As far as I understand, Scikit-learn currently tries to link against the system (or possibly numpy) BLAS, and if it is not found falls back to the bundled CBLAS. The latter case results in degraded performance, and can lead to packaging issues (e.g. #10105).

In the future, once support for scipy <0.16 is dropped it might be worthwhile to use the Scipy BLAS API in all cases and remove the bundled CBLAS. At the time of writing, the minimal supported scipy version is 0.13.2.