Explainer: Generative AI environmental footprint — what it is, why it matters, and how to evaluate options

Generative AI models consumed an estimated 4.3 TWh of electricity globally in 2025, roughly equivalent to the annual electricity consumption of Cyprus, and this figure is projected to reach 15 to 20 TWh by 2028. Behind every ChatGPT query, Midjourney image, and Copilot code suggestion lies a physical infrastructure of GPUs, cooling systems, and data centers with a measurable carbon, water, and materials footprint. For sustainability professionals evaluating whether and how to deploy generative AI tools, understanding these environmental costs is no longer optional. It is a prerequisite for responsible adoption and accurate Scope 2 and Scope 3 emissions reporting.

Why It Matters

The scale of generative AI deployment is growing exponentially. OpenAI reported 200 million weekly active users in early 2025, up from 100 million in late 2023. Google, Microsoft, Meta, Amazon, and Apple collectively committed over $300 billion in AI-related capital expenditure for 2025 and 2026, the majority flowing into data center construction and GPU procurement. The IEA estimated that global data center electricity consumption reached 460 TWh in 2025, approximately 2% of global electricity demand, with AI workloads representing the fastest-growing share.

This matters for sustainability professionals for three specific reasons. First, the EU Corporate Sustainability Reporting Directive (CSRD), effective for large companies from fiscal year 2024, requires disclosure of energy consumption and greenhouse gas emissions from digital operations, including cloud computing and AI services. Companies using generative AI through cloud providers must account for these emissions in their Scope 3 reporting under Category 1 (purchased goods and services) or Category 11 (use of sold products). Second, the European AI Act, which entered into force in August 2024 with phased compliance deadlines through 2027, requires providers of high-risk AI systems to document energy consumption and environmental impact. Third, investor scrutiny is intensifying: CDP reported that 78% of institutional investors now request data on companies' digital infrastructure emissions, up from 42% in 2022.

Organizations that fail to measure and manage their AI-related environmental footprint face regulatory risk, reputational exposure, and inaccurate emissions inventories that undermine the credibility of net-zero commitments.

Key Concepts

Training vs. Inference Energy represents the most fundamental distinction in understanding AI environmental impact. Training is the computationally intensive process of building a model by processing vast datasets across thousands of GPUs over weeks or months. Training GPT-4 consumed an estimated 50 to 80 GWh of electricity, equivalent to powering 5,000 US homes for a year. However, inference, the process of running trained models to generate outputs, collectively consumes far more energy at scale because it occurs billions of times daily across millions of users. A single ChatGPT query consumes approximately 3 to 10 Wh of electricity, roughly 10 times a standard Google search. At 200 million weekly users generating multiple queries each, inference energy for ChatGPT alone reaches an estimated 1 to 2 TWh annually.

Power Usage Effectiveness (PUE) measures total data center energy consumption divided by IT equipment energy consumption. A PUE of 1.0 would mean all electricity powers computing equipment with no overhead for cooling, lighting, or infrastructure. The industry average PUE in 2025 was approximately 1.55, meaning 55% additional energy is consumed for non-computing functions. Hyperscale operators like Google (PUE 1.10) and Meta (PUE 1.08) achieve significantly better efficiency through advanced cooling, optimized airflow, and custom hardware design. When evaluating AI providers, PUE directly impacts the total emissions attributable to each query or model run.

Water Usage Effectiveness (WUE) captures the liters of water consumed per kilowatt-hour of IT energy for data center cooling. Microsoft disclosed that its global water consumption increased 34% from 2021 to 2023, reaching 7.8 billion liters, largely driven by AI workload growth. Evaporative cooling, the most common approach in large data centers, consumes 1 to 5 liters of water per kWh of IT load depending on climate and system design. In water-stressed regions, this consumption creates material environmental and social risks.

Embodied Carbon refers to the greenhouse gas emissions from manufacturing, transporting, and eventually disposing of the hardware that runs AI workloads. A single NVIDIA H100 GPU has an estimated embodied carbon footprint of 150 to 200 kg CO2e, and training a frontier model may require 10,000 to 25,000 GPUs. The total embodied carbon of the hardware needed to train one frontier model can reach 2,000 to 5,000 tonnes CO2e, a figure frequently omitted from published analyses that focus exclusively on operational energy.

Carbon Intensity of Electricity varies dramatically by location and time. Running AI workloads on a grid powered primarily by hydroelectric and nuclear generation (as in Quebec, Norway, or France) produces 10 to 50 g CO2e per kWh, while the same workload on a coal-heavy grid (as in parts of Poland, India, or the US Midwest) generates 600 to 900 g CO2e per kWh. The carbon footprint of identical AI computations can vary by a factor of 20 or more depending solely on where the data center is located.

Decision Framework: Evaluating AI Options by Environmental Impact

Sustainability professionals selecting generative AI tools or cloud providers should evaluate options across five dimensions.

Energy Transparency. Does the provider publish energy consumption per query, per token, or per model call? Google Cloud Carbon Footprint, Microsoft Emissions Impact Dashboard, and AWS Customer Carbon Footprint Tool all provide account-level emissions data, but granularity varies significantly. Providers that disclose per-workload energy data enable more accurate Scope 3 accounting.

Grid Carbon Intensity. Where are the data centers processing your workloads physically located, and what is the carbon intensity of the local electricity grid? Google matches 100% of its global electricity consumption with renewable energy purchases on an annual basis and has committed to 24/7 carbon-free energy matching by 2030. Microsoft achieved 100% renewable energy matching in 2025. However, annual matching does not mean every computation is powered by clean energy in real time. Ask providers whether they offer region selection that allows routing workloads to lower-carbon grids.

Model Efficiency. Smaller, task-specific models consume dramatically less energy than frontier models. Running a 7-billion parameter model for text summarization consumes approximately 1% of the energy required by a 1.8-trillion parameter model for the same task. Sustainability professionals should match model capability to task requirements rather than defaulting to the largest available model. Fine-tuned smaller models frequently outperform general-purpose large models on domain-specific tasks while consuming a fraction of the energy.

Hardware Generation. NVIDIA H100 GPUs deliver approximately 3 times the performance per watt of the previous-generation A100 for inference workloads. Google's custom TPU v5e achieves approximately 2 times the efficiency of TPU v4 for large language model inference. Providers using current-generation hardware deliver lower per-query emissions even if their overall energy consumption is higher due to scale.

Water Stewardship. In water-stressed regions, data center cooling creates meaningful competition for scarce water resources. Evaluate whether providers use air cooling, closed-loop liquid cooling, or evaporative systems, and whether they operate in regions classified as high or extremely high water stress by the World Resources Institute Aqueduct tool.

Real-World Benchmarks

Hugging Face published lifecycle carbon assessments for its open-source models in 2024, setting a transparency benchmark. Training BLOOM (176B parameters) on the Jean Zay supercomputer in France produced 24.7 tonnes CO2e, benefiting from France's low-carbon nuclear-dominated grid. An equivalent training run on a US average grid would have produced approximately 120 tonnes CO2e, illustrating the decisive impact of grid location.

Salesforce researchers demonstrated in 2025 that a fine-tuned 3B parameter model matched GPT-4 performance on customer service classification tasks while consuming less than 0.5% of the inference energy. Their findings were consistent with broader industry research showing that 80 to 90% of enterprise generative AI use cases can be served by models with fewer than 13 billion parameters.

Microsoft reported in its 2025 Environmental Sustainability Report that AI workloads accounted for approximately 15% of its total data center energy consumption but were growing at 40% year-over-year, the fastest-growing workload category. The company invested $1.3 billion in carbon-free energy procurement specifically to offset AI-driven demand growth.

Action Checklist

• Inventory all generative AI tools and services in use across your organization, including shadow IT deployments
• Request per-workload energy and emissions data from each cloud and AI provider using their sustainability dashboards
• Map AI workloads to data center locations and assess the carbon intensity of local electricity grids
• Evaluate whether frontier models are necessary for each use case or whether smaller, fine-tuned models can deliver equivalent performance at lower environmental cost
• Include AI-related emissions in Scope 3 Category 1 reporting and verify alignment with CSRD and ISSB disclosure requirements
• Establish procurement criteria that weight energy transparency, renewable energy commitments, and water stewardship alongside capability and cost
• Set organizational targets for AI energy efficiency improvement, such as reducing energy per inference by 20% annually through model optimization and hardware refresh
• Monitor regulatory developments including the EU AI Act environmental disclosure requirements and emerging carbon accounting standards for digital services

FAQ

Q: How much carbon does a single ChatGPT query produce? A: A single ChatGPT-4 query produces approximately 3 to 10 g CO2e depending on query complexity, response length, and the carbon intensity of the data center's electricity grid. For comparison, a standard Google search produces approximately 0.2 to 0.3 g CO2e. An organization generating 10,000 AI queries per day would produce roughly 10 to 35 tonnes CO2e annually from AI inference alone, before accounting for training, fine-tuning, or embodied hardware emissions.

Q: Are renewable energy claims from cloud providers reliable? A: Most major cloud providers match 100% of annual electricity consumption with renewable energy purchases, but this is annual volumetric matching, not real-time carbon-free energy. On a given hour, a data center may draw 80% of its power from fossil fuel generation while balancing that consumption with renewable energy certificates purchased from a wind farm operating at a different time and location. Google publishes hourly carbon-free energy percentages for each data center region, providing the most granular transparency currently available.

Q: Should we avoid using generative AI for sustainability reasons? A: Not necessarily. The relevant question is whether the value generated by AI use justifies its environmental cost and whether that cost is minimized through responsible deployment choices. AI applications that enable emissions reductions elsewhere, such as optimizing building energy systems, improving supply chain efficiency, or accelerating materials discovery, may produce net-positive environmental outcomes even after accounting for their computational footprint. The goal is informed, intentional use rather than blanket avoidance.

Q: How do I account for AI emissions in corporate sustainability reporting? A: Under the GHG Protocol and CSRD, AI-related emissions from cloud services fall under Scope 3 Category 1 (purchased goods and services). Use provider sustainability dashboards to obtain account-level emissions data. For more granular reporting, request workload-level data and apply provider-specific emission factors. The Partnership for Carbon Transparency (PACT) Pathfinder Framework provides guidance on obtaining and verifying product-level carbon data from cloud providers.

Sources

• International Energy Agency. (2025). Data Centres and Data Transmission Networks: Tracking Report. Paris: IEA Publications.
• Luccioni, A. S., Viguier, S., & Ligozat, A. L. (2024). "Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model." Journal of Machine Learning Research, 24(253), 1-15.
• Microsoft. (2025). 2025 Environmental Sustainability Report. Redmond, WA: Microsoft Corporation.
• Google. (2025). 24/7 Carbon-Free Energy: 2025 Progress and Methodology. Mountain View, CA: Google LLC.
• Dodge, J., et al. (2024). "Measuring the Carbon Intensity of AI in Practice." Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency, 1877-1894.
• European Commission. (2024). European AI Act: Environmental Impact Assessment Requirements. Brussels: EC Directorate-General for Communications Networks.
• Patterson, D., et al. (2024). "The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink." IEEE Computer, 57(1), 18-28.