Depending on who you ask, AI will either break the electric grid or usher in a new era of cheap, clean energy.
But data scientist Jeff Willert says there is a big missing middle.
Support our work by subscribing to our free weekly newsletter or Substack.
“They’re arguing about two ends of the distribution, but there’s a lot of space in the middle where a lot of good work is being done,” he said on the Energy Changemakers podcast.
Willert, who leads the data science and AI team for Schneider Electric’s SE Advisory Services, centers his argument on a term emerging in energy circles: Frugal AI.
First question: Do I even need AI for this?
The term might conjure Ebenezer Scrooge hoarding compute capacity. But it’s not about miserliness; it’s practical.
And it shouldn’t be mistaken for energy efficiency. Efficiency means doing the same work using less energy. Frugal AI begins even earlier.
It starts with: “Should we be doing this with a model at all?” Willert says.
Or another way to put it: Frugal AI means using the right size tool for the job — the right level of compute to solve the problem.
For example, companies don’t necessarily need massive compute to catch common mistakes on a utility bill. Sometimes a smaller AI model does the trick, or traditional machine learning, or even lookup tables or decision trees.
“We are looking at the individual problem and saying, what’s the simplest solution we can come up with to solve this problem to a high degree of accuracy?” he says.
He isn’t saying don’t use the newest, largest and most energy-hungry models. He’s saying use them only if you need them. His company, Schneider, uses powerful frontier AI models when a problem warrants that level of compute, he says, but other times relies on smaller models or simpler mechanisms — whatever delivers the required accuracy with the least complexity.
Should we feel guilty?
So how does that translate to the average person prompting AI to find cheap airline tickets or the best chicken parmesan recipe? Are we being extravagant?
It all adds up, he says. But that average person going on a Chatbot diet is probably the smallest lever. “A prompt is a tiny thing.”
AI use is counted in tokens, units it converts into numbers. As a rule of thumb, one token equals roughly 100 words, according to OpenAI. Willert estimates that a question to a common chatbot, like Gemini or Claude, runs tens or hundreds of tokens. By contrast, more powerful agentic AI, which can independently carry out tasks, uses far more tokens. An individual agent request can consume about 15 times as many tokens as a typical human chat.
“While it’s important that we are cautious about how we use it, it’s just a drop in a very large ocean,” he said.
“I’m actually happy when I get to solve a problem without AI,” he adds.
The instinct goes back to his doctoral work on nuclear reactor modeling.
“The whole idea back then, 15 years ago, was that high-fidelity simulation is enormously expensive. And so we were working to pair it with cheaper models that do most of the work, and only paying for the expensive part where you actually need it,” he says. “And that is very much parallel to what I’m doing today. It’s the frugal AI argument. What is the cheapest computation that still gets me the right answer?
The missing middle
The term frugal AI isn’t Willert’s. In fact, France published a frugal AI framework two years ago to measure and reduce AI’s environmental impact.
But it’s a term for the times, with data centers expected to drive development of more power plants — particularly natural gas-fired plants. What does that mean for emissions and environmental impact?
Data center critics are right to worry that they lack information about the impact, Willert says. Most chatbot queries last year went to models whose environmental footprint hasn’t been disclosed, which creates an environment “ripe for greenwashing.”
That’s where he sees a more frugal approach coming into play via use of AI tools designed for a specific decision or workflow. Rather than relying on powerful, energy-hungry AI models for every job, companies can break work into smaller tasks and use the simplest tool capable of handling each one accurately.
We’re measuring AI wrong
Another problem, he says, is the way industry measures AI use. It focuses on how much computing power AI consumes. As he sees it, a better measure is what we get in return for the electricity AI uses.
Consider an international manufacturer trying to cut greenhouse gases. An AI system might identify ways to cut energy to avoid thousands of tons of carbon. Of course, using AI to do this also creates carbon emissions, but it might only amount to grams or kilograms.
That doesn’t mean AI’s electricity consumption should be ignored. “AI can have massive benefit, but we can’t ignore that we are generating CO2 while we’re doing it,” he says.
But will it play at town hall?
Willert isn’t confident in the forecasts driving today’s build-out of power plants, transmission, and data centers. Models are getting more efficient. But then there is the Jevons paradox, or what efficiency circles call the rebound effect. The cheaper something gets, the more people use it. That, he says, is exactly why frugality matters.
It wouldn’t be surprising to see “frugal AI” become a marketing term, a proof point for companies showing they are performing with responsibility.
But to ensure its veracity, AI providers need to be as transparent as possible, which means publishing emissions, the problem being solved, the model used, where it’s hosted, and how it was validated. For his own shop, Schneider Electric, he foresees a public dashboard comparing the energy Schneider consumes with the emissions that the AI insights help clients avoid.
One caveat: Willert doubts the frugality argument will necessarily win over communities fighting data centers. Why? Communities tend not to care about things like how much intelligence a model delivers per watt. Instead, it’s electricity rates, water, noise, traffic and land use issues keeping them up at night. Frugal AI may curb unnecessary demand, but it doesn’t answer those questions.
Listen to the full podcast: Can Frugal AI Ease Grid Strain from Data Centers
Read Willert’s white paper: The Missing Middle in the AI Energy Debate
