Google Just Locked In 20 Years of Nuclear Power. Electricity Is the AI Boom's Scarcest Resource.

Google has a power problem, and this week it decided to buy its way out of it. On October 6, the company signed a deal with Constellation Energy that will fund the expansion of nuclear plants across three states, adding 890 megawatts of new nuclear capacity to the grid. The agreement runs for 20 years, unlocks more than $4.3 billion in investment, and makes one of the world's biggest AI companies one of the world's biggest private funders of nuclear power.

It is the latest sign that the scarcest resource in the AI boom is no longer chips. It is electricity.

What the deal actually is

The agreement has two parts. First, a 20-year power purchase agreement for 890 megawatts of new nuclear capacity, produced by upgrading ("uprating") 11 of Constellation's existing nuclear units across six sites in Illinois, Pennsylvania and New Jersey. Uprates squeeze more power out of existing reactors with new equipment and technology, and Constellation says the first upgrades should start delivering power in 2028.

Second, a separate 15-year energy supply agreement covering 2,700 megawatts from Constellation's existing fleet, giving Google steadier long-term pricing on power it already needs. Together, the arrangement totals roughly 3,590 megawatts aimed at Google's data centers and regional operations, all inside PJM Interconnection, the regional grid covering much of the mid-Atlantic and Midwest and serving some 67 million people.

There is also a software dimension. As part of the deal, the two companies plan to co-develop AI applications for grid operations, and Constellation will use Google Cloud and Gemini Enterprise software to help manage and optimize its plants. The deal funds concrete and turbines on one side, and runs the plants on Google's software on the other.

The scarcest resource in the AI boom is no longer chips. It is electricity, and Google just locked in 20 years of it.

Why nuclear, and why existing plants

AI data centers have one non-negotiable requirement: power that never blinks. A training run on tens of thousands of accelerators cannot tolerate the variability of wind and solar without expensive storage, and grid outages at that scale are measured in millions of dollars per hour. Nuclear is the only large-scale carbon-free source that runs around the clock, which is why it has become the default answer for tech companies with 24/7 load and emissions pledges.

The strategic choice here is uprates over new builds. New reactors take a decade or more and carry famous cost-overrun risk. Upgrading an existing, licensed reactor is faster, cheaper and politically simpler. The new capacity enabled by Google's deal is expected to unlock roughly as much electricity as a brand-new nuclear reactor, according to reporting on the agreement, without breaking ground on a single new plant.

"We're committed to meeting our growth responsibly by actively investing in clean, reliable power that brings new capacity to our nation's grids," said Amanda Peterson Corio, Google's global head of energy and power. Constellation CEO Joe Dominguez framed the partnership as a model for how tech companies and the energy industry can jointly invest in infrastructure in a way that "delivers grid-wide benefits for all, funded by private entities."

The political trick: new power, not just shuffled power

The cleverest part of the deal is what it avoids. When a giant new data center plugs into the grid, regulators and the public have a fair question: will my electricity bill go up to pay for it? Americans are already angry about fast-rising electricity prices, and the data center boom looks like the culprit. Deals like this one answer the question structurally.

Because the power Google is buying is entirely new capacity, added to the grid through the uprates, it does not take electricity away from anyone. Under the arrangement, Google pays a set rate for what it uses while the plants stay hooked up to the broader grid, so the average homeowner keeps their supply. If the structure works as designed, the AI boom's power appetite gets satisfied without raising residential rates, defusing the thorniest political fight in power markets.

Wall Street noticed. Constellation shares jumped about 12 percent on the news, and the rally pulled up other power stocks: Talen Energy rose 12 percent, Vistra 11 percent, NRG 7 percent. The market has figured out that long-term contracts with investment-grade tech buyers are the safest revenue in the electricity business.

Not a one-off: tech's nuclear buying spree

Google's deal is the second mega-agreement of its kind in a week. On September 30, Amazon and Constellation announced a 20-year power purchase agreement tied to expanding the Calvert Cliffs nuclear plant in Maryland, backing more than $3 billion in infrastructure investment. Earlier, Constellation signed a 20-year, 690-megawatt deal with Amazon that included 190 megawatts of new nuclear capacity.

Oracle is running a different, even more inventive play. It struck an agreement with We Energies in Wisconsin to subscribe to 10 to 20 percent of the electricity from the Point Beach nuclear plant, powering Project Lighthouse, the data center campus it is co-developing with OpenAI. Strikingly, Oracle says the arrangement will absorb about $300 million in rising nuclear fuel costs between 2027 and 2033, lowering bills for more than a million Wisconsin households. A cloud company is voluntarily paying so that utility customers pay less, because the alternative, no reliable power at all, is worse.

Even Washington is in the game: Vistra got a Department of Energy offer of $4.2 billion in low-interest government loans to upgrade its nuclear reactors. Constellation itself, which generates roughly 80 percent of the country's nuclear power output, has seen its shares more than triple over the past three years in anticipation of exactly this demand wave.

What could go wrong

The deals are 20-year bets on an industry famous for overpromising demand. If AI investment cools, the hyperscalers could find themselves locked into power they don't need, though the pricing structures make the downside bounded. There is execution risk too: uprates still have to clear engineering and regulatory milestones, and 2028 is a target, not a guarantee.

But the direction of travel is unmistakable. The AI economy has moved through two bottlenecks in three years: first GPUs, then power. When Google, Amazon and Oracle all start buying nuclear reactors the way they used to buy server racks, the constraint isn't temporary. It is structural. The companies training the future's models are now in the business of making the future's electricity, and they are doing it because nobody else will build it fast enough.