Micron said it is mass-producing and shipping sixth-generation high-bandwidth memory (HBM4). The company directly refuted recent chatter about being dropped from "Vera Rubin" supplies, timing the announcement with the opening of Nvidia's annual developer conference (GTC 2026).
Micron said in a press release on the 16th (local time) that it has been mass-producing and shipping 36-gigabyte (GB) 12-high HBM4 since the first quarter of this year. Micron said the product was "designed for Nvidia Vera Rubin." Some in the semiconductor industry had said HBM4 from Samsung Electronics and SK hynix would go into Nvidia's Vera Rubin AI chip launching this year.
Micron said the pin (data transfer lane) speed of its 36GB 12-high HBM4 exceeds 11 gigabits per second (Gb/s). It also delivers bandwidth exceeding 2.8 terabytes per second (TB/s). Compared with fifth-generation HBM (HBM3E), it offers 2.3 times the bandwidth and 20% better power efficiency, the company said. The 36GB 12-high HBM4 specifications Micron disclosed meet the performance reportedly requested by Nvidia for its memory suppliers.
Micron also said it has shipped samples of a 48GB 16-high HBM4 to customers. Compared with the 36GB 12-high HBM4, it boosts capacity by 33% to address new demand.
Sumit Sadana, Micron's chief business officer (CBO and senior vice president), said, "We are working closely with Nvidia to architect from the outset so compute and memory can scale together," adding, "Micron's HBM4 delivers unprecedented bandwidth, capacity and power efficiency."
Micron also said it is in volume production of SOCAMM2, a low-power memory module for AI servers. The company said this product was also designed for Nvidia's AI Superchip and rack system "Vera Rubin NVL72" and the central processing unit "Vera." Micron SOCAMM2 provides up to 2TB of memory per CPU and 1.2TB/s of bandwidth. The company also said it is mass-producing the Micron 9650, a PCI Express 6th generation (PCIe Gen6) data center solid-state drive (SSD).