“It is relatively easy to track AI chips to know if they are on or off and where they are located, but more research is needed to distinguish between training and inference workloads on those chips.”