The spectrum of dileptons produced by the quark-gluon plasma in an ultrarelativistic nucleus-nucleus collision depends only, to a good approximation, on the transverse mass M_t of the dilepton. This scaling is exact as long as transverse flow is negligible, and the system is in local thermal equilibrium. We implement a state-of-the-art modelization of kinetic and chemical equilibration in the early stages of the evolution to study the modifications of the spectrum. Violations of M_t scaling resulting from these effects are evaluated as a function of the shear viscosity to entropy ratio (eta/s) that controls the equilibration time. We determine the dependence of the spectrum on system size, centrality, rapidity, and collision energy. We show that the quark-gluon plasma produces more dileptons than the Drell-Yan process up to invariant masses of order M = 4 GeV. Due to different kinematics, for a given M_t , the dependence of the dilepton yield on M is opposite for the two processes, so that experiment alone can in principle determine which process dominates.