Micron Starts High Volume Production of HBM4, PCIe Gen6 SSDs and SOCAMM2 for NVIDIA Vera Rubin

Micron has officially entered high volume production with a trio of products designed to strengthen NVIDIA’s Vera Rubin ecosystem, including its HBM4 36GB 12H, the Micron 9650 PCIe Gen6 data center SSD, and 192GB SOCAMM2 memory modules. In its official announcement, Micron said volume shipment of the HBM4 stack began in the first quarter of calendar 2026, and that the product is specifically designed for the NVIDIA Vera Rubin platform.

The headline memory product is clearly Micron’s HBM4 36GB 12H, which the company says reaches over 11 Gb/s pin speeds and delivers greater than 2.8 TB/s of bandwidth. Micron also says this represents a 2.3x bandwidth increase and more than 20% better power efficiency compared with its HBM3E. For the Rubin generation, this matters because memory bandwidth is increasingly one of the most decisive constraints in large AI training and inference clusters, especially as model size and context demands continue to scale.

Micron is also using this announcement to signal that it is already thinking beyond the first HBM4 configuration. The company says it has shipped customer samples of HBM4 48GB 16H, demonstrating 16 die stacking capability and delivering a 33% capacity increase per HBM placement compared with the 36GB 12H version. That is a notable roadmap signal, because it suggests Micron is not only supporting Rubin class systems now, but is also preparing for the even higher density memory demands that future AI accelerators will require.

On the storage side, Micron says the 9650 is the industry’s first PCIe Gen6 SSD in high volume production. The company positions it as a data center SSD optimized for NVIDIA BlueField 4 STX reference architecture and agentic AI workloads, with up to 28 GB/s sequential read throughput and 5.5 million random read IOPS. Micron also claims the drive offers up to 2x the read performance of Gen5 along with 100% higher performance per watt, making it one of the more important non GPU announcements tied to the Rubin platform story.

The third pillar is SOCAMM2, where Micron says 192GB modules are now in high volume production as part of a broader portfolio ranging from 48GB to 256GB. These modules are designed for NVIDIA Vera Rubin NVL72 systems and standalone NVIDIA Vera CPU platforms, enabling up to 2TB of memory and 1.2 TB/s of bandwidth per CPU. That makes SOCAMM2 a meaningful piece of the Rubin era infrastructure puzzle, particularly for CPU heavy AI and HPC workloads where capacity, power efficiency and density all need to move together.

Strategically, this announcement shows how the Vera Rubin platform is becoming much more than a GPU story. Micron is effectively positioning itself as a foundational memory and storage partner across the Rubin stack, spanning accelerator memory, CPU side modular memory, and data center storage. In the current AI market, where hyperscalers and AI cloud operators are racing to optimize every layer of throughput and power efficiency, that kind of vertically aligned component support can be just as important as the compute silicon itself. Market coverage following the announcement also noted the significance of Micron reaching HBM4 production at this stage, since high bandwidth memory supply remains one of the most closely watched parts of the AI infrastructure cycle.

For NVIDIA, Micron’s announcement helps validate that the Vera Rubin ecosystem is moving from roadmap language into real supply side execution. For Micron, it is a strong statement that the company intends to compete aggressively at every critical tier of AI memory and storage, from HBM to SOCAMM2 to Gen6 enterprise SSDs. As AI infrastructure moves deeper into the Rubin generation, this kind of ecosystem readiness may prove just as important as the accelerators themselves.

Do you think memory and storage will become just as important as GPU performance in deciding who leads the next AI data center cycle?