We use longitudinal microdata to estimate the urban density premium for U.S. establishments, controlling for observed establishment characteristics and dynamic establishment behavior. Consistent with previous studies, we estimate a density premium between 6 and 10 percent, even after controlling for establishment composition, local skill mix, and the endogeneity of location choice. More importantly, we find that the estimated density premium is realized almost entirely at birth and is constant over the life of establishments. We find little evidence that the endogenous entry or exit of establishments can account for any of the estimated density premium. We interpret our results as implying that the returns to agglomeration diffuse within a city through a reallocation channel rather than through an increase in the productivity of existing firms.