Meta Platforms has partnered with Micron Technology to adapt LPDDR5X memory chips — originally designed for smartphones — for use in large-scale AI data centers, a project that could ease the industry's chronic shortage of high-bandwidth memory while slashing power consumption.
In a research paper published this month, Meta engineers demonstrated that LPDDR5X modules consume roughly one-third the power of conventional DDR5 server memory while delivering two to three times the throughput per watt. The chips also reduce latency in key AI inference benchmarks, including Spark SQL workloads that underpin Meta's advertising and recommendation systems.
The memory shortage has become one of the most acute bottlenecks in AI infrastructure. Nvidia's latest Blackwell GPUs and AMD's MI300 accelerators both require vast quantities of high-bandwidth memory, and demand from cloud providers has outstripped supply for more than a year. By turning to a memory standard already produced in enormous volumes for mobile phones, Meta and Micron hope to tap an underutilized supply chain and diversify away from the expensive HBM3E modules that dominate the AI server market.
The collaboration reflects a broader trend among hyperscalers to rethink every layer of the data-center stack as AI workloads consume an ever-larger share of global electricity. Meta has already committed tens of billions of dollars to expanding its computing footprint, and chief executive Mark Zuckerberg has said the company plans to double its AI capacity to fourteen gigawatts by the end of next year.
Industry analysts greeted the announcement with cautious optimism. While LPDDR5X offers compelling efficiency advantages, it lacks the error-correction features and module-density options that enterprise customers typically demand. Micron has said it is developing ruggedized packages specifically for data-center environments, with sampling expected in the first quarter of 2027.
If the approach scales, it could pressure traditional server-memory suppliers such as Samsung and SK Hynix to accelerate their own low-power product roadmaps. It might also open the door for smaller AI training clusters to achieve competitive performance without the premium price tag of HBM-equipped systems.
Image source: i.ibb.co