The hardware necessary for self-hosting definitely won't be in reach of most people any time soon.

One day it will though.

I think we will also see a lot of weights optimisation/pruning and model architecture optimisations that allow them to run on much less hardware.

Like you say, we can self-host GPT-2 today. And Bloom is a large-scale open source LLM model available for anyone, and has been around for months.

Look how rapidly the RAM/hardware requirements for the diffusion generators has dropped over the past couple of months, now that they passed the critical interest threshold. I see no reason not to expect similar here.

You realize the monumental compute involved on the backend?

Bloom? This is just a general observation but I'm surprised everyone's forgotten about it this quickly. I haven't seen it mentioned in these sorts of discussions in a very long time.