Boise, Idaho, March 5, 2026, 14:42 MST
- Micron has begun sending out customer samples of its 256GB low-power memory module, targeting data center use.
- The company says its module pushes CPU-attached memory up to 2 terabytes for every eight-channel server CPU.
- A Futurum analyst sees the launch as indicative of a broader change underway in how AI server memory gets designed.
Micron Technology on Thursday announced it’s now sending out customer samples of its 256GB SOCAMM2 low-power server memory module. The chip is aimed squarely at AI data centers struggling with power and capacity constraints.
The product is targeting a less-than-glamorous issue—AI workloads keep pushing up the demand for memory near the CPU, yet data centers are getting squeezed by power limits, heat, and rack capacity.
Micron pointed out that modern AI setups lean hard on large language model LLM inference—which means running these models after they’re trained. As context windows expand and key-value (KV) caches stick around, memory requirements keep climbing.
SOCAMM2—system-on-chip attached memory module—links LPDRAM directly to server CPUs using a modular setup, diverging from the registered DIMMs (RDIMMs) that dominate current server memory configurations.
Micron claims its 256GB module, based on a monolithic 32Gb LPDDR5X design, uses just a third of the power and space of comparable RDIMM setups. The company also says the module offers a third more capacity than its earlier 192GB SOCAMM2 product.
According to the company, the module supports up to 2TB of LPDRAM for every eight-channel server CPU. In Micron’s own testing, it reduced “time to first token”—that early wait before a model replies—by 2.3x in select long-context inference scenarios. Micron Technology
Micron’s new “256GB SOCAMM2” offers a CPU-attached memory that cuts energy use for AI and high-performance computing, according to Raj Narasimhan, senior vice president over the cloud memory unit, in a company statement. Micron Technology
Nvidia’s Ian Finder says this higher-capacity, lower-power module targets “the next generation of AI CPUs,” with system designers now focused on squeezing out performance and efficiency “at every layer.” Micron Technology
Futurum analyst Brendan Burke called the announcement “significant not only for the capacity milestone itself,” noting it signals a broader shift—CPU-attached LPDRAM emerging as its own tier for newer AI workloads, he said. Futurum
Micron is collaborating with Nvidia on memory design and taking part in the development of the JEDEC SOCAMM2 specification, the company said. JEDEC is the industry standards group responsible for semiconductor and memory interface standards.
Micron slipped roughly 0.9% to $397.05 during late trading in the U.S.
Even so, Micron’s performance numbers come from its own labs, and actual customer deployments hinge on how quickly servers qualify the new modules and how fast platforms sign on. That window could open up, especially if field results lag or if buyers prefer familiar memory setups.
