Artificial Intelligence (AI) is transforming the lives of billions and plays a crucial role in solving complex problems and boosting productivity across industries. However, the rapid proliferation of AI technologies also raises significant environmental concerns. Training large-scale AI models requires vast computational resources, resulting in substantial energy consumption, carbon emissions, and water use. As AI capabilities grow, energy demand doubles approximately every ten months, outpacing improvements in hardware efficiency and renewable integration. This research investigates how the environmental footprint of AI can be mitigated through the application of circular economy principles and policy interventions inspired by extended producer responsibility. We explore the potential for reusing AI models via transfer learning, distillation, and repurposing, as well as the role of open-source sharing in prolonging model lifecycles. We also draw parallels between AI systems and consumer electronics, highlighting similarities in their environmental impacts and obsolescence dynamics. We develop a game-theoretic model comparing firm-level incentives with socially optimal outcomes and evaluate the effectiveness of various policy tools, including training and operations taxes, minimum usage requirements, and decreasing eco taxes designed to extend model longevity. Our findings reveal key trade-offs in AI model development and deployment decisions and identify regulatory approaches that can align private incentives with environmental goals. This study contributes to the emerging discourse on sustainable AI by proposing a structured framework for quantifying environmental costs and assessing regulatory levers that could guide the industry toward greener practices.