Graph self-supervised learning has attracted plenty of attention in recent years. However, most existing methods are designed for homogeneous graphs yet not tailored for bipartite graphs, and their objectives could induce cluster-level errors since they only consider instance-wise topological information. In this paper, we introduce a novel co-cluster infomax (COIN) framework to capture the cluster-level information by maximizing the mutual information of co-clusters. Different from previous infomax methods which estimate mutual information, COIN directly calculates mutual information. Besides, COIN is an end-to-end method which can be trained jointly with other objectives. Furthermore, we theoretically prove that COIN could effectively increase the mutual information of node embeddings and it is upper-bounded by the prior distributions of nodes. Experimental results show that COIN outperforms state-of-the-art methods on various downstream tasks.