Micron's CEO Says Even His Own Customers Couldn't See This Coming. Now He's Spending $200 Billion

Micron CEO Sanjay Mehrotra said AI-memory demand exceeded even customers’ forecasts, while the company reported Q3 FY2026 revenue of $41.46B and an 80.4% operating margin. Micron announced a long-term $200B U.S. investment plan to add supply, with fiscal 2026 CapEx expected above $25B and construction-related CapEx rising by more than $10B YoY in FY2027, amid a structural DRAM supply gap estimated at the highest ever. Management guided Q4 revenue to $50.0B (±$1B) and non-GAAP EPS to $31 (±$1), with gross margin around 86% and HBM4 already in high-volume shipments to a lead AI accelerator customer.

Analysis

Micron’s real edge is not “AI exposure”; it is that memory has temporarily stopped behaving like a commodity and started behaving like a bottlenecked infrastructure layer. That changes the earnings mix for the whole AI stack: MU captures scarce-margin economics now, while downstream buyers like NVDA/AMD and hyperscalers absorb a higher HBM bill that can quietly pressure gross margin or force design tradeoffs. The second-order beneficiaries are the equipment names feeding the capex supercycle — AMAT, LRCX, KLAC, and to a lesser extent ASML — because this is not a one-quarter inventory snapback; it is a multi-year capacity race.

The market is likely right to re-rate the durability of MU’s earnings power, but it may still be underestimating how long the constraint can persist because new fabs do not solve demand over the next 2-4 quarters. The main near-term risk is not demand collapse; it is a narrative shock from any sign that customer commitments are softer than advertised, or that HBM intensity per accelerator stabilizes faster than expected. Over 6-18 months, the bigger reversal risk is supply response: if competitors catch up on DRAM/HBM or if hyperscalers redesign around memory efficiency, MU’s margin peak could arrive before the revenue peak.

Consensus is probably still too anchored to the old “memory is cyclical, therefore the multiple must stay low” framework. That misses the contract lock-in and the strategic value of being the default supplier in a constrained market, which argues for a higher quality-of-earnings multiple even if the absolute growth rate normalizes. But after a 700%+ run, the stock can still compress on sentiment even while fundamentals improve, so the cleanest expression is not chasing outright but buying duration on pullbacks and isolating the relative winner versus the broader semi complex.

Key Decisions for Investors

• Long MU on 2-4 week weakness; use a staged entry rather than chasing the gap. Base case is continued earnings beats and backlog visibility over the next 1-3 quarters; thesis is invalidated if management cuts FY2026/FY2027 gross margin or revenue guidance by >10% or if HBM supply commentary turns incremental rather than tight.
• Express the capex supercycle with a basket long in AMAT/LRCX/KLAC vs. short SOXX. The setup is that memory buildout converts directly into equipment orders and services revenue over the next 12-24 months, while the ETF remains exposed to broader semi multiple compression.
• Pair trade: long MU / short NVDA into the next earnings window. The risk/reward is attractive if HBM scarcity persists, because MU monetizes the bottleneck while NVDA bears the input-cost squeeze; this breaks if NVDA demonstrates faster-than-expected memory cost pass-through or if accelerator demand re-accelerates enough to absorb margin pressure.
• Use Jan-2027 MU call spreads instead of outright common for a cleaner 12-18 month expression. This limits downside if the market starts discounting a 2027 supply response while preserving upside if strategic customer agreements keep pricing power elevated.
• Set a watch item on HBM pricing and any evidence of competitor capacity additions from SK Hynix/Samsung. If HBM spot or contract pricing rolls over before the next two quarterly reports, trim MU and rotate into equipment names.