Quantifying Cloud Cost Efficiency in Federated Learning: An Empirical Comparison with Centralized Training

Machine learning has become an important component of data driven applications. Most large-scale machine learning jobs are now executed on cloud platforms. These deployments often face high bandwidth and storage cost issues which can overside computation expenses. The present study investigates whether FL (Federated Learning) can overcome such costs without compromising predictive performance. A controlled experiment was conducted using two clients and ten communication rounds on the California Housing dataset to compare a centralized configuration with a federated one. In the FL set up, only model parameters were exchanged thus eliminating repeated data transfer. Realistic cloud pricing models for data transfer, compute and temporary storage were applied to recorded execution statistics. Results showed that centralized training consumed nearly 5,000 × more network bandwidth than FL while producing similar accuracy (MSE = 0.542 vs 0.556; R2 = 0.587 vs 0.576). Under limited-bandwidth conditions and when local data are available, FL substantially lowers operating expenses without compromising on model quality. The study also highlights deployment variables such as round budgeting, update sizing, and client selection that influence overall efficiency. These findings indicate that federated learning which was popular for its privacy preservation is a practical, economically feasible solution for cloudbased machine learning systems.