Didn't see it mentioned but it's because of Moore's law. It's slowing down and the way around it is more cores and more focused transistors to perform specific jobs that will increase IPC since frequency scaling is either not existent or frequencies will fall as we move to smaller node processes.
SMT4 is a useless feature. They would be better off implementing the Bfloat16 support into their hardware to speed up machine learning algorithms significantly. It has massive advantages over traditional FP16 and nearly on par with FP32 in accuracy. They did say they are going to support it in the interview and it's great to hear that. In my computer science class I took when Google announced it we were all pretty excited about it.
