Eco-Friendly Federated Learning: Designing a FinOps Framework for Multicloud Resource Coordination

Federated Learning (FL) is increasingly adopted as a privacy-preserving machine learning paradigm, enabling decentralized model training across geographically distributed and siloed data sources. Its extension to multi-cloud environments, involving multiple cloud service providers, remains a nascent, yet highly relevant research area, particularly due to the operational and architectural heterogeneity it introduces. This scenario stresses the need for efficient coordination of computational resources, cost governance, and sustainability, which are not adequately addressed in current FL implementations. This work presents ECOFED (Ecologically and Cost-Optimized Federated Learning), a FinOps-driven framework for FL in multicloud contexts, aiming to optimize cloud resource utilization while minimizing financial waste and environmental impact. By incorporating financial observability, resource allocation strategies, and sustainability metrics into the FL orchestration layer, the proposed approach enables dynamic cost-performance tradeoff analysis and supports environmentally responsible model training. To empirically validate the framework, a series of controlled experiments is conducted using the CIFAR10 dataset for image classification tasks. The experiments are deployed across the three major public cloud providers: Google Cloud Platform (GCP), Microsoft Azure, and Amazon Web Services (AWS). The results demonstrate that applying FinOps principles to multicloud FL setups improves financial transparency and control, while significantly enhancing resource efficiency and environmental sustainability. This work provides a concrete path toward scalable, cost-aware, and green federated learning.