Considerable uncertainty and debate exist in projecting the future capacity of forests to sequester atmospheric CO2. Here we generate spatially explicit patterns of biomass loss by mortality (LOSS) from a dataset (n = 2676) of long-term (1951 to 2018), largely unmanaged forest plots to constrain projected (2015–2099) net primary productivity (NPP), heterotrophic respiration (HR) and net carbon sink in six dynamic global vegetation models (DGVMs) across North and South America, Africa, Asia, and Australia. This approach was motivated by the higher accuracy of field-derived LOSS estimates and the strong relationship among LOSS, NPP, and HR at continental or biome scales. The field observations showed higher values of LOSS in tropical regions (0.53 Kg m− 2 y− 1) than in North America (0.22 Kg m− 2 y− 1). The upscaled gridded LOSS map show hotspots in Southern Asia & Australia, Northwestern South America, and the western coast of North America. The DGVMs overestimated LOSS, particularly in tropical regions and eastern North America by as much as 0.5 Kg m2− y1. The spread of DGVM-projected NPP and HR uncertainties was substantially reduced after constraining the simulations with the observed LOSS patterns. The observation-constrained models show a decrease in the tropical forest carbon sink by the end of the century, particularly across South America (from 2 to 1.3 PgC y1), and an increase in the sink in North America (from 0.75 to 0.97 PgC y1). These results suggest that forest demographic data can be used to constrain land carbon sink projections.