// NATURE NEWS — SPAZIO & SCIENZA
AI’s energy paradox: will the AI boom strain energy systems or be the source of solutions?
Melissa Hobson is a freelance marine-science journalist based in Hastings, UK.
Search author on:
PubMed
Google Scholar
Ginelle Greene-Dewasmes is working to make AI energy use more sustainable.
Artificial intelligence is exploding in popularity, and many people have concerns about the sustainability of its energy consumption. AI models need vast amounts of electricity: according to the International Energy Agency, a typical AI data centre can consume as much energy as 100,000 households each year. Energy use by US data-centre servers more than tripled between 2014 and 2023 — reaching 4.4% of the nation’s total electricity consumption. This figure is expected to at least double by 20281.
However, AI might also hold the key to improving energy efficiency: offsetting some of the demand it creates. An article published by the World Economic Forum suggests that the use of AI tools across industries — in automating tasks or monitoring real-time energy use, for example — can reduce energy consumption by up to 60%. Ginelle Greene-Dewasmes, a sustainable-development specialist and an author of the article, has been working to lower AI’s energy demands while finding ways to use the tools to reduce wider emissions.
Originally from Trinidad, Greene-Dewasmes became interested in solving energy-security issues for island states with few resources. In 2021, she completed a PhD on the green-energy transition across small island developing states at the University of York, UK, before joining the World Economic Forum in 2022. Now based in Geneva, Switzerland, Greene-Dewasmes has been the Forum’s AI and Energy Lead since 2024. Her work involves bringing together AI and energy leaders from around the world to identify shared challenges, align priorities and develop practical actions that can improve AI’s energy impact.
As well as looking at improving AI’s energy efficiency, Greene-Dewasmes and partners are exploring how models can find innovations that improve existing energy infrastructure, such as that in buildings that are already in use. Nature spoke to her about navigating the increasing energy needs of the technology and whether its benefits could outweigh this demand.
AI needs cloud networks, chips, graphics-processing units and cooling systems, all of which consume large amounts of energy. Developing and training models requires high levels of processing power to crunch the data, but AI use is much more energy intensive overall because of the scale. ChatGPT, the chatbot from technology firm OpenAI in San Francisco, California, has more than 800 million users each month and receives more than 2.5 billion queries daily. As AI becomes mainstream, we need to build more energy infrastructure to keep up with the surge in demand.
But, when it comes to scaling up AI, I think the conversation is too skewed towards the energy needed to power centralized infrastructure, when there are lots of other things that will contribute to AI’s energy requirements. Large data centres will exist in only the few countries that have the money, space and resources to host them. For most of the world, AI infrastructure will be disaggregated and decentralized. For example, if one million people are using their phones more intensively because of AI, that affects the energy infrastructure.
Although more complex LLM queries are currently directed to remote data centres, some smaller tasks can be completed directly on new devices that have AI chips. As this becomes more common, what will that mean for infrastructure planning for energy? It’s much harder to measure this than the outputs from one data centre.