Standard lossless compression (without further assumptions) is already very close to being as optimal as it can get: At some point the pure entropy of these huge datasets just is not containable anymore.

The most likely savior in this case would be procedural rendering (i.e. instead of storing textures and meshes, you store a function that deterministically generates the meshes and textures). These already are starting to become popular due to better engine support, but pose a huge challenge from a design POV (the nice e.g. blender-esque interfaces don't really translate well to this kind of process).