Explainer: Responsible AI and its environmental impact

Why It Matters

Training a single large language model can emit 300 to 500 tonnes of CO₂, roughly equivalent to 60 round-trip transatlantic flights (Strubell et al., 2019; updated estimates by Luccioni et al., 2024). The International Energy Agency (IEA, 2025) projects that global data-centre electricity consumption will more than double between 2022 and 2026, reaching approximately 1,000 TWh annually, with AI workloads accounting for a rapidly growing share. Goldman Sachs Research (2024) estimates that AI-driven demand alone could push data-centre power consumption to 4.5 percent of worldwide electricity generation by 2030. For sustainability professionals, these numbers signal a paradox: AI is simultaneously one of the most powerful tools for decarbonisation and one of the fastest-growing sources of electricity demand. Responsible AI frameworks attempt to resolve this tension by embedding environmental accountability into every stage of the AI lifecycle, from model design and hardware procurement through deployment and end-of-life disposal of compute infrastructure.

Key Concepts

Carbon footprint of AI. The environmental cost of an AI system spans two phases. The embodied carbon phase covers semiconductor fabrication, server manufacturing, and facility construction. The operational carbon phase covers the electricity consumed during training and inference. Research from the AI Index at Stanford University (2025) shows that inference now dominates total compute demand because models are queried billions of times after a single training run. A single query to a large language model consumes roughly ten times the energy of a standard web search (IEA, 2025).

Power Usage Effectiveness (PUE). PUE measures data-centre energy efficiency as the ratio of total facility energy to IT equipment energy. The global average PUE in 2024 was approximately 1.58, meaning 58 percent of energy was consumed by cooling, lighting, and other overhead (Uptime Institute, 2024). Hyperscale operators such as Google and Microsoft report PUEs between 1.10 and 1.20, but edge and colocation facilities often exceed 1.80.

Green AI vs. Red AI. The concept introduced by Schwartz et al. (2020) distinguishes research that prioritises computational efficiency ("Green AI") from research that pursues accuracy gains through ever-larger models regardless of resource cost ("Red AI"). Green AI encourages reporting FLOPs, energy consumption, and carbon emissions alongside accuracy metrics.

Responsible AI governance. Responsible AI extends beyond environmental concerns to encompass fairness, transparency, privacy, and safety. However, the EU AI Act (2024) and the OECD AI Principles (updated 2024) increasingly recognise environmental sustainability as a governance dimension. Article 40 of the EU AI Act requires providers of general-purpose AI models above a computational threshold to disclose energy consumption and known environmental impacts.

Water footprint. AI cooling systems consume substantial freshwater. Researchers at the University of California, Riverside (Li et al., 2023) estimated that training GPT-3 consumed approximately 700,000 litres of freshwater for cooling. Microsoft's own 2024 sustainability report disclosed a 34 percent increase in water consumption between 2021 and 2023, driven largely by AI infrastructure expansion.

How It Works

Reducing AI's environmental footprint operates across four layers.

Hardware efficiency. Chip manufacturers design processors that deliver more computation per watt. NVIDIA's Blackwell B200 GPU (2024) delivers up to four times the training performance per watt compared with its predecessor, the H100. Google's TPU v5p achieves similar gains through custom silicon optimised for transformer architectures.

Model architecture optimisation. Techniques such as model distillation, pruning, quantisation, and mixture-of-experts routing reduce the number of floating-point operations required for a given task. Hugging Face (2025) demonstrated that distilled models can retain 95 percent of a large model's accuracy while using 60 percent fewer FLOPs. Sparse mixture-of-experts architectures, used in models like Mixtral (Mistral AI, 2024), activate only a fraction of total parameters per inference pass, cutting energy use proportionally.

Infrastructure decarbonisation. Operators procure renewable energy, deploy liquid cooling, and locate facilities in cold climates to reduce PUE. Google (2025) reported matching 64 percent of its global data-centre electricity with carbon-free energy on an hourly basis. Microsoft committed to being carbon-negative by 2030 and has signed over 13.5 GW of clean energy agreements since 2024.

Carbon-aware scheduling. Workloads shift in time and geography to coincide with periods of high renewable generation. Research from Allen Institute for AI (2024) showed that carbon-aware job scheduling across multiple regions can reduce training emissions by 30 to 40 percent without increasing wall-clock time.

What's Working

Hyperscale transparency is improving. Google, Microsoft, and Meta now publish annual environmental reports disclosing energy consumption, water use, PUE, and carbon-free energy percentages. Google's 2025 Environmental Report revealed that the company reduced its Scope 2 emissions by 50 percent relative to 2019 despite a threefold increase in compute capacity, largely through 24/7 carbon-free energy matching.

Efficient model design is gaining traction. The emergence of small language models (SLMs) and on-device inference has demonstrated that many enterprise use cases do not require frontier-scale models. Microsoft's Phi-3 family (2024) showed that a 3.8-billion-parameter model can match GPT-3.5 on several benchmarks while consuming a fraction of the compute. Hugging Face's open-source leaderboards now rank models by efficiency metrics alongside accuracy, shifting community incentives.

Regulatory momentum is building. The EU AI Act's energy-disclosure requirements for general-purpose models create a compliance baseline that will propagate globally as multinational companies standardise reporting. The UK AI Safety Institute (2025) has begun evaluating environmental impacts as part of its frontier model assessments.

Industry coalitions are scaling best practices. The Green Software Foundation, whose members include Accenture, Microsoft, GitHub, and Thoughtworks, has published the Software Carbon Intensity (SCI) specification, enabling developers to measure and reduce the carbon intensity of software per unit of work.

What Isn't Working

Rebound effects undermine efficiency gains. As models become cheaper to run, usage scales faster than efficiency improves. The IEA (2025) notes that despite hardware efficiency gains of roughly 20 percent per year, total AI energy demand is growing 25 to 35 percent annually, producing a net increase in absolute consumption.

Scope 3 reporting remains opaque. Most AI companies disclose Scope 1 and 2 emissions but struggle to account for the embodied carbon of semiconductor fabrication, rare-earth mining for magnets and cooling systems, and end-of-life e-waste. TSMC, which fabricates the majority of advanced AI chips, does not publicly disaggregate the carbon footprint of individual chip lines (Climate TRACE, 2025).

Water stress is intensifying. Data centres are often sited in water-stressed regions because of grid access or tax incentives. Meta's data centre in Mesa, Arizona, and Google's facilities in The Dalles, Oregon, have drawn scrutiny from local communities concerned about freshwater diversion during drought conditions (The Guardian, 2024).

Voluntary commitments lack enforcement. Corporate net-zero pledges in the AI sector often rely on carbon offsets rather than absolute reductions. Amazon, for example, reported reaching 100 percent renewable energy matching in 2023, yet its absolute carbon emissions rose 34 percent between 2019 and 2023, driven in part by AWS data-centre expansion (Amazon Sustainability Report, 2024).

Key Players

Established Leaders

• Google DeepMind — Pioneer in energy-efficient AI; reduced data-centre cooling energy by 40% using reinforcement learning; leads 24/7 carbon-free energy matching
• Microsoft — Committed to carbon-negative by 2030; signed 13.5 GW+ in clean energy deals; developed Phi-series efficient SLMs
• NVIDIA — Dominant AI chip supplier; Blackwell B200 delivers 4x training performance per watt vs. prior generation
• Meta — Open-sourced LLaMA model family; published detailed data-centre energy and water disclosures

Emerging Startups

• Hugging Face — Open-source AI platform promoting model efficiency leaderboards and carbon tracking tools
• Mistral AI — Paris-based startup deploying sparse mixture-of-experts models that reduce inference energy
• WattTime — Provides real-time grid emissions data enabling carbon-aware compute scheduling
• Cerebras Systems — Wafer-scale AI chips designed for energy-efficient training at scale

Key Investors/Funders

• Breakthrough Energy Ventures — Bill Gates-backed fund investing in sustainable compute and data-centre efficiency
• Green Software Foundation — Industry coalition (Accenture, Microsoft, Thoughtworks) developing open-source carbon measurement standards
• EU Horizon Europe — Funded research programmes on sustainable AI infrastructure and Green AI methodologies

Sector-Specific KPI Benchmarks

Action Checklist

• Audit your AI carbon footprint. Use tools such as CodeCarbon, ML CO2 Impact, or the Green Software Foundation's SCI specification to measure training and inference emissions across all deployed models.
• Adopt efficient architectures first. Evaluate whether a distilled or small language model meets your use-case requirements before defaulting to a frontier-scale model. Benchmark accuracy against FLOPs, not just accuracy alone.
• Require PUE and CFE disclosures from cloud providers. When negotiating cloud contracts, request hourly carbon-free energy percentages, PUE, and WUE data for the specific regions hosting your workloads.
• Implement carbon-aware scheduling. Use WattTime or Electricity Maps APIs to shift non-urgent training jobs to times and regions with the highest renewable penetration.
• Set science-based targets for AI operations. Align AI infrastructure emissions with SBTi pathways and include Scope 3 estimates for hardware embodied carbon.
• Disclose AI environmental impacts in sustainability reports. Follow the EU AI Act's disclosure template and report energy consumption, water use, and emissions per model or product line.
• Engage suppliers on embodied carbon. Request lifecycle assessments from chip manufacturers and server OEMs; factor embodied carbon into procurement decisions.

FAQ

How much energy does training a large AI model actually consume? Training a frontier large language model with hundreds of billions of parameters can require 10 to 50 GWh of electricity, depending on hardware, data-centre efficiency, and training duration. For context, 50 GWh is roughly the annual electricity consumption of 5,000 average US households. Inference costs per query are individually small (0.5 to 5 Wh) but multiply rapidly at scale: a model serving one billion queries per month can consume as much energy as its entire training run every few weeks.

Does the EU AI Act require environmental reporting? Yes, partially. Article 40 of the EU AI Act requires providers of general-purpose AI models that exceed a defined computational threshold (generally interpreted as 10²⁵ FLOPs of training compute) to report energy consumption and known environmental impacts. The European Commission is expected to issue detailed reporting guidelines by late 2026. Even organisations below the threshold are encouraged to disclose voluntarily, as institutional investors and enterprise buyers increasingly request this data.

Can AI actually help reduce emissions despite its own footprint? Substantial evidence suggests yes. The Boston Consulting Group (BCG, 2024) estimated that AI applications in energy, transport, agriculture, and industry could help abate 5 to 10 percent of global greenhouse gas emissions by 2030, a figure that far exceeds AI's own projected footprint of roughly 2 to 4 percent of global electricity. However, realising this net benefit requires deliberate governance: deploying AI for high-impact decarbonisation use cases while simultaneously minimising the environmental cost of the AI systems themselves.

What is carbon-aware computing and how effective is it? Carbon-aware computing shifts workloads in time or across geographic regions to coincide with periods of high renewable energy generation on the grid. Research from the Allen Institute for AI (2024) demonstrated 30 to 40 percent reductions in training-run emissions using multi-region carbon-aware scheduling. Tools like WattTime and Electricity Maps provide real-time marginal emissions data for hundreds of grid regions worldwide, enabling automated scheduling decisions.

How should organisations choose between on-premise and cloud AI infrastructure for sustainability? Cloud hyperscalers typically operate more energy-efficient facilities (PUE of 1.10 to 1.20) than enterprise on-premise data centres (PUE of 1.50 to 2.00) and have greater purchasing power for renewable energy. However, data-sovereignty requirements, latency constraints, and total-cost considerations may favour on-premise or colocation. The key is to compare actual carbon intensity per unit of compute rather than relying on headline PUE alone, and to factor in the embodied carbon of any new hardware purchases.

Sources

• Strubell, E., Ganesh, A., & McCallum, A. (2019). Energy and Policy Considerations for Deep Learning in NLP. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics.
• 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.
• International Energy Agency. (2025). Electricity 2025: Analysis and Forecast to 2027. IEA, Paris.
• Goldman Sachs Research. (2024). AI, Data Centers and the Coming US Power Demand Surge. Goldman Sachs Global Investment Research.
• Stanford University Human-Centered AI. (2025). AI Index Report 2025. Stanford HAI.
• Uptime Institute. (2024). Global Data Center Survey 2024. Uptime Institute.
• Li, P., Yang, J., Islam, M. A., & Ren, S. (2023). Making AI Less Thirsty: Uncovering and Addressing the Secret Water Footprint of AI Models. arXiv preprint arXiv:2304.03271.
• Schwartz, R., Dodge, J., Smith, N. A., & Etzioni, O. (2020). Green AI. Communications of the ACM, 63(12), 54–63.
• European Parliament & Council. (2024). Regulation (EU) 2024/1689 Laying Down Harmonised Rules on Artificial Intelligence (AI Act). Official Journal of the European Union.
• Allen Institute for AI. (2024). Carbon-Aware Multi-Region Scheduling for Large-Scale Model Training. AI2 Technical Report.
• Google. (2025). 2025 Environmental Report. Alphabet Inc.
• Microsoft. (2024). 2024 Environmental Sustainability Report. Microsoft Corporation.
• Amazon. (2024). 2023 Sustainability Report. Amazon.com Inc.
• Boston Consulting Group. (2024). How AI Can Be a Powerful Tool in the Fight Against Climate Change. BCG.
• Green Software Foundation. (2024). Software Carbon Intensity (SCI) Specification v1.0. GSF.
• Climate TRACE. (2025). Global Emissions Monitoring: Semiconductor Manufacturing Sector. Climate TRACE Coalition.