WHY THIS IS A STRUCTURAL SHIFT Several interlocking changes have reduced cyclicality and created a higher demand floor with longer visibility: AI-driven inelastic demand from hyperscalers and data centers: Generative and Agentic AI workloads (training + inference) require massive, continuous high-bandwidth memory (HBM). HBM consumes ~3–4× the wafer capacity of standard DRAM per bit due to complex stacking/packaging (TSV, advanced processes). Hyperscalers and neoclouds prioritize securing supply over price, signing multi-year long-term agreements (LTAs). Data centers are projected to absorb a very large share of memory output estimates of up to 70% of prduction. Memory content per AI server has risen sharply, and context/KV-cache growth in agentic systems further multiplies needs. This creates a persistent demand floor unlike prior cycles. Supply-side constraints Only three major players (Samsung, SK Hynix, Micron) dominate advanced DRAM/HBM. Prior downturns led to underinvestment; new fabs take 2–3+ years and are extremely capital-intensive. Capacity is prioritized toward high-margin HBM, crowding out conventional DRAM/NAND. Manufacturing complexity continues to rise. Meaningful new supply is not expected to fully catch demand until late 2027 or later in many forecasts. The AI data-center supercycle provides the near-term structural backbone, while physical AI (robots/humanoids, autonomous vehicles/EVs, drones, and space systems) extends the demand runway into the edge and real-world embodiment. This broadens the end-market base beyond traditional cyclical consumer electronics, supports higher utilization and pricing discipline, and makes memory a more strategic, infrastructure-like asset. Supply constraints and the physics of advanced memory production mean the imbalance is expected to persist for years even as capacity expands. LOOONG $MU $DRAM $SNDK $SKHY #memory #AIinfrastructure #AgenticAI