---
title: "What is the environmental impact of an AI-generated video?"
description: "A 5-second AI video emits about 41 gCO2e (290 metres by car). A video 2 times as long needs 4 times the energy. No published measurement covers commercial services."
canonical: https://projetcelsius.com/en/blog/ai-video-environmental-impact/
author: "Guillaume Pakula"
datePublished: 2026-10-05
dateModified: 2026-10-05
language: en
translationOf: https://projetcelsius.com/blog/impact-environnemental-video-ia/
---

# What is the environmental impact of an AI-generated video?

**Author**: Guillaume Pakula
**Published**: 2026-10-05
**Section**: Explained
**Canonical URL**: https://projetcelsius.com/en/blog/ai-video-environmental-impact/
**French original**: https://projetcelsius.com/blog/impact-environnemental-video-ia/
**Source**: Projet Celsius (https://projetcelsius.com/en/), a climate strategy consultancy of engineers and PhDs based in Paris and Marseille.

> A 5-second AI-generated video emits **about 41 gCO2e**, equivalent to **driving 290 metres** and to the emissions of 31 [AI-generated images](https://projetcelsius.com/en/blog/ai-image-carbon-footprint/). **A video 2 times as long needs 4 times as much energy.** Depending on the generator and its settings, measurements range from 0.06 to 190 gCO2e, and none of them covers a commercial service: the **"1 kWh per Sora video"** is an estimate.

## Key takeaways

- A 5-second AI-generated video emits **about 41 gCO2e**, equivalent to driving 290 metres.
- A video **2 times as long** needs 4 times as much energy to generate.
- Measurements span a ratio of **1 to 3,000**, depending on the generator and its settings.
- Generating **1 video a day for a year** emits 15 kgCO2e, equivalent to driving 110 km.

Generating a 5-second video with artificial intelligence (AI) emits **about 41 gCO2e** (grams of CO2 equivalent) with world average electricity, equivalent to driving **290 metres**. According to a [study published in September 2025](https://arxiv.org/abs/2509.19222), the most downloaded open video generator (also known as an open text-to-video model) uses [90 Wh (watt-hours) of electricity per video (in French)](https://projetcelsius.com/ia/chiffres/electricite-video-generee/), equivalent to **31 AI-generated images** or **300 written questions**.

A video 2 times as long needs **4 times as much energy**. Published measurements range from **0.06 to 190 gCO2e** depending on the generator and its settings, a ratio of 1 to 3,000. Google, OpenAI and the Chinese video apps publish no **measurement** for their own services.

## How much energy does an AI-generated video use?

### About 41 gCO2e for a 5-second video at 720p

The measurements come from 3 researchers at Hugging Face. In September 2025, they published the energy use of **[7 open video generators (in French)](https://projetcelsius.com/ia/graphiques/video-generee-7-modeles/)** (generators whose code is public), tested on a data centre graphics card. Their reference model is the most downloaded of them: it uses **90 Wh for a video of 81 frames at 720p** (the resolution of an HD screen), a little over 5 seconds long. It is **the second most energy-intensive of the 7**, while the median model, which produces lower-resolution videos, uses **25 Wh**, or 12 gCO2e.

> **The calculation** Multiplying 90 Wh by 458 gCO2e per kWh, the [average carbon intensity of world electricity in 2025](https://ember-energy.org/latest-insights/global-electricity-review-2026/) according to Ember, an energy think tank, counting the construction of power plants and the extraction of fuels, gives 41 gCO2e. According to [Impact CO2](https://impactco2.fr/), the comparison tool of ADEME (the French Agency for Ecological Transition), a petrol or diesel car emits 142 gCO2e per km, so 41 g corresponds to 290 metres. With French electricity, 41 gCO2e per kWh according to [Ember data](https://ourworldindata.org/grapher/carbon-intensity-electricity), the same video emits 3.7 gCO2e. This Projet Celsius calculation gives an order of magnitude only: it counts the electricity used by the servers as measured in the laboratory, and excludes their cooling and manufacture.

### A ratio of 1 to 3,000 depending on the generator and its settings

In the same study, the least energy-intensive generator uses **0.14 Wh**, or 0.06 gCO2e, for a video of less than 2 seconds at low resolution, while the most energy-intensive uses **415 Wh**, or 190 gCO2e, for 5 seconds at 720p. This ratio of **1 to 3,000** reflects the **default settings** (duration, resolution, number of computation passes) as much as the generators themselves. The published figures also disagree with one another: for the same model, the [International Energy Agency (IEA)](https://iea.blob.core.windows.net/assets/de9dea13-b07d-42c5-a398-d1b3ae17d866/EnergyandAI.pdf) puts the graphics card alone at **115 Wh**, while the 2025 study measures **22**, and nothing explains this difference of 1 to 5.

### Equivalent to 31 images or 300 written questions

An [AI-generated image](https://projetcelsius.com/en/blog/ai-image-carbon-footprint/) uses 2.9 Wh on average, and a [written question](https://projetcelsius.com/en/blog/chatgpt-query-energy-consumption/) put to an AI assistant (also called a query) uses 0.3 Wh. [The 5-second video (in French)](https://projetcelsius.com/ia/graphiques/electricite-requete-texte-image-video/) is therefore equivalent to **31 images, or 300 questions**.

The same video also uses as much electricity as **1 hour 10 minutes of streamed video**, screen and network included, based on the 77 Wh per hour that the [IEA](https://www.iea.org/commentaries/the-carbon-footprint-of-streaming-video-fact-checking-the-headlines) calculated for 2019. Data centres account for only 5% of that hour: counting the servers alone, generating 5 seconds is equivalent to streaming **23 hours**. Our article on the [carbon footprint of TikTok](https://projetcelsius.com/en/blog/tiktok-carbon-footprint/) covers video viewing.

### No measurement published by Google, OpenAI or the Chinese apps

In early October 2026, there was **no published measurement** from Google for Veo, from OpenAI for Sora (shut down in 2026), from Kuaishou for Kling or from Runway. These values therefore describe open models, and **nothing indicates whether commercial services do better or worse**.

They also come from **a single team**, Sasha Luccioni's group at Hugging Face, which is behind the measurements picked up by the IEA and by MIT Technology Review. The [environmental impact of AI](https://projetcelsius.com/en/blog/environmental-impact-of-ai/) also plays out at the scale of data centres, which per-video measurements do not capture.

## Why an AI-generated video uses so much energy: frames, duration and resolution

### Dozens of frames calculated together

A video is a sequence of frames, **from 8 to 30 per second** for the generators in the 2025 study. The generator calculates **the whole sequence in one go**, so that objects and movements stay consistent from one frame to the next. **Every additional frame** therefore enters the calculation for all the others.

### A video 2 times as long, 4 times as much energy

The 2025 study measured this on the most downloaded generator: **doubling the duration multiplies the energy by about 4**. A video 2 times as long (nearly 11 seconds) would therefore emit **170 gCO2e** instead of 41. The rule was verified **up to 100 frames**, or nearly 7 seconds; beyond that, the figures are extrapolations.

### From 720p to 1080p, 5 times as much energy

The same rule applies to the number of pixels. Going from 720p to 1080p (full HD) multiplies the number of pixels by 2.25 and **the energy by 5**, and doubling both height and width multiplies the energy **by 16**. The number of computation passes (the steps through which the generator refines the image) acts in proportion: dividing the number of passes by 2 divides the energy by 2. The effects of duration, resolution and passes **multiply with one another**.

## "1 kWh per video" and "4 litres of water": 2 estimates and no measurement

### "1 kWh per Sora video": an estimate made without data from OpenAI

The figure has 2 origins. In May 2025, [MIT Technology Review](https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/) published measurements made with Sasha Luccioni on an open model, CogVideoX: **944 Wh for 5 seconds** with the newer version, against 30 Wh with the older one. In November 2025, a [blog post](https://reclaimedsystems.substack.com/p/every-sora-ai-video-burns-1-kilowatt) titled "Every Sora AI video burns 1 Kilowatt hour" estimated that a 10-second video from Sora, OpenAI's generator, uses **0.94 kWh**, or 940 Wh, based on an assumption by analysts about computing time.

**Neither of these 2 values is a measurement of Sora**, whose energy use OpenAI has never published. Extrapolating the rule from the 2025 study beyond what has been verified gives **0.31 kWh at 720p** and **1.6 kWh at 1080p** for 10 seconds, and the blog post's estimate falls within that range. The service no longer exists: OpenAI [announced the shutdown](https://engadget.com/ai/openai-is-shutting-down-its-sora-video-generation-app-211023358.html) on 24 March 2026 and closed the app on **26 April 2026**, according to its [help centre](https://help.openai.com/en/articles/20001152-what-to-know-about-the-sora-discontinuation).

### How much water does an AI-generated video use? Less than half a litre for 5 seconds

**No study measures water per video.** Applying the method from our article on images, which uses [1.1 litres per kWh](https://arxiv.org/abs/2508.15734) to cool the servers and [3.1 litres](https://arxiv.org/abs/2304.03271) to produce the electricity, a 5-second video uses **0.1 litre** of cooling water and nearly **0.4 litre** in total. The November 2025 blog post puts it at a little over 4 litres for 10 seconds, using an almost identical factor. The gap comes from the energy it assumes, which is **3 times higher** than the 0.31 kWh in our extrapolation (that value would give **1.3 litres**). The scope of these factors is detailed in our article on [water consumption by AI](https://projetcelsius.com/en/blog/ai-water-footprint/).

## 1 video a day for a year: equivalent to driving 110 km

### From 15 to 450 kgCO2e a year depending on use

Generating 1 video of 5 seconds a day for a year emits **15 kgCO2e**, equivalent to driving **110 km**. At 30 videos a day, the free limit Sora offered in October 2025 ([according to Business Insider](https://www.aol.com/articles/openai-explains-plans-start-making-165332295.html)), the total would reach **450 kgCO2e** with our reference figure. That is **5.5%** of the [annual carbon footprint of a person in France](https://www.statistiques.developpement-durable.gouv.fr/lempreinte-carbone-de-la-france-de-1990-2024), or a fifth of a Paris-New York return trip. At the time, the head of the service called its economics "currently completely unsustainable". There is no reliable count of videos generated worldwide.

### 3 ways to emit less

- **Shorten the video**: dividing the duration by 2 divides the energy by 4.
- **Lower the resolution for test runs**: 480p needs 5 times less energy than 720p.
- **Prepare your prompt** (the request you give the generator): each abandoned attempt uses as much energy as the video you keep.

The first of these appears in our ranking of [actions that reduce the impact of AI](https://projetcelsius.com/en/blog/reduce-ai-environmental-impact/). A company that produces videos with AI can include them in its **purchases of digital services**, as our guide to [AI in the Bilan Carbone® (the French carbon accounting method)](https://projetcelsius.com/en/blog/ai-emissions-carbon-accounting-scope-3/) explains. Our [AI carbon footprint calculator](https://projetcelsius.com/en/tools/ai-carbon-footprint-calculator/) gives a first estimate.

## Key takeaways

- **About 41 gCO2e for 5 seconds at 720p**, equivalent to driving 290 metres, for the most downloaded open generator.
- **From 0.06 to 190 gCO2e** depending on the generator and its settings: a ratio of 1 to 3,000.
- **A video 2 times as long needs 4 times as much energy**; going from 720p to 1080p needs 5 times as much.
- **No published measurement** exists for commercial services: the "1 kWh per Sora video" is an estimate.

Our articles on [AI electricity consumption](https://projetcelsius.com/en/blog/ai-electricity-consumption/) at world scale, on choosing [a more eco-friendly AI](https://projetcelsius.com/en/blog/most-eco-friendly-ai/) and on [whether you should boycott AI](https://projetcelsius.com/en/blog/should-you-boycott-ai/) take these orders of magnitude further.

## Frequently asked questions

### How much electricity does an AI video use?

A video of a little over 5 seconds at 720p uses about 90 Wh, according to a measurement published in 2025 on the most downloaded open generator. That is nearly 5 phone charges, or 41 gCO2e with world average electricity. The median model in the same study uses 25 Wh, and the 7 models tested range from 0.14 to 415 Wh. No commercial service has published a measurement.

### Does a Sora video use 1 kWh?

There is no published measurement: OpenAI has never disclosed how much energy Sora uses, and the service was shut down in April 2026. The figure in circulation comes from a measurement made in 2025 on another generator (944 Wh for 5 seconds) and from a blog post's estimate (0.94 kWh for 10 seconds). Our extrapolation gives 0.31 to 1.6 kWh for 10 seconds, depending on the resolution.

### How much water does it take to generate an AI video?

Less than half a litre for 5 seconds, according to our calculation: about 0.1 litre to cool the servers and nearly 0.4 litre once the water used to produce the electricity is counted. No study measures water per video directly. The 4-litre figure in circulation covers 10 seconds and assumes 3 times more energy than our extrapolation.

### Which AI video generator uses the least energy?

There is no ranking of commercial services, because none has published data. Among 7 open generators tested in 2025, the least energy-intensive uses 3,000 times less energy than the most energy-intensive, but it produces shorter, lower-resolution videos. For a given generator, duration and resolution account for most of the gap.

### Is generating an AI video worse for the environment than streaming it?

Yes. Generating 5 seconds of video uses as much electricity as watching 1 hour 10 minutes of streamed video, screen and network included: 90 Wh, against 77 Wh per hour watched according to the International Energy Agency (IEA), using its 2019 figure. Counting data centres only, the gap is about 20 times larger.

## Sources

- **Measurements per video**: [Delavande, Pierrard and Luccioni, Video Killed the Energy Budget: Characterizing the Latency and Power Regimes of Open Text-to-Video Models](https://arxiv.org/abs/2509.19222), Hugging Face, 23 September 2025, tables 2 and 4, sections 5.2 and 6; [MIT Technology Review, We did the math on AI's energy footprint](https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/), 20 May 2025; [IEA, Energy and AI](https://iea.blob.core.windows.net/assets/de9dea13-b07d-42c5-a398-d1b3ae17d866/EnergyandAI.pdf), April 2025, pp. 45 and 46.
- **Conversion to CO2e and equivalents**: [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); [ADEME, Impact CO2](https://impactco2.fr/), accessed 2 October 2026; [EPA, Greenhouse Gas Equivalencies Calculator](https://www.epa.gov/energy/greenhouse-gas-equivalencies-calculator-calculations-and-references) (US Environmental Protection Agency), accessed 2 October 2026; [IEA, The carbon footprint of streaming video](https://www.iea.org/commentaries/the-carbon-footprint-of-streaming-video-fact-checking-the-headlines), December 2020; [SDES, France's carbon footprint from 1990 to 2024](https://www.statistiques.developpement-durable.gouv.fr/lempreinte-carbone-de-la-france-de-1990-2024) (SDES is the statistical service of the French Ministry of Ecological Transition). The ADEME and SDES pages are in French.
- **Image and written question**: [Luccioni, Jernite and Strubell, Power Hungry Processing](https://arxiv.org/abs/2311.16863v3), 2024; [Epoch AI, How much energy does ChatGPT use?](https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use), 7 February 2025.
- **Figures in circulation**: [Alistair Alexander, Every Sora AI video burns 1 Kilowatt hour](https://reclaimedsystems.substack.com/p/every-sora-ai-video-burns-1-kilowatt), 21 November 2025; [Google, Measuring the environmental impact of delivering AI at Google Scale](https://arxiv.org/abs/2508.15734), August 2025; [Li, Yang, Islam and Ren, Making AI Less Thirsty](https://arxiv.org/abs/2304.03271), 2023, revised version of 2025.
- **Sora**: [Business Insider, republished by AOL, 31 October 2025](https://www.aol.com/articles/openai-explains-plans-start-making-165332295.html); [OpenAI, What to know about the Sora discontinuation](https://help.openai.com/en/articles/20001152-what-to-know-about-the-sora-discontinuation), accessed 2 October 2026; [Engadget, 24 March 2026](https://engadget.com/ai/openai-is-shutting-down-its-sora-video-generation-app-211023358.html).
- **Images**: Opening photo: Mark J Sebastian, [Wikimedia Commons](https://commons.wikimedia.org/wiki/File:White_with_red_strips_clapperboard.jpg), CC BY-SA 2.0, cropped.
