Existing learning-based surface reconstruction methods from point clouds are still facing challenges in terms of scalability and preservation of details on point clouds of large scales. In this paper, we propose the TSRNet, a novel scalable learning-based method for surface reconstruction. It first takes a point cloud and its related octree vertices as input and learns to classify whether the octree vertices are in front or at back of the implicit surface. Then the Marching Cubes (MC) is applied to extract a surface from the binary labeled octree. In our method, we design a scalable learning-based pipeline for surface reconstruction. It does not consider the whole input data at once. It allows to divide the point cloud and octree vertices and to process different parts in parallel. Our network captures local geometry details by constructing local geometry-aware features for octree vertices. The local geometry-aware features enhance the predication accuracy greatly for the relative position among the vertices and the implicit surface. They also boost the generalization capability of our network. Our method is able to reconstruct local geometry details from point clouds of different scales, especially for point clouds with millions of points. More importantly, the time consumption on such point clouds is acceptable and competitive. Experiments show that our method achieves a significant breakthrough in scalability and quality compared with state-of-the-art learning-based methods.