It's not limited to just the mobile device. You could have a MacBook/mini/studio that is part of your local "cluster" and the inference runs across all of them and optimized based on power source.