Boise, Idaho, March 5, 2026, 14:42 MST
- Micron said it has started shipping customer samples of a 256GB low-power memory module for data centers.
- The company said the module can scale CPU-attached memory to 2 terabytes per eight-channel server CPU.
- A Futurum analyst said the launch points to a wider shift in AI server memory design.
Micron Technology said on Thursday it has begun shipping customer samples of its 256GB SOCAMM2 low-power server memory module, targeting AI data centers that are running into power and capacity limits. 1
The product aims at a messy, practical problem: as AI workloads grow, servers need more memory close to the CPU, but data centers are already constrained by electricity, heat and rack space. 1
Micron said modern AI systems increasingly depend on large language model (LLM) inference — running models after training — where larger “context windows” and persistent key-value (KV) caches can push memory demand higher. 1
SOCAMM2, short for system-on-chip attached memory module, is designed to connect low-power DRAM (LPDRAM) to server CPUs in a modular format, a step away from standard RDIMMs, or registered dual in-line memory modules, used broadly in servers. 1
Micron said the 256GB module is built on a monolithic 32Gb LPDDR5X design and delivers one-third the power consumption and one-third the footprint of equivalent RDIMM configurations, while providing one-third more capacity than its prior 192GB SOCAMM2 module. 1
The company said the module can enable up to 2TB of LPDRAM per eight-channel server CPU and cut “time to first token” — the delay before a model starts responding — by 2.3 times in certain long-context inference tests, based on Micron’s internal measurements. 1
“Micron’s 256GB SOCAMM2” provides a more energy-efficient CPU-attached memory option for AI and high-performance computing, Raj Narasimhan, a senior vice president who leads Micron’s cloud memory business, said in the company statement. 1
Nvidia executive Ian Finder said the higher-capacity, lower-power module helps “the next generation of AI CPUs,” as system designers try to optimize performance and efficiency “at every layer.” 1
Futurum analyst Brendan Burke wrote that the announcement is “significant not only for the capacity milestone itself,” pointing to what he described as a broader move toward CPU-attached LPDRAM as a distinct tier for newer AI workloads. 2
Micron said it is working with Nvidia on memory design and is involved in defining the JEDEC SOCAMM2 specification, referring to the industry standards group that publishes semiconductor and memory interface standards. 1
Micron shares were down about 0.9% at $397.05 in late U.S. trading.
Still, the performance claims are based on Micron’s internal testing and customer rollouts depend on qualification timelines and server platform adoption — a gap that can widen if real-world deployments fall short or if buyers stick with more established memory configurations. 1
