Nvidia tensor cores are small systolic arrays. They'd have to throw out a lot of their ecosystem investments and backwards compatibility to make effective use of them as the main GPU compute, and there's really no need given how competitive their chips are right now.