The Current24:10AI needs a lot of power. So Google’s going nuclear
Big tech companies are scrambling to secure nuclear energy deals worth billions of dollars in order to meet the growing demands of generative artificial intelligence (AI) — but critics say they need to rethink that and slow down.
“Tech companies have gotten away with a lot just because it’s a new area,” Sasha Luccioni, AI researcher and climate lead at New York-based AI developer HuggingFace, told The Current’s guest host Peter Armstrong.
“The approach tends to be, ‘move fast and break things,’ in start-ups and Silicon Valley. And so what worries me is that approach transposed to nuclear energy, because nuclear energy is something that has to involve a lot of care.”
In September, Microsoft announced their plans to purchase power from the Three Mile Island plant in Pennsylvania, where the worst nuclear accident in U.S. history took place in 1979. The site shut down its remaining operating unit in 2019 but plans to reopen for this project.
Last week, Google announced a deal with nuclear startup Kairos Power, buying 500 megawatts of electricity from seven small modular reactors (SMRs) that have yet to be built. And just a few days ago, Amazon struck a similar deal with the startup X-Energy.
Small modular reactors are considered safer than traditional nuclear plants due to their smaller size and power output, according to proponents of the technology. They’re also touted as cheaper and faster to build, which will ideally provide the tech companies with a source of renewable energy — in Google’s case, as early as 2030.
Companies are racing toward nuclear because they are using exponentially more power now, due to the upsurge of generative AI use. Both Microsoft and Google missed their net zero emissions goals for 2024 — a first, because these companies tend to set goals they can meet, according to Luccioni.
“This year they’re like, “Oh, oops — AI caught us by surprise,” she said.
Some see the Microsoft, Google and Amazon investments in nuclear energy as a positive step toward reducing Big Tech’s emissions.
“I think all of that is great for the environment,” said Jacopo Buongiorno, a nuclear science and engineering professor and director of MIT’s Center for Advanced Nuclear Energy Systems.
How much power does generative AI use?
Generative AI can create text or multimedia such as images, music or video based on a user’s request for a dinner recipe or something more technical, like a house offer letter to win a bidding war.
The systems are trained on large datasets and programmed with specific rules and algorithms — so with each request, the AI identifies patterns in order to provide an answer.
“They’re trying to stick that into anything that moves, anything that a customer can interact with,” Luccioni said of the big tech companies.
Generative AI takes more power to perform than, for example, streaming shows and scrolling through your social media feed, says Luccioni, who has researched AI’s environmental costs and is currently developing a rating system for the technology.
Put another way, one ChatGPT query uses as much electricity as a lightbulb needs to be on for 20 minutes, according to a researcher cited by NPR.
Luccioni says companies should be mandated to release how much energy its AI tech uses, and to allow users to opt out of using it.
Once there is more information about generative AI’s energy usage, she says that governments will be able to develop regulation around its use.
“We should be curbing the demand,” she said.
Instead, as demand rises, tech companies are investing in nuclear energy and building more AI data centres, which are the mega-warehouses full of servers, routers and cooling systems that power AI.
“There is going to be a couple of years when the demand is going to be outstripping any kind of renewable energy supply,” Luccioni said.
‘An ideal energy source’: Nuclear energy prof
Tech companies have been actively looking for a renewable energy source to sustain the rapid growth of generative AI. In 2023, Google built a geothermal power project in Nevada to power its nearby data centres. And Amazon recently became the world’s top corporate buyer of solar and wind energy.
“The use of nuclear reactors is almost like an ideal energy source,” Buongiorno said. “And that’s because nuclear reactors produce electricity 24/7.”
Buongiorno believes it is “quite realistic” for nuclear power to run constantly, without carbon emissions. He says that the small modular reactors that the companies will be using produce less energy than the larger reactors currently on the grid, but are reliable enough to power the data centres that perform AI calculations.
Wind and solar energy, comparatively, need to be backed up with energy storage batteries and that will add to the cost and land usage associated with that power generation, Buongiorno said.
“Building a new nuclear plant is expensive. The good news is that once you have built it and amortized it, then the marginal cost of operation and maintenance plus fuel are actually pretty low and predictable over decades.”
As for concerns about radiation from the SMRs, Buongiono says he has “virtually none,” due to the smaller size and energy output than conventional nuclear power plants.
They also have “passive safety systems,” according to one proponent of Canada’s investment into SMRs.
“If things go wrong, there is no need for human intervention, things just shut down. And so that’s probably one of the biggest differences,” Rumina Velshi, former president and CEO of the Canadian Nuclear Safety Commission, told CBC’s Edmonton AM earlier this month.
The reactor at Three Mile Island is expected to be back up and running by 2028, pending approval from the U.S. Nuclear Regulatory Commission. Google’s deal with Kairos Power projects the first SMR to be ready by 2030, whereas Amazon and X-Energy are planning for 2039.
However, the first SMR project in North America is under construction right now, operated by Ontario Power Generation (OPG) at the Darlington plant in Clarington, Ont. It’s expected to be the first in North America to go online and it is on schedule to become operational by 2029.
“If the investment is there, and the need is there, the demand is there, it’s absolutely feasible,” Buongiorno said.
