What Is the Jevons Paradox and What Does It Mean for Micron and Sandisk Investors After Google's Revolutionary AI Breakthrough?

Google unveiled TurboQuant, claiming at least 6x memory reduction and up to 8x speedup (reducing memory needs by as much as ~83%), which sparked immediate selloffs in memory names (Micron -10%, Sandisk -14%). Mizuho analyst Vijay Rakesh reiterated outperform ratings, citing the Jevons paradox that efficiency-driven price declines could spur greater memory demand; Micron trades at P/E 17 and PEG 0.04 after a >500% three-year run with Q3 guide $33.5B (+260% YoY), GM +660bps to ~81% and adj. EPS ~$19.15, while Sandisk (spun off Feb 2025) is up ~1,850% but trades at P/E 15 and PEG 0.01 with Q3 midpoint revenue $4.6B (+171% YoY) and GM ~65.9%.

Analysis

The Jevons mechanism here is not magic — it is elasticity across a multi-layer billing stack. Lowered memory-per-inference reduces marginal cost per token, which will expand use cases where memory was the gating factor (real-time personalization, long-context agents, and multi-model ensembles). If per-inference memory falls 30–80% for the same model, hyperscalers can either cut price-per-inference or run 2–3x more models within fixed datacenter memory budgets; either path structurally boosts chip volumes even if ASPs compress by 20–40%.

Second-order supply effects favor scale-efficient, low-cost producers with capex discipline. DRAM and HBM demand reacts differently: large model inference shifts spend toward HBM/DRAM bandwidth and persistent caching (NAND), not just raw DDR DRAM. That rebalances value between Micron/Sandisk and incumbents in HBM (Samsung/SK Hynix) and forces OEMs to buy more modules, adapters, and packaging — an inventory-led recovery that typically shows up in supplier bookings 3–9 months after software adoption accelerates.

Big risks are fast, open-source, cross-stack optimizations (which can retro-fit models across fleets), and cyclical oversupply if memory suppliers preemptively flood the market. Watch the hyperscalers’ procurement cadence and component-level spot pricing: if spot DRAM/NAND fall >30% in 90 days without commensurate demand growth, the Jevons lift can be muted. Near-term (days–weeks) volatility will dominate; medium-term (3–12 months) is where demand elasticity should reveal itself and capex cycles will determine realized pricing power.

Key Decisions for Investors

• Buy MU (equity) — 6–12 month hold. Allocate 0.75–1.5% portfolio. Rationale: pure-play DRAM/narrow-moat scale benefits if demand elasticity plays out; target +30–50% if guidance holds and hyperscaler bookings pick up; hard stop -20% from entry or if company revises down capex/revenue guidance.
• Long SNDK via 3–9 month call spread (buy calls or call spread sized to risk 1–2% portfolio) — event trade ahead of next quarterly bookings commentary. Rationale: higher leverage to NAND tailwinds and cleaner balance sheet after spin; target 50–100% option return if revenue/margins accelerate; risk: total premium loss if adoption lags or implied vol collapses.
• Relative-value pair: Long MU + Long SNDK, Short NVDA (equal notional) — 6–12 month horizon. Rationale: isolate memory-volume upside vs GPU multiple compression risk; target net 20–40% absolute return if memory earnings re-rate independently; stop if NVDA outperforms combined memory names by >25% or if market-wide multiple expansion narrows.
• Hedge: buy 9–18 month protective puts on MU (or use collars) sized to 25–50% of long position to limit downside from rapid ASP collapses. Rationale: caps downside from a sudden supply-driven price collapse while leaving upside to Jevons-driven volume expansion.