The AI Power Trade: What's Inside the Nuclear, Uranium, and Data-Center ETFs

Holdings data as of August 6, 2026, straight from each fund's reported book. Weights are what the ETFs actually hold, not estimates.

The AI story everyone knows is about chips. Buy NVDA, buy SMH, ride the compute boom. Fair enough. But the bottleneck that keeps data-center operators up at night isn't silicon. It's power, and the physical space to put the machines. A single large AI data center can draw as much electricity as a small city, and you can't train a model on hardware you can't plug in. That's the second-order trade, and three ETFs let you own it. They hold wildly different things.

Think of the AI buildout in three layers. The chips (SMH). The power to run them (nuclear and uranium, through NLR and URA). And the buildings that house them (SRVR, the data-center landlords). Here's what's actually inside the power and real-estate layers.

NLR: the diversified nuclear bet

NLR, the VanEck Uranium and Nuclear Energy ETF, is the closest thing to a one-ticker bet on the whole nuclear economy. It's spread evenly, with no single name above 8.6%. The top of the book: Cameco at 8.52%, Constellation Energy at 8.39%, BWX Technologies at 6.93%, and Public Service Enterprise Group at 6.92%.

What makes this an AI-power trade and not just an energy fund is the mix. Constellation is the utility that struck a deal to restart Three Mile Island to feed a hyperscaler's data centers. It's the poster child for the entire thesis. Below the utilities sit the fuel and reactor builders, and then the small modular reactor names the market is betting on for the next wave: Oklo at 5.22% and NuScale Power at 4.69%. Uranium miners round it out. NLR isn't a bet on one company. It's a bet that the AI buildout drags the whole nuclear supply chain up with it.

URA: the concentrated uranium bet

URA, the Global X Uranium ETF, sounds similar and behaves nothing like it. Where NLR spreads the money, URA piles it into one name: Cameco is 24.6% of the fund. One in four dollars. After that it thins out fast, to Oklo at 6.49% and NexGen Energy at 6.09%.

URA is a bet on the uranium commodity itself and the miners who dig it, and a big chunk of that book is listed overseas, from Kazakhstan's Kazatomprom to Australian and Canadian miners. If the uranium price runs, URA runs harder than NLR. If Cameco stumbles, you feel it more than you'd want to for something that's supposed to be a basket. Same theme as NLR, very different risk. One is diversified nuclear, the other is a concentrated commodity bet wearing a nuclear label.

One name worth flagging: Oklo shows up in both funds, 5.22% of NLR and 6.49% of URA. Own both ETFs thinking you're diversifying and you're stacking the same small-reactor bet twice. Cameco too, under two different tickers: it's CCJ inside NLR and its Toronto listing inside URA.

SRVR: owning the actual buildings

SRVR, the Pacer Data and Infrastructure REIT ETF, owns the physical layer as real estate: the data centers and the towers. It's the most top-heavy of the three. Three landlords make up 44% of the entire fund: Equinix at 15.03%, Digital Realty at 14.98%, and American Tower at 14.29%.

This is the least glamorous way to play AI and maybe the most direct. Equinix and Digital Realty are the two largest data-center landlords on earth. If AI needs more compute, it needs more square footage, in more buildings, with more power and cooling, and these are the companies that own it and rent it back. The rest of the fund adds cell towers (Crown Castle, SBA Communications), a records-storage business turned data-center operator (Iron Mountain), and a handful of overseas names. But make no mistake, SRVR is a concentrated bet on three landlords.

How the three layers fit together

Put them side by side and the concentration tells the story. SRVR jams 44% into three names. URA puts a quarter of the fund in Cameco alone. NLR is the balanced one, nothing over 8.6%. Same macro thesis, three different ways to express it, three different risk profiles.

And they barely overlap. Chips in SMH, power in NLR and URA, buildings in SRVR. Own all three and you've covered the AI buildout end to end, from the processor to the reactor to the loading dock. Own one and you've made a specific call about where the bottleneck, and the returns, actually land. Just know which layer you're buying.

Want to go deeper? Here's what power density means for AI data centers, and the full ETF library to compare these against semis, robotics, and clean energy.