Small Modular Reactors Fit Industry Better Than Grid, Study Finds

A roadmap that cuts against the hype

Most of the recent excitement about small modular reactors (SMRs) is about electricity: who will power the next round of AI data centers, and whether factory-built nuclear plants can arrive before the grid runs out of capacity. A new study from the University of Michigan argues that this focus may be misplaced. The work, published in Nature Communications and funded by Idaho National Laboratory, concludes that SMRs could make money in American heavy industry but cannot compete on wholesale electricity prices in the power sector.1

The researchers built an optimization model to work out how many SMRs, and of which type, would be needed to meet energy demand economically at 34 ammonia plants, 9 steel plants, 47 refineries and 955 facilities that use process heat. A second model tested how SMRs would perform when dispatched into a wholesale power market.1 Co-author Brendan Kochunas described the paper as the first comprehensive look at where building and running SMRs actually makes business sense.1

The headline result is clear. Making hydrogen for ammonia, refining and steel is the most profitable early use for these reactors.21 Selling electricity is not. In the power-market model, the study estimates annual electricity revenues that are negative in every case, ranging from about −$4.9 million to −$78 million.21 That holds even though some designs show fairly low levelized costs, including about $46 per megawatt-hour for an integral molten salt reactor.21 The study says SMRs do not appear able to beat renewables or large light-water reactors such as the AP1000 at current wholesale prices. Without subsidies or big cuts in construction costs, the power-sector route looks narrow.1

The tax credit holds everything up

The industrial case rests on a single policy: the Hydrogen Production Tax Credit created by the 2022 Inflation Reduction Act, which pays up to $3 per kilogram of clean hydrogen.1 With the credit, the model finds about 91 gigawatts-electric (GWe) of SMR capacity could be built profitably in a first wave. Without it, the figure is about 4 GWe.1 In a second wave, where mass production cuts construction costs by roughly 20%, the gap grows to 171.9 GWe with the credit versus 7.8 GWe without.1

Projected margins differ widely by sector. Hydrogen for ammonia shows projected profit margins of around 300%. Most refineries come in above 50%, and steel sits near 20%.1 Facilities that only need process heat do much worse: only a few are profitable, and most of those earn margins below 10%.1 The full paper puts the number of industrial sites where SMRs can profitably supply process heat at just 33.21

The study also tests what happens when costs overrun, which matters because overruns are common in nuclear construction. With the credit in place, ammonia and refining stay profitable even if capital costs more than double. Steel and process heat become vulnerable once costs rise by 10% to 50%.21 Every modeled ammonia deployment clears a 7% internal rate of return, the cost-of-capital benchmark the authors use for engineering and construction.21 Without the credit, natural gas with carbon capture beats SMRs at first-of-a-kind costs. Only at mature, nth-of-a-kind costs could SMRs compete with carbon capture for ammonia hydrogen.21

Timing is the obvious weak point. The paper notes that deployment is most attractive before the hydrogen credit expires in 2033.21 Few, if any, commercial SMRs are expected to be running at scale much before then. Co-author Jessica Lovering presents the industrial route as a niche market where SMRs could prove themselves and lower their costs until they can compete in the power sector later.1 That is a reasonable path. But it means the industry's best near-term business case depends on a policy window that could close before the first fleet is built.

The climate gains are significant but limited. US industry emitted 1,360 million metric tons of CO₂ in 2020. The researchers project SMRs could cut industrial emissions by 8% in the first wave and 14% in the second.1 The cost per ton is not cheap: for process heat, the minimum abatement cost is around $500 per ton of CO₂, more than twice the EPA's $235 social cost of carbon.21

Meanwhile, the data center deals keep coming

The study lands in a market where hyperscalers are signing nuclear contracts at a rapid pace. Several deal trackers count about 9.8 GW of nuclear capacity committed to AI infrastructure across 13 announced deals, and every major US hyperscaler has signed at least one.1314 Meta has the largest commitment, up to 6.6 GW across TerraPower, Oklo, Vistra and Constellation, with delivery aimed at 2032 to 2035.14 Its Oklo campus in Pike County, Ohio, is planned to hold 16 Aurora reactors of 75 MW each.11 Amazon backs X-energy's Xe-100 design. Google has a 500 MW fleet agreement with Kairos Power.14

The demand behind these deals is real. Goldman Sachs Research estimated US data center demand at 31 GW in 2025, expected it to reach 41 GW in 2026, and projected 66 GW by 2027.13 One industry newsletter, citing the International Energy Agency, reports that conditional offtake agreements between data center operators and SMR projects grew from 25 GW at the end of 2024 to 45 GW by April 2026.19

The gap between commitments and delivered power is large. One tracker finds that as of July 2026, only about 1.92 GW of the committed capacity was actually supplying power. All of it comes from Amazon's agreement with Talen Energy's Susquehanna plant, which is a conventional reactor, not an SMR.14 The next tranche is also an existing plant: Microsoft's restart of Three Mile Island Unit 1, targeted for the second half of 2027.14 Even the leading SMR developers, including Kairos, X-energy and TerraPower, expect their first fully working units to deliver electricity around 2030.12

Reconciling the two pictures

On the surface, the Michigan findings and the hyperscaler deals seem to contradict each other. If SMRs lose money selling electricity, why are the world's richest companies paying for SMR electricity?

The most likely answer is that data center deals are not wholesale-market deals. The study modeled reactors selling into competitive power markets at prevailing prices.1 Hyperscalers are signing long-term bilateral contracts and paying a premium for firm, carbon-free supply on a fixed schedule. Some industry analysis describes the calculation as the cost of inaction: an AI campus that cannot get power from a congested grid loses more than the extra cost of nuclear electricity.15 Seen this way, the AI buyers are acting like a heavily subsidized industrial customer. In practice, they are doing for SMR electricity what the hydrogen credit does for SMR hydrogen.

The cost evidence supports this reading. First-of-a-kind SMR electricity is benchmarked at $80 to $150 per megawatt-hour, above the $60 to $80 vendors say is needed to be competitive.14 Lux Research has estimated that early SMRs could cost almost three times as much as natural gas once overruns and delays are included.19 Ontario's Darlington project, the first SMR budget approved by a G7 government, puts the first unit plus shared site infrastructure at about $18,646 per kilowatt. That is above the average of recent Western large-reactor builds, although the fleet average falls as later units are added.26 None of this resembles a technology that can win on wholesale price.

Industrial and data center buyers are also closer to each other than the headline suggests. Dow Chemical's planned deployment of four X-energy units at its Seadrift site in Texas is designed to deliver 80 MW of electricity and 200 MW of process heat per unit, partly to make hydrogen on site.8 X-energy is also Amazon's SMR partner.14 The Energy Information Administration also groups AI, data centers and industrial sites together as possible behind-the-meter customers that may not need a grid connection.2 High-temperature gas and molten salt designs, in particular, could serve industrial heat and hydrogen production as well as generate power.2

What it means

The conclusion is not that SMRs have no role in the AI power buildout. It is that their early economics depend on customers willing to pay above market rates, whether those customers are hyperscalers with large capital budgets or ammonia plants receiving a tax credit. Neither model shows SMRs competing head-to-head on today's grid. The 9.8 GW of data center commitments are better read as risk capital for first-of-a-kind reactors than as near-term relief for grid capacity, especially since the operating share still comes from existing reactors.14

For policymakers, the Michigan study points to a specific lever. If the aim is to bring SMR costs down to where they can eventually compete in the power sector, as Lovering suggests, the hydrogen credit may do more than AI enthusiasm alone.1 The two demand sources reinforce each other: every reactor built for a data center or a refinery helps move the industry toward cheaper nth-of-a-kind units. The risk is that both the incentive and investor patience run out before that point is reached.