I'd assume there is integration work involved to train on TPU, so some companies would want to stay on NVIDIA. AWS also offers their own training chip instance

I'd assume there is integration work involved to train on TPU, so some companies would want to stay on NVIDIA. AWS also offers their own training chip instances (Trainium) while also providing NVIDIA instance types (P).