AI Data Center Spending Is Outpacing Every Forecast on Wall Street. These 2 Stocks Are the Best Pick-and-Shovel Plays.

Goldman Sachs now forecasts hyperscalers’ 2026 AI capex to reach about $527B (from $465B three months earlier), while spend this year is projected around $750B and could rise further—supporting “pick-and-shovel” beneficiaries Micron and TSMC. Micron reported a 345% Q3 FY2026 revenue jump to $41.5B and a >1,300% surge in adjusted EPS to $24.67, with AI-driven memory demand described as still in early innings. TSMC’s Q1 revenue rose ~41% to nearly $36B, with adjusted EPS up ~58% to $3.49/ADR and gross margin reaching 66%, reinforcing sustained demand for leading-edge AI silicon.

Analysis

The most important read-through is that AI infrastructure is still in the re-acceleration phase, which tends to favor the suppliers with the tightest near-term capacity rather than the platforms spending the money. That means MU can keep comping harder than the hyperscalers for the next 1-2 quarters if memory lead times stay stretched, while TSM remains the cleaner secular beneficiary because every additional accelerator, custom ASIC, or inference chip still flows through its advanced-node bottleneck. The market often underprices this stage of the cycle because it extrapolates spending into lower ROIC; in practice, the first beneficiaries are usually component vendors with pricing power and scarce wafer capacity.

The second-order loser is not necessarily a named stock but the capex-heavy cloud complex: incremental dollars spent on data centers can compress near-term FCF yield and support multiple skepticism if investors see capex outpacing monetization. The key risk for MU is that memory is still a cyclical market underneath the AI narrative; if hyperscaler order patterns normalize into mid-2026, the stock can give back sharply even if secular demand remains intact. For TSM, the main tail risk is geopolitical and customer-concentration related, but in the 6-18 month window the bigger swing factor is whether advanced packaging and leading-edge capacity stay constrained enough to preserve margin.

Contrarian view: consensus may be too focused on the absolute size of AI capex and not enough on the mix. If incremental spending shifts toward lower-margin infrastructure, power, or networking instead of compute and memory, the upside to MU/TSM can be less linear than headline capex suggests. Falsifiers are straightforward: any rollback in hyperscaler 2026 capex guides, memory ASPs rolling over, or TSM margin commentary implying capacity has caught up. If those happen, the trade should de-risk quickly.

Key Decisions for Investors

• Go long MU for the next 1-3 months on the thesis that memory pricing still has the tightest translation from capex to earnings; use a trailing stop if memory ASP checks turn down or if the stock re-rates beyond the next guide-up.
• Stay long TSM as the higher-quality way to express AI capex durability over 6-18 months; the falsifier is any sign that advanced-node utilization or gross margin peaks before the next capacity wave is absorbed.
• Pair trade: long MU / short a basket of capex-heavy hyperscalers (MSFT, AMZN, GOOGL, META) if the market starts penalizing FCF compression; best entry on a weak tape after earnings, with a 1-2 quarter horizon.
• Use NVDA as a watch item rather than an outright chase: if next-channel checks show accelerator demand broadening beyond top-tier hyperscalers, NVDA regains leverage; if not, supplier scarcity may be better expressed in MU/TSM.
• Alert: reduce exposure if 2026 hyperscaler capex estimates stop rising for two consecutive reporting cycles; that would likely mark the inflection from scarcity-driven upside to digestion risk.