To me it's an umbrella for all of the projects that fell out of an initial realization (by Jeff Dean and his circle) that Google's infrastructure was large enough to put into practice a class of older ML ideas that had spent years being thought infeasible. They developed the means of training huge models, then useful models that resulted from that, and follow-on project like TPUs.