In this paper we study the nerves of two types of coverings of a sphere $S^{d-1}$: (1) coverings by open hemispheres; (2) antipodal coverings by closed hemispheres. In the first case, nerve theorem implies that the nerve is homotopy equivalent to $S^{d-1}$. In the second case, we prove that the nerve is homotopy equivalent to a wedge of $(2d-2)$-dimensional spheres. The number of wedge summands equals the Möbius invariant of the geometric lattice (or hyperplane arrangement) associated with the covering. This result explains some observed large-scale phenomena in topological data analysis. We review the particular case, when the coverings are centered in the root system $A_d$. In this case the nerve of the covering by open hemispheres is the space of directed acyclic graphs (DAGs), and the nerve of the covering by closed hemispheres is the space of non-strongly connected directed graphs. The homotopy types of these spaces were described by Björner and Welker, and the incarnation of these spaces appeared independently as "the poset of orders" and "the poset of preorders" respectively in the works of Bouc. We study the space of DAGs in terms of Gale and combinatorial Alexander dualities, and propose how this space can be applied in automated machine learning.