The hyperscale, cost-efficient compute protocol for the world's deep learning models.