Pretty much everything this comment said. Nvidia had nothing to do with the first few uses of CUDA for machine learning.

It could be done on FPGAs and stuff but that wasn't easily accessible to the masses.