Steel structures are prone to inelastic local instabilities that cause deterioration in their strength and stiffness. To inform decisions on whether deformed steel members can remain in service in the aftermath of extreme events, this work develops a deep learning-based framework that quantifies the reserve capacity of hot-rolled wide-flange steel beams directly from point cloud data capturing their deformed shapes due to local buckling. A dataset of wide-flange steel beams under monotonic and cyclic loading is generated via physics-based continuum finite element simulations, comprising 14,921 point cloud samples with reserve capacity annotations. The primary failure mode of the steel beams is inelastic local buckling. Geometric feature engineering, including curvature-based descriptors sensitive to local buckling, is used to enhance the structural representation. Several graph neural network architectures are evaluated on the task of predicting the reserve capacity, demonstrating reliable estimates of the reserve capacity with a mean prediction error of only 0.6% for beams with reserve capacities above 40% relative to their peak resistance. To enable reliable decisions regarding structural repairs, the predictive uncertainty is quantified by modeling both aleatoric and epistemic components: aleatoric uncertainty is learned through an auxiliary variance prediction, and epistemic uncertainty is captured using an ensemble of independently trained models. Overall, the presented work demonstrates that the reserve capacity of steel beams can be accurately inferred from geometric data alone, paving the way for automated, data-driven assessment methods that facilitate the inspection of steel structures.