Hi Colin,
I'm not sure where I got it from either. It has been in my head for a long time and so I think it had solidified. I'm probably regurgitating rubbish I read somewhere else. I apologise since I do not want to misinform anyone. I'll edit my original post to note this error.
It makes no sense that the image would be demosaiced before storage as DNG. The file is simply a 14-bit value per pixel (or whatever the bit depth is). It should not have to be demosaiced to be stored elsewhere as raw. It would be for an RGB TIFF (or any other RGB file). If it was demosaiced, it would not be raw. I think the gains in smaller file storage are due to the use of a better compression algorithm than that used in the manufacturer raws. They have to be compressed on-the-fly within camera and so it is advantageous to use a faster less efficient algorithm.
The demosaicing is done before storage in the ACR cache. However this will be specific to the ACR version and so will be done again if ACR gets updated to a new process version anyway.
Incidentally I read that the ACR cache for LR 4 stores the files as Adobe RGB JPEGs and so the cache files are much smaller. I was puzzled by this since it means any heavy lifting within ACR development will always have to go back to the original raw and do the demosaicing again. I am not sure how the new ACR will use the cache files. Perhaps they are now only used for thumbnail tiling previews where the lossy JPEG compression is not noticeable. Or more probably they are using the lossless compression part of the JPEG specification without using any of the lossy compression routines that reduce colour bit depth using fixed sized lookup tables for 8x8 grids (hence the blocky appearance of highly compressed JPEGs). As usual there are lots of things you could do to the image and they never tell me enough about the process for my curiosity.
Alex