I think the particular focus on FSD has to do with what I heard on the Adam Jonas (of MS) institutional client call: Tesla is *far* ahead of everyone else and maybe has an insurmountable lead. It's a very valuable part of the company in the eyes of outsiders.
BTW, as a software developer (not in AI/NN), I think it'll be 20-30 years before there's even a reasonable FSD capability. The basics of accurate object detection and tracking aren't there, other than in ideal situations, and we all know that just doesn't cut it in the real world. In my area there was a rush to NN for analysis of real time data about a decade ago. It petered out when the researchers concluded that they had no idea what the NN was latching onto during its detection phase. Basically, they couldn't trust it so they've gone back to more deterministic detection and analysis methods. Anyway, if I were to make a WAG on what it would take, hardware-wise, for minimal FSD, it would be a Tesla NN system on each of the car's cameras - probably binocular camera systems and generate some sort of complex intermediate output. That output would feed into the "brain" NN system in order to make sense of the car's surroundings.