The VanEck Semiconductor ETF (SMH) closed at $647.10 on June 15, 2026, capping a one-year return of approximately 136%. That number is not a performance footnote. It is a statement about where institutional capital has been going, and more importantly, where it intends to stay.
Who Is in This Trade
The composition of today’s chip rally is worth examining carefully. SMH’s top holdings, including NVIDIA, TSMC, Broadcom, and others such as Marvell, represent different points along the same supply chain, and investors have been treating them as a single thesis rather than individual companies. That has made the ETF a vessel for conviction trading: a single ticker that captures the entire AI infrastructure stack from fab to accelerator to networking switch.
Goldman Sachs tracks a basket of AI infrastructure companies that returned an average of 44% year-to-date through mid-2026, compared with a 9% increase in consensus two-year forward earnings estimates for the same group. The gap between price appreciation and earnings growth is the tell. This is not a fundamental rerating. It is a positioning trade, driven by the belief that the infrastructure buildout is durable enough to justify multiples that have run ahead of the underlying numbers.
Retail flows into semiconductor ETFs have been consistent, but the institutional picture shows continued directional conviction. Options activity on SMH has reflected sophisticated capital using leverage to amplify exposure in a high-conviction environment.
Where the $630 Billion Is Going
The four largest hyperscalers, Amazon, Microsoft, Alphabet, and Meta, are projected to spend approximately $630 billion on infrastructure in 2026 (with broader analyst estimates ranging from roughly $600 billion to $725 billion), roughly triple what they spent two years ago. Of that figure, only about 25% flows directly to chips. The remaining 75% goes to data centers, power systems, cooling infrastructure, networking hardware, and land acquisition.
That breakdown matters for investors who think they are buying the AI trade through NVIDIA alone. The real capital deployment is broader, and increasingly, it is finding its way into power utilities, REITs with data center exposure, and specialized networking hardware companies like Arista and Marvell. The chip is the headline, but the picks-and-shovels story has expanded well beyond the fab.
TSMC is the clearest example of how durable this spending looks from the supply side. The company holds a 72% market share in pure-play semiconductor foundries and projects AI accelerators (high-teens percent of its 2025 revenue, around 18%) to grow at a 55–60% compound annual rate through 2029. Those projections come with multi-year customer agreements already signed. TSMC’s 2026 capital expenditure guidance of $52–56 billion reflects how seriously they are taking the demand signal.
The Risk Nobody Is Pricing
Goldman Sachs noted it plainly in their research: “The timing of an eventual slowdown in capex growth poses a risk to these companies’ valuations.” That is the quiet threat sitting underneath the strong one-year performance in SMH.
Infrastructure spending of this scale requires continuous justification from the hyperscalers’ own earnings. If AI-driven revenue does not grow fast enough to absorb the capex, the cycle stalls. It has happened before: the fiber optic overbuilding of the late 1990s, the server farm expansion cycle of 2015–2017. In both cases the infrastructure was real and the demand was real, but the timing mismatch between supply and monetization created violent corrections in the companies that had priced in perfection.
The current AI trade is pricing in a remarkably clean outcome: sustained hyperscaler spending, consistent chip demand, no major geopolitical disruption to TSMC’s Taiwan fabs, and a steady march toward AI-generated revenue that justifies the multiples. Each of those assumptions is individually reasonable. All of them being true simultaneously is a different calculation entirely.
What Comes Next
Capital does not stay still. The movement into semiconductor ETFs and AI infrastructure names over the past 18 months represents one of the largest thematic rotations in equity markets since the post-COVID tech surge. The question is whether this is a structural reallocation, institutions permanently repricing AI infrastructure as a core holding, or a cyclical trade that will unwind when the first hyperscaler cuts its capex guidance.
The answer is probably both, in sequence. The structural thesis is real: AI computation is not a temporary demand spike. But current valuations reflect the structural thesis running into a near-term positioning trade, and those two things are priced together in the same ETF. Investors who understand the difference are better positioned to hold through the volatility that will come when the cycle eventually pauses.
Follow @UTXOMacro for daily Bitcoin, macro, and AI breakdowns.