The military does not rent rifles
A soldier draws a rifle from the arms room. It has a serial number and a hand receipt with her name on it. It fires whether the manufacturer is having a good quarter or a bad one. Nobody in a distant office can push a policy update that stops the bolt from cycling. When she walks outside the wire, the rifle does not phone home to ask permission to work.
Defense procurement settled on this model a long time ago for anything that matters in a fight. The force owns its rifles and its trucks and its radios. It owns them because a rented capability is a capability someone else controls. And yet the newest capability entering the force, the one being wired into planning cells and intel shops and staff work, is almost entirely rented. Most military AI use today is an API call to a server owned by a commercial vendor, billed by the token, governed by terms of service written for consumers.
Three hands on the off switch
The first hand belongs to the vendor. Cloud AI pricing moved constantly through 2025. Models were deprecated, rate limits shifted, and usage policies were rewritten on the vendor's schedule rather than the customer's. None of this was malicious. It was just business. But a capability that changes terms whenever its owner needs a better quarter is not a capability a commander can plan around.
The second hand belongs to the adversary. In May 2024 the Washington Post reported that Russian electronic warfare had degraded American precision weapons in Ukraine so badly that the success rate of one GPS-guided artillery shell fell below 10 percent. A round that once hit 7 times in 10 was hitting less than 1 in 10. The shell was fine. The link it depended on was not. Any intelligence tool that requires a live connection to a distant server inherits this exact failure mode, and CSIS analysis shows China building dense electronic warfare infrastructure around the South China Sea for the same purpose. Cut the link and the rented intelligence simply stops existing.
The third hand belongs to the budget. Metered pricing means capability rises and falls with spend. Need does not enter the equation. The unit facing the worst month of its deployment is the unit running up the largest bill, and when the money runs out the intelligence runs out with it. No other piece of military equipment works this way. Trucks do not drive fewer miles in the fourth quarter.
What open weights changed
Through 2024 and 2025 something unusual happened. Research labs released capable models as open weights under permissive licenses. Full models, downloadable as files, licensed for commercial and government use. A weights file sitting on a drive you own answers to nobody. No vendor can revoke it, no jammer can reach it, and no meter runs while it works.
The quality gap between open and closed models narrowed all year. For summarization, translation, document drafting, and question answering over local files, open models crossed the threshold of useful. So the math flipped. For the first time an organization could own its intelligence outright, the same way it owns its rifles, and the marginal cost of a question dropped to the electricity required to answer it.
An engine is not a vehicle
But a weights file is an engine sitting on a garage floor. It is real power and it is useless by itself. Nobody hands a rifleman a crate engine and calls him mobile. What the engine needs is a vehicle. The vehicle is mission software, the layer that lets a team actually use an owned model under field conditions. It has to run on the hardware a unit already carries. It has to keep working when the network drops. It has to know who is allowed to ask what, log what was asked, and hold up when the person using it is tired, gloved, and in a hurry.
As of late 2025 that vehicle does not exist. The labs that release open weights are not building it. The defense primes are optimized for decade-long programs, and the cloud AI vendors have no incentive to build software whose whole point is that it never calls them. So the engine sits on the floor while the force keeps renting. The bet worth watching is not whether open models get better. They will. The bet is who builds the vehicle around them, and whether it gets built before someone learns the cost of the off switch the hard way.