Eventually we will not only need on-device ML assistants but even more so then, to own the whole stack from hardware bottom up.