MOTIVATION: Phylogenetic placement of a query sequence on a backbone tree is increasingly used across biomedical sciences to identify the content of a sample from its DNA content. The accuracy of such analyses depends on the density of the backbone tree, making it crucial that placement methods scale to very large trees. Moreover, a new paradigm has been recently proposed to place sequences on the species tree using single-gene data. The goal is to better characterize the samples and to enable combined analyses of marker-gene (e.g., 16S rRNA gene amplicon) and genome-wide data. The recent method DEPP enables performing such analyses using metric learning. However, metric learning is hampered by a need to compute and save a quadratically growing matrix of pairwise distances during training. Thus, the training phase of DEPP does not scale to more than roughly 10 000 backbone species, a problem that we faced when trying to use our recently released Greengenes2 (GG2) reference tree containing 331 270 species. RESULTS: This paper explores divide-and-conquer for training ensembles of DEPP models, culminating in a method called C-DEPP. While divide-and-conquer has been extensively used in phylogenetics, applying divide-and-conquer to data-hungry machine-learning methods needs nuance. C-DEPP uses carefully crafted techniques to enable quasi-linear scaling while maintaining accuracy. C-DEPP enables placing 20 million 16S fragments on the GG2 reference tree in 41 h of computation. AVAILABILITY AND IMPLEMENTATION: The dataset and C-DEPP software are freely available at https://github.com/yueyujiang/dataset_cdepp/.