Normally I agree with you, but I don't think that's the case this year at least. Then on-device ML chip (Tensor G3) is a legit difference-maker. A lot of the ML features they are rolling out wouldn't be possible on older hardware. You could do the non-realtime portions in the cloud, but the realtime audio cleanup wouldn't work for example (way too much latency). Also, personally I much prefer on-device. I frequently have a spotty data connection so it's very disruptive when stuff relies on cloud connectivity.