Training and inference

This is the first thing you have to know about supervised learning:

training is when you learn model parameters from input. This literally means learning the best value we can for a bunch of number input numbers of the model. This can easily be on the hundreds of thousands.
inference is when we take a trained model (i.e. with the parameters determined), and apply it to new inputs

Both of those already have hardware acceleration available as of the 2010s.