There’s a simple way we could drastically cut AI energy use

If we choose AI models to solve problems more wisely, we could potentially save 31.9 terawatt-hours of energy this year alone, which is equivalent to the power of five nuclear reactors.

Thiago da Silva Barros from the University of the Côte d'Azur in France and his colleagues studied 14 different tasks for which people use generative artificial intelligence tools, from text generation to speech recognition and image classification.

They then examined public leaderboards, including those hosted by machine learning center Hugging Face, for the performance of various models. The energy efficiency of the models during inference—when the AI ​​model produces an answer—was measured using a tool called CarbonTracker, and the overall energy consumption of that model was calculated by tracking user uploads.

“Based on the size of the model, we have estimated the energy consumption and from this we can try to make our estimates,” says da Silva Barros.

The researchers found that across all 14 tasks, moving from the most efficient to the most energy efficient models for each task reduced energy consumption by 65.8 percent, while the result was only 3.9 percent less useful—a trade-off they believe may be acceptable to the public.

Since some people already use the most energy-efficient models, if people in the real world switched from high-performing models to the most energy-efficient models, they could achieve an overall reduction in energy consumption of 27.8%. “We were surprised at how much we could save,” says a team member Frederic Girouard at the French National Center for Scientific Research.

However, this will require changes from both users and artificial intelligence companies, says da Silva Barros. “We have to think in the direction of using smaller models, even if we lose some performance,” he says. “And for companies, when they develop models, it's important to share some information about the model that allows users to understand and evaluate whether the model is very energy-intensive or not.”

Some artificial intelligence companies reduce the power consumption of their products through a process called model distillation, where large models are used to train smaller models. This is already having a significant impact, says Chris Preist at the University of Bristol in the UK. For example, Google recently stated 33 times higher energy efficiency in Gemini over the past year.

However, allowing users to choose the most efficient models “is unlikely to limit the increase in power consumption in data centers, as the authors suggest, at least in the current AI bubble.” says Preist. “Reducing the energy cost of a hint will simply allow us to serve more customers faster using more complex reasoning options,” he says.

“Using smaller models can definitely result in lower energy consumption in the short term, but there are many other factors that need to be taken into account when making any meaningful predictions for the future,” says Sasha Luccioni in “The Hugging Face”. She warns that “rebound effects, such as increased consumption, as well as the broader impact on society and the economy need to be taken into account.”

Luccioni notes that any research in this area is based on external assessments and analysis due to a lack of transparency on the part of individual companies. “To do this kind of more complex analysis, we need more transparency from artificial intelligence companies, data center operators and even governments,” she says. “This will allow researchers and policymakers to make informed predictions and decisions.”