I'm hopeful that the costs are mostly fixed and up front in the training of the deep GPT. Subsequent tuning or execution of a generic model is way way cheaper, hopefully sustainably so.
As much as I care about global warming I don't think "stop using energy for technological advancement" is a credible position. AI training isn't Bitcoin, it's doing something useful. I've tipped over to thinking LLMs are a real breakthrough and worth the investment.