I feel like they would not be too niche, many AI researchers I know complain about the amount of RAM on a single GPU, especially the ones they or their labs can afford.

Very often multi-GPU training is maxing out RAM on each GPU and isn't making full use of compute capacity.