The increasing adoption of Model-Driven Engineering in in- dustrial contexts highlights scalability as a critical limitation of several MDE tools. Most of the current model-to-model transformation engines have been designed for one-shot translation of input models to output models, and present efficiency issues when applied to very large models. In this paper, we study the application of a lazy-evaluation approach to model transformations. We present a lazy execution algorithm for ATL, and we empirically evaluate a prototype implementation. With it, the elements of the target model are generated only when (and if) they are accessed, enabling also transformations that generate infinite target models. We achieve our goal on a significant subset of ATL by extending the ATL compiler.