The Underlying Growth in AI Infrastructure
Interestingly, the current market growth isn’t focusing on chipmakers but rather on the layers beneath them.
Keith Kaplan, CEO of TradeSmith, has been exploring what he describes as the “challenges” in developing AI—essentially, the physical limitations faced by a sector grappling with surging demand. He believes the real value in this AI boom lies not with the obvious players but with the companies critical to their operations.
A Massive Financial Challenge
By the end of 2025, four major U.S. tech giants—Microsoft, Google, Amazon, and Meta—are anticipated to invest over $700 billion in AI infrastructure. NVIDIA predicts this investment might balloon to $1 trillion by 2027, implying a potential economic impact of $3–4 trillion within three to five years.
Much of this funding isn’t directed toward software. Instead, it’s funneled into enormous facilities, vast power systems, and infrastructure that is still in the works. Visualize demand being funneled through a narrow pipe; the total available water upstream may be vast, yet the flow remains constrained at the bottleneck.
Kaplan pointed out five significant challenges facing this growth, along with their respective stocks.
Memory: The Core of AI Functionality
High Bandwidth Memory (HBM) is crucial for GPU functionality. It’s worth noting that Micron Technology is the only HBM manufacturer in the U.S., and its entire 2025 production capacity was sold out before the year even began. The forecast for 2026 production is already nearly set.
Micron’s role in the supply chain is expected to strengthen as NVIDIA shifts its chips from HBM3E to HBM4. The company’s annual revenue is projected to reach $58 billion, with a net income of $24 billion. Kaplan anticipates this stability will hold for the next few years, and he’ll be watching the upcoming earnings report on June 24th closely.
Photonics: Speeding Up Communication
In AI hardware, after data is transmitted from one chip, it must communicate with many others. Traditional copper wires can’t accomplish this efficiently; the answer lies in silicon photonics, where light travels through fibers at incredible speeds unachievable by copper.
Coherent Corp produces the optical transceivers that facilitate this, converting electrical signals to light pulses and back. NVIDIA recently acquired a $2 billion stake in Coherent, which is seen as pivotal for AI’s next phase.
Heat Management: Keeping Chips Cool
The most advanced AI chips can draw up to 1,200 watts. Installing 72 of these chips can emit the same heat as a small apartment within the confines of a refrigerator. Air cooling won’t suffice here; direct liquid cooling is necessary for efficient heat transport.
Vertiv Holdings provides cooling distribution units for many large hyperscalers. With a market cap over $100 billion and annual revenues exceeding $10 billion, it’s a substantial player. Kaplan sees the recent drop in stock price as a potential entry point before the earnings report on July 29th.
Power Generation: Return of Nuclear
AI data centers consume as much power as medium-sized cities. For instance, Meta’s Hyperion facility is currently operating at 2 gigawatts and is set to expand to 5 gigawatts. To sustain such demands, natural gas and nuclear energy are becoming increasingly essential.
Constellation Energy, the largest nuclear operator in the U.S., is making significant moves—like signing a 20-year agreement with Microsoft to restore the Three Mile Island plant for substantial capacity. This shift also highlights how tech companies are now among the largest consumers of nuclear power, a notable change from a decade ago.
Despite flat stock prices, Kaplan interprets this as preparation for the company’s early August earnings report.
The Grid: Completing the Connection
After power plants are constructed, it can take years to connect them to data centers. For example, transformer lead times have increased from 12 months to 2.5 years. Some utility projects in Northern Virginia won’t see grid connections until at least 2028.
Eaton Corporation has been addressing these challenges for over a century, providing the infrastructure needed to link data centers to power grids. Data centers have become Eaton’s fastest-growing market, indicating significant future promise.
Early Stages of Megatrends
In just two years, spending on memory has jumped from 8% to 30% of hyperscalers’ budgets. By 2030, U.S. data centers could account for 17% of national electricity consumption, a striking increase from the current 4%.
This raises a question for investors: are we missing out? Kaplan asserts that the infrastructure isn’t built yet, so there’s still opportunity in the gap.
Mentioning Micron Technology, it’s noted that while analysts rate it positively, they also recommend other stocks that may be more advantageous at this time.
