There is a large literature on state of health (SOH) analysis that aims to be used for predicting the remaining useful life (RUL) of in-service commercial Li ion batteries. These studies generally rely on libraries of off-line experiments to correlate battery operating parameters to RUL. However, the accuracy of these predictions must ultimately be limited by the statistical variation of RUL among nominally identical cells. There are relatively few publications that assess capacity decline in enough commercial cells to adequately quantify this variation (e.g. >10 cells), but those that do show a cell-to-cell variability in capacity fade that is surprisingly wide. Furthermore, capacity vs cycle number curves generally cross each other many times in large samples, a challenge to RUL predictions. In this work we analyze the capacity fade statistics for 24 nominally identical commercial pouch cells, finding a wide cell-to-cell variability that is qualitatively similar to that seen in other work. Our data indicate that RUL predictions based on remaining capacity can be reasonably accurate only after a large number of failures have already been observed. Analysis of our failure data with normal and with 2- and 3-parameter Weibull probability density functions provide uniformly good fits using a variety of definitions of failure, although we argue against using a 3-parameter Weibull function for our data. We see preliminary indications in our data that the pdf fitting parameters asymptotically approach constant values. We suggest a strategy whereby testing continues until a predesignated confidence level for failure probability has been achieved, rather than using a predesignated number of cycles (Type I censoring) or a predesignated number of failures (Type II censoring). An increased emphasis on making batteries whose lifetimes are more reproducible could lead to improvements in battery cost and safety.