I don’t think a payment model can be figured out until the utility of the technology justifies the true cost of training and running the models. As you say, right now it’s all subsidized based on the belief it will become drastically more useful. If that happens I think the payment model becomes simple.

There's enough solid FOSS tooling out there between vLLM and Qwen3 Apache 2.0 models that you can get a pretty good assistant system running locally. That's still in the software creation domain rather than worldwide impact, but that's valuable and useful right now.