[INTERVIEW] AI-fueled rally at Samsung, SK hynix faces sustainability test: Harvard economist

Samsung Electronics and SK hynix are benefiting from surging AI-chip demand and record highs tied to HBM and data center spending, but the article warns that current AI valuations may prove unsustainable without scalable profit models. Harvard economist John Y. Campbell argued that high prices imply lower future returns, with risks from oversupply and intensified Chinese competition if AI infrastructure spending cools. He also highlighted structural flaws in consumer finance, including Korea’s jeonse rental-deposit system and the need for simpler, standardized financial products.

Analysis

The market is conflating two very different earnings regimes: model-layer AI monetization, which is still optionality-heavy and prone to valuation compression, and the picks-and-shovels memory cycle, which is already being converted into cash flow by infrastructure buildout. The more durable trade is not “AI wins everywhere,” but that hyperscaler capex is being forced into a narrower set of bottlenecks, which supports pricing power for high-spec memory and advanced packaging while leaving model developers exposed to faster competitive erosion and lower switching costs.

The second-order risk is that today’s strongest beneficiaries also create their own reversal signal. If capacity additions in HBM and adjacent memory run ahead of enterprise AI demand by even one budgeting cycle, margins can mean-revert quickly because supply is relatively concentrated and expansions are visible months in advance. That means the earnings duration for semiconductor winners is likely measured in quarters, not years, unless utilization stays exceptionally tight.

The broader equity implication is that high AI-related multiples are most vulnerable in the software/platform cohort where revenue quality is still unproven. Investors are paying for future monopoly rents in an ecosystem that may instead normalize toward a multi-model, multi-vendor equilibrium with lower take rates and higher customer churn. That setup favors mean reversion in the most crowded AI leaders while preserving upside in firms with actual input bottlenecks and balance-sheet-supported buybacks.

Contrarian read: the consensus is underestimating how quickly capital can migrate from “AI narratives” to “AI infrastructure.” If monetization disappoints, the first money rotates out of software-like AI exposures and into semis, power, networking, and select industrial enablers rather than out of AI entirely. In that scenario, the real loser is not the whole theme, but the expensive, differentiation-light names that need perpetual multiple expansion to justify their prices.

Key Decisions for Investors

• Tactically stay long semiconductor infrastructure over AI platform exposure: favor a basket of memory and equipment beneficiaries on a 3-6 month horizon, while using rallies to trim any AI software names trading on 2026 revenue stories rather than current earnings.
• Initiate a relative-value short on the most crowded large-cap AI platform names versus semiconductor beneficiaries (pair structure), targeting 10-15% spread compression over the next 1-2 quarters if capex remains strong and monetization evidence stays thin.
• Use downside hedges on GOOGL, AMZN, and META via put spreads into the next earnings cycle; the setup is not a fundamental sell thesis, but a multiple-risk trade if investors start questioning AI payback periods.
• Watch for signs of memory oversupply in 2H: if channel checks show inventory build or ASP stabilization, reduce long exposure quickly—this is a high-beta trade with abrupt margin downside once the market prices a capex pause.
• For more convex exposure, prefer call spreads on semiconductor enablers over outright longs in the AI application layer; the risk/reward is better because the thesis depends on visible procurement budgets rather than uncertain end-user monetization.