We present a heuristic model to quantitatively explain the suppression of deep convection in convection-resolving models (CRMs) with small domains. We distinguish between “computational” smallness (few grid columns) and “physical” smallness (representing a small geographic area). Domains that are computationally small require greater instability to sustain convection because they force a large convective fraction, driving strong compensating subsidence warming. Consequently, detrainment occurs lower for undiluted convection. Both computationally and physically small domains limit the physical updraft width, increasing entrainment dilution. This enhancement of entrainment strengthens the sensitivity to domain size beyond that for undiluted deep convection. Coarsening grid spacing to expand the physical domain and physical updraft width can reduce domain size sensitivity. Simulations using the System for Atmospheric Modeling (SAM) confirm the heuristic model results. We also present simulation results for two shallow convection cases, which are less sensitive to domain size, but also exhibit sensitivities.