If "LLMs" includes reasoning models, then you're already wrong in your first paragraph:

"something that is just MatMul with interspersed nonlinearities."