---
title: "How and why does AI harm the environment?\nWhere do its emissions come from?"
description: "In Mistral AI's life cycle assessment of an AI model, training and answers weigh most for climate and water, server hardware for mineral resources."
canonical: https://projetcelsius.com/en/blog/why-ai-is-bad-for-the-environment/
author: "Guillaume Pakula"
datePublished: 2026-10-05
dateModified: 2026-10-05
language: en
translationOf: https://projetcelsius.com/blog/comment-ia-pollue/
---

# How and why does AI harm the environment? Where do its emissions come from?

**Author**: Guillaume Pakula
**Published**: 2026-10-05
**Section**: Explained
**Canonical URL**: https://projetcelsius.com/en/blog/why-ai-is-bad-for-the-environment/
**French original**: https://projetcelsius.com/blog/comment-ia-pollue/
**Source**: Projet Celsius (https://projetcelsius.com/en/), a climate strategy consultancy of engineers and PhDs based in Paris and Marseille.

> AI has an impact at 4 points: manufacturing the chips, building the data centres, training the model, then the electricity and water behind each answer. In the life cycle assessment published by Mistral AI, training and answers account for **85.5% of greenhouse gases**, and server hardware for **61% of mineral resources**.

## Key takeaways

- An AI model harms the environment at **4 stages**, from chip manufacturing to the answers it gives.
- In Mistral AI's life cycle assessment, training and answers account for **85.5% of greenhouse gas emissions**.
- Server hardware accounts for **61% of mineral resource depletion** in Mistral AI's life cycle assessment.
- Trial runs added **50%** to training emissions in a study that measured them.

Artificial intelligence (AI) harms the environment because it relies on a deeply physical industry: **chips to manufacture**, data centres to build, models to train, then **electricity and water for every answer**. In the results published by [Mistral AI](https://mistral.ai/news/our-contribution-to-a-global-environmental-standard-for-ai) for one of its models, training and answers account for **85.5% of greenhouse gas emissions**, and server hardware for **61% of mineral resource depletion**.

This study is a [life cycle assessment](https://projetcelsius.com/en/blog/lca-vs-carbon-footprint/): it counts a product's impacts from the extraction of materials to the end of life of the equipment. It shows that the dominant stage **changes with the impact considered**: running the servers for climate and water, their hardware for mineral resources.

## Why AI harms the environment: a life cycle assessment of one model

### Behind the screen, servers in a data centre

An AI answer is computed by **graphics processors (GPUs)**, chips specialised for this kind of calculation and mounted in servers lined up in their thousands in data centres, the buildings known as "the cloud". Our article on the [environmental impact of AI](https://projetcelsius.com/en/blog/environmental-impact-of-ai/) puts figures on these impacts worldwide; this one explains where they come from.

### 4 stages, from chip to answer

An AI model goes through **4 stages**: [manufacturing the chips and servers](https://projetcelsius.com/en/ai/analyses/what-is-ai-made-of/), building the data centre, training the model (tuning it on vast quantities of text) and answering questions, a stage the industry calls "inference". The first 3 take place mostly **before the first question**, while stage 4 is repeated with every use.

### A stage-by-stage life cycle assessment, published by Mistral AI

In July 2025, Mistral AI published the results of the life cycle assessment of [its Large 2 model (in French)](https://projetcelsius.com/ia/graphiques/acv-mistral-large-2/). The study was carried out with the firm Carbone 4 and ADEME (the French Agency for Ecological Transition), then **reviewed by 2 engineering consultancies**. It follows the frugal AI framework from AFNOR, France's national standards body: [AFNOR Spec 2314](https://afnor.org/en/news/referential-for-measuring-and-reducing-environmental-impact-of-ia).

As of January 2025, after 18 months of use according to the company, the model had emitted **20,400 tCO2e** (tonnes of CO2 equivalent) and consumed **281,000 m³ of water**, the annual household water use of about [5,400 people](https://economie.eaufrance.fr/sites/default/files/2026-06/rapport_sispea_2024_vf.pdf) in France. The study also counts the minerals and metals extracted, expressed as a reference metal in the same way greenhouse gases are expressed as CO2: **660 kg of antimony equivalent**. According to [ADEME](https://librairie.ademe.fr/economie-circulaire-et-dechets/9495-avis-de-l-ademe-l-intelligence-artificielle-generative-des-impacts-environnementaux-importants.html), 20,000 tCO2e is the annual footprint of **more than 2,400 people in France**.

These figures apply to **a single model**, from one developer, at one date. As of October 2026, no other developer has published equivalent results, broken down by stage and covering several impacts.

## Before the first question: manufacturing, construction and training

### Chips etched in East Asia, a building to put up

According to [Greenpeace East Asia](https://www.greenpeace.org/static/planet4-eastasia-stateless/2025/04/5011514f-greenpeace_chipping_point.pdf), the AI chips from Nvidia and AMD that the organisation studied are etched in Taiwan, and their memory chips are produced in South Korea and Japan. [The electricity used for this manufacturing (in French)](https://projetcelsius.com/ia/graphiques/fabrication-puces-ia-electricite/) **rose 4.5-fold in a year**, reaching **984 GWh** (gigawatt hours) in 2024. Nvidia sells its GPUs in groups of 8 mounted on a motherboard, the board that links them. Making one such motherboard emits [1,312 kgCO2e](https://images.nvidia.com/aem-dam/Solutions/documents/HGX-H100-PCF-Summary.pdf), and [2,274 kgCO2e](https://images.nvidia.com/aem-dam/Solutions/documents/HGX-B200-PCF-Summary.pdf) for the next generation. Memory accounts for **42% of the total in the first case and 49% in the second**.

In Mistral AI's assessment, server hardware (manufacturing, transport and end of life) accounts for 11% of greenhouse gases and 5% of water consumed, but **61% of mineral resource depletion**. Construction of the data centre stays **under 1% of emissions** and accounts for 1.5% of mineral resources.

### Training: millions of hours of computation

Meta reports **39 million GPU hours** and [11,390 tCO2e](https://huggingface.co/meta-llama/Llama-3.1-405B) to train its 3 Llama 3.1 models in 2024, the equivalent of **4,500 GPUs running for a year**. Google researchers estimated [the training of GPT-3 (in French)](https://projetcelsius.com/ia/graphiques/entrainement-modeles-emissions/), the OpenAI model released in 2020, at [552 tCO2e](https://arxiv.org/abs/2104.10350): **about 16 times less** than the largest of Meta's 3 models (8,930 tCO2e).

These figures cover the training of published models. For a family of models whose authors also counted the trial runs, development added **50%** to the emissions of the final training according to a [2025 study](https://arxiv.org/abs/2503.05804), a share that developers **do not generally publish**.

## With every answer: electricity and water

### 1.14 gCO2e and 45 mL of water for an answer of about 320 words

Once the model is trained, every question triggers a computation. Mistral AI puts an answer of about 320 words from its assistant at **1.14 gCO2e**, the equivalent of driving a car for 8 metres, and **45 mL of water**, **for the answer alone**, not counting the user's device. Our article on [how much energy a ChatGPT query uses](https://projetcelsius.com/en/blog/chatgpt-query-energy-consumption/) covers this stage, and the one on [AI electricity consumption](https://projetcelsius.com/en/blog/ai-electricity-consumption/) shows what it represents across data centres.

### The country hosting the servers changes the result

In France, 1 kWh of electricity generated emitted **41 gCO2e** in 2025, including the construction of power plants and the extraction of fuels, against 458 g on average worldwide, according to data from [Ember](https://ourworldindata.org/grapher/carbon-intensity-electricity), an energy think tank. French electricity therefore emits **11 times less**. Mistral AI's page **does not specify the country of its servers**; it states that electricity is counted at the carbon intensity of the local grid, without accounting for renewable electricity purchase contracts.

### Water cools the servers, but most of it is used to generate their electricity

Servers heat up and must be cooled, often with water. According to the [International Energy Agency (IEA)](https://iea.blob.core.windows.net/assets/de9dea13-b07d-42c5-a398-d1b3ae17d866/EnergyandAI.pdf), however, this direct cooling accounts for only **a quarter of the water consumed** by data centres: **two thirds goes to generating their electricity**, and the rest to manufacturing chips. Our article on [the water consumed by AI](https://projetcelsius.com/en/blog/ai-water-footprint/) sets out these scopes in detail.

## The stage that weighs most depends on the impact measured

### For climate and water: training and answers

In Mistral AI's assessment, training and answers, meaning the servers' electricity and cooling, account for **85.5% of greenhouse gases and 91% of water**. **The published page does not separate the 2 stages**, so it is impossible to say which weighs more. Worldwide, the [GreenIT](https://greenit.eco/nos-etudes-et-essais/impacts-environnementaux-et-sanitaires-de-lintelligence-artificielle-dans-le-monde-en-2025-et-2030/) association finds the same order with a different method and scope (servers and data centres): the use phase accounts for **92% of a score that aggregates 15 impacts**, and manufacturing for 7%.

### For mineral resources: server hardware

For mineral resources, the order is reversed: server hardware accounts for **61%**, training and answers for 29% and users' devices for 7%. This result points to **the lifespan of servers and chips** as the main lever for this impact. At the end of its life, this equipment becomes electronic waste (also known as e-waste): **1.2 to 5 million tonnes** in cumulative terms between 2020 and 2030 for generative AI, according to a [2024 study](https://www.nature.com/articles/s43588-024-00712-6).

### What this assessment does not say

Mistral AI presents its study as **a first approximation**: in 2025, no reliable inventory of GPU manufacturing existed yet (in May 2026, [ADEME](https://librairie.ademe.fr/economie-circulaire-et-dechets/9357-life-cycle-assessments-of-gpus-for-artificial-intelligence.html) published a life cycle assessment of 8 GPUs), which makes the hardware share less solid than the others. **The full report is not public**, and the results cover 3 impacts without detailing **which metals are involved**. The shares would differ for another model, trained elsewhere or used by more people. Nor does the study assess the indirect effects of AI, such as what its users do with it.

## Key takeaways

- **AI harms the environment at 4 stages**, from chip manufacturing to answers, including data centre construction and model training.
- **Training and answers account for 85.5% of greenhouse gases** and 91% of water in Mistral AI's assessment.
- **Server hardware accounts for 61% of mineral resource depletion**, compared with 11% of emissions.
- **These shares come from a single model**, and are the only ones published by stage: they do not apply to all AI.

Knowing which stage weighs most helps you prioritise [the actions that count for reducing the impact of AI](https://projetcelsius.com/en/blog/reduce-ai-environmental-impact/), and read what developers publish before choosing [a more eco-friendly AI](https://projetcelsius.com/en/blog/most-eco-friendly-ai/). Our [comparison of ChatGPT, Claude and Gemini](https://projetcelsius.com/en/blog/chatgpt-vs-claude-vs-gemini-environmental-impact/) brings together the measurements available for each assistant. Our guide on [AI in the Bilan Carbone®](https://projetcelsius.com/en/blog/ai-emissions-carbon-accounting-scope-3/), the French carbon accounting method, explains how a company counts them.

## Frequently asked questions

### Why is ChatGPT bad for the environment?

Like all AI assistants, ChatGPT runs on servers fitted with GPUs in data centres. Those chips had to be manufactured, the buildings built and the models trained, and each answer then consumes electricity and water. OpenAI has not published a life cycle assessment: the known orders of magnitude come from other developers and independent measurements.

### Does AI training or AI use have the bigger environmental impact?

The results published by Mistral AI do not separate the two: together, they account for 85.5% of greenhouse gases. Training happens once, whereas answers accumulate with use, so the longer a model is in use, the larger the share of answers becomes. The trial runs that precede training, which are rarely published, added 50% to training emissions in a study that measured them.

### Why does AI use water?

For 2 main reasons: to cool the servers and to generate the electricity they consume. According to the International Energy Agency, two thirds of the water used by data centres goes to electricity generation, a quarter to direct cooling and the rest to manufacturing chips. Mistral AI counts 45 mL of water for an answer of about 320 words.

### Is manufacturing AI chips bad for the environment?

Nvidia puts the manufacture of a motherboard of 8 GPUs at 1,312 kgCO2e (about 160 kgCO2e per GPU, with the board and cooling included) and that of the next generation at 2,274 kgCO2e. For comparison, ADEME counts about 193 kgCO2e for a laptop. Over the life cycle of a model, server hardware accounts for a minority of greenhouse gases (11% at Mistral AI) but dominates for mineral resources (61%).

### Does where an AI is hosted change its carbon footprint?

For the climate, yes, other things being equal: French electricity emits 11 times less CO2e per kWh than the world average, and running the servers accounts for most of the emissions. For mineral resources, barely: 61% come from server hardware, which is the same wherever it is installed.

## Sources

- **Life cycle assessment**: [Mistral AI, with Carbone 4 and ADEME, Our contribution to a global environmental standard for AI](https://mistral.ai/news/our-contribution-to-a-global-environmental-standard-for-ai), 22 July 2025, reviewed by Resilio and Hubblo; [AFNOR Spec 2314, general framework for frugal AI](https://afnor.org/en/news/referential-for-measuring-and-reducing-environmental-impact-of-ia), June 2024; [ADEME, opinion on generative artificial intelligence](https://librairie.ademe.fr/economie-circulaire-et-dechets/9495-avis-de-l-ademe-l-intelligence-artificielle-generative-des-impacts-environnementaux-importants.html) (in French), 22 July 2026, p. 6.
- **Manufacturing and end of life**: [Nvidia, HGX H100 Product Carbon Footprint Summary](https://images.nvidia.com/aem-dam/Solutions/documents/HGX-H100-PCF-Summary.pdf) and [HGX B200 Product Carbon Footprint Summary](https://images.nvidia.com/aem-dam/Solutions/documents/HGX-B200-PCF-Summary.pdf), 2025, p. 3; [Greenpeace East Asia, Chipping Point](https://www.greenpeace.org/static/planet4-eastasia-stateless/2025/04/5011514f-greenpeace_chipping_point.pdf), April 2025, pp. 7 and 20; [ADEME, Life cycle assessments of GPUs for artificial intelligence](https://librairie.ademe.fr/economie-circulaire-et-dechets/9357-life-cycle-assessments-of-gpus-for-artificial-intelligence.html), May 2026; [Wang et al., E-waste challenges of generative artificial intelligence](https://www.nature.com/articles/s43588-024-00712-6), Nature Computational Science, 2024; [ADEME, Impact CO2](https://impactco2.fr/), accessed 2 October 2026.
- **Training**: [Meta, Llama 3.1 model card](https://huggingface.co/meta-llama/Llama-3.1-405B), 2024; [Patterson et al., Carbon Emissions and Large Neural Network Training](https://arxiv.org/abs/2104.10350), 2021, p. 7; [Morrison et al., Holistically Evaluating the Environmental Impact of Creating Language Models](https://arxiv.org/abs/2503.05804), 2025.
- **Electricity and water**: [Ember, Global Electricity Review 2026](https://ember-energy.org/latest-insights/global-electricity-review-2026/), and [Our World in Data, based on Ember](https://ourworldindata.org/grapher/carbon-intensity-electricity), carbon intensity of electricity in 2025, life cycle (458 gCO2e per kWh worldwide, 41 in France); [IEA, Energy and AI](https://iea.blob.core.windows.net/assets/de9dea13-b07d-42c5-a398-d1b3ae17d866/EnergyandAI.pdf), April 2025, pp. 242 and 243; [Association Green IT (GreenIT), Impacts environnementaux de l'intelligence artificielle dans le monde](https://greenit.eco/nos-etudes-et-essais/impacts-environnementaux-et-sanitaires-de-lintelligence-artificielle-dans-le-monde-en-2025-et-2030/) (in French), September 2026; [Observatoire Sispea, report on 2024 data](https://economie.eaufrance.fr/sites/default/files/2026-06/rapport_sispea_2024_vf.pdf) (in French), June 2026.
- **Images**: Opening photo: Enrique Jiménez, [Wikimedia Commons](https://commons.wikimedia.org/wiki/File:Silicon_Wafer_20190210.jpg), CC BY-SA 2.0, cropped. Chuquicamata: Diego Delso, [Wikimedia Commons](https://commons.wikimedia.org/wiki/File:Mina_de_Chuquicamata,_Calama,_Chile,_2016-02-01,_DD_110-112_PAN.JPG), CC BY-SA 4.0. The Dalles: Tony Webster, [Wikimedia Commons](https://commons.wikimedia.org/wiki/File:Google_Datacenter_-_The_Dalles,_Oregon_%2817832143871%29.jpg), CC BY 2.0, cropped.
