History Says This Chip ETF Keeps Beating the S&P 500. It Has Trailed the Index Only Twice Since 2016.

VanEck Semiconductor ETF (SMH) has returned about 59% in 2026 versus 12% for the S&P 500 and outperformed the index in eight of the 10 years through 2025, generating nearly 1,400% cumulative return since year-end 2015. The $72B fund remains materially concentrated, with Nvidia representing about 23% of assets, Taiwan Semiconductor about 10%, and the five largest holdings roughly half of the portfolio. The article highlights downside risk: SMH fell about 9% in 2018 and roughly one-third in 2022, about twice the S&P 500's losses, while shares remain 15% below their June 52-week high.

Analysis

The relevant exposure is not broad “semis” but a highly correlated AI-capex factor: NVDA, TSM, AVGO and equipment suppliers are all levered to the same hyperscaler procurement cycle. That makes SMH a poor diversifier for an existing megacap-growth book; its apparent diversification masks simultaneous downside from a cloud capex pause, export-control tightening, or a valuation-driven duration selloff. In a risk-off regime, the ETF’s concentration can mechanically amplify redemptions and force selling of the same liquid leaders that dominate major growth benchmarks.

Near term, the key catalyst is not another AI demand narrative but confirmation that hyperscaler capex converts into sustained orders beyond leading GPUs: TSM monthly sales, NVDA data-center guidance, and commentary from MSFT/GOOGL/AMZN/META should determine whether the current drawdown is consolidation or an earnings-estimate reset. Over 1-3 months, a rebound in NVDA without broadening into TSM, ASML and memory would be technically fragile and suggest positioning rather than improving industry breadth. Over 6-18 months, the bigger risk is that custom ASIC adoption and inference optimization shift profit pools away from merchant accelerators, compressing NVDA’s incremental pricing power even if aggregate AI spending remains robust.

Consensus is too binary on “AI up or AI down.” The more likely adverse outcome is continued AI infrastructure growth paired with lower returns on incremental capital, which hurts the highest-multiple compute vendors while benefiting lower-multiple foundry, networking, memory and equipment franchises. A broad ETF therefore may underperform a selectively constructed AI supply-chain basket if spending rotates from training clusters toward deployment, networking and edge inference.

Key Decisions for Investors

• Do not add outright SMH exposure until breadth improves: require NVDA to stabilize while TSM and ASML outperform SMH for 2-3 weeks. Without that confirmation, treat SMH as concentrated growth beta rather than a diversified semiconductor allocation.
• For a 3-6 month relative-value expression, consider long TSM versus short SMH in matched beta size. TSM retains AI volume participation but has less single-customer/product valuation sensitivity; exit if TSM monthly revenue decelerates materially or if NVDA raises forward data-center guidance enough to reaccelerate SMH earnings revisions.
• For existing NVDA longs, buy 3-6 month downside puts or implement a collar around the next earnings cycle rather than selling core exposure into weakness. The hedge is warranted if implied volatility remains below the level associated with prior earnings-driven gap risk; reduce it if hyperscaler capex guidance is raised broadly.
• Watch memory and networking as deployment-stage beneficiaries: MU and AVGO merit a research watchlist, not immediate purchases, pending evidence of improving HBM pricing, switch demand, and non-NVDA revenue growth. Those data would validate a rotation within semis rather than a new broad-cycle peak.
• Use an SMH/S&P 500 relative-strength break below the prior 3-month low as a portfolio-risk trigger: cut semiconductor overweight toward neutral, because that would signal earnings-revision deterioration rather than an isolated leader correction.