[LUM#22] AI's energy bulimia: a blessing in disguise?

The deployment of AI is contributing to the rapid growth of digital's ecological footprint. In the face of criticism, the sector is aiming for rapid gains to achieve carbon neutrality. Promises that need to be proven.

The deployment of AI is increasing the ecological footprint of digital technology. This is hardly surprising, given that AI draws its performance from the processing of ever-growing masses of data, driving up computing requirements. This, in turn, leads to an ever-increasing number of processors and data centers, all of which increase the sector's energy consumption. According to the International Energy Agency (IEA), the electricity required to run AI and cryptocurrencies is set to double between 2022 and 2026.

"The model's training phase is the most power-hungry. The training of Chat-GPT 3 is said to have consumed 1300 megawatt-hours, i.e. 200 times more than the annual consumption of a French household. Knowing that improving an AI model requires repeating this training phase several times... The query, also known as the inference phase, consumes ten times more energy than a standard Google query," explains Edmond Baranes, researcher at the Montpellier Research in Economics Laboratory (MRE).1).

A bidding war to invest in AI

This energy consumption contributes to the sector's growing carbon footprint. Greenhouse gas emissions from the digital sector are set to triple by 2050, according to the trend scenario in a March 2023 joint report by the French Environment and Energy Management Agency (Ademe) and the French Regulatory Authority for Electronic Communications, Posts and Press Distribution (Arcep). And frugality is not on the agenda. " No one envisages a reduction in digital energy consumption," confirms Edmond Baranes. One of the reasons for this is that the business model of the digital giants, the MAMAAs (Microsoft, Amazon, Meta, Apple and Alphabet) is based precisely on data. " There's a bidding war to invest in AI, because mastering it means producing more data, and therefore more value", explains the economist who has been working for years on digital markets, and more specifically on the relationships between telecom operators and digital content providers.

Faced with criticism of its energy-hungry nature, the sector has two responses. Firstly, technological innovations to reduce the power consumption of IT systems. " The arrival of a new generation of processors that consume less power than GPUs (TPUs, NPUs, etc.), as well as improved algorithms - faster and more efficient - will improve the energy performance of AI", stresses Edmond Baranes. However, energy consumption will continue to rise, not least because of the "rebound effect", a well-known mechanism that describes the continuation of the negative effects of an improving technology. The use of renewable energies to power data centers is also widely emphasized by MAMAA. For the time being, the sector's carbon neutrality remains wishful thinking. In 2020, Microsoft pledged to achieve negative emissions by 2030, but by 2024 its emissions had risen by 30% (Le MondeJune 15, 2024).

Finding the right indicators

AI should also contribute indirectly to reducing our consumption. An OECD report from 2022 praises the "double transition" of green and digital technologies, based on the idea that the products and services rendered by AI are sources of efficiency gains and therefore energy savings, by helping to manage energy systems ("smart grids") or by improving the optimization of transport and mobility networks. However, these gains must be set against other negative effects of AI deployment, such as the increase in certain "recreational" uses, largely based on the generation of texts, images and videos, and which consume more and more energy.

"Today, even if discussions are developing, the expertise needed to assess the energy gains linked to AI is not yet well stabilized. One initiative by Ademe and Arcep is precisely to provide an annual survey for a sustainable digital future, in order to track the evolution of digital energy consumption. The challenge is in particular to find the right indicators and make them evolve in line with technological advances", points out Edmond Baranes.

In the meantime, the sector is not subject to any very strict restrictions on its energy consumption. "There are general policy guidelines on the digital responsibility of companies, with incentives that can be taken up in their CSR, but there are no targets for lowering consumption. Fortunately, corporate initiatives are developing around green computing, data center optimization, sustainable product design and recycling," stresses Edmond Baranes. The European Union passed an AI regulation at the end of May 2024, the world's first binding law on artificial intelligence. "The text lists digital services in degrees of risk, but the place given to energy consumption is very limited".

Progressive pricing of digital services according to the volume of data used also seems out of the question. Even though telecom operators have to manage network capacity, and thus keep pace with providers' appetite for data traffic. "The few awareness-raising campaigns to encourage users to stream during off-peak hours are a response to this concern on the part of operators. But today, any hindrance to Internet access is hardly audible, as it would be perceived by many as an infringement of a fundamental freedom", concludes the economist.

UM podcasts are now available on your favorite platforms (Spotify, Deezer, Apple podcasts, Amazon Music...).

  1. MRE (UM)