The descriptions below come from the strategy memorandum (opens in a new tab) and a detailed analysis by Holland & Knight (opens in a new tab).
| Project | What it buys |
|---|---|
| Swarm Forge | A competition that pairs warfighting units with technology firms to test AI-enabled tactics |
| Agent Network | AI agents for battle management and decision support, from planning through execution |
| Ender's Foundry | Faster simulation development and tighter feedback between simulation and operations |
| Open Arsenal | A pipeline from technical intelligence to fielded countermeasures |
| Project Grant | AI-assisted deterrence, moving from static posture to dynamic pressure |
| GenAI.mil | Frontier model access for roughly 3 million personnel at multiple classification levels |
| Enterprise Agents | A playbook for agent design and deployment in business workflows |
The names are memorable. The operating model matters more. A single leader per project means someone can be asked why a milestone slipped. Monthly reporting means the answer reaches the Deputy Secretary. A 30-day window for fast-follow nominations means components cannot wait for a perfect plan.
The 30-day rule is where this gets hard
The strategy sets "AI model parity" as a procurement criterion: the latest models should be deployable within 30 days of public release. It also directs objectivity benchmarks within 90 days and standard "any lawful use" contract language within 180 days.
Thirty days is a reasonable update cadence for a commercial API. It is a difficult cadence for a classified enclave. A new model arriving on a classified network needs an evaluation, a provenance record, a configuration change, and an authorization path. A program that has not built those steps in advance will not compress them into a month because a memo asked.
The parity rule is not a model requirement. It is a pipeline requirement, and most contractors do not have the pipeline.
That is the gap to sell against, and the gap to worry about. A demo shows the model can do the task. A deployment record shows the model can be swapped, evaluated, rolled back, and explained. The second artifact is what a 30-day cadence demands.
The "any lawful use" language adds a second filter. A model with a use policy that refuses a lawful military query becomes a program risk, regardless of its benchmark scores. Contractors building on a vendor's model need to know where that vendor draws the line before the solicitation asks.
Data is the real gate
The strategy enforces the department's data decrees and gives the Chief Digital and AI Office authority to release data to cleared users with a valid purpose. It also pushes modular open architectures so components can be replaced at commercial speed.
For contractors, this changes the evaluation. Data access, interface exposure, and the ability to plug into the department's data platform become part of the offer. A product that keeps its schema closed to protect a moat is now a product that slows the integration the strategy is trying to buy.
The same logic applies to the seven projects. None of them is a standalone tool. Each depends on data moving between a sensor, a network, a model, and a person who can act. The contractor who can show that path, with permissions and provenance, has a better story than the contractor with the highest benchmark.
What to do with the list
Do not chase all seven. Pick the one where you have a system of record, a user, and a deployment path. Then build the evidence that the strategy asks for:
- A named model version and a record of what changed between versions.
- An evaluation set that reflects the mission, not a public leaderboard.
- An authorization artifact that shows how the system is bounded.
- A rollback plan that works without a vendor's help.
- A data map that shows where the data comes from and who can see it.
That list is boring. It is also the difference between a project that reports monthly progress and a project that reports a demo.
The seven projects will be judged by whether they put a capability in front of a user, on a schedule, with an audit trail. The same test applies to every contractor who wants to support them.