Micron launches HBM4 production for Nvidia Vera Rubin with 2.3x bandwidth boost and 20% efficiency gain

Micron Accelerates HBM4 Production for Next-Generation AI Infrastructure

Micron has entered high-volume production of its HBM4 36GB 12H memory, marking a significant milestone in AI infrastructure development as the industry prepares for Nvidia's Vera Rubin platform. Announced at GTC 2026 in San Jose, California, the memory giant simultaneously confirmed it has achieved mass production across three critical product categories: HBM4 memory, SOCAMM2 modules, and PCIe Gen6 SSDs, making it the first memory supplier to bring all three products to volume shipment for the Nvidia Vera Rubin platform ecosystem at the same time 1

1

3

3

.

The HBM4 36GB 12H configuration runs at over 11 Gb/s pin speeds, delivering memory bandwidth greater than 2.8 TB/s. Compared to Micron's HBM3E at the same 36GB 12H configuration, this represents a 2.3 times bandwidth improvement alongside more than 20% improvement in power efficiency, according to Micron's internal power calculator data 1

1

. For AI accelerators operating in dense server environments, this combination of bandwidth and power efficiency directly affects how much useful work can be sustained without hitting thermal or power constraints 2

2

.

Expanding Memory Capacity with 48GB 16H Stack Samples

Micron has already shipped samples of a 48GB 16H stack to customers, adding four additional die layers to the standard configuration. This higher-capacity variant increases memory capacity per HBM placement by approximately 33% over the 36GB 12H product, a milestone that points toward denser configurations in future AI accelerator generations 1

1

2

2

. The 48GB 16H stack signals Micron's roadmap for scaling memory capacity as AI models continue to grow in size and complexity.

SOCAMM2 Modules Address System Memory Requirements

At the system memory level, Micron confirmed that its 192GB SOCAMM2 modules are now in high-volume production. These modules target Vera Rubin NVL72 systems and standalone Vera CPU platforms, supporting up to 2TB of memory capacity and 1.2 TB/s of bandwidth per CPU 1

1

3

3

. The broader SOCAMM2 family spans capacities from 48GB to 256GB, giving server builders flexibility depending on workload class and platform configuration. In AI computing solutions where host-side bandwidth and capacity are becoming increasingly important alongside accelerator-attached HBM pools, SOCAMM2 modules help address growing memory demands 2

2

.

PCIe Gen6 SSD Doubles Storage Performance

Micron also updated the storage side of its roadmap with the 9650 PCIe Gen6 SSD, which entered mass production last month. Designed for Nvidia's BlueField-4 STX reference architecture, the drive supports up to 28 GB/s sequential read throughput and 5.5 million random read IOPS, doubling PCIe 5.0 read performance while delivering 100% higher performance per watt 1

1

3

3

. The drive targets AI inference, training, and agentic workloads in liquid-cooled environments, addressing data center deployments where storage throughput must scale alongside faster compute and larger memory footprints 2

2

.

Strategic Integration Across the Memory and Storage Stack

What distinguishes this announcement is Micron's coordinated approach to subsystem design. HBM4 targets the accelerator memory tier, SOCAMM2 modules address CPU-attached system memory, and the 9650 extends bandwidth at the storage layer. This represents a more complete infrastructure strategy than competing in a single niche, as AI servers increasingly require balanced design to avoid bottlenecks that can appear anywhere from on-package memory to host memory to storage 2

2

.

"The next era of AI will be defined by tightly integrated platforms developed through joint engineering innovations across the ecosystem. Our close collaboration with NVIDIA ensures that compute and memory are designed to scale together from day one," said Sumit Sadana, executive vice president and chief business officer at Micron Technology 1

1

. As Vera Rubin-era platforms approach deployment through 2026, Micron's strategy positions the company across the full memory and storage stack, forming a core foundation for next-generation AI servers entering service this year 2

2

.