In dynamic nuclear polarization, polarization is transferred from electron to nuclear spins at low temperatures. So far, it is not possible to predict the relaxation of hyperpolarized solids, and applications are developed largely by trial and error. Here, we study the low-field nuclear spin relaxation in pyruvic acid, doped with trityl. We find that, at low fields, the relaxation time constants for both 13 C and 1 H scale linearly with the applied field. We model the data using a thermodynamic approach, in which heat is transferred via triple-spin-flips involving two electron spins and one nuclear spin. The triple-spin-flip rate is calculated from first principles using a formalism developed by Wenckebach. The heat capacity of the Non-Zeeman reservoir is determined from the 1 H relaxation data at intermediate fields, which leads to a parameter-free, yet nearly quantitative description of the observed 1 H relaxation rates from 5 mT to 2 T. The observed 13 C relaxation is consistent with the formalism, provided that the direct energy exchange between the 1 H and 13 C reservoirs, and the diffusion barrier are accounted for at low and high fields, respectively. The formalism also describes the observed electron-mediated 1 H-13 C thermal mixing.