The fact that this is running on tpus is a huge point. Counting those against the other available datacenter hardware used by others, it puts google at a huge advantage, and compute > * while scaling is still working