Industry sources say NVIDIA is considering reducing the high-bandwidth memory (HBM) stack on its next-generation Rubin Ultra AI accelerator from a planned 12 layers to eight. With rising memory costs, NVIDIA is prioritizing bandwidth-per-dollar and o

2026-09-10

Industry sources say NVIDIA is considering reducing the high-bandwidth memory (HBM) stack on its next-generation Rubin Ultra AI accelerator from a planned 12 layers to eight. With rising memory costs, NVIDIA is prioritizing bandwidth-per-dollar and overall system economics.