The rapid expansion of artificial intelligence technologies has raised significant environmental concerns, with projections indicating substantial increases in energy consumption, carbon emissions, and water usage through 2030. Despite growing awareness, the field lacks a unified approach to defining, measuring, and governing the environmental impacts of AI systems. Current frameworks suffer from fragmented metrics, incomplete lifecycle coverage, and insufficient integration with policy mechanisms. This paper addresses these gaps by proposing a comprehensive theoretical framework for sustainable AI that integrates environmental assessment across five distinct lifecycle phases: design and planning, development and training, deployment and inference, operation and maintenance, and decommissioning. Drawing on Life Cycle Assessment theory, systems theory, and stakeholder perspectives, the framework provides standardized metrics for energy, carbon, water, and embodied impacts while incorporating governance mechanisms at each phase. The framework offers actionable guidance for researchers in designing sustainability-aware studies, practitioners in implementing green AI solutions, organizations in strategic planning, and policymakers in developing effective regulations. By bridging technical implementation with policy governance, this work contributes to the advancement of sustainable AI scholarship and provides a foundation for future empirical validation and sector-specific adaptations.