They mention that they use them for inference in one of the TF papers; also the focus on per watt efficiency rather than speed/memory means it's probably for inference.