Makes sense I guess. The "listen for the trigger" bit is done locally on those Motos (I've got the 2014 variant) but really all it needs to do is keep an "ear" open for that one phrase so it's probably got a better margin, only helped by local processing. The bit of audio that comprises the query has to get compressed and interpreted on the server side (not to mention the fact that it consists of a much wider variety of possible content to be parsed) so that can't help. I guess a dedicated chip that's just constantly asking "did anyone say 'OK Google?" has an easier job than a remote process asking "wtf does that fuzzy bit of compressed audio mean?"
(obviously being silly and oversimplifying but same idea)
(obviously being silly and oversimplifying but same idea)