Memory panic amid TurboQuant is a buying opportunity in these stocks: BofA

Micron shares fell 7% and the SOX dropped ~5% after concerns around Google’s TurboQuant (which can reduce KV cache memory footprint up to 6x). Bank of America argues the selloff is overdone, noting Google simultaneously doubled its 2026 capex outlook to ~$180B and that upcoming chips (Google TPU v7 with 192GB HBM vs TPU v6 at 32GB, plus Nvidia/AMD/Meta roadmaps) point to rising memory needs. BofA sees the episode as a buying opportunity and favors exposure to AI compute, semicap equipment, networking and memory names such as Nvidia, Broadcom, AMD, Lam Research, Applied Materials, Marvell and Credo; it also cites a 66% rise in China AI capex in 2025 and the temporary nature of prior efficiency fears.

Analysis

Compression algorithms change the unit economics of training but do not linearly reduce total infrastructure demand; in fact, lower per-parameter cost tends to catalyze model scale and experimentation, increasing aggregate FLOPs and memory bytes consumed over 12–36 months. This creates a bifurcated market: vendors that sell more memory per system (HBM, on-package buffers, high-density DIMMs) and those selling the upstream equipment to produce them (EUV-capable foundry inputs, HVM test and assembly) will capture most upside, while pure spot-DRAM cyclical exposure remains vulnerable to sentiment-driven swings.

Timing and catalysts are specific: expect headline volatility in days–weeks around software announcements and earnings, inventory and pricing adjustments in quarters 1–4, and structural demand shifts over 12–36 months as hyperscalers deploy larger models. Watch cloud capex guides, wafer-start cadence, HBM spot ASPs and foundry allocation notices as primary catalysts; reversal risks come from rapid, cross-industry adoption of aggressive compression standards, a sudden surge in HBM supply (capacity ramp >20% YoY), or a pivot to alternative memory fabrics (CXL/persistent memory) that change content-per-box math.

The market reaction appears anchored to a single narrative; the contrarian view is that efficiency is a demand accelerant, not a degenerative force. That implies preferencing vendors that expand content-per-server or control the upstream capacity cycle (NVDA, AVGO, LRCX, AMAT) over commodity DRAM bets that face pricing and inventory cyclicality. Position sizing should distinguish short-term sentiment trades from multi-quarter structural exposures to avoid being whipsawed by memory-price noise.

Key Decisions for Investors

• Long NVDA (6–12 month horizon): buy a Jan-2027 call spread (debit) sized to risk 1% portfolio notional; target 3x premium if cloud capex/AI model deployments accelerate, stop-loss if spread premium falls 50% within 60 days.
• Semi-cap equipment exposure via LRCX/AMAT (9–18 months): buy LRCX outright or LEAP calls (Jan-2028) with a 15% stop and +30–50% price target tied to wafer-start recovery and HBM packaging demand; risk limited to premium/position size.
• Pair trade (3–9 months): long NVDA / short MU equal-dollar to isolate compute vs memory-price risk — expect asymmetric upside if compute content per box rises. Close or hedge if the spread compresses by >10% intraday or if MU issues positive inventory drawdown guidance.
• Opportunistic MU mean-reversion (6 months): accumulate on >10% intra-day weakness from current levels with a 20% hard stop and a 40–60% upside target if hyperscaler capex is reiterated; keep position size tactical (<=1% portfolio) due to inventory/cycle risk.