--- title: "How big will AI's environmental\nimpact be by 2030, 2035 and 2050?" description: "By 2030, data centres could use almost as much electricity as Japan. Electricity, CO2, water and e-waste: the IEA and RTE scenarios, and what we know about 2050." canonical: https://projetcelsius.com/en/blog/ai-environmental-impact-2030-2050/ author: "Guillaume Pakula" datePublished: 2026-11-03 dateModified: 2026-11-03 language: en translationOf: https://projetcelsius.com/blog/impact-environnemental-ia-2030-2050/ --- # How big will AI's environmental impact be by 2030, 2035 and 2050? **Author**: Guillaume Pakula **Published**: 2026-11-03 **Section**: Explained **Canonical URL**: https://projetcelsius.com/en/blog/ai-environmental-impact-2030-2050/ **French original**: https://projetcelsius.com/blog/impact-environnemental-ia-2030-2050/ **Source**: Projet Celsius (https://projetcelsius.com/en/), a climate strategy consultancy of engineers and PhDs based in Paris and Marseille. > By 2030, the electricity used by data centres, driven by AI, would almost double to **945 TWh** according to the [International Energy Agency](https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary), **close to Japan's consumption**. In 2035, it would stand at around **1,200 TWh**, as much as Russia, and up to 1,637 in the highest scenario. The CO2 from this electricity and the water consumed would follow the same slope. For 2050, **no serious projection** covers AI: the only studies are foresight exercises on digital technology as a whole, built before generative AI. ## Key takeaways - In 2030, the world's data centres would use **almost as much electricity as Japan**. - In 2035, the gap between the IEA scenarios is **larger than France's entire electricity consumption**. - The CO2 from their electricity would rise from **180 to 320 million tonnes** by 2030. - For 2050, **no reference projection exists for AI**: only foresight studies of digital technology as a whole. In 2025, the world's data centres used **485 TWh** of electricity (1 TWh, or terawatt-hour, is 1 billion kilowatt-hours), **slightly more than the whole of France**. Artificial intelligence (AI) has become the main driver of this growth: the consumption of data centres dedicated to it jumped by **50% in a year**, according to the [April 2026 update](https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary) from the International Energy Agency (IEA). For 2030 and 2035, the agency publishes **4 scenarios**, shown here against the consumption of familiar countries. These figures cover every use of data centres, from cloud storage to video. AI is growing faster than the rest: data centres filled with GPU servers would **triple their consumption** by 2030. Beyond 2035, **no reference projection** covers AI. ## How much electricity will AI use? The IEA publishes the reference series for the world's data centres: **scenarios up to 2035**, set out in its [Energy and AI](https://www.iea.org/reports/energy-and-ai) report of April 2025 and confirmed in April 2026. ### In 2030: almost Japan In the IEA's Base Case, data centres would use **945 TWh in 2030**, almost double the 2025 figure. Japan used **1,027 TWh** in 2025, according to the [annual data from Ember](https://ember-energy.org/data/yearly-electricity-data/), an energy think tank from which we take our country benchmarks, except for France. The 3 other scenarios frame the Base Case **from 833 to 1,008 TWh**. These 945 TWh would be **about 3%** of global electricity, against 1.5% in 2024. ### In 2035: as much as Russia, or far more The agency extends its scenarios to 2035, and itself describes them as **exploratory**. The Base Case reaches **1,193 TWh**, roughly **Russia's consumption** (1,178 TWh in 2025). The highest scenario, known as Lift-Off, climbs to **1,637 TWh**, between Russia and India. The lowest stops at 942 TWh, the level the Base Case already reaches in 2030. ### And how much of that total is AI alone? No meter separates AI from other uses. The IEA tracks the data centres it calls AI-focused, which line up GPU servers: they would reach **465 TWh in 2030**, half the total, and on their own **as much as France** (451 TWh in 2025 according to RTE, the French electricity transmission system operator, the benchmark used in all our articles). Other studies go higher. The French collective [GreenIT](https://greenit.eco/wp-content/uploads/2026/10/2026-10-02-greenit-etude-ia-synthese-v2.pdf) projects **1,169 TWh** for AI alone, because it assumes servers running at 69% of their capacity all year round. Our article on [AI electricity consumption](https://projetcelsius.com/en/blog/ai-electricity-consumption/) compares these studies one by one, and the [chart of the 3 projections for 2030 (in French)](https://projetcelsius.com/ia/graphiques/centres-de-donnees-2030-trois-projections/) (in French) puts them side by side. The hub note ["AI as much as Japan" (in French)](https://projetcelsius.com/ia/analyses/l-ia-autant-que-le-japon/) (in French) explains what the comparison measures. ## Why do AI energy projections vary so much? In 2035, **695 TWh** separate the IEA's highest scenario from its lowest, **far more than all of France's electricity**. The gap comes down to 4 questions nobody can yet answer, and each scenario answers them in its own way. ### How many people will use AI, and for what? In a year, the leading model developers saw their active users **triple** and their revenue **increase fivefold**, according to the IEA. Uses are also getting heavier: agents, programs that chain dozens of steps on their own, models that reason before answering and [AI-generated video](https://projetcelsius.com/en/blog/ai-video-environmental-impact/) need **far more than a written question**. The Lift-Off scenario assumes stronger adoption and constraints lifted. ### Will models become more efficient? The High Efficiency scenario assumes the same demand, served by better-sized models and by more computing done on phones and computers themselves. It puts consumption at **1,013 TWh in 2035**, 180 less than the Base Case. The gains are real: Google cut the energy of a median Gemini prompt **33-fold in a year**, according to its [2026 environmental report](https://sustainability.google/files/google-2026-environmental-report.pdf). Growing use quickly absorbs them, as our article on [the energy of a ChatGPT query](https://projetcelsius.com/en/blog/chatgpt-query-energy-consumption/) shows. ### Will the electricity grid keep up? In 2025, the IEA estimated that around **20% of data centre projects** risked delays from an electricity grid already under strain. Lead times for transformers and cables had **doubled in 3 years**, and gas turbines had to be ordered several years ahead. In 2026, the agency adds a **shortage of memory chips** for AI, expected to last at least until the end of 2027. ### Will the money keep flowing? Data centres now cost too much to be paid for from companies' own funds alone: they depend on **financial markets**, and therefore on confidence in AI's profitability. The Headwinds scenario assumes AI that struggles to pay for itself, high interest rates and projects blocked locally. It stops at **942 TWh in 2035**. At the other end, the French think tank [The Shift Project](https://theshiftproject.org/publications/intelligence-artificielle-centres-de-donnees-rapport-final/), which extends announced capacity, sees **up to 1,500 TWh** as early as 2030, cryptocurrencies included. ## How much CO2, water and waste by 2030? What electricity emits depends on the power plants that generate it. Data centres also consume water and replace their hardware quickly. For a single building, see our article on [the carbon footprint of a data centre](https://projetcelsius.com/en/blog/data-center-carbon-footprint/). ### CO2: as much as the United Kingdom in 2030 The electricity used by data centres emitted **180 million tonnes of CO2** in 2024, according to the IEA: two thirds of the CO2 emitted by France that year (264 million, according to the [Global Carbon Project](https://ourworldindata.org/co2-emissions)). In the Base Case, these emissions would reach **320 million tonnes in 2030**, **as much as the United Kingdom** (313 million in 2024). The 2026 update sees them rising further, towards 350 million in 2035. Gas would supply **30% of this electricity** in 2030, coal 20%. These figures count only the CO2 from power plants. GreenIT adds server manufacturing and the other greenhouse gases, for AI alone: **636 million tonnes of CO2 equivalent** in 2030, if electricity remains as carbon-intensive as in 2025. That is **far more than all of France's greenhouse gas emissions** (367 million tonnes of CO2 equivalent in 2024, according to [Citepa](https://www.citepa.org/fr/secten/), the body that compiles France's emissions inventory). Our article on [the environmental impact of AI](https://projetcelsius.com/en/blog/environmental-impact-of-ai/) details the method, and [the one on AI in a Bilan Carbone®](https://projetcelsius.com/en/blog/chatgpt-carbon-footprint/) (the French carbon accounting method) brings these totals down to the scale of an SME. ### Water: more than double between 2023 and 2030 Data centres consume water to cool themselves, and indirectly to produce their electricity and their chips. The IEA estimates this consumed water, meaning water evaporated and not returned to rivers, at **560 billion litres** in 2023 and **1,200 billion in 2030**. Two thirds came from **energy production**. In the United States, a large 100-megawatt data centre consumes on average **2 million litres a day**, including the water behind its electricity, as much as 6,500 households. Local conditions matter more than the total: **the same litre** weighs little in Scandinavia and a lot in Arizona. Our article on [the water footprint of AI](https://projetcelsius.com/en/blog/ai-water-footprint/) goes from the drop per question to the whole site. ### E-waste: up to 5 million tonnes A study published in [Nature Computational Science](https://www.nature.com/articles/s43588-024-00712-6) in October 2024 estimates that generative AI could produce **1.2 to 5 million tonnes** of cumulative e-waste between 2020 and 2030, depending on how fast it develops. That is small against the **62 million tonnes** the world discarded in 2022, all devices combined, according to the [Global E-waste Monitor 2024](https://ewastemonitor.info/the-global-e-waste-monitor-2024/). This flow would grow faster if servers were **replaced more often** or if restrictions on chip imports tightened. According to the authors, reusing, repairing and recycling data centre equipment would reduce this waste **by 16 to 86%**. Our article on [why AI is bad for the environment](https://projetcelsius.com/en/blog/why-ai-is-bad-for-the-environment/) shows the weight of manufacturing. ## And in France, what is forecast for 2035? In France, the projections come from RTE, the electricity transmission system operator, which tracks grid connection requests from data centres. They cover the data centres that host AI, not AI alone. ### From 4 or 5 TWh in 2024 to more than 30 in 2035 According to its [Bilan prévisionnel 2025](https://assets.rte-france.com/prod/public/2025-12/2025-12-23-bilan-previsionnel-2025-consommation.pdf) (long-term outlook, in French), data centres located on dedicated sites used **4 to 5 TWh in 2024**, less than Luxembourg. RTE retains 2 trajectories, depending on the share of projects that go ahead: **10 to 15 TWh in 2030**, then **21 to 32 TWh in 2035**. That is between the consumption of Croatia and that of Ireland. A third trajectory considers **stagnation**, if investment moves to other European countries, if municipalities impose a moratorium or if a financial bubble bursts. The 32 TWh of the high trajectory would be about **7% of the electricity** consumed in France in 2025 (451 TWh according to RTE's [Bilan électrique](https://assets.rte-france.com/prod/public/2026-02/Bilan-electrique-2025-principaux-resultats.pdf), its annual electricity report, in French). The [hub chart on France (in French)](https://projetcelsius.com/ia/graphiques/centres-de-donnees-france-2035/) (in French) shows the 2 trajectories. ### Electricity almost 9 times less carbon-intensive A kilowatt-hour consumed in France emits **51.9 grams of CO2 equivalent**, including power plant manufacturing and losses, according to the [Base Carbone](https://data.ademe.fr/datasets/base-carboner) of ADEME (the French Agency for Ecological Transition). Worldwide, a kilowatt-hour generated emits 458 on average, according to Ember. Even at 32 TWh, the data centres located in France would emit in the order of **1.7 million tonnes of CO2 equivalent** a year, by our calculation. That is **less than 0.5%** of the 367 million emitted by the country in 2024. French users also rely on data centres located elsewhere: **53% of their digital uses** were hosted abroad in 2022, according to an [ADEME study](https://ecoresponsable.numerique.gouv.fr/docs/2024/etude-ademe-impacts-environnementaux-numerique.pdf) published in 2025. Counting them, the agency estimates that AI accounted for **13% of the electricity** used by the data centres serving France in 2025, according to a 2026 study cited by [Arcep](https://www.arcep.fr/uploads/tx_gspublication/rapport-IA-et-environnement_210526.pdf) (the French electronic communications regulator). Our guide to [the digital carbon footprint of a company](https://projetcelsius.com/en/blog/corporate-digital-carbon-footprint/) shows how to count these uses hosted elsewhere. ## Can we know what AI will weigh in 2050? No institution publishes a projection specific to AI beyond 2035, and most stop at 2030. For 2050, the only studies are foresight exercises on digital technology as a whole. ### What exists: a foresight study of digital technology The reference foresight study on French digital technology to 2050 is the one by ADEME and Arcep, published in [March 2023](https://www.arcep.fr/uploads/tx_gspublication/note-synthese-Arcep-gouvernement-prospective-2030-2050_mars2023.pdf) (in French). It covers all digital technology: phones, screens, networks and data centres. With no change, its carbon footprint in France would rise from **17.2 million tonnes of CO2 equivalent in 2020 to 49 million in 2050**, almost 3 times more. Depending on the model of society chosen, it would range from 9.3 million if sufficiency prevails to **81 million** if technology comes first. This study was built **before the rise of generative AI**, as Arcep pointed out in May 2026. Its scope was also narrower: ADEME's update, which counts data centres located abroad, already finds **29.5 million tonnes for 2022**. France nonetheless plans to stabilise the footprint of digital technology by 2030 to 2035, in its [national low-carbon strategy](https://projetcelsius.com/en/blog/snbc-3-business-obligations/). Our article on [digital pollution statistics](https://projetcelsius.com/en/blog/digital-pollution-statistics/) details these trajectories. ### What cannot be said Nobody knows how many times AI will be used in 2050, how much electricity each use will need, or where that electricity will come from. The IEA itself considers its 2035 figures exploratory, and notes that growth could **surprise on the upside** after 2030. A figure for AI in 2050 circulating without a scenario or a source would be speculation. Everything will depend on **3 things**: the volume of use, the electricity per use and the carbon intensity of that electricity. > **What about the emissions AI could help avoid?** According to the IEA, existing AI applications, if widely adopted, could reduce energy-related emissions by around 5% in 2035. That is far more than the emissions of data centres, and far less than the climate requires. At Projet Celsius, we watch **the gap between the 2035 scenarios** instead: year after year, it will show which way the trajectory is leaning. The signals appear every spring in the IEA and RTE updates: **chip deliveries, signed grid connections**, and the share of gas in the power plants supplying new sites. For an organisation, the useful question is closer to home: **how much its own use weighs**, which is growing faster than models are becoming efficient. That is the subject of our article on [the Bilan Carbone® of AI in a company](https://projetcelsius.com/en/blog/ai-carbon-footprint-in-your-company/), and our [Base IA (in French)](https://projetcelsius.com/ia/base/) (in French) publishes the figures needed to calculate it. The steps that reduce it are in our guide to [reducing the environmental impact of AI](https://projetcelsius.com/en/blog/reduce-ai-environmental-impact/). ## Key takeaways - The world's data centres used 485 TWh in 2025, slightly more than France. The IEA projects 945 in 2030, close to Japan's consumption. - In 2035, the IEA scenarios range from 942 to 1,637 TWh, a gap larger than France's consumption. - The CO2 from their electricity would rise from 180 to 320 million tonnes in 2030, and the water consumed from 560 to 1,200 billion litres. - In France, RTE projects 21 to 32 TWh in 2035 for data centres, with electricity almost 9 times less carbon-intensive than the global average. - For 2050, no projection covers AI: the ADEME and Arcep foresight study, which triples the footprint of digital technology, predates generative AI. To find out which AI uses the least, see our [comparison of ChatGPT, Claude and Gemini](https://projetcelsius.com/en/blog/chatgpt-vs-claude-vs-gemini-environmental-impact/); for one person's use, our article [should you boycott AI?](https://projetcelsius.com/en/blog/should-you-boycott-ai/). ## Frequently asked questions ## Frequently asked questions ### How much electricity will AI use in 2030? In the IEA's Base Case (April 2026), the world's data centres would use 945 TWh in 2030, against 485 TWh in 2025, or about 3% of global electricity. AI-focused data centres would account for about half of it, 465 TWh. The agency's other scenarios range from 833 to 1,008 TWh. ### Will AI use as much electricity as Japan? Almost, but the comparison covers all data centres, not AI alone. The 945 TWh the IEA projects for them in 2030 approach the 1,027 TWh used by Japan in 2025 (Ember). AI-focused data centres, at 465 TWh, would use roughly as much as France (451 TWh in 2025 according to RTE). ### How much CO2 will AI emit in 2030? The IEA projects about 320 million tonnes of CO2 in 2030 for the electricity of all data centres, against 180 million in 2024, then about 350 million in 2035. GreenIT, which counts server manufacturing and the other greenhouse gases for AI alone, reaches 636 million tonnes of CO2 equivalent, with electricity unchanged. ### How much water will data centres use in 2030? About 1,200 billion litres a year according to the IEA, more than double the 560 billion estimated for 2023. This is consumed water, not returned to rivers, two thirds of it linked to energy production in 2023. In the United States, a large 100-megawatt data centre consumes on average 2 million litres a day, as much as 6,500 households. ### How much electricity will data centres in France use in 2035? According to RTE's Bilan prévisionnel 2025, data centres on dedicated sites would use 21 to 32 TWh in 2035, against 4 to 5 TWh in 2024, depending on the share of projects that go ahead. The high trajectory would be about 7% of French electricity. RTE also considers stagnation. ### What does ADEME say about the impact of digital technology in 2050? The ADEME and Arcep foresight study (March 2023) projects, with no change, a digital carbon footprint of 49 million tonnes of CO2 equivalent in 2050 in France, almost 3 times that of 2020. Depending on the scenario, it would range from 9.3 to 81 million. This study was built before the rise of generative AI. ## Sources - International Energy Agency, [Energy and AI](https://www.iea.org/reports/energy-and-ai), April 2025: data centre consumption in 2024, CO2 emissions (chapter 2), water (box 5.4). - International Energy Agency, [Key Questions on Energy and AI](https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary), 16 April 2026: 2030 and 2035 scenarios (Table A.1), AI-focused data centres (figure 1.5), emissions in 2035. - Ember, [yearly electricity data](https://ember-energy.org/data/yearly-electricity-data/), electricity demand by country in 2025 (Japan, Russia, United Kingdom, Croatia, Ireland, Luxembourg, India), CC BY 4.0 licence. For France, the benchmark is RTE's, as in our other articles. - Global Carbon Project, via Our World in Data, [CO2 emissions by country](https://ourworldindata.org/co2-emissions), 2024. - Wang et al., ["E-waste challenges of generative artificial intelligence"](https://www.nature.com/articles/s43588-024-00712-6), Nature Computational Science, 28 October 2024. - ITU and UNITAR, [Global E-waste Monitor 2024](https://ewastemonitor.info/the-global-e-waste-monitor-2024/), March 2024. - RTE, [Bilan prévisionnel 2025-2035, consommation](https://assets.rte-france.com/prod/public/2025-12/2025-12-23-bilan-previsionnel-2025-consommation.pdf) (in French), December 2025, section 2.5; [Bilan électrique 2025](https://assets.rte-france.com/prod/public/2026-02/Bilan-electrique-2025-principaux-resultats.pdf) (in French), February 2026 (451 TWh consumed in 2025, consumption adjusted for weather variations). - ADEME and Arcep, [Évaluation de l'impact environnemental du numérique en France, analyse prospective à 2030 et 2050](https://www.arcep.fr/uploads/tx_gspublication/note-synthese-Arcep-gouvernement-prospective-2030-2050_mars2023.pdf) (in French), summary note, March 2023. - ADEME, [Évaluation de l'impact environnemental du numérique en France](https://ecoresponsable.numerique.gouv.fr/docs/2024/etude-ademe-impacts-environnementaux-numerique.pdf) (in French), January 2025. - Arcep, [Intelligence artificielle générative : quels défis environnementaux ?](https://www.arcep.fr/uploads/tx_gspublication/rapport-IA-et-environnement_210526.pdf) (in French), May 2026. - ADEME, [Base Carbone](https://data.ademe.fr/datasets/base-carboner), electricity 2024, average consumption mix, mainland France. - Citepa, [Secten 2026 report](https://www.citepa.org/fr/secten/) (in French), France's emissions in 2024. - GreenIT, [Impacts environnementaux et sanitaires de l'IA dans le monde en 2025 et 2030](https://greenit.eco/wp-content/uploads/2026/10/2026-10-02-greenit-etude-ia-synthese-v2.pdf) (in French), summary, September 2026. - The Shift Project, [Intelligence artificielle, données, calculs : quelles infrastructures dans un monde décarboné ?](https://theshiftproject.org/publications/intelligence-artificielle-centres-de-donnees-rapport-final/) (in French), October 2025. - Google, [Environmental Report 2026](https://sustainability.google/files/google-2026-environmental-report.pdf), June 2026, p. 17.