They don’t need to if they can process a significant chunk of the queries on-device. Llama3-level inference works fine on M2-level chips today and the M4 is already in a mobile device.