When the P2P program is first installed, the program can populate its database in one of three ways. It can (1) rely exclusively on hard-wired data, (2) rely on hard-wired data at first but supplant it with dynamically-learned data after contacting the network, (3) gather all of its data dynamically. In practice, a spectrum of variations are commonly used which encompass the first and second option. The third option is simply not practicable because it would require scanning every conceivable Internet address (yes, all 255,255,255,255 of them).
At the two ends of this spectrum are systems like Usenet and Gnutella. In the case of Usenet, clients upload messages to a particular server which then replicates those messages and sends them to other servers within the system for mirroring, making them available to users throughout the system. This is essentially decentralised hosting. The servers themselves relate on a peer-to-peer basis, but the user clients do not.
In the case of Gnutella, however, peer-to-peer is extremely decentralised. You cannot join the network unless you know at least one member. This member then "introduces" you to his/her circle of friends, thus populating the database of peers for your Gnutella installation (for acting both as client and as host). Few, if any, ever get the full database of potential peers because addition to the ring of trust requires knowing someone else.