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

Are we sure this will be the case? Perhaps the sweet spot for hardware that can train/run language models is the GPU already, especially with the years of head start Nvidia has?