Their biggest (m2 max in studio) are more like chiplets with interposers, which exponentially lowers the yield problems depending on how much you can split it up. Also larger chips can still set a threshold on gpu cores to disable due to defects and that kind of thing, where a mobile one might just throw out the chip with a much lower number of (or any?) defects.

Nvidia can make tons of tiers out of the same chip by just setting different thresholds on the number of usable cores, it isn't all just price discrimination (though sometimes I think they have been found to be fusing off much more than needed for the number of defects as a pure price discrimination play, or that might have been Intel with cache size).