8 July 2025
Nieuws
0 reacties
5x viewed
A more sustainable application of generative AI? If you ask Gen AI applications like ChatGPT and Claude about that, they quickly come up with well-reasoned recommendations. Meanwhile, they themselves still consume an alarming amount of energy and water. So that has to change. And it can be done differently. Capgemini therefore published a report earlier this year with practical tips.
"As soon as you ask a question, powerful mathematical models are put into action."
Training GPT-4 was estimated to require as much electricity as the annual consumption of 5,000 American households. And that's not all: now that that model is operational, every demand from a user leads to additional energy consumption. Moreover, to prevent overheating, servers in data centers must be properly cooled. So that costs a lot of water as well.
Inevitable consequence?
American technology companies in recent years have focused primarily on driving the performance of their generative AI models. They invested heavily in advanced chips and building new data centers. And the sharply rising energy and water consumption they caused by doing so? They saw that as an inevitable consequence of that new development.
Growing ecological footprint
Users of Gen AI applications also seem to have little regard for the sustainable aspects. For example, the Capgemini report Developing Sustainable Gen AI found that while many companies' use of Gen AI applications is already leading to a larger carbon footprint, they are adjusting their sustainability goals rather than limiting the use of that technology.
A totally new approach
"Meanwhile, all indications are that the use of generative AI is only going to rise sharply," emphasizes Joost Carpaij, Gen AI expert at Capgemini. "In doing so, you see that those AI models are getting more and more sophisticated and therefore are going to provide more comprehensive answers. But the longer an answer, the more energy it takes. Unless you come up with a totally new approach that makes it possible to arrive at more efficient AI models. And early this year we saw a development in that area that you can safely call a breakthrough."