The rapid rise of Generative AI (GenAI) technologies has brought transformative capabilities across industries but has also raised serious concerns about their environmental sustainability. As the computational demands of training and deploying large-scale AI models continue to escalate, so too does their carbon footprint. This paper adopts a comprehensive Life Cycle Assessment (LCA) approach to evaluate the environmental impact of GenAI models throughout their lifecycle— from hardware manufacturing and data center infrastructure to model training, deployment, and inference. We analyze and compare the energy efficiency and performance of five widely adopted GenAI models: GPT-3, ChatGPT (GPT-4), LLaMA 2, PaLM 2, and DistilBERT. Emissions are modeled using publicly available energy benchmarks, ML CO2 calculators, and estimation methodologies where direct data is unavailable. Beyond analysis, we introduce a Green AI Benchmarking Framework that integrates sustainability metrics, such as energy consumption and carbon emissions, into model evaluation standards, alongside traditional performance metrics. Our findings aim to guide researchers, developers, and policymakers toward more energy-conscious and environmentally responsible AI development practices.
Tegwen Malik (Mon,) studied this question.