Completely right. Applying a neural network is much faster than training one. The main trick here is fitting the trained model into cache (or smaller) so that the matrix multiplies are fast.