AI Infrastructure ETFs Quietly Outran Nvidia This Year

The year's best AI trade was not the one on the front page#

Nvidia is the stock that stands in for the whole AI story, and on 9 September 2026 it was worth about $5.4 trillion (Yahoo Finance). It closed 2025 at $186.50 a share and finished trading on 31 August 2026 at $220.78, a price gain of roughly 18% (Macrotrends; StockAnalysis). A good eight months by any normal standard, and a mediocre way to own the artificial intelligence build-out.

Over the same period the VanEck Semiconductor ETF returned 54.56% on net asset value. The Defiance AI & Power Infrastructure ETF, which mostly owns switchgear and turbine makers rather than chip designers, returned 26.35%. A fund launched back in November 2009, the First Trust Nasdaq Clean Edge Smart Grid Infrastructure Index Fund, returned 17.02% and grew to $12.01 billion in assets by 9 September, after taking in more than $3.1 billion of new money in the first four months of the year alone (Bloomberg).

Search behaviour tracked the money late. US searches for "grid etf" ran at about 2,400 a month in August 2025 and hit 14,800 in May 2026, while "ai infrastructure etf" went from 880 to 4,400 (Google Ads data retrieved via DataForSEO, 11 September 2026).

What an AI infrastructure ETF actually holds#

An exchange-traded fund is a basket of shares you can buy in a single trade, priced continuously through the day. Most of the funds here are passive, following a published index rule rather than a manager's hunches. You pay an annual fee, quoted as an expense ratio, taken out of the fund's value a little at a time rather than billed to you.

The phrase "AI infrastructure" covers four layers, and funds sitting on each behave nothing like each other. There is silicon, meaning the processors and memory that do the computing. There is the box: servers, racks, cooling and the power units that keep them alive. There is electricity, from gas turbines and transformers to the high-voltage cables that carry it. And there is property, the windowless warehouses themselves, usually owned by real estate investment trusts.

The holdings show the gap. GRID's five biggest positions on 9 September were Schneider Electric at 9.12%, Eaton at 8.84%, Johnson Controls at 8.43%, ABB at 7.70% and Quanta Services at 7.51%, out of 119 holdings. Nvidia does not appear at all. The Defiance fund mixes layers, with Eaton at 8.95%, GE Vernova at 8.88% and Nvidia at 4.51%.

One term before the table. Standardised performance means returns published in a fixed regulatory format at month end, on net asset value, after fees. It is the only figure that compares two funds fairly, which is why every fund number below carries the same date.

The scoreboard to 31 August 2026#

Fund figures are standardised NAV total returns to 31 August 2026, published by VanEck, Defiance, First Trust, State Street and VanEck again. Nvidia is a price return and excludes its small dividend. Past performance is not a guide to future returns.

Interpretation, offered as opinion: the spread from 54.56% down to minus 6.27% inside one theme is the story. Four funds all marketed against the same headline trend produced outcomes that had almost nothing in common.

Why the money moved down the stack#

The order books explain most of it.

GE Vernova reported second-quarter orders of $24.2 billion on 22 July 2026, up 88% organically, with backlog rising $13.0 billion in three months. Its gas power equipment backlog and reserved factory slots went from 100 GW to 116 GW in a single quarter. Data centre orders passed $5 billion in the first half, more than double the company's entire 2025 total.

A week later Vertiv reported quarterly net sales of $3.27 billion, up 24%, and raised full-year guidance to between $13.8 and $14.2 billion. Chief executive Giordano Albertazzi said demand "continues to intensify and with each technology advancement, deployments grow more complex and more infrastructure-intensive."

A chip is useless without a building, a cooling loop and a connection to a substation. The firms supplying those things book revenue on contracts signed long before a single graphics card is installed.

The demand behind the orders is measurable. The International Energy Agency's Energy and AI report put data centre electricity use at about 415 terawatt hours in 2024, roughly 1.5% of the world total, and projected it would more than double to around 945 TWh by 2030. Data centre investment reached about half a trillion dollars in 2024, nearly twice the 2022 figure.

Grid operators are repricing too. In its capacity auction for the 2026/2027 delivery year, PJM, which runs the wholesale electricity market across thirteen US states, cleared at $329.17 per megawatt-day, the maximum the regulator allows, against $269.92 a year earlier. Forecast peak demand rose more than 5,400 megawatts, driven largely by data centre expansion. A capacity auction pays generators simply to be available on the worst day of the year, so a price stuck at its ceiling is the system saying it is short.

The label tells you almost nothing#

The "AI power" label has also been attached to funds that did not work.

The VanEck Uranium & Nuclear ETF lost 6.27% over the first eight months of 2026, despite holding Constellation Energy and Cameco, the two names most often cited as the nuclear answer to data centre demand. The Utilities Select Sector SPDR returned 0.19%. Investors noticed. State Street's mid-year flow review recorded $823 million leaving utilities ETFs in the first half, while technology funds took $44.8 billion and industrials $9.9 billion, out of record US ETF inflows above $1 trillion.

Interpretation: utilities sell electricity at rates regulators approve, so a surge in demand does not automatically become a surge in profit, and it can mean years of heavy capital spending first. Equipment makers get paid on the way in.

Two things are worth checking before buying any of these. Overlap is one. SMH holds 22.39% in Nvidia, so an investor who already owns Nvidia and then buys the fund for diversification has mostly bought more Nvidia, plus Taiwan Semiconductor at 9.70% and Broadcom at 5.81%.

Cost is the other. Defiance charges 0.69% a year and GRID 0.56%, against 0.0945% for a plain S&P 500 tracker. On $10,000 that gap is about $60 a year, every year, whether the theme works or not. Thematic funds also tend to launch once a story is established, so the fee starts running at the point of maximum enthusiasm.

What would have to go wrong#

All of this rests on capital spending by a handful of buyers, and capital spending is a decision rather than a law of physics.

The Bank of England's Financial Policy Committee said in its July 2026 Financial Stability Report that "the risk of a sharp correction in equity markets remains high", and that AI companies now account for around half of the S&P 500 by value, up from roughly a quarter in 2022. The excess cyclically adjusted price-to-earnings yield on the index has drifted towards levels last seen in the dot-com bubble.

Two fragilities follow from the orders themselves. Backlogs are promises, and orders booked in today's mood can be delayed or cancelled. The customer list is also short, so one cloud company slowing a build programme lands on many suppliers at once.

Hypothesis rather than observed fact: the layer that gains most from a slowdown scare may be the regulated one, which is exactly the layer that lagged this year.

Key takeaways#

1. To 31 August 2026, a semiconductor ETF returned 54.56% and an AI-and-power fund 26.35%, against roughly 18% for Nvidia's share price and 13.04% for the S&P 500 tracker.
2. The same theme produced a 6.27% loss in a uranium and nuclear fund and a 0.19% return in utilities. The label is not the exposure.
3. Equipment makers led because they book revenue early. GE Vernova's data centre orders passed $5 billion in the first half of 2026, more than double its 2025 total.
4. The demand is documented, not hypothetical: the IEA expects data centre electricity use to more than double to about 945 TWh by 2030, and PJM's capacity price hit its regulatory ceiling.
5. Fees run from 0.0945% to 0.69%, overlap with shares you may already own can be substantial, and the Bank of England rates the risk of a sharp equity correction as high.

Frequently asked questions#

What is an AI infrastructure ETF? A fund holding companies that supply the physical inputs to artificial intelligence: chips, servers, cooling, electrical equipment, power generation or data centre property. There is no standard definition, so two funds with similar names can hold entirely different companies.

Did AI infrastructure funds really beat Nvidia in 2026? Some did. To 31 August, SMH returned 54.56% and AIPO 26.35% against about 18% for Nvidia's share price. GRID returned 17.02%, slightly behind. NLR and XLU were well behind.

Why did utilities and nuclear funds lag if AI needs so much power? Regulated utilities earn returns set by regulators and must spend heavily on new capacity before they earn from it. Equipment suppliers are paid on contracts signed now. That is interpretation of the published returns, not a prediction that the gap persists.

How do I check what a fund actually owns? Every issuer publishes a daily holdings file and a monthly fact sheet on its own site. Read the top ten weights and the number of holdings before the performance chart.

Are these funds riskier than a broad index fund? They are more concentrated, usually in one or two sectors and often in a small number of companies, which raises both the potential gain and the potential loss. They also charge more.

Is it too late to buy this theme? Nobody can answer that, and anyone who claims to should be ignored. What is knowable: valuations across AI-linked equities are historically high by the Bank of England's measures, and the funds that ran hardest are the ones most exposed if spending slows.

Glossary#

Exchange-traded fund (ETF): a basket of investments that trades on an exchange like a single share, usually tracking a published index.

Expense ratio: the annual fee, expressed as a percentage of your holding, deducted from the fund's value rather than charged separately.

Net asset value (NAV): the per-share value of everything a fund owns, calculated once a day after the market closes.

Standardised performance: returns published in a regulator-mandated format at month end, which makes funds comparable.

Backlog: the value of orders a company has received but not yet delivered. A guide to future revenue, not a guarantee of it.

Capacity auction: a market in which a grid operator pays generators to be available at peak demand, separate from payment for the electricity itself.

Real estate investment trust (REIT): a listed company that owns income-producing property, in this case data centre buildings, and passes most of its rent to shareholders.

Terawatt hour (TWh): one billion kilowatt hours. A household electricity meter measures in kilowatt hours, so a terawatt hour is a billion of those.

This is reporting, not investment advice. Figures attributed to named sources are facts as those sources published them. Sentences that weigh those figures are interpretation, and are flagged as such.

References#

1. VanEck, VanEck Semiconductor ETF (SMH) performance, month-end standardised NAV returns to 31 August 2026; holdings and net assets to 9 September 2026.
2. Defiance ETFs, Defiance AI & Power Infrastructure ETF (AIPO), standardised performance to 31 August 2026, holdings to 10 September 2026, net assets to 9 September 2026.
3. First Trust, Nasdaq Clean Edge Smart Grid Infrastructure Index Fund (GRID) summary, standardised performance to 31 August 2026, net assets to 9 September 2026.
4. First Trust, GRID holdings, as of 9 September 2026.
5. State Street Global Advisors, SPDR S&P 500 ETF Trust (SPY), standardised performance to 31 August 2026.
6. State Street Global Advisors, Utilities Select Sector SPDR Fund (XLU), standardised performance to 31 August 2026.
7. VanEck, Uranium & Nuclear ETF (NLR) performance, standardised performance to 31 August 2026.
8. Macrotrends, Nvidia stock price history (2025 year-end close of $186.50), and StockAnalysis, Nvidia historical prices (31 August 2026 close of $220.78).
9. GE Vernova, Second quarter 2026 financial results, 22 July 2026 (orders, backlog, gas slot reservations, data centre orders).
10. Vertiv Holdings, Second quarter 2026 results, 29 July 2026 (net sales, guidance, chief executive quotation).
11. International Energy Agency, Energy and AI, executive summary, 2025 (415 TWh in 2024, 945 TWh by 2030, investment figures).
12. PJM Interconnection, Auction procures 134,311 MW of generation resources, 22 July 2025 (2026/2027 clearing price, prior-year price, peak load forecast).
13. State Street Global Advisors, ETF inflows set records in first half, 6 July 2026 (total flows, technology, industrials and utilities flows).
14. Bank of England, Financial Stability Report, July 2026 (correction risk, AI share of the S&P 500, CAPE yield).
15. Bloomberg, Clean energy ETF GRID draws $3.1 billion in inflows on AI power demand, 27 April 2026 (year-to-date inflows into GRID).
16. Yahoo Finance, NVIDIA Corporation (NVDA), quote data for 9 September 2026 (market capitalisation).
17. Google Ads search-volume data for the United States, retrieved via DataForSEO on 11 September 2026 (monthly volumes for "grid etf" and "ai infrastructure etf").