For weights, yes, and for inference. If you can decompose and distribute a model across enough nodes, then you can get meaningful compute out of CPUs too — for instance for tokenization and smaller models.