It's entirely possible they've overspec'd the compute in current vehicles. Give yourself 10-20x the compute you think you'll need, because the cost relative to that of the vehicle is still small. Optimize in future generations.
With moore's law it would make more sense to just replace the processing in the future than try and future proof it.
I think you're grossly underestimating the cost of doing that to a fleet of vehicles sold with the promise of being self driving capable with the hardware package on board.
Change the promise to "we'll upgrade the hardware free over time".