Predicting terrestrial carbon, C, budgets and carbon‐climate feedbacks strongly relies on our ability to accurately model interactions between vegetation, C and water cycles, and the atmosphere. However, C fluxes simulated by global, process‐based terrestrial biosphere models (TBMs) remain subject to large uncertainties, partly due to unknown or poorly calibrated parameters. This is because TBMs have not routinely been confronted against C cycle related datasets within a statistical data assimilation (DA) system. In this review, we present 15 years' development of a C cycle DA system for optimizing C cycle parameters of the ORCHIDEE TBM. We analyze the impact of assimilating multiple different C cycle related datasets on regional to global‐scale gross and net CO2 fluxes. We find that assimilating atmospheric CO2 data is crucial for improving (increasing) ORCHIDEE predictions of the terrestrial land C sink. The improvement is predominantly due to the global‐scale constraint these data provide for optimizing initial soil C stocks, which are likely in error due to inaccurate assumptions about steady state spin‐up and incomplete knowledge of land use change histories. When comparing the data‐constrained ORCHIDEE land C sink estimates to the CAMS atmospheric inversion, we show that while the two approaches agree on the global C sink magnitude, they continue to differ in how the global C sink is partitioned between the northern hemisphere and tropics. We also discuss technical challenges faced in our C cycle DA studies, in particular the difficulty in characterizing the error covariance matrix due to unknown observation biases and/or model‐data inconsistencies. We offer our perspectives on how to tackle these challenges that we hope can serve as a roadmap for other TBM groups wishing to develop C cycle DA systems. Key Points Considerable progress has been made in constraining ORCHIDEE terrestrial biosphere model regional to global scale CO2 fluxes within a data assimilation system Results highlight the importance of optimizing initial C stocks ‐ in addition to C cycle related parameters ‐ using global‐scale datasets Challenges remain in utilizing the wide variety of available data, particularly when characterizing the observation error covariance matrix