The shape of the problem

AI math is simple — there's just an absurd amount of it.

A model is billions of multiplications per token. None of them is hard; the challenge is doing them all at once. That's a parallel problem, and CPUs are built for the opposite.