Comment by zkmon

I guess the idea is, gains from inference speed could offset the cost of upgrading the chips to a new model when really required. I think general purpose models would consolidate and release frequency might flatten out, favoring this strategy.