AI may soon account for half of data center power use if trends persist

May 24, 2025 by Bob Yirka, Phys.org

Collected at: https://techxplore.com/news/2025-05-ai-account-center-power-trends.html

Alex de Vries-Gao, a PhD candidate at VU Amsterdam Institute for Environmental Studies, has published an opinion piece about the results of a simple study he conducted involving the possible amount of electricity used by AI companies to generate answers to user queries. In his paper published in the journal Joule, he describes how he calculated past and current global electricity usage by AI data centers and how he made estimates regarding the future.

Recently, the International Energy Agency reported that data centers were responsible for up to 1.5% of global energy use in 2024—a number that is rising rapidly. Data centers are used for more things than crunching AI queries, as de Vries-Gao notes. They are also used to process and store cloud data, notably as part of bitcoin mining.

Over the past few years, AI makers have acknowledged that running LLMs such as ChatGPT takes a lot of computing power. So much so, that some of them have begun to generate their own electricity to ensure their needs are met. Over the past year, as de Vries-Gao notes, AI makers have become less forthcoming with details regarding energy use. Because of that, he set about making some estimates of his own.

He started by looking at chips manufactured by the Taiwan Semiconductor Manufacturing Company, a company that makes most of the chips for companies like Nvidia. He then used estimates by noted analysts, earnings reports and details regarding the devices bought, sold and used to build AI data centers. He next looked at publicly available electricity consumption reports for the hardware used to run AI applications, as well as their utilization rates.

De Vries-Gao then used all the data he had amassed to make rough estimates for electricity usage by different AI providers and then added them all together, arriving at an estimate of 82 terawatt-hours of electricity consumed for all of this year, based on current demand—roughly equivalent, he notes, to all the power used by a country like Switzerland.

He then did the same arithmetic with the assumption that the demand for AI would double over the course of the rest of this year. If things turn out that way, AI applications could consume approximately half of all the power used by data centers around the world.

De Vries-Gao notes that there is more at stake with AI data center power use than the increase in demand, which could lead to increases in power prices. There is also the environmental impact. If most AI providers use electricity from the grid to power their data centers, there could be a huge increase in the release of greenhouse gases because so much electricity is still generated by burning coal, leading to more global warming.

More information: Alex de Vries-Gao, Artificial intelligence: Supply chain constraints and energy implications, Joule (2025). DOI: 10.1016/j.joule.2025.101961

Journal information: Joule