Fairer AI Carbon Accounting: Incorporating Market-based Attribution and Uncertainty in Embodied and Operational Carbon Footprint

The computational demands of large-scale AI models raise significant concerns about their carbon footprint. Most carbon accounting methods for large-scale AI models suffer from three key limitations: they overlook embodied carbon (from hardware manufacturing) or model it simplistically, rely on location-based carbon attribution that fails to reflect individual corporate efforts to decarbonize (e.g., via Power Purchase Agreements (PPAs)), and are deterministic, ignoring inherent uncertainties. This paper proposes PUMA, a Probabilistic Uncertainty Market Attribution carbon accounting model for large-scale AI models. PUMA integrates market-based carbon intensity to accurately account for the impact of PPAs and employs probabilistic modeling to capture uncertainties in the carbon accounting for AI models arising from spatiotemporal variations in manufacturing and operation, as well as evolving efficiency. We make an effort to develop a comprehensive carbon dataset by aggregating related data from diverse sources, and then we implement a simple yet effective Kernel Density Estimate (KDE) on the distribution of the parameters from the collected dataset. We compare PUMA with LLMCarbon, the state-of-the-art carbon accounting model for large AI models. The deviation of the accounting result is significant, reaching up to around 201%. CCS Concepts • Social and professional topics → Sustainability; • Computing methodologies → Artificial intelligence.