I run the models here, on machines I own.
Not for ideology. It started because I wanted to know how the thing worked, and renting an endpoint teaches you nothing about that. You send text, you get text back, and everything worth understanding happens somewhere you are not allowed to look.
So there is a rack in this house with more graphics cards in it than anyone needs, and for the first month the only thing it taught me was that I had bought the wrong ones. That is also learning, it is just the expensive kind.
The trade is honest. It is slower than the good hosted models. It is mine to fix at two in the morning. The electricity is not free and I can hear it from the next room. What I get back is that nothing leaves the building, which for some of the work I do is not a preference but the whole requirement.
The part I did not expect is what it does to your sense of the limits. When you rent, a limit feels like the state of the art — this is simply what is possible today. When you own the machine, a limit is a number. You can go and look at it. Sometimes you can move it.
That difference is most of why I keep doing it.
Local is not there yet. Everyone who runs it knows that and says so. But the gap closes in one direction only, and on the day it closes I will not have to go anywhere — it will close on the machines already sitting in this room.