Recommender systems are increasingly evaluated not only by recommendation quality, but also by their computational and environmental footprint. This work investigates whether data reduction can lower the energy consumption of recommender systems without disproportionately degrading performance. We compare random, temporal and structure-aware coreset strategies across three representative recommendation algorithms. Energy consumption is measured separately for data preprocessing and model training. Our results show that the sustainability benefits of data reduction depend strongly on both the recommendation model and the data reduction strategy. Reducing the training data can lower energy use while preserving competitive performance, but poorly chosen reductions may substantially harm accuracy or user coverage. Evaluation also shows that data reduction cannot be assumed to be cost-free. The energy required to construct a reduced dataset may offset part of the savings achieved during model training. These findings highlight the need to evaluate data reduction from a broader system-level perspective that jointly considers recommendation quality, user coverage, preprocessing overhead, and training costs.