Nvidia unveiled the Vera Rubin AI computing platform at CES 2026, claiming up to 10x lower inference token costs and faster training for MoE models.