How much energy does AI consume?

The development of AI requires an enormous amount of computing power and therefore energy. In addition, more and more products are coming onto the market that are being used by more and more people.

Every image you generate with Midjourney uses almost as much energy as one full charge for your phone. Every video you make with an AI model like Sora consumes even more energy. In addition, big tech also requires big water: water is needed to build all those computers and servers and to keep them cool. ‘GPT-3 needs to “drink” (i.e., consume) a 500ml bottle of water for roughly 10-50 responses, depending on when and where it is deployed,’ researchers at the University of California have written. And GPT-4 uses even more.

What developments have we seen so far?

1%

The International Energy Agency (IEA) predicts that electricity consumption by data centres, AI and cryptocurrencies could double by 2026. Bear in mind, though, that data centres currently account for just 1% of global energy consumption. So even if this number doubles, it’s still fairly under control.

Nuclear fusion

Sam Altman, the CEO of OpenAI, says we need nuclear fusion to keep developing AI without depleting the planet’s resources. Nuclear fusion releases a vast amount of energy. Right now, though, we can’t generate enough energy from nuclear fusion.

More use

Scientists are now mainly concerned about the energy consumption of using AI models, and not so much about training them. An AI model is only trained once, but it then has to answer millions of questions from users, and that takes energy each time.

Positive note

After today, you know in which cases you should use a search engine without generative AI and when you should use ChatGPT. If we all use generative AI conscientiously, we can avoid wasting a lot of energy!

We can pick smart locations to build data centres: in deserts in order to run them on solar energy, at the poles for natural cooling. Or we can link them to heating networks so that the heat they release can be used for heating. Flemish scientists are currently also investigating whether it would be a good idea to put data centres in space. Then they could run on solar energy continuously.

Source: www.vrtnws.be

Legislation can help too. In Europe, data centres must report their energy consumption and emissions.

Engineers are working hard to make more efficient computers and software. This will mean that AI needs less energy to give better answers with lower CO₂ emissions.