Comment by TOMDM

My intuition is that there would be a fairly stable base load, but doing something like switching on a new training run of a frontier model would be incredibly spiky, thousands of GPUs going from somewhat idle to 100% in seconds.