Decentralised storage networks are often assumed to guarantee that a file lasts forever. They change how files are addressed and who holds them, and persistence depends on the incentives layered on top.
Addressing by content rather than location
Conventional web addresses describe where a file sits: a server, a path, a domain. Whoever controls that location controls what is returned.
Content addressing instead derives an identifier from a hash of the file itself, so the address depends on the data rather than on the host.
Requesting that identifier can only ever return the original content, because different data would produce a different identifier and fail verification.
Retrieval becomes a search rather than a lookup
With no location in the address, a client asks the network which peers hold the content and retrieves it from whoever responds.
Any node with a copy can serve it, so popular content is served from many places and no single host is a point of failure.
The corollary is that content nobody has requested recently may be held by very few nodes, and retrieval slows or fails accordingly.
Nothing persists without a reason to keep it
A node caching a file will eventually discard it, since storage is finite and there is no obligation attached to a cached copy.
Pinning services solve this commercially by contracting to hold specified content, which reintroduces a paid provider into the arrangement.
Incentivised networks instead pay storage providers and require them to prove periodically that they still hold the data, with payment contingent on those proofs.
Proofs are what make the payment meaningful
A provider is challenged to demonstrate possession of the data at unpredictable intervals, and a correct response requires actually holding it.
Failing the challenge forfeits payment and any collateral posted, which makes discarding the data more expensive than storing it.
Redundancy across independent providers covers the case where one fails, since the same content is stored several times over.
What this changes for applications
Content-addressed references cannot be silently altered, which is why they are preferred for anything whose integrity matters after publication.
Availability still has to be paid for by someone, whether through a pinning contract, a storage deal or a community that keeps copies for its own reasons.
The guarantee is therefore about integrity rather than permanence, and treating the two as the same thing is how references end up resolving to nothing years later.