I'd suggest trying t-SNE [1] instead; you'll be losing almost all of the variance by projecting onto the first two eigenvectors produced by PCA.

[1] http://karpathy.github.io/2014/07/02/visualizing-top-tweeps-...