Buy these memory stocks to play the red-hot CPU trade: Analyst

Mizuho’s Jordan Klein said he is "more bullish" on DRAM and memory as AI CPU demand accelerates, citing tight supply and no meaningful new capacity until 2H27. He highlighted Micron at 3-4x buyside EPS and said SanDisk could benefit from rising SSD demand and potential quarterly EPS of $25-30 with 80%+ gross margins. The article also notes stronger CPU attach rates in agentic AI workloads and favors AMD and Nvidia’s ARM-based CPU opportunities over Intel’s more capacity-constrained position.

Analysis

This is less a generic AI-semiconductor upside story than a re-pricing of memory as the gating factor for AI capex. The second-order effect is that every incremental CPU/agentic deployment increases not just compute demand but the working-capital intensity of the whole stack: more DRAM per server, tighter SSD inventory, and better pricing power for the memory cohort before any meaningful greenfield supply can arrive. That creates a cleaner earnings revision cycle in MU than in the more crowded AI compute names, because memory is moving from “cyclical commodity” to “scarcity premium” over the next 4-8 quarters.

The market is likely underestimating how much agentic AI changes the CPU mix. If orchestration-heavy workloads materially raise CPU attach rates, then AMD and NVDA benefit twice: first through unit demand, then through higher-value platform lock-in around their respective ecosystems. TSM is the quiet structural winner here because it is the bottleneck that monetizes both the accelerator and CPU wave without taking end-demand risk directly; however, that also means any slowdown in AI capex would transmit quickly into foundry utilization sentiment.

The contrarian risk is positioning and timing. The easiest money may already have been made in the obvious AI names, while memory is still cheap enough to rerate if guideposts remain firm into the next print cycle. The reversal case is a digestion phase in AI server orders or any sign that cloud customers are optimizing memory per node rather than expanding footprints, which would hit the trade within days in high-beta names and over 1-2 quarters in the supply chain.

Near term, the highest-risk assumption is that the market extrapolates straight-line demand into 2026/27 without a pause. If memory margins overshoot, expect capital allocation pressure to increase and eventual supply response to cap the upside, but that is a later-cycle issue; the next catalyst set is earnings/guidance, not capacity additions. In other words, this is a “buy the scarcity window” setup rather than a durable secular-growth rerating across all AI semis.

Key Decisions for Investors

• Long MU for a 3-6 month trade into the next earnings cycle; upside is best if management confirms pricing firmness and HBM/DRAM tightness. Use a starter position on any post-rally consolidation rather than chasing strength.
• Long SNDK on pullbacks as a higher-beta way to express SSD demand strength; risk/reward is attractive if the market continues to ignore ownership and volume risk, but trim if the stock gaps on pre-earnings optimism.
• Pair trade: long MU / short AMD or long MU / short ARM over 1-2 quarters. The thesis is that memory is still under-owned and earlier in the earnings revision cycle, while CPU/AI platform names are already more consensus-owned.
• Long NVDA with a tighter horizon only if you want the agentic AI CPU-attach optionality; best expressed via call spreads 3-6 months out to limit downside if the narrative proves overstated.
• Add TSM on weakness as the lowest-volatility beneficiary of the CPU/accelerator mix shift; hold 6-12 months, but size modestly because foundry exposure is still one step removed from the final demand upside.