Training is the expensive part here. It seems much more likely that the training of these models slows down drastically and is written off as a sunk cost, a few companies continue running inference on years-old models, and the free versions go away.