Dade, there is no mesh export here, only spheres and cylinders (just a pair of them per hair material used). However in the long run we have to move to converting hairs to mesh on the fly. First, for GPGPU kernels; second, with the current approach performance drops linearly with the number of hair objects (on QBVH) what is really bad.
Doug already made some optimizations, but not enough I think. The main problem is not in hair export function (this just collects all the needed primitive objects), but in the main object export code which simply chokes on those tenth of thousands of objects.
Linux builds packager