AI’s Energy Demands Threaten a Nuclear Waste Nightmare

AI’s Energy Demands Threaten a Nuclear Waste Nightmare

By Michael Riordan edited by Dan Vergano

Long in decline, the U.S. nuclear industry is hoping for resurrection at two sites of its greatest failures: Three Mile Island in Pennsylvania and the Hanford Site in Washington state. Nuclear power, the industry claims, will help satisfy the surging power demands from data centers and the growing AI economy. But such a wrong turn ignores the long-unresolved problems of radioactive nuclear wastes that AI cannot wish away.

In September Constellation Energy announced plans to restart a shuttered reactor at Three Mile Island, prodded by Microsoft, which will need many gigawatts of power to perform extensive AI calculations in its expanding fleet of data centers. Amazon followed suit and announced in November that it will invest $334 million to develop small modular nuclear reactors (SMRs) at Hanford, site of the world’s first plutonium-production facility.

Google and Meta are also hoping to bring nuclear power back. In October 2024 Google announced it eventually plans to purchase 500 megawatts of electricity from Kairos Power, which is developing a novel SMR in Oak Ridge, Tenn., on the site of the national lab that long refined uranium for the nuclear industry. And Facebook parent Meta is seeking bids for nuclear power plants for its data centers.

On supporting science journalism

If you’re enjoying this article, consider supporting our award-winning journalism by subscribing. By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today.

These tech giants recognize that the next generation of microprocessors to be used for AI calculations at data centers will require oodles of electricity to power and cool them. A single Nvidia Blackwell chip, for example, can draw up to two kilowatts, more than what is needed for a typical house. Cram thousands of them in servers inside a data center, and they will need as much power as a small city.

So-called hyperscale data centers require over 100 megawatts (100 MW)—a sizeable fraction of the output of a major power plant. And that power should be cheap, steady and reliable.

An authoritative December 2024 report from the U.S. Department of Energy, written by energy experts at the Lawrence Berkeley National Laboratory, is especially illuminating. The growth in U.S. data-center energy usage over the next five years, they state, would correspond “to a total power demand for data centers between 74 and 132 [gigawatts].” That would represent some 7 to 12 percent of the U.S. electricity consumption forecast for 2028.

Where on Earth is all this power going to come from? Given the challenges electric utilities face in supplying electricity to meet other growing needs, including electric vehicles, it’s small wonder that big tech has turned back to the atomic nucleus. But the power demands outlined in the DOE report would require building or resurrecting the equivalent of at least 40 Three Mile Island reactors over the next five years. That’s impossible.

Several years ago Amazon, Google, Meta and Microsoft promised not to exacerbate atmospheric carbon-dioxide levels. But that laudable goal is becoming increasingly difficult to achieve, given their data centers’ exploding electricity consumption. So they have instead begun touting a return to nuclear power to avert this thorny problem. That’s a huge mistake.

Nuclear power is indeed a source of carbon-emission-free energy, but it is hardly a clean energy source, and it is definitely not renewable. All along the uranium supply chain—from mining to enrichment to the fabrication of fuel rods or pellets—opportunities abound for radioactive releases. In South Texas, for example, landowners worry about contamination of their groundwater by renewed uranium mining activities nearby.

Since 1989, the DOE has spent hundreds of billions of taxpayer dollars cleaning up the original nuclear complex—including the gargantuan Oak Ridge factory that enriched much of the uranium used for commercial nuclear power. And despite decades of trying, the department has yet to fully clean up and dismantle the oozing, disintegrating tanks of highly radioactive wastes left over from plutonium processing at Hanford.

The storage and containment of spent nuclear fuel is in fact the crucial unresolved challenge of the U.S. nuclear industry. Over 90,000 tons of these wastes are stored at 77 sites in 35 states—an amount increasing by over 2,000 tons a year.

Small modular reactors, promoted by Microsoft co-founder Bill Gates and others, will only add to this growing burden. As former U.S. Nuclear Regulatory Commission (NRC) chair Allison Macfarlane and Rodney C. Ewing of Stanford University stated, “In some cases these new reactors may make it worse by creating more waste that’s more costly to manage, new kinds of complex waste, or just more waste, period.”

Elsewhere, Macfarlane stressed the procedural and practical difficulties faced by novel nuclear reactor technologies in gaining NRC acceptance and achieving commercial success. Shortly after it had received NRC certification in 2023, for example, the much-touted NuScale SMR project was abandoned after anticipated construction costs more than doubled to $9.3 billion. Leaving aside the waste problem, a commercially successful SMR design is probably over a decade away.

But the relentless AI gold rush, if left unchecked, will impose unattainable demands on projected power supplies well before that. Meanwhile, electricity rates will rise inexorably in light of the law of supply and demand. That looming energy crisis explains big tech’s efforts to slow shutdowns of fossil-fueled power plants and to resurrect shuttered reactors.

Amazon, Google, Meta and Microsoft executives should instead take a deep breath and begin reevaluating their options. Do they really need to build and upgrade data centers at such a breakneck pace? Or is this devil-take-the-hindmost AI arms race just the result of bitter competition, prodded by recent advances in semiconductor technology?

And what about the truly clean, renewable energy sources they once embraced—especially solar, wind and geothermal? Yes, the variability of solar and wind energy makes them a poor match to the steady power requirements of data centers. But energy storage has come a long way recently and has a promising future. And the recent startling success of the Chinese DeepSeek AI program demonstrates that software efficiency will play an important role in this effort.

Given the dark clouds still lingering over nuclear power, especially its unresolved waste problem, these renewable alternatives deserve renewed consideration.

This is an opinion and analysis article, and the views expressed by the author or authors are not necessarily those of Scientific American.