They have super sustainable revenue. They are deadly supply constrained on compute, and have a really difficult balancing act over the next year or two in which they have to trade off spending that limited compute on model training so that they can stay ahead, while leaving enough of it available for customers that they can keep growing number of customers.