Morgan Stanley sends clear message on semiconductor stocks after selloff

Semiconductor sector (SOXX) fell roughly 10% over the past month after Google rolled out TurboQuant, which Google says can cut certain AI memory needs by up to 6x, triggering >10% drops in memory names like Micron and SanDisk. Morgan Stanley disagrees that TurboQuant materially reduces overall memory demand, keeps overweight ratings on Micron ($520 PT) and SanDisk ($690 PT), and argues memory shortages will persist and support elevated earnings. Micron reported a record 81% gross margin and $23.86B revenue for fiscal Q2 2026; the stock is up ~25% YTD, closing at $357 on March 27.

Analysis

Compression and quantization advances are acting like a demand multiplier, not a pure subtractor: by lowering cost-per-inference these techniques widen the set of viable use-cases and shorten model iteration cycles, which drives incremental memory capacity per data center over 12–36 months. That dynamic amplifies demand heterogeneity — HBM and advanced-packaged die-to-die stacks become more critical than commodity DDR for large-scale inference clusters, shifting margin leverage to packaging and test suppliers with constrained capacity.

Corporate procurement behavior is the overlooked amplifier. When hyperscalers respond to perceived scarcity by signing forward take-or-pay commitments, they both smooth near-term revenue for suppliers and materially shorten available spot supply, increasing the probability of episodic shortages and tighter pricing for at least one capex cycle (6–18 months) while new capacity ramps. Conversely, a coordinated acceleration of capex by multiple suppliers remains the largest medium-term reversion risk because wafer lead times and tool delivery can flood the market within 12–24 months.

From a cross-chain perspective, beneficiaries are not just memory fabs — equipment vendors, advanced packaging houses, and wafer-foundry partners will capture outsized economics as content shifts toward HBM and vertically integrated modules. The key fragility is technological commoditization: a ubiquitous, low-cost compression standard adopted across model stacks would compress revenue per memory-bit and favor incumbents with scale in DRAM/NAND manufacturing, creating a winner-take-most outcome in two years.

Key Decisions for Investors

• Long selective memory fabs (MU) via a 12-month call spread (buy 12m ATM call, sell 12m +30% call) to capture sustained tightness while capping cost; position size sized to risk 3–5% portfolio. Rationale: asymmetric upside if scarcity persists; capped loss if market re-prices within a year.
• Pair trade: Long advanced-packaging/equipment names (e.g., ASML, AMAT) and short broad semiconductor ETF (e.g., SOXX) for a 6–18 month horizon to express structural mix shift toward HBM/stacking where tool and packaging scarcity creates outsized margins. Target 2:1 notional to the long leg; rebalance on quarterly inventory prints.
• Sell short-dated volatility on large cloud providers (GOOGL) around product messaging windows (30–60 days) using covered-call overlays or short-dated iron condors — premium tends to overreact to headline efficiency announcements; cap tail risk with defined-loss spreads.
• Tactical short: commodity NAND/consumer flash names with high inventory exposure for 3–6 months if spot ASPs show sequential declines on OEM destocking; size small and use buy-stop limit to cap loss given potential scarcity-driven squeezes.
• Risk management: set portfolio-level stop if memory ASP index falls 20% from current levels or if announced new wafer start capacity implies >10% supply growth within 12 months; take partial profits at +40% on long memory directional trades.