You probably don’t think much about it, but the ChatGPT query you just ran uses a lot of power. According to the Allen Institute for Artificial Intelligence, a single query to one of the most popular AI chatbots on the market uses as much power as a lightbulb for 20 minutes—more than 10 times the power consumption of a simple Google search.

Now, consider that experts predict that the 2,700 AI data centers in the United States, which process millions of AI requests from users every day, will consume six percent of the country’s total electricity consumption by 2026, a full two percent more than the total consumption in 2022. According to the International Energy Agency, data centers, AI, and the cryptocurrency sector consumed an estimated 460 terawatt hours in 2022, more than a tenth of total U.S. electricity consumption that year. As companies like Apple, Google, and Microsoft launch ever more powerful AI tools and look to build even more data centers to support them, a power problem could be looming.

Gregory Nemet, a professor at the La Follette School of Public Affairs whose research focuses on the state of sustainable energy, says the question may be less about the search for future power sources and more about the future of artificial intelligence itself.

“We don’t know what’s going to happen with AI,” he says. “There’s a lot of growth right now, but there might be a point where it starts to taper off because we’re not getting any real benefit from it anymore. And that’s the big open question that nobody really has an answer to.”

Since the 2008 recession, Nemet says, demand for electricity in the U.S. has grown by 1% annually—at least until recently. The combination of the proliferation of AI data centers and the advent of electric vehicles has already pushed up demand and could potentially double it over the next few decades. Still, Nemet is confident the U.S. will find ways to keep up with the increase.

“One of the advantages of AI is that it is relatively location-independent,” says Nemet. “So you can build data centers and servers for AI near wind and solar power generation and then simply lay fiber optic cables instead of expensive transmission cables.”

This is already the case in some parts of the country. The U.S. Department of Energy’s National Laboratories has already built its own AI data center near a nuclear reactor, which provides much of the power for the data center. Amazon is doing the same. Nemet said companies like Microsoft, Apple and Google – three of the key frontrunners in AI development – have become major buyers of renewable energy as part of a broader strategy to reduce carbon emissions.

But the picture is bleaker elsewhere. Public utilities in states like Georgia, Virginia, Washington and Texas are already struggling to meet the rising electricity demands of AI data centers. Some of them are turning to fossil fuels like natural gas to fill the gap. But this strategy requires large and sustained investments in infrastructure and runs counter to nationwide efforts to reduce carbon emissions.

“Meeting that electricity demand can be done well or badly,” says Nemet. “You can do it with natural gas, which will cause a lot of problems later, or you can do it with clean energy, which will avoid those problems.”

Earlier this year, Matt Sinclair, assistant professor in the Department of Computer Science and Computer Architecture, spoke at a National Science Foundation workshop on sustainable computing, which featured lively discussions about the growing potential of AI—and the rising costs of running it. Sinclair was not surprised. He has watched in recent years as the maximum wattage of graphics processing units (GPUs) from companies like NVIDIA and AMD has increased as they try to run massive new AI models.

“Companies are aware of this and are developing more efficient hardware. However, they are also using larger models that require more and more power,” he says.

The situation reminds Sinclair of Jevons’s paradox, the idea that increasing the efficiency of a resource increases its overall use rather than decreasing it. The paradox was once used to describe coal consumption on trains in the 19th century, but it seems even more apt when applied to AI and electricity.

“As AI becomes more efficient and uses less energy, paradoxically it increases overall energy consumption,” he explains. “Because now you can use it in ways that weren’t possible before. This in turn enables greater use of AI and increases overall energy consumption.”

To find a balance, Sinclair believes it will require collaboration between groups that have not traditionally worked together: the utilities building power transmission lines, the computer scientists developing the software and hardware that makes artificial intelligence possible, and the U.S. government, which may need to regulate the entire system.

“And that leads to some really interesting research questions,” he says.

Elsewhere in the computer science department, Karu Sankaralingam, the Mark D. Hill and David A. Wood Professor of Computer Science, is tackling the efficiency problem head-on. Sankaralingam has spent most of his career finding ways to make computer chips and processors do more while using less power.

Sankaralingam says that the companies that are rushing to market their AI products to capitalize on their enormous popularity are not necessarily bringing the most efficient tools to market. Ten years ago, it was possible to make efficiency gains by improving the functionality of the transistors that pass electricity through an operating system. Those gains have now largely been exhausted.

Some engineers have chosen to develop customized efficiency solutions for specific AI programs.

“AI algorithms change very, very quickly,” explains Sankaralingam. “And it takes a long time to build and sell a chip. By the time you get something to market, that algorithm has already changed, making the chip you just built irrelevant.”

Instead, Sankaralingam and graduate students in his lab have chosen a different approach, which he calls “efficient generalization,” which involves applying a set of energy efficiency principles to different applications rather than focusing on a single one.

This approach has opened up new horizons for Sankaralingam’s research, which revolves around a central question: How much can efficiency be increased without losing its generality? Initial simulations in the laboratory show that there is still room for improvement. Microsoft, for example, recently improved its chatbot servers so that they consume ten times less energy.

“I think people will develop other solutions that use less energy, and I think we’re starting to see some of that,” he says. “People are aware that energy consumption needs to be controlled for this technology to achieve its ultimate vision and reach.”