Interpretability research is a thing that is both useful and academia would be very good at it.
Also you probably don't need moleFLOP computers to do it.
a Schelling point for those who seek one