Small Modular Reactors: The Nuclear Bet That Could Power the AI Age

Nuclear power has had a complicated few decades. After Chernobyl and Fukushima, the political and public appetite for new nuclear construction collapsed across much of the Western world. Costs ballooned on the large plants that did get built. Timelines stretched from years into decades. The industry looked, to many observers, like a relic.

Then AI arrived and changed the math entirely.

The energy demands of modern AI infrastructure are so large, so continuous, and so carbon-sensitive that tech companies are being forced to think about power generation in ways they never had to before. And increasingly, the answer they keep arriving at is nuclear — specifically, a new generation of smaller, factory-built reactors called Small Modular Reactors, or SMRs.

Tech giants have committed over $10 billion to nuclear partnerships, with 22 gigawatts of SMR projects in development globally. This is not a niche experiment. It's a full-scale industry bet on a technology that doesn't yet exist commercially at scale — and understanding what SMRs actually are is increasingly essential context for anyone following the AI infrastructure story.

The definition is right there in the name, but it's worth unpacking each word.

A small modular reactor is a nuclear fission reactor with a rated electrical power output of less than 300 megawatts, which uses modular design principles to achieve streamlined construction and enhanced scalability compared to large light-water reactors.

SMRs have a power capacity of up to 300 MW per unit — about one-third of the generating capacity of a traditional nuclear power reactor. To put that in perspective, a conventional large nuclear plant like the ones you'd recognize from news coverage typically generates 1,000 to 1,600 MW. An SMR produces a fraction of that — but that's precisely the point.

The "modular" part is where the real innovation lives. Prefabricated units of SMRs can be manufactured and then shipped and installed on site, making them more affordable to build than large power reactors, which are often custom-designed for a particular location — sometimes leading to significant construction delays. Instead of building a bespoke, one-of-a-kind reactor on-site over 15 years, the idea is to manufacture standardized components in a factory, ship them to location, and assemble them — the same logic that made aircraft manufacturing and semiconductor production dramatically cheaper over time through repetition and scale.

SMRs offer additional flexibility in operation and wider deployment opportunities, allowing nuclear to be used in more locations and for a greater range of applications. You can place them on land that wouldn't suit a traditional plant. You can add modules incrementally as power demand grows. And you can site them closer to the users — like a data center campus — rather than needing to transmit power across hundreds of miles of grid infrastructure.

At their core, SMRs do the same thing all nuclear reactors do: they split atoms to generate heat, use that heat to produce steam, and use the steam to spin turbines that generate electricity. The physics is the same as a conventional plant — and the same as a coal plant, for that matter, minus the carbon.

SMR safety principles largely rely on simple phenomena such as natural circulation to cool the reactor core, even during incidents or accidents, requiring little or no operator intervention to bring the reactor to a safe state. This passive safety design is one of the most important engineering improvements over older reactor generations. Rather than requiring active intervention by operators or powered systems to prevent a meltdown, many SMR designs are engineered so that basic physics — gravity, convection, natural circulation — automatically bring the reactor to a safe state if something goes wrong.

These passive safety systems also allow the elimination of a range of components such as valves, safety-grade pumps, pipes, and cables, thereby reducing the risk of their failure. Fewer moving parts means fewer things that can go wrong — a meaningful argument in an industry where public trust in safety is foundational.

SMR concepts encompass various reactor types, including Generation IV designs, thermal-neutron reactors, fast-neutron reactors, molten salt, and gas-cooled reactor models. Not all SMRs are the same technology. Some use water as a coolant and moderator, similar to existing reactors. Others use helium gas, molten salt, or liquid metal. Each design has different efficiency, fuel, and safety profiles — making the SMR landscape more of a family of technologies than a single product.

The connection to AI infrastructure is not subtle. As we covered in our data center cooling piece, the power demands of modern AI compute are reaching a scale that the existing grid — and certainly renewable energy alone — cannot reliably meet.

AI data centers will consume 945 terawatt-hours annually by 2030 — equivalent to the entire electricity consumption of Japan — while demanding 24/7 carbon-free power that only nuclear can reliably provide. That last phrase is critical: 24/7 carbon-free. Renewable energy — solar and wind — is intermittent. The sun doesn't always shine and the wind doesn't always blow. Battery storage technology is improving but remains insufficient at the scale required to backstop a hyperscaler's power needs through multiple cloudy, windless days. Natural gas can fill the gap, but it produces carbon and increasingly conflicts with the environmental commitments tech companies have made publicly.

Nuclear, uniquely, produces large amounts of carbon-free electricity continuously, regardless of weather or time of day. It runs at capacity factors above 90% — meaning it's generating power more than 90% of the time, a figure no renewable source approaches without storage. For a data center that needs reliable, always-on power to run GPU clusters worth billions of dollars, that reliability has enormous value.

The roster of companies moving into SMRs reads like a who's-who of the AI buildout.

Google made headlines in October 2024 with a groundbreaking agreement to purchase 500 MW of carbon-free power from Kairos Power's SMRs by 2030. In May 2025, Google doubled down, committing early-stage capital to Elementl Power for three U.S. reactor sites totaling 1.8 GW.

Amazon Web Services partnered with nuclear developers to explore SMR deployment near its Virginia data centers, with plans to deploy 5 GW by 2040, including a $500 million investment in X-Energy's 320 MW SMR project in Washington state.

Microsoft is exploring SMRs and microreactors for its global network alongside its headline deal to restart Three Mile Island in Pennsylvania, which will deliver over 800 MW of clean power to its data centers by 2028.

Meta issued a request for proposals to secure 1 to 4 GW of new nuclear capacity in the U.S., and Oracle is planning SMR-powered data centers by 2030.

Globally, the Nuclear Energy Agency's SMR Dashboard identifies 127 SMR technologies globally, with 74 featured in its most comprehensive review to date. Russia and China are already ahead on deployment — Russia and China connected their first SMRs to the grid in 2019 and 2021 respectively — while Western nations are still working through design certification and regulatory approval processes.

It would be dishonest to present SMRs as a solved problem. They are not, and the bullishness of the tech companies backing them needs to be weighed against some hard realities.

Cost is the central challenge. The promise of SMRs is that factory fabrication and standardized design will drive costs down through repetition — the same learning curve that made solar panels 90% cheaper over two decades. But that learning curve requires actually building units at scale first, and right now the economics of first-of-a-kind SMRs are difficult. A conventional nuclear plant has an overnight cost of roughly $6,600 per kilowatt, while an SMR currently runs closer to $10,000 per kilowatt. The argument is that shorter construction timelines — five years for an SMR versus fifteen for a large plant — reduce financing costs enough to make the total investment comparable. But if SMR construction timelines were to extend by just two years, total costs would surpass those of conventional nuclear — a sobering illustration of how thin the margin for error is.

The NuScale warning. The most prominent cautionary tale in the space is NuScale, which had its first planned U.S. SMR project cancelled when costs surged to $9.3 billion — nearly double the original estimate of $5.3 billion. The dramatic cost overruns and delays on recent large nuclear projects in the U.S., U.K., France, and Finland explain why building nuclear is regarded as a high-risk operation. SMRs are supposed to solve this problem — but the confidence is theoretical until the first batch are actually built on time and on budget.

Timeline gaps are real. The earliest commercial SMR operations in the United States are expected in the late 2030s at best, given the time required for design certification, combined licensing, and construction. The tech companies signing deals today are securing power for a decade from now — a long time in an industry moving at AI speed.

The renewables comparison. Critics argue that by the time SMRs come online, solar and battery storage may be cheap enough to make them economically unnecessary. Australian analysis estimates the 2030 cost of solar and wind with storage at A$89–125 per MWh, compared to A$230–382 for SMRs — suggesting that in markets with abundant renewable resources, the cost case for SMRs is weak.

Regulatory and political risk. The political appetite for nuclear power can change, sometimes dramatically, as Germany and Japan have demonstrated. A change in administration, a public incident, or shifting energy policy could reshape the regulatory landscape that SMRs depend on.

SMRs sit at the intersection of three of the most consequential macro themes of the next decade: AI infrastructure buildout, energy security, and decarbonization. That intersection is what makes them worth understanding regardless of whether you believe the technology will ultimately succeed at scale.

The investment thesis is not that SMRs are a sure thing — they're not. It's that the energy problem facing AI infrastructure is real and getting worse, that nuclear is the only carbon-free baseload option available at the required scale, and that a decade of serious capital from the world's most cash-generative companies is going to fund a lot of learning on that cost curve.

The companies best positioned in this story are not necessarily the SMR developers themselves, whose timelines and cost projections carry real uncertainty. The more durable positioning may be in the supply chain: fuel enrichment, nuclear engineering services, specialized materials, and the broader energy infrastructure that has to be upgraded to accommodate a wave of new generation capacity coming online in the 2030s.

The nuclear industry spent 40 years in decline. AI may be what brings it back.

Educational content only. Not financial advice. Brezco Analytics is an independent research and media platform.