Paste the identifier bafkreifzjut3te2nhyekklss27nh3k72ysco7y32koao5eei66wof36n5e into the gateways of two unrelated companies and both hand back the same eleven bytes: “hello world”. Neither could send anything else under that name, because the name is the SHA-256 fingerprint of the content.
Now the catch, which is the story of decentralized storage in 2026. At 10:06 UTC on 17 November 2026 the best-known gateway of all, ipfs.io, did not answer a plain request for that identifier. Its sister, dweb.link, replied with error 429 and one line saying the gateway is “switching to a service worker gateway only”. The address still works. The free service behind it has changed.
A blockchain is a poor hard disk, so nearly everything a dapp or an NFT shows you is kept somewhere else, and one question decides whether it will still be there next year: who is paying to keep this file? IPFS, Filecoin and Arweave give three different answers. With IPFS nobody is paid. Filecoin charges rent and makes the landlord prove the file is there. Arweave takes one payment that is meant to last two centuries.
The previous lesson of this path, ENS names and decentralised identity, showed that a .eth name can point to a website through its contenthash record. This lesson is about what that record points to. It also picks up where the lesson on what an NFT really is stopped: the token says “the picture is on IPFS” or “on Arweave”, and here is what those words do and do not guarantee.
A blockchain is a ledger, not a hard disk
Every full node of a chain keeps a copy of everything the chain has agreed to remember. That is the source of its security, and the reason space on it is rationed and priced by the byte.
One megabyte on Ethereum: 724 million gas
Take a file of one mebibyte, 1,048,576 bytes, a small photo. The lesson on how the EVM works gave the price of permanent contract storage: 22,100 gas to fill one empty 32-byte slot. The file needs 32,768 slots.
The table below is an EXAMPLE, built on readings taken at about 09:45 UTC on 17 November 2026: base fee 0.128 gwei at Ethereum block 26,433,954 (public RPC node), a gas price of 0.145 gwei once a small tip is added, and ether at $1,812.53 (Coinbase spot). The 21,000 gas of each transaction is ignored.
| Way of storing 1 MiB | Gas schedule | Gas needed | What it takes | Cost in this example |
|---|---|---|---|---|
| Contract storage | 22,100 gas per new 32-byte slot | 724,172,800 | At least 44 transactions, the room of 12 full blocks | About 0.105 ETH, about $190 |
| Calldata (data carried by a transaction) | Floor of 40 gas per non-zero byte (EIP-7623) | 41,943,040 | 3 transactions | About 0.0061 ETH, about $11 |
| Amazon S3 Standard, US East, for comparison | $0.023 per GB a month | none | One upload | About $0.00002 a month |
Read the dollar column with care. Gas was unusually cheap that morning. At 20 gwei, an example of a busy day, the same megabyte of contract storage would cost about 14.5 ETH. The price is not even the hard limit. A block holds 60 million gas and one transaction at most 16,777,216, so a single megabyte of storage would take over twelve consecutive blocks of the whole network, and a gigabyte about 12,360 blocks, some 41 hours with nothing else getting in.
Calldata is cheaper because it is not state: contracts cannot read it later, and nodes are no longer bound to keep it. Since 8 July 2025 every Ethereum execution client can run without the block bodies from before the Merge, the Ethereum Foundation announced that day.
Blobs and inscriptions do not solve it
Rollups post their data in blobs, cheaper again than calldata. Blobs are deleted by design after about 18 days, as the lesson on blobs and EIP-4844 explains. They prove that data was published. They do not keep it.
Bitcoin is the counter-case. An inscription writes a whole file into a transaction, and every full node then carries it for good, at block-space prices. The lesson on Ordinals, Runes and BRC-20 covers that route. It works for a small image. It does not scale to a film archive.
So the usual pattern is modest. Put 32 bytes on the chain, a fingerprint of the file, and put the file elsewhere. Three questions remain: where is “elsewhere”, who pays for it, and how do you check that you received the right file? The third one has the cleanest answer.
Decentralized storage starts with a file named by its content
From URL to CID
A web address says where a file is: a server, then a path. Whoever runs the server can change what sits there, or remove it, and the address stays the same. Most broken links and most swapped NFT pictures come from that.
A content identifier, or CID, says what the file is. It is computed from the bytes themselves. Change one byte and the CID changes; the same bytes give the same CID on any machine using the same settings. A CID carries no location at all, so anyone holding the file can serve it, and the receiver can check it without trusting the sender.
The tool behind this is the hash function from the lesson on cryptography: SHA-256 turns any input into 256 bits that act as its fingerprint. IPFS uses SHA-256 by default, according to its documentation on content identifiers.
Inside a CID: version, codec, hash
Here is the “hello world” identifier taken apart. The SHA-256 of those eleven characters, with no line break, is a string of 64 hexadecimal characters. Four short prefixes are placed in front of it, then the whole is written in base32 with a leading letter b.
text hello world (11 bytes)
sha-256 b94d27b9934d3e08a52e52d7da7dabfac484efe37a5380ee9088f7ace2efcde9
prefix 01 CID version 1
prefix 55 codec "raw": the bytes are the file itself
prefix 12 the hash function is sha2-256
prefix 20 the digest is 32 bytes long
base32, "b" bafkreifzjut3te2nhyekklss27nh3k72ysco7y32koao5eei66wof36n5e
Two families exist. A 46-character identifier that starts with Qm is a CIDv0, the older form. CIDv1, the one above, adds those leading identifiers so that software knows which encoding, which format and which hash function it is looking at. If SHA-256 is ever retired, the format can name its successor.
Large files become a tree of blocks
Hash a photo with an ordinary tool and you will usually not find the result inside its CID. IPFS does not hash a large file in one go. It cuts it into blocks, typically between 256 KiB and 1 MiB each according to the documentation, hashes every block, then builds a root block that lists the hashes of the others. The CID is the hash of that root.
It is the Merkle tree of the lesson on what a block contains, applied to pieces of a file instead of transactions. A downloader can fetch blocks from several machines at once and check each one on arrival; a single corrupted block is spotted and fetched again elsewhere. One side effect surprises people: the same file cut with different settings gets a different CID.

IPFS finds files, it does not promise to keep them
The InterPlanetary File System comes from a paper that Juan Benet submitted to arXiv on 14 July 2014, “IPFS – Content Addressed, Versioned, P2P File System”. His company, Protocol Labs, built the first software and later Filecoin. IPFS itself has no token and no payment of any kind. It is a way to name, find and move files.
Providers, a hash table and bootstrap nodes
A node that holds a block announces it in a distributed hash table, a shared directory with no central server of the kind described in the lesson on peer-to-peer networks. Someone who wants a CID asks that directory, or an indexing service over HTTP, which machines provide it, then downloads the blocks and checks each against its hash.
A newcomer finds the network through a short list of bootstrap nodes, and somebody has to run them. The directory itself only records who offers a file right now. If the last provider goes offline, the CID is still valid and still leads nowhere.
Pinning and garbage collection
A node keeps what it has been told to keep. That instruction is a pin. Everything else it has merely passed along sits in a cache, and garbage collection deletes cached data when disk space is needed. Add a file from the command line and your own node pins it automatically; nobody else’s does.
The IPFS documentation on persistence is blunt about the consequence: the network “doesn’t guarantee that any content is persistently available”. A popular file survives because many nodes choose to hold it. An unpopular one survives only as long as its one sponsor keeps a machine running, or pays a company to do so.
Public by default
The privacy page of the documentation is just as direct: “All traffic on IPFS is public, including the contents of files themselves, unless they’re encrypted”. The connection between two nodes is encrypted in transit; the content is not, and which node asks for which CID can be observed.
Encryption is therefore the user’s job, before the file is added. Even then, think twice about anything sensitive. A ciphertext that many strangers can copy is only as safe as its cipher remains, for as long as the copies exist.
IPNS and DNSLink: stable names for changing content
Content addressing has an obvious cost: update a website and its CID changes. Two layers give a fixed name to moving content. An IPNS name is the hash of a public key; its owner signs a record pointing to the current CID and republishes it regularly, because records in the hash table expire after 48 hours. DNSLink does the same with an ordinary domain: a TXT record on _dnslink.example.org holds the current path. An ENS contenthash, seen in the previous lesson, plays that role on Ethereum.
2026, the year the free gateways stopped
Browsers do not speak IPFS, so most people have always reached it through a gateway: a web server that fetches the blocks for you and returns the file over HTTPS. The address comes in two shapes, the path form https://gateway/ipfs/CID and the subdomain form https://CID.ipfs.gateway/. The documentation asks web apps to use the second, which gives each CID its own browser origin.
A gateway that fetches anything for anyone is expensive to run and easy to abuse. For years a few sponsored ones did it for free, and much of the “decentralized” web quietly depended on them. This is how that ended.
| Date | What happened | Source |
|---|---|---|
| 14 August 2024 | Cloudflare’s public gateway hostnames stop connecting to IPFS; traffic had been handed to ipfs.io and dweb.link from 14 May | Cloudflare blog |
| 16 April 2025 | NFT.Storage Ltd. announces it is winding down; its products pass to partners | NFT.Storage blog |
| 31 January 2026 | Fleek, a popular host for IPFS websites, ends its hosting as it turns to AI | IPFS blog |
| 31 May 2026 | Date announced for the end of Storacha’s upload and IPFS service | IPFS newsletter, 16 April 2026 |
| 26 July 2026 | Storj Labs opens a court-supervised financial restructuring | Storj open letter |
| 15 August 2026 | Date given to Infura’s customers for the end of its IPFS API and dedicated gateways | Filebase blog, 29 July 2026 |
| 24 and 25 August 2026 | Shipyard says it will stop its IPFS work; rate limiting starts on ipfs.io and dweb.link | Shipyard and IPFS blogs |
| 30 September 2026 | Shipyard’s last day of IPFS work | Shipyard blog |
| 14 October 2026 | All in-browser traffic of the public gateways now goes through a service worker that fetches from peers | IPFS blog |
Shipyard was the independent team that maintained the main IPFS software and operated ipfs.io, dweb.link and the bootstrap nodes. On 24 August 2026 it wrote in a post on the end of its IPFS work that Protocol Labs would not renew its funding, and that Protocol Labs, “as the owner of the associated domains and infrastructure, will determine their future”.
The next day the IPFS project published its plan to move beyond the sponsored gateways. Its line: “The IPFS mission was never to run a big gateway.” People who open ipfs.io or dweb.link in a browser are now redirected to a gateway that runs inside their own browser, fetches blocks from peers and verifies them locally. Sites and back-end services that used the public gateways as a free file server are throttled, which is the 429 error from the opening test.
Other free doors closed in the same period. Storacha, the successor of web3.storage, was due to shut its upload and IPFS service on 31 May 2026, the IPFS project’s newsletter said on 16 April; on 17 November 2026 its address redirected to Fil One, a paid storage product in the Amazon S3 style priced at $5.99 per terabyte a month, with private buckets only. Infura told its customers that its IPFS API and dedicated gateways would end on 15 August 2026, according to Filebase, a pinning company that competes with it.
The protocol did not close. A CID resolves wherever someone still provides the data, and the two company gateways in the opening test answered at once. But the same IPFS post says core maintainership is being reorganised, and the latest releases of Kubo, the main node software, and of IPFS Desktop both date from 15 September 2026, two weeks before Shipyard’s last day. The documentation now describes the public gateways as offered “on a best-effort basis” by the IPFS Foundation.
What a pinning service costs today
If your file must stay reachable and you do not want to run a node around the clock, you pay a pinning service. These are list prices in dollars, read on the companies’ own pricing pages on 17 November 2026.
| Service | Free tier | First paid plan | Extra storage |
|---|---|---|---|
| Pinata | 1 GB, 500 files, 10 GB of bandwidth | $20 a month: 1 TB, 500 GB of bandwidth | $0.07 per GB |
| Filebase | 5 GB, 500 pinned files, 10 GB of IPFS egress | $7.50 a month: 500 GB, 250 GB of IPFS egress | $0.015 per GB |
| NFT.Storage | No free uploads; the April 2025 handover said files from the old free product stay accessible, possibly more slowly | A one-time fee for long-term storage, run by Lighthouse | Not published per GB |
| Fil One (the former Storacha address) | 1 TB for 30 days | $5.99 per TB a month, $5.99 minimum; an S3-style store, not IPFS pinning | Same rate |
A pin is a subscription. Stop paying, or see the company disappear, and the file is one garbage collection away from vanishing unless somebody else pinned it too. The IPFS documentation says so itself: there is no guarantee that these companies will keep their services running. After Fleek’s exit the project advised site owners to build so that “no single provider change requires starting over”: keep your own export of the content and keep the name that points to it, DNS or ENS, in your own hands.
Filecoin rents space and makes providers prove it
IPFS leaves open the question of who is paid. Filecoin, from the same company (see our Filecoin project page), answers with a blockchain whose miners are storage providers. Its mainnet launched on 15 October 2020, the date given in the SEC’s 2023 complaint against Binance, and its token FIL has a stated maximum of 2 billion units.
Deals, sealing and daily proofs
A client pays a provider to hold data for a set period. The provider locks FIL as collateral. It then has to prove two things, both described in Filecoin’s documentation on proofs.
Proof of Replication comes first, when the data arrives: the provider encodes its own unique copy, a slow step called sealing, and shows that it did. Proof of Spacetime follows for the rest of the contract. In each proving period the provider must answer a challenge on every sector it claims, within an assigned 30-minute window. Miss it and, in the documentation’s words, pledge collateral is forfeited and storage power is reduced.
This is what IPFS lacks: a penalty, enforced by a chain, for losing a file. It has a price. Sealed data is cold storage, slow to read back, and the original deal system was built for large archives. Most applications therefore used a friendlier layer on top, web3.storage and NFT.Storage among them, which is how the closures in the table above reached Filecoin users too.
From 17 exbibytes to one
Filecoin’s raw capacity reached an all-time high of nearly 17 exbibytes in the second and third quarters of 2022, and its count of storage providers a high of over 4,100, according to the research firm Messari’s report on the network for the second quarter of 2024. On 17 November 2026 at 09:44 UTC the Filfox explorer showed 1.11 exbibytes of raw capacity, about 1,135 pebibytes, and 432 active providers. That is roughly one fifteenth of the peak.
The fall says less about lost files than about empty shelves. Block rewards once paid providers for committing space whether or not anyone used it. As rewards shrank and commitments expired, the unused space left. What remains is mostly space holding verified client data: Filecoin counts such sectors ten times, and the weighted figure stood at 10.6 exbibytes the same morning.
Who stores there? Starling Lab, a Stanford and USC research centre, says it runs a 22-petabyte Filecoin node at USC alongside the university’s libraries, where it works on preserving the USC Shoah Foundation’s testimonies. Filecoin’s review of 2025 also names the Smithsonian, MIT Open Learning and the Digital Public Library of America among sources of archived data.
The 2017 sale, smart contracts and the Onchain Cloud
Filecoin was financed by one of the largest token sales of 2017. Protocol Labs reported raising more than $205 million, the SEC recalled in its Binance complaint; the company’s own account gives $205.8 million. The sale ran in two parts, an advisor round in July and a round in August and September, through agreements for future tokens sold to accredited investors under a Regulation D exemption. The tokens arrived three years later.
Since 14 March 2023 the chain runs smart contracts compatible with Ethereum’s. In 2025 it added faster finality and a second kind of proof, Proof of Data Possession, meant for data that must be readable at once.
That proof underpins the Filecoin Onchain Cloud, live on mainnet since 26 March 2026. Storage is paid through contracts, proofs are checked on chain every 24 hours and payments halt automatically if proofs stop. The launch post gives a price from $2.50 per tebibyte a month per copy, with two copies by default, which is about half a cent per gigabyte a month. It also gives the size at launch: 49.41 TiB stored across 478 datasets, with 81 payer wallets connected. The offer is cheap, and its use so far is small.
FIL does four jobs in this machine. It pays for storage and transaction fees, it is locked as collateral (about 57.2 million FIL on 17 November 2026, per Filfox), it is created as block rewards, about 53,000 FIL a day at that date, and it is burnt through fees and penalties, about 43.2 million so far. Nothing in that list says anything about its price.
Arweave sells two hundred years in one payment
Arweave, a network launched in 2018, reverses the logic. There is no rent and no contract with a provider. You pay once, the data joins a single shared archive, and miners are rewarded for proving that they hold copies of that archive.
The endowment and its two assumptions
Part of each payment goes into a reserve called the storage endowment, released to miners slowly over time. Arweave’s protocol paper, whose first author is Sam Williams, puts the terms in one sentence: “Users pay upfront for the storage of 20 replicas for 200 years, at present costs.”
Two hundred years becomes “permanent” through a bet on falling costs. If storing a given amount of data gets cheaper by only 0.5% a year, the paper argues, the endowment never runs dry, and it sets that against a decline it measures at about 38.5% a year over the last half-century.
The margin is wide. It still rests on two assumptions the paper states itself. Storage must keep getting cheaper. And the calculation holds “in the absence of token price changes”: the reserve is held in AR, so what it can buy moves with AR. Arweave describes its goal as storing data “for the maximum possible term”. That is an engineering target backed by an economic bet, and nobody can test it before decades have passed.
What a gigabyte costs today
The network quotes its own price. At about 09:45 UTC on 17 November 2026, Arweave’s public interface asked 13.52 AR to store one gibibyte, about $33.30 with AR at $2.463 (Coinbase spot), paid once. Through Turbo, a service by ArDrive that bundles uploads and takes cards, the same gibibyte was quoted at $51.18.
For scale, $33.30 is what Amazon S3 Standard bills for one gigabyte over roughly 120 years at its list price of 17 November 2026. Take that as an order of magnitude only. Amazon’s price will change many times in 120 years, and Arweave’s promise has been running for eight.
The archive itself is small. That morning the network reported about 393 terabytes in total at block 2,044,080, where Filecoin’s capacity is counted in the thousands of petabytes.

Nobody can delete, anybody can refuse to serve
Arweave has no delete function. What it has is refusal. The paper describes a system in which each participant chooses “what data they would like to store and serve, without any mandates from the protocol”. Miners, it notes, are subject to the laws of their own state, and gateways decide independently what they index and serve.
So unlawful content can be filtered out by miners, by gateways and by apps, and in practice much of it becomes hard to reach. No authority can order every copy erased, though. For a user the rule is short: an upload cannot be taken back, by anyone.
Storj, Sia, Swarm, Walrus: other answers
Storj encrypts files on the user’s machine, splits them and spreads the pieces across independent node operators, behind an interface compatible with Amazon S3. Its list price on 17 November 2026 was $7 per terabyte a month, plus $7 per terabyte downloaded. The company behind it, Storj Labs, began a court-supervised financial restructuring on 26 July 2026, and its website says its products are being sold through that process. The open letter announcing it states: “The network continues to operate normally.”
Walrus, built by Mysten Labs on the Sui blockchain, went live on 27 March 2025. It is a rent model with a twist: the price is fixed in dollars, $0.023 per gigabyte a month according to its documentation, but billed on the encoded size of a file, about 4.5 times the original plus a fixed amount of metadata. Its own example works out at roughly $62 to keep a 50 GB file for a year. Storage can be bought about two years ahead at most.
Two older networks are still maintained. Sia, where renters sign storage contracts with hosts, published a beta version of its host software on 4 November 2026. Swarm, the storage project that grew up alongside Ethereum, released version 2.8.3 of its Bee client on 15 October 2026.
Side by side: who pays, for how long, at what price
All prices are list prices in dollars read on 17 November 2026. Network fees, download charges and monthly minimums are left out.
| Criterion | IPFS alone | Pinning service | Filecoin Onchain Cloud | Arweave | Walrus | Amazon S3 Standard |
|---|---|---|---|---|---|---|
| Model | None: you host | Rent, monthly | Rent, checked by proofs every 24 hours | One payment | Rent, prepaid up to about 2 years | Rent, monthly |
| Who keeps the file | Whoever pins it | The company | Providers under contract | Miners paid from the endowment | Storage nodes, for the paid period | Amazon |
| Promise | None | While you pay and the company lives | While you pay; payment stops if proofs stop | 20 replicas for 200 years paid upfront: an economic bet | Until the paid period ends | While you pay |
| Can it be deleted? | Unpin it; other copies may remain | Unpin it | End the contract | No | Yes, if stored as deletable | Yes |
| Indicative price | Your own machine | $0.015 per GB a month (Filebase, 500 GB plan); $0.02 (Pinata, 1 TB plan) | From $2.50 per TiB a month per copy, two copies by default | About $33 per GiB, once | $0.023 per GB a month on the encoded size | $0.023 per GB a month, first 50 TB, US East |
| Token | None | None | FIL | AR | WAL, fees in SUI | None |
The table leaves one row out on purpose: speed. A file on S3 arrives in milliseconds from a known address. A file on IPFS arrives as fast as its best provider allows, and one held only in sealed Filecoin sectors can take much longer.
What it is used for, and where it lets you down
NFT media is the best-known use. A token’s metadata and picture usually live off the chain, and a content address guarantees that the file cannot be swapped. It does not guarantee that anyone still hosts it. When the marketplace Foundation closed on 27 April 2026, it told collectors its IPFS gateway would be maintained only until 27 April 2027. After that date the pictures exist wherever someone else has pinned them.
Front ends are the second use. A dapp’s website can be published on IPFS and named through ENS, so that no single hosting company can take it down. As the lesson on dapps points out, this spreads the hosting; it does not stop anyone from opening a fake copy at another address.
Archives are the third, and the one with named users: the Starling Lab node and the museum and library collections cited above. Here the proofs matter more than speed.
The weak points are the same each time. Retrieval is slower than a commercial cloud. Gateways are ordinary web servers that can be throttled, blocked or closed, as 2026 showed. Links rot when the last pin lapses. Privacy is absent unless you add it yourself.
Securities law, takedowns and the right to erasure
For a US reader the tokens come first. In June 2023 the SEC alleged, in its complaints against Binance and Coinbase, that FIL had been offered and sold as a security. Neither claim was tested. The agency announced the dismissal of the Coinbase case on 27 February 2025, in a release saying the decision rested “not on any assessment of the merits of the claims”, and on 29 May 2025 the SEC and Binance filed a joint request to dismiss the other case with prejudice, which means it cannot be filed again, Cointelegraph reported from the court filing. No court has ruled on FIL, and Protocol Labs was not a defendant in either case.
Unlawful content is handled at the door, not in the network. The public gateways run for the IPFS Foundation apply a denylist called Bad Bits, a list of hashed CIDs flagged for reasons such as malware or copyright violation, and a blocked CID returns error 410 there. The documentation on these public services states the limit plainly: the Foundation does not host the content and can only block it at its own gateways. The data stays on any node that holds it, and another gateway may still serve it.
Personal data is the hardest case. The European Data Protection Board adopted the final version of its guidelines on blockchains on 7 July 2026. They state that “technical impossibility cannot be invoked to justify non-compliance with GDPR requirements”. On deletion they leave no room: “The rights to erasure and to object must be complied with by design.” Their advice is to keep personal data off the chain, in storage that can be erased.
Those guidelines address blockchains and do not name IPFS, Filecoin or Arweave. Their logic applies with more force to a network built so that nothing can be erased. A company that serves European users and writes a customer’s name into Arweave has created a problem it cannot fix later.
Try it for free: one CID, two gateways, one vanishing file
None of the first three steps costs money or needs a wallet.
- Open the “hello world” CID through two company gateways, for example
https://ipfs.filebase.io/ipfs/bafkreifzjut3te2nhyekklss27nh3k72ysco7y32koao5eei66wof36n5eandhttps://bafkreifzjut3te2nhyekklss27nh3k72ysco7y32koao5eei66wof36n5e.ipfs.4everland.io/. Both returned the text at 10:06 UTC on 17 November 2026. They are named as examples, not recommended, and either may change its terms. - Compute the SHA-256 of the eleven characters hello world with any offline tool on your computer. You should get the digest shown earlier, the one wrapped inside the CID. You have just verified a file without trusting the server that sent it.
- Find where an NFT’s media lives. Read the token’s tokenURI on a block explorer, as shown in the lesson on buying an NFT safely, and see whether it starts with ipfs://, ar:// or https://. Then ask the question of this lesson: who is paying for that file?
- Watch a file vanish. IPFS Desktop packs a node and a file manager into one application for Windows, macOS and Linux. Take it only from the releases page of the ipfs/ipfs-desktop repository on GitHub; version 0.50.1 of 15 September 2026 was the latest on 17 November 2026. Import a harmless test file, because it becomes public the moment you add it. Copy its CID and open it through a gateway from your phone. Then quit the application and try again a little later: the CID is unchanged and no longer leads anywhere, because your computer was its only provider.
Two further steps cost real money and are described here, not recommended. A paid pin is a subscription at the prices in the table above. An Arweave upload is billed by size at the rates given earlier, about $33 per gibibyte on 17 November 2026, so a one-megabyte file costs a few cents.
Before any Arweave upload, remember that it cannot be deleted. Never send personal data, identity documents, someone else’s work or anything you might regret. No upload requires typing a seed phrase into a website or an app: a tool that asks for one is a trap.
Before trusting “stored on IPFS”, find who pins it
A CID proves which file you received. It says nothing about whether anyone still holds it, and 2026 showed the difference: the links kept working while the free hosting behind them closed, one service after another.
So the check is the same for an NFT collection, a dapp front end or an archive. Find the name of whoever pins or pays for the file, and what happens when they stop. One date is already known: Foundation’s gateway is due to run only until 27 April 2027.
Storage is one of several physical services now sold through tokens, with computing power and wireless coverage beside it. The next lesson of this path covers those networks, known as DePIN.



