We consider a random walk on a homogeneous space $G/Λ$ where $G$ is a non-compact simple Lie group and $Λ$ is a lattice. The walk is driven by a probability measure $μ$ on $G$ whose support generates a Zariski-dense subgroup. We show that the random walk equidistributes towards the Haar measure unless it is trapped in a finite $μ$-invariant set. Moreover, under arithmetic assumptions on the pair $(Λ, μ)$, we show the convergence occurs at an exponential rate, tempered by the obstructions that the starting point may be high in a cusp or close to a finite orbit. The main challenge is to show that the dimensional properties of a given probability distribution on $G/Λ$ improve under convolution by $μ$. For this, we develop a new method, which combines a dimensional interpolation result and a dimensional increase alternative. This approach allows us to bypass inherent geometric obstructions. To show dimensional interpolation, we establish a general subcritical projection theorem under optimal non-concentration assumptions on the projector, and a corresponding submodular inequality for irreducible representations which allows its application to random walks. Both are of independent interest. The dimensional increase alternative aligns with the spirit of Bourgain's projection theorem. It is fine-tuned for random walks and has the advantage of being valid in situations lacking transversality.