This suggests two interesting things to me:
1) Google/Waymo, the "software and ML can solve this problem" company is admitting that hardware is important to solving the problem.
2) They are likely much further away from L5 or even public release than some believe given that they are now completely revamping their sensor suite.
Most companies want lidar to get cheaper not better. Very rarely has the solution to a perception problem been "better sensors" as opposed to "more ml".
If Waymo can revamp the car's sensor suite this late in the process, all the while ramping up their fleets and consumer services, it suggests to me that their software is sufficiently powerful and flexible not to need massive and costly retraining; basically I think they have "solved" the problem on the software side (given appropriate constraints such as good weather, excellent mapping, etc.) and are now working on rationalizing the hardware costs.
Especially when the graph shows that they've completed 6M of those 10M in the last year! (https://www.theatlas.com/charts/S1uAcpj97)