When you write software for a car, you always have to think about the impact of your code in the car, and pressing enter or crashing the sw is not the safe way to respond. So if the comms fails, you have to ask yourself what should the thing you're writing do and what the driver would see or think.
The problem with this architecture is that when you have a physical sensor, it is obvious when it fails and you can easily add contingencies. When you have a sensor other CAN or whatever, you are adding new failure modes: sensor OK but communication not reliable, or sensor info not yet received. You thus need more code to handle those new failure modes which are tricky for the average sw guy.
The hardware is simplified and the software becomes more sophisticated. In this case, the recurring costs are minimized (simplified wiring harnesses that are installed via automation) in favor of an increase in non-recurring costs (more up-front engineering time).
The failure modes of these kinds of designs are by now a solved problem. The fact that automated assembly of simpler wiring harnesses occurs might mean the hardware reliability goes up.
I expect because tesla is doing it in SV, they have enough software guys available to figure things out (and maybe frontload the diagnostics to make servicing things easier).
That's how technology works - make cheaper and less sophisticated hardware do more and better.