They don't need to worry about whether local models can do it. All they have to do is make difficult or remove the power to run a local model in the first place. By, for example, taking your RAM and putting it in a datacenter.