Where do you think those models are trained?

And most personal computing has tended towards only having a smartphone, and a laptop for a real PC. You can't run much in the way of an LLM on those (at least not with any real speed).