5 minute read

TL;DR: In six months the Department of War went from talking about AI to mandating it: a new AI strategy with 30 day deployment clocks, a $58.5 billion budget request, eight frontier AI vendors cleared onto classified networks, and over a million uniformed and civilian users on GenAI.mil. The money and the mandates are real. The open question, and the one that matters most to those of us who field capability for a living, is whether test and evaluation can keep up.

I have spent my career turning defense data into fielded capability, and I have never seen the Department move on a technology the way it is moving on AI right now. Not in cyber. Not in digital engineering. What follows is my read on the developments that actually matter from the past six months, and what they mean if your job, like mine, is making sure the thing we field actually works.

From Governance to Velocity

In January the Department of War released its Artificial Intelligence Strategy alongside a memo on transforming the defense innovation ecosystem. Read past the vision statements and you find something unusual: enforceable clocks. Vendors will be judged on their ability to deploy the latest models within 30 days of public release. The military departments were given 30 days to deliver federated data catalogs to the Chief Digital and AI Office, and any refusal of a CDAO data request must be justified to the Under Secretary within 7 days.

That last part deserves a pause. The Department has struggled with data ownership fights for as long as I have been in this business. Putting a 7 day shot clock on saying no to data sharing is a bigger cultural change than any model contract.

The strategy also stood up seven pace setting projects, with names like Swarm Forge, Agent Network, and GenAI.mil, and initial demonstrations are due right about now, July 2026. And the organizational chart moved underneath it all: CDAO was realigned under Research and Engineering last August, and in January the Department installed Cameron Stanley, an AWS veteran and former Maven lead, to run it.

The Money Is Real Now

The FY27 budget request asks for $58.5 billion for AI and CJADC2. Inside that number sits a $29.5 billion bet on a sovereign AI arsenal: government owned, SCIF accredited data centers stocked with state of the art GPUs and AI supercomputers. That is the largest AI infrastructure investment in the Department’s history, and it tells you where the opportunity is shifting: away from model licenses and toward secure compute, data engineering, and integration.

Autonomy got an even more dramatic number. Replicator, which the Congressional Research Service found delivered only hundreds of the promised thousands of uncrewed systems, was absorbed into the Defense Autonomous Warfare Group. DAWG’s FY27 request: $54.6 billion, up from roughly $226 million the year before. And Golden Dome’s battle management layer has been publicly anchored on AI, with $17.9 billion requested and $3.2 billion in space interceptor agreements already awarded.

From Pilots to Programs of Record

For years the fair criticism of defense AI was that it lived in pilot purgatory. That era is ending.

• GenAI.mil launched in December and hit 1.1 million users across five of six branches within two months. Pentagon personnel built more than 103,000 AI agents in about five weeks using its low code tools.
• Maven Smart System is converting from prototype to a formal program of record by the end of FY26, with the contract ceiling grown to nearly $1.3 billion and users in all eleven combatant commands.
• The Army signed a ten year enterprise agreement with Anduril worth up to $20 billion, centered on AI enabled counter UAS command and control. In my corner of the defense world, layered defense and counter UAS, that award is the clearest signal yet that autonomy at the tactical edge is now a funded line of business.

When the Air Force sunset NIPRGPT, its 700,000 user pathfinder, with three weeks notice, barely anyone blinked. That is what it looks like when generative AI stops being an experiment and becomes infrastructure.

Vendor Politics Is Now Program Risk

On May 1 the Department announced agreements with eight companies, including AWS, Google, Microsoft, NVIDIA, OpenAI, Oracle, Reflection, and SpaceX, to run AI on IL6 and IL7 classified networks. The Department’s CTO said plainly that it would be irresponsible to rely on any one partner.

The name missing from that list tells the other half of the story. After a public rupture over usage policies on classified networks, the Department designated Anthropic a supply chain risk, terminated its agreement worth roughly $200 million, and the company took the Department to federal court. I am not going to litigate who was right here. The lesson for program managers is simpler: a vendor’s usage policy is now a program risk you underwrite, the same way you underwrite technical maturity or supply chain exposure. Put it in your risk register.

The Part That Keeps Me Up at Night: Testing

Here is the tension nobody has resolved. Fielding velocity is now mandated. Evaluation rigor is not.

The failures are already instructive. Thirty AI enabled drone boats launched with the Navy last year ended up idling when their systems rejected inputs. A counter drone interceptor test ended in a mechanical failure that started a 22 acre fire. The Congressional Research Service could not even establish what Replicator meant by the word fielded.

The institutions are responding, quietly. The updated DOT&E manual now tracks system safety and unexpected behavior for AI enabled and autonomous systems. The Army awarded a contract in January specifically to assess AI’s unpredictable behaviors and safeguard autonomous systems. Those are the right instincts, but they are a fraction of the investment flowing to the fielding side.

I have run test programs at open air ranges for two decades, and I will say this as plainly as I can: an AI capability that has not been tested against unexpected behavior is not a capability. It is a liability with a contract number. The Department is buying velocity, and velocity is the right call given the threat. But test capacity, not model access, is now the binding constraint on how much of that $58 billion becomes fielded advantage rather than fielded risk.

So What

• If you build or buy for the Department: the opportunity has moved from models to the unglamorous middle, secure compute, data catalogs, integration, and evaluation infrastructure.
• If you manage programs: treat vendor usage policies and classified network authorization as first class risks, and demand a definition of fielded before you report a system as such.
• If you work in test and evaluation: your function just became the rate limiter for the largest technology bet in Department history. Staff it, instrument it, and say so loudly when the schedule does not include you.

The Department of War has decided AI is worth moving fast for. Now comes the harder discipline: proving, system by system, that fast and trustworthy can be the same thing.

Tags: AI