The site is a few miles from Terre Haute, Indiana, on the grounds of a coal plant that has been demolished. What is planned there is 960 megawatts of nuclear capacity, twelve Xe-100 reactors supplied by X-energy, operated by American Electric Power, with Amazon as the first customer for the electricity. The company confirmed the arrangement on October 19.
The size is three times a 2024 agreement between the two companies, and it is part of a plan to have 5 gigawatts of nuclear capacity in the United States by 2039. Twelve reactors at 80 megawatts each is a specific kind of bet: not one large plant with a single construction schedule, but a series of identical units that can be built in sequence and brought online one at a time.
The choice of a retired coal site is the least romantic and most practical part of the plan. A coal plant already has a high-voltage connection, a switchyard, transmission rights, water access and a county government that has spent decades collecting property tax from it. Those assets can take a decade to assemble at a greenfield site and are, in the United States, the binding constraint on almost every large electrical project.
The reactors are another matter. X-energy’s Xe-100 is a high-temperature gas-cooled design that uses tritium-coated particle fuel, a configuration that is passively safe by construction and entirely unproven at commercial scale. The company has a construction permit application before the Nuclear Regulatory Commission and a first project planned at a Dow chemical site in Seadrift, Texas.
That project is the gate. Nothing in Indiana moves until Seadrift works.
That sequencing is the honest part of the announcement. Amazon is not buying a plant. It is buying the output of a plant that will be built if a different plant is built first, on a timeline that depends on a regulator, a fuel supply chain and a supply chain for pressure vessels that does not yet exist at the volume required.
The fuel is its own bottleneck. TRISO fuel is manufactured at a facility in Oak Ridge, Tennessee, at a scale meant for demonstration, and a dozen reactors running for decades would require fabrication capacity that has not been financed. X-energy has said it intends to build a commercial fuel plant. It has not said when.
For Amazon, the appeal is straightforward. A data center draws power at a flat rate around the clock, and the grid’s carbon intensity varies by hour. Buying nuclear output gives the company a supply that is both firm and carbon-free, which matters for the corporate emissions pledges it has made and for the electricity contracts its cloud customers ask about.
The structure also shifts risk in a way that has become standard. Amazon is funding development and committing to buy power, not owning the plant or holding the operating licence. A long-term purchase agreement is what makes a small reactor bankable: no lender will finance a first-of-a-kind design without a creditworthy buyer signed to twenty years of payments.
That is the trade both sides are making. The utility gets a customer that improves the project’s cost of capital. The buyer gets a price hedged against a volatile electricity market and a claim on capacity in a region where interconnection queues are years long. The reactor vendor gets an order book it can show to the next customer.
The precedent is now well established. Microsoft signed an agreement to buy power from the Crane Clean Energy Center, the reopened unit at Three Mile Island. Meta has solicited proposals for nuclear capacity.
Google has a deal with Kairos Power for advanced reactors. In each case a hyperscaler contracted for output years before construction, and in each case the contract was the event that made financing possible.
The record on delivery is less encouraging. The two reactors at Vogtle in Georgia came in seven years late and more than 15 billion dollars over the original budget. NuScale’s flagship project, which had a Department of Energy award and a utility partner, was cancelled in 2023 when the cost estimates rose faster than the subscriptions.
Small modular reactors are supposed to avoid both failures by being smaller, factory-built and repeatable. None has been built at scale.
The 2039 target gives the program thirteen years, which is roughly the time Vogtle took for two units. Twelve reactors in thirteen years would be a manufacturing achievement rather than a construction one, and it assumes the first unit at Seadrift is finished on schedule and that the NRC’s review of a design it has not licensed before proceeds without a long pause.
What Amazon has secured is optionality, priced as a commitment. If the reactors arrive, the company owns a slice of firm clean power in a state where the alternatives are gas and a grid connection that may not come. If they do not, the purchase agreement has conditions, and the cost of the option is the development spending.
The larger point is about how electricity is now bought. A decade ago a data center signed a utility tariff and waited. Now the largest buyers sign for capacity that exists only on paper, at sites chosen for their switchyards, from vendors whose designs are still in review. The capacity is sold before it is built because the alternative is to build nothing and wait for a queue that does not move.
