i'm guessing that the data is just plain too different (probably no more mobileye sensors, for example) and not worth adapting to the older ML models (which itself would require extensive testing) when the new system is going to end up with different ML models anyways.
[edit] Tesla previously hired Jim Keller (chip designer) into the autopilot team. considering the kinds of things he may be working on, i'd be surprised if the differences in either the sensors or GPUs aren't significant.