Micron rides the AI boom - The Economist

Micron's AI memory positioning is fundamentally a race against a 12-18 month HBM3E production window. SK Hynix has already locked in over 50% share through its early NVIDIA partnership. Micron's genuine breakthrough lies not in training-side HBM but in inference-side demand — as AI shifts from large-model training to enterprise deployment, memory bandwidth requirements grow non-linearly, and Micron's DDR5/LPDDR5X capacity elasticity aligns precisely with this demand structure. On the technical front, HBM4's TSV stack jumps to 12-16 layers, making yield the existential variable. If geopolitical pressure continues to constrain Micron's advanced packaging operations, HBM4 production will be forced to migrate to U.S. soil, deterministically amplifying cost disadvantages. This is not a risk scenario; it is the established trajectory. Competitively, Samsung's repeated HBM3E qualification delays mean it poses no near-term threat. The critical variable: whether Micron secures exclusive HBM supply for AMD's MI400 by Q3 2025. Success would reframe its valuation from cyclical to growth — a qualitative, not quantitative, shift. The 12-24 month tail effect: HBM will consume 15-20% of conventional DRAM capacity, potentially triggering structural shortages by 2026. Micron's strategic bet is "AI inference memory" over "AI training memory" — if this thesis holds, its revenue mix undergoes an irreversible structural transformation.