Samsung Electronics (005930.KS) has secured an early position as a key supplier partner for NVHBM, Nvidia’s custom high-bandwidth memory, and is developing an eight-layer version of HBM4E, described as the seventh-generation product, to meet Nvidia’s requirements. The eight-layer design is shorter than the 12-layer and 16-layer versions Samsung originally prepared. Nvidia has specified speeds of 17 to 18 gigabits per second, about 20% above the 14.4 Gbps achieved by Samsung’s initial HBM4E samples. HBM (high-bandwidth memory) increases capacity by stacking DRAM layers, but taller stacks raise thinning, precision-stacking, manufacturing, and yield challenges. Nvidia’s lower-layer approach instead prioritizes speed and could make higher-volume supply easier amid a memory shortage. NVHBM is optimized for Nvidia’s proprietary NVLink standard, which was unveiled a day earlier, and is likely to be used starting with Rubin Ultra, the next-generation AI GPU due for release next year. Rubin Ultra expands GPU-to-GPU connections eightfold, from 72 to as many as 576, allowing faster HBM-equipped GPUs to be linked into a single unit even if memory capacity per GPU declines. Samsung is more competitive than SK hynix (000660.KS) or Micron in custom HBM because it can design and produce both DRAM and logic-based base dies, handling the work end to end.