It took a look on the fpack utility. It is very effricient on fits files. Reducing the size to half "in a trice" with the Hcompress algorithm on the images I tested.
Regarding use case for cloud storage: thought it could be great for group projects, where users like us agree on a target (e.g. a certain galaxy) and upload all images they have taken of that target. Maybe, leveraging also on all the metadata related to the file, all such images can be stakked. So we could efficiently generate images with long exposure times. I am not sure if technically that could work, and maybe it is not a new idea, so apologies in a advance if this is a dumb proposal...
Right now, I'm testing with Amazon S3, and it's going OK. Of course, needs to figure out how this all scales. Hopefully I'll have something ready for beta-testing in the next couple of weeks.