Stay up up to now with free updates
Simply log in Artificial intelligence myFT Digest – delivered straight to your inbox.
Informed readers already know this many Energy might be needed to fuel hoped-for advances in machine learning, large language models, and various other things that tech salespeople classify as “artificial intelligence.”
Barclays has one nice summary out today on the energy requirements of AI:
Most of the greater than 11,000 registered data centers worldwide will not be yet involved in AI-related activities. According to an IEA2 report from mid-2024, their combined electricity consumption (excluding cryptocurrencies) is around 1.0 to 1.5% of the world's total.
However, data center energy needs could change dramatically in the approaching years because of the spread of AI.
According to a June 2024 evaluation from Barclays Research – which uses a bottom-up approach based on utilities' forward-looking supply contracts – annual data center power demand within the US could increase by 14-21% every year through 2030. That would mean US data center demand would roughly triple by 2030, from 150-175 terawatt hours (TWh) per 12 months 2023 to as much as 150 terawatt hours (TWh). 560 TWh – akin to 13% of current US electricity demand.
It's value noting that the info center boom got here at a really opportune time for the business real estate market, but that could be one other story. From the bank:
headlines Virginia have already shown how “bottlenecks“can arise in energy supply, particularly as data center construction tends to be concentrated in certain critical regions.
And the one energy source that actually meets demand? It seems that this is principally nuclear energy:
If the web must be flooded with AI junk, we would as well get clean energy from it.
