Basically, you have a representation of the entire scene on the GPU at all times. If you think about it, that is already the case, since shaders, vertex buffers, and textures are all on GPU memory anyway. For RT you add the locations of the meshes and their bounding boxes, and which shaders they are using, which are executed when the rays hit a triangle.
This means many view-based optimization tricks used during rasterization like deferred shading or clustered forward lighting don’t apply during raytracing so new tricks need to be developed to optimize that.