# The Environmental Impact of Artificial Intelligence > AI uses a lot of electricity, water and hardware, but it can also help the climate and nature. This article gives the 2026 figures on the environmental impact of AI: how much energy data centers and a single prompt use, how much water AI needs, where AI cuts emissions and how the sector is becoming more sustainable. > URL: https://datanorth.ai/blog/the-environmental-impact-of-artificial-intelligence > Last modified: 2026-10-09T09:05:21+00:00 AI affects the environment in two directions. Training and running AI models takes electricity, water and hardware, and the data centers behind them are one of the fastest-growing sources of electricity demand in the world. At the same time, AI helps organizations cut emissions, protect nature and use energy and materials more efficiently. Whether AI is good or bad for the environment depends on how much we use it, what we use it for and how clean the power behind it is. This article gives the current figures for both sides, updated in October 2026. ## Key Learnings These are the main points of this update. - Data centers used about 485 TWh of electricity in 2025, according to the IEA, and are expected to reach 950 TWh by 2030, around 3% of global electricity demand. - A short text prompt uses little energy, about 0.24 to 0.34 Wh according to Google and OpenAI’s CEO. Long reasoning tasks, images and especially video use far more. - GPT-3 used about a 500 ml bottle of water for every 10 to 50 medium-length responses, according to a peer-reviewed study. - AI also has a positive impact, from warning about illegal deforestation months in advance to weather forecasts that need about 1,000 times less energy than traditional models. - Big tech’s footprint grows with AI: Google’s electricity use rose 37% in 2025 and Microsoft’s total emissions rose 25% in a year. ## The rapid rise in popularity of AI AI has moved from pilot projects to everyday use in a few years. According to Eurostat, 20% of EU enterprises with 10 or more employees used AI in 2025, up from 13.5% a year earlier. OpenAI says ChatGPT has more than 1 billion weekly active users. Every request those users send runs on servers in a data center, which is why the environmental impact of AI has become a mainstream question. ## Is AI bad for the environment? AI is not bad for the environment by definition, but its footprint is real and growing fast. The negative impact comes from the electricity, water and hardware that data centers need. The [International Energy Agency (IEA)](https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary) reports that electricity demand from data centers grew 17% in 2025, and from AI-focused data centers 50%. Data center emissions are projected to reach around 350 million tons of CO2 a year by 2035, about 2% of global electricity-sector emissions. That is still a small share, and AI can help reduce emissions in other sectors. The concern is the speed of the growth and the concentration of new data centers in a few regions, where they put pressure on power grids and water supplies. So the answer depends on the use: AI that cuts energy waste in a building or delivery network can save more than it consumes, while generating large volumes of images and video mainly adds to the footprint. *AI uses electricity, water and hardware. Whether it helps or harms on balance depends on how it is used. ## The environmental impact of AI AI affects the climate both negatively and positively. The goal is to keep the negative impact small and get the most out of the positive effects. We start with the negative side. ### Negative impact: Energy Consumption & CO2 Emission The negative impact of AI is mostly linked to data centers, the large buildings full of servers that train and run AI models. According to the US Department of Energy, data centers use 10 to 50 times the energy per floor space of a typical office building. The exact carbon footprint of AI is still hard to pin down, because companies report in different ways and there is no single standard for measuring it. These statistics give the best indication in 2026: - Globally, data centers used about 485 TWh of electricity in 2025, and the IEA expects 950 TWh in 2030. - In the United States, data centers used 4.7% of all electricity in 2024, which could rise to 11.8% by 2030, according to Lawrence Berkeley National Laboratory. - In the EU, data centers used 68 TWh in 2024, and the European Commission expects 114 TWh by 2030. - Meta reports 8,930 tons of CO2 equivalent for training Llama 3.1 405B, based on the local electricity mix. That equals the yearly emissions of about 1,900 gasoline cars, at the 4.6 tons per car the US EPA uses. Training a foundational model from scratch is done by hyperscalers which always costs more energy, when fine-tuning or running inference, the energy consumption is a fraction of that. - Researchers writing in [Nature Computational Science](https://www.nature.com/articles/s43588-024-00712-6) estimate that generative AI could add 1.2 to 5 million tons of e-waste in total between 2020 and 2030. ### How much energy does one AI prompt use? One short text prompt uses about as much energy as watching TV for 9 to 13 seconds. The impact comes from scale and from the type of task. TaskEstimated energy per taskSourceMedian Gemini text prompt0.24 WhGoogle, 2025Average ChatGPT query0.34 WhSam Altman (OpenAI), 2025Long reasoning query on a large modelabout 3.9 WhMicrosoft Research, 2025One response from Llama 3.1 405Babout 1.9 WhMIT Technology Review, 2025One image (Stable Diffusion 3 Medium)about 0.6 WhMIT Technology Review, 2025One 5-second AI videoabout 940 WhMIT Technology Review, 2025 The Google and OpenAI figures are self-reported, and all of these figures use different methods, so read them as orders of magnitude. Efficiency is improving fast: Google says the energy per median Gemini prompt fell by a factor of 33 in one year. Microsoft has not published a per-prompt figure for Copilot; its researchers estimate 0.31 Wh per query for large models in general. The popular claim that one AI image costs a full phone charge overstates the 2023 test behind it, in which the least efficient model used about half a charge per image. The Llama, image and video figures come from measurements by MIT Technology Review on open models, and video stands out by far. ### Negative impact: Water Consumption Data centers use water to cool their servers, and power plants use water to generate the electricity. Researchers from UC Riverside and UT Arlington estimate that GPT-3 needed about 500 ml of water for every 10 to 50 medium-length responses, and that training GPT-3 in Microsoft’s US data centers could consume 5.4 million liters of water, 700,000 liters of it on site ([Li et al.](https://arxiv.org/abs/2304.03271)). Newer figures depend heavily on what is counted. Google reports 0.26 ml, about five drops, for a median Gemini prompt. Mistral AI, which had its life-cycle analysis peer-reviewed, reports 45 ml of water and 1.14 grams of CO2 equivalent for a 400-token answer from Le Chat. To understand AI’s footprint better, we need more data from the companies that build the models. Regulation is starting to require it. Under the [EU AI Act](https://datanorth.ai/blog/eu-ai-act-august-2026-the-step-by-step-compliance-checklist-for-enterprises), providers of general-purpose AI models have had to document the known or estimated energy consumption of their models since August 2, 2025. EU data centers of 500 kW and more report their energy and water use every year, and in September 2026 the European Commission adopted an EU rating scheme for data centers. Public, comparable figures per model are still missing. ### Positive impact: Green AI Green AI is a movement within AI research. A 2019 paper by researchers at the Allen Institute for AI popularized the term for AI research that delivers new results without increasing computing costs, ideally while reducing them. A recent example is Hugging Face’s AI Energy Score, which rates AI models from one to five stars, like the energy label on a household appliance. AI is also good at processing large amounts of satellite and weather data, which helps to monitor climate change and predict natural disasters. Since February 2025, the European weather center [ECMWF](https://www.ecmwf.int/en/about/media-centre/news/2025/ecmwfs-ai-forecasts-become-operational) has run an AI forecasting system that uses about 1,000 times less energy per forecast than its traditional model. NASA and IBM have released Prithvi WxC, an open-source foundation model for weather and climate applications trained on 40 years of NASA data. ### Positive Impact: Preserve Nature AI can help against climate change, especially by protecting nature. - **Combat deforestation:** WWF’s [Forest Foresight](https://www.worldwildlife.org/news/magazine/winter-2023/could-ai-help-stop-deforestation-before-it-starts/) analyzes satellite images to warn about illegal deforestation up to six months in advance, with 80% accuracy according to WWF in 2023. It detects signals such as new roads for logging machinery and is now used by governments in six countries. - **Fight wildfires:** In California, the ALERTCalifornia network of more than 1,050 cameras uses AI to spot fires. In its first two months it identified 77 fires before any 911 call came in. The FireSat satellites, the first three of which were deployed in July 2026, can detect fires as small as five by five meters. - **Track biodiversity:** Scientists combine AI with camera traps, drones and satellite images to follow animal populations. Google’s open-source model SpeciesNet recognizes nearly 2,500 categories of mammals, birds and reptiles in camera trap photos. - **Prevent overfishing:** The Environmental Defense Fund’s Smart Boat Initiative uses AI cameras to recognize fish species and sizes. Using AI and satellite imagery, Global Fishing Watch found that about 75% of industrial fishing vessels are not publicly tracked. That insight helps authorities find illegal fishing. ### Positive Impact: Circular Economy A circular economy keeps products, components and materials in use for as long as possible through durability, reuse, remanufacturing and recycling. AI helps with this. **Identifying and recovering recyclables:** AI-guided sorting robots recognize recyclable materials faster and more consistently than people. [AMP](https://ampsortation.com/articles/global-cleantech-100-2026) reports that its AI has identified more than 200 billion items and its systems have processed 2.8 million tons of recyclables, with robots that sort up to four times faster than manual sorting. **Monitor water quality:** AI makes it easier to monitor water quality and spot contamination early, so water companies can act before pollution spreads. ### Positive Impact: Optimization of Operations AI consumes a lot of energy, but it can also save energy. The [IEA](https://www.iea.org/reports/energy-and-ai/ai-and-climate-change) calculates that wider use of existing AI applications in industry, transport and buildings could cut 1,400 million tons of CO2 emissions in 2035. It adds that there is currently no momentum to make that happen, so the real effect could stay marginal. **Agriculture:** By combining data on weather, soil and crops, AI helps farmers use less water, fertilizer and crop protection. Camera-based sprayers, for example, recognize weeds and spray only where needed. **Transportation:** AI can plan efficient routes and reduce congestion. Google’s Project Green Light advises cities on traffic light timing, with early results showing up to 30% fewer stops and 10% lower emissions at the intersections involved. **Waste Management:** AI reduces food waste with dynamic pricing on products close to their expiry date. Dutch supermarket chain Albert Heijn uses AI to predict daily sales per product and gives discounts of up to 70% as the expiry date approaches. AI can also improve [inventory management](https://datanorth.ai/use-case/inventory-management) and [predict maintenance](https://datanorth.ai/use-case/predictive-maintenance), which prevents waste and costly repairs. ## How AI becomes more sustainable More efficient models, hardware and data centers reduce the energy per task, although total demand still rises faster than efficiency improves. For the technical options, read our article on [the environmental impact of LLMs and how to reduce it](https://datanorth.ai/blog/future-proofing-llms-environmental-impact). ### Green Data Centers Green data centers run on renewable energy and use smart cooling to reduce energy use, emissions and costs. Back in 2016, DeepMind’s machine learning system cut the energy used to cool Google’s data centers by 40%. Even so, large tech companies struggle to meet their climate goals as AI grows. According to its [2026 environmental report](https://blog.google/company-news/outreach-and-initiatives/sustainability/2026-environmental-report/), Google’s electricity demand grew 37% in 2025. That year its operational emissions fell 2%, but its supply chain emissions grew 25%. Google no longer claims to be carbon-neutral: it stopped maintaining operational carbon neutrality in 2023, and it calls its goal of net zero and 24/7 carbon-free energy by 2030 a “moonshot”. Microsoft reports that its total emissions increased 25% year over year, while it keeps its goal of becoming carbon negative by 2030. ### Reuse of water and heat The water use of data centers is a problem too. Closed-loop cooling reuses the same water instead of evaporating it, and the heat data centers produce can be reused. Engineering firm Ramboll designed a heat pump installation that recovers 215,000 MWh of surplus heat from Meta’s data center in Odense, Denmark, for the district heating of more than 12,000 homes. According to the European Commission, reusing about half of the waste heat from European data centers would cover the heating demand of 4 million households. ### Underperforming AI models By stopping underperforming models early during a search for the best model settings, MIT Lincoln Laboratory cut the energy used for that training by 80%. Capping the power of GPUs, the chips that do most AI computing, cut energy use by 12% to 15%, while tasks took only about 3% longer. Choosing a smaller model also helps: for many business tasks a [small language model](https://datanorth.ai/blog/small-language-models-vs-llm) is accurate enough and needs a fraction of the computing power. ## What does DataNorth do to reduce its environmental impact? At DataNorth we also invest in sustainability. Most importantly, we plant trees through Trees for All to compensate for our emissions, and in 2025 we compensated for 125 tons of CO2. Our clients can do the same through our [CO2 compensation service for AI](https://datanorth.ai/service/co2-compensation-program-for-ai). Do you want to know what kind of positive impact AI can have for your business? [Contact us](https://datanorth.ai/contact) and we will look at it together. ## Frequently Asked Questions (FAQ) **Is AI bad for the environment?** AI can help reduce emissions in energy, transport and agriculture. However, it also has a real environmental impact: data centers require substantial electricity, water and hardware, and their electricity demand grew by 17% in 2025. Ultimately, AI’s net environmental effect depends on what it is used for and how clean the electricity powering it is. **How does AI affect the environment positively?** AI helps the environment by processing large amounts of data quickly. It warns about illegal deforestation up to six months in advance, spots wildfires early, tracks illegal fishing with satellite images, sorts recyclables up to four times faster than people and makes weather forecasts with about 1,000 times less energy than a traditional forecasting model. **How much energy does one AI prompt use?** A short text prompt uses about 0.24 to 0.34 Wh, according to Google and OpenAI CEO Sam Altman, roughly what a TV uses in 9 to 13 seconds. A long reasoning task can use around 3.9 Wh, an image about 0.6 Wh and a 5-second AI video about 940 Wh. **How much water does AI use?** AI’s water use depends on what is counted. A peer-reviewed study estimates that GPT-3 used about 500 ml of water for every 10 to 50 medium-length responses. Google reports 0.26 ml for a median Gemini prompt, and Mistral AI reports 45 ml for a 400-token answer from Le Chat. **How does AI training compare to car emissions?** Training a large model today can equal the yearly emissions of more than a thousand cars. Meta reports 8,930 tons of CO2 equivalent for training Llama 3.1 405B, based on the local electricity mix, about the yearly emissions of 1,900 gasoline cars. A widely quoted 2019 study from the University of Massachusetts Amherst found that one extreme experiment, a full neural architecture search, emitted about 284 tons of CO2 equivalent, as much as five cars over their lifetime. **How can you use AI more sustainably?** You can use AI more sustainably by choosing the smallest model that does the job, avoiding images and video you do not need, and keeping prompts and answers focused. Organizations can also choose providers or cloud regions that run on renewable energy and run suitable tasks on local models. *Sign up for our NewsletterGet updates about our latest AI related blogs, research, services and much more! Newsletter Sign Up Form (Blog) (EN)EmailRadio Field I would like to receive marketing-related emails of DataNorth.Subscribe!