Recent advances have made it feasible to apply the stochastic variational paradigm to a collapsed representation of latent Dirichlet allocation (LDA). While the stochastic variational paradigm has successfully been applied to an uncollapsed representation of the hierarchical Dirichlet process (HDP), no attempts to apply this type of inference in a collapsed setting of non-parametric topic modeling have been put forward so far. In this paper we explore such a collapsed stochastic variational Bayes inference for the HDP. The proposed online algorithm is easy to implement and accounts for the inference of hyper-parameters. First experiments show a promising improvement in predictive performance.