3 of the Top 5 Performers in the S&P 500 in the First Half Offer 1 Particular Thing for AI Customers. (Hint: It Isn't the GPU, and Nvidia Isn't Among Them.)

The article argues the AI infrastructure winners are broadening from GPUs to memory/storage, citing first-half S&P 500 outperformance for Sandisk (SNDK), Micron (MU), and Western Digital (WDC). It highlights rapid AI data-center growth: SNDK data center revenue up 640% YoY to $1.4B, Micron data center revenue up 650% to $11B, and WDC total revenue up 45% to over $3B. It also reiterates Nvidia’s dominance, with revenue last year exceeding $215B and NVDA/AMD both up more than 350% over the prior three years, framing memory names as a next phase of AI demand driven by inference (2/3 of compute).

Analysis

The important shift is not “AI is broadening” but that the profit pool is moving from one-time model buildouts toward recurring data retention and retrieval economics. That favors memory/storage vendors with pricing power and high operating leverage, while making the GPU duopoly less of a singular trade as hyperscalers rebalance capex toward inference infrastructure. In practice, this can create a relative-performance window where MU and SNDK re-rate faster than NVDA because investors start paying for durability of demand, not just growth rate.

The second-order effect is on capex mix, not total capex. As inference becomes a larger share of AI workload, customers generate more data per deployed dollar, which increases attachment rates for NAND, DRAM, and storage systems; that should also support adjacent networking and data-center power names over time. WDC’s upside is more diluted than MU/SNDK because HDD remains more cyclical and less clearly tied to the highest-value AI workloads, so the market may overestimate how much of the AI spend converts into structural margin expansion there.

The near-term catalyst is earnings guidance over the next 1-3 quarters, especially whether memory vendors can sustain tight supply and disciplined pricing. The main contrarian risk is that memory is still a commodity cycle disguised as a secular theme: if inventory was pulled forward or cloud capex slows, these names can de-rate faster than NVDA because the market is paying a premium for perceived scarcity. The thesis is falsified if spot memory pricing rolls over, data-center revenue growth decelerates meaningfully, or hyperscaler commentary shifts back toward compute-heavy rather than storage-heavy deployment.

Key Decisions for Investors

• Long MU vs. short NVDA on a 1-3 month relative-value basis: express the view that AI spend is broadening into storage while keeping beta to the AI theme; size modestly because NVDA remains the central tollbooth and would outperform if compute demand re-accelerates.
• Add SNDK on pullbacks as a cleaner second-order beneficiary of inference growth; use a 3-6 month horizon and take profits if memory pricing weakens or the stock trades far ahead of next earnings.
• Avoid chasing WDC into strength; if anything, pair long MU / short WDC to isolate higher-quality AI memory exposure versus a legacy storage name with weaker structural pricing power.
• For defined risk, buy MU 3-6 month call spreads ahead of the next data-center commentary cycle; thesis breaks if gross margin guidance comes in below trend or if customer inventory signals soften.
• Set a watch item on DRAM/NAND spot pricing and hyperscaler capex mix: if either signals a return to compute-only spend or >10% pricing deterioration, cut the memory overweight.