A typical computer used by a typical person has only a fraction of energy use of GPU farms utilized to train DNN models. We're not at the point where you can pick an existing DNN off the shelf and just use it, you have train (at least partially) a new one for each new task, often many times.