This study proposes a two-phase research framework to address the problem of capacity-demand mismatch in the high-tech industry. In the first stage, due to the characteristics of small data, we apply the virtual data generation process (VDGP) to support the data learning for demand forecast. In the second stage, based on the demand scenarios, a robust capacity decision is provided by the stochastic programming (SP) technique and the minimax regret (MMR) technique addressing demand uncertainty. We conduct an empirical study of a TFT-LCD firm to validate the proposed framework. That result shows that the proposed framework, in particular the SP technique, provides a robust capacity levels addressing the problem of capacity-demand mismatch.