Hi,
I think this depends heavily on the amount of groups you need to support and the way you imagine to maintain your Galaxy server.
We are supporting approximately 30 groups as of today with ~300 researchers and for us it is easier to maintain one instance for all. Also because we try to run Galaxy in reproducible mode, this means we keep old versions of tools, data, workflows and so on ... If you want to run Galaxy in "run-and-kill" mode, Docker is a very good solution even for a large user-base.
Don't know if you know this project, but here you have a Galaxy Docker container with all what you need.
https://github.com/bgruening/docker-galaxy-stable
This one is used by many working groups in reproducible-mode, but they maintain them by there own, more or less for one group or as dedicated webserver (NAR-webserver issue) to suite one task. I also know one project that is offering a handful of Galaxy flavours in a web-browser. You can pick for example an RNA-seq flavour and gets redircted to your personal Galaxy instance. Use it once and kill it afterwards. This is nice for the 'move tools to your data' idea.
Here is a (incomplete) list of already existing Galaxy flavours:
https://github.com/bgruening/docker-galaxy-stable/tree/16.01#list-of-galaxy-flavours
and here is our first attempt to click together your own flavour without knowledge of Dockerfile:
http://galaxyflavorsgenerators.github.io/galaxyFlavorGenerator/
Hope this helps,
Bjoern