Well you can definitely bet it will be faster not slower than the first two, especially given the basic (i.e. shared/city-agnostic) engineering required and the policy component, which will get easier and easier with each city as risk aversion turns to FOMO.

I would not bet on bureaucracy speeding up anytime anywhere soon.

It also seems like a process that should scale horizontally pretty predicably.

My argument is based on theory. We know that a lot of the learning is facing unusual situations (trucks delivering traffic lights, etc) that can happen in many places. And we have some idea of how long the mapping takes.