AI model sizes are growing faster than memory bandwidth improves. GPT-4 class models require hundreds of gigabytes of high-bandwidth memory for inference. Training runs require even more. Each new GPU generation increases the memory requirement per accelerator — Blackwell uses 192GB of HBM3E per GPU, up from 80GB on Hopper.
The production of HBM (High Bandwidth Memory) is technically challenging: it involves stacking multiple DRAM dies vertically using through-silicon vias (TSVs), then bonding the stack to the GPU through advanced packaging. Yield rates are lower than standard DRAM, and capacity expansion takes 12-18 months. This creates a persistent supply deficit as AI demand grows faster than HBM capacity comes online.
SK Hynix was first to market with HBM3E and holds the largest market share. Their partnership with NVIDIA gives them a privileged position in allocation.
Micron (Closelook Sentinel Ticker) entered HBM later but is ramping aggressively. Micron's quarterly earnings calls provide the best public data on HBM pricing, demand, and allocation — making it the most useful real-time signal for AI memory demand.
Samsung has struggled with HBM yield rates and fallen behind. Their recovery trajectory is a key variable — if Samsung catches up, duopoly pricing power weakens. If they continue to lag, Micron and SK Hynix retain exceptional margins.
Memory is tracked through Layer 2 of the 6-Layer Model. When HBM pricing holds or increases quarter-over-quarter, it confirms AI demand strength. When pricing softens or inventory builds, it's an early warning that the AI CapEx cycle may be cooling — which feeds into the CapEx Cliff analysis.
Micron's earnings calls are the single most important data point for the memory layer. Closelook monitors HBM pricing, allocation splits, and capacity expansion timelines quarterly. When Micron's HBM gross margins expand, it confirms the constraint thesis.
Sentinel Tickers — Why Micron →Functional Index — Memory Layer →CapEx Cliff Question →