I believe that's also why there are so many options for turning fully trained TensorFlow graphs into C++ code. You use the expensive GPU for building the AI, t

I believe that's also why there are so many options for turning fully trained TensorFlow graphs into C++ code. You use the expensive GPU for building the AI, then dumb it down for mass-deployment on cheap CPUs.