Small Modular Reactor Permit Clears TVA to Build First U.S. Unit

A permit, not a power plant

On September 29, the Nuclear Regulatory Commission announced a construction permit that lets the Tennessee Valley Authority build a small modular reactor at its Clinch River site near Oak Ridge, Tennessee. The agency called it the path to the nation's first commercial SMR.7 The unit is a 300-megawatt BWRX-300, a boiling-water reactor designed by GE Vernova Hitachi.2 The NRC finished its review in 14 months, four months ahead of its own schedule.4 NRC Chairman Ho K. Nieh said the early finish is "what the new NRC looks like."4

The permit lets TVA build the reactor. It does not let TVA run it. Before loading fuel, TVA still needs a separate operating license, which means a second application and a second review.15 The permit document, numbered CPAR-2, is dated September 28 and gives TVA until September 30, 2032 to finish construction.10 TVA has not announced when construction will start or what the project will cost.5 It is aiming for commercial operation in the early 2030s.6

Reports differ on one small detail. Some place the site in Oak Ridge and others in Roane County.12 Both are correct, because the site sits in Roane County near Oak Ridge.5 The bigger difference is in how outlets framed the story. One technology outlet called the BWRX-300 a "data center-specific" reactor.20 The NRC and TVA said something narrower: the unit is meant to meet growing regional power demand.2 That gap between the official description and the headlines says a lot about how the permit is being received.

Why the AI power crunch made this a headline

TVA did not apply for this permit because of artificial intelligence. It got an early site permit for Clinch River back in 2019, which covers up to 800 megawatts of reactor capacity.10 Even so, the permit arrives as electricity demand climbs fast. A Lawrence Berkeley National Laboratory study projects that data centers could use about 12 percent of U.S. electricity by 2030, up from 4.4 percent in 2023.6

Nuclear fits what data centers need. Many AI facilities run around the clock and need steady power. Solar and wind vary with the weather and need storage or backup to match that profile, while a reactor produces steady output in any conditions.5 That is why tech companies have signed a long list of nuclear deals:

• Google: has an agreement with Kairos Power for up to 500 megawatts.11
• Amazon: with X-energy, has a stated goal of 5 gigawatts of new nuclear capacity by 2039.18
• Meta: has agreed with TerraPower on up to eight Natrium plants.11

The deals keep coming. On October 19, Amazon said it will fund a proposed 960-megawatt plant made up of 12 X-energy Xe-100 reactors. The plant would sit on the site of a demolished coal plant near Terre Haute, Indiana, and American Electric Power would operate it.18 Weeks earlier, Amazon signed a 20-year power purchase agreement with Constellation for Maryland's Calvert Cliffs plant. The deal funds a $3 billion investment and includes plans to explore SMRs on the site.16

The Clinch River design also has a direct link to data centers. Blue Energy has applied to connect up to five BWRX-300 units, about 1.5 gigawatts in total, to the Texas grid run by ERCOT. The power would supply a Crusoe data center, with the first unit expected in service in 2033.13 ERCOT said it was the first SMR application it had ever received.13 Blue Energy also filed the first part of its own NRC application in September. Being second means it can follow a licensing path TVA has already cleared.10

The timing problem

The data center angle has a basic weakness: SMRs arrive late compared with the demand driving the interest. As of mid-October, none of the U.S. designs in hyperscaler deals was producing power, and most companies' first-power targets cluster around 2030.11 One analysis found only a single power purchase agreement with a new advanced reactor: TVA's deal to buy up to 50 megawatts from Kairos's Hermes 2 plant, which serves Google data centers.12 The same analysis put that in perspective. A campus with 1 gigawatt of computing load, running at a power usage effectiveness (PUE) of 1.3, draws about 1,300 megawatts in total.12 PUE is a building's total power divided by the power its computers use.

For now, the nuclear power actually reaching data centers comes from existing plants and restarts, not new reactors.12 Natural gas is filling most of the near-term gap. Global Energy Monitor counts 378 gigawatts of planned U.S. gas capacity, up 50 percent in six months, and half of that is meant for data centers.6 The reporting broadly agrees that Clinch River will not ease today's power shortage.5

Markets are showing the same caution. Holtec, which wants to build two 300-megawatt SMRs at its Palisades site in Michigan, has delayed its IPO indefinitely. It cited weak market conditions, including uncertainty about data centers.14 X-energy's shares have fallen by nearly two-thirds since its April listing, even after the new Amazon deal.18 Investors appear willing to fund nuclear, but they are pricing in long timelines.

Grid power, not private power

One trend in the latest deals gets less attention than it should. Most of the new nuclear capacity is planned to feed the public grid rather than sit next to a data center and supply it directly. The Indiana Xe-100 plant will connect to the grid, while its power also serves Amazon's data centers.18 At Calvert Cliffs, Constellation said the power will keep flowing to the PJM regional grid, and Amazon gets a fixed-price hedge instead of a direct supply.16 Amazon had dropped a plan for a data center at Calvert Cliffs after strong local opposition.16 Clinch River is a utility project serving TVA's seven-state territory.6

This matters for grid planning. Nuclear developers seem to be avoiding the politically risky model of reactors that serve only one customer. Instead, they are adding firm capacity that serves everyone, with tech companies paying for it and getting long-term price certainty. Data center operators will still need grid connections either way.12 Our view is that this approach will last longer than the private on-site reactor idea that dominated SMR coverage a year ago.

Cooling: on both sides of the meter

Cooling shows up twice in this story. On the reactor side, the BWRX-300's main feature is natural circulation. Hot water rises out of the core, cools, and sinks back without the large electric pumps a conventional boiling-water reactor uses.10 It also has passive safety systems that rely on gravity, pressure, and stored water.5 The design is meant to cool itself in an emergency without pumps or operator action.6 Removing the recirculation pumps means fewer parts that can fail or need their own safety case.10 That simplification is a big part of how GE Vernova Hitachi argues the design will cost less.

On the data center side, cooling largely sets how much power a site needs to buy. The Uptime Institute's 2026 survey puts the industry's average PUE at 1.52, and large new builds often reach 1.3 or better.12 Better cooling therefore shrinks the amount of generation a campus has to secure. Every improvement in PUE reduces the gap that reactors, gas turbines, or the grid have to fill. Improvements in cooling will reach data centers years before most SMRs do.

The cost question

The economic doubts are significant. Georgia's two newest Vogtle reactors cost about $35 billion combined and finished years late.5 Getting from licensing to selling power took about 15 years.10 Canada is building the first BWRX-300 at Darlington. That unit is estimated at C$6.1 billion, and the four-unit program at C$20.9 billion.6 One analysis put the cost at about C$25,667 per kilowatt for the first unit.12 The best-known earlier U.S. SMR project, NuScale's plant in Utah, was cancelled in 2023 after its expected power price rose from $55 to $89 per megawatt-hour.6 David Schlissel of the think tank IEEFA argues that experience with real projects shows SMRs are not cost-competitive.6

The site itself is a reminder. Clinch River once hosted the Clinch River Breeder Reactor. Its cost estimate grew from about $400 million to $8 billion, and the Senate cancelled it in 1983 after roughly $1.6 billion had been spent.10

Supporters say the answer is building the same design many times. GE Vernova Hitachi CEO Jason Cooper called fleet deployment "essential."6 The Department of Energy is backing Clinch River with $400 million.6 TVA is weighing up to four units at the site.20 The design is also planned for Poland and Texas.6

Our reading

The permit's real achievement is regulatory, not about electricity. A 14-month review suggests U.S. nuclear licensing can now run on a schedule, which has been rare for decades.10 TerraPower's Natrium reactor in Wyoming got its permit in March after an 18-month review, so this is a pattern, not a one-off.10

Licensing is no longer the main obstacle. Cost, construction, and fuel are. Whether Clinch River can set a believable benchmark for cost and schedule will shape how utilities and data center developers judge SMRs for years.5 AI demand gave the permit its headlines. Whether these reactors ever power data centers at scale depends on how construction goes, and TVA has not yet said when it will begin.