The TPU is not a GPU nor is it commercially available. It is a chip optimized around a limited featureset with a limited software layer on top of it. It's an impressive demonstration on Google's behalf to be sure, but it's also not a shot across the bow at Nvidia's business. Nvidia has the TSMC relations, a refined and complex streaming multiprocessor architecture and actual software support their customers can go use today. TPUs haven't quite taken over like people anticipated anyways.
I don't personally think CUDA is impossible to replace - but I do think that everyone capable of replacing CUDA has been ignoring it recently. Nvidia's role as the GPGPU compute people is secure for the foreseeable future. Apple wants to design simpler GPUs, AMD wants to design cheaper GPUs, and Intel wants to pretend like they can compete with AMD. Every stakeholder with the capacity to turn this ship around is pretending like Nvidia doesn't exist and whistling until they go away.
I don’t disagree with what you are saying but I want to point out that the fact that the TPU is not a GPU is not really relevant. In the end what matters most is whether or not it can accelerate PyTorch.
I don't personally think CUDA is impossible to replace - but I do think that everyone capable of replacing CUDA has been ignoring it recently. Nvidia's role as the GPGPU compute people is secure for the foreseeable future. Apple wants to design simpler GPUs, AMD wants to design cheaper GPUs, and Intel wants to pretend like they can compete with AMD. Every stakeholder with the capacity to turn this ship around is pretending like Nvidia doesn't exist and whistling until they go away.