My point was to show how difficult and time consuming it is to deploy actual operational compute, while demand is increasing at a faster pace because the models are becoming more token hungry, and more GPUs will collect dust for a longer time, which I discussed in my last article, “Short on Compute.”