The AI Buildout Needs Power It Cannot Get: 4 Energy Stocks Already Have It

Key Points

• Stansberry Research analyst Gabe Marshank argues the AI data center buildout cannot earn its cost of capital once compute pricing resets, putting neocloud business models like CoreWeave, Nebius Group and IREN at risk
• Atlas Energy Solutions and Liberty Energy are building behind-the-meter power that lets data centers skip a five to seven year grid interconnect wait
• Ormat Technologies and Fervo Energy anchor an enhanced geothermal story where the open question is engineering execution, not demand

Every valuation in the AI infrastructure trade rests on one quiet assumption: that compute stays scarce enough to keep prices high. That assumption is already showing cracks, and most investors haven't repriced for it.

Gabe Marshank, lead analyst of Market Maven at Stansberry Research, is not an AI skeptic. He uses the technology daily and calls the product brilliant. His argument is narrower and harder to wave off. The technology works. The business model behind the buildout may not. And if he's right, the capital chasing data centers is headed somewhere very different from where the capital chasing power is headed.

The Return-on-Capital Math Behind a $4.5 Trillion Buildout

Marshank puts the data center buildout at roughly $4.5 trillion by 2030. On his numbers, the compute layer would need to generate something near $1.7 trillion in annual revenue to earn an adequate return on that capital.

For scale, he pegs the existing platform economy—Alphabet NASDAQ: GOOGL, Microsoft NASDAQ: MSFT, Amazon.com NASDAQ: AMZN, Meta Platforms NASDAQ: META and Oracle NYSE: ORCL, stripping out consumer products—at roughly $1 trillion in revenue today. So the ask is a standing start to $1.7 trillion in four years, just to break even on the spending.

That matters because the market is treating capital expenditure as if it were value created. Marshank has seen that movie. He points to the biodiesel buildout, where plants were valued at a multiple of the money sunk into the ground until commodity economics showed up and the equity went to zero.

AI Compute Pricing Is Already Rolling Over

His core claim is that data centers produce a commodity: AI compute. Supply of that commodity is set to rise enormously, and capitalism is likely to do what it always does when supply surges.

The key detail is that compute doesn't idle. Operating costs run near 3% of a data center's total capital cost, so the machines run regardless. Supply equals demand, and AI compute pricing is the clearing mechanism, the way an airline sells the last seat at whatever price travelers are willing to pay, rather than flying it empty.

Marshank tracks published rate cards. By his count, prices were about $3.50 per kilowatt in 2022, climbed to roughly $4 to $4.25 as AI demand took off, and are now beginning to correct even as demand headlines stay loud. Long-term contracts muddy the picture with de-escalators and discounts to list, but he's convinced prices are headed down.

The comparison he keeps returning to is the bandwidth bubble of 1996 to 2001. Capital went into the ground, supply came on, pricing collapsed roughly 90%, and Level 3, Global Crossing and WorldCom blew up on revenue that never matched the model.

Neocloud Stocks and the Missing Moat

This is where he gets blunt about neocloud stocks. He sees returns going negative at CoreWeave NASDAQ: CRWV, Nebius Group NASDAQ: NBIS and IREN Limited NASDAQ: IREN, and he struggles to identify what differentiates them.

His framing: the business model is spending itself. If a neocloud can commit $50 billion and Alphabet can commit far more, scale alone isn't a moat. He notes that IREN's $9.7 billion GPU cloud agreement with Microsoft, disclosed in a securities filing, implied returns he'd describe as merely adequate in the hottest market anyone has seen.

The risk, in his telling, is simple. The capital was spent on the belief that rising volume lifts revenue. If volume rises and pricing falls, revenue never meets expenses.

Behind-the-Meter Power Offers a Shortcut Around the Grid

The bullish half of his case sits in energy, where the AI bottleneck is real and physical. Two names come straight out of the oil patch: Atlas Energy Solutions NYSE: AESI and Liberty Energy NYSE: LBRT.

Both are building behind-the-meter power, meaning on-site generation that sidesteps the five-to-seven-year wait in the utility interconnect queue.

The U.S. grid is slow to absorb new sources, because every addition has to be balanced.

Pressure pumping taught these companies how to deploy enormous horsepower on short notice. Pointing it at a data center is the same skill set with a better customer.

Geothermal Stocks Trade on an Engineering Question

Liberty Energy also stepped sideways into geothermal under founder Chris Wright, now U.S. energy secretary. Marshank calls geothermal the most overlooked energy source in the world, and his two names are Ormat Technologies NYSE: ORA and Fervo Energy NASDAQ: FRVO.

Ormat brings decades of legacy assets. Fervo is the pure play, applying horizontal drilling and multi-stage completions to create enhanced geothermal systems. Its Cape Station project in Beaver County, Utah is permitted for up to 2 gigawatts, and Google has signed a 396-megawatt power purchase agreement there with an option for roughly 600 megawatts more, sitting inside a broader 3-gigawatt framework running through 2033.

The stock trades below its $27 May IPO price, pressured by interconnect delays that pushed revenue recognition to the right. Marshank shrugs at that. The real question, he says, is engineering: whether a closed-loop system holds the water it pumps down. If it does, the upside is large. If water starts leaking out of the wells, that's the sell signal.

Compare the clock to nuclear, which he was early and right on years ago. New nuclear takes five to seven years minimum in the United States. Geothermal proof points could arrive within two.

What Should Actually Change Your Mind

Marshank doesn't use stop-losses, and he doesn't trade on headlines. Macro inputs are the most efficiently priced information in the market, so an edge is far easier to find in a single security than in the bond market. His horizon runs three to five years, looking for three-to-five-times outcomes.

The risk is that pricing in geothermal and distributed power disappoints the same way he expects it to disappoint in compute. The upside is that both sit on physical scarcity rather than on a commodity whose supply is multiplying.

His discipline is worth borrowing regardless of whether you buy the call: decide in advance what would make you sell, then watch that—not the tape. Because if your only exit plan is panicking after a 50% drawdown, you may as well panic now. For investors who want to run that kind of screen themselves, Marshank's team has built an AI-powered research tool for finding asymmetric setups at Stansberry Research.