The network room at Google’s Council Bluffs, Iowa data center helps connect Google’s data centers to each other. Google
When Microsoft bought a 407-acre pumpkin farm in Mount Pleasant, Wisconsin, it wasn’t to grow Halloween jack-o’-lanterns. Microsoft is growing data centers—networked computer servers that store, retrieve, and process information. And those data centers have a growing appetite for electricity.
Microsoft paid a whopping $76 million for the pumpkin farm, which was assessed at a value of about $600,000. The company, which has since bought other nearby properties to expand its footprint to two square miles, says it will spend $3.3 billion to build its 2-million-square-foot Wisconsin data center and equip it with the specialized computer processors used for artificial intelligence (AI).
Microsoft and OpenAI, maker of the ChatGPT bot, have talked about building a linked network of five data centers—the Wisconsin facility plus four others in California, Texas, Virginia, and Brazil. Together they would constitute a massive supercomputer, dubbed Stargate, that could ultimately cost more than $100 billion and require five gigawatts of electricity, or the equivalent of the output of five average-size nuclear power plants.
Smashing pumpkins. In early 2024, Microsoft purchased the Creuziger family’s popular "Land of the Giants" pumpkin patch (top, as seen in September 2021), adding 407 acres to its nearly two-mile-square data center campus in Mt. Pleasant, Wisconsin. Bottom, architectural renderings of buildings for the facility’s first construction phase, already underway about two miles further south. Google Streetview; Burns & McDonnell
Microsoft, Amazon, Apple, Google, Meta, and other major tech companies are investing heavily in data centers, particularly “hyperscale” data centers that are not only massive in size but also in their processing capabilities for data-intensive tasks such as generating AI responses. A single hyperscale data center can consume as much electricity as tens or hundreds of thousands of homes, and there are already hundreds of these centers in the United States, plus thousands of smaller data centers.
In just the past year, US electric utilities have nearly doubled their estimates of how much electricity they’ll need in another five years. Electric vehicles, cryptocurrency, and a resurgence of American manufacturing are sucking up a lot of electrons, but AI is growing faster and is driving the rapid expansion of data centers. A recent report by the global investment bank Goldman Sachs forecasts that data centers will consume about 8 percent of all US electricity in 2030, up from about 3 percent today.
Bill Gates, Jeff Bezos, Elon Musk, Mark Zuckerberg, Larry Ellison, and other so-called “tech bros” who also happen to be among the world’s richest men have thought about how the energy industry can—or must, in their view—keep pace with AI’s rapid growth while also enabling Big Tech to meet its climate commitments. They have all come to the same conclusion: Nuclear energy, whatever it costs, is the only viable solution.
In a rash of recent announcements, Big Tech companies have declared that they will either be reviving existing nuclear power plants, developing next-generation nuclear reactors, or both. Dollars are also flowing to nuclear fusion projects—even though many physicists think commercial fusion power plants that generate electricity are at least decades in the future, if they ever can be built. The federal government is not only supporting this nuclear-powered vision but also subsidizing it in the name of “clean energy.” However, both the government and the tech industry are largely ignoring the known and significant downsides of nuclear power—including high costs, long construction times, accidents, nuclear weapons proliferation risks, and environmental contamination from uranium mining and radioactive waste disposal.
Back to the future. The Three Mile Island reactor Unit 1 in Pennsylvania is one of two shutdown nuclear reactors (along with one at the Palisades plant in Michigan) that may be brought back online to supply electricity—a first in the United States. The two cooling towers at left are from Unit 2, which was permanently damaged in the 1979 accident. Constellation Energy
Betting on nuclear. Again.
In Pennsylvania, Microsoft has plans to revive Three Mile Island. For people old enough to remember that name, it’s synonymous with the demise of nuclear power in the United States. Forty-five years ago, a partial reactor meltdown at the Three Mile Island nuclear power plant 10 miles south of Harrisburg, Pennsylvania, gripped the nation and exposed nearly two million people to radiation. It was the worst accident in the history of the US commercial nuclear power industry.
The failed reactor never operated again, but a similar reactor built on the same island in the Susquehanna River was restarted six years after the accident and later received a license extension until 2034. That reactor was shut down in 2019 after its owner, Constellation Energy, was unable to secure subsidies from the state of Pennsylvania and deemed the reactor a financial albatross. Now, however, Constellation plans to reopen the reactor and sell 100 percent of the electricity that will be generated by it—enough to power 800,000 homes—to Microsoft.
About 80 miles upriver from Three Mile Island, Amazon recently purchased a new data center next to the two-reactor Susquehanna nuclear power plant. Amazon wanted to increase the amount of electricity flowing directly from the nuclear plant to the data center, but the Federal Energy Regulatory Commission ruled against the change, with one commissioner warning that it “could have huge ramifications for both grid reliability and consumer costs.”
Meta, the company that owns Facebook and Instagram, intended to build a new data center dedicated to AI next to another existing nuclear plant, according to recent reports. But the discovery of a rare bee species at the site threw a monkey wrench into those plans. Had Meta succeeded, it would have been the first Big Tech company to deploy nuclear-powered AI, CEO Mark Zuckerberg reportedly told staffers at a recent all-hands meeting.
Zuckerberg did not say where Meta wanted to build its data center. At least one entomologist has speculated that rusty-patched bumble bees—the first bumble bee species listed as federally endangered—have been spotted near the Diablo Canyon power plant in California, which had been slated to begin decommissioning this year but received a life extension until at least 2030.
In Michigan, the already-shuttered Palisades nuclear plant could be brought back online as early as next year. The reactors at Palisades and Three Mile Island would be the first ones ever restarted after decommissioning.
Prime power supply. Amazon’s recently acquired data center (foreground) in Salem Township, PA, is a stone’s throw from the Susquehanna nuclear power plant. In November, the Federal Energy Regulatory Commission blocked the company’s request to obtain more electricity directly from the plant, bypassing the grid “behind-the-meter.” Talen Energy
The sudden interest in nuclear energy is largely due to AI, which is rapidly transforming the tech industry. Electric utilities are forecasting the nation will need the equivalent of 34 new, full-size nuclear power plants over the next five years to meet power requirements that are rising sharply after several decades of falling or flat demand.
Microsoft, Amazon, and other tech giants are not interested only in reviving existing nuclear plants. They are also funding the development of next-generation nuclear reactors. On October 14, Google announced a deal to purchase nuclear energy from small modular reactors (SMRs) that will be developed by Kairos Power. Two days after Google’s announcement, Amazon said it had signed agreements to invest in four SMRs to be constructed, owned, and operated by Energy Northwest, a consortium of public utilities in Washington state. Amazon hopes the new reactors can power a cluster of energy-gobbling data centers in eastern Oregon. And Oracle is designing an AI data center to be powered by three SMRs, an announcement that Oracle chairman Larry Ellison characterized as seemingly “bizarre” but necessary to meet AI’s “crazy” energy demands.
The first of these next-generation reactors are expected to become operational in the early 2030s, but only three SMRs have been built to date, none of them in the United States.
Keep cool and carry on. An employee diagnoses an overheated CPU at Google’s data center in The Dalles, Oregon. Cooling systems for servers are the largest portion of data centers’ electricity need. Google
Counting “compute”
Globally, electricity demand is also soaring and is now expected to be 6 percent higher in 2035 than the International Energy Agency forecast just a year ago. Electricity consumption by data centers, of which there are already 11,000 worldwide, could reach more than 1 million gigawatt-hours in 2027—about as much total electricity as Japan now uses annually, according to a recent analysis by the agency.
Alex de Vries, who works for the Netherlands’ central bank and in his spare time writes a blog about the unintended environmental consequences of digital technologies, published a peer-reviewed analysis last year that looked at the growing energy consumption of AI. De Vries estimated that if every Google search became an AI-powered Google search, Google’s AI alone could potentially require as much electricity as all of Ireland. Realistically, however, Google’s energy consumption is likely to be constrained by how much “compute” the company can buy.
Prompt response. Estimated energy consumption per request for various AI-powered systems compared to a standard Google search. Chart by Thomas Gaulkin / Datawrapper, from Alex de Vries, "The growing energy footprint of artificial intelligence," Joule, Volume 7, Issue 10, October 2023.
In the AI community, “compute” is everything—even a noun and an adjective. “Compute” is tech-speak for computing power or computing resources: the high-performance processing units that make AI possible.
“I think compute is gonna be the currency of the future,” OpenAI CEO Sam Altman said in a podcast earlier this year. AI developers crave compute, and compute craves electricity.
Tech companies talk about “AI-powered” products as if AI itself isn’t powered by something else, but AI consumes electricity in multiple ways. First, there’s the training required to create AI models such as ChatGPT. Training starts with “scraping” vast amounts of text, images, video, and other data from the internet—essentially taking a gargantuan snapshot of online books, news articles, encyclopedias, patents, photos, and other information found on millions of websites. And because scraping captures only one moment in time, it must be done repeatedly. The article you are reading will likely be scraped by AI.
Developers feed this mountain of raw material to AI models, and they digest it by analyzing patterns in the data—what word tends to follow a series of other words, for example—and using that analysis to form “intelligent” responses to prompts. The models are graded on how well they mimic human-created content (regardless of its accuracy) and then tested repeatedly to fine-tune the answers.
Training is an energy-intensive process. The data sets used for training have grown dramatically over the past few years, and the largest AI models are now trained using hundreds of billions of words, which can take months of processing by tens of thousands of specialized computer chips working day and night.
An analysis done by OpenAI in 2018 found that the amount of “compute” required to train the largest AI models was doubling every three to four months. An analysis of more recent models reports that the training requirements multiplied by four to five times annually during the past four years.
It does compute. The resources, or “compute,” used to train recent AI models grew by as much as five times per year from 2010 to 2024. Compute is a measure of the processing needed to run supercomputer and AI systems, typically expressed in trillions of floating-point operations per second, or teraflops. When OpenAI launched GPT-1 in 2018, it was trained on 18,000 teraflops; when GPT-4 launched five years later, it used 21,000,000,000 teraflops. Chart: Thomas Gaulkin / Datawrapper; Source: Epoch AI, CC-BY
Electricity is also required to process AI queries. A ChatGPT-powered Google query, for example, uses almost 10 times as much energy as a traditional Google search, according to the Electric Power Research Institute. ChatGPT alone responds to approximately 200 million requests per day.
A recent paper by Sasha Luccioni, a researcher at the AI firm Hugging Face, and two co-authors estimates that generating a single AI image can use almost as much energy as fully charging your smartphone.
Quantifying the “compute” used by a particular AI model is easier than estimating the energy used to make the hardware, software, and infrastructure for data centers—and to keep them cool. It's even more difficult to estimate the broader energy impacts of a technology that is already reshaping the labor forces and consumer behaviors of modern societies—for example, using digital devices to perform tasks previously done by hand.
Based solely on current trends, power consumption at US data centers is projected to grow by about 10 percent annually between now and 2030. By one estimate, the exponential growth of AI could consume nearly all the world’s energy production by 2050.
Electric data land. Estimated combined gigawatts consumed by enterprise, colocation, and hyperscale data centers in the United States. Chart: Thomas Gaulkin / Datawrapper; Source: McKinsey & Company
There are ways to increase computing efficiency, and an initiative involving several national labs is focused on getting the semiconductor industry “back on the path of doubling energy efficiency every two years.” However, the AI field is currently more focused on performance—which requires ever-bigger models, training datasets, and processing capability—than on efficiency.
In the meantime, data centers are being built faster than energy capacity is expanding. The rapid growth of this sector has not been adequately figured into climate models and is rarely mentioned as a safety concern about AI. In the March 2023 “pause” letter that called on AI labs to stop training the most powerful AI systems for at least six months, tech experts expressed concern about losing jobs—or even control of civilization—but not about climate impacts.
The AI boom is heavily dependent on power-hungry graphics processing units, or GPUs—specialized computer chips that can process enormous amounts of data. These chips are in short supply, and a multinational corporation called Nvidia commands almost 90 percent of the market. Nvidia, which is based in Delaware but sells chips manufactured in Taiwan and Mexico, recently surpassed Apple and Microsoft to become the world’s most valuable corporation. It is valued at $3.43 trillion, up from $1 trillion only a year ago.
Nvidia’s most advanced platform, which is called Blackwell and is used to train AI models, is a cluster of eight GPUs that together consume 15 kilowatts of power—about half of what a typical US household uses. The entire supply for the next 12 months is already sold out. Demand for it is “insane,” the company’s CEO Jensen Huang said in an interview with CNBC in October.
A doomsday device? Nvidia CEO Jensen Huang announces the new Blackwell architecture at the company's annual GPU Technology Conference on March 18, 2024. He has described demand for the new system as “insane.” Nvidia / YouTube
The financial services company Morgan Stanley estimates that Nvidia will produce 450,000 Blackwells during just the fourth quarter of this year and sell them for about $22,000 each. That would amount to almost $10 billion in revenue, and Blackwell is just one of the GPU models sold by Nvidia. Tech writer John Loeffler calls Blackwell “nothing short of a doomsday device,” because he fears there will not be enough carbon-free energy to power the millions of these devices that are being produced.
And Nvidia will soon have competition. The giant tech companies are working to build AI chips of their own. OpenAI CEO Sam Altman traveled to the Mideast a year ago to solicit between $5 trillion and $7 trillion from investors, including the United Arab Emirates, for a chip-building venture known as Tigris, the Wall Street Journal reported, although that project currently appears to be stalled. Saudi Arabia recently launched a $100 billion fund to invest in AI, and China is challenging US dominance with its Qwen AI system.
“Next year, you will see that they just can’t find enough electricity to run all the chips,” said Elon Musk, speaking at the 2024 Bosch Connected World Conference in February. Musk, who wears many hats in business and will soon assume control of the Trump administration’s so-called Department of Government Efficiency, co-founded OpenAI and recently rolled out an “anti-woke” AI chatbot to compete with OpenAI’s ChatGPT.
Fair is foul and foul is fair. Built in 1955, the Lawrence Energy Center is a coal plant in Lawrence, Kansas, now owned by the Evergy utility company. Originally slated for retirement last year, Evergy announced in June 2023 it would stay open another five years, until 2028. Evergy
A dirty secret
Whether it’s chip manufacturing or bot training and chatting, where will the energy for AI activity come from? Big tech companies have been prominent in efforts to move toward a carbon-free economy. But with the rise of AI, tech-related emissions are going up.
Apple has pledged to become carbon-neutral by 2030. Google set a “moonshot goal” of running its data centers entirely on carbon-free energy by 2030. Microsoft pledged to become carbon-negative by 2030. Amazon, Google, Microsoft, Meta, and Amazon have collectively purchased half of the global corporate renewables market.
But the AI boom has pushed climate goals aside. Microsoft’s emissions, for example, are up by 30 percent since 2020. Google’s emissions have risen by almost 50 percent over the past five years. “As we further integrate AI into our products, reducing emissions may be challenging due to increasing energy demands from the greater intensity of AI compute, and the emissions associated with the expected increases in our technical infrastructure investment,” Google acknowledged in its 2024 environmental report.
At the inaugural AI + Energy Summit held on September 26, former Google CEO Eric Schmidt opined that the tech industry is “not going to hit the climate goals anyway.” Schmidt suggested it would be better to “bet on AI” solving the climate problem than to constrain AI.
“AI has a dirty secret. It’s dirty,” said Leslie Miley, technical advisor to Microsoft’s chief technology officer, in a keynote speech at the QCon Conference in late March. “Generative AI is amazingly energy-intensive, even more so than normal cloud services. ... Google and Meta and Microsoft are all doing their level best to buy green energy, to buy carbon credits. The fact of the matter is, there's not going to be enough.”
In Wisconsin, for example, Microsoft has announced a 250-megawatt solar project to help power its new AI data center in the former pumpkin field. Not mentioned in the announcement: To meet the enormous, round-the-clock electricity demands of the data center, the regional utility has applied to build two new natural-gas-burning power plants, with a total capacity of 1.3 gigawatts. “It’s a massive buildout that will push our state’s climate goals out of reach, locking us into 30 more years of fossil fuels at a time when we all know we must rapidly transition to clean energy,” 14 environmental and health advocacy groups wrote in a letter to Microsoft.
In its April 2024 analysis, Goldman Sachs predicted that global data centers’ demand for power will more than double by 2030, estimating this level of growth would require utilities to invest $50 billion in new power generation capacity. “We assume a 60/40 split between gas and renewables,” the firm reported.
“I can’t recall the last time I was so concerned about the US energy trajectory, as major utilities maneuver for mass gas capacity expansion in the face of load growth,” Tyler Norris, a doctoral fellow at Duke University, tweeted in March. Without a course correction, US emissions goals are “effectively dead,” he wrote.
AI is even breathing new life into old coal plants.
AI already has some beneficial climate applications and could help make renewable energy and the power grid (and data centers themselves) more efficient in the future. With increased efficiency, it’s possible that the power demands of AI could turn out to be lower than currently projected. A November 2023 report by Boston Consulting Group, commissioned by Google, asserted that AI “has the potential to unlock insights that could help mitigate 5% to 10% of global greenhouse gas emissions by 2030.”
But there is no solid evidence that AI can deliver a quick fix for the climate crisis. In fact, AI is also helping the oil and gas industry increase production of fossil fuels. In Guyana, for example, ExxonMobil is using AI “to determine the ideal parameters for drilling” in deep water. For now, at least, AI’s massive environmental footprint is more of a climate problem than a solution.
Recommissioning. The Palisades nuclear power plant on the shore of Lake Michigan. The reactor was shut down for decommissioning in May 2022. But with support from the Department of Energy, including a $1.52 billion loan, Holtec International now plans to restart the plant’s 800-megawatt reactor and install two small modular reactors on the site too. Jim West / Alamy
Is this the “nuclear renaissance”?
As AI’s energy demands grow more intense, and it becomes increasingly clear that the expansion of wind and solar power cannot keep pace, tech leaders have set their sights on nuclear energy.
So nuclear hype has flowed like champagne at a wedding reception.
Proponents of nuclear power have been predicting a “nuclear renaissance” for nearly a quarter-century. But nuclear has never been cost-competitive with other energy sources, and that is unlikely to change anytime soon. The US Energy Information Administration’s Annual Energy Outlook 2023 projected that renewable power would continue to outcompete nuclear, even in scenarios that predict aggressive cost declines for nuclear.
The Biden administration embraced subsidies to keep existing nuclear power plants online and reopen closed ones—for example, a $1.52 billion loan guarantee from the Energy Department is what made it possible for the owner of the shuttered Palisades nuclear plant to announce plans for a reopening. “In 2022, utilities were shutting down nuclear reactors; in 2024, they are extending reactor operations to 80 years, planning to uprate capacity, and restarting formerly closed reactors,” the Energy Department approvingly noted in a report released at the end of Climate Week NYC in late September.
The White House also recently offered $900 million in new funding for small reactors. In its initiative for AI, the Energy Department waves vaguely at plans to “unlock new clean energy sources, optimize energy production, and improve grid resilience.”
“We’re looking at a chance to build new nuclear at a scale not seen since the ‘70s and ‘80s,” Secretary of Energy Jennifer Granholm said at the opening plenary of the American Nuclear Society annual conference in June.
The Energy Department sees potential for a “commercial liftoff” that could triple US nuclear capacity by 2050 and puts a positive spin on AI’s role in boosting nuclear: “AI and data center load growth is aligning the fundamentals for new nuclear with requirements for 24/7 power, valuing decarbonization, and investment in new generation assets.”
Green energy. US Secretary of Energy Jennifer Granholm speaking at an Amazon event on nuclear energy investments, October 16. Granholm announced $900 billion in funding for development of small modular reactors, saying, “The technology companies know that in order for these data centers to achieve great community buy-in, bringing their own power with them is an important piece of that.” The Climate Pledge / YouTube
Despite this federal support, the nuclear renaissance so far lacks an order book for new nuclear plants that are actually being constructed. What it does have, as noted in the White House’s “liftoff” report, is “a set of customers who are willing and able to support investment in new nuclear generation assets.” Namely, big tech companies that can afford to pay big electricity bills.
A team at BestBrokers, which evaluates financial brokerages, looked at how much electricity the 10 largest tech companies that disclose financial data used in their last fiscal year and estimated how long it would take each of those companies to cover that expense based on their average daily revenue. According to this evaluation, Apple, which consumed an estimated 3,487 gigawatt-hours of electricity during its 2023 fiscal year—more than some entire nations—could pay off its entire power bill with about 10 and half hours of its 2023 revenue. Nvidia would need around 11 hours.
Although tech titans currently have ample funds to invest in energy, the cost curve for AI is going up. The expense of powering chatbots is already climbing so fast that companies are holding back their newest versions from the public.
Existing nuclear power can’t satisfy the demand for energy that is not only more abundant but also cheaper. “We still don’t appreciate the energy needs of this [AI] technology,” lamented OpenAI CEO Sam Altman at the World Economic Forum in Davos in January. "There's no way to get there without a breakthrough." Altman, who has warned that AI’s “compute costs are eye-watering,” called for increased investment in nuclear fusion as well as fission.
Breaking new ground, eventually. Bill Gates, center, at a ceremony marking the launch of TerraPower's Natrium reactor project, flanked by the Governor of Wyoming and representatives of TerraPower, the Department of Energy, and engineering and utility companies. Construction of the reactor has not yet begun, and has yet to be approved by the Nuclear Regulatory Commission. TerraPower
The fission (and fusion) frenzy
In addition to OpenAI, Altman also chairs Oklo, a nuclear power startup that went public last year when it merged with a special purpose acquisitions company that Altman also chairs. Oklo plans to build its first liquid metal-cooled sodium fast reactor at Idaho National Laboratory in 2027. However, the company’s initial application for a license was denied—for lack of information—by the Nuclear Regulatory Commission in January 2022 and has not yet been re-submitted.
In August 2023, the Pentagon announced an “intent to award” a contract to Oklo for a small modular reactor at an Air Force base in Alaska. However, the deal was quietly revoked a month later.
Despite setbacks like these, Altman sees the future of nuclear energy and AI as inextricably linked. “I don’t see a way for us to get there without nuclear,” he told CNBC last year.
Retired Microsoft co-founder Bill Gates is not as worried about AI’s energy demands as Altman is, but he too is bullish on nuclear energy. Among his multiple investments in nuclear startups is a company called TerraPower, which has received funding from the Energy Department and Los Alamos National Laboratory to develop a sodium-cooled fast reactor similar to Oklo’s.
Gates has invested more than $1 billion in a TerraPower plant that broke ground in Kemmerer, Wyoming, in June. TerraPower says the reactor will be operational by 2030. But construction of the plant’s Natrium reactor has not yet begun, nor has it been approved by the Nuclear Regulatory Commission (NRC), which is still conducting safety and environmental reviews.
Gates issued a celebratory announcement calling the science behind the reactor “super cool.” Not mentioned in the announcement is the estimated price for the 345-megawatt reactor: $4 billion, of which the federal government is contributing half. Even if the project comes in on budget (which would make it exceptional among US nuclear reactors of the past several decades), it will be more expensive than comparable gas or renewable projects.
Microsoft and Google are also placing bets on nuclear. Earlier this year, Microsoft hired a director of nuclear technologies and a director of nuclear development acceleration to lead the company’s strategy for powering AI advances with small, onsite nuclear reactors—as well as buying energy from larger conventional reactors such as Three Mile Island. Microsoft, which has invested $13 billion in OpenAI and owns almost half of its equity, plans to use AI to expedite the process of getting nuclear plants approved and has been training an AI model on regulatory and licensing documents.
Don‘t be evil. A clip from a promotional video for Google’s partnership with Kairos to develop small modular reactors to “help Google and the world meet electricity demand with carbon free energy every hour of every day.” Google / YouTube
Google last month signed an agreement to buy a total of 500 megawatts of generating capacity—about half the output of a conventional nuclear reactor—from six to seven Hermes small modular reactors designed by Kairos Power. Google aims to deploy the reactors next to Google data centers by 2030. This past summer, Kairos broke ground on an NRC-permitted demonstration reactor in Oak Ridge, Tennessee. It was the first non-water-cooled US reactor approved for construction in more than 50 years.
Michael Terrell, senior director for energy and climate at Google, said the agreement with Kairos could help the company support AI technologies and “reliably meet electricity demands with carbon-free energy every hour of every day.”
Wealthy tech companies and individuals are investing in nuclear fusion as well as fission. Peter Thiel, who co-founded PayPal and was the first outside investor in Facebook, joined Altman in backing a fusion startup called Helion, which claims it will begin producing electricity from its first commercial reactor by 2028 and will sell it to Microsoft.
Breakthrough Energy Ventures, a venture capital firm founded by Bill Gates, has invested in Helion and three other fusion startups. One of those ventures, an MIT spinoff company called Commonwealth Fusion Systems, is also backed by Amazon CEO Jeff Bezos and has received $1.8 billion in second-round venture capital. Commonwealth announced earlier this week that it has leased land to build a commercial-scale fusion power plant in Virginia, but the company has not yet secured any permits or customers.
Critics such as Daniel Jassby, the former principal research physicist at the Princeton Plasma Physics Lab, have called the excitement surrounding these fusion projects “over-the-top and unjustified.”
In an interview with the Bulletin published earlier this month, physicist Bob Rosner, a former president of the American Physical Society and former director of a national laboratory, highlighted the biggest problem with counting on fusion to power the AI revolution: “Climate change is a serious problem which needs to be addressed by ‘decarbonizing’ our energy generation systems sooner rather than later—by 2050 at the absolute latest. And we’re not going to have practical fusion energy in that kind of time frame. There’s no way.”
Tread carefully. A gate at the site of Idaho National Laboratory’s Experimental Breeder Reactor I. On December 10, 1951, it became the first power plant to produce electricity from atomic energy. It was decommissioned in 1964. Dave Meyers / Unsplash
The downsides of nuclear
Despite massive infusions of money from corporations, billionaires, and governments, nuclear is not a sure bet. Around the world, large reactors have repeatedly come in over budget and behind schedule, and although they require lower initial capital investment, smaller reactors are likely to be even less economic than the larger ones that now exist, in terms of the cost of the electricity they produce.
“Very few of the proposed SMRs have been demonstrated, and none are commercially available, let alone licensed by a nuclear regulator,” wrote Allison Macfarlane, who chaired the NRC a decade ago and now directs the School of Public Policy and Global Affairs at the University of British Columbia, in an essay published last year by IAI News.
The only SMR reactor design certified by the NRC is the NuScale Power reactor, which received more than $200 million in federal support and was slated to be built at the Idaho National Laboratory. But the projected cost of building the reactor ballooned between 2020 and 2023; its only committed utility customer dropped out, and the project was canceled a year ago.
The only new nuclear reactors that have been built in the United States in the past 30 years are Vogtle Units 3 and 4 in Georgia. These Westinghouse AP1000 pressurized water reactors were the nation’s first “advanced” reactors, but they ended up costing $35 billion (more than twice as much as originally projected) and were completed seven years behind schedule.
There are more than 50 designs for new reactors, and the tech giants investing in nuclear do not seem to be working toward a standardized design—something that many nuclear experts recommend as the best way to get plants approved and built quickly and affordably.
Reactor proliferation. A rendering of Oklo’s Aurora Powerhouse that will be built on the Idaho National Laboratory site and is planned to be operational in 2027. Reactor concepts like this one produce fissile material that bad actors could use to make nuclear weapons. Gensler
Cost and time are not the only obstacles that must be overcome if nuclear is to meet the AI-energy problem. The work force needed to build a host of nuclear projects has dwindled as plants have closed and not been replaced. Also, none of the proposed new projects has included any new ideas for what to do about the radioactive waste generated not only by reactors by also by uranium mining. There is still no permanent repository for long-lived radioactive waste in the United States.
Radioactive waste isn’t just a disposal problem, either. “Bill Gates should be worried about reprocessing and proliferation,” said Alex Glaser, a nuclear security expert at Princeton University and a member of the Bulletin’s Science and Security Board.
Concepts like the Oklo fast reactor would produce fissile material that bad actors could use to make nuclear weapons. Oklo and other nuclear startups propose to reprocess their waste to keep costs down. But that reprocessing produces plutonium that could be diverted for use in nuclear weapons. The United States rejected reprocessing in the 1970s after determining that the potential for proliferation made it too risky for commercial use.
In all the hype about AI and nuclear, there is scant mention of nuclear weapons proliferation to more countries or the risk that fissile material could be acquired by (or provided to) terrorists. Nor is there much attention to the vast amounts of raw materials and water required for the growth of both AI and nuclear energy, or the electronic waste generated by chip manufacturing and data centers.
A conversation of a few dozen questions with an AI chatbot may require a half-liter of water. A large data center consumes more than a million gallons of water daily, and some data centers are being built in places where water is already scarce.
Developers are loathe to reveal how much water they use, but after a legal battle with an Oregon newspaper, Google finally agreed to reveal that its data centers in The Dalles consume 29 percent of the town’s water supplies. Google plans to build two more data centers there.
Thirsty data. Google’s water use at its facilities in The Dalles, Oregon, increased sharply after it brought a third data center online in 2018. By 2022, the company‘s operations represented 29 percent of the entire city’s water consumption. Chart: Thomas Gaulkin / Datawrapper. Source: The Oregonian / City of the Dalles
No roadmap for “responsible” AI
AI has been compared to electricity—a utility that people soon won’t be able to live without. But there is currently no framework for regulating this new utility, and AI’s energy demands have been given short shrift in the many discussions of AI’s safety risks.
Even researchers who have called for the formation of a new international body to regulate or keep tabs on AI research, perhaps modeled after the International Atomic Energy Agency or the Intergovernmental Panel on Climate Change, have failed to note the connections between AI, energy, and climate change.
The Biden White House has said it wants “responsible” AI. The current National Artificial Intelligence R&D Strategic Plan includes a paragraph noting the dramatic increase in computation demands for AI, and the need to make AI “sustainable,” but has no suggestions for preventing AI’s growth from exacerbating the climate crisis.
The Organisation for Economic Co-operation and Development has likewise raised “sustainability concerns.” But the recommendations in the OECD report on AI’s environmental footprint offered only ideas for measuring AI’s impacts and supporting AI applications to fight climate change.
Generative governance. President Joe Biden speaks before signing an executive order on “Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence” on October 30, 2023. White House photo by Adam Schultz
Democrats in the US Senate and House have proposed legislation to assess the current environmental impacts of AI and adopt a system for reporting future impacts. “The development of the next generation of AI tools cannot come at the expense of the health of our planet,” said Massachusetts Senator Edward Markey, a sponsor of the bill.
But there was little momentum for the bill even before last month’s election. In the United States, at least, deregulation now seems more likely than any regulation of AI or its climate impacts. The nuclear industry may also benefit from deregulation in the next administration; support for nuclear energy is a rare area of bipartisan agreement. Donald Trump has vowed to eliminate federal support for climate mitigation, but has occasionally spoken favorably about nuclear energy.
Both incoming vice president JD Vance and Elon Musk, who have the president-elect’s ear at the moment, are strongly pro-nuclear. “Broligarchs” Musk and Peter Thiel played significant roles in the recent presidential campaign. Thiel reportedly had a hand in pushing for JD Vance, who had previously worked for him, as Trump’s vice presidential pick.
Last month, Thiel spoke with journalist and podcaster Bari Weiss, telling her that Musk—who campaigned extensively for Trump—provided “a great deal of cover” for other Silicon Valley billionaires and business leaders to support Trump.
The tech bros now have a clear path to the unfettered growth of AI and are already pressing Trump to review federal AI policy and weed out laws and regulations that “may be unnecessarily impeding AI adoption.”
Silicon Valley’s AI gold rush aligns almost perfectly with aspirations at Mar-a-Lago, where AI is seen as a must-win race with China. But a second race is also afoot, one in which skyrocketing US electricity demand may outpace supplies, perhaps leading to power outages and utility rate increases of up to 70 percent by 2029.
History suggests nuclear will be a slow starter in that race.
Dawn Stover is a contributing editor at the Bulletin of the Atomic Scientists. She began her career at Harper's magazine and worked... Read More
Dawn Stover
Keywords: AI, Advanced nuclear reactors, Amazon, Google, Meta, Microsoft, NVIDIA, OpenAI, Three Mile Island, apple, big tech, fusion, nuclear plants, proliferation, small modular reactors
Topics: Artificial Intelligence, Climate Change, Nuclear Energy
Get alerts about this thread
0 Comments
Oldest
Bulletin Daily
