This paper investigates the trade-off between model performance and energy efficiency in supervised learning by varying the size of the training data, motivated by the increasing energy consumption of machine learning models. We establish empirical energy scaling laws for neural network training and develop optimization frameworks for balancing accuracy (the percentage of correctly classified samples on the test set) and energy consumption through stratified random sampling. We demonstrate that energy consumption per training epoch scales linearly with dataset size for CNNs, with architecture-dependent coefficients ranging from \(2.143 \times 10^{-5}\) kJ/sample (MNIST) to \(1.356 \times 10^{-4}\) kJ/sample (CIFAR-10). Critically, we show that model complexity, not dataset characteristics, determines energy scaling patterns. Our comprehensive analysis across MNIST, Fashion-MNIST, and CIFAR-10 reveals optimal subset sizes for energy efficiency: 5% for MNIST and Fashion-MNIST and 24% for CIFAR-10 when maximizing accuracy per unit energy consumed. We extend these findings to federated learning with 20 clients, validating the framework’s generalizability across distributed training scenarios. For practitioners targeting specific accuracy thresholds, we provide energy budgeting strategies—for example, achieving 80% accuracy on MNIST requires only 7.5% of the data while saving 82% energy. The derived energy models enable solving three classes of optimization problems: minimizing energy subject to accuracy constraints, maximizing accuracy within energy budgets, and optimizing efficiency metrics. This work provides actionable guidance for sustainable AI development and establishes a methodology for deriving energy scaling laws for new datasets and architectures.