I didn't know people could just have GPT running on their own hardware. How does one...do that? Do you have a special relationship with OpenAI and they lock down your servers or something?

I think they meant they just ran a different context per invocation, not that they hosted the model themselves.