The euphoria around artificial intelligence masks an environmental issue that is starting to make headlines : the explosion in the electricity consumption of data centers, essential for training chatbots. In 2024, while the planet faces unprecedented environmental issues, the proliferation of ever-larger AI models questions their ecological viability.
In this race for gigantism led by large groups in the tech sector; Meta, Google and OpenAI in the lead; a new wave of French and European entrepreneurs and researchers is exploring alternative paths, favoring efficiency over raw power.
The insatiable appetite for large language models like GPT-4 and its 2,000 billion parameters is worrying, and there is reason for it. A simple query on ChatGPT swallows up ten times more energy than a traditional Google search and its power consumption is simply gargantuan.
Wells Fargo’s projections are alarming: by 2030, AI power consumption in the United States could increase 80-fold. This situation is pushing technology giants to extreme solutions, such as Microsoft, which plans to reactivate a nuclear reactor on Three Mile Island to power its data centers.
The problem is that many companies that use AI as an internal solution are not aware of this impasse. “ Many companies perceive LLMs as extremely powerful tools that they must deploy to be in the race for innovation, and most have not yet seen the excess energy consumption that this causes. But, this will probably become a real subject in 2025-2026 » explains Frédéric Brajon, co-founder of the Saegus firm, an entity specializing in the AI transition for professionals.
Slowing down the growth of AI currently seems very complicated as it has become essential; the search for alternatives or the optimization of existing technologies is therefore a path that we must take at all costs.
The emergence of frugal alternatives
Faced with this excess, young companies chart a different path. The start-up Illuin, in collaboration with CentraleSupélec, has developed “ Croissant LLM”, a model 200 times lighter than ChatGPT. In the field of hardware, Ampere Computing, an American company, is rethinking the architecture of processors for AI, achieving three times the energy efficiency of traditional GPUs. Notably those of the giant NVIDIA, which today sits at the top of artificial intelligence.
In France, the PEPR-IA program (Priority Research Program and Equipment for AI), brings together the efforts of Inria, CNRS and CEA to develop more economical solutions, aimed at multiply the energy efficiency of generative AI by 1,000. A project supported by an envelope of 73 million euros from the French State.
This quest for energy sobriety is not just an environmental concern, because European regulations will soon get involved. From 2025, the CSRD (Corporate Sustainability Reporting Directive) will force companies to reveal their carbon footprint, including their use of AI. Its main objective: Encourage the sustainable development of businesses and identify those that are disciplined in this area. The information collected will make it possible to better assess the impact of the company and its activity on the environment. » can we read on the official government website.
This regulatory development, coupled with the foreseeable increase in user costs, could accelerate the adoption of solutions that are more respectful of the planet. However, some experts fear a rebound effect: these more efficient technologies could paradoxically democratize the use of AI, ultimately increasing overall energy consumption. The equation between energy efficiency and performance will certainly take a long time to resolve.
- AI models consume enormous quantities of electricity that are incompatible with contemporary environmental issues.
- Several European start-ups and programs are developing lighter and more economical solutions.
- From 2025, companies will have to reveal their carbon footprint under the yoke of the CSRD directive.
