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.
