3 Reasons Why Micron and Sandisk Investors Can Ignore What History Says Will Happen

The article argues that AI-driven memory demand has not yet peaked, supporting a multi-year period of elevated utilization and margins for Micron and Sandisk. Nvidia estimates the five largest AI hyperscalers will spend $1.3T on data-center capex next year, up from $800B, with cumulative data-center spending projected at $3T-$4T by 2030. Micron's additional capacity is not expected online until mid-2027 or 2028, while Sandisk plans a $31B production facility, reinforcing the view that supply constraints and favorable memory pricing may persist.

Analysis

The key underwriting distinction is DRAM/HBM versus NAND: MU’s AI exposure is concentrated in high-bandwidth memory, where qualification cycles, packaging complexity and customer concentration create materially tighter effective supply. SNDK is more exposed to NAND, where capacity additions are easier to replicate and pricing remains more vulnerable to smartphone/PC inventory corrections. A broad “memory supercycle” narrative therefore risks assigning HBM-like scarcity economics to flash suppliers that may not earn them.

Near term, the relevant catalyst is quarterly contract-price progression and evidence that hyperscaler memory content per server is rising faster than bit supply. Over 1-3 months, MU can rerate if HBM mix, yields and gross-margin guidance continue to improve; NVDA’s platform demand is supportive but its own supply-chain leverage increasingly shifts toward memory and advanced packaging. Over 6-18 months, announced greenfield capacity becomes a risk rather than a benefit: memory projects tend to create synchronized supply, and a modest AI-capex digestion phase could sharply compress spot and contract pricing before fabs are fully utilized.

Consensus appears too willing to extrapolate aggregate data-center capex into memory revenue. The more important variable is memory dollars per incremental compute unit, which can fall if customers optimize model inference, deploy cheaper accelerators, or shift from training clusters to utilization-heavy workloads. The upside case is not simply sustained volumes; it requires disciplined industry wafer starts and HBM supply remaining constrained through the next capacity ramp.

Key Decisions for Investors

• Prefer MU over SNDK as a 3-6 month relative-value expression: long MU / short SNDK in equal dollar amounts. MU has the cleaner exposure to high-value HBM scarcity, while SNDK carries greater NAND pricing and new-capacity risk; reassess if NAND contract prices outperform DRAM for two consecutive monthly resets.
• Do not chase a broad semiconductor-beta trade through NVDA on this signal alone. Use NVDA only as a read-through: a data-center revenue beat without corresponding upward memory-content commentary would weaken the memory thesis rather than validate it.
• Add MU only following a quarterly guide that shows gross-margin expansion alongside stable HBM shipment/qualification commentary; target a 15-20% upside over 3-6 months versus a 8-10% stop if management signals lower HBM mix, yield delays, or weaker contract pricing.
• Monitor DRAM and NAND contract-price data, hyperscaler capex guidance, and Korean/Taiwanese producer wafer-start commentary. A sequential decline in DRAM pricing or a material capex reduction by major cloud buyers is the falsification trigger for memory longs and favors reducing exposure before new supply ramps.