I've only done a quick scan but I don't think this has been posted:-
https://twitter.com/p_ferragu/status/1295804869119533056
More generally Tesla has never said Dojo is needed for working Robo-taxis...
They probably have a prototype version of Dojo which speeds up training 10X, the initial production version of Dojo probably speeds up training 100X-1000X.
Faster training means a faster "march of nines" once FSD is "feature complete".
So I don't think they need a production version of Dojo to be "feature complete". or start the process of gaining regulatory approval. It maybe that they need a working version of Dojo to gain regulatory approval.
As far as we know HW3 is sufficient for FSD, or more accurately it is not close to 100% utilization.
HW4 - is smaller-cheaper-faster-better as per normal chip industry optimization, it makes sense to make the best possible version of the chip whenever industry R&D makes suitable gains.
Tesla doesn't need to actively drive chip R&D, others are doing that....
Tesla is possibly making their own chip for Dojo for the same reason they made HW3, industry standard R&D would not deliver the result they wanted on time...However once Tesla has the initial version of a chip, it seems something more like industry standard processes are used.
They may eventually be in chip design and fabrication in a big way especially if we consider Tesa/SpaceX/Neuralink sharing a chip R&D / fab facility.