Abstract

The rapid evolution of Large Language Models (LLMs) and generative Artificial Intelligence (AI) technologies has significantly increased power consumption, raising concerns about their environmental impact. This study proposes a methodology to estimate local Carbon Dioxide Equivalent (CO2eq) emissions by recording GPU power consumption and utilizing carbon intensity data from Electricity Maps. Additionally, a case study involving a server equipped with two NVIDIA H100 GPUs is presented. The analysis reveals that users tend to train AI models during periods when energy is predominantly sourced from thermal power plants, potentially increasing the carbon footprint as AI adoption becomes more widespread.