The more efficient AI models become, the more their environmental impact is likely to rise. This paradox lies at the heart of a debate the industry often prefers to sidestep — and that research is only beginning to confront. Can we reduce AI’s carbon footprint without slowing its growth?
A carbon footprint that grows heavier as models get bigger
The meteoric rise of artificial intelligence comes with a mounting environmental cost. Training a large language model harnesses thousands of processors for several weeks. Training GPT-3 generated 552 tonnes of CO₂, the equivalent of more than 300 round-trip flights between Paris and New York.[1] Models have kept growing ever since: the size of large language models increased ten-thousandfold between 2018 and 2023, with each generation drawing on considerably greater computing resources.[2] As for GPT-5, released in 2025, no official figures have been disclosed. OpenAI is among the ten major players that have released no key environmental metrics, according to the Foundation Model Transparency Index 2025 published by researchers from Stanford, Berkeley, Princeton and MIT.[3] This lack of transparency amounts to a governance shortfall that regulators, particularly in Europe, are beginning to take into account.
A proliferation of models that shows no sign of slowing
At the same time, models are proliferating at an equally sustained pace. According to Stanford University’s AI Index Report 2024, there are ninety-six of them (59 in the United States, 35 in China and 2 in Europe).[4] The adoption of consumer tools is just as striking: ChatGPT reached 400 million weekly active users in February 2025 and one billion monthly active users in June 2026.
Data centres whose consumption could double by 2030
Globally, data centres accounted for 1.5% of the world’s electricity consumption in 2024, or 415 TWh, according to the International Energy Agency.[5] That consumption is expected to more than double by 2030, reaching 945 TWh, driven by the widespread use of AI. In the United States, half of the growth in electricity demand between now and 2030 is tied to data-centre needs.[6]
Unprecedented infrastructure investment that shows what is at stake
The infrastructure investments now under way show what is at stake. The Stargate project, launched in January 2025 by OpenAI, Oracle and SoftBank with a target of 500 billion dollars and 10 gigawatts of capacity — the energy equivalent of the annual consumption of a country like Belgium — illustrates the scale of the commitments involved.[7]
System architecture and the invisible footprint of rare metals
Real efficiency gains have nonetheless been achieved. Industry players are optimising their models and investing in specialised hardware architectures. In its 2024 sustainability report, NVIDIA states that the shift to GPU accelerated computing makes it possible to handle the same workloads with 10 to 20 times fewer servers than a conventional CPU infrastructure, cutting overall energy consumption accordingly.[8] In August 2025, Google published Gemini’s environmental data on a per-query basis — 0.24 Wh and 0.03 g of CO₂ equivalent for a median text query[9] — the first transparency exercise of its kind among the major players. To date it remains without equivalent in the sector.
These gains run up against a structural economic dynamic, however: the rebound effect. As Lou Welgryn and Théo Alves Da Costa point out in their analysis published on Bon Pote, the case of DeepSeek illustrates this strikingly. Launched in January 2025, this Chinese open-source model was presented as far less resource-hungry than its competitors. Early independent tests, however, showed that the real efficiency gains were more limited than announced, and that the model’s mass adoption generates a rebound effect that cancels out part of the expected benefits.[10] More fundamentally, the growth in usage follows the growth in the economic value generated: McKinsey & Company estimates that generative AI could represent between 2,600 and 4,400 billion dollars in annual economic value.[11] A potential of this magnitude implies a multiplication of deployments, of queries and, mechanically, of energy consumption.
The question of system architecture then becomes decisive. Do all applications justify the use of massive models? Work compiled by Stanford University shows that compact, specialised models can compete, on certain tasks, with far heavier generalist architectures, for a fraction of the resources deployed.[12] One point of caution, however: the cumulative emissions of inference can exceed those of training once models are queried at scale.[13] The issue, then, is less the choice of model than the relevance of each use. Yet these figures reflect only part of the real footprint. As researcher Kate Crawford shows in her Atlas of AI (Yale University Press, 2021), AI rests on a triple extraction — minerals, data, human labour — whose environmental cost far exceeds the electricity consumption of data centres alone. Manufacturing the chips needed to train models draws on rare metals whose extraction generates considerable local impacts, often invisible in tech companies’ carbon accounts.
French public research is tackling the issue head-on
Faced with this reality, French public research has organised itself. The PEPR IA programme, co-led by INRIA, the CEA and the CNRS as part of France 2030, is devoting 73 million euros over five years to developing frugal AI (https://www.pepr-ia.fr/projet/sharp/) — defined as the ability to carry out complex tasks while drawing on as few resources as possible, without sacrificing performance. This challenge is all the sharper because, for a task as simple as image recognition, a machine consumes four to six orders of magnitude more energy than a human brain.[14] Efficiency is a scientific challenge of the first order.
A carbon equation that technological performance alone will not solve
So, can AI’s carbon impact be reduced? Yes, partially. The efficiency gains are real, and the room for technical progress is considerable. But without keeping the growth in usage in check, these gains will remain structurally insufficient. AI’s carbon equation will not be solved by technological performance alone: it demands explicit strategic choices, from companies and states alike, about what we really want to optimise.
—
To go further: the report by Lou Welgryn and Théo Alves Da Costa published on BonPote, « Intelligence artificielle : le vrai coût environnemental de la course à l’IA » (September 2025, updated April 2026), is one of the most comprehensive syntheses available in French on this subject. It covers, among other things, water consumption, mineral resources, the real emissions of the digital giants and local conflicts over use. https://bonpote.com/intelligence-artificielle-le-vrai-cout-environnemental-de-la-course-a-lia/
[1] « Carbon Emissions and Large Neural Network Training », Google / UC Berkeley / Google Brain, April 2021. https://arxiv.org/abs/2104.10350
[2] Lou Welgryn and Théo Alves Da Costa, « Intelligence artificielle : le vrai coût environnemental de la course à l’IA », BonPote, September 2025 (updated April 2026). Primary source cited: Stanford University, AI Index Report 2025. https://bonpote.com/intelligence-artificielle-le-vrai-cout-environnemental-de-la-course-a-lia/
[3] Foundation Model Transparency Index 2025, December 2025. https://crfm.stanford.edu/fmti/
[4] Stanford University Human-Centered Artificial Intelligence, AI Index Report 2024, April 2024. https://aiindex.stanford.edu/report/
[5] International Energy Agency (IEA), « Energy and AI », special report, April 2025. https://www.iea.org/reports/energy-and-ai
[6] IEA, ibid. Base-case scenario: global data-centre consumption would reach 945 TWh in 2030, compared with 415 TWh in 2024.
[7] OpenAI, official press release « Announcing The Stargate Project », 21 January 2025; https://openai.com/index/announcing-the-stargate-project/
[8] NVIDIA, Sustainability Report Fiscal Year 2024, https://images.nvidia.com/aem-dam/Solutions/documents/FY2024-NVIDIA-Corporate-Sustainability-Report.pdf
[9] Google Cloud Blog, « How much energy does Google’s AI use? We did the math », Amin Vahdat and Jeff Dean, 21 August 2025. https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference
[10] Lou Welgryn and Théo Alves Da Costa, « Intelligence artificielle : le vrai coût environnemental de la course à l’IA », BonPote, September 2025 (updated April 2026). https://bonpote.com/intelligence-artificielle-le-vrai-cout-environnemental-de-la-course-a-lia/
[11] McKinsey & Company, « The Economic Potential of Generative AI », June 2023. https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/the-economic-potential-of-generative-ai-the-next-productivity-frontier
[12] Stanford University, AI Index Report 2024, op. cit.
[13] Stanford University, AI Index Report 2024, op. cit.
[14] Benjamin Guedj, Inria researcher, « Machine learning : comment allier performance d’apprentissage et sobriété numérique ? », Inria, April 2022. https://www.inria.fr/fr/machine-learning-ia-numerique-frugal.
