“NVIDIA's Rubin platform is designed to train Mixture-of-Experts (MoE) models with one-fourth the number of GPUs and reduce inference token costs by up to 10x compared to the Blackwell platform.”