What Happened
Emil Michael, the Undersecretary of Defense for Research and Engineering (effectively the Pentagon's CTO), went public about what he found when he started reviewing AI contracts inherited from the previous administration.
The short version: AI vendors had embedded dozens of operational restrictions into contracts governing models used in sensitive military operations. These included prohibitions on moving satellites and restrictions on planning operations that might lead to kinetic strikes. Military commands including Central Command, INDOPACOM, and SOUTHCOM were running a single vendor's model in critical positions.
The worst part: vendors retained the theoretical ability to disable their models mid-operation if terms were violated. In a combat scenario, that means someone in Silicon Valley could pull the plug while soldiers are in the field.
The trigger for Michael's public comments was the Maduro raid. After the operation, a senior executive at one of the AI vendors reportedly questioned whether their software had been used, implying potential concerns about its application in the mission.
Why It Matters
This story cuts to a fundamental tension that affects everyone building on AI, not just the military. When you rely on a single vendor's model and that vendor retains control over when and how you can use it, you're operating on borrowed ground.
For enterprise AI users, the parallel is clear. If your vendor updates terms of service, changes pricing, or decides your use case falls outside acceptable use policies, your workflow breaks. The Pentagon's situation is extreme, but the structural problem - vendor lock-in with operational kill switches - exists at every scale.
Michael's proposed fixes mirror what smart organizations are already doing: multi-vendor strategies, fixed-price contracts instead of cost-plus, and simplified procurement. The SpaceX model he referenced (risk-sharing, efficiency-focused contracting) has proven effective in aerospace. Applying it to AI procurement makes sense.
Our Take
The fact that the Department of Defense - the single largest technology buyer on Earth - got locked into restrictive single-vendor AI contracts tells you everything about how fast this market moved before procurement caught up.
Michael is right that AI is infrastructure, not a discretionary product. You don't let your power company dictate which appliances you can plug in. The same logic should apply to AI models embedded in critical operations.
For the rest of us, the lesson is straightforward: build for portability. Use abstraction layers. Test with multiple models. The organizations that treat AI like interchangeable infrastructure will survive vendor changes. The ones that hardwire a single provider into their stack will eventually get a "holy cow" moment of their own.
The shift from cost-plus to fixed-price contracting is worth watching too. It forces vendors to compete on delivery rather than billing hours. If the Pentagon can make that shift work for defense AI, it sets a precedent for how large organizations buy AI services across every sector.