Key Definitions
| Term | Definition |
|---|---|
| Green AI | The practice of developing and deploying AI systems with minimized environmental impact through energy-efficient training, optimized inference, responsible hardware lifecycle management, and selection of low-carbon computing infrastructure. |
| Carbon Footprint of AI | The total greenhouse gas emissions attributable to an AI system across its lifecycle, including hardware manufacturing, energy consumption for training and inference, cooling, data storage, and end-of-life disposal. |
| CSRD | The EU Corporate Sustainability Reporting Directive requiring large and listed companies to report on sustainability matters including environmental impact, applying to entities meeting defined size thresholds from financial years starting 2024 onwards. |
| European Sustainability Reporting Standards (ESRS) | The mandatory reporting standards developed by EFRAG that specify the sustainability information companies subject to CSRD must disclose, including climate-related metrics, energy consumption, and resource use. |
| Scope 1 Emissions | Direct greenhouse gas emissions from sources owned or controlled by the reporting organization, such as on-site generators powering AI infrastructure. |
| Scope 2 Emissions | Indirect greenhouse gas emissions from purchased electricity, heating, or cooling used to power AI systems and data centers. |
| Scope 3 Emissions | All other indirect greenhouse gas emissions in the organization's value chain, including cloud computing services, purchased AI tools, business travel, and supply chain emissions. |
| Power Usage Effectiveness (PUE) | A metric measuring data center energy efficiency, calculated as total facility energy divided by IT equipment energy, where a value of 1.0 represents perfect efficiency. |
| Model Efficiency | The ratio of AI model performance to the computational resources required, with more efficient models achieving comparable results using less energy and computing power. |
| EU Taxonomy | The EU classification system establishing which economic activities qualify as environmentally sustainable, relevant to AI systems that support climate change mitigation or adaptation. |
| Life Cycle Assessment (LCA) | A systematic evaluation of the environmental impacts of a product or system across its entire lifecycle from raw material extraction to end-of-life disposal. |
| Carbon Offsetting | The practice of compensating for greenhouse gas emissions by funding equivalent emission reductions elsewhere, subject to increasing scrutiny regarding quality and additionality. |
Chapter 1: The Environmental Impact of AI
AI systems consume significant amounts of energy throughout their lifecycle — this chapter provides detailed guidance on compliance and best practices.
1-1. Understanding AI's Energy Consumption
AI systems consume significant amounts of energy throughout their lifecycle.
Training large language models requires enormous computational resources.
A single training run for a large model can consume as much energy as several households use in a year.
The energy cost of AI training has been doubling approximately every 3.4 months.
Inference — the ongoing use of trained models — accumulates energy consumption over time.
While individual inference requests use little energy, billions of daily requests across deployed systems create substantial aggregate consumption.
Data storage for training datasets, model checkpoints, and operational data requires continuous energy.
Cooling systems for data centers add significantly to total energy consumption.
The power usage effectiveness of data centers determines how much overhead energy is consumed beyond computation.
Global data center energy consumption is estimated at 1 to 2 percent of global electricity generation.
AI workloads are growing faster than data center efficiency improvements.
Without intervention, AI could significantly increase global energy demand within a decade.
The environmental impact depends heavily on the carbon intensity of the electricity source.
AI training in Norway using hydroelectric power produces a fraction of the emissions of training in a coal-dependent grid.
Geographic and temporal optimization of AI workloads can significantly reduce carbon footprint.
1-2. Water Consumption of AI Systems
AI data centers consume large quantities of water for cooling.
Evaporative cooling systems use water directly to reduce temperatures.
Even closed-loop cooling systems consume water indirectly through electricity generation.
A single large model training run can consume thousands of liters of fresh water.
Water consumption is a growing concern in water-stressed regions.
Data center operators are increasingly required to report water consumption.
Alternative cooling technologies such as immersion cooling can reduce water consumption.
Locating data centers in cool climates reduces cooling requirements.
Organizations should consider water footprint alongside carbon footprint when evaluating AI infrastructure.
1-3. Hardware Lifecycle and E-Waste
AI hardware — GPUs, TPUs, and specialized accelerators — has significant embodied carbon from manufacturing.
The production of semiconductor chips requires energy-intensive processes and rare materials.
Rapid hardware obsolescence driven by AI performance demands creates e-waste challenges.
GPU refresh cycles for AI workloads are typically shorter than for general computing.
Responsible hardware lifecycle management extends useful life through refurbishment and repurposing.
End-of-life recycling of AI hardware recovers valuable materials but has its own environmental cost.
The EU Waste Electrical and Electronic Equipment (WEEE) Directive governs e-waste management.
Organizations should include hardware lifecycle management in their AI sustainability planning.
Second-life markets for AI hardware enable continued use at lower performance tiers.
Designing AI systems to work efficiently on existing hardware reduces upgrade pressure.
1-4. The Carbon Footprint of the AI Supply Chain
AI's carbon footprint extends beyond direct energy consumption to the entire supply chain.
Hardware manufacturing emissions constitute a significant portion of total lifecycle emissions.
Raw material extraction for semiconductor production involves mining and chemical processing.
Transportation of AI hardware from manufacturing to deployment adds supply chain emissions.
Cloud service provider infrastructure has its own embodied carbon and operational emissions.
Data acquisition and preparation processes consume energy and may involve travel.
The full supply chain perspective reveals emissions that operational measurement alone would miss.
Organizations should request supply chain emission data from AI vendors and cloud providers.
Life cycle assessment methodology provides a framework for comprehensive carbon accounting.