it’s clearly very expensive to train gpt-4-ish models. but how expensive is it to run them? if openai permitted, what kind of hardware would you need to run a local gpt-4?