Why the SMH ETF Could Be a Core AI Investment for 2026

VanEck Semiconductor ETF (SMH) provides concentrated exposure to AI infrastructure leaders—Nvidia, TSMC, Broadcom, Micron and AMD comprise ~49.8% of assets and the top 10 ~73%—and returned roughly 49% in 2025 with a ~30.9% annualized gain over the past decade (as of Jan. 8, 2026). Goldman Sachs estimates $527 billion in global AI-related data center capex in 2026 and Deloitte projects inference will represent two-thirds of AI compute demand by 2026, underpinning sustained demand for GPUs, memory, networking and other semiconductor-related hardware. SMH trades at about 33x trailing 12-month earnings, making it an efficient broad AI-infrastructure proxy but one that carries above-market valuation risk.

Analysis

Market structure: Hyperscalers, NVIDIA (NVDA), TSMC (TSM), memory suppliers (MU) and equipment leaders (ASML, LRCX, KLAC) are primary beneficiaries as AI data‑center capex (Goldman $527B 2026) concentrates spend. SMH’s top 5 = ~49.8% of assets, so ETF flows amplify winners and create single‑name concentration risk; smaller fabless, legacy CPU vendors and non‑AI memory suppliers are relative losers. Supply constraints (EUV tool lead times, HBM supply, TSMC node capacity) imply pricing power and multi‑quarter revenue visibility for fabs/equipment; commodity impact: higher copper/energy demand, tighter helium/NF3 supply chains; risk‑on semantics support wider credit spreads tightening but could push real yields higher if capex inflates investment demand.

Risk assessment: Tail risks include tightened export controls/China decoupling or a cyclical memory glut; a 10–30% demand shock would materially compress earnings for MU/TSM in 6–12 months. Immediate catalysts (days–weeks): NVDA/TSM/ASML earnings and customer capex commentary; short term (3–6 months): memory price inflection and foundry booking cadence; long term (2026+) inference accounting for ~66% of compute signals durable demand but opens the risk of inference ASICs and edge acceleration disintermediating GPU unit growth. Hidden dependencies: hyperscaler procurement cadence, power/thermal ceilings in racks, HBM supply; monitor booking lead times, ASML delivery schedules, and US export policy over next 90 days.

Trade implications: Direct: establish tactical longs in NVDA (2–3% portfolio) and ASML (1–2%) for 6–18 months, trim into strength; use MU (1%) as cyclical leveraged memory exposure but size small. Pair: long NVDA / short QCOM (0.8/0.8%) to express data‑center GPU upside vs smartphone modem cyclicality through next 12 months. Options: sell 30–60d cash‑secured puts on TSM sized to 1–2% at strikes 5–8% below market to collect premium and gain entry on weakness; consider buying NVDA Jan‑2027 LEAPS (~delta 0.30–0.40) for convex exposure if conviction >12 months. Rotate out of late‑cycle software multiple expansion positions into semis on 5–12% pullbacks.

Contrarian angles: The market underestimates medium‑term memory cyclicality and the risk that efficient inference (quantized models, NPU ASICs) reduces GPU unit growth — this could compress NVDA forward multiples by 15–30% if realized. Conversely, SMH at ~33x TTM already embeds significant growth; a 10% correction would present a disciplined buy window for diversified capex exposure. Historical parallel: 2016–18 GPU cycles showed equipment lead times and memory cycles can flip returns quickly; unintended consequence — hyperscalers optimizing server density and power may favor specialized accelerators, creating dispersion within SMH rather than uniform upside.

Key Decisions for Investors

• Establish a 2–3% long position in NVDA for a 6–18 month horizon to capture inference-driven GPU demand; set a stop‑loss at 15% below entry and scale out 30–50% of the position on a 25–40% gain.
• Add ASML (1–2% portfolio) as a long-term (12–36 months) structurally defensive play on EUV scarcity and tool lead times; accumulate on any >8% intra‑month pullback and trim half on a 30% appreciation.
• Deploy cash‑secured put selling on TSM sized to 1–2% of capital with 30–60 day expiries at strikes 5–8% below market to generate yield and/or acquire TSM at a discount; close if puts trade >2x premium paid or if TSM issues negative 2‑quarter capacity guidance.
• Implement a relative pair: long NVDA (1.6%) / short QCOM (1.6%) to express data‑center vs smartphone exposure over 12 months; rebalance if spread narrows by >20% or on material policy shifts (US‑China export controls) within 90 days.
• Hold a 1–2% tactical SMH position but only add to reach 4–6% total semiconductor allocation if SMH retraces to a forward/TTM PE ≈28x or drops >10% from current levels, signaling a durable buying opportunity rather than chasing momentum.