In both academia and industry, the multiple-input multiple-output (MIMO) techniques have shown enormous gains in spectral efficiency by exploiting the spatial degrees of freedom. So far, an underlying assumption in most of existing multi-user multiple-input multiple-output (MU-MIMO) design has been that all the users in the system use infinite blocklength, so that they can achieve the Shannon capacity. This setup, however, is not suitable considering the presence of delay-constrained users whose the blocklength tends to be finite. In this paper, we consider a heterogeneous setting in a MU-MEMO system where delay tolerant users and delay constrained users coexist. To maximize the sum spectral efficiency in this system, we first present the spectral efficiencies for delay-tolerant user and delay-constrained user as the Rayleigh quotient. We then derive a first-order optimality condition of the optimization problem that maximizes the sum spectral efficiency of total users and satisfies the latency requirement of delay-constrained users, and propose the generalized power iteration preceding algorithm. In the simulation results, we prove that the proposed method outperforms baseline schemes.