[daviddumenil]

I think it was more a case that Google had a large existing investment in CPU-based compute.

A lot of Dean's initial work was getting models effectively training on that kind of hardware using distributed SGD.