Decentralized storage priced on put, keep and exit cost

Storage quotes are designed to be compared wrongly. The headline is the put price ($/TB/month), but every real bill has three layers: put (upload), keep (renewal, sealing, minimum retention), and get (retrieval, egress, per-request API fees). The ranking above prices durability and exit costs as first-class dimensions, because in this market the exit is where the money goes.

The three-layer cost stack

• Put - one-time upload and sealing cost. Filecoin's sealing economics make initial deals cheaper at scale but bundle a renewal story [1].
• Keep - monthly $/TB (subscription style) or a one-time endowment. Different risk profiles: recurring bills compound, endowments bet on token economics [3].
• Get - the decisive number. Retrieval per GB varies by an order of magnitude across networks, and archive-heavy workloads discover this cost every time someone actually reads the data.

The honest comparison framework: take your real read/write ratio and price all three layers at that ratio. The storage cost tool does exactly this - a "cheaper" provider that charges multiples for retrieval loses at any read-heavy workload. Price the hold over your real holding period: a five-year keep on a subscription model and a five-year endowment are different products with different failure modes.

Durability: three mechanics, not one number

Durability is a mechanism, not a percentage. The networks in this population take three different bets:

1. Economic collateral (Filecoin) - providers post collateral that gets slashed on failed proofs. Durability rests on the penalty being bigger than the cheat [1].
2. Endowment (Arweave) - your one-time payment funds a trust that pays for replication going forward [3]. Elegant design, real track record - and still a bet, denominated in a token.
3. Redundancy and replication (erasure coding) - shards across independent operators with audits. The most S3-like mental model.

None of these are backups. A storage network guarantees against node loss; it does not guarantee against your own deletion, bad client code, or a gateway outage. Design retention accordingly - and for archive-class data, redundancy across two providers on different networks beats any single network's guarantees, because the failure modes are uncorrelated and the cost is arithmetic. On paper this market looks deep; the honest count of providers that document what they actually do is the 88% in the strip above.

Live from our database - caveats and sources in each linked review.

Retrieval: price the read, not the gigabyte

The get-layer decides real cost, and it behaves differently per model [3]:

• Per-request APIs (S3-style) - small reads are cheap, bulk restore is where bills spike. Test with your real object-size distribution.
• Gateway reads (Arweave-style) - reads ride public gateways; the cost is latency variance and gateway trust, not dollars [3].
• Deal retrieval (Filecoin-style) - retrieval is a negotiation with the miner holding the deal; latency and price vary per provider [1].

The honest test: fetch your own data back before the workload is committed. Retrieval speed, price, and API behavior differ enough to reorder the ranking - and the storage cost tool prices the full round trip at your read ratio. Our archive keeps 4 dead or scam operators in this vertical on record; 60% of the live ones still let you pay without identity checks. The durable names are the ones whose keep-layer mechanics survived their first token winter.

Permanent storage: what you are really buying

The permanent-storage hub covers the endowment mechanics in depth. The short version for this page: Arweave-style permanence bets that the endowment's returns exceed future storage costs [3]. That is a strong design with a real track record - and it is still a bet denominated in a token. For data that must survive a specific period, contract-based storage with verifiable renewals is the conservative choice; for data that must outlive institutions, the endowment model is the only serious answer in this market.

S3-compatible: what the label really means

"S3-compatible" covers a spectrum - the s3-compatible hub separates them [1]:

• Full API surface - versioning, multipart, lifecycle policies. The genuine article.
• Core object operations - get/put/list works, advanced features silently no-op. This is where client code breaks.
• Gateway-only - the S3 surface is a thin layer over token mechanics; rate limits and consistency semantics differ under load.

Test with your actual client library, not curl. The failure modes all live at the edges - multipart uploads, list pagination, consistency timing under load. One more reading note: the durability mechanisms above map to very different failure surfaces, so a five-year hold across one network is not the same product as a five-year hold across two [1].