> their huge customers will be able to produce ASICs hat will be faster and cheaper to operate than their GPUs

Why? NVIDIA is better positioned to produce faster and more efficient ML ASICs any of their huge customers (except possibly Google). And on top of that, the fact that there is a huge library of CUDA code that will run out of the box on NVIDIA hardware is a big advantage.

Arguably, this shift has already happened. Modern NVIDIA datacenter GPUs, like the H100, only bear a passing resemblance to a GPU -- most of the silicon is dedicated to accelerating ML workloads.