A network of $N$ elements is studied in terms of a deterministic globally coupled map which can be chaotic. There exists a range of values for the parameters of the map where the number of different macroscopic configurations is very large, and there is violation of selfaveraging. The time averages of functions, which depend on a single element, computed over a time $T$, have probability distributions that do not collapse to a delta function, for increasing $T$ and $N$. This happens for both chaotic and regular motion, i.e. positive or negative Lyapunov exponent.