Why Does AI Need Nuclear Power?

As the new wave of generative AI scales up, so does its hunger for energy. Data centers are expanding at an unprecedented rate, putting immense strain on power grids and worsening years-long backlogs for electrical connections. With AI adoption accelerating across industries, big tech companies are tapping nuclear power as an alternative low-emissions energy source — going so far as to revive a dormant nuclear plant and bet big on new, compact reactors to secure their power supply.

While nuclear energy is clean, its history is far from spotless. Have technological advances made it safer and more practical? And can these nuclear solutions arrive fast enough to power the future of AI?

Big Tech, Big Energy

Data centers are booming. As of 2024, there were over 7,000 facilities worldwide, or double the number in 2015. Bloomberg estimates they consume more energy each year than Italy…as in, the entire country…and by 2034, they may rival India’s energy use.¹

What’s driving this surge? The rising demand for new forms of generative AI and cloud services, which require massive computational power. Electricity use by Meta, Amazon, Microsoft, and Google — the giants powering this digital ecosystem — more than doubled between 2017 and 2021.²

And it’s not slowing down. Generative AI is being called a game-changer, with consulting firm McKinsey & Company estimating it could add up to $4.4 trillion to the global economy.³ But there’s one major catch: we need the power to run it.

Generative AI services guzzle energy. A single ChatGPT request uses about 2.9 watt-hours of electricity. That’s 10 times more than a Google search and enough energy to power a 60-watt incandescent light bulb for 3 minutes. Multiply that by millions of daily requests, and the energy use adds up faster than you can say “write up a recipe for lasagna that doesn’t use ricotta cheese.” Advanced tasks like image, audio, and video generation push these demands even higher.²

Efficiency gains are in the works, with custom chips being developed to handle specific workloads more efficiently than traditional graphic processing units.⁴ But breakthroughs in cutting-edge models still come with big increases in computational and energy demands.⁵

Generative AI’s growing appetite could account for up to 40% of increased electricity needs between now and 2030.³ By then, data centers might be consuming 11 to 12% of all U.S. electricity, requiring an additional 50 GW of power, or roughly the output of 50 large power plants, alongside a staggering $500 billion investment in infrastructure.³ That’s a shocking electricity bill.

With demand soaring, it’s unclear whether energy production will scale up fast enough. Lead times to power new data centers are growing, and some city hubs, like Amsterdam, Dublin, and Singapore, have at one point temporarily banned new data center construction.³ The problem isn’t just the availability of power; upgrading or installing critical infrastructure, like substations, transformers, and transmission lines, often takes years to complete.

And then there’s the environmental cost. All this energy draw coupled with a shortage of clean power is making it harder for Big Tech to curb emissions. Microsoft pledged to be carbon negative by 2030, but its emissions jumped 29% from 2020 to 2023, largely due to the construction of new data centers for AI.⁶ Google’s emissions climbed even faster, rising 48% between 2019 and 2023. The company still aims for net-zero emissions by 2030, but it’s going to need a lot of clean energy to approach that.⁷

Embracing Nuclear

Solar and wind power are cornerstones of clean energy, but they clock out when the sun isn’t shining or the wind isn’t blowing. Meanwhile, data centers keep humming 24/7. So, what’s the backup plan? For many, it’s nuclear power.

Enthusiasm for nuclear power in the U.S. took a significant blow after the Three Mile Island accident in 1979 and the Chernobyl disaster in 1986. Fun fact, I was born not too far away from Three Mile Island. Even though public confidence in nuclear power waned, other reactors at Chernobyl and the twin reactor at Three Mile Island continued operating. Today, 94 reactors across 28 states still crank out about one-fifth of the nation’s energy.⁸ Not bad for a technology some thought was headed for extinction.

And confidence in nuclear power is rising again, reflecting a growing realization that air pollution caused by emissions from fossil fuel plants have caused far more deaths per terawatt-hour than nuclear power.⁹ Even politicians are coming around: California Governor Gavin Newsom has pushed to keep open the Diablo Canyon nuclear plant he once sought to close, and five states recently reversed bans on building new reactors.¹⁰

That might surprise anyone who’s watched the price of solar and wind power plummet even as nuclear power has gotten more expensive.¹¹ But for Big Tech, the expense is justified: data center construction costs can run into the billions and that doesn’t include the cutting-edge tech—the servers, processors, and storage hardware—that add an additional 30% or so to the price tag and need replacing every five to six years¹²¹³¹⁴. Keeping these massive investments running at full capacity is far more critical than the cost of electricity itself.

Nuclear Options

It’s no surprise, then, that major companies are already cozying up to nuclear power, even locating data centers next to power plants to bypass grid bottlenecks and secure the massive energy supply their operations demand. In March 2024, Amazon Web Services paid $650 million to Talen Energy for its 960 MW Cumulus data center campus directly connected to the ~2,500 MW Susquehanna nuclear plant in Pennsylvania.

The datacenter already had rights to 300 MW but requested an additional 180 MW. The Federal Energy Regulatory Commission rejected this, citing concerns over grid reliability and cost fairness.¹⁵ Its hesitancy makes sense. A recent survey by the Electric Power Research Institute found that more than a quarter of utility companies experienced thermal and voltage violations as a result of the operational impacts of connected data centers.¹⁶

This growing tension between data center demands and grid limitations isn’t deterring other tech giants from trying to tap into nuclear power. But sometimes nature has other plans: Meta’s request for an AI data center near a nuclear facility hit a real stinger when reports of endangered bees in the area shut down the project.⁸

So what’s a tech giant to do? Well, if you’re Microsoft, you dust off an old reactor and bring it back to life. In a big move, Microsoft signed with Constellation Energy, a company already in charge of a dozen other nuclear power facilities across the US, to restart the remaining 800 MW reactor at Three Mile Island in Pennsylvania.¹⁷¹⁸
The reactor shut down in 2019 due to competition from cheaper natural gas and subsidized renewables. But with demand for reliable energy surging, Microsoft has committed to buying all its power for twenty years. Meanwhile, Constellation Energy will pour $1.6 billion into restoring the reactor.¹⁸

Microsoft’s going nuclear, but it’s not a bargain. Analysts estimate the power will cost $110-115 per MWh, well above typical low $100s rates for co-located energy.¹⁹ So why pay a premium? Because locking in constant power at a steady rate insulates Microsoft from future price surges and helps it hit its emissions targets.

Other closed nuclear plants are being considered for revival, too, like the Duane Arnold nuclear power plant in Iowa.²⁰ As the founder of Radiant Energy Group’s Mark Nelson puts it:

“Existing nuclear plants are the hottest thing in power right now. They’re going to be able to nearly name their price to build out to data centers that are parked right at their gate.”²¹

Can Small Modular Reactors Spark a Nuclear Renaissance?

With so much renewed interest in nuclear, why not build large, new facilities from scratch? That’s because the only two reactors to be approved for construction in the US since the ‘70s finished terribly late and grossly over budget.²² Two ~1,000 MW reactors added to the Vogtle power plant in Georgia broke ground in 2009 with a $14 billion price tag, but weren’t operational until 2023 and 2024, after $35 billion had been poured into the project. By the time the reactors came online, the average consumer had plonked down $900 and would continue paying $9 more per month for their energy bill.¹⁰

Yet nuclear expert and senior adviser at the plant’s Energy Department, Julie Kozeracki, would go on to say:

“The real tragedy with Vogtle is that we stopped after two units. Three of the biggest issues were starting with an incomplete design and construction plan, an untrained workforce and an immature supply chain. We solved all three and then stopped.”¹⁰

She raises an important point: the approach to nuclear plant construction needs to move beyond bespoke, one-off projects. That’s why a new generation of nuclear entrepreneurs aims to emulate Henry Ford by introducing assembly-line-style production for nuclear reactors, relying on skilled, repetitive labor and well-developed supply chains to drive efficiency and scalability.

The latest efforts at re-energizing nuclear power in the US are focused on SMRs, or small modular reactors: small because they’re up to 300MW, about a third the size of a traditional reactor; modular because they’re made from factory-built components designed for transport and quick assembly at the project location; and reactors because… well, splitting atoms is kind of the point.

SMRs are flexible. They can plug into the grid in rural areas with less infrastructure or deliver power directly to industries like data centers or desalination plants. And because they’re modular, you can add more reactors as energy needs grow, keeping upfront costs lower and making power production more adaptable.

The modern designs of SMRs also offer a bunch of practical and safety benefits. They typically operate at lower pressures, need to be refueled less often, and are constructed with passive safety shut-offs.²³
And passive safety shut-offs are a game-changer in ensuring future nuclear power plants avoid the disasters of the past.

Take Fukushima, Japan, in 2011: it wasn’t the 9.0 magnitude earthquake that caused the nuclear disaster. Backup generators kicked in as planned to keep cooling systems running. But when the earthquake-triggered tsunami breached the sea wall, it flooded the plant, knocked out the generators, and halted the cooling of the reactor cores. The result? Overheating, meltdowns in three reactors, and the release of radioactive material.²⁴ Passive safety systems, however, eliminate these vulnerabilities. They don’t rely on backup generators or human intervention to keep reactors safe.²³

The SMR designed by California-based Kairos Power uses a molten fluoride salt coolant that’s barely pressurized. This design minimizes the risk of high-pressure failures and eliminates the need for constant coolant replenishment, since the molten salt doesn’t evaporate like water in traditional water-cooled reactors.²⁵

The SMR from Maryland-based X-energy is also a low-pressure design, but with helium gas as its coolant. If the reactor overheats, a graphite moderator within the system becomes less effective at sustaining the nuclear reaction, preventing an accident without the need for external intervention.²⁶

Both reactors use what’s called tri-structural isotropic (or TRISO) particle fuel made of uranium, carbon, and oxygen, wrapped in multiple layers of ceramic and graphite. Think of each TRISO particle as its own tiny containment vessel, capable of trapping nuclear fission products inside. The particles also stay intact at extremely high temperatures, keeping the fuel from melting even in worst-case scenarios.²⁶ With coolants and fuel stable across wide temperature ranges, safety is built directly into the design of these SMRs.

In October 2024, Google commissioned seven Kairos reactors to add 500 MW to the US grid, operated by Kairos themselves. The first SMR is expected to come online by 2030, with additional units planned through 2035.²⁷ Meanwhile, Amazon has teamed up with the power utility consortium Energy Northwest to commission and run four SMRs from X-energy, totaling 320 MW in the initial phase. Amazon ultimately aims to bring 5 GW of X-energy SMRs online in the US by 2039.²⁸

Google and Amazon’s approach of adding energy to the grid offers far more flexibility than Microsoft’s strategy of purchasing the entire output of a single nuclear reactor. They purchase only the energy they need, with excess energy sold into the grid by the utilities operating the reactors. This boosts access to low-carbon energy in local communities and helps offset the hefty upfront costs of building SMRs. Plus, additional reactors can be added over time, scaling power production as energy needs grow.²⁸

But there’s a catch: these shiny new SMR designs still need regulatory approval from the U.S. Nuclear Regulatory Commission (or NRC) before they can be built.²⁸²⁹

Can Regulations Keep Pace with Innovation?

A major barrier to the low-cost, assembly-line manufacture of SMRs is the regulatory complexity involved in sourcing materials for a nuclear reactor. They also conflict with global ambitions to boost nuclear power. At COP28, 22 countries, including the United States, pledged to triple nuclear capacity by 2050 compared to 2020 levels.³⁰

And then there’s the fuel issue. Since the 1990s, most nuclear fuel production has been outsourced to Russia. After the invasion of Ukraine, companies like X-energy have had to delay licensing and start dates due to a lack of nuclear fuel.^{31 32} In response, five nations, including the U.S., are mobilizing $4.2 billion to establish a global enriched uranium market independent of Russia, but these solutions will take time.³³

So far, only one SMR design has received NRC certification: NuScale Power’s 77 MW reactor which used the same technology as larger, traditional reactors. Getting that approval took a decade and cost $500 million¹⁰, but inflation and rising interest rates nearly doubled NuScale Power’s electricity prices, and the company couldn’t secure enough customers to keep the project afloat. It failed.³⁴

The NRC claims it’s working to improve, citing its recent approval of Kairos Power’s Hermes test reactor in just 18 months. To be clear, this is a non-powered test reactor, not a commercial SMR. But it does signal a shift towards performance-based evaluations — the kind of approach needed to bring advanced SMR technologies from concept to reality.¹⁰³⁵

The rapid rise of generative AI and explosive growth of data centers have sparked renewed interest in decommissioned nuclear plants and the development of newer, safer SMRs to meet their enormous energy demands. I’m actually working on visiting one of these facilities to get an inside look, so stay subscribed and keep an eye out for that. As the push for decarbonization accelerates, nuclear energy seems to be standing out to tech giants as a reliable, low-carbon solution. While the timeline for new nuclear plants will depend on both technological progress and regulatory approval, one thing is clear: if AI is going to reshape the world, it’ll need a lot of watts to do it.