Our understanding of the dynamics of plate motions is based almost entirely upon modeling of present-day plate motions. A fuller understanding, however, can be derived from consideration of the history of plate motions. Here we investigate the kinematics of the last 120 Myr of plate motions and the dynamics of Cenozoic motions, paying special attention to changes in the character of plate motions and plate-driving forces. We analyze the partitioning of the observed surface velocity field into toroidal (transform/spin) and poloidal (spreading/subduction) motions. The present-day field is not equipartitioned in poloidal and toroidal components; toroidal motions account for only one third of the total. The toroidal/poloidal ratio has changed substantially in the last 120 Myr with poloidal motion decreasing significantly after 43 Ma while toroidal motion remains essentially constant; this result is not explained by changes in plate geometry alone. We develop a self-consistent model of plate motions by (1) constructing a straightforward model of mantle density heterogeneity based largely upon subduction history and then (2) calculating the induced plate motions for each stage of the Cenozoic. The slab heterogeneity model compares rather well with seismic heterogeneity models, especially away from the thermochemical boundary layers near the surface and core-mantle boundary. The slab model predicts the observed geoid extremely well, although comparison between predicted and observed dynamic topography is ambiguous. The midmantle heterogeneities that explain much of the observed seismic heterogeneity and geoid are derived largely from late Mesozoic and early Cenozoic subduction, when subduction rates were much higher than they are at present. The plate motion model itself successfully predicts Cenozoic plate motions (global correlations of 0.7--0.9) for mantle viscosity structures that are consistent with a variety of geophysical studies. We conclude that the main plate-driving forces come from subducted slabs (>90%), with forces due to lithospheric effects (e.g., oceanic plate thickening) providing a very minor component (<10%). For whole mantle convection, most of the slab buoyancy forces are derived from lower mantle slabs. Unfortunately, we cannot reproduce the toroidal/poloidal partitioning ratios observed for the Cenozoic, nor do our models explain apparently sudden plate motion changes that define stage boundaries. The most conspicuous failure is our inability to reproduce the westward jerk of the Pacific plate at 43 Ma implied by the great bend in the Hawaiian-Emperor seamount chain. Our model permits an interesting test of the hypothesis that the collision of India with Asia may have caused the Hawaiian-Emperor bend. However, we find that this collision has no effect on the motion of the Pacific plate, implying that important plate boundary effects are missing in our models. Future progress in understanding global plate motions requires (1) more complete plate reconstruction information, including, especially, uncertainty estimates for past plate boundaries, (2) better treatment of plate boundary fault mechanics in plate motion models, (3) application of numerical convection models, constrained by global plate motion histories, to replace ad hoc mantle heterogeneity models, (4) better calibration of these heterogeneity models with seismic heterogeneity constraints, and (5) more comprehensive comparison of global plate/mantle dynamics models with geologic data, especially indicators of intraplate stress and strain, and constraints on dynamic topography derived from the stratigraphic record of sea level change. ¿ 1998 American Geophysical Union