They’re testing. Rides are only open to the public in Phoenix, which has been the case since 2017. If the model was easily scalable, every city in the world would have it by now.

ADAS levels 2 to 5 require exponential levels of investment, not linear in any way. Even the L2+ systems are pretty disappointing. GM had to scan the entire highway network to get their system to work, and Tesla’s and Comma.AI don’t work that well at all. Mercedes has a L3 system for the easiest use case (stop and go traffic on the highway). Waymo’s L4 seems like it works, but the fact that it’s in a controlled environment and they haven’t scaled the service after years of operation is pretty telling about the state of the technology. We’ve been “18 months away” from the promised land for a decade. As it turns out, AI is way harder than anyone thought.