Video of the

Typical Generator Model for all sorts

of Projects

AI guzzles energy. Can scientists do something about it?

While the use of artificial intelligence is increasing rapidly, the energy consumption of these programs remains sky-high for the time being. Two scientists explain how it can be more economical.

Trouw Jose Stoop May 3, 2025, 10:00

When a person tries to remember something, the brain stores bits of information. Our head does this a lot more efficiently than how artificial intelligence works. That energy-guzzling monster needs enormous amounts of information to function. For example, new AI models are trained with large amounts of data, a process that can take months. In addition, the powerful computers and data centers that keep AI running ( Alfred: Why? ) consume a lot of power, especially for cooling. And then there is the actual use of the model; ( so putting the AI ​​on standby already costs an immense amount of energy? ) something that consumes relatively more and more energy now that all the large models are trained on all the information that humanity has ever produced, while the answers they generate are becoming increasingly extensive.

Smarter software saves energy

One way to make AI less energy-intensive is to make it smarter about information, inspired by the human brain. Roy Schwartz, a researcher at the Hebrew University of Jerusalem, is looking at how AI models can become more efficient. For example, his team is investigating whether (1) an AI model does not always have to process a complete text, but sometimes only needs a summary. “Just like humans, who do not remember everything literally – that would cost too much energy,” he explains. Another method revolves around (2) ‘embedding’, focusing on the fact that an AI model converts text into numerical values ​​(vectors) in order to generate an answer. Common words have their own vector, but long or less frequently used words consist of multiple vectors, and it takes extra computing power to translate them back from vectors to text. Schwartz’s team is looking for ways to add these words directly to the model’s dictionary, making AI faster and more energy-efficient. (3) What is already widely used is a large AI model to train a smaller, more efficient model. “Think of a pasta recipe, for example. The small model will initially give the best-known recipe, but then check with the large model to see if there are any more options. The Chinese DeepSeek makes extensive use of this.” (4) ‘Quantization’ is also gaining popularity: an AI model is optimized by calculating with simpler numbers, which saves computing power and energy. According to Schwartz, the last two ( 3 + 4 ) methods are particularly interesting because they make AI more efficient without compromising the accuracy of the model too much.

Energy-efficient hardware and data centers

Hardware also plays a major role in making AI more sustainable. This is a research area that Rajendra Bishnoi of TU Delft focuses on. “Most of the energy in AI systems is not consumed by calculations, but by moving data between storage and processing units,” he explains. (This is important information because generator AI has to retrieve and write a lot of data. If you can retrieve all that data at once, then immediately calculate it, and then write it away, that scales time and energy) “That is why optimizing data storage is essential to create an energy-efficient system.” His research focuses on new computational methods that allow data processing and storage to be done directly in the memory units (that's what I would do...process everything directly in the internal memory), eliminating the need for data transfer – this efficient approach is inspired by the workings of the human brain.

And finally, there is also a battle to be made in making the data centers that keep AI running more sustainable (I probably won't have any problems with that because Generator AI is based on DNA), something that big tech companies are increasingly focusing on. They would benefit from becoming more energy efficient anyway with their sky-high energy bills. But although they are increasingly getting their energy from sustainable sources, their energy demand is simply too high for the supply. It is therefore not surprising that Google and Amazon, among others, have plans to build their own small nuclear power plants (I am in favor of much smaller nuclear power plants there). Big tech is not very transparent about energy consumption (it should be because the climate suffers from it and then we do), which makes it difficult to make a statement about the total amount of emissions. Schwartz: “I’m in favor of more government regulation like a global carbon tax—not just on AI, but also on air travel and industry and so on. The concept of ‘AI’ is quite flexible, and it’s also unclear how much energy is used for what. That makes reporting energy consumption easy to manipulate. We could enforce that with legislation.”

What We Do

About Us

Paragraphs are the main building blocks of web pages. To change what this one says, just double-click here or hit Edit text. You can change the style here, too. Of course, dragging the text will move the paragraph on the page.

Paragraphs are the main building blocks of web pages. To change what this one says, just double-click here or hit Edit text. You can change the style here, too. Of course, dragging the text will move the paragraph on the page. One paragraph may be enough to introduce your project. If it gets too long, break up the text into several paragraphs. Long text is perfect for explaining what's unique about your website. For example, you can tell visitors how it all got started or what makes you special today. Plain text is often the simplest way to engage your audience. Imagine your reader as a person in front of you, and tell them something to peak their interest. There's nothing more powerful than a great story!

Become A Volunteer

Paragraphs are the main building blocks of web pages. To change what this one says, just double-click here or hit Edit text. You can change the style here, too. Of course, dragging the text will move the paragraph on the page.

Paragraphs are the main building blocks of web pages. To change what this one says, just double-click here or hit Edit text. You can change the style here, too. Of course, dragging the text will move the paragraph on the page. One paragraph may be enough to introduce your project. If it gets too long, break up the text into several paragraphs. Long text is perfect for explaining what's unique about your website. For example, you can tell visitors how it all got started or what makes you special today. Plain text is often the simplest way to engage your audience. Imagine your reader as a person in front of you, and tell them something to peak their interest. There's nothing more powerful than a great story!

Thanks To Our Sponsors

Want to get involved? Contact us today.