Google Bought 890 Megawatts of Nuclear to Feed Its Data Centers

On October 6, Google and Constellation Energy announced a deal to add 890 megawatts of generating capacity by upgrading 11 nuclear units in the PJM region, the grid that covers much of the eastern United States. Constellation put the investment at more than $4.3 billion, and the two companies signed a 20-year power purchase agreement to back it. The first uprate is due in 2028, with Google saying the full capacity should be online by the end of 2032.

The headline number is the 890 megawatts of new capacity. The more interesting number is 2,700 megawatts, which is a separate 15-year supply arrangement between the same two parties covering output from PJM's existing fleet. Google is financing new generation and contracting for a large block of power that already exists, which means it is bidding against every other large buyer in the region for a scarce resource.

Why uprates instead of new plants

Uprating means raising a reactor's output by improving its components and operating limits rather than building a new facility. It is the cheapest and fastest route to more nuclear power, because the plant, its licence and its grid connection already exist. A new reactor takes a decade and a half and a regulatory process that few utilities want to start. A 10 percent uprate on an existing unit takes a few years.

That is why Constellation is the counterparty. It operates the largest nuclear fleet in the United States, so it has 11 units sitting behind a single grid operator and a track record of getting uprates approved. For Google, the deal converts a long-run cost problem into a long-run supply contract. Data center economics turn on power, and locking a fixed price for 20 years removes the volatility that makes a five-year investment case hard to sign.

The demand picture is not a forecast anymore

The timing of the announcement lines up with a pile of data showing the load is already here. The Energy Information Administration's short-term outlook, released the same day, put US electricity consumption up 4 percent year over year in the third quarter of 2026, with the commercial sector up 5 percent. PJM's own long-term forecast projects summer peak demand growing about 3.6 percent a year for the next decade.

The International Energy Agency's Electricity 2026 report, published October 1, is blunter. Global electricity consumption is rising at its fastest rate in 15 years, at an average of 3.6 percent a year through 2030. Electricity demand from data centers grew 17 percent in 2025, and consumption from AI-focused facilities grew 50 percent. In the United States, data centers now account for roughly half of all incremental demand growth.

The bottleneck is not generation alone. Of the capacity that submitted interconnection requests between 2000 and 2019, only 13 percent had reached commercial operation by the end of 2024. About 77 percent had been withdrawn. Grid planning tools built for slow, spread-out residential growth do not handle clusters of demand that arrive on a data center timetable, which is why buyers are increasingly going around the queue and paying for their own power.

Why nuclear specifically

Uprates are not the only option, and the choice of nuclear says something about what large buyers value. Gas turbines are faster to permit and cheaper to build, and several operators have ordered their own gas engines to skip the interconnection queue. But gas carries fuel price risk over a 20-year horizon, which is the opposite of what a company wants when the contract is meant to remove volatility. Renewables are cheaper per megawatt-hour and intermittent, and a training cluster cannot run on a generation profile that depends on weather unless it pays for storage on top.

Nuclear gives a flat output curve and a long operating life, and an uprate delivers both without a new site. That combination is why the same playbook keeps showing up. Microsoft signed a deal to restart a reactor at Three Mile Island. Amazon has been buying power directly from nuclear operators. The pattern reflects a search for a supply contract where the price is knowable in 2046, more than any enthusiasm for atomic energy as such.

Who pays for the wires

That workaround is where the political fight sits. The Energy Department announced almost $2 billion for grid upgrades, adding 7.5 gigawatts of capacity and rebuilding more than 1,500 miles of line. That is federal money covering a cost that would otherwise land on ratepayers. The House passed a bipartisan Ratepayer Protection Act in September that would push state regulators to keep residential customers from absorbing the grid costs of AI data centers. Households get the bill either way, through taxes or through rates, and the argument is about which one.

The World Bank added a number on October 7 that makes the stakes concrete for developing economies. A 500-megawatt AI data center, which is the industry benchmark for training workloads, would consume nearly 15 percent of the average baseload power available across emerging market and developing economies in East Asia and the Pacific. The same facility in an advanced economy takes up roughly four times less of the available base. Energy is about 65 percent of a data center's annual operating cost, so power access, not silicon, is the long-run constraint.

Europe is going for disclosure instead of supply

The European Commission proposed a mandatory sustainability rating in September for data centers above 500 kilowatts, with annual labels starting in August 2027. The grades cover energy and water efficiency, renewable sourcing, grid-support functions, local water stress and waste-heat reuse. They will not cap energy use or require total power-draw reporting.

Read that scope carefully and it tells you what the EU is doing. The label is a comparison tool, not a limit. It gives buyers a way to prefer efficient sites, and it leaves the supply problem to the market. The US approach is closer to buying capacity directly, through federal loans and long-term corporate contracts like the Google deal. Same problem, two very different instruments.

What the deal template looks like

Google's contract is likely to become a pattern rather than a one-off. The structure has three parts: a long-term purchase agreement that finances new capacity, a separate supply deal for existing output, and a counterparty that already owns the asset and the permits. Any hyperscaler that wants power at scale will be negotiating a version of it, because the alternative is waiting for a grid connection that may never clear.

The risk in that template is that it concentrates supply. If a handful of companies can sign 20-year contracts for existing generation, they can effectively take that output off the market for everyone else, and the price a smaller operator pays goes up. Nuclear uprates add real capacity, which helps. Contracting for the existing fleet does not. The second part of the Google order is the part other buyers should worry about.

For now, the deal says something simple about where the industry is. A year ago, the constraint on AI was chips. Today, a company is spending $4.3 billion to squeeze more power out of reactors that have been running for decades, and signing away two decades of price certainty to do it. The compute is available. The electricity is the thing that has to be arranged, years in advance, the way a supply chain is arranged.