Hyperscalers guided capital spending sharply higher; consensus 2026 capex for the five largest US spenders now sits at $844 billion, up 81% year over year. Spending that much, that fast, pushes near-term free cash flows negative. The market priced that gap today and discounted the revenue pool being built.

Click here for SMH Standardized Performance

Past performance is no guarantee of future results. Investment return and principal value will fluctuate; shares may be worth more or less than original cost when redeemed. Current performance may be lower or higher. Call 800.826.2333 or visit vaneck.com for month-end performance.

How Does the Current AI Chip Drawdown Compare to Prior Semiconductor Cycles?

This is the fourth deepest of seven semiconductor drawdowns since 2011. When compared to 2022, when demand fell and earnings fell with it, this time earnings more than doubled, and what fell was the multiple. TSMC (Taiwan Semiconductor Manufacturing Company Limited, "TSM") outperformed on revenue and profit, guided capex above its prior forecast, and still fell more than 3% the next session.

Historic Semiconductor Drawdown

Source: FactSet Global Prices, SMH-US monthly and daily closes, USD, price return (excludes dividends), retrieved August 11, 2026. Past performance is not a guarantee of future results. Fund performance current to the most recent month end is available by visiting vaneck.com or by calling 800.826.2333.

Source: FactSet Global Prices. Peak, trough and recovery months are VanEck's identification at a 15% threshold. Past performance is not a guarantee of future results. Depth figures reflect SMH share price on a price-return basis. Recovery periods shown are historical and are not an indication of any future recovery. Please see important disclosures and definitions at the end of the commentary.

What Is $844 Billion a Year of AI Infrastructure Spending Buying?

The key distinction in the current AI cycle is between the capex level and the capex growth rate. The market is trading the growth rate. The companies supplying this build get paid on the level.

Capex for the five largest hyperscalers is predicted to climb to approximately $1.37 trillion by 2030. The growth rate decelerates sharply but never decreases. A hyperscaler spending $1.15 trillion a year and never growing that number still represents $1.15 trillion of annual demand for every company selling into it.

Free cash flow turns negative in 2026 and 2027 because capex growth is running ahead of operating cash flow growth. That gap closes mechanically in 2028 when capex growth comes back down with no demand change. Two negative years in the middle of the largest infrastructure build in corporate history is telling a story about timing rather than demand.

Capex Level Keeps Rising While FCF Recovers ($bn)

Source: FactSet Estimates, consensus mean CAPEX and CF_OP, for MSFT, GOOGL, AMZN, META and ORCL; estimate date August 7, 2026. Forecasts are subject to change and may not be realized. For illustrative purposes only

Source: FactSet Estimates, consensus mean, estimate date August 7, 2026. Forecasts are not guarantees of future results. For illustrative purposes only

How Far into the AI Deployment Cycle Are We, and What Drives Compute Demand From Here?

Most large organizations have touched AI. Very few have finished deploying it. 88% say they use AI somewhere, but only 6% have crossed McKinsey's threshold of meaningful earnings before interest and taxes (EBIT) impact. (Source: McKinsey & Company, “The State of AI in 2025: Agents, Innovation, and Transformation,” November 2025.) That gap between having AI somewhere and having it everywhere is the runway for compute demand.

The internet analogy is useful here: by 2010, almost 72% of Americans were online, but only 4.2% of retail dollars had moved there. (Sources: NTIA, "Digital Nation: Expanding Internet Usage," February 2011; U.S. Census Bureau, "Quarterly Retail E-Commerce Sales, 4th Quarter 2010," February 2011.) ChatGPT launched in 2022, meaning we are only in year three of adoption. The next leg of compute demand is agents, software that plans, retrieves, attempts, and retries rather than answering once. One agentic task can consume hundreds of times the compute of a single chat, and Goldman Sachs Research forecasts token volume dominated by enterprise agent workloads from 2028 onward.

Predicted token growth by 2030

Source: Goldman Sachs Research, 'AI Agents Forecast to Boost Tech Cash Flow as Usage Soars,' May 20, 2026. Forecasts are subject to change and may not be realized. For illustrative purposes only.

Which Parts of the AI Supply Chain Benefit Most from Hyperscaler Spending?

The supply chain runs from chip design at the top, most sensitive to how fast spending grows, down to power infrastructure and data centers at the base, which get paid on the level regardless of the growth rate.

Why Chip Design Sits at the Top of the AI Spending Chain

Fabless semiconductor companies design chips and outsource manufacturing to foundries like TSMC. Without fabs to maintain, they put capital into R&D and move faster as compute architecture evolves. NVIDIA is a prime example; they pivoted from gaming GPUs to a dominant AI training chip, made possible by the flexibility fabless design allows. The model covers both training and inference, including the custom silicon now being designed by Amazon, Google, Meta, and OpenAI.

SMHX, VanEck's fabless semiconductor ETF, provides pure-play exposure to this layer, excluding foundries, equipment makers, and vertically integrated manufacturers.

Capex to Operating Cash Flow

R&D Expenses to Sales

Source: MarketVector as of 2025. "Fabless" is the MarketVector™ US Listed Fabless Semiconductor Index (MVSMHX); "IDM" is a MarketVector-constructed set of companies classified by MarketVector as integrated device manufacturers. Past performance is not a guarantee of future results. Index performance is not illustrative of fund performance. It is not possible to invest directly in an index.

Why Every Layer of the Data Center Supply Chain Is Under Simultaneous Strain

Demand for chips, memory, cooling, and power is constrained at once, and each bottleneck carries multi-year order book visibility for the companies resolving it.

Source: WSTS/SIA; Deloitte; Gartner; TrendForce; Bloomberg; Vertiv; Grand View Research; MarketsandMarkets; U.S. Department of Energy.

Power is the most durable of those constraints. US data center grid-power demand is projected to more than double from 62 GW in 2025 to 134 GW by 2030, and the grid is not built for it. Leading hyperscalers have signed nuclear agreements in the past 18 months because it is the only clean energy source that scales to data center load with minimal downtime.

US Data Center Power Demand (GW)

Source: S&P Global Commodity Insights / 451 Research; EPRI, Powering Intelligence (2026); U.S. Department of Energy; company announcements.

RACK, VanEck's data center supply chain ETF, spans this entire chain, from semiconductors and memory through cooling, power, and nuclear energy. It targets the companies the hyperscalers are spending with, not the hyperscalers themselves.

How Does Falling Robot Cost Connect to AI Deployment?

The cost of industrial robots has fallen dramatically over the past two decades, shifting adoption from a strategic consideration to an increasingly standard practice for many manufacturers. As AI moves from answering questions to completing workflows, a portion of that work happens in factories and warehouses. IBOT, VanEck's robotics ETF, captures this layer.

What Makes the Space Economy an Opportunity Independent of AI Capex?

Space sits off the hyperscaler capex chain by design. The Space Shuttle costs $54,500 per kilogram to reach orbit. Falcon 9 does it for under $2,720. (Source: NASA, "The Recent Large Reduction in Space Launch Cost,” 2020.) That cost reduction turned access to orbit from a government monopoly into a commercial industry. Space economy revenues are expected to triple by 2035. (Source: World Economic Forum and McKinsey & Company, "Space: The $1.8 Trillion Opportunity for Global Economic Growth," April 2024.) WARP, VanEck's pure-play space ETF, takes advantage of the growing space economy.

What Does a Decelerating AI Capex Growth Rate Mean for Investors?

The growth rate is going to come down. That hits the top of the supply chain hardest. But a slower growth rate is not the same as a cut. At 3% growth on $1.15 trillion, the spend is still above $1.15 trillion. No hyperscaler management team or analyst consensus is pointing to actual cuts.

The drawdown repriced multiples on a business whose earnings more than doubled and whose customers accelerated spending. The depth of AI deployment has barely started, the infrastructure supporting it is under strain at every layer, and the spending funding it is not going down. The growth rate is normalizing. The spend is not.

To receive more Thematic Investing insights, sign up in our subscription center.