<p>Machine-learning interatomic potentials have demonstrated power-law scaling in predictive accuracy as training data and model capacity increase, but it remains unclear whether models trained at scale acquire interpretable chemical concepts. Here we show that an E(3)-equivariant machine-learning interatomic potential learns local bond information without direct supervision of bond properties. To expose this information, we develop an edge-wise emergent energy-decomposition framework and apply it to an Allegro neural-network potential trained on SPICE2, a dataset composed of stable molecular structures. The framework analyzes edge-wise energy contributions obtained from the trained model and their distributions together with information entropy to examine how data size, data composition and model-training scenarios shape the internal representation of chemical bonds. The resulting bond-dissociation energy estimates for archetypal bond types agree quantitatively with literature values and are consistent across models trained on organic and inorganic datasets. We further examine a hybrid training set and find that combining complementary chemical data improves transition-state energy-prediction accuracy while reshaping the learned bond representations. These results indicate that scalable interatomic potentials can acquire transferable bond concepts without explicit bond-energy labels, providing a way to analyze the transition-state energy-prediction problem.</p>