The average predictability time (APT) method is used to identify the most predictable components of decadal sea surface temperature (SST) variations over the Southern Ocean (SO) in a 4000-yr unforced control run of the GFDL CM2.1 model. The most predictable component shows significant predictive skill for periods as long as 20 years. The physical pattern of this variability has a uniform sign of SST anomalies over the SO, with maximum values over the Amundsen–Bellingshausen–Weddell Seas. Spectral analysis of the associated APT time series shows a broad peak on time scales of 70–120 years. This most predictable pattern is closely related to the mature phase of a mode of internal variability in the SO that is associated with fluctuations of deep ocean convection. The second most predictable component of SO SST is characterized by a dipole structure, with SST anomalies of one sign over the Weddell Sea and SST anomalies of the opposite sign over the Amundsen–Bellingshausen Seas. This component has significant predictive skill for periods as long as 6 years. This dipole mode is associated with a transition between phases of the dominant pattern of SO internal variability. The long time scales associated with variations in SO deep convection provide the source of the predictive skill of SO SST on decadal scales. These analyses suggest that if the SO deep convection in a numerical forecast model could be adequately initialized, the future evolution of SO SST and its associated climate impacts are potentially predictable.