The rapid expansion of artificial intelligence workloads has intensified reliance on centralized data centers, resulting in significant energy consumption and infrastructure constraints. Concurrently, billions of edge devices including smartphones, laptops, and IoT systems operate at low average utilization, representing a vast pool of underutilized computational capacity. This paper presents Netara, a decentralized AI inference framework that applies circular economy principles by transforming idle edge devices into a distributed computing network without requiring additional physical infrastructure. The proposed system consists of a lightweight device agent that activates during idle periods, a task orchestration engine that decomposes workloads into micro-tasks, a consensus-based verification mechanism ensuring reliability across heterogeneous nodes, and a REST-based interface for seamless integration. By extending the lifecycle utilization of existing hardware resources, the framework aims to reduce dependence on energy-intensive data centers and improve sustainable compute accessibility. Preliminary prototype analysis suggests the feasibility of parallel task execution and fault-tolerant distributed inference, with projected potential cost reductions relative to comparable centralized inference systems. The proposed approach contributes to sustainable AI infrastructure aligned with circular economy principles, particularly for resource-constrained regions and the Global South, while introducing the concept of circular compute as a model for reusing underutilized consumer computational resources for AI inference.