Samsung Electronics has secured a key partner role supplying NVHBM, NVIDIA’s custom high-bandwidth memory, and is developing an eight-layer HBM4E stack at 17 to 18 gigabits per second, about 20% faster than its 14.4 Gbps initial samples. NVHBM moves the memory controller to the HBM base die, targeting 30% higher bandwidth than HBM4E, up to 15% lower power use and more GPU space for AI compute cores. The shorter stack eases manufacturing versus 12- and 16-layer plans and is meant to improve volume during a memory shortage that NVIDIA says could limit growth through fiscal 2028. NVHBM is likely to start with Rubin Ultra, expanding GPU-to-GPU links from 72 to as many as 576, and will be co-developed with Amazon’s Annapurna Labs for multi-supplier production. NVIDIA’s second-quarter revenue hit $96.221 billion as memory-related supply commitments jumped to $279 billion, while consensus operating-profit estimates for Samsung and SK Hynix have been revised sharply higher amid sustained HBM pricing power.