Constellation's Amazon and Google Nuclear Deals: Hyperscalers Pay for Added Megawatts

Constellation signed two tech giants in six days. On September 30 Amazon agreed to buy 690 MW from the Calvert Cliffs plant in Maryland for 20 years. The stock traded 3.8% higher before the open on October 1. On October 6 Google signed a two-part deal, and the shares rose about 5% before the open again. The easy reading is simple: AI still needs nuclear power, reactors are scarce, so anyone who owns one wins.

Look at the middle column, not the last one. Neither contract sells existing output on its own. Both pay for an uprate, which means new equipment that lets an existing reactor produce more power. Amazon's contract also pays to renew the licences of two reactors that would otherwise expire in the next decade. Google's 2,700 MW of existing output is a financial contract: the power keeps flowing to PJM, the grid that serves Pennsylvania, New Jersey and 11 other states. It comes attached to 890 MW of new output that Constellation will spend more than $4.3 billion to build.

Two deals in a week are not a pattern. The table below lists every current nuclear contract the five covered companies have signed with a named tech buyer, plus the White House pledge. Talen's 2024 supply to an Amazon data centre next to its plant is left out, because its 2025 contract replaced it.

Seven of the nine contracts add something new: a restart, an uprate, a licence renewal or a new reactor. Only two sell existing output alone, Talen's and Vistra's at Comanche Peak. Both were signed in 2025, and both only promise to explore uprates later. Talen's contract was itself a rewrite. Regulators at FERC had turned down a plan to send more of its plant's power straight to the Amazon site. The pattern did not start with the pledge. The pledge wrote it down: signers promised to buy from "new or otherwise additive power plants" and to pay the full cost.

A running reactor is scarce, but its power already serves someone. Moving that power to one buyer for 20 years takes it away from today's customers. The pledge says the buyer must not push the cost of replacing it onto households. A financial contract, like Google's 2,700 MW, avoids the problem by leaving the power on the grid. An uprate or a licence renewal creates power that would not exist without the contract. So what matters for the next deal is a company's room to add megawatts, not how many it already runs.

• Uprates. Most reactors can produce a little more with new turbines, pumps or fuel. Google's 890 MW is spread across 11 reactors, so a large fleet can offer an addition no single plant could.
• Licence renewals. A plant near the end of its licence can be sold as new life. Clinton and Calvert Cliffs were both sold that way. A company with older reactors has more of these to offer.
• Restarts. Crane and Duane Arnold were plants that had already closed. Very few closed plants can be restarted, so this is a one-off rather than a pipeline.
• New reactors. The purest addition, and the slowest. Oklo's Meta project is all new, with first power no earlier than 2030.

The strongest objection comes from Constellation itself. In Q2 2026 it signed 920 MW of 15–20 year contracts with what it called investment-grade customers. The American Nuclear Society reports that about 890 MW of that is existing output. Only Walmart's 176 MW comes with an uprate. So existing nuclear power still finds buyers after the pledge. The argument here holds only in a narrower form. The Q2 buyers who took existing output alone were not named. And since 2025, no named tech giant has bought existing output without an addition in the same deal. If the Q2 buyers include a pledge signer, the rule is weaker than this note says.

The second objection is the size of Google's existing-output piece. At 2,700 MW, it is three times the new capacity. Making it a financial contract may just be good politics. If existing power only needs a small addition attached to sell, the easy reading is mostly right. Room to add would then matter only at the edges, and every reactor owner would benefit.

The third objection is the bill. Constellation will spend more than $4.3 billion on the Google uprates, and the Calvert Cliffs deal enables more than $3 billion more. Neither release discloses a price. A 20-year contract that pays back a large build looks more like a utility's regulated return than a windfall. If the returns on uprates are thin, the company with the most room to add also has the most money at risk. The early rallies on both deal days would then be cheering the wrong half of the contract.

The fourth objection is geography. Comanche Peak sold existing output to Amazon in Texas, which has its own grid outside PJM. The deal only included options on uprates. If the rule is mostly about PJM, Texas may keep selling existing output on its own. A week of contracts settles none of this. Each objection needs its own evidence: the Q2 buyers named, a price for the 2,700 MW, a disclosed return on uprate spending (perhaps at Constellation's November results), or a new contract in Texas.

Sort the five companies by their room to add and the order is clear. Constellation has the most reactors to uprate and licences to renew, and has just used both in back-to-back deals. Vistra has done it once, with Meta at three plants, and holds an option at Comanche Peak. NextEra is the middle case. It runs four nuclear sites with room of their own, but nuclear is a small part of a company built around a regulated Florida utility. Its one tech nuclear deal is a restart. Talen owns a large gas fleet but only one nuclear plant, and Amazon's contract already covers about 85% of its share. Oklo has only new reactors and no running plant to pair them with.

The framework's ratings measure how broad each contract book is, not room to add, but here the two agree. Constellation and Vistra are rated strong on transaction embedding, because their contracts span several tech buyers. Talen is held at intact: one buyer, one nuclear plant, and most of the ramp still ahead. NextEra is also strong, on contracts that are mostly renewables. Reading the gap between Constellation and Talen as a gap in room to add is this note's view, not the framework's.

The GE Vernova note argued that having already sold the megawatt is what sets a power company apart. This note partly reverses that. A sold megawatt is a contract already won. It says little about whether the seller can win the next one. Talen's Amazon contract is signed and ramping up, and it is the very reason Talen has the least nuclear left to sell. Constellation's two deals say the opposite about Constellation, at the cost of a building bill whose return nobody has disclosed.

When a buyer's promise changes what it can sign, the scarce thing moves. Here it moved from the reactor itself to the room left to add one more megawatt.