5 Stocks Winning the AI Race While Everyone Watches NVIDIA

Key Points

• Keith Kaplan of TradeSmith identifies five physical "choke points" in the AI build-out—high bandwidth memory, photonics, thermal management, power generation, and grid transmission—where demand is already outrunning the world's ability to supply.
• The four largest U.S. hyperscalers are on track to spend more than $700 billion on AI infrastructure this year alone, with global data center electricity demand projected to more than double by 2026.
• Despite significant recent gains, each of the five featured stocks still has demand fundamentals that haven't been fully priced in, with multi-year backlogs and committed capacity pointing to a sustained growth runway.

The biggest growth stories in the market right now aren't happening with the chip makers. They're happening one layer beneath them.

Keith Kaplan, CEO of TradeSmith, has spent recent months mapping what he calls the "choke points" of the AI build-out—the physical bottlenecks where trillion-dollar demand is running into a world that can't supply fast enough. His argument: the largest fortunes of this AI cycle won't go to the visible players. They'll go to the companies that those players can't function without.

The $700 Billion Problem

By the end of 2025, the four largest U.S. hyperscalers—Microsoft NASDAQ: MSFT, Alphabet NASDAQ: GOOGL, Amazon NASDAQ: AMZN, and Meta Platforms NASDAQ: META—are expected to spend more than $700 billion on AI infrastructure. That number climbs toward $1 trillion by 2027 and, according to projections shared by NVIDIA NASDAQ: NVDA during its most recent earnings call, could reach $3 to $4 trillion in total economic impact within three to five years.

That money isn't going into software. It's going into football-field-sized buildings, gigawatt-scale power systems, and transmission lines that don't exist yet. Think of it as demand being squeezed through a very narrow pipe: no matter how much water is upstream, the flow is whatever the choke point allows.

Kaplan identifies five of those choke points, and a stock for each.

Memory: The Chip Inside the Chip

High-bandwidth memory (HBM) is what allows a GPU to actually function.

Without it, AI chips don't work. Micron Technology NASDAQ: MU is the only American producer of HBM, and its output for 2025 was completely sold out before the year began. Capacity for 2026 is already largely committed.

As NVIDIA transitions its Rubin chips from HBM3E to HBM4, Micron's position in the supply chain only strengthens. Annual revenue is tracking toward $58 billion, and net income is already at $24 billion. Kaplan sees this as a three-to-five-year hold, with the next earnings report on June 24 as an early signal of where demand is heading.

Photonics: The Speed of Light Between Chips

When data leaves one chip, it has to communicate with thousands of others, and copper wire can't keep up. The solution is silicon photonics: light passed through fiber at speeds copper simply can't match.

Coherent Corp. NYSE: COHR is the leading supplier of the optical transceivers that make this happen, converting electrical signals into pulses of light and back again at every connection in an AI cluster. NVIDIA took a $2 billion stake in Coherent earlier this year—a signal, Kaplan argues, of just how central the company is to the next phase of AI build-out.

Thermal Management: Keeping It From Melting

Top-end AI chips now draw up to 1,200 watts each.

Put 72 of them in a rack, and you've got the heat output of a small apartment in a space about the size of a refrigerator. Air cooling can't handle it. Direct-to-chip liquid cooling can carry heat 3,500 times more efficiently than air at the same flow rate.

Vertiv Holdings NYSE: VRT already supplies most of the large hyperscaler build-outs with its cooling distribution units. With a market cap above $100 billion and annual revenue over $10 billion, it's not a speculative name—and a recent pullback of more than 10% over the week before Memorial Day is what Kaplan would view as an attractive entry point ahead of its July 29 earnings.

Power Generation: Nuclear Is Back Online

A single AI data center now consumes the power of a mid-sized city. Meta's Hyperion facility is currently operating at 2 gigawatts and is expected to scale to 5. Meeting that kind of baseload demand 24 hours a day, seven days a week means natural gas and nuclear. Both are booming.

Constellation Energy NASDAQ: CEG is the largest nuclear operator in the U.S., with 21 reactors and a 20-year deal signed with Microsoft in September 2024 to bring Three Mile Island back online. Constellation is investing $1.6 billion to revive the plant and deliver 835 megawatts of dedicated capacity.

The shift in who's buying nuclear power tells the story: a decade ago, the biggest buyers were utilities. Today, they're software companies. Amazon's 17-year, $18 billion power purchase agreement with Talen Energy NASDAQ: TLN underscores just how aggressively Big Tech is moving to lock up baseload supply.

The stock is flat over the past three months, a consolidation Kaplan views as a setup ahead of early August earnings, not a sign the thesis has broken.

The Grid: The Last Mile That Takes the Longest

Even after the power plant is built, getting electricity to the data center can take years. Transformer lead times have stretched from 12 months to two and a half years. Heavy gas turbines are running up to seven years out. In some northern Virginia utility territories, grid hookups for projects filed after 2024 won't be available until at least 2028.

Eaton Corporation NYSE: ETN has been solving exactly this problem for more than 100 years. Transformers, switchgear, power distribution—the infrastructure that connects every data center to the actual grid.

Data centers are now Eaton's fastest-growing end market, and the company is sitting on multi-year backlogs with a $148 billion market cap and nearly $30 billion in annual revenue. The stock has been largely flat for the past three months, which Kaplan reads as quiet accumulation. Next earnings: August 4.

A 5-Year Mega Trend Still in Its Early Innings

Memory spend as a share of hyperscaler budgets has shifted from 8% to 30% in just two years. U.S. data centers could account for 17% of all national electricity consumption by 2030, up from 4% today. These aren't projections being whispered—they're numbers already reshaping capital allocation across the economy.

The question most investors are asking is whether they've missed it. Kaplan's view: not even close. The stocks may be up, but the infrastructure hasn't been built yet. That gap is exactly where the opportunity lives.