US tech firms are building vast networks of data centres across the globe and say AI is fuelling the growth, throwing the spotlight on the amount of energy the technology is sucking up and its impact on the environment.
Every time a user punches a request into a chatbot or generative AI tool, the request is fired off to a data centre.
Even before that stage, developing AI programs known as large language models (LLMs) needs a huge amount of computer power.
All the while, the computers are burning through electricity and the servers get hotter, meaning more electricity to cool them.
The International Energy Agency (IEA) said in a report earlier this year that data centres in general used roughly 40 percent of electricity on computing and 40 percent on cooling.
Big tech firms have been rushing to pack all their products with AI ever since OpenAI launched its ChatGPT bot in late 2022.
Plenty of experts are concerned these new products will cause electricity usage to spike.
This is firstly because AI services require more power than their non-AI analogues.
For example, various studies have shown that each request made to ChatGPT uses roughly 10 times the power of a single Google search.
So if Google switches all search queries to AI — about nine billion a year — it could hugely inflate the company’s electricity usage.
And most of these new services and products rely on LLMs.
Programming these algorithms is extremely intensive and usually requires high-powered computer chips.
They in turn require more cooling, which uses more electricity.
Before the era of AI, estimates generally suggested data centres accounted for around one percent of global electricity demand.
The IEA report said data centres, cryptocurrencies and AI combined used 460 TWh of electricity worldwide in 2022, almost two percent of total global electricity demand.
The IEA estimated that the figure could double by 2026 — the equivalent of Japan’s usage figures.
Alex De Vries, a researcher who runs the Digiconomist website, modelled the electricity used by AI alone by focusing on sales projections from the US firm NVIDIA, which has cornered the market in AI-specialised servers.
He concluded in a paper late last year that that if NVIDIA’s projected sales for 2023 were correct and all those servers ran at full power, they alone could be responsible for between 85.4–134.0 TWh of annual electricity consumption — an amount similar to Argentina or Sweden.
“The numbers I put in that article were already conservative to begin with because I couldn’t include things like cooling requirements,” he told AFP.
