Google's TPU and Groq's LPU are the only real, commercially viable ways to provide all the compute required for the inference people want.

Just having an ungodly amount of capex by blanketing the Midwest with datacenters full of GPUs is a disaster in slow motion.