Schrödinger’s Carbon: Until Measured, Operational Emissions Remain Uncertain

Abstract

Rapidly improving generative artificial intelligence has led to significant investments in datacenter infrastructure, driving power demand, and raising environmental concerns. This has led to a growing body of research towards modeling embodied and operational carbon of datacenter servers across a variety of paradigms. However, most existing models take in deterministic inputs and output a singular average value that does not capture the inherent variability in estimating embodied and operational carbon emissions. Further, these average outputs obscure the impact of interacting factors, such as those related to deployment or software characteristics; each of which has its own underlying uncertainty distribution. This means in most cases, these averages do not accurately represent a particular server’s context. This thesis explicitly parameterizes and quantifies the full probabilistic distribution of operational carbon in AI inference tasks. It explores several factors of variability— deployment, spatiotemporal, and computational profile— and quantifies their impact on the overall carbon footprint through statistical and sensitivity analysis. While this work focuses on operational carbon, uncertainty propagation and understanding of variability should be used across a datacenter server’s entire life cycle. When this methodology is used alongside the existing uncertainty-aware embodied carbon measurements, it enables a holistic assessment from cradle to grave. This facilitates informed decision-making in server replacement, workload scheduling, hardware procurement, capacity planning, and more scenarios.