That's assuming Apple isn't training their models on Nvidia hardware.

Inference is on Apples own silicon though.