Label Propagation (LP) is a classic method in semi-supervised learning, where labels diffuse across graph edges until convergence. Its main obstacle on large dense graphs is scalability: classical LP requires an n x n adjacency, with quadratic memory and runtime costs. We focus on the common class of low-rank graphs, where the adjacency has the form VV⊤ for an embedding matrix V ∈ ℝ> n x d. Such graphs arise naturally in applications including similarity graphs from embeddings, recommender systems, kernel methods, and dense affinity graphs in vision and biology. We introduce VLP (Vector-Based Label Propagation), which operates entirely in the embedding space without explicit edges. VLP reduces memory from O (n2) to O (nd), runs efficiently on GPUs, and is mathematically equivalent to classical LP. We further give the first convergence proof of LP in this low-rank setting. VLP makes label propagation feasible for graphs with millions of nodes, far beyond the reach of traditional methods.