Overview
- Memory for AI workloads is consuming a dominant share of capacity, with reports that high-bandwidth memory is sold out for the rest of 2026 and AI servers taking most available DRAM.
- Micron says next-generation production nodes will not reach volume until after mid-2027, which the company warns will keep DRAM and NAND markets tight beyond 2027.
- Micron CEO Sanjay Mehrotra forecasts a longer-term structural wave of demand from humanoid robots that he estimates could need roughly ten times the memory of today’s L2+ vehicles, and he presented this as a forecast rather than a confirmed product demand level.
- Memory makers are enjoying strong pricing power and profits while device makers face higher component costs that squeeze margins, prompting big buyers such as Tesla to propose very large, multi‑decade fixes like the reported $119–$122 billion Terafab plan whose effects would not appear before 2030.
- HBM is technically costly to produce because it stacks dies and uses through‑silicon vias, consuming about four times the wafer area per gigabyte of standard DRAM, and global supply is concentrated in Micron, Samsung, and SK Hynix with fab expansion slowed by scarce equipment such as ASML EUV tools.