[CuriouslyC]

There are a variety of GPT models with different numbers of parameters. I can run GPT-J locally right now, and as model architectures progress and the killer app for video cards goes from games/mining to AI, we will see a combination of increased card VRAM and memory optimized models that will make running this stuff locally feasible. Training is expensive, but inference is usually fairly cheap, you just need to be able to load the model in VRAM. Thankfully there are open source groups and communities working together to train and fine tune models of this sort so we won't be completely reliant on commercial entities.