DRAM makers will only meet 60% of demand by 2027, with new fabs prioritizing AI HBM over consumer memory.
Nikkei Asia reports that Samsung, SK Hynix, and Micron are ramping DRAM production, but new fab capacity won't come online until 2027-2028 at the earliest. Supply is projected to meet only 60% of demand by end of 2027, with SK Group's chairman warning shortages could persist until 2030. Production needs to grow 12% annually to meet demand, but only 7.5% growth is planned. New fabs will primarily produce high-bandwidth memory (HBM) for AI data centers, not general-purpose DRAM for consumer devices.
HBM prioritization means AI inference and training infrastructure stays expensive longer than most roadmaps assumed. Cloud providers absorbing higher HBM costs will pass them downstream — expect GPU instance pricing to remain elevated or increase through 2027. On-device AI features requiring high RAM (e.g., local LLM inference on phones/laptops) will face hardware bottlenecks for years.
Audit your model serving stack this week: benchmark memory footprint of your current inference setup and test quantized variants (INT4/INT8) using llama.cpp or vLLM to identify if you can cut GPU memory requirements by 30–50% without meaningful accuracy loss.
Clone and build llama.cpp: git clone https://github.com/ggerganov/llama.cpp && cd llama.cpp && make
Tags