The Real AI Growth Story Has Nothing to Do With Chips
Listen to the audio version of this article (generated by AI).
Editor’s note: U.S. markets are closed today for the Juneteenth holiday. And the TradeSmith offices are also closed, including customer service.
Today, we’re sharing with you a recent interview our CEO, Keith Kaplan, did with MarketBeat’s Bridget Bennett about his AI chokepoints thesis.
Keith says the best way to profit from the AI boom isn’t to buy big software stocks like Microsoft and Google. It’s to buy stocks in companies that are solving critical chokepoints in the AI infrastructure buildout.
Below, he breaks down each chokepoint – from the advanced cooling systems that keep AI chips from overheating to the power lines that feed them – and names the companies set to capture the most value from each one.
Keith’s AI Chokepoints Q&A
Bridget Bennet: Everyone seems focused on AI chipmakers right now. But the real growth story is shifting rapidly toward the buildout of the physical infrastructure for AI.
With so much capital pouring into semiconductor companies, where are you looking to find the most significant growth in the market today?
Keith Kaplan: I’m looking at the physical world that has to wrap itself around those chips.
By the end of this year, just within the U.S., the four hyperscale giants —Microsoft, Google, Amazon, and Meta – are on track to spend more than $700 billion on AI infrastructure.
That’s almost four times what they spent back in 2024, and the figure is projected to top $1 trillion by 2027.
This is the largest sustained surge in industrial infrastructure spending since the post-World War II buildout. That’s how massive this is. The money isn’t going into software or marketing. It’s pouring into football-field-sized buildings, gigawatt-scale power systems, and transmission lines that do not even exist yet.
But this spending is running straight into a wall. Demand has completely outrun the world’s ability to manufacture, build, and approve, creating severe physical chokepoints across the entire AI supply chain.
Bridget: Talk a bit more about what you mean by chokepoints and where that friction is showing up the most.
Keith: Think of a water pipe. The narrowest part of that pipe restricts the total flow, no matter how much water is waiting upstream. Right now, we are trying to squeeze immense market demand through a very narrow pipe.
The mainstream media loves to tell the chip story, but it’s totally incomplete. To run high-end AI chips, you need specialized memory fast enough for AI tasks, power systems capable of handling hundreds of watts per chip, advanced thermal cooling to keep it from melting, a massive power plant nearby, and heavy-duty transformers to deliver the electricity.
I’m focusing on five specific points where the supply chain is breaking down under the weight of all this demand: high-bandwidth memory, silicon photonics, thermal management, power generation, and the electrical grid.
Industrial booms always play out this way. The picks-and-shovels suppliers – the ones building the physical infrastructure – often outperform the headline names once the initial gold rush cools.
Bridget: There are plenty of skeptical investors out there who argue that this entire AI theme has run too far, too fast. How much realistic runway do these infrastructure plays actually have left?
Keith: They have a tremendous amount of runway because people are misjudging the timing.
The AI trend is playing out in three distinct waves. The first wave was the chips, led by the massive runup in Nvidia from 2023 to 2025. I would argue that specific wave is mostly closed because a lot of that money is already priced into the stock.
Wave two is software, featuring companies like Salesforce and Adobe, which are actually facing headwinds right now because AI is shifting their traditional business models.
We’re now entering the third wave, which is the physical buildout. Look at the numbers. Memory has gone from making up just 8% of hyperscaler budgets in 2024 to roughly 30% of their total spend this year. That’s a four-fold shift in two years, an expansion we almost never see.
Global data center electricity demand is projected to double over the next few years. In the U.S. alone, data centers could account for 17% of all electricity consumption by 2030, compared to just 4% today.
This AI infrastructure wave could easily run for five years, outlasting the initial chip and software waves. Even if a stock is sitting at an all-time high, its long-term forward earnings are likely not fully baked in yet.
Bridget: Let’s break down your specific stock selections across these categories. What is the first major bottleneck on your radar?
Keith: The first critical bottleneck is right inside the server chassis itself, and that’s high-bandwidth memory, or HBM.
It stacks 12 layers of memory directly next to the AI chip – without it, these AI chips don’t work. The pick here is Micron Technology (MU). It designs and manufactures memory and storage chips, and it’s the only U.S. producer of high-bandwidth memory.
The company’s entire 2025 manufacturing output is sold out, and its 2026 capacity is already largely committed. As the industry transitions to advanced architectures like Nvidia’s Rubin chips, Micron’s competitive positioning is only going to strengthen.
Its annual revenue and net income are soaring, making it a fantastic three- to five-year hold. Some investors get nervous when a stock like Micron drops significantly on short-term market news, but if your investment horizon is long-term, those pullbacks are excellent opportunities to buy the dip. The underlying structural demand for its hardware is locked in.
Bridget: What is the second chokepoint you’re tracking?
Keith: That would be photonics and interconnects. AI chips have to move massive amounts of data to thousands of other chips at once. Copper connections can’t keep up. Silicon photonics – which transmits data as pulses of light rather than electrical signals – is the solution the industry is betting on.
My top selection in this space is Coherent (COHR), which is the leading global supplier of optical transceivers. These are the tiny, essential components that handle that light conversion in AI clusters.
To give you an idea of how central Coherent is to this ecosystem, Nvidia took a $2 billion strategic stake in the company. That piece of information tells you everything you need to know. Nvidia, the world’s dominant AI chip designer, is anchoring itself to Coherent. Investors should probably pay close attention to them, too.
Bridget: Your third category is thermal management. Just how severe is the heat issue inside these massive new data centers?
Keith: It’s mind-boggling, and it comes down to basic physics. Top-end AI chips can now draw up to 1,200 watts of power each, whereas a high-end consumer gaming chip draws a fraction of that.
When you pack 72 of these chips into a single server rack, that tiny physical footprint throws off the same amount of heat as a small apartment building.
Traditional air cooling is dead – it just doesn’t work at this scale. The industry is being forced to shift to direct-to-chip liquid cooling, which carries heat away roughly 3,500 times more efficiently than air at the same flow rate.
The dominant player here is Vertiv (VRT). It’s the leading supplier of coolant distribution units, which are about the size of a refrigerator and pump liquid coolant directly through these dense AI server racks.
Vertiv already supplies most of the large hyperscaler buildouts globally, and it’s highly profitable. So pullbacks in this stock will be fantastic entry points for a company with this kind of institutional backing.
Bridget: Let’s transition to the energy story, which has captured a lot of attention recently. A single cutting-edge data center can consume as much electricity as a mid-sized city. Where is that power going to come from?
Keith: AI computing requires uncompromised, constant baseload power 24 hours a day, 7 days a week. Intermittent renewable sources like solar or wind can’t back up a critical data center cluster without massive battery backup systems that do not exist yet.
That leaves the market with two primary choices: natural gas and nuclear energy. To play this, I really like Constellation Energy (CEG). It’s the largest nuclear operator in the U.S. with 21 active reactors.
A decade ago, the primary buyers of nuclear energy were public utility companies. Today, the biggest buyers are trillion-dollar software companies. Microsoft signed a historic 20-year deal to bring the Three Mile Island nuclear plant back online exclusively to power their data centers.
Amazon executed an $18 billion long-term deal with nuclear power producer Talen Energy for similar reasons.
Every single gigawatt of new baseload demand flows directly to Constellation’s bottom line. Investors sometimes worry that nuclear projects take too long to build. But because Constellation is reviving existing infrastructure and signing direct corporate deals with Big Tech, it’s positioned to capture this revenue much faster than most people realize.
Bridget: Even if a company generates the electricity, getting it to the facility is a massive problem. What is happening with the electrical grid right now?
Keith: This is arguably the most severe chokepoint of all, and the data behind it really got me excited about this trade.
Even if you build a dedicated power plant tomorrow, the current regulatory queue means a new project might wait five to eight years just to get a physical hookup to the grid.
The manufacturing backlogs for grid equipment are wild. Industrial transformers that used to ship in 12 months now take an average of two and a half years to deliver. Heavy-duty gas turbines are taking up to seven years to ship.
In high-density tech hubs like Northern Virginia, local utilities are explicitly warning that new projects won’t see a grid connection until late in the decade or later.
To capitalize on this, you look at a company like Eaton Corporation (ETN). Eaton is a power management giant that has been around for over 100 years. It makes the transformers, the switchgear, and the power distribution units that connect a data center to the electrical grid.
Data centers have rapidly become Eaton’s fastest-growing market. When a company has a multi-year backlog for products the world completely relies on to build out the future, that’s when you want to buy it.
Bridget: It’s fascinating to look at the physical architecture required to back up the digital revolution. Keith, thank you so much for breaking down the thesis for us today.
Editor’s note: This transcript was lightly edited for flow and clarity.