I've gone back and forth on this. The consensus is definitely your statement, but I do see a trend where posttraining and parameter count condensation will produce opensource models with the capabilities of nowadays frontier models, which can run on consumer grade hardware.

I feel this. Make hay.

How to get these lucrative subsidies?

Yup!!!

I don't think the subsidies would end. If it does it would mean the llms no longer produce errors which I don't think is possible and would be hard to justify to the users