This study compares two generations of the IPSL-EPOC decadal climate prediction system based on the IPSL-CM6A-LR Earth System Model, focusing on differences in ocean data assimilation. The first configuration (DCP1), developed for CMIP6-DCPP, is carbon concentration-driven, while the second (DCP2), operational since 2023 under the World Meteorological Organization (WMO), incorporates updated observations and a fully interactive carbon cycle. Both systems assimilate sea surface temperature (SST) and salinity (SSS), but differ in data sources and spatial coverage: DCP1 uses the Atlantic-focused Reverdin dataset, whereas DCP2 relies on the global EN4 dataset with climatological filling in data-sparse regions. Over the common 1980–2014 period, DCP2 improves SST and SSS initialization and enhances predictive skill, particularly for ocean temperatures during the first five forecast years. Improvements are most evident in the North Atlantic, tropical Pacific, and Southern Ocean. However, skill declines beyond 5 years, especially for large-scale indices such as the Atlantic Multidecadal Variability (AMV). This loss of skill is partly linked to the use of climatological SSS in EN4 in gridpoints where observations are missing, which appears to damp ocean variability, reduce deep-water variability, and limit fluctuations of the Atlantic Meridional Overturning Circulation (AMOC). These results highlight the need for careful treatment of sparse observational data in decadal prediction systems to preserve the low-frequency variability of ocean circulation.