The AI Boom Can Buy Every Chip It Wants. It Still Has to Wait for Constellation's Nuclear Plants.

You can buy every GPU you want. You can lease data center space. You can even get a building permit in some states in under a year. But connecting a power plant to the grid now takes 40 months in the markets where AI data centers are being built.

That is why Constellation Energy, a nuclear power company that most investors would call a utility, has quietly signed some of the most consequential contracts in the AI build-out.

Constellation operates the largest nuclear fleet in the United States — roughly 21 gigawatts of zero-carbon capacity already built, already licensed, and already connected to the grid. Nuclear plants run at more than 90 percent of their theoretical maximum. You cannot simply turn them up. But you can sell what they produce to someone who cannot build a replacement fast enough.

Microsoft is that someone. In September 2024, Microsoft signed a 20-year agreement with Constellation to purchase 835 megawatts from Three Mile Island Unit 1, a plant that had been shut down in 2019 because it was losing money. Constellation is restarting it. Microsoft will buy every megawatt-hour it produces. The contract, valued by some estimates at roughly $16 billion over its life, is Constellation's largest power purchase agreement ever.

Then in June 2024, Meta signed a separate 20-year deal for 1,121 megawatts from Constellation's Clinton plant in Illinois. In the second quarter of 2026 alone, Constellation locked up approximately another one gigawatt of new nuclear contracts, with an average duration of 18.5 years.

These are not symbolic letters of intent. They are long-term power purchase agreements with investment-grade technology companies willing to pay a premium for the only energy source that is firm, 24/7, and carbon-free.

The physical wall

AI data centers are not energy-efficient. A single cluster designed for artificial intelligence can consume as much electricity as a small town. U.S. data center peak demand is projected to nearly triple, with AI workloads alone accounting for up to 40 percent of data center power demand by 2026. BloombergNEF sees a shortfall of 19 gigawatts between what the grid can deliver and what data centers need by 2035.

Wind and solar cannot fill this gap on their own. They produce power intermittently, and AI racks need electricity whether the sun is shining or the wind is blowing. Natural gas is dispatchable and cheap to build, but it carries emissions that violate the sustainability commitments every major technology company has made.

Nuclear is the only source that is simultaneously firm and carbon-free. But new nuclear plants cost between $6,400 and $12,700 per kilowatt to build — five to ten times what natural gas costs. Even with regulatory acceleration, new reactors will not arrive until the early 2030s at the earliest. Small modular reactors, which have drawn the most startup interest, are years from first commercial deployment.

Goldman Sachs estimates that massive new nuclear capacity would be needed by 2030 to meet data center demand. Less than 10 percent of that is expected to be available globally by then.

Capital arrived at the problem. Nuclear capacity did not.

Who owns the constraint

Constellation emerged from Exelon in 2022, inheriting its nuclear fleet. It was a pure-play nuclear company with about 21 gigawatts of operating reactors. In January 2026, it acquired Calpine for approximately $22 billion, adding natural gas and geothermal capacity to create the nation's largest private-sector power producer. The combined company — before regulatory-required divestitures — is the largest private-sector power producer in the United States, with roughly 55 gigawatts of total capacity.

The nuclear side is the one that matters for AI. Here is why:

Constellation's plants are already interconnected. They have grid connections that bypass the 40-month interconnection queue that now strangles new generation. They run at a 92 percent capacity factor, above 90 percent of their theoretical maximum. You cannot squeeze more output from a nuclear reactor without filing for an uprate — which Constellation is doing, adding 30 megawatts at Clinton as part of the Meta deal.

The plants are also the ones hyperscalers cannot replicate themselves. A technology company can build a data center in two years. It cannot obtain a Nuclear Regulatory Commission license, design, construct, and commission a nuclear plant in less than a decade. The bottleneck is regulatory and physical, not financial.

Being necessary gets you orders. Being scarce decides who keeps the money.

The economics of scarcity

Constellation reported second-quarter 2026 revenue of $7.5 billion and adjusted operating earnings per share of $2.55, and raised its full-year 2026 guidance to between $11.50 and $12.50 per share. For the first quarter of 2026, operating margins expanded to 22 percent — up from 7 percent a year earlier, though a portion of that improvement reflects Calpine's high-revenue output landing on a cost structure that had not yet fully scaled.

Management projects more than 20 percent base earnings growth between 2026 and 2029, excluding the additional upside from contracts that have not yet started delivering revenue. The nuclear production tax credit, which grows with inflation, adds a floor under margins.

But there is a gap between the deals and the cash.

Constellation estimates that roughly 30 to 40 percent of its earnings framework still depends on spot power prices — what analysts call merchant exposure. The Microsoft and Meta contracts are locked, but many of the new agreements signed in 2026 do not begin delivering revenue until 2029. Until then, Constellation is carrying debt, funding restarts, and integrating Calpine while earning a significant portion of its income at whatever the wholesale market pays.

That is why the stock has pulled back roughly 22 percent from its 52-week high of $413 and is down 21.7 percent year-to-date, trading near $277. The market is not losing faith in the nuclear thesis. It is pricing a waiting room.

The balance sheet question

The Calpine acquisition was a bet that the AI power shortage would be broad, not just nuclear-specific. Natural gas is the fastest source of dispatchable power to deploy, and Constellation needed it to fill the gap between today and 2029.

It was also expensive. Total debt sits at approximately $66 billion, with $700 million in cash, for a debt-to-equity ratio of 76 percent. Free cash flow over the trailing twelve months was just $309 million — positive, but a fraction of the $3.9 billion the company spent on capital expenditures. Operating cash flow was $4.2 billion, which covers debt service but leaves little margin for error if power prices fall, execution stumbles, or the restart costs more than planned.

Compare this with Vistra, the closest peer in the AI power trade. Vistra also signed a 20-year, 2.6 gigawatt nuclear deal with Meta. It spent $4.7 billion to acquire Cogentrix and its 5.5 gigawatts of natural gas plants. But Vistra's balance sheet carries higher leverage — 357 percent debt-to-equity — with only $5.5 billion in total equity against $37 billion in debt. Vistra generates more free cash flow at $1.1 billion over the trailing twelve months, but its revenue growth of 3.8 percent is a fraction of Constellation's 26 percent.

Both companies are spending heavily to position themselves. The question is whether the contracted revenue that arrives in 2029 justifies the debt and capex carried between now and then.

What the stock price is telling you

At $277, Constellation trades at roughly 28 times trailing earnings and 51 times forward earnings. That forward multiple reflects the market's view that earnings in the next twelve months do not yet include the new contracts. By that same logic, the stock looks expensive for what the company is earning today.

Peer comparison shows the premium. American Electric Power, a traditional regulated utility, trades at 21 times trailing earnings and yields 3 percent in dividends. Constellation's dividend yield is less than 0.6 percent. Investors are not buying this stock for income. They are buying a growth option on nuclear scarcity, priced at a growth-company multiple for a business that is still 30 to 40 percent merchant generator.

The valuation clock matters here. If contracted revenue arrives on schedule, margins hold at current levels, and the nuclear premium persists, the stock has a long runway. If power prices fall, restarts delay, or hyperscalers find a workaround, the forward multiple compresses and the debt burden becomes the dominant risk.

The capacity cure

Every scarcity eventually attracts supply. For Constellation, the relief comes from three directions.

First, the interconnection queue. In June 2026, the Federal Energy Regulatory Commission ordered six major grid operators to explain how they would connect data center power faster. The orders put a 60-day clock on grid rules. If interconnection timelines compress, new gas and renewable plants reach the market sooner, increasing competition for the load that Constellation's plants currently monopolize through their existing connections.

Second, small modular reactors. Companies like Kairos Power, TerraPower, and Oklo are in various stages of licensing and design. None will ship power before the early 2030s, and first-of-a-kind execution risk is enormous. But if one succeeds and scales, it introduces new nuclear capacity that does not require buying existing plant output.

Third, customer insourcing. Amazon has invested over $20 billion to convert a nuclear site into an AI campus. Microsoft is investing alongside BlackRock and Global Infrastructure Partners in energy infrastructure. Hyperscalers are learning that the energy companies are themselves the bottleneck. If they bypass the middle layer and own or co-locate at nuclear plants, the PPA model becomes less durable.

None of these alternatives arrives quickly. But a hidden winner without a capacity cure is not an investment thesis — it is a late theme trade.

Where the story goes from here

Constellation Energy is the purest public-market play on nuclear scarcity in the AI power trade. It owns operating plants, it has signed multi-decade contracts with the companies building the most AI, and it is restarting dormant capacity to capture premium pricing from customers who have no alternative.

The tension is timing. The contracts that make this story compelling do not begin delivering revenue until 2029. In the interim, Constellation is a highly leveraged company earning a significant portion of its income from volatile wholesale markets. The stock has already fallen from its highs as investors weigh the gap between today's merchant exposure and tomorrow's contracted certainty.

The confirmation signal: watch quarterly revenue growth and operating margin expansion as the Calpine integration completes. If Constellation maintains margins above 20 percent while adding contracted nuclear capacity each quarter, the thesis holds. The normalization signal: if new interconnection reforms cut queue times below 24 months, or if a competitor secures a multi-gigawatt nuclear PPA at similar terms, Constellation's scarcity premium begins to erode.

The famous products get the headlines. The plants that were already built — and already connected — decide how much power the AI economy can actually use.