Theoretical models aimed at describing magic-angle-spinning (MAS) dynamic nuclear polarization (DNP) NMR typically face a trade-off between the scientific rigor obtained with a strict quantum mechanical description, and the need for using realistically large spin systems, for instance using phenomenological models. Thus far, neither approach has accurately reproduced experimental results, let alone achieved the generality required to act as a reliable predictive tool. Here, we show that the use of aggressive state-space restrictions and an optimization strategy allows full-scale ab initio MAS-DNP simulations of spin systems containing thousands of nuclei. Our simulations are the first ever to achieve quantitative reproduction of experimental DNP enhancements and their MAS rate dependence for both frozen solutions and solid materials. They also revealed the importance of a previously unrecognized structural feature found in some polarizing agents that helps minimize the sensitivity losses imposed by the spin diffusion barrier.