We develop a characteristic-based asset demand model in which cross-asset risk-return trade-offs vary with asset characteristics. The model relaxes the uniform substitution structure of the multinomial logit (MNL), accommodates large price elasticities, and enables recovery of investor-specific primitives, including alphas and factor loadings, from structural demand estimates. Applied to U.S. institutional equity holdings from 2000 to 2022, the model reveals meaningful deviations from MNL substitution patterns, particularly along the market equity dimension. The estimated average own-price elasticity is 77 percent higher than under the MNL, driven largely by investors whose portfolios imply cross-asset complementarity. Nonetheless, both elasticity estimates are substantially lower than those implied by CAPM calibrations. The model also uncovers heterogeneity in investor alphas: hedge funds earn near-zero alphas, while brokers earn up to five basis points annually.