Google Locks In 3,590 MW From Constellation, 890 MW New Nuclear

Google just bought the output of a new nuclear reactor without anyone building one. On October 6, 2026, Google and Constellation Energy, the largest US nuclear operator, announced a 20-year power purchase agreement for 890 MW of new nuclear capacity, created by upgrading 11 existing reactor units across six sites in Illinois, Pennsylvania and New Jersey. A second 15-year agreement covers 2,700 MW from Constellation's existing fleet in the PJM grid. Total: 3,590 MW.

The first upgraded capacity arrives in 2028, and all 890 MW will be online before the end of 2032. Constellation is investing more than $4.3 billion. The deal shows where AI power is actually coming from in the near term: not small modular reactors, which are still years away, but squeezing more out of reactors that already exist.

What Google Signed

The Google-Constellation deal is a pair of long-term electricity contracts: one that pays for new nuclear capacity created by upgrading existing reactors, and one that buys power from Constellation's current fleet.

Neither company disclosed a price. Google also said it will deploy Gemini Enterprise to help Constellation optimise nuclear plant operations, so part of the relationship runs in the other direction.

What a Nuclear Uprate Is

A nuclear uprate is an upgrade that increases the electrical output of an existing reactor by improving its thermal or electrical efficiency, using new turbines, generators, pumps, instrumentation or fuel strategies, without building a new reactor.

Uprates typically add a few percent to a few tens of percent of a unit's capacity. Spread across 11 units, they add up to 890 MW, which Google says is "roughly the equivalent to adding an entirely new large reactor" and Constellation compares to the output of several small modular reactors. The Nuclear Regulatory Commission must approve each one, but the approval path is far shorter than licensing a new plant.

Reports identify plants including Byron, Braidwood, LaSalle, Quad Cities and Dresden in Illinois, Peach Bottom and Limerick in Pennsylvania, and Salem in New Jersey. Constellation has not published the final list.

The Numbers in Context

Our Analysis: What 3,590 MW Actually Powers

Big megawatt numbers are hard to picture. Here is our rough translation, with assumptions stated.

Annual energy. 3,590 MW running all year is about 31.4 terawatt-hours. At a typical nuclear capacity factor of around 90%, that is roughly 28 TWh a year, comparable to the annual electricity use of a mid-sized European country.

GPUs. If each deployed AI accelerator draws about 1.4 kW all-in, including its share of servers, networking and cooling, 3,590 MW could support roughly 2.5 million accelerators running at full load. Real numbers vary widely with chip generation and cooling efficiency, so treat this as an order of magnitude, not a forecast.

Cost per megawatt. $4.3 billion for 890 MW works out to about $4.8 million per MW of new capacity. Recent new large reactors in the US have cost several times that per MW and taken well over a decade. That gap is why uprates and restarts are the fastest nuclear option for AI demand.

Who carries the risk. The two contracts do different jobs. The 890 MW uprate PPA is a 20-year commitment that gives Constellation the revenue certainty to spend $4.3 billion; Google effectively underwrites new supply. The 2,700 MW fleet contract mostly locks in existing output for 15 years, which protects Google from price spikes but adds little new power to the grid. Read the deal as roughly one-quarter new supply and three-quarters hedge. That is still valuable, but critics of data center demand will focus on the larger, existing-fleet half.

Timing gap. First uprated power arrives in 2028 and full capacity by 2032, while Google's AI capacity needs are growing every quarter. In the meantime, Google will still rely on gas-heavy grid power and other contracts. Nuclear deals like this one fix the 2030s; they do not solve 2027.

Why PJM Matters

PJM is the grid operator for much of the Mid-Atlantic and Midwest, including Northern Virginia, the densest data center market on earth. Data center demand has pushed PJM capacity prices to record highs, and household bills have followed. Google says this deal adds power "at no cost to other ratepayers," which is a direct answer to that political pressure. We covered the backlash in the Big Tech ratepayer pledge.

There is also a market-design question. PJM has been considering rules that let large buyers bring their own new capacity. Whether these uprates can count under those rules has not been confirmed. The 2,700 MW fleet contract, by contrast, mostly keeps existing plants delivering into the market rather than adding new supply.

The Big Tech Nuclear Race

Google is not alone. In the same week, Oracle agreed to subscribe to output from the Point Beach nuclear plant in Wisconsin to support its Port Washington "Project Lighthouse" data center. Microsoft has its Three Mile Island restart deal with Constellation, and Amazon has backed nuclear projects of its own.

The common thread: AI data centers need firm, 24/7 power, and solar plus batteries cannot yet deliver that at scale. Nuclear is the only large carbon-free source that runs around the clock. Existing reactors are the only nuclear capacity available this decade.

What It Means for Developers and AI Buyers

1. Power cost is AI cost. Electricity is one of the largest operating costs for AI data centers. Long-term fixed-price power contracts help Google keep inference prices stable while competitors on spot power face volatility. That eventually shows up in API pricing; track it on our LLM API Pricing Tracker.

2. Expect region constraints. New AI capacity will cluster where firm power is available. If you need large GPU allocations in 2027 and 2028, regions backed by long-term power deals will likely have better availability.

3. The US-China energy gap is real. China has been adding generating capacity far faster than the US, which gives it an edge in powering AI despite chip restrictions, as we covered in China's energy advantage. Uprates are one of the few ways the US can add firm capacity quickly. See also our AI data center power wall analysis.

What To Watch Next

• NRC approvals for each of the 11 uprates
• The final list of plants from Constellation
• Whether PJM counts the uprates as new capacity for large-load rules
• More uprate deals from Microsoft, Amazon and Meta
• Google's next nuclear announcement; it has now enabled over 1.5 GW

Key Takeaways

• Oct 6, 2026: Google and Constellation signed deals totalling 3,590 MW in the PJM grid
• 890 MW of new nuclear capacity via upgrades at 11 reactor units, under a 20-year PPA
• 2,700 MW from Constellation's existing fleet under a 15-year agreement
• First uprate in 2028, all 890 MW by end of 2032; Constellation investing $4.3B+
• About 7,200 construction jobs and 4,400 jobs sustained
• Our estimate: enough for roughly 2.5 million AI accelerators at full load, about 28 TWh a year
• Uprates are the fastest nuclear option for AI power this decade, far ahead of SMRs

Sources

• Google and Constellation joint press release (Oct 6, 2026)
• Google blog: "Why we're backing America's existing nuclear plants" (Oct 6, 2026)
• POWER Magazine and World Nuclear News on the 11-unit uprate programme
• Enerdatics analysis of the two contracts and reported plant list (Oct 9, 2026)
• World Nuclear News on Oracle and Point Beach