Everybody that pins it or retrieved the file from someone else (second one to maximum cache size that is configurable). Files and folders are referred to by hash.

Usually you still need to run a server to seed the files, but if a file becomes popular it will be served from anybody who also retrieved it and likely persist in the network even if your server goes down.

Think of it as BitTorrent for the web. It has Gateways that allow current browsers to access the network and pointers to update content with a new version (that's /ipns/ not /ipfs/ - the old version persists as long as someone hosts it).

What's also interesting is that due to the nature of the hashing algorithm (it's a merkle tree), you can "ipfs add" a directory, preserving the file structure inside, so websites on IPFS can use relative paths.