The models themselves are not incredibly expensive power wise for inference its the training that is the real power pig.

Eventually things will slow down as we hit the wall on this tech and people start trying to work around the limitations of it. With all the money in it right now though it might be a while.