Why AI Data Centers Are Bringing Nuclear Power Back From the Dead

Something happened in the energy industry in 2026 that would have seemed unlikely five years ago and impossible ten years ago: nuclear power became cool again.

Not cool in an abstract, policy-document kind of way. Cool in the sense that the biggest, most powerful technology companies on Earth are signing long-term contracts for nuclear electricity as fast as the deals can be structured. Microsoft restarted Three Mile Island. Amazon signed agreements tied to the Susquehanna nuclear plant in Pennsylvania. Google contracted with Kairos Power for small modular reactors. Meta, Oracle, and a dozen others are in various stages of nuclear procurement conversations.

The reason is not a sudden ideological conversion to clean energy. The reason is AI data centers, and a simple, brutal physics problem that no amount of corporate sustainability messaging can solve.

The Problem Nobody In Silicon Valley Wants to Talk About

A single ChatGPT query uses nearly 10 times the electricity of a standard Google search. That fact alone would be manageable. What is not manageable is the scale.

By 2026, global data center electricity consumption is approaching 1,000 terawatt-hours annually. To put that in context: if data centers were a country, they would be the fifth largest electricity consumer on Earth, sitting between Japan and Russia. The IEA projects that figure could reach 1,050 TWh before the end of the year under high-growth scenarios.

The compound annual growth rate of data center electricity consumption since 2017 has been 12% - more than four times faster than total global electricity consumption growth. That trajectory is not flattening. It is accelerating. Goldman Sachs Research projects a 165-175% increase in global data center power demand by 2030 compared to 2023 levels.

And the hardware is getting hungrier, not more efficient. Nvidia's current Blackwell GPU generation already pushed rack power densities to levels that require liquid cooling rather than air. The Rubin generation expected in 2027 is projected to boost thermal design power by 50% over Blackwell, at up to 180 kilowatts per rack. The upgraded Vera Rubin generation after that doubles this to 360 kilowatts per rack.

Each generation of AI hardware consumes more power than the last, at the same time that more of those racks are being deployed. The power demand is not just growing. It is compounding.

Why Renewables Cannot Solve This Alone

The standard answer to energy demand growth in the 2010s was renewables: more solar, more wind, the grid gets cleaner, everyone wins. The AI data center problem breaks this answer in a specific way that is worth understanding.

Solar and wind are intermittent. The sun does not always shine. The wind does not always blow. In most renewable energy deployments, this intermittency is managed through grid balancing - when your solar panels are not producing, the grid draws from other sources. The grid absorbs the variation.

An AI data center cannot absorb variation. It runs 24 hours a day, 7 days a week, 365 days a year, at or near maximum load. It cannot slow down because cloud cover reduced solar output for three hours. It cannot pause training runs because wind speeds dropped overnight. It needs baseload power - generation that runs continuously and predictably regardless of weather conditions.

Goldman Sachs research put the specific number on this: wind and solar, even when paired with battery storage, can meet roughly 80% of a data center's power demand. The remaining 20% requires baseload generation that does not depend on weather. And when you are building facilities that consume hundreds of megawatts continuously, 20% of that is not a rounding error. It is a significant, constant power requirement that renewables cannot reliably fill.

Natural gas can fill the gap. It is dispatchable - you can turn it on and off as needed - and it is relatively fast to deploy. But natural gas is carbon-intensive, and the hyperscalers have made public commitments to carbon-free energy that make large-scale natural gas expansion politically and reputationally complicated.

Which leaves nuclear. Carbon-free. Runs continuously. Output does not depend on weather. Exactly what AI data centers require.

The Deals Being Signed Right Now

The scale of nuclear procurement happening in 2026 is striking enough to be worth documenting specifically.

The conditional offtake pipeline - agreements between data center operators and companies developing small modular reactors - nearly doubled in eighteen months, from 25 gigawatts at the end of 2024 to 45 gigawatts by April 2026, according to IEA data. That is not announcements or letters of intent. That is signed agreements representing committed future capacity.

Microsoft's Three Mile Island restart is the most visible single deal. The Constellation Energy agreement to bring the 837-megawatt plant back online - at a cost of $1.6 billion and under a 20-year power purchase agreement - made Three Mile Island the symbolic centrepiece of the tech-nuclear relationship. Three Mile Island's name carries obvious cultural weight: it was the site of the worst nuclear accident in US history in 1979. Microsoft restarting it to power AI infrastructure is a story that writes itself.

Amazon's nuclear agreements include commitments tied to the Susquehanna Steam Electric Station in Pennsylvania. Google has contracted with Kairos Power for small modular reactors, with the first units expected to come online around 2030. Meta has announced nuclear procurement intent. Oracle, which is building a 1-gigawatt data center campus, has specifically designed the site with nuclear power supply in mind.

The hyperscalers secured more than 10 gigawatts of new or restarting nuclear capacity in the US over the eighteen months through 2025. With the SMR pipeline now at 45 gigawatts, the long-term commitment is significantly larger.

Goldman Sachs forecasts that 85 to 90 gigawatts of new nuclear capacity would be needed to meet all of the data center power demand growth expected by 2030 relative to 2023 levels. The US is nowhere near being able to build that much nuclear in that timeframe - but the direction of investment is unmistakeable.

What Small Modular Reactors Actually Are

Most of the nuclear capacity being contracted for AI data centers is not traditional large-scale nuclear plants. It is small modular reactors, or SMRs - and understanding what makes them different is important context for why tech companies are interested specifically in this technology rather than conventional nuclear.

Traditional nuclear power plants are enormous, expensive, and slow to build. The most recent US nuclear plant completion - Vogtle Unit 4 in Georgia, finished in 2024 - cost around $35 billion and took over a decade to construct. The capital cost and construction timeline make traditional nuclear unsuitable for the pace at which data center operators need to secure power.

Small modular reactors are designed to solve these problems through a different engineering approach. Rather than building a unique, site-specific reactor at enormous scale, SMRs are designed to be standardised, factory-manufactured, and assembled on-site in modular units. Each unit produces less power than a traditional plant - typically between 50 and 300 megawatts - but multiple units can be deployed together, and the factory manufacturing approach is intended to reduce costs and construction time through repetition and standardisation.

The promise of SMRs has existed in various forms since the 1950s. What is different in 2026 is that tech company procurement commitments are providing the demand signal and the financing guarantees that could finally enable SMR developers to build at the volume needed for costs to come down.

NuScale, Kairos Power, TerraPower, X-energy, and a handful of others are the primary US SMR developers with active programmes. None have yet delivered a completed commercial SMR. The IEA projects that the first US SMRs will be commissioned after 2030, meaning the nuclear capacity being contracted for right now will not actually be generating power for AI data centers for at least four to six years.

The Gap Between Now and Nuclear

This creates a significant short-term problem that the industry is managing imperfectly.

The data center buildout is happening now. The capital is being spent now. The GPU clusters are being installed now. The nuclear power to run them sustainably is, at best, four to six years away for SMR capacity and longer for any new large-scale nuclear.

The gap is being filled primarily by natural gas, with renewables handling what they can. In 2024, natural gas was the largest electricity source for data centers at 40%, followed by renewables at 24%, nuclear at 20%, and coal at 15%. That energy mix is not compatible with the public carbon commitments most hyperscalers have made, and the companies involved know it.

The honest framing from Goldman Sachs captures the situation accurately: "nuclear is the preferred option for baseload power, but the difficulty of building new nuclear plants means that natural gas and renewables are more realistic short-term solutions." Natural gas bridges the gap while nuclear capacity is being developed. Whether that bridging period lasts four years or ten depends heavily on whether SMR development programmes deliver on their timelines - which, in the history of nuclear construction, they often have not.

Morgan Stanley Research forecasts a potential shortfall of 49 gigawatts in available US power access by 2028, against a projected demand of 74 gigawatts. If that shortfall materialises, it does not just slow down AI development. It shapes which companies can build in which locations, since data center siting increasingly depends on available grid capacity rather than land availability or tax incentives.

The Grid Overhaul Nobody Is Talking About

The nuclear story gets most of the attention, but it is actually a smaller part of the energy infrastructure story than the grid upgrade that is simultaneously underway.

A $1.4 trillion overhaul of US electricity infrastructure is in progress, driven substantially by data center demand. Fifty-one utilities are involved. The scale of transmission and distribution upgrades required to connect the power being generated to the data centers being built represents the largest upgrade to American electrical infrastructure since the rural electrification programmes of the mid-twentieth century.

This is not optional. The nuclear plants being contracted for need transmission lines to carry their power to data center campuses. The solar and wind farms providing the renewable baseline need grid connections. The data centers themselves need substations and distribution infrastructure capable of handling hundreds of megawatts of continuous load.

The grid is the invisible bottleneck in AI development that gets far less coverage than the chips, the models, and the data centers themselves. In several US regions, the binding constraint on data center construction in 2026 is not land, not capital, not permits for the building - it is available grid capacity. The queue for new grid connections in major data center markets like Northern Virginia stretches years into the future.

What This Means Beyond the Technology Industry

The AI-nuclear energy story has consequences that extend well beyond the technology companies driving it.

For electricity consumers, the surge in data center demand is already contributing to upward pressure on electricity prices in regions with high data center concentration. Northern Virginia - home to the largest concentration of data centers in the world - has seen electricity demand from data centers put pressure on regional capacity that has contributed to rate discussions at the state utility level. As data centers spread to new markets, the same dynamic follows.

For communities hosting data centers and nuclear plants, the economic and environmental questions are layered. Data centers bring construction jobs and tax revenue. They also consume enormous amounts of local grid capacity and significant water resources for cooling. Nuclear plants bring reliable employment and baseload power. They also carry the political complexity that has made nuclear siting contentious in the US for decades.

For energy investors, the AI-nuclear thesis has become one of the clearest investment narratives of 2026. Uranium prices have moved significantly. Nuclear energy companies - Constellation Energy in particular, following the Three Mile Island deal - have seen substantial equity appreciation. The SMR developers are attracting venture and private equity capital at a scale that was not available to the sector two years ago.

The Honest Assessment

The AI-nuclear energy story is real, the scale is significant, and the direction is clear. The hyperscalers are not going to stop building data centers, the data centers are not going to become less power-hungry, and the commitment to carbon-free baseload power makes nuclear the only long-term answer that checks all the required boxes.

The timing reality is also real: SMRs are years from commercial deployment, traditional nuclear takes a decade or more to build, and the grid upgrades needed to connect everything are multi-year infrastructure projects. Between now and the nuclear future being planned, natural gas is doing more of the work than anyone's public sustainability commitments would suggest.

This is not hypocrisy, exactly. It is the predictable consequence of deploying a power-hungry technology faster than clean power infrastructure can be built to support it. The companies involved are aware of the gap and are making long-term commitments to close it. Whether those commitments arrive on schedule, or are pushed by the same forces that have historically made nuclear construction timelines optimistic, is the open question.

What is not open is whether AI's energy demands are creating a fundamentally different relationship between the technology industry and the power sector. That relationship has already changed. The tech companies that once bought renewable energy certificates to offset their consumption are now signing nuclear power purchase agreements, financing SMR development programmes, and shaping investment in electricity infrastructure at a scale that makes them significant actors in energy policy.

The data center and the nuclear plant, once unrelated industries, are becoming structurally dependent on each other. That dependency will shape both industries for the next twenty years.

It already is.