I investigate labour earnings in the top tail (± top 12%) of the South African income distribution from 1995 to 2007, using a new harmonised data set constructed from the OHSs and LFSs (Kerr et al 2011). Nonparametric techniques suggest that the distribution is well approximated by a Pareto distribution. Surprisingly, this distribution seems to have been remarkably stable over the entire period. Parametric estimates suggest that the tail parameter is around 1.8, which suggests that the distribution is “fat tailed”. This implies that extreme outcomes are more common than with the standard “normal” distribution. I discuss some of the implications of such fat tails for the way we think about inequality.