You probably don’t think much about it, but there’s a lot of power consumed by that ChatGPT query you just executed. According to the Allen Institute for Artificial Intelligence, a single query to one of the market’s most popular AI chatbots consumes the electrical equivalent of powering a lightbulb for 20 minutes — more than 10 times the power cost of a simple Google search.
Now consider that experts project the 2,700 AI data centers located in the United States that handle millions of AI queries from users each day will consume six percent of the country’s total electricity by the year 2026, up two full percent from the total percentage consumed in 2022. According to the International Energy Agency, data centers, AI and the cryptocurrency sector sucked up an estimated 460 terawatt hours in 2022, more than a tenth of the overall U.S. electrical consumption that year. With companies like Apple, Google and Microsoft releasing ever more powerful AI tools and looking to build even more data centers to support them, a power problem could be looming.
Gregory Nemet, a professor with the La Follette School of Public Affairs whose research focuses on the state of sustainable energy, says the question may be less about finding future sources of electricity than the future of AI itself.
“We don’t know what’s going to happen with AI,” he says. “There’s big growth now, but there could come a point where it slows down because we’re not really getting any more benefits out of it. And that’s the big open question that no one really knows the answer to.”
Since the 2008 recession, Nemet notes that demand for electricity in the United States has experienced a flat 1% bump annually — that is, until recently. The combination of proliferating AI data centers and the advent of electric vehicles has already supercharged demand and could potentially double it over the next few decades. That said, Nemet remains confident the U.S. can find ways to keep up with the increase.
“One of the benefits of AI is that it’s fairly location independent,” Nemet says. “So, you can build data centers and servers for AI close to wind and solar energy production and then just be building fiber optics instead of expensive transmission cables.”
In some parts of the country, this is already happening. The U.S. Department of Energy’s National Laboratories have already constructed their own AI data center near a nuclear reactor that provides much of its power. Amazon is also doing this. Nemet says companies like Microsoft, Apple and Google — three of the key leaders in AI development — have become big buyers of renewable energy as part of a larger carbon emissions-reduction strategy.
But in other parts of the nation, the situation is bleaker. Public utilities in states like Georgia, Virginia, Washington and Texas are already struggling to meet increased demand for electricity fueled by AI data centers. Some of them are turning to fossil fuels like natural gas to cover the gap, a strategy that demands a large and sustained investment in infrastructure and runs counter to nationwide efforts to reduce carbon emissions.
“Fueling this demand for electricity can be done well, or it can be done badly,” says Nemet. “It can be done with natural gas which is going to create a lot of problems later, or can be done with clean energy, which can avoid those problems.”
Earlier this year, Matt Sinclair, an assistant professor in the Department of Computer Sciences who studies computer architecture, presented at a National Science Foundation sustainable computing workshop, where spirited discussions about the growing potential of AI — and the growing cost to power it — were frequently front and center. Sinclair wasn’t surprised. He’s seen the maximum wattage consumed by graphics processing units (GPUs) manufactured by companies like NVIDIA and AMD increase in recent years, as they try to run massive new AI models.
“Companies are aware of it, and they’re designing more efficient hardware, but they’re also using bigger and bigger models that need more and more power,” he says.
The situation reminds Sinclair of Jevon’s Paradox, the idea that an increase in a resource’s efficiency increases its overall use instead of decreasing it. The paradox was once used to describe coal use in trains in the 1800s, but it seems even more apropos when applied to AI and electricity.
“Making AI more efficient, making it consume less energy, paradoxically increases the overall energy consumption,” he explains. “Because now you can use it in ways you couldn’t before. In turn, this enables more use of AI, increasing overall energy consumption.”
Sinclair believes that finding a balance will require collaboration between groups that have traditionally not worked together — the utilities that construct electric transmission lines, the computer scientists who design the software and hardware that makes AI 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 Sciences Department, Karu Sankaralingam, the Mark D. Hill and David A. Wood Professor of Computer Sciences, is tackling the efficiency problem head-on. Sankaralingam has spent most of his career working on finding ways to make computer chips and processors do more while using less power.
Sankaralingam says that as companies race to get their AI products out the door to cash in on their red-hot popularity, the tools they’re releasing aren’t necessarily the most efficient. As little as 10 years ago, it was possible to wring efficiency gains by improving the functionality of the transistors that conduct electricity through an operating system. Those gains have largely topped out.
Some engineers have opted to create efficiency solutions tailor-made for specific AI programs.
“AI algorithms are changing very, very fast,” Sankaralingam explains. “And it takes a long time to build a chip and sell a chip. By the time you’ve got something out the door, this algorithm has changed, making the chip you just built irrelevant.”
Instead, Sankaralingam and the graduate students in his lab have opted for a different approach he calls efficient generalization, in which a set of power-efficiency principles are applied across different applications rather than targeted at one.
The approach has opened new horizons for Sankaralingam’s research, centered around a central question: How far can the efficiency be pushed without losing its generality? Early simulations in the lab show there are still potential gains. Microsoft, for example, recently improved their chatbot servers to consume 10 times less energy.
“People will, I think, build solutions differently that don’t require as much energy, and I think we are seeing nascent forms of that,” he says. “People are aware that the energy consumption needs to be controlled for this technology to reach its ultimate vision and scope.”