Google Signs 3.6 GW Constellation Deal to Fund Nuclear Uprates for AI Power

Google has agreed to buy roughly 3.6 gigawatts of electricity from Constellation Energy in the PJM Interconnection, the largest power grid in the United States, under two long-term contracts announced on Tuesday, October 6. About a quarter of that supply will come from new capacity that Constellation plans to add by upgrading existing nuclear reactors, according to Reuters. For enterprise technology leaders, the deal is another reminder that the supply and price of AI compute is now shaped as much by power contracts as by chips.

What did Google and Constellation agree to?

The agreement has two parts. The first is a 20-year power purchase agreement for 890 megawatts of new output from "uprates" at 11 Constellation nuclear units in Illinois, Pennsylvania and New Jersey. Constellation will put more than $4.3 billion into the work, which covers turbines, steam generators and digital control systems, the American Nuclear Society's Nuclear Newswire reported. The first upgraded unit is expected to deliver power by 2028, and the full 890 MW is expected by the end of 2032, according to POWER Magazine.

The second part is a 15-year agreement for 2,700 MW from Constellation's existing PJM fleet. That portion adds no new capacity to the grid. Reports describe it a little differently: Enerdata called it nuclear capacity, while Reuters described a long-term supply agreement not tied to a specific source of generation.

Constellation has not formally published its list of units. A company spokesman told Nuclear Newswire the work will take place at Braidwood, Byron, LaSalle and Quad Cities in Illinois, Limerick in Pennsylvania and Salem in New Jersey. The two companies also expanded a five-year technology agreement under which Constellation will use Google Cloud and AI services, including agentic workflows in Gemini Enterprise.

Why are tech companies paying to upgrade nuclear plants?

An uprate gets more electricity out of a reactor that is already licensed, built and connected to the grid. That makes it one of the faster ways to add firm, carbon-free supply, compared with new plants that can take many years to permit and build. Constellation said the 890 MW would be "entirely new" capacity, POWER reported.

Grid rules are a second driver. PJM has proposed an approach it calls "Bring Your Own New Capacity," which would let new large loads of 50 MW or more at a single site reduce their exposure to a proposed Interim Resource Adequacy Service, under which uncovered loads could be curtailed during emergencies starting in June 2027, if they bring new supply with them. The proposal is pending at the Federal Energy Regulatory Commission, and neither company has said whether the new nuclear output will be registered under it, according to POWER. Reuters described the stakes more bluntly: under PJM management's proposal, data centers that do not bring their own power could risk being cut off from the grid during peak demand, and the companies said the deal responds to it.

Google is not alone. Constellation signed an uprate agreement with Amazon for its Calvert Cliffs plant in Maryland about a week earlier, and Google announced a separate uprate agreement with Georgia Power covering the Vogtle and Hatch plants in September, Nuclear Newswire reported.

What does the deal mean for enterprise AI buyers?

Few enterprises will ever sign a nuclear contract. But nearly all of them buy compute from companies that do, and power has become one of the main limits on how much AI capacity cloud providers can offer, where, and at what price. The new nuclear output arrives gradually between 2028 and 2032, so it does little for capacity this year or next. The more immediate signal is how aggressively hyperscalers are locking up long-term supply in the country's largest grid region.

Practitioners at AI Connect events have described the same constraint from inside the data center. "A thousand watt GPU needs a thousand watts of cooling," Steve Heibein, Federal AI Chief Technologist at Hewlett Packard Enterprise, said during a AI Connect webinar on why rising GPU power is pushing data centers toward liquid cooling. His point was that rack power, not just the number of chips a company can buy, now decides how many AI servers a facility can hold.

That has practical consequences for planning:

• Cloud capacity is tied to power. GPU availability by region may depend on where providers have secured electricity, so capacity reservations and multiregion options belong in cloud negotiations.
• Private AI faces the same math. Companies running AI on their own infrastructure hit power and cooling limits at a smaller scale, which is one reason vendors at AI Connect events have pushed prebuilt AI factory designs and bringing AI to where sensitive data already lives instead of building from scratch.
• Emissions reporting will follow the contracts. Sustainability teams should ask providers how nuclear and other clean power purchases are reflected in the emissions figures that enterprises report for their own cloud use.

What to watch next

PJM has asked FERC to make its new large-load rules effective by October 12 so they can apply to the capacity auction scheduled to begin in December, POWER reported. Whether regulators sign off, and whether other hyperscalers follow with uprate contracts of their own, will shape how quickly new power reaches data centers in the Midwest and Mid-Atlantic. Constellation also still needs Nuclear Regulatory Commission approval for the upgrades, with filings planned between late 2026 and 2030, and the pace of those reviews will decide whether the first new megawatts arrive on schedule in 2028.

Frequently asked questions

What is a nuclear uprate?

An uprate is an upgrade that lets an existing, licensed reactor produce more electricity, using new equipment such as turbines, steam generators and control systems. It is generally faster than building a new plant.

How much new power does the Google and Constellation deal add?

About 890 MW of the roughly 3.6 GW is new capacity from uprates at 11 units. The other 2,700 MW comes from Constellation's existing PJM fleet and adds no new capacity.

Why does this matter to enterprises that do not buy power directly?

Cloud providers' access to power increasingly limits how much AI compute they can offer and where. Long-term power deals shape GPU capacity, regional availability and pricing for enterprise customers.

Filed under: Data & AI Infrastructure, AI Strategy & ROI