"Du tableau noir au data center : quelle empreinte pour l'IA ?" par Anne-Laure Ligozat

"Du tableau noir au data center : quelle empreinte pour l'IA ?" par Anne-Laure Ligozat

From the blackboard to the data center: what footprint for AI? by Anne-Laure Ligozat

by Anne-Laure Ligozat · View the channel on PSI

Applied Sciences & Engineering AI & Society UBDigital and Information technology: general topicsUBJDigital and information technologies: social and ethical aspects
🎙 Anne-Laure Ligozat 👥 16K 📅 October 21, 2026 ⏱ 30 min 👁 32 📄 literature review 🧭 2026-10-21
Available in: English (current) Français

Keywords

AIenvironmental footprintdata centerslife cycle analysiscarbon emissions

Summary

In this 30-minute seminar, Anne-Laure Ligozat presents the environmental footprint of artificial intelligence, focusing on the life cycle of AI models and infrastructures. She explains the shift from small-scale computations to massive supercomputer training, highlighting the exponential growth in computational requirements. Using life-cycle assessment methodology, she shows that the manufacturing of GPUs and other equipment contributes significantly to environmental impacts, often exceeding use-phase impacts on indicators like human toxicity and mineral resource depletion. She presents case studies, including the BLOOM language model, which emitted about 60 tonnes of CO2 during training, and a detailed LCA of a GPU. The talk emphasizes the importance of multiple environmental indicators, as carbon intensity alone is insufficient: even in low-carbon electricity grids, water consumption and resource extraction remain high. She reviews trends showing that despite efficiency gains, absolute energy consumption and carbon footprints of AI training continue to rise, driven by increased usage and larger models. The presentation includes data from the IEA on data center electricity consumption, projections to 2030, and examples of national regulations (Ireland, Denmark, France). The speaker concludes that AI’s environmental impact is not sustainable on its current trajectory and calls for systematic integration of environmental considerations into AI research and development.

202 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides a substantive overview of the environmental costs of AI, supported by numerous quantitative examples and peer-reviewed studies. The speaker skillfully combines her own research (e.g., BLOOM model analysis, doctoral work by Clément Morin and Martha Lopez) with broader literature to illustrate key points. She argues convincingly that focus on carbon footprint alone is misleading, as other impact categories like resource depletion and water usage are equally critical. The use of life-cycle assessment as a rigorous framework strengthens the argument, though she acknowledges limitations such as incomplete data on GPU manufacturing. The argumentation is well-structured, progressing from methodology to specific examples and finally to systemic trends. However, some projections (e.g., future energy consumption) rely on assumptions that could be challenged, but the overall case for urgent attention is compelling.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is high; the speaker references multiple academic studies, including her own published work, and uses established methodologies like LCA. She clearly differentiates between measured data and estimates, and she explicitly mentions uncertainties in data availability. The title accurately introduces the scope, moving from traditional mathematical contexts to data center impacts. No major discrepancies between title and content emerged. The talk is based on a well-recognized research area, and the speaker’s credentials (professor, co-director of an observatory) lend additional credibility. The description links to the seminar program and membership support, which are relevant but not sources for the content.

247 words

Title / Content Match

The title accurately reflects the content: the talk moves from traditional mathematical concerns to the environmental footprint of AI data centers.

Quality & Reliability

8/10

The talk is grounded in scientific methodology, using life-cycle analysis and citing multiple peer-reviewed studies. The speaker presents her own research alongside external work, but some projections and assumptions are not fully verified.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

Contribution & Novelties

This talk synthesizes the growing body of evidence on the environmental impact of AI, making it accessible to a mathematical audience. It goes beyond simple carbon footprint reporting by highlighting underappreciated impact categories and the nonlinearity of AI’s environmental cost. The speaker’s own research on training emission decomposition and on overlooked phases (e.g., all experiments before final training) provides genuinely new perspectives.

Pour aller plus loin :

  • Life-cycle assessment — The core methodology used to evaluate environmental impacts across product stages.
  • Kaya identity — Decomposition framework used to analyze drivers of CO2 emissions.
  • Green computing — Broader context of efforts to reduce the environmental footprint of computing.

107 words

Radar Profile

The profile shows high scores in information quantity, quality, and reliability, with a slightly lower technical level, indicating that the talk was dense and well-supported but remained accessible to a broad scientific audience.

Reliability 8/10