When Does Decentralization Pay Off? An Empirical and Simulation Study of Data Mesh Operating Models
Whether a large organization should adopt data mesh hinges on how data dependencies between its domains are structured and grow, yet this structure has not been measured and existing evidence on data mesh is almost entirely qualitative. We mine 566 public infrastructure manifests of Norway's labour and welfare administration (NAV) to reconstruct its declared cross-team data-dependency graph and five years of its growth. The graph is sparse (3.7% of team pairs) and hub-concentrated (degree Gini 0.53 versus 0.24 in matched random graphs); edges grew quadratically at a stable low density (a = 2.03, 95% CI [1.99, 2.31]), contradicting our hypothesis of sub-quadratic growth, while catalogued data products grew sub-quadratically (a = 1.37). We then prove that a central data team's capacity is bounded under communication overhead regardless of headcount, that mesh stability is governed by the most connected domains, and that relabeling teams as domains without transferring ownership can never beat centralization. A verified discrete-event simulation of central, mesh, "theater" and hub-and-spoke operating models at equal headcount, pre-registered before any comparative output, shows that the all-pairs dependency assumption penalizes the central model rather than the mesh, and that the mesh's delivery advantage depends chiefly on diseconomies of scale in central teams: without them, the two models deliver within 8% of each other. Cross-domain demand, dependency structure and reuse matter far less. Data, code and pre-registration are public.