Tatton Teaser
Posted 21 August 2024
AI Growth = Pending Energy Crisis?
The rapid advancement of artificial intelligence (AI) is fuelling unprecedented investment, but it’s also creating a potential energy crisis. Training and operating AI models requires an enormous amount of computational power, leading to substantial energy consumption.
This problem has compounded as AI models have become more complex. For example, the rise of large language models, behind the likes of ChatGPT, demand even greater amounts of energy. Training a model like GPT-3 is estimated to use ~1,300 megawatt hours (MWh) of electricity – nearly as much consumed by 130 US homes annually. To put this in context, streaming an hour of Netflix consumes about 0.0008 MWh of electricity. In other words, you would be required to watch 1,625,000 hours of Netflix to consume the same amount of energy used to train GPT-3.
Data centres, the backbone of AI operations, are becoming major energy consumers. Their rapid expansion, coupled with the increasing energy demands of AI hardware, is putting immense pressure on the grid. Morgan Stanley and Goldman Sachs estimate that global data centre energy consumption will more than triple this year, from ~15 terawatt hours (TWh) in 2023 to ~50 TWh in 2024. This coincides with the fast growth seen in the global AI market, boosted by Big Tech’s spending spree on AI-related investments – the majority of which was buying land for the construction of even more data centres.
Improving energy efficiency in AI hardware and data centre operations will be crucial. The effort towards this has been evident, with the latest Olympics as a prime example: a data centre in Paris was connected to the energy system heating the Olympic pools. However, critics of such projects, aimed to recycle data centres’ excess heat, have argued that this won’t be enough to curb the long-term energy issues surrounding AI.
Ultimately, a balance must be struck between technological advancement and sustainable energy practices, to ensure that the benefits from AI can be realised without triggering serious environmental consequences.
Thank you Chris Oufi for the analysis.
