Feeding systems worldwide use metabolizable protein as the unifying unit for computing protein availability in feeds and the protein requirements for various physiological functions. Metabolizable protein is the sum of microbial true protein synthesized in the rumen from RDP and endogenous sources and RUP. In practical feeding situations, these quantities are seldom measured, so mathematical models are needed to predict the amount of ruminally synthesized microbial protein. The objective of this study was to develop a model to predict microbial protein production in the rumen using a more mechanistic and generalizable approach by representing the supply of both RDP and rumen-degraded carbohydrates. A bisubstrate Michaelis-Menten functional form was used to associate the amounts of rumen-degraded NDF (RDNDF) and starch (RDST) with 2 Michaelis-Menten constants and the amount of RDP with the asymptotic response (Vmax). The mechanistic biological reasoning was that, from a mass balance standpoint, the amount of RDP determines the limiting microbial protein produced while the amount of degraded carbohydrates are associated with the affinity (Km) of substrate used for carbon precursors in cellular growth (derivation of cell constituents for dividing cells). The equation is an efficiency derivation with actual microbial N, and actual RDNDF and RDST derived using diet-level prediction factors (White et al., 2017; J. Dairy Sci. 100: 3611–3627). The model was fitted with 583 treatment means from 154 publications using a Bayesian nonlinear hierarchical modeling approach. The fitted model was Microbial N (g/d) = (117.0 + 0.071 × RDP)/[(1 + 0.043/RDNDF)(1 + 0.051/RDST)] where RDP is in g/d and both RNDF and RDST are in kg/d. The Bayesian hierarchical approach allowed successful introduction of between study variability in all model parameters. Model evaluation suggested low slope and mean bias (<2% of MSE). Likewise, a 5-fold cross-validation suggested good predictive ability (Concordance Correlation Coefficient, CCC = 0.52) and relatively low root mean square prediction error (29% of mean) when compared against models from the literature.