The Environmental Cost of Generative AI: A Growing Concern

Generative AI's environmental impact in figures

PARIS (AFP) - The surge in generative artificial intelligence (AI) is being met with growing fears about the technology's ecological footprint, one of the top questions up for discussion at a global summit in Paris on February 10-11. Here are some key figures on the state of play in early 2025: Every request made to OpenAI's chatbot, which is able to generate all kinds of responses to natural-language queries, consumes 2.9 watt-hours of electricity. That is 10 times more than the equivalent figure for a Google search, according to the International Energy Agency (IEA). OpenAI claims that ChatGPT now has 300 million weekly users making a total of one billion requests every single day. Beyond ChatGPT, which fronted generative AI's emergence into public consciousness in 2022, there are thousands of chatbots. One survey by French pollsters Ifop found that 70 per cent of 18- to 24-year-olds in the country said they used generative AI. In America, a Morning Consult poll found that 65 per cent of 13- to 17-year-olds used generative AI, with the number close to half for the general population. Generative AI would not function without data centres hosting vast reserves of information and computing power. In 2023, data centres accounted for almost 1.4 per cent of global electricity consumption, according to a study by consultancy Deloitte. But with massive investments planned into generative AI, the figure is expected to reach three per cent by 2030 - or 1,000 terawatt-hours (TWh). Deloitte said that was comparable with the combined annual consumption of France and Germany. The IEA forecast a more than 75 per cent increase in data centre power consumption by 2026 compared with 2022's levels, to 800 TWh. American consultancy Gartner said the vast power demands meant that up to 40 per cent of data centres built for AI applications could face electricity shortages by 2027. Training one of the large language models (LLMs) that power chatbots generates around 300 tonnes of greenhouse gas carbon dioxide, researchers at the University of Massachusetts Amherst estimated in 2019. That is around the same output as 125 return flights between New York and Beijing. Two years later, Oxford University researchers put the figure at 224 tonnes for a single training session for OpenAI's GPT-3 model. Developers have to train thousands of models to push their technology forward. Despite such estimates, researchers say judging generative AI's overall greenhouse emissions is challenging. Experts and institutions have pointed to a lack of information on how models are produced, as well as an absence of global measurement standards. Beyond energy, generative AI also consumes water, especially for cooling computer hardware. GPT-3 requires around half a litre (one pint) of water to generate between 10 and 50 responses, according to a conservative estimate from researchers at the University of California Riverside and University of Texas at Arlington. Overall, increased AI demand for water is forecast to amount to between 4.2 billion and 6.6 billion cubic metres (155 billion - 233 billion cubic feet). That is four to six times the annual water consumption of Denmark, according to the same 2023 study. Around 2,600 tonnes of electronic waste such as graphics cards, servers and memory chips emerged from generative AI applications in 2023, according to a study from the Nature Computational Science journal. The researchers extrapolated that figure to 2.5 million tonnes by 2030 if current trends continue and nothing is done to limit waste. That would be the equivalent of around 13.3 billion discarded smartphones. And like much computer hardware, AI equipment including chips requires rare metals to manufacture. Mining for such metals, often in Africa, can involve heavily polluting processes.

Generative AI's environmental impact in figures

Each interaction with OpenAI's chatbot uses 2.9 watt-hours of electricity. In 2023, data centres were responsible for nearly 1.4% of global electricity consumption, according to a Deloitte. With substantial investments in generative AI, this share is projected to rise to 3% by 2030, equating to 1,000 terawatt-hours (TWh) of energy.The surge in generative artificial intelligence (AI) is being met with growing fears about the technology's ecological footprint, one of the top questions up for discussion at a global summit in Paris on February 10-11. Here are some key figures on the state of play in early 2025: Ten times Google's power Every request made to OpenAI's chatbot, which is able to generate all kinds of responses to natural-language queries, consumes 2.9 watt-hours of electricity. That is ten times more than the equivalent figure for a Google search, according to the International Energy Agency (IEA). OpenAI claims that ChatGPT now has 300 million weekly users making a total of one billion requests every single day. Beyond ChatGPT, which fronted generative AI's emergence into public consciousness in 2022, there are thousands of chatbots. One survey by French pollsters Ifop found that 70 percent of 18- to 24-year-olds in the country said they used generative AI. In America, a Morning Consult poll found that 65 percent of 13- to 17-year-olds used generative AI, with the number close to half for the general population. Bigger than France and Germany Generative AI would not function without data centres hosting vast reserves of information and computing power. In 2023, data centres accounted for almost 1.4 percent of global electricity consumption, according to a study by consultancy Deloitte. But with massive investments planned into generative AI, the figure is expected to reach three percent by 2030 -- or 1,000 terawatt-hours (TWh). Deloitte said that was comparable with the combined annual consumption of France and Germany. The IEA forecast a more than 75 percent increase in data centre power consumption by 2026 compared with 2022's levels, to 800 TWh. American consultancy Gartner said the vast power demands meant that up to 40 percent of data centres built for AI applications could face electricity shortages by 2027. Hundreds of flights in CO2 Training one of the large language models (LLMs) that power chatbots generates around 300 tonnes of greenhouse gas carbon dioxide, researchers at the University of Massachusetts Amherst estimated in 2019. That is around the same output as 125 return flights between New York and Beijing. Two years later, Oxford University researchers put the figure at 224 tonnes for a single training session for OpenAI's GPT-3 model. Developers have to train thousands of models to push their technology forward. Despite such estimates, researchers say judging generative AI's overall greenhouse emissions is challenging. Experts and institutions have pointed to a lack of information on how models are produced, as well as an absence of global measurement standards. Rivers of water Beyond energy, generative AI also consumes water, especially for cooling computer hardware. GPT-3 requires around half a litre (one pint) of water to generate between 10 and 50 responses, according to a conservative estimate from researchers at the University of California Riverside and University of Texas at Arlington. Overall, increased AI demand for water is forecast to amount to between 4.2 billion and 6.6 billion cubic metres (155 billion - 233 billion cubic feet). That is four to six times the annual water consumption of Denmark, according to the same 2023 study. Heaps of electronic waste Around 2,600 tonnes of electronic waste such as graphics cards, servers and memory chips emerged from generative AI applications in 2023, according to a study from the Nature Computational Science journal. The researchers extrapolated that figure to 2.5 million tonnes by 2030 if current trends continue and nothing is done to limit waste. That would be the equivalent of around 13.3 billion discarded smartphones. And like much computer hardware, AI equipment including chips requires rare metals to manufacture. Mining for such metals, often in Africa, can involve heavily polluting processes.