Comment by bitexploder

Part of the reason for this is likely due to customers preference to have CUDA available which TPUs do not support. TPU is superior for many use cases but customers like the portability of targeting CUDA