Impacts Écologiques de l'IA

I : Introduction aux termes techniques

I.1 : Définitions

I.2 : Au CEMS

I.3 : Schéma récapitulatif

II : Zoom sur les chatbots basés sur LLM

II.1 : Termes techniques

II.2 : Cycle de vie

III : Impacts écologiques de chaque phase

III.1 : Collecte des données

III.2 : Bandwidth Infrastructure Energy

III.3 : Pré-entraînement

III.4 : Entraînement

III.5 : Inférence

III.6 : Tableau récapitulatif

IV : Impacts écologiques au global

IV.1 : Énergie

IV.2 : Gaz à effet de serre

IV.3 : Eau

IV.4 : Minerais

IV.5 : Biodiversité et sols

IV.6 : Rapport de l'ADEME

V : Conclusion

V.1 : Sujets laissés de côté

V.2 : Pour aller plus loin...

V.3 : Solutions

V.4 : Discussion

Impacts Écologiques de l'IA

Baptiste ChanusSéminaire au vert du CEMS30 juin 2026

I : Introduction aux termes techniques


I.1 : Définitions



I.2 : Au CEMS

L'usage omniprésent : les chatbots basés sur des LLM.

Generative AI accounted for over 20% of the global AI market in 2026, and is projected to reach 40% by 2030. 1

Par exemple :

Chatbot : Interface conversationnelle permettant à un utilisateur d'interagir en langage naturel avec un système. Peut être basé sur des règles (chatbot classique) ou sur un LLM (chatbot génératif).


I.3 : Schéma récapitulatif

NLP Deep Learning LLM Chatbots IA ML


II : Zoom sur les chatbots basés sur LLM


II.1 : Termes techniques

Modèles entraînés spécifiquement à 'réfléchir' avant de répondre. Génèrent des 'chaînes de pensée' internes longues avant la réponse finale. Consommation par requête très supérieure aux modèles standards 2 à 10x plus de tokens (~ mots)

Plusieurs LLM collaborent, s'échangent des messages, utilisent des outils. AutoGPT, LangGraph, CrewAI. La multiplication des appels API démultiplie la consommation énergétique. Et augmente le nombre de tokens

Agents capables d'agir sur l'environnement : navigation web, écriture de code, gestion de fichiers. Boucles de planification-action-observation. Chaque étape donnent des appels LLM supplémentaires.


II.2 : Cycle de vie

Collecte Pré-entraînement Fine-Tuning Renforcement Inférence Récupération de données sur internet (wikipédia, reddit, arXiv, libgen, github) Première détermination des poids du modèles : Le modèle apprend à prédire le token suivant Spécialisation (avoir des conversations, programmer...) Sur des exemples annotés par des humains Des humains évaluent la qualité des réponses du modèle, si elles sont 'éthiques' etc Réponses aux requêtes des utilisateurs


III : Impacts écologiques de chaque phase


III.1 : Collecte des données

Scraping : récupération du contenu de pages internet.


Infrastructure :

  1. Crawlers (serveurs qui font tourner les logiciels de scraping)
  2. Réseau (transfert de données entre les crawlers et les hébergeurs)
  3. Hébergeurs (Gère les requêtes des bots)
  4. Raffinage des données (nettoyage et parsing du contenu)
  5. Stockage (archivage des données d'entraînement) 2

Résumé :

Scaling to AI training datasets:

GPT-4 reportedly trained on ~13 trillion tokens (estimated 5-10 billion web pages scraped).

5 billion pages × 119 tons / billion = 595 metric tons CO₂ from scraping alone.

Doesn't include training compute (which dwarfs scraping—GPT-3 training estimated 552 tons CO₂, GPT-4 likely 10-50× higher).

Scraping represents 2-5% of total AI model carbon footprint.

Small percentage, but absolute tonnage is significant (equivalent to hundreds of households).

Training is one-time (retrain every 12-18 months).

Scraping is continuous (Perplexity, real-time answer engines scrape constantly for fresh data).

Annual scraping emissions for answer engine:

If engine scrapes 100M pages/day for current information:

365 days × 100M pages/day = 36.5 billion pages/year

36.5 × 119 tons / billion = 4,343 tons CO₂/year from scraping alone.

Exceeds one-time training cost if model lifespanyears.

3


Crawlers :

Power Usage Effectiveness (PUE): Ratio de l'énergie utilisée par les data centers sur l'énergie utilisée uniquement par le matériel informatique

1.6 de PUE en moyenne dans l'industrie (pour 1 kWh dépensé pour les serveurs, 0.6 kWh sont dépensés pour le refroidissement et autre), 1.1-1.2 pour les plus optimisée (Google/Microsoft)

Donc pour les crawlers :

Émission de carbonne (sur le réseau US):

GPT-4 s'est entraîné sur 13 billions (en français) de tokens (soit 5-10 milliards de pages web scrapées).

5-10 milliards de page × 46 320 kg CO₂ / milliard de page = 231.6-463.2 metric tons CO₂ due au scraping.


Réseau:

Serveur hébergeur → Router du fournisseur d'accès → Réseau internet → Datacenter de l'entreprise IA


Charge de l'hébergeur:

Carbon emissions (U.S. grid average 0.386 kg CO₂/kWh):

309,000 kWh × 0.386 kg = 119,274 kg CO₂ (~119 metric tons)

Typical scenario:

Server capacity impact:

If server handles 10K requests/hour peak:

Bots add 400K requests/month ÷ 720 hours = 555 requests/hour average

Peak overlap: If bots scrape during human peak hours, compete for resources.

Energy consumption increase:

Server power draw (idle): 150W

Server power draw (80% load): 300W

Server power draw (90% load with bots): 330W

Incremental power from bots: 30W sustained

Monthly: 30W × 720 hours = 21.6 kWh

Annual: 259 kWh

Carbon (U.S. grid): 100 kg CO₂/year

Seems small, but scales: 1,000 publishers experiencing this = 100 tons CO₂/year collectively.

III.2 : Bandwidth Infrastructure Energy

CDNs consume power.

Cloudflare, Fastly, Akamai operate global edge networks (hundreds of PoPs—points of presence).

Each PoP:

CDN energy model:

Serving 1TB from CDN ≈ 15-20 kWh (includes all infrastructure overhead)

If AI crawlers consume 500GB/month from your CDN:

0.5 TB × 18 kWh/TB = 9 kWh/month

Carbon: 3.5 kg CO₂/month = 42 kg/year

Small per publisher, but CDNs serve thousands of publishers.

Aggregate CDN energy serving AI crawlers (industry estimate):

AI crawler traffic = 5% of global CDN traffic

Global CDN traffic = ~200 exabytes/month

AI crawlers = 10 exabytes/month

10 million TB × 18 kWh/TB = 180 million kWh/month

Annual carbon: 833,000 metric tons CO₂

For comparison: 180,000 passenger vehicles/year.

https://web.archive.org/web/20260410195102/https://aipaypercrawl.com/articles/ai-crawler-environmental-impact


III.3 : Pré-entraînement

Consommation électrique : - GPT-3 : 1.3 GWh - DeepSeek-V-3 (2024): 2.1-4.2 GWh (selon leur document technique) - Meta’s LLaMA 3–405B (2024): 21 GWh - GPT-4 : 50-70 GWh (40-55x plus que GPT-3) - GPT-5 (estimation) : 100GWh (1.5-2x plus que GPT-4)

Gaz à effet de serre : - GPT-4 : 25 000 tonnes de CO2e (CO2 equivalent) 4 - GPT-5 (estimation) : 42 000 tonnes de CO2e

Eau : - GPT-4 : 600 millions de litres 5 - GPT-5 (estimation) : 1 milliards de litres

6


III.4 : Entraînement

A priori moins pbmatique que l'entraînement écologiquement mais se fait à coup de travail humain


III.5 : Inférence


Différence selon le type de requête : - Classification de texte (analyse de sentiment, codage en catégorie, proximité sémantique) et extraction de courtes citations de document sont les moins énergivores - Classification d'image (Google Vision or Azure Computer Vision) : pas beaucoup plus énergivore, suivi par la détection d'objets - Interaction textuelle "classique" avec un petit modèle (quelques phrases) ~24x plus énergivore que la classification de texte - Interaction textuelle "classique" avec ChatGPT (réponse courte, résumé) ~200x plus énergivore (soit ~0.42Wh) - Pour une interaction longue ça peut monter à 1000x plus - Taille du contexte (le modèle prend toute la conversation en entrée) : la quantité de calcul augmente avec le carré du nombre de tokens en entrée - Choix du modèles (petit ou grand, instantanée ou raisonnement, tel ou tel pays)


Facteurs clés (variation : 0.3WH à >10Wh/prompt): - Taille du modèle - Type de requête (c'est l'aspect génératif qui est consommateur) - Taille du contexte - Paramètres d'utilisation (instantané vs raisonnement) - Pays où le modèle est déployé


III.6 : Tableau récapitulatif

Phase CO₂ / Énergie Eau Matériaux / Déchets
Collecte données Énergie serveurs scraping ; faible mais non nul Refroidissement serveurs de collecte Stockage sur disques durs (terres rares)
Pré-entraînement Phase la plus intensive : centaines à milliers tCO₂ Millions de litres (refroidissement) GPU neufs : cobalt, terres rares, PFC
Fine-tuning / RLHF Modéré : ~10-30x moins que pré-entraînement Proportionnel au compute utilisé Matériel existant, impact marginal
Déploiement Carbone incorporé dans les serveurs de prod Installation physique, câblage Serveurs, câbles cuivre, racks
Inférence Dominant sur durée de vie : milliards de requêtes Continue 24h/24 : millions L/jour Usure accélérée du matériel (e-waste)
Fin de vie Énergie de recyclage Eau de décontamination DEEE : e-waste toxique, métaux lourds

IV : Impacts écologiques au global


IV.1 : Énergie

In 2025, global data centers were estimated to consume 448 TWh of electricity. If data centers’ electricity use were considered a country, it would have ranked 11th globally by electricity consumption. If data centers’ electricity consumption were a country in 2025, their 448 TWh of electricity use would have ranked 11th globally behind the electricity consumption of France (468 TWh)

[@aczelEnvironmentalCostAIs2026]

Data centers’ electricity use in 2025 had a carbon footprint of 189 million tonnes of CO2e, which would require 3.2 billion tree seedlings grown over 10 years to offset, roughly the total number of trees in the entire United Kingdom.

Data centers’ 2025 electricity consumption had a water footprint of 4.5 trillion liters of water—enough to fill 1.8 million Olympic-sized pools or meet the annual basic domestic water needs of over 600 million people in Sub-Saharan Africa.

The land footprint of 2025 data centers’ electricity demand was 6,900 km2, nearly 4.5 times the size of Greater London.

AI workloads accounted for roughly 20% of total data centers’ electricity use in 2025 and are projected to grow to 40% by 2030. If AI’s share of data center energy rises to 40%, electricity consumption could reach 378 TWh—over 9 times the electricity consumption of Nigeria (world’s 6th largest nation by population, with 224 million).

AI is now one of the largest drivers of data center energy consumption. In 2025, AI workloads alone were estimated to account for over 20% of total electricity use in data centers, amounting to approximately 93 TWh61.

If AI workloads alone were a country, its electricity use in 2025 would rank 39th in the world, above industrialized countries like Finland (83 TWh) and Belgium (82 TWh).

Projected global data centers’ electricity consumption could exceed 945 TWh by 2030, accounting for almost 3% of projected global electricity use, If treated as a country, 945 TWh would rank 6th globally by electricity consumption.

The associated water footprint of projected 2030 electricity consumption of data centers is 9.3 trillion liters, or enough to meet the minimum annual domestic water needs of all 1.3 billion residents of Sub-Saharan Africa for a full year.

The associated land footprint of data centers’ expected electricity use in 2030 would be over 14,500 km2, roughly 10 times the size of Mexico City or about twice the Jakarta metropolitan area, home to over 32 million people.

The physical lifecycle of AI hardware presents a growing crisis. AI infrastructure could generate up to 2.5 million metric tons of e-waste annually by 2030, equivalent to discarding nearly 250 Eiffel Towers every year. Without effective recycling, heavy metals (such as lead, cadmium, and mercury) can contaminate soil and water.

Data centers are not uniformly distributed across the globe. Nearly half of the world’s facilities are in the United States, which hosts over 4,000 sites. The next tier includes Germany and the United Kingdom (roughly 500 each), followed by Mainland China, then France, Canada, Australia, India and Japan. Other countries and territories include Italy, Brazil, the Netherlands, Spain, Indonesia, the Russian Federation, Ireland, Switzerland, Malaysia, Sweden, and Hong Kong (SAR), followed by Poland.

Because the electricity mixes of the world's data center hubs differ dramatically—from coal-heavy to nuclear- and hydro-dominated systems—the environmental footprints of data center and AI operations across them also vary, even when facilities share identical engineering designs and hardware. Carbon intensities vary by an order of magnitude across the major data center hubs of the world. Indonesia, India, and Hong Kong (SAR) are among the most carbonintensive grids with carbon footprints 62%, 51%, and 43% higher than the global average, respectively. Poland and Mainland China rank lower with carbon intensities at 30% and 21% higher than the global average. By contrast, the carbon footprint of electricity in the United States, Germany, and Italy is 18%, 24%, and 32% below the global average. At the lowest end, France, Sweden, and Switzerland (37 g) fall about 88–91% below the global average, reflecting their reliance on nuclear energy and hydroelectricity.

Beyond operational electricity, constructing data centers and manufacturing GPUs, batteries, and servers rely on critical minerals (lithium, cobalt, and other rare earths). Early in the life cycle, mining is energy-, water-, and land-intensive and often occurs in jurisdictions with weaker environmental protections.


IV.2 : Gaz à effet de serre

As Kaack et al. (2022) note, most corporate sustainability commitments focus on operational emissions (“Scope 2”) while excluding embodied impacts and external supply chains (“Scope 3”). This narrow focus perpetuates a form of technocratic environmentalism that prioritizes efficiency within corporate boundaries rather than systemic sufficiency across the entire AI ecosystem. [@hlabisaEcologyArtificialIntelligence2025]


IV.3 : Eau

. Data centers, which host most AI systems, often rely on water for cooling—sometimes withdrawing millions of liters per day. In many cases, these withdrawals occur in regions already facing drought or groundwater depletion, amplifying both environmental and social stress. Google’s Mesa data center in Arizona, for example, holds a permit to use 5.5 million cubic meters of water annually41, enough to fill about 2,200 Olympic-sized swimming pools. This volume could meet the basic annual water needs of about 753,000 people in Sub-Saharan Africa at 20 liters per person per day42.[@aczelEnvironmentalCostAIs2026]

Even when some withdrawn water is returned, large-scale withdrawals can strain aquifers and river systems, particularly in arid or groundwater-depleted regions.


IV.4 : Minerais

En comptabilité, la dépréciation permet d'étaler le coût d'un actif sur sa durée de vie utile estimée. Pour leurs serveurs et GPU, la plupart des hyperscalers ont adopté une durée de vie de cinq à six ans.

Michael Burry, l'investisseur qui avait prédit la crise des subprimes de 2008, suggère que la durée de vie réelle de ces équipements pour l'intelligence artificielle serait plutôt de deux à trois ans.

Nvidia est en effet passé d'un cycle d'innovation de deux ans à un rythme annuel, accélérant mécaniquement l'obsolescence de son propre matériel.

https://www.generation-nt.com/actualites/gpu-depreciation-ia-nvidia-michael-burry-datacenter-2066297


IV.5 : Biodiversité et sols

At end of life, poorly managed e-waste can expose frontline communities to hazardous substances. By 2030, AI infrastructure could generate up to 2.5 million metric tons of e-waste each year, roughly equivalent to discarding 250 Eiffel Towers annually. [@aczelEnvironmentalCostAIs2026]

Extracting these materials is energyintensive and environmentally damaging, often depleting water sources and polluting ecosystems43. At the end of their lifecycle, AI systems generate hazardous e-waste. When improperly managed, this waste exposes frontline communities—especially in the Global South—to toxic substances.[@obringerOverlookedEnvironmentalFootprint2021]

AI’s land footprint is not limited to the built environment; it extends upstream through resource extraction and supply chains. Semiconductor manufacturing depends on high-purity silicon, copper, cobalt, and rare earth elements such as neodymium and terbium (Ali et al. 2017; Wübbeke 2013). The mining and processing of these materials often occur in ecologically sensitive areas, such as China, Africa, and South America, resulting in deforestation, soil contamination, and riverine pollution. For instance, tailings and acid mine drainage from rare-earth processing in Inner Mongolia have created toxic waste lakes, which have destroyed local vegetation and compromised groundwater quality (Wübbeke 2013). Such upstream degradation represents a hidden land cost; a displacement of ecological burden from highincome, infrastructure-consuming regions to resourceproducing peripheries. This externalization of impact reinforces global environmental inequality, whereby the biophysical consequences of AI’s growth are disproportionately borne by ecosystems and communities far removed from centers of consumption.

The siting of AI infrastructure often coincides with peri-urban biodiversity hotspots, due to its proximity to transport corridors and grid infrastructure. Large industrial estates alter microclimates, disrupt ecological connectivity, and contribute to urban heat islands, further stressing local flora and fauna.

The mining and smelting of critical minerals involve the destruction of habitats, the introduction of invasive species through road networks, and long-term toxic effects on aquatic ecosystems (Edwards et al. 2014). In biodiversity terms, the AI industry therefore acts as both a direct agent of land transformation and an indirect amplifier of extractive frontiers, accelerating the rate at which natural systems are converted into industrial metabolism.

Within fabrication, hazardous by-products such as silicon tetrachloride pose additional pollution risks if not recycled or neutralized (Thomas 2019). [@hlabisaEcologyArtificialIntelligence2025]


IV.6 : Rapport de l'ADEME


V : Conclusion

Point clés : - Modèles avec des durées de vie basse (donc entraînement souvent) - Scraping permanent - Matériel avec durée de vie basse - Usage massif des plus gros modèles de raisonnement

The synthesis of 46 selected sources reveals that artificial intelligence infrastructures exhibit substantial and interlinked demands for energy, water, materials, and land. These dependencies are mutually reinforcing, forming a socio-ecological metabolism in which each resource dimension intensifies the others. The findings are organized below around these four domains and their crossdomain interactions.

The four resource domains interact in non-linear ways. Higher energy use increases cooling demand, which in turn leads to increased water withdrawals. Water scarcity can restrict energy production and chip fabrication, both of which depend on material extraction that transforms landscapes. These feedbacks form a coupled resourcenexus system, in which local efficiency gains may produce global environmental costs. [@hlabisaEcologyArtificialIntelligence2025]

A central insight from this review is that the environmental burden of artificial intelligence (AI) cannot be understood through any single variable (energy, water, or materials) taken in isolation. Instead, these three domains interact as a resource nexus that defines the thermodynamic and ecological metabolism of AI infrastructures. Each input reinforces or constrains the others: electricity generation depends on water for cooling; semiconductor fabrication consumes both energy and ultrapure water; and the production of advanced materials requires extensive energy and chemical processing. The overall effect is a multiplicative coupling of environmental costs, where local improvements in one dimension may worsen pressures in another. Data centers illustrate this nexus most clearly. As computational loads increase, more servers and cooling equipment are required, resulting in higher electricity and water consumption. Facilities that switch from air to evaporative cooling systems to improve energy efficiency often increase their water usage (Mytton 2021). Conversely, water-free cooling solutions such as closed-loop chillers can increase energy demand by 10–15 percent (Shehabi et al. 2024). These trade-offs are context-specific: in arid regions such as Arizona or the Gulf states, electricity generation itself relies on thermoelectric plants that withdraw vast volumes of water, compounding local scarcity (Yang & Lant 2011). Thus, reducing one environmental metric often amplifies another, creating governance dilemmas that are more about substitution than resolution. The semiconductor sector magnifies this coupling. A single 300-mm wafer may require up to 2.5 m3 of ultrapure water and several hundred kilowatt-hours of electricity (Ruberti 2023; TSMC 2022). Purification and recirculation systems depend on chemical reagents and high-pressure pumps, which themselves consume energy. At the same time, the chemicals used in wafer cleaning (hydrofluoric acid, hydrogen peroxide, and silicon tetrachloride) entail significant embodied emissions and toxicity risks (Shen et al. 2018). The production, transportation, and disposal of these materials form a continuous loop of energy-material exchange that sustains the informational order of the digital economy.

Within this coupled nexus, efficiency gains often fail to reduce aggregate throughput. Improvements in chip architecture or data-center design lower the resource cost per computation but stimulate larger computational demand. As seen with the transition from NVIDIA’s A100 to H100 accelerators, each watt saved per operation enables the proliferation of more servers and training runs (NVIDIA 2022b; Schneider et al. 2025). The same logic applies to water: reductions in cooling water intensity invite higher server density and greater total heat generation. This rebound effect, which illustrates why efficiency-enabling growth remains inadequate when sustainability strategies focus solely on technological optimization (Sorrell 2009), is a key example.

Although successive generations of processors have achieved efficiency gains, Google (2024) reports a more than 30-fold improvement in energy per inference between 2023 and 2024.

Similar trends are visible in NVIDIA’s shift from the A100 to the H100 architecture, where performance per watt has increased dramatically (NVIDIA 2022a; 2022b) [@hlabisaEcologyArtificialIntelligence2025]


V.1 : Sujets laissés de côté


V.2 : Pour aller plus loin...

https://huggingface.co/blog/sasha/environmental-impact-disclosures

https://www.lemonde.fr/les-decodeurs/article/2025/06/08/pourquoi-notre-utilisation-de-l-ia-est-un-gouffre-energetique_6611132_4355770.html

https://bonpote.com/intelligence-artificielle-le-vrai-cout-environnemental-de-la-course-a-lia/#Linference_pour_lusage_quotidien

CRAWFORD K., 2021, Atlas of AI [@crawfordAtlasAIPower2021]

https://theshiftproject.org/publications/intelligence-artificielle-centres-de-donnees-rapport-final/

https://theshiftproject.org/app/uploads/2025/09/Synthese-RF-PIA-1.pdf

https://dl.acm.org/doi/epdf/10.1145/3715275.3732007

https://theconversation.com/lia-generative-est-elle-soutenable-le-vrai-cout-ecologique-dun-prompt-269432

https://lareleveetlapeste.fr/ia-les-data-centers-transforment-les-territoires-en-deserts/

https://www.sciencedirect.com/science/article/pii/S2667010024001690?fr=RR-2&ref=pdf_download&rr=9d8e8aaa3f9499bd

https://www.researchgate.net/publication/403073048_The_data_heat_island_effect_quantifying_the_impact_of_AI_data_centers_in_a_warming_world

https://korben.info/etude-ia-moins-energie-recherche-web.html

https://www.generation-nt.com/actualites/gpu-depreciation-ia-nvidia-michael-burry-datacenter-2066297

V.3 : Solutions

Every kilowatt-hour of electricity used to train or run an AI model carries environmental footprints, including a carbon footprint from the generation mix; a water footprint from electricity production and cooling; and a land footprint from energy infrastructure, reservoirs, and fuel extraction. These three footprints do not always shift in the same direction47,48. For example, switching from coal to bioenergy can, on average, reduce the carbon footprint by 72%, but this comes at the cost of much larger water and land footprints. On average, the water footprint of bioenergy is more than 30 times greater than that of coal and its land footprint is 100 times greater [@aczelEnvironmentalCostAIs2026]

The environmental footprint of energy production in a given location depends on the share of each source in its electricity supply portfolio. For example, in Brazil, where hydropower dominates the grid, the carbon footprint of electricity generation is 77% below the global average. But this comes with other environmental costs: the water and land footprints are nearly triple the global mean.

ADEME


V.4 : Discussion

Dans les conclusions : - engage une réflexivité sur la science - communauté de lecture - question bête - en local - envoyer de la merde

-> le care : homme/femme -> disponibilité/non-réciprocité -> dire qu'on centre sur l'éco car souve, liberté etc pas possible -> la communauté -> usages spé vs humain dispo h24 -> pb de rigueur (balancer tableau sur chatgpt vs générer un code)

-> qcq ça veut dire de faire de la science : une activité individuelle, rationnelle

  1. Aczel, M., Chamanara, S., Matin, M., Farsi, A., Marwala, T., & Madani, K. (2026).  The Environmental Cost of AI’s Energy Use: Carbon, Water and Land Footprints . https://doi.org/10.53328/INR26RMA002
  2. https://web.archive.org/web/20260410195102/https://aipaypercrawl.com/articles/ai-crawler-environmental-impact
  3. Tous les chiffres viennent de https://web.archive.org/web/20260410195102/https://aipaypercrawl.com/articles/ai-crawler-environmental-impact
  4. Ordre de grandeur : 420 000 arbres sur 10 ans
  5. Ordre de grandeur : 237 piscines olympiques
  6. Aczel, M., Chamanara, S., Matin, M., Farsi, A., Marwala, T., & Madani, K. (2026).  The Environmental Cost of AI’s Energy Use: Carbon, Water and Land Footprints . https://doi.org/10.53328/INR26RMA002
  7. Aczel, M., Chamanara, S., Matin, M., Farsi, A., Marwala, T., & Madani, K. (2026).  The Environmental Cost of AI’s Energy Use: Carbon, Water and Land Footprints . https://doi.org/10.53328/INR26RMA002
  8. https://www.amnesty.org/fr/latest/news/2026/05/global-enormous-data-pipelines-powering-major-generative-ai-systems-are-rooted-in-mass-invasions-of-privacy-by-design/

Appendix

  1. Aczel, M., Chamanara, S., Matin, M., Farsi, A., Marwala, T., & Madani, K. (2026).  The Environmental Cost of AI’s Energy Use: Carbon, Water and Land Footprints . https://doi.org/10.53328/INR26RMA002
  2. https://web.archive.org/web/20260410195102/https://aipaypercrawl.com/articles/ai-crawler-environmental-impact
  3. Tous les chiffres viennent de https://web.archive.org/web/20260410195102/https://aipaypercrawl.com/articles/ai-crawler-environmental-impact
  4. Ordre de grandeur : 420 000 arbres sur 10 ans
  5. Ordre de grandeur : 237 piscines olympiques
  6. Aczel, M., Chamanara, S., Matin, M., Farsi, A., Marwala, T., & Madani, K. (2026).  The Environmental Cost of AI’s Energy Use: Carbon, Water and Land Footprints . https://doi.org/10.53328/INR26RMA002
  7. Aczel, M., Chamanara, S., Matin, M., Farsi, A., Marwala, T., & Madani, K. (2026).  The Environmental Cost of AI’s Energy Use: Carbon, Water and Land Footprints . https://doi.org/10.53328/INR26RMA002
  8. https://www.amnesty.org/fr/latest/news/2026/05/global-enormous-data-pipelines-powering-major-generative-ai-systems-are-rooted-in-mass-invasions-of-privacy-by-design/