It is usually considered that evolutionary algorithms are highly parallel. In fact, the theoretical speed-ups for parallel optimization are far better than empirical results; this suggests that evolutionary algorithms, for large numbers of processors, are not so efficient. In this paper, we show that in many cases automatic parallelization provably provides better results than the standard parallelization consisting of simply increasing the population size lambda. A corollary of these results is that logarithmic bounds on the speed-up (as a function of the number of computing units) are tight within constant factors. Importantly, we propose a simple modification, termed log(lambda)-correction, which strongly improves several important algorithms when lambda is large.