Micron Develops Edge Inference Memory for Physical AI, Targeting Humanoid Robots and Autonomous Driving

Micron Technology (MU) is developing Edge Inference Memory (EIM) designed for physical AI applications, targeting autonomous driving and humanoid robots, Vice President Lee Cheol said Monday at SEDEX 2026 in Seoul. Industry expects mass production in the second half of 2028. Micron is working to apply EIM technology to LPDDR6 memory to meet physical AI's low-power DRAM requirements, and is discussing products with major customers. The technology stacks low-power DRAM atop logic chips to process AI inference data at high speed, though the compute architecture has not been fully disclosed. Lee said high-bandwidth memory (HBM) will eventually be used in humanoid robots. Current robot AI models use roughly 1 billion to 2 billion parameters, keeping bandwidth demand modest, but advanced vision-language models will exceed LPDDR's capability. Micron estimates future high-performance humanoid robots will require about 200GB of DRAM and 4TB of storage, versus 12GB and 256GB for current service robots. Counterpoint Research projects global humanoid robot DRAM demand will grow more than 20-fold, from about 8,300 terabits this year to roughly 172,000 terabits by 2030.