Abstract Supercontinent cycles dictate Earth’s long‐term tectonic and magmatic evolution, yet the variations in the lithologies of the mantle sources of global basalts remain poorly constrained. Tracking the balance between these two source lithologies, peridotite and pyroxenite, is notoriously difficult as extensive fractional crystallization obscures primary mantle signatures in global basalt records. Here, we present a machine learning framework trained using experimental and thermodynamically modeled primary melts to directly distinguish peridotite‐ and pyroxenite‐derived magmas. Applying this method to a global basalt data set spanning the past 3 Ga shows that the source lithologies of global basalts shift cyclically with supercontinent cycles, with pyroxenite contributions peaking during the assembly stages and peridotite contributions dominating during the breakup stages. These shifts coincide with variations in lithospheric thickness and δ18O of continental crust, pointing to a persistent association between the supercontinent cycle, lithospheric dynamics, and the sources of global magmatism throughout Earth’s geological history.