Constellation's $4.3B uprate deal shows the real AI power is an old reactor

890 megawatts. That is what Google contracted for on October 6, in a 20-year power purchase agreement with Constellation Energy — enough firm, carbon-free output to stand in for one large new reactor.

The megawatts come from no new reactor. They come from upgrades to turbines, steam generators and digital control systems at 11 units Constellation already operates in Illinois, Pennsylvania and New Jersey, all on the PJM Interconnection.

Constellation will spend more than $4.3 billion on the work. Google is the anchor customer that makes the spending bankable. The first incremental megawatt is dated 2028; the full 890 MW should reach the grid by the end of 2032.

New capacity from plants built in the 1970s and 1980s

An uprate is the least dramatic way to add power to a grid. Engineers replace or retune the non-nuclear side of a plant — low-pressure turbines, generators, condensers, feedwater pumps, control electronics — so the reactor's existing heat turns into more electricity.

The nuclear island does not change. The fuel load does not change. The site does not change.

Constellation has run this play before, on its own balance sheet. It is spending $800 million to lift output at its Byron and Braidwood plants in Illinois by roughly 158 MW, with Byron starting in 2026 and Braidwood following in 2027. The Google agreement scales the same method to six sites and eleven units at once.

Two figures from Canary Media's reporting sketch the geography: a 348 MW uprate at the 2,317-MW Limerick plant near Philadelphia, and a 97 MW uprate at the 2,327-MW Salem station in southern New Jersey. The remainder is spread across the fleet, which spans Illinois, Pennsylvania and New Jersey.

Per kilowatt, the whole program lands at just over $4,800. A greenfield reactor is a different order of magnitude, and a different decade.

Three routes to firm power, three timelines

Constellation's deal sits at the fast end of a spectrum the industry is still sorting out. The routes to firm clean electricity differ mainly in when the power arrives, not in what it costs to run once it is there.

Uprates win on schedule because the interconnection already exists. The transmission, the site permit, the operating licence and the workforce are in place; the operator only has to prove the added capacity to the grid.

The federal government is leaning on the same lever. One day before the Google announcement, the Department of Energy closed a $4.2 billion loan to extend the lives of three Vistra nuclear plants in Pennsylvania and Ohio and to underwrite 433 MW of uprates there, with Meta as the offtaker. Between the Vistra and Constellation programs, most of PJM's near-term uprate capacity is now spoken for.

Restarts sit in the middle. Microsoft is funding the return of Three Mile Island Unit 1, now the Crane Clean Energy Center, at 835 MW and a 2027 target. Google is backing NextEra's 615-MW Duane Arnold restart in Iowa. Both reuse whole plants that already had grid connections and, in Crane's case, an existing licence that Constellation wants to extend to 2054.

New reactors sit at the far end. Almost every data center agreement tied to a fresh design — Kairos Power's Hermes 2, X-energy's Xe-100, TerraPower's Natrium — is a framework, an option or a licence application rather than construction. Kairos's 50 MW Hermes 2 is the only advanced unit actually being built, and it will not be finished before 2030.

The demand behind the deal is not subtle

PJM serves 67 million people and the densest data center corridor in the country. Capacity prices in the market have climbed more than eleven-fold since 2024 as plant retirements collided with a wave of new large loads.

In response, PJM has proposed that data centers connecting to its 13-state grid either bring their own generation or accept remote curtailment at peak. Google and Constellation framed their PPA as a direct answer to that "bring your own power" proposal.

The framing carries political weight. The deal is built so the new capacity is paid for by the buyer rather than the residential ratepayer — the claim both companies and the White House Ratepayer Protection Pledge lean on. Google says it has committed roughly $25 billion in PJM territory over two years for the same reason.

There is a quieter second layer. Beyond the 890 MW of uprates, Google signed a 15-year supply agreement for 2,700 MW from Constellation's existing PJM fleet. That tranche adds no new capacity. It buys revenue certainty for plants already running: nuclear again, but also the gas, oil and renewables that made up about 30% of Constellation's generation last year.

Read together, the two agreements do two jobs. One adds megawatts. The other steadies a balance sheet, and keeps existing assets from retiring early in a market that cannot spare them.

Who is paying, and for how long

Constellation has now signed long-term nuclear supply with all four hyperscalers. Google's 890 MW uprate program, Amazon's 690 MW Calvert Cliffs agreement (including a 190 MW uprate backed by more than $3 billion), Meta's Clinton deal and Microsoft's Crane restart together cover roughly 1,080 MW of announced uprates and 1.64 GW of restarts.

Add the earlier operating-plant contracts, and about 8.5 GW of US reactor output is now locked to Amazon, Google, Meta and Microsoft. The International Energy Agency counts 45 GW of conditional data center agreements with small modular reactors. Only a sliver of that second figure is under construction.

The gap is the story. The existing fleet delivers before 2030. The new-build pipeline is a decade of paperwork from its first commercial megawatt.

The uprate route has a low ceiling. Jefferies estimates total uprate potential in PJM and across the country at roughly 3 to 5 GW. The Google and Amazon programs consume a large share of that room. There are only so many working reactors with headroom to stretch.

The same contract, one ocean away

The model is not staying in PJM. In September, Google agreed to buy as much as half the output of Fortum's Loviisa nuclear plant in Finland under a 22-year deal, part of a €13 billion AI-infrastructure plan. That revenue underpins roughly €1 billion of investment to keep the plant online to 2050.

That is the same trade as the Constellation PPA, moved to a market with different rules and different politics: a hyperscaler exchanging a long, predictable payment stream for the certainty that an existing plant keeps generating. Google frames it as a first of its kind in Europe, and the structure carries over cleanly.

Europe sharpens the comparison because its nuclear fleet is thinner and older, and the new-build pipeline is even slower than America's. If Loviisa's life extension turns on a US cloud company's balance sheet, the power plant and the data center start to look like one asset with two balance sheets.

What every one of these deals shares is a buyer that values uptime above unit cost. A data center cannot park a GPU cluster when the wind drops; it needs firm output around the clock and will pay a premium to lock it in. Nuclear, already reluctant to ramp up and down, is the natural counterparty.

What they also share is a timing problem. Each agreement pays for capacity that arrives late in the decade or after it, while the demand arrives now. The bridge in the meantime is gas, storage and demand response — the less-praised half of the hyperscaler energy strategy, and the part that decides whether the grid stays stable between now and 2032.

What decides whether the model spreads

Three things.

First, the federal uprate program, known as UPRISE, has to turn goodwill into approvals. Constellation says the Google-backed tranche advances it, with the first unit dated 2028.

Second, the count. The Department of Energy wants 5 GW of new nuclear from uprates and restarts by 2030 without a single new reactor. Google claims its uprate and restart deals alone cover nearly a quarter of that target. If the timelines slip past 2030, the arithmetic breaks.

Third, the alternative. Gigawatt-scale gas is faster to build and, today, cheaper per kilowatt. If capacity prices keep climbing and regulators keep squeezing data center loads, nuclear uprates stop being a premium product and start being the default.

The pattern is not new to this archive. As we wrote in our September note on ONE Nuclear, a nuclear label on a portfolio does not make the megawatts firm — 4 GW of that company’s 5 GW plan was gas. Constellation’s difference is that its raw material is already licensed, already connected, and already staffed.

That is also the limit. An uprate is a bet that the existing fleet still has slack, and the fleet does not grow. Once the last turbine is retuned, the next delivery of firm clean power has to come from new concrete — and concrete does not arrive on a data center schedule.