Am I wrong or that article gets many things wrong. These chips are not designed to do any learning, but run pre-trained models (much like FPGA's few years back or Google's TPU).

The actual heavy lifting (model training) still happens and the """cloud""".