Probably that there are already several generations of TPU hardware - the best ones go to internal use, while the older hardware gets rented out to gcp to amortize the development costs.