Dryland ecosystems occupy ~40% of the land surface and are thought to dominate the inter-annual variability (IAV) and long-term trend of the global carbon (C) cycle. Therefore, it is imperative that global terrestrial biosphere models (TBMs), which form the land component of IPCC earth system models, are able to accurately simulate dryland vegetation and biogeochemical processes. However, compared to more mesic ecosystems, TBMs have not been widely tested or optimized against in situ dryland ecosystem CO2 fluxes. Here, we address this gap using a Bayesian data assimilation system and 89 site-years of daily net CO2 flux (net ecosystem exchange - NEE) data from 12 southwest US Ameriflux sites spanning forest, shrub and grass dryland ecosystems to evaluate and optimize the C cycle related parameters of the ORCHIDEE TBM. We find that the default (prior) model simulations drastically underestimate both the mean annual NEE and the NEE IAV. By testing different assimilation scenarios, we showed that optimizing phenology parameters dramatically improves the model ability across all sites to capture both the magnitude and sign of the NEE IAV. At high elevation forested sites, which are a mean C sink, optimizing parameters related to C allocation, respiration and turnover reduces the underestimate in simulated mean annual NEE. Our study demonstrates that all TBMs need to be calibrated specifically for dryland ecosystems before they are used to determine dryland contributions to global C cycle variability and long-term carbon-climate feedbacks.