Skip to content
Get in touchFR
Updated in October 2026
Explained

What is the carbon footprint of an AI-generated image?
CO2, water and comparisons

What an AI-generated image represents in CO2, water and energy, based on measurements published between 2023 and 2025. The article includes a calculator, shows how far generators differ and traces the source of the widely quoted figures on phone charges and litres of water.

Guillaume Pakula
By Guillaume Pakula, co-founder of Celsius. Since 2020, he has helped 80+ organisations with their Bilan Carbone® and climate strategy.
October 2026
Updated October 2026 · 8 min
An AI-generated image emits about 1.3 gCO2e, the equivalent of 9 metres by car and 10 times the emissions of a written question. The energy it uses equals 15% of a phone charge: the often-quoted "whole charge" was the figure for the most energy-intensive generator. Across all studies, measurements range from 0.04 to 5.3 gCO2e, and none of these measurements covers ChatGPT, Gemini or Midjourney.
Key takeaways
  • 1An AI-generated image emits about 1.3 gCO2e, the equivalent of 9 metres by car.
  • 2Across all studies, measurements range from 0.04 to 5.3 gCO2e per image.
  • 3The energy used by an average image equals 15% of a phone charge.
  • 4700 million images would represent about 930 tCO2e if each used 2.9 Wh.

Generating an image with artificial intelligence (AI) emits about 1.3 gCO2e (grams of CO2 equivalent) with global average electricity, the equivalent of 9 metres by car. That is the average in the most cited study, published in 2024: 2.9 Wh (watt-hours) of energy per image. The carbon footprint of an AI-generated image is therefore 10 times that of a written question put to ChatGPT, while a 5-second video emits as much as 31 images.

AI-generated image

What your images emit, and what that equates to

Set the number of images, the generator and the server electricity.

Generator
Electricity

1.3gCO2efor 1 image

  • 9 metresby petrol or diesel car
  • 15%of a phone charge
  • 10written questions to an AI

2.9 Wh of electricity. Order of magnitude, based on open models tested in the laboratory.

Luccioni, Jernite and Strubell, Power Hungry Processing (2024); Ember, Global Electricity Review 2026 (AI Impact Database); ADEME (the French Agency for Ecological Transition), Impact CO2; EPA; Epoch AI. Calculation by Projet Celsius, server electricity only, data as at 2 October 2026.

Across all studies, published measurements range from 0.04 to 5.3 gCO2e per image, and on the same machine the spread between generators is 1 to 46. OpenAI, Google and Midjourney publish no measurement for their services: the values available come from open models, whose code is public, tested in the laboratory between 2023 and 2025.

1How much CO2 does an AI-generated image emit?

About 1.3 gCO2e per image on average

The most cited measurement comes from 3 researchers at Hugging Face and Carnegie Mellon University. They had several open generators produce 1,000 images and found 2.9 Wh per image on average (in French) and 1.35 Wh for the median generator (0.62 gCO2e), which splits the models tested into 2 halves. The median is quoted because a few very energy-intensive generators pull the average up.

An Nvidia Tesla computing card lying flat on a white background, with its black and green casing
An Nvidia Tesla graphics card for computing, photographed in 2012. Images are generated on cards of this type: the 2024 study ran its generators on 8 Nvidia A100 cards, a more recent generation. Photo by Raysonho @ Open Grid Scheduler / Grid Engine, Wikimedia Commons, CC0

From 0.04 to 5.3 gCO2e across all studies

The lowest measurement, 0.04 gCO2e (0.086 Wh), is for the lightest of the 17 generators tested in 2025 by researchers at the University of Florence on a consumer graphics card. The highest, 5.3 gCO2e (11.5 Wh), is for the least efficient generator in the 2024 study, run on data centre graphics cards. Between these 2 limits, MIT Technology Review estimated in May 2025 that a recent generator, Stable Diffusion 3 Medium, used 0.29 gCO2e (0.63 Wh). These values come from 3 different protocols: they give the range of published measurements, a spread of 1 to 130, and do not rank the generators against each other.

An image equals 10 written questions, and a 5-second video equals 31 images

According to Epoch AI's estimate, a written question put to an assistant (in French), also known as a query, uses about 0.3 Wh, or 0.14 gCO2e, so an average image equals 10 questions. A 5-second video (in French), measured in 2025 on the most downloaded open generator, uses 90 Wh, or 41 gCO2e, which equals 31 images, or 300 questions. The environmental impact of AI therefore depends above all on what it is asked to produce.

No measurement published by OpenAI, Google or Midjourney

Since August 2025, Google has published the energy consumption of a written question put to Gemini, at 0.24 Wh. For images, in early October 2026, there was no official data from OpenAI, Google, Midjourney or Black Forest Labs, the developer of Flux. The gap with commercial services therefore remains unknown, in either direction.

2Generator, size, passes and electricity: what makes an image's carbon footprint vary

The result depends on 4 factors: the generator, the size of the image, the number of computation passes and the electricity that powers the servers.

AI-generated image

From the lowest to the highest measurement, a spread of 1 to 130

7 published measurements for one image, converted into CO2e using global average electricity.

1 image, depending on the generator and the study
The lightest generator · 17 models tested in 20250.04 gCO2e0.086 Wh
Stable Diffusion 3 Medium · MIT Technology Review, 20250.29 gCO2e0.63 Wh
The median generator · most cited study, 20240.62 gCO2e1.35 Wh
Stable Diffusion XL · IEA, 2025, graphics card only0.78 gCO2e1.7 Wh
The average of the generators · most cited study, 20241.3 gCO2e2.9 Wh
The heaviest generator · 17 models tested in 20251.9 gCO2e4.08 Wh
The heaviest generator · most cited study, 20245.3 gCO2e11.5 Wh
Luccioni, Jernite and Strubell, Power Hungry Processing (2024); Bertazzini et al., The Hidden Cost of an Image (2025); MIT Technology Review (May 2025); IEA, Energy and AI (2025). Conversion: calculation by Projet Celsius.

The generator: a spread of 1 to 46 on the same machine

The University of Florence researchers tested 17 generators on the same machine, with the same graphics card. The lightest uses 0.086 Wh per image and the heaviest 4.08 Wh, a spread of 1 to 46. The lightest is a model designed to produce an image in a few computation passes, the steps through which a generator gradually refines an image of noise.

Size and number of passes: up to 4.7 times as much energy

An image 1,024 pixels on a side has 4 times as many pixels as one of 512. In the same study, that change multiplies the energy used by up to 4.7 depending on the generator, and one of the models tested is unaffected. Energy use rises in proportion to the number of passes: when MIT Technology Review ran the same model at 25 and then 50 passes, consumption doubled, from 0.63 to 1.2 Wh.

The country of the servers: 11 times less CO2e with French electricity

The same image emits 0.12 gCO2e if the servers run on French electricity (41 gCO2e per kWh in 2025 according to Ember data), which is 11 times lower than with the global average. With US electricity, 384 gCO2e per kWh according to the same source, the figure is 1.1 gCO2e.

3How much electricity and water does a generated image use?

"One image = one phone charge": where the mistake comes from

The comparison comes from the first version of the most cited study, dated November 2023. It stated that the least efficient generator used as much energy as 950 smartphone charges for 1,000 images, or nearly one charge per image, on the basis of 12 Wh per charge. That statement concerned the most energy-intensive of the generators tested, but the articles that repeated it applied it to every image.

The US Environmental Protection Agency (EPA), the source of this value, made 2 revisions to it in 2024. The published version of the study uses 22 Wh and puts the most energy-intensive generator at about half a charge; the value in force since October 2024 is 19 Wh. With this value, that generator accounts for 60% of a charge, the average image for 15% and the median image for 7%.

AI-generated image

On average, an image uses 15% of a phone charge

The share of a phone charge (19 Wh) that one image represents, depending on the generator.

Luccioni, Jernite and Strubell, Power Hungry Processing (2023 and 2024 versions); US Environmental Protection Agency (EPA), value for one phone charge in use since October 2024. Calculation by Projet Celsius.

How much water does an image use? A few millilitres

No study measures the water consumed per image. The order of magnitude can be calculated from the water used per kWh: Google states that a question of 0.24 Wh consumes 0.26 mL, or about 1.1 litres per kWh, to cool its servers. At that rate, an average image uses 3 mL. If we add the water consumed to produce the electricity, which averages 3.1 litres per kWh in the United States according to researchers at the University of California, Riverside, our calculation reaches 12 mL, a little over 2 teaspoons.

"2 to 5 litres of water per image": a figure with no identified source

This figure has been circulating since at least April 2025. One instance, dated 2 April 2025, already attributes it to the same Riverside team, whose study contains no per-image figure: it estimates that 500 mL of water is needed for 10 to 50 written responses from an older model. In June 2025, the Spanish fact-checking site Maldita found that no scientific study addressed the subject. The low value, 2 litres, is 170 times our high estimate. Our article on the water consumed by AI details the scope of these measurements.

4What do millions of generated images add up to?

10 images a day for a year: less than a meal with beef

A person who generates 10 images a day for a year emits 4.9 kgCO2e, a little less than a meal with beef (4.97 kgCO2e according to ADEME). Set against the average carbon footprint of a person in France, 8.2 tCO2e in 2024, this amounts to less than 0.1%.

700 million images in 9 days in spring 2025

On 3 April 2025, Brad Lightcap, OpenAI's chief operating officer, stated that more than 130 million users had generated more than 700 million images since the launch of ChatGPT's new image generator on 25 March, according to TechCrunch. That is about 80 million a day. On 19 May 2026, Google announced more than 50 billion images produced by its generator, up from 5 billion 7 months earlier, or about 200 million a day over the period. Neither of the 2 figures has an independent count behind it.

High-angle view of the Digital Realty data centre buildings in the port of Marseille, with ferries at the quayside
Digital Realty's MRS2, MRS3 and MRS4 data centres in the port of Marseille, in November 2024: an example of a French site, unrelated to the generators cited. Since 2021, they have been cooled by water at 15 °C from a former mining gallery that leads to the sea, according to Data Center Dynamics. Photo by LaMèreVeille, Wikimedia Commons, CC BY-SA 4.0, cropped

2 GWh and 930 tCO2e if every image used 2.9 Wh

If each of these 700 million images had used 2.9 Wh, those 9 days would have required 2 GWh of electricity and emitted about 930 tCO2e with global average electricity: the equivalent of 450 Paris-New York return trips. For Google's 50 billion images, produced in about 9 months, the same calculation gives about 150 GWh, or 0.03% of the electricity France consumes in a year. This Projet Celsius calculation relies on a strong assumption: neither of the 2 services publishes its consumption, and their generators are not the ones the studies measured.

3 ways to emit less when generating images

  • Be precise in your request, or prompt, before running it: every new attempt is one more image.
  • Use a small size for test runs: going from 512 to 1,024 pixels multiplies the energy used by up to 4.7.
  • Keep the standard quality where the tool has a setting for it: doubling the number of computation passes doubles the energy used.

We rank 10 actions, including generating fewer images, in our article on reducing the impact of AI, and we examine what 4 AI services publish in which eco-friendly AI to choose. A company can count its generated images among its purchases of digital services: see our guide to AI in the Bilan Carbone® (the French carbon accounting method) and the AI carbon footprint calculator.

5Key takeaways

  • About 1.3 gCO2e per image on average, the equivalent of 9 metres by car; 0.62 gCO2e for the median generator.
  • From 0.04 to 5.3 gCO2e depending on the study, the generator and the size of the image; 11 times lower with French electricity.
  • 15% of a phone charge for an average image, and 3 to 12 mL of water according to our calculation.
  • No published measurement from the major services: the volume calculations remain assumptions.

Our article on AI electricity consumption gives the orders of magnitude at world scale, and our comparison of ChatGPT, Claude and Gemini sets out, assistant by assistant, what each provider publishes.

Further resources

Frequently asked questions

or: guillaume@projetcelsius.com

More articles to read

View all →
French environmental cost label for clothing: what changed on 1 October 2026
Regulation

French environmental cost label for clothing: what changed on 1 October 2026

6 min read
What is the French textile eco-score?
Who displays it, and has it become mandatory?
Regulation

What is the French textile eco-score? Who displays it, and has it become mandatory?

12 min read
Removable batteries and battery passports: who is affected in 2027?
Regulation

Removable batteries and battery passports: who is affected in 2027?

14 min read

A senior consultant reads every message

Tell us where you stand and what is pressing, and we will reply within 24 working hours with a first view on how to proceed.

Write to a consultant →
Our toolkit

A full range of free assessment tools

Cost estimator, obligation and eligibility checkers, footprint calculator.

Estimate costs and impacts
Ref. 2026·BC·0184REPORT · FINANCIAL YEAR 2026Bilan Carbone®complete 1·2·3Presentation to the Executive Committee · 2027 action planCONTENTS01Executive summaryp. 402Scope and methodologyp. 1203Emissions by scopep. 24042027 action planp. 5605Methodological appendicesp. 78VOLUME84 pages · ConfidentialSTANDARDISO 14064 · GHG ProtocolQUOTATION · BC.2026.0184Issued 05·05·2026 · Valid for 30 daysCelsius consultancy · Paris 3rd arrondissementINDICATIVE COST ESTIMATEYour companyMid-sized company · 180 employees · CSRD 2027SERVICEBilan Carbone®Scope 1·2·3 · 13 weeksBPIFRANCE FUNDING · - 60%INDICATIVE RANGE EXCL. VAT€16,800to €22,400PROJECT DETAILS01 · SCOPING3 weeks~€4,20002 · MODELLING8 weeks~€9,80003 · PRESENTATION2 weeks~€4,800Indicative estimate · Refined after scopingSIRET 891 234 567 00012

Bilan Carbone® cost estimator

Your price range in 1 minute, based on 2026 market rates and public funding.

1 minStart
Estimate costs and impacts
AI FOOTPRINT · MONTHLY USEYOUR TOTAL FOOTPRINT4.2 kgCO₂eq for 1,000 typical prompts · May 2026MODEL COMPARISON · 1,000 TYPICAL PROMPTSGPT-4oOpenAI6.8 kgClaude OpusAnthropic2.1 kgGemini ProGoogle4.5 kgMistral LargeMistral AI · FR1.6 kg

AI carbon footprint calculator

The climate impact of your AI queries, by model and task. Ecologits methodology.

3 minStart
Check an obligation
DIAGDÉCARBON'ACTIONACCELERATE THETRANSFORMATIONOF YOUR COMPANYIN ENERGY ANDTHE ENVIRONMENTOFFICIAL SCHEME · BPIFRANCE × ADEMECHECKER · ELIGIBILITYELIGIBLEYou meet the 3 Diag Décarbon'Action criteria.YOUR FUNDING€10,000 excl. VAT€6,000excl. VATAmount payable by you after the Bpifrance subsidyFUNDING BREAKDOWNBPIFRANCE 40%YOU 60%€4,000€6,000SCHEME SUBSIDISED BY

Diag Décarbon'Action eligibility

Check in 30 seconds whether your business is eligible for Bpifrance funding covering 40% of a Bilan Carbone® assessment.

30 secStart
Check an obligation
FRENCH REPUBLICMinistry for the Ecological TransitionLEGAL REQUIREMENTGreenhouse GasEmissions AssessmentArticle L. 229-25 of the French Environmental CodeDecree 2022-982 · published 1 July 2022FREQUENCY · 4 YEARSFILING · ADEMECHECKER · BEGES REQUIREMENTREQUIRED720 employees · BEGES requirementPOTENTIAL PENALTY€50,000maximum fine for failing to complete itart. L. 229-25OFFICIAL REFERENCE

BEGES checker

Does the BEGES requirement apply to you? An immediate answer, with the deadline and penalty.

30 secStart
Check an obligation
9:42SCANSCANNING DPP...DPP IDENTIFIEDOrganic cotton T-shirtSKU TX-CB-220 · Batch L-26-04781ESPR COMPLIANT42 DATA POINTS · 5 SECTIONSORIGINIndia · GOTSOrganic cottonFOOTPRINT5.2 kg CO₂eqLCA ISO 14040RECYCLABILITY85%Cotton recycling schemeREPAIRABILITY7.5 / 10Accessible partsEND OF LIFETextile recycling scheme (TLC) · Bordeaux/FRRefashion-approved recovery operatorEU · ESPR 2024/178112.05.2026

DPP checker

Is your product covered by the Digital Product Passport?

1 minStart
Check an obligation
COSTENVIRONMENTAL386POINTS257PER 100 GOfficial methodEcobalyse, v7.0.0CHECKER · ENVIRONMENTAL LABELLINGELIGIBLETextiles · voluntary labellingENVIRONMENTAL COST386 ptsT-shirt 150 g · 257 pts per 100 gExample: Ministry for the Ecological Transition FAQ

Textile environmental labelling 2026

Voluntary labelling, but third parties can publish it from October 2026: where do you stand?

1 minStart