Remember when every whitepaper in 2018 had a diagram with a little cloud labeled “IPFS” bolted onto the side? That was the shorthand for serious. You weren’t just dumping JPEGs on S3 like some centralized peasant, you were pinning content-addressed hashes to a distributed network that would outlive nation-states. The permanence was the whole pitch.
Turns out permanence has a payroll.
Shipyard, the team maintaining much of IPFS’s core infrastructure, is winding down. Their last day is September 30, 2026, after Protocol Labs declined to renew their funding. Separately, Infura is shutting down its IPFS service on August 15, 2026. Two dates, roughly six weeks apart, both landing on the same layer of the stack.
Why this lands differently for tooling
I review AI toolkits for a living, which means I spend a lot of time looking at dependency trees that nobody wants to look at. And IPFS shows up in more of them than you’d expect. Dataset distribution. Model weight hosting. Provenance claims for generated media. Anything that needed a content hash and a story about verifiability reached for IPFS, usually through a hosted gateway, usually without anyone on the team thinking hard about who was keeping the lights on.
That’s the pattern I keep running into when I evaluate tools. The interesting failure mode isn’t the thing that breaks loudly. It’s the thing three layers down that everyone assumed was infrastructure — the same category as DNS or NTP — that turns out to be a specific team with a specific funding agreement and a specific end date.
IPFS the protocol doesn’t disappear when Shipyard closes. Protocols don’t work that way. But protocols also don’t merge pull requests, triage security issues, or answer questions about why your gateway is timing out. People do that, and the people are leaving.
The gateway problem nobody planned for
The Infura shutdown is the more immediate one for anybody shipping product. If your tool talks to IPFS through a hosted gateway, you are not running a decentralized system. You are running a centralized system with extra hashing. That was always true, and it was always fine, right up until the gateway operator picked a date.
What I’d actually check if I were auditing a stack right now:
- Search your codebase for hardcoded gateway URLs. Not just the obvious ones — check config files, docs, and any SDK you pulled in that quietly defaults to a public endpoint.
- Figure out whether you’re pinning content anywhere or just hoping someone else is. Unpinned content on IPFS has a half-life, and it’s shorter than people think.
- Ask whether content addressing was load-bearing for your product or just aesthetic. If you can swap in ordinary object storage and nobody notices, that answers the question about how much the decentralization was really doing.
- Look at what else in your dependency list is maintained by a single funded team. This is not a one-off situation.
The uncomfortable part
Shipyard’s own framing was that they were grateful for the support and disappointed by the outcome. That’s a gracious way to describe getting your funding pulled after more than two years of maintaining infrastructure that a lot of other projects treated as free.
And that’s the structural issue. IPFS was funded by an organization with its own priorities, not by the thousands of projects that depended on it. None of those projects were paying for maintenance. They were consuming it. When the single upstream funder reallocated, the whole arrangement went with it, and every downstream team gets to discover their exposure at the same time.
I bring this up in toolkit reviews and it makes people tired. Nobody wants a maintenance-funding analysis when they’re trying to pick a vector database. But the thing that determines whether a tool is still working in three years usually isn’t the feature list. It’s whether somebody is getting paid to care about it.
What to actually do
You have months, not weeks, which is more warning than most infrastructure gives you. Use it. Move deliberately rather than waiting for a broken build to force the decision, because migrations done under deadline pressure are how you end up with two half-finished storage layers and a mystery bug in production.
If content addressing genuinely matters to your product, there are self-hosting paths and other providers, and you should be evaluating them now while you still have the luxury of testing things properly. If it doesn’t matter, simplify. Fewer moving parts you don’t control is a real feature, even if it looks less impressive in a diagram.
The broader lesson is one I’ll keep repeating in these reviews. Decentralized architecture doesn’t mean decentralized funding, and the second one is what keeps your dependencies alive. Ask who’s paying before you build on top of something. It’s a boring question with expensive consequences.
đź•’ Published: