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

The article argues that AI-driven memory-chip demand has not peaked, with hyperscaler data-center capital expenditure projected by Nvidia to rise to $1.3T next year from $800B and reach $3T-$4T by 2030. Micron's additional capacity is not expected online until mid-2027 or 2028, while Sandisk is investing $31B in new production, supporting the view that supply constraints and elevated memory margins could persist for years. The author views Micron and Sandisk as attractive despite historically cyclical memory markets and low forward earnings multiples.

Analysis

The investable distinction is not “memory” broadly but HBM/DRAM versus NAND. MU has materially higher AI sensitivity through HBM and server DRAM mix, where qualification cycles, packaging complexity, and customer-specific supply commitments can sustain pricing longer; SNDK is more exposed to NAND, where technology transitions are faster, inventory corrections are historically sharper, and Chinese supply remains a greater medium-term swing factor. A rising AI capex envelope therefore supports MU’s earnings durability more directly than it supports SNDK’s multiple.

New fab announcements should be read as a future-cycle risk rather than confirmation of durable scarcity. The market will capitalize 2027-28 supply additions well before output arrives; if hyperscaler capex shifts from training clusters toward lower-memory inference architectures, memory pricing could soften while fixed-cost absorption deteriorates. The key 1-3 month catalysts are HBM allocation commentary, DRAM contract-price revisions, and MU’s gross-margin/bit-shipment outlook; the 6-18 month risk is simultaneous capacity ramping across global DRAM and NAND producers.

Consensus appears too willing to extrapolate current spot tightness into all memory categories. The more defensible thesis is that AI raises the DRAM profit floor, not that it eliminates cyclicality: higher bit content per accelerator is real, but hyperscalers retain purchasing power and will force pricing concessions once qualified supply broadens. NVDA remains a useful read-through, but a GPU shipment beat without upward HBM-content or memory-supply commentary would not validate a broad memory long.

Falsification for the MU-over-SNDK view would be two consecutive quarters of weakening HBM mix or DRAM ASP guidance at MU, alongside NAND contract-price stabilization and improving SNDK gross-margin guidance. A meaningful reduction in hyperscaler capex plans, or evidence that custom ASIC deployments require less memory per dollar of compute, would warrant reducing sector exposure immediately.

Key Decisions for Investors

• Initiate a 3-6 month pair: long MU / short SNDK in equal dollar amounts. The trade isolates superior HBM/DRAM exposure from broad semiconductor-beta risk; target 15-20% relative upside, with a 7% relative stop if SNDK NAND pricing and margin guidance improve faster than MU DRAM metrics.
• Add to MU only following independently verifiable upward revisions to HBM volumes, DRAM contract prices, or gross-margin guidance at the next earnings update. Avoid chasing a headline-driven move absent these data points; the relevant risk/reward deteriorates if the stock rerates before estimate revisions.
• Use NVDA earnings as a sector timing event: maintain MU exposure if management identifies memory availability or HBM supply as a shipment constraint; trim if GPU demand remains strong but supply commentary shifts toward easing component availability, which would signal the pricing peak may be approaching.
• Do not treat SNDK as a clean AI-memory proxy. Keep any standalone long contingent on evidence of sustained enterprise SSD demand and NAND pricing discipline; otherwise it is a more appropriate funding short against MU than a directional AI position.