The quick answer is that IPFS doesn't have any automatic replication by default. If I add a to my local machine it gets indexed on my local machine and I get back a hash pointing at that file. The data ONLY exists locally.
Another IPFS user on the internet could use my hash to retrieve my content and the data would be transferred directly from my machine to theirs. At this point, the data exists in 2 places and new users could fetch it from either location. IPFS is like a browser cache, it contains the things you have been looking at most recently and older things might be deleted. (pinning can optionally be used to keep your local node from deleting something)
In the future, Filecoin can be used on top of IPFS to hire a group of IPFS users to fetch your content and replicate it. This won't be needed for popular public content since that will get replicated anyway, but if you wanted to store backups on IPFS you could need some way to induce other machines to store your data.