Abstract: Microbes form complex multi-species communities that play many critical roles across the biosphere. Building quantitative models that may help us predict the composition and function of a microbial community in a given habitat is a long-standing aspiration. To this end, we must understand what features of microbial communities (and at what level of description) are predictable, which are not, and why. We have addressed this question through a combination of quantitative experiments and simple models. Experimentally, we cultivate large numbers of natural communities in environments with a defined nutrient composition. Our experiments indicate that the coarse-grained metabolic organization of microbial communities is reproducible in identical habitats, following simple equations that are consistent over ecological and evolutionary timescales, despite substantial variability at finer levels of resolution.