About your first point, IMO the usefulness of AI will remain relatively limited as long as we don’t have continuously learning AI. And once we have that, the disparity between training and inference may effectively disappear. Whether that means that such AI will become more accessible/affordable or less is a different question.
