If you bake the model onto the chip itself, which is what should be happening for local LLMs once a good enough one is trained eventually, you’ll be looking at orders of magnitude reduction in power consumption at constant inference speed.