BEST: Blockchain-Enabled Sustainable Task Scheduling for Optimizing Large AI Model Workloads at the Edge

The proliferation of Large AI Models (LAMs) to the network edge promises unprecedented intelligence in real-time applications. It also introduces a critical sustainability challenge due to the immense energy footprint of model inference on resource-constrained devices. While model-level optimizations have made progress, they fail to address the systemic issue of aggregate energy consumption across a network. This paper introduces BEST (Blockchain-Enabled Sustainable Task Scheduling), a novel framework that reimagines edge resource management to explicitly tackle this sustainability crisis. BEST establishes a decentralized marketplace for underutilized edge resources by introducing a novel application of Real-World Asset (RWA) tokenization to abstract and commoditize computational capacity and battery power. These tokenized resources can be securely and transparently traded among peer devices via a blockchain-based infrastructure. To navigate this dynamic marketplace, we develop a cooperative multi-agent deep reinforcement learning (MADRL) algorithm based on Proximal Policy Optimization (PPO). The agents learn a sophisticated scheduling policy that dynamically balances task latency, system-wide energy consumption, and task success rate. Extensive simulations against compared baselines demonstrate BEST’s unique strengths. It reduces total energy consumption by up to 26% — 46% against other schemes, while maintaining a high task success rate and latency competitive with the state-of-the-art scheme. By creating a market-driven, collaborative ecosystem, BEST provides a scalable and effective pathway toward achieving truly sustainable LAM at the edge.