Samsung ships 12-layer HBM4 to Nvidia in February; 8-layer variants planned for 2027

Samsung and SK hynix plan to ship 8-layer HBM4 memory to Nvidia in February. That puts both of South Korea’s memory heavyweights on the same delivery calendar for the world’s most demanding chip customer. What’s actually shipping High-bandwidth memory, or HBM, is DRAM stacked vertically and placed right beside a processor. The “layers” are the floors in that building. More floors means more capacity per stack, but also more heat and more chances for something to go wrong during manufacturing. HBM4 is the newest generation. Samsung began mass shipments of HBM4 to Nvidia in February 2026, aimed at the Vera Rubin AI accelerator. That early batch was mostly 12-layer stacks. Those chips reportedly hit speeds of up to 11.7 Gbps. The JEDEC industry standard calls for 8 Gbps, so Samsung’s parts cleared the bar with plenty of room to spare. Both companies are preparing to ramp up 8-layer HBM4 for Nvidia’s Rubin Ultra NVL systems, with volume shipments targeted around the second quarter of 2027. Why fewer layers can be the smarter move The research points to two drivers: thermal management and yield. Taller 12-layer stacks trap more heat and are harder to manufacture without defects. Yield i...