The Three Mile Island Nuclear Power Plant has two reactors. The second reactor was the site of an infamous partial meltdown in 1979. and it hasn’t worked since. But the first reactor continued to operate without incident until 2019, when it was shut down for financial reasons — mainly due to competition from gas- and wind-powered electricity. Kotek says there are relatively few idle reactors that can also be brought back online relatively quickly, but that many power plant owners are interested in extending their licenses to operate their existing plants to try to ride the AI wave of power .
Part of the enthusiasm of power plant operators is due to government incentives to keep low-carbon power online. The deflation law includes tax credits tied to electricity production at existing nuclear power plants, but Kotek says the industry will also need to build new reactors if it wants to capture that projected energy demand. The number of operating nuclear reactors in the United States peaked at 112 in 1990. and down to 92 by 2022, and the most recently built reactors in the US – at the Vogtle power plant in Georgia – took more than 14 years to build and came in at more than twice the expected budget.
“The U.S. showed Vogtle that we’re not very good at building plants,” said Todd Allen, chair of nuclear engineering and radiological sciences at the University of Michigan. But Allen points out that China appears to be building nuclear power plants much faster than the US, so acceleration is possible, and that if demand for data center power continues to grow, then building entirely new plants will look an increasingly attractive option.
Those potentially long timelines are part of why Microsoft is interested in small modular reactors, which should be faster and cheaper to build. But tech firms tend to focus on finding new energy sources instead of improving the efficiency of their AI operations, says Sascha Luccioni, head of AI and climate at Hugging Face, a company that develops tools to build apps that use machine learning. “Regulation can be one way to incentivize [great efficiency]starting with mandatory reporting and transparency for companies providing AI tools and services,” she says.
At the event at Carnegie Mellon University, Pichai said work on improving AI’s energy consumption is still in its “early stages.” “We’re all inefficiently pre-training these models, absolutely,” he said, but added that inference — actually asking an AI model to perform a task — could become “dramatically more efficient over time.” Google Feeds in 2023 were 48 percent higher than their 2019 baseline, primarily due to increased data center energy use and supply chain emissions, setting Google’s goal of reaching net zero emissions by 2030. under increasing threat. “AI’s energy needs are growing right now,” Luccioni says, but renewable or low-carbon energy to power AI isn’t keeping pace fast enough.
For some, the prospect of the site of America’s most famous nuclear disaster being used to power the AI revolution may prove unsettling. But Allen points out that the first reactor was not shut down due to operational problems. Restarting the reactor, he says, will mostly be a matter of making sure it’s still in good working order and that there are enough trained personnel to run it smoothly.
