This paper presents a pre-training technique called query-as-context that uses query prediction to improve dense retrieval. Previous research has applied query prediction to document expansion in order to alleviate the problem of lexical mismatch in sparse retrieval. However, query prediction has not yet been studied in the context of dense retrieval. Query-as-context pre-training assumes that the predicted query is a special context for the document and uses contrastive learning or contextual masked auto-encoding learning to compress the document and query into dense vectors. The technique is evaluated on large-scale passage retrieval benchmarks and shows considerable improvements compared to existing strong baselines such as coCondenser and CoT-MAE, demonstrating its effectiveness. Our code will be available at https://github.com/caskcsg/ir/tree/main/cotmae-qc .