If 8-bit training methods take off, I think the calculus is going to change rapidly, with newer cards that have decent amounts of memory and 8-bit acceleration starting to become dramatically more cost and time effective than the venerable A100s.