If the models are more efficient, the tasks that still need a server will get the same result at a lower cost. OpenAI can also pivot to building more local models and license them to device makers, if it wants.