Maybe your point is the majority of the time that's not needed, fair, then maybe the approach is both?
This is circular reasoning. We don’t have latency-sensitive applications because sufficiently powerful local AI is not profitable to entertain such applications.
MooresLawIsDead and his spidey web / senses thinks it already popped: