When they were using A100 I knew that H100 would be the next ones. But what are the next ones now?

can someone explain why this is nvidia and not TPU ?

I thought that TPU were much better than GPU at ML workloads. And cheaper overall for Google too. Why is this nvidia instead of TPU ?

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).