You might not realise it, but using an AI assistant like ChatGPT requires hundreds, sometimes thousands, of processing cores in a data centre to respond to your query. These computers must be on standby 24/7 to respond to your query instantly. They use vast amounts of electricity and generate heat that must be cooled. All for every bit of text, image, or suggestion that appears on your screen in seconds.
“AI is incredibly powerful, but also a huge energy guzzler,” says Fernando Castor, Associate Professor of Green Software Development. “Some language models use five times more energy than others, even though they produce results of similar quality for some tasks.”
And it’s not just large language models. Image recognition, voice assistants, and automatic recommendations on platforms like YouTube or Spotify also rely on AI algorithms. The more they’re used, the more global energy demand grows.
Saving energy starts with smarter code
Castor conducts research on how software itself can be made more energy-efficient. He calls it green software development: creating code that offers the same functionality, but in a smarter and more efficient way.
“Software may seem intangible, but every line of code tells hardware what to do”, says Castor. “If that code is inefficient, everything underneath consumes more power. Efficient code can make a real difference - for your battery, and for the power grid behind a data centre.”
His team is even exploring whether AI can be trained to write more energy-efficient code. It sounds paradoxical - using energy to save energy - but it works. In tests, some AI models have already generated better, more efficient code than experienced programmers.
Chips that think like a brain
On the hardware side, Amirreza Yousefzadeh, Assistant Professor of Edge AI, is developing neuromorphic chips: processors inspired by the way the human brain works. Like the brain, they function quickly, but use extremely little energy.
“A brain uses about 10 watts of energy to perceive, understand, learn and make complex decisions”, Yousefzadeh explains. “But AI models use thousands of watts to perform similar tasks.”
He’s designing this energy-efficient hardware for use in hearing aids, AR glasses and wearables. AI data centres could also benefit from the same technology in future.
Can you do anything about it?
To some extent, yes. As a user, you can be more conscious of how you use AI. Now that you know that a single prompt may require serious computing power, and thus energy. Generating a single image, for instance, can use enough electricity to charge a laptop dozens of times. That might make you think twice before hitting “generate” again for that extra image or unnecessary query.
However, the real impact lies in how we design these systems. The choices made when writing software, building chips or configuring data infrastructures significantly impact how much energy AI consumes.
The future lies with the designers
AI is part of our everyday lives. But how we build the technology is still up to us. At the University of Twente, researchers combine technical innovation with societal responsibility: how can we create systems that are not only powerful, but also energy-efficient, fair and future-proof?
By considering impact and sustainability from the very start, they are building technology that works for the world of tomorrow. And maybe, one day, you’ll help shape that too!
