Edge inference memory (EIM) as a new category, and humanoid robot DRAM demand that could grow 20-fold by 2030
Micron is extending its moat from data centre HBM to physical AI at the edge. At SEDEX 2026, vice president Li Zhe disclosed that the company is co-developing a new memory category with major customers: Edge Inference Memory (EIM), which stacks low-power LPDDR on a logic die.
Read more
It aims to combine the power consumption of LPDDR with the low latency and high bandwidth of HBM. The target applications are physical AI uses such as humanoid robots, autonomous driving and AR/VR, with volume production in the second half of 2028 in step with the arrival of the LPDDR6 standard. Co-development with customers continues the custom path of HBM4E (what Micron sells is no longer a standard part). On the demand side, Counterpoint forecasts that DRAM per robot will rise from about 19GB in 2026 to about 39GB in 2030, and with shipments ramping, overall DRAM demand is expected to grow about 20-fold by 2030. CEO Mehrotra set the tone last quarter, saying humanoid robots will need 200GB or more of memory plus terabyte-class storage, comparable to AI servers. EIM is the product that stakes out a position at the inflection point where edge AI memory content rises systematically.
Key evidence
- Micron vice president Li Zhe disclosed the new EIM category at SEDEX 2026 (the Korea semiconductor exhibition)
- Technical approach: low-power LPDDR stacked on a logic die, combining LPDDR's power consumption with HBM's low latency and high bandwidth
- Designed for battery-powered devices, it addresses the pain points of HBM's power draw and the insufficient bandwidth of pure LPDDR
- Volume production is targeted for the second half of 2028, in step with the arrival of the LPDDR6 standard
- It differs from HBM by focusing on power saving, and from PIM in that it does not add compute units inside the memory
- The co-development model with customers continues the custom path set by HBM4E
Show 4 more
- Counterpoint forecast: DRAM capacity per robot of about 19GB in 2026, rising to about 39GB in 2030
- Combined with rising shipments, overall DRAM demand is expected to grow about 20-fold by 2030
- CEO Mehrotra on last quarter's earnings call: humanoid robots will need 200GB or more of memory plus terabyte-class storage, comparable to AI servers
- Current robot AI models have only 1 billion to 2 billion parameters, so bandwidth requirements are not high (LPDDR and NAND dominate today)
Catalysts to watch
- The pace of EIM volume production in the second half of 2028
- When the LPDDR6 standard arrives
- Disclosure of the identity of the major customers co-developing it (large players in humanoid robots, autonomous driving and AR/VR)
- The ramp in humanoid robot shipments
- The inflection point at which a jump in edge AI model parameter counts drives memory content higher
Risks
- The 2028 volume production timeline for EIM could slip
- If commercialisation of physical AI use cases is slower than expected, demand will take longer to materialise
- Samsung and SK Hynix catching up in similar custom memory
- A delay or split in the LPDDR6 standard would hold back EIM
- The 20-fold DRAM demand forecast depends on assumptions about humanoid robot shipments, and long-range forecasts are highly uncertain
Sources (3)
- Micron vice president Li Zhe's public disclosure of EIM at SEDEX 2026
- Counterpoint forecast of DRAM capacity per device and 2030 demand
- Micron CEO Mehrotra's comments on last quarter's earnings call on humanoid robot memory demand
A brief is a lead for our research, not a verified conclusion or advice.