AI and the environment - Energy Use and Carbon Emissions

Energy Use and Carbon Emissions

AI models require a lot of electricity to “train” and run, especially large models like those used for chatbots and image generators. AI models are housed in massive data centres – essentially huge warehouses filled with computers.  In 2024, data centres in the USA used around 200 terawatt hours of electricity, which is roughly the annual energy usage of the whole of Thailand!  

Google’s data centres emitted 48% more CO2 in 2023 than they did in 2019, mostly as a result of the growth of its AI capacities. Meanwhile, Microsoft’s total carbon footprint grew by over 23% between 2020 and 2024 as a result of the expansion of its AI and cloud services.

According to MIT’s Technology Review, by 2028, energy being used for AI-specific purposes will be between 165 and 326 terawatt hours per year.  That’s enough to power over 28 million households in the USA for a year!  Of course, that energy could come from renewable sources like wind and solar power, or even from nuclear energy.  But if it came from fossil fuel-powered power stations, it would produce emissions of Co2 equivalent to driving over 300 billion miles – or to the sun and back 1,600 times!

A single large AI model can emit over 626,000 pounds (about 284,000 kg) of carbon dioxide just during its training.  That’s the same as 300 round-trip flights between New York and San Francisco. That maybe doesn’t sound too bad, considering the many uses of a large language model (LLM) like ChatGPT.  But the training is only the start.  Each query to ChatGPT emits an estimated 2-3 grams of CO2.  Given that ChatGPT already processes over 1 billion queries every day – a number that’s growing all the time – that’s at least 2,000 tonnes of CO2 per day.  And that’s equivalent to the annual emissions from over 900 average gas-heated houses, every day!

‍