$AMD Lisa Su gives her opinion on $GOOG TPUs at the UBS Conference:

"Google has done a good job with the TPU architecture over the years. But it is a more purpose-built design. It lacks the programmability, model flexibility, and balanced training and inference capabilities that GPUs offer. GPUs combine a highly parallel architecture with high programmability, which enables fast innovation.

From our perspective, there is room for all types of accelerators. However, over the next five years, GPUs should remain the clear majority of the market because we are still early in the cycle, and software developers want the flexibility to experiment with new algorithms.

You simply cannot know ahead of time what to hard-code into an ASIC. That is the difference. So a 20%–25% share for ASIC-style accelerators seems reasonable. It is also important to recognize that this is a large and expanding market, and we will see strong innovation in both silicon and software, which will drive further differentiation across the industry."