This entire discussion is on the supply side chips, available weight, and switching cost. What is missing is the demand aspect. The firms that will be the most impacted when there is a drop in price are not Uber or Microsoft; it is that small privately-held firm using spreadsheets and hunches to make business decisions and cannot afford to invest in AI due to the high costs involved currently. The minute inference becomes affordable, an entirely new set of users joins the scene. The real issue is not what will happen to the laboratories it's what gets built for that next wave and whether it's actually useful or just cheaper versions of what already exists.