AI's power demand and emissions, with figures from the IEA and company reports, set beside the climate uses with measured results in forecasting and grids.
Yes. AI adds emissions through the electricity its data centres draw and through the chips, servers and buildings it needs. The IEA's April 2026 report, Key Questions on Energy and AI, finds that global data centre electricity demand grew by 17% in 2025, to 485 TWh. Consumption by AI-focused data centres rose by 50% in the same year.
The IEA puts data centres at 1.5% of global electricity consumption in 2025. In the states, metropolitan areas and regions where they cluster, they account for as much as 20 to 30% of electricity demand, which is where grid operators feel the pressure first.
The agency's central case has data centre consumption roughly doubling to 950 TWh by 2030, around 3% of global electricity demand. In its projections, the emissions associated with data centres double to around 350 million tonnes in 2035, still about 2% of global electricity sector emissions at that date.
How much carbon that electricity carries depends on how it is bought. The tech sector signed around 40% of all corporate renewable power purchase agreements worldwide in 2025, the IEA reports. The same report expects 15 to 27 GW of onsite natural gas generation to be powering data centres by 2030, mostly in the United States, where developers are working around slow grid connections.
Two sustainability reports published in 2026 show how this growth lands on individual companies.
Google reports total ambition-based emissions of about 14.5 million tonnes of CO2 equivalent for 2025, up 18% on 2024 and 81% on its 2019 base year. Its total electricity consumption grew by 37% in 2025. Operational emissions fell 2% over the year, which Google credits to certificates from its long-term clean energy contracts and others bought on the market, and it attributes the overall rise mainly to supply chain activity behind its expansion.
Microsoft reports that its total emissions across Scopes 1, 2 and 3 rose 25% year over year in fiscal 2025. It attributes the rise mainly to the expansion of its data centre infrastructure and to its decision to stop using unbundled renewable energy certificates. Scope 2 emissions, from purchased energy, went from nearly 2% of its total to 13%.
A company buying compute inherits both kinds of emissions: the electricity its workloads draw, and a share of the emissions from making the hardware they run on. The choice of accelerator affects both, and our guide to GPUs and TPUs explains how the main chip families differ.
The IEA's answer starts with efficiency: "Measured per individual task, the energy efficiency of AI is improving at a rate unprecedented in energy history." Energy use per AI task has fallen by at least an order of magnitude a year recently, and a simple text query now typically uses less electricity than running a television for the same length of time. If every conventional internet search were run as a simple AI text query, the IEA puts the total at under 4 TWh a year, less than 1% of today's data centre consumption.
The growth in demand comes from heavier uses. Video generation, reasoning and agentic tasks can consume hundreds or thousands of times more energy per query than simple text generation. Agentic systems, which plan and carry out multi-step work (see AI agents vs agentic AI), are among the uses the IEA names. The agency calls AI's energy demand "the result of three rapidly evolving and uncertain trends: improvements in efficiency, surging uptake, and changing model capabilities".
Weather forecasting offers some of the best documented results. In 2023, researchers at Google DeepMind and Google Research described GraphCast in Science. Trained on historical reanalysis data, it predicts hundreds of weather variables for the next 10 days at 0.25 degree resolution across the globe in under a minute. The paper reports that it significantly outperformed the most accurate operational deterministic systems on 90% of 1,380 verification targets, and that it improved prediction of severe events, including tropical cyclone tracks, atmospheric rivers and extreme temperatures.
The European Centre for Medium-Range Weather Forecasts (ECMWF) took its Artificial Intelligence Forecasting System (AIFS) into operations on 25 February 2025, alongside its physics-based model. ECMWF reports gains of up to 20% on tropical cyclone tracks and "a reduction of approximately 1,000 times in energy use for making a forecast". An ensemble version, which runs 51 forecasts to show the range of possible outcomes, became operational on 1 July 2025.
ECMWF also sets out the limits. The AI ensemble runs at 31 km resolution, against 9 km for the physics-based ensemble, which it says "remains indispensable for high-resolution fields and coupled Earth-system processes". It is exploring hybrid systems, and its AI forecasts start from initial conditions produced by physics-based data assimilation.
ECMWF names surface solar radiation and wind speeds at turbine height among the quantities AIFS will help predict for the renewable energy sector. The IEA reports that AI-enhanced weather forecasting is already helping grids integrate variable renewables and reduce curtailment.
Among the deployed examples in the IEA's 2026 report, a forecasting model in Chile up to 15% more accurate than previous industry standards has led to more effective grid management, more wind power on the grid and less curtailment. In Europe, AI manages heating, ventilation and air conditioning across more than 1.5 million devices in 250,000 buildings in eight countries, responding to real-time market conditions; the IEA credits the approach with 2 GW of demand response.
Across the economy, the IEA estimates that well-documented AI use cases could save over 13 exajoules of energy by 2035, equal to 3% of global final energy consumption, if barriers to wider uptake are overcome. An IEA survey of energy companies found the lack of digital skills to be the single largest of those barriers, and only 10% of global electricity consumption is covered by open electricity data policies.
For adaptation, Google reports that its flood forecasting information covers more than two billion people in around 150 countries for the most significant riverine flood events. The 2022 review Tackling Climate Change with Machine Learning, by David Rolnick and colleagues in ACM Computing Surveys, maps a much wider field, from smart grids to disaster management, and its authors stress that machine learning is no silver bullet and that technology alone is not enough.
Not with the data available today. The two IEA figures above use different units and rest on different conditions: around 350 million tonnes of data centre emissions in 2035 in its projections, and potential energy savings of 13 EJ by 2035 that depend on wider adoption. On adoption, the IEA's own summary heading reads: "the energy sector is not yet taking full advantage of AI opportunities".
The IEA also observes that specialist scientific models tend to be much smaller and less computationally intensive than frontier generative AI models. In public debate the word AI covers both a weather model and a frontier chatbot, which draw on very different amounts of compute.
The IPCC's 2022 mitigation report frames the question this way. It found, with high confidence, that sensors, robotics and AI can improve energy management and efficiency in all sectors, and that some of these gains can be reduced or counterbalanced by growth in demand for goods and services. Its summary for policymakers puts it in one line: "Digital technology supports decarbonisation only if appropriately governed."
Applied to a single AI programme, the company reports above point to electricity and its origin first, then the hardware and buildings in the supply chain. A climate benefit claimed for the programme can be held to the standard the forecasting results set: a published, measured comparison.
Yes. Data centres used about 485 TWh of electricity in 2025, around 1.5% of global consumption, and the IEA projects the emissions associated with them to double to around 350 million tonnes in 2035, about 2% of global electricity sector emissions. Chips, servers and construction add supply chain emissions on top, which Google names as the main driver of the rise in its reported footprint.
Each use adds a small amount. The IEA says a simple text query typically uses less electricity than running a television for the same time. Demand grows with heavier uses such as video generation, reasoning and agentic tasks, which can use hundreds or thousands of times more energy per query than simple text generation.
Through better weather forecasts, grid management, demand response, industrial efficiency and flood warnings. ECMWF's AI forecasting system has been in operation since February 2025, and the IEA estimates that documented AI uses could save over 13 EJ of energy by 2035 if barriers to uptake are overcome.
Not on its own, on the evidence here. The IEA's savings estimate equals 3% of global final energy consumption by 2035, the authors of the main machine-learning review write that no one solution will fix climate change, and the IPCC finds that digital technology supports decarbonisation only if appropriately governed.
Data centre power and the climate uses of AI fall within two of the four panels at the Davos Super Intelligence Summit 2027 on 19 January 2027: the panel on AI infrastructure, which includes data centres, energy and scaling, and the panel on driving social impact through innovation. The programme sets out all four.