Abstract: Fast and accurate simulation of nonlinear seismic responses to urban building clusters under batch earthquake loading is of great importance for post-earthquake regional loss assessment and rapid emergency response. Intelligent algorithms are a novel and effective solution. However, they still have prohibitively high training costs when applied to building clusters at the scale of hundreds of thousands. To address this issue, a “reduction–prediction” scheme is proposed. Large-scale building clusters are first reduced to a small set of representative buildings using an adaptive distance-threshold mechanism. A machine learning model is then applied to efficiently predict the nonlinear seismic responses of the building clusters. Case studies show that the Mahalanobis distance effectively characterizes sample similarity in a feature space based on Principal Component Analysis. Combined with the Weighted Maximum Mean Discrepancy, it enables adaptive selection of the representative building set. The proposed “compression–prediction” framework achieves fast seismic nonlinear response prediction for building clusters at the scale of hundreds of thousands, with a mean absolute error of 2.23% and a 99% reduction in computation time, providing a practical reference for rapid seismic damage assessment and emergency decision-making.