Micron -- How the Business Works

Memory is an oligopoly with the highest barriers to entry in semiconductors. Three companies -- Samsung, SK Hynix, and Micron -- control roughly 92% of the global DRAM market. Building a new memory fab from scratch costs $20 billion or more and takes over three years. The IP required to manufacture at leading-edge nodes (1-gamma DRAM, HBM4, 200+ layer NAND) represents decades of cumulative process engineering that cannot be replicated. No new competitor has successfully entered DRAM manufacturing in over 20 years. This structural oligopoly means that when demand surges -- as it has with AI -- the supply response is inherently slow, creating extended periods of pricing power for incumbents.

HBM has transformed memory from commodity to differentiated product. High Bandwidth Memory is not a commodity -- it requires advanced packaging, through-silicon vias, and tight integration with GPU/accelerator designs. Each HBM stack consumes roughly 3x the DRAM wafer area of conventional modules, permanently constraining bit supply growth. Micron achieved an industry first with HBM4 volume production for NVIDIA Vera Rubin, with yields ramping faster than the prior generation. The HBM annualized run rate reached ~$8B by FQ4 2025, and Micron has sold out its entire CY2026 supply with committed price and volume. HBM4E (CY2027) introduces a customizable base logic die that further differentiates the product and increases margins. The HBM TAM is forecast to grow from ~$35B in CY2025 to ~$100B by CY2028 at a ~40% CAGR.

The cyclicality question is the central investment debate. Memory has never sustained peak margins for more than 4-6 quarters historically. From FY2022 peak ($8.35 non-GAAP EPS) to FY2023 trough (-$4.45 EPS) was a 152% swing. The bull case argues this time is structurally different: HBM constrains wafer supply, greenfield fabs take 3+ years, AI demand is secular not cyclical, and SCAs may provide margin floors. The bear case notes that every prior memory cycle has ended in oversupply, capex above $25B in FY2026 (stepping up meaningfully in FY2027) follows the classic pattern of peak-cycle overinvestment, and customers receiving only 50-65% of demand today could normalize quickly. Management insists supply-demand tightness persists "beyond calendar 2026," but the Street remains skeptical -- the stock trades 22% below its 52-week high despite record results.

SCAs and the "memory as strategic asset" thesis could change the game. The introduction of 5-year Strategic Customer Agreements is the most important business model development at Micron since the company began. If these agreements include meaningful downside margin floors and expand to a significant portion of revenue, they would structurally reduce earnings volatility and justify a permanent re-rating of the stock. Memory would evolve from a commodity subject to boom-bust cycles into a contracted, mission-critical input for AI infrastructure -- similar to how long-term power purchase agreements stabilized the utility sector. However, SCA terms remain confidential. Until investors can verify the downside protection provisions, the "structural shift" narrative will be met with healthy skepticism. The FQ3 2026 guide of $33.5B in revenue and $19.15 EPS -- exceeding full-year FY2024 revenue in a single quarter -- will test whether the market is ready to believe.