Vertical integration has many benefits. It’s the same reason Apple started making their own chips and ditched intel. The SOCs will be cheaper and better optimized for Tesla’s use case.

Every Nvidia GPU I’m aware of ships with both CUDA cores and Tensor cores. For a pure ML application those CUDA cores are almost useless. Tesla would be paying Nvidia’s premium for transistors that aren’t optimized for ML inference.

A Tesla designed chip can be 100% dedicated to tensor multiplication. You’re paying less per transistor and every transistor is utilized to the fullest extent.