AI’s Picks and Shovels Problem: The Energy Companies That Can’t Afford a Downturn

It’s a clean narrative, and NVIDIA’s balance sheet backs it up. Over $26 billion in cash and equivalents. Gross margins above 70%. No debt dependency. If AI demand slows, NVIDIA sells fewer chips and margins compress, but the company survives comfortably. They’ve been through cycles before.

But NVIDIA isn’t actually the deepest infrastructure dependency in the AI buildout. GPUs are important, but they’re portable, fungible, and produced by a company with a fortress balance sheet. The real picks and shovels — the layer everything else depends on, the one with no substitutes and no fallback — is power. Electricity. The companies generating, transmitting, and contracting the energy that keeps data centers running.

And those companies look nothing like NVIDIA.

Constellation Energy is the most visible name in the AI power story. The company operates the largest fleet of nuclear plants in the United States and has signed a 20-year power purchase agreement with Microsoft to restart the Three Mile Island Unit 1 reactor specifically for data center demand. The stock has roughly tripled since 2023 on the AI power narrative. But Constellation carries approximately $5.5 billion in long-term debt against operating cash flow that has historically fluctuated between $2–3 billion annually. The company is profitable and investment-grade, but it’s not a cash fortress. A sustained downturn in data center demand that leads to PPA renegotiations or cancellations would hit revenue while the debt obligations remain fixed.

Vistra Corp tells a more concentrated story. The Texas-based power generator has aggressively positioned itself for AI demand, and the stock has been one of the best performers in the S&P 500. But Vistra carries roughly $12 billion in total debt. Its natural gas and nuclear fleet is being re-contracted toward data center customers at premium rates. The bull case assumes those premium rates persist. The bear case asks what happens when data center buildout slows, power demand plateaus, and Vistra is left with a debt load sized for a growth trajectory that didn’t materialize. The company’s debt-to-EBITDA ratio leaves meaningfully less cushion than what you’d see at the hyperscalers funding the demand side.

Talen Energy is where the risk profile gets sharp. The company emerged from bankruptcy in 2023 and has since repositioned around data center power, including a direct supply arrangement with Amazon’s data center operations at its Susquehanna nuclear plant in Pennsylvania. Talen carries significant debt from its restructuring, and its revenue is becoming increasingly concentrated in a small number of large data center contracts. The company’s liquidity position — the cash and credit available to weather a downturn — is a fraction of what the hyperscalers sitting on the other side of those contracts hold. If Amazon scales back, delays, or renegotiates, Talen has limited room to absorb the impact.

NRG Energy, another major player positioning for AI power demand, carries over $8 billion in long-term debt. The company has been investing in generation capacity and grid infrastructure to serve data centers, but its balance sheet reflects a legacy utility model: asset-heavy, debt-funded, with margins that depend on long-term contract stability. NRG’s interest coverage is adequate in a stable environment but compressed compared to the technology companies driving the demand. The gap between NRG’s financial resilience and Microsoft’s financial resilience is enormous, yet both sit on opposite sides of the same contract.

The structural problem is asymmetry. When a hyperscaler signs a 15-year power purchase agreement with an energy company, both sides are betting on the same future: that data center demand will grow steadily for decades. But the consequences of being wrong are distributed unevenly. If demand slows, Microsoft can delay a data center buildout, write down some infrastructure costs, and absorb the hit against $80 billion in cash. The energy company on the other side of that PPA has sized its debt, its capital expenditure, and its generation capacity against projected revenue that may not arrive. They don’t have $80 billion in cash. They have debt covenants.

This asymmetry creates a specific kind of fragility. The energy companies are making irreversible capital commitments — restarting nuclear reactors, building natural gas plants, upgrading transmission infrastructure — based on demand signals from customers who can change their minds. The power contracts have terms and conditions, but contract renegotiation in a downturn is a well-documented pattern in energy markets. When your largest customer has more lawyers and more leverage than you do, the contract is only as strong as the customer’s continued willingness to honor it at the original terms.

The market is pricing these energy companies as AI winners. And they may be — if the buildout continues at its current pace for the duration of their contract terms. But the same market is not pricing the scenario where AI infrastructure spending decelerates and these companies are left holding generation capacity they built for demand that slowed down. In that scenario, the companies with the least liquidity break first. Not NVIDIA. Not Microsoft. The power companies.

This is where the AI Infrastructure Stress Index framework applies. Our credit spreads signal tracks the cost of borrowing for AI infrastructure broadly. As energy companies take on more debt to build generation capacity for data centers, their borrowing costs become part of the same credit ecosystem. If AI infrastructure credit spreads widen further, refinancing gets more expensive for exactly the companies that can least afford it. The GPU spot pricing signal matters here too: falling GPU prices signal oversupply, which signals potential data center demand deceleration, which flows downstream to the energy contracts those data centers depend on.

The picks and shovels of the AI buildout are real. But they’re not sitting in NVIDIA’s cash-rich balance sheet. They’re in the balance sheets of energy companies that are betting everything on a demand curve they don’t control. If the AI infrastructure cycle turns, these are the companies where the stress shows up first — and the market isn’t pricing that risk.