> Nvidia will no doubt sell custom Groq-like chips for inference now.

For $0bn they could have sold an Nvidia-like chip for inference.