As an alternative to variable selection or shrinkage priors in high dimensional regression problems, we propose to randomly compress the predictors to dramatically reduce storage and computational bottlenecks. As opposed to existing Bayesian dimensionality reduction approaches, the exa