Hugging Carbon: Quantifying the Training Carbon Emissions of AI Models at Scale

Xinlei Wang, Ruibo Ming, Jing Qiu, Junhua Zhao, Jinjin Gu

Keywords: machine learning, artificial intelligence, language model, large language models, environmental impact, carbon emissions

Abstract: The scaling-law era has propelled artificial intelligence (AI) from research into a global industry, but its rapid growth raises concerns over energy demand, carbon emissions, and environmental sustainability. Unlike traditional sectors, AI still lacks systematic methodologies for comprehensive carbon accounting, leaving open the questions of how large the problem is today and how large it might be in the near future. We propose a FLOPs-based framework to estimate training emissions of open-source models on Hugging Face, introducing a tiered approach to handle uneven disclosure quality. Compute is converted to energy using hardware efficiency characteristics and then to emissions using the carbon intensity of the relevant grid, which we summarize as an AI Training Carbon Intensity (ATCI, emissions per compute) and for which we report an empirical reference value to enable quick model-level estimates. Our results show that training the most popular 5,234 models (with over 5,000 downloads) emitted approximately 5,8000 tons of carbon emissions. These findings provide the comprehensive industry-scale estimate of AI’s training footprint and a practical methodology to guide future standards and sustainability strategies.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 3754