[jlebar]

Without speaking to the capabilities of TPUs, note that most ML models today are mostly convolutions and matrix multiplies.