Gross primary productivity (GPP) is the largest carbon flux in the Earth system, playing a crucial role in supporting ecosystem metabolism and sequestering atmospheric carbon dioxide (CO₂). However, existing GPP estimates present significant uncertainties and discrepancies, partly due to the underrepresentation of the CO₂ fertilization effect, which has enhanced the terrestrial carbon sink over recent decades. We introduce CEDAR (sCaling Ecosystem Dynamics with ARtifical intelligence) -GPP, the first global upscaled GPP dataset integrating the direct CO2 fertilization effect on photosynthesis. CEDAR-GPP was generated by synthesizing globally distributed eddy covariance measurements of carbon fluxes, multi-source satellite observations, and climate variables using machine learning models. Our models effectively predicted monthly GPP (R² ~ 0.74), the mean seasonal cycles (R² ~ 0.79), and spatial variabilities (R² ~ 0.67). Importantly, incorporating direct CO2 effects substantially improved the models’ capability to reproduce long-term GPP trends across global flux sites. While aligning with existing satellite-based products on the global patterns of annual mean GPP, seasonality, and interannual variability, CEDAR-GPP shows elevated long-term trends, particularly in the tropics. This reflects a strong temperature control on the response of light use efficiency to CO₂. Available at a 0.05º monthly resolution from 1982 to 2020, CEDAR-GPP offers a comprehensive representation of global GPP spatiotemporal dynamics, providing valuable insights into ecosystem-climate interactions.