Title: Understanding the Environmental Impact of Generative AI Services
Subtitle: Developing Effective Methods to Measure the Sustainability of Generative AI
Resource URL: https://cacm.acm.org/sustainability-and-computing/understanding-the-environmental-impact-of-generative-ai-services/
Publication Date: 2025-06-03
Format Type: Blog Post
Reading Time: 19 minutes
Contributors: Laurent Lefèvre;Mathilde Jay;Eddy Caron;Adrien Berthelot;
Source: Communications of the ACM (Association for Computing Machinery)
Keywords: [Generative AI, Life Cycle Assessment, Sustainable Computing, Carbon Footprint, Metal Depletion]
Job Profiles: DevOps Engineer;Data Scientist;Chief Information Officer (CIO);Artificial Intelligence Engineer;Chief Technology Officer (CTO);
Synopsis: In this blog post, Professor Eddy Caron, PhD student Mathilde Jay, and researchers Adrien Berthelot and Laurent Lefèvre present a new methodology for assessing the environmental impact of GenAI services, focusing on energy consumption, material usage, and life cycle effects.
Takeaways: [Generative AI services impose substantial environmental pressures not only from data-center operations but also from end-user terminals and network infrastructure, which together account for over 85% of metal depletion potential., Replicating a fraction of the training process and applying regression yields an accurate, reproducible estimate of total energy consumption, avoiding the impracticality of complete retraining., Thermal design power estimates can overstate GPU energy use by around 6%, while software-based power meters may underreport actual consumption by roughly 20% compared to external meters., A multi-impact life cycle assessment—including global warming potential, primary energy demand, and abiotic depletion—reveals that embodied emissions and resource extraction remain significant even after decarbonizing datacenter electricity., Extending the operational life of hardware and software emerges as a key lever for reducing both embodied and operational environmental impacts, mitigating rebound effects associated with efficiency gains.]
Summary: Over the past decades, digital transformation has placed growing pressure on environmental resources, and the arrival of generative artificial intelligence (AI) services like ChatGPT and Stable Diffusion has intensified concerns about electricity consumption and material use. The authors argue that current methods—often limited to thermal design power estimates or partial life-cycle assessments—fail to capture the full environmental footprint of such systems.

To address this, they adapt life cycle assessment (LCA) standards (ISO 14040/44) to GenAI services, broadening the scope beyond carbon emissions to include metal depletion and primary energy demand. They present a service-level model, tracing user requests from personal terminals through networks, web servers, inference hardware, and data storage, while separating operational from embodied impacts. A novel sampling methodology replicates a fraction of the training process—validated by experiments on the Grid’5000 Sirius cluster—to derive regression equations that accurately estimate total training energy consumption without full retraining.

Applying this approach to Stable Diffusion, the authors estimate multi-category impacts over one year of service: global warming potential, abiotic depletion, and cumulative energy demand. Their results reveal that end-user terminals and networks account for over 85% of metal depletion potential and more than 45% of carbon emissions, underscoring the importance of including peripheral infrastructure. They also demonstrate that thermal design power often overestimates consumption, while software-based meters can underreport by 20% compared to external power meters. The study highlights the need for extending hardware and software lifecycles, improving transparency in deployment data, and adopting holistic assessment tools to guide sustainable GenAI development.
Content: ## Introduction

Digital technologies have expanded relentlessly over recent decades, a phenomenon commonly referred to as digital transformation. While these advances promise greater efficiency and novel applications, they also impose a growing burden on the environment. The latest wave—generative artificial intelligence (AI), or GenAI—focuses on producing human-like text, code, and images via services such as ChatGPT and Stable Diffusion. As these services gain popularity, questions about their sustainability intensify, yet existing evaluation methods remain fragmented and incomplete.

In this article, researchers from multiple European institutions propose a transparent and reproducible methodology to assess the environmental impact of GenAI services. By combining life cycle assessment (LCA) principles with experimental energy measurements, this approach accounts for both carbon emissions and resource depletion across the entire service life cycle—from hardware manufacturing to user interactions.

## 1. Limitations of Current Environmental Impact Assessments for AI

### 1.1 Shortcomings of Training-Phase Estimates

Early studies have focused primarily on the machine learning training phase, typically estimating electricity use via thermal design power (TDP) ratings multiplied by training duration. Although straightforward, this method neglects auxiliary components—central processing units (CPUs), memory, fans, network switches—and often overestimates consumption. Software-based power meters (e.g., Intel’s Running Average Power Limit or NVIDIA Management Library) offer greater granularity but still require live measurements and exclude non-computing infrastructure.

### 1.2 Neglect of Full Life Cycle Impacts

Even fewer analyses incorporate the full life cycle of equipment, limiting themselves to carbon costs during training and inference. Yet digital services drive demand for rare metals and involve end-user devices and network infrastructures. Without a multi-phase, multi-category assessment, key environmental pressures—such as metal depletion and embodied emissions—remain hidden.

## 2. Life Cycle Assessment for Digital Services

Life cycle assessment (LCA), governed by ISO 14040 and 14044 standards, evaluates potential environmental impacts across all product phases: manufacturing, use, and end of life. The LCA process comprises four steps: defining goals and scope, compiling an inventory of flows, characterizing impacts, and interpreting results under uncertainty. Though long applied in sectors like manufacturing, LCA has only recently been adopted for information and communication technologies (ICT), where it can expose hidden sources of environmental burden.

## 3. Methodology for Measuring GenAI Environmental Impact

### 3.1 Service-Level Model

Extending prior frameworks, the proposed model represents a GenAI service as a sequence of components: personal terminals, networks, web servers for inference, data storage, and the dedicated training phase. User requests traverse these elements, each contributing to electricity use and material consumption. While the current study omits the opaque process of training-data production, it includes all hardware life cycle phases.

### 3.2 Estimating Training Electricity via Replication

To overcome the impracticality of full retraining, the authors replicate a fraction of the training process on the Grid’5000 “Sirius” cluster, monitoring power with external meters (0.1 W precision at 1 Hz) and open-source tools at 2 Hz. Experiments on a single GPU node demonstrate that energy per step remains constant, enabling linear regression models to project total energy use from sample runs. For Stable Diffusion versions 1-1 through 1-5, they derive equations of the form:

Energy (kWh) = a × N + b

where _N_ is the number of training steps, with coefficients fitted to experimental data (R² > 0.99).

### 3.3 LCA-Based Modeling

The assessment targets three impact categories: abiotic depletion potential (mineral and metal resource use), global warming potential (carbon equivalent emissions), and primary energy demand (cumulative energy use). Device footprints and electricity-mix emissions are sourced from public inventories (e.g., ADEME Base Empreinte), consortium data (NegaOctet), and open projects (Boavizta). By mapping measured or estimated electricity consumption and equipment lifespans to these databases, the model yields multi-category impact results.

## 4. Application to Stable Diffusion

### 4.1 Experimental Setup and Energy Estimates

Stable Diffusion’s training phases (versions v1-1, v1-4, v1-5) involve hundreds of GPUs and CPUs, with total steps ranging from 1.94 × 10³ to 5.95 × 10⁵. Regression-based projections estimate per-node energy between 401 and 1 060 kWh, scaling to 12.8–33.9 MWh across 32 GPUs. These figures align with existing literature on comparable models.

### 4.2 Full-Year Service Impact

Over one year (August 2022–August 2023), the service received measured traffic, with half of site visits resulting in one four-image generation at 512 × 512 resolution. Assuming a U.S. average electricity mix for datacenter operations and a user-weighted mix for terminals and networks, the LCA reveals:

• Total carbon emissions: 463 tCO₂e
• Terminals and networks: >85% of abiotic depletion potential, >45% of carbon footprint, >30% of primary energy demand

Operational and embodied impacts are separated, highlighting that decarbonizing electricity sources alone cannot eliminate embodied emissions in hardware and user devices.

## 5. Discussion

### 5.1 Infrastructure Footprint and Rebound Effects

The disproportionate impact of terminals and networks grows with user volume, emphasizing that service-scale assessments must include peripheral infrastructure. Although improving efficiency and decarbonizing datacenters reduce operational emissions, they risk a rebound effect, whereby lower per-unit costs drive increased overall usage.

### 5.2 Hardware and Software Obsolescence

Training resembles a one-time production cost analogous to hardware manufacturing, reinforcing the value of prolonging equipment and software lifecycles. Avoiding premature replacement curbs embodied emissions and resource depletion.

### 5.3 Accuracy of Energy Estimation Methods

Contrary to expectations, TDP overestimates GPU consumption by around 6%, while software power meters underreport by approximately 20% when scaled to full training. The authors conclude that replicating training steps with external meters yields the most reliable estimates; in their absence, software meters are acceptable, and TDP can serve as a rough fallback.

## 6. Conclusion

This work advances the evaluation of generative AI sustainability by integrating service-level modeling, multi-category life cycle assessment, and enhanced energy estimation techniques. It uncovers the substantial contributions of end-user devices and networks to overall impact and demonstrates the critical need for transparent, reproducible methods. To guide informed decisions on GenAI development and deployment, the community must improve data transparency—including detailed electricity consumption, deployment practices, and equipment lifespans—so that future assessments can refine these estimates and drive meaningful environmental improvements.

## Acknowledgments

The researchers thank public and open-source inventory providers for life cycle data, as well as the Grid’5000 platform for experimental resources. Funding support was received from national and regional research programs in sustainable information and communication technologies.