ChatGPT carbon footprint: what a query costs and how much AI weighs in a company's footprint

Google's 2026 environmental report shows emissions 81% higher than in 2019, including an 18% rise in 2025 alone, driven by the construction of data centres for AI. This is the most visible effect of AI on the climate, along with the water AI consumes to cool these facilities. The question is what this means for a company that uses these tools.

In most cases, it matters far less than the headlines suggest, as long as you think in orders of magnitude (in French) and compare AI with the other emission sources in the footprint.

1An AI query uses about as much energy as a Google search

For a long time, a query to an AI model was said to use 2 to 3 Wh, or 10 to 15 times as much as a Google search. That figure is now out of date: recent analyses have divided it by about 10, and according to Epoch AI the old estimate was about 10 times too high. For a model-by-model breakdown, the ChatGPT, Claude and Gemini comparison ranks the published figures.

A short text query to ChatGPT, Gemini or Claude uses about 0.3 Wh: Google measured a median Gemini query at ~0.24 Wh (arXiv study 2508.15734), Sam Altman puts ChatGPT at ~0.34 Wh, and a benchmark of GPT-4o, a model withdrawn from ChatGPT in February 2026, gives ~0.43 Wh. An AI query and a Google search are therefore of the same order of magnitude, nowhere near a factor of 10. For more on this comparison, see ChatGPT query vs Google search.

• AI query (generative AI, short text): ~0.3 Wh, not the 2-3 Wh that used to be quoted
• Standard Google search: same order of magnitude, ~0.3 Wh
• Life cycle of a large model: 20,400 tCO2e for Mistral Large 2, from training to January 2025, including 18 months of use, according to the life cycle assessment (LCA) that Mistral published with Carbone 4 and ADEME (the French Agency for Ecological Transition)
• Data centres worldwide: ~1.5% of global electricity consumption in 2024, projected to reach ~3% in 2030 (base case of the IEA, the International Energy Agency)

2AI makes up a small share of an SME's carbon footprint

For an SME with 200 employees that uses AI tools intensively (100 queries per employee per day, which is already a lot), the annual footprint of that use stays below 1 tonne of CO2, very little next to the main emission sources:

Provider LCA - Mistral × Carbone 4 × ADEME

1.14 gCO2e for 400 tokens: a figure to read within its boundary

See the Mistral Large 2 LCA

Mistral has published the first complete life cycle assessment of a large model, covering production, training, inference (meaning use of the model to answer queries) and the end of life of the servers. The figure is useful but cannot be compared point by point with Google's per-query measurement (~0.24 Wh), which covers only the data centre's electricity. A broad boundary on one side, a production indicator on the other.

• Purchases and subcontracting (Scope 3 (in French)): often 40-70% of the total footprint
• Building energy: 10-20% on average
• Business travel: 5-15% for the executive committee alone (in French)
• IT equipment (manufacturing): 70-80% of the IT carbon footprint
• AI and cloud: less than 1% for most SMEs

3AI use is growing fast, and its trajectory is worth monitoring

If AI use doubles every year (as it does in many companies) and efficiency gains in models do not offset that growth, digital could become a significant emission source within 5 to 10 years. For now, they do not: the gains are absorbed by the increasing size of models.

Including AI in your Bilan Carbone® (the French carbon accounting method) now gives you a baseline against which to measure this change. Methodologies exist, although they are still young: the ADEME digital reference framework and the database of the French organisation Boavizta.

45 levers to reduce your company's digital carbon footprint

In descending order of impact:

The cloud's physical footprint - French presidency (Élysée), February 2025

Where the "cloud" lands in France

Read the Le Parisien report (in French)

About thirty sites announced by 2030, concentrated in Hauts-de-France, Île-de-France and Grand Est. Behind the word "cloud" lie hectares of land, a grid connection and a local electricity mix, which together decide the real footprint. The issue has moved beyond energy and become one of land-use planning.

• Extend the lifespan of equipment from 4 to 6 years: the annualised manufacturing footprint falls by about 33%. This is the number 1 lever, and it saves money.
• Include refurbished equipment in the IT purchasing policy: a refurbished PC has a footprint ~70% lower than a new one
• Choose a cloud hosting provider in France or Scandinavia: emissions are 5 to 10 times lower than with a US data centre powered by natural gas
• Use AI selectively: a chatbot that answers questions the FAQ already covers wastes energy. Reserve AI for high-value tasks
• Choose smaller models when the task allows it: a small model uses 5 to 10 times less energy than a frontier model on simple tasks

5Key takeaways

For most organisations, AI has a real but still marginal carbon cost. Within digital, manufacturing equipment weighs far more than the use of cloud services; what needs watching in the coming years is how AI use develops.

• An AI query uses about as much energy as a Google search (~0.3 Wh each): the old factor of 10 has not survived recent measurements, and usage remains marginal at the scale of a company's carbon footprint.
• AI accounts for less than 1% of the total footprint of most SMEs, even when it is used intensively
• The most effective digital lever is extending the lifespan of equipment: going from 4 to 6 years cuts the annualised manufacturing footprint by about 33%.
• The choice of data centre matters: hosting in France or Scandinavia emits 5 to 10 times less than hosting in the United States, on a life cycle assessment basis.
• Google's emissions have risen by 81% since 2019: the issue is real at the macro level, but it is not (yet) a priority for an SME.

Further resources

Frequently asked questions

1How much energy does a ChatGPT query use?

About 0.3 Wh for a short text query. The long-standing estimate of 2-3 Wh (10 times a Google search) has been divided by about 10 by recent analyses: Google measured a median Gemini query at ~0.24 Wh, and Sam Altman puts ChatGPT at ~0.34 Wh. An AI query and a Google search are now of the same order of magnitude.

2How much does AI weigh in a company's carbon footprint?

For most companies, AI is not a priority emission source: it accounts for less than 1% of the footprint in most cases. The priorities remain purchases (Scope 3 (in French)), energy and travel. AI is something to include and monitor, not an emergency.

3How do you calculate the carbon footprint of AI tools?

ADEME's digital reference framework and the Boavizta database provide emission factors for each type of cloud service. In practice, your cloud provider (AWS, Azure, GCP) also publishes emissions data by service. These figures go into your Bilan Carbone® under digital services in Scope 3.

4Should we stop using AI for climate reasons?

No. The aim is to use AI selectively rather than systematically, and to choose hosting providers with a low-carbon electricity mix. Giving up AI to save 1 tonne when the executive committee's travel (in French) emits 24 would be a poor trade-off.

5Is a data centre in France lower-carbon than one in the US?

Yes, 5 to 10 times lower on a life cycle assessment basis, thanks to the French electricity mix (90% low-carbon, mainly nuclear). It is a simple but rarely used lever: choose a hosting provider with data centres in France or Scandinavia.

6Can AI help reduce a company's emissions?

Yes, if it is used well: it can optimise logistics, analyse consumption data and model reduction scenarios. Its net effect depends on what it saves compared with what it consumes.

7What is the carbon footprint of generative AI?

The footprint arises in 2 phases: training, which happens once, and use (inference, meaning running a trained model to answer a query), which recurs with every query. The LCA published by Mistral puts Mistral Large 2 at 20,400 tCO2e, including training and the first 18 months of use. A short text query uses about 0.3 Wh, the same order of magnitude as a Google search: Google measured 0.24 Wh for a median Gemini query. Worldwide, data centres accounted for about 1.5% of electricity consumption in 2024, and the IEA projects a share of around 3% in 2030.

8How can a company reduce the carbon footprint of its AI use?

The first lever is targeted use: reserving AI for high-value tasks rather than using it by default (for example, to answer questions that an FAQ already covers). Next come smaller models, where the task allows it (5 to 10 times less energy than a frontier model on simple tasks), and hosting in France or Scandinavia, which can cut emissions by up to a factor of 5 to 10 compared with a US data centre powered by gas. Within a company's digital carbon footprint, extending the lifespan of equipment from 4 to 6 years matters more: the annualised manufacturing footprint falls by about 33%.

Ask Guillaume your questionA personal reply within 24 working hours

Send my questionor: [email protected]