We’re Back and We Brought an S3 Cloud With Us
We’ve been busy building a world-class S3 layer on top of the D2 distributed cloud. Price, no egress fees, extreme durability, AI native/friendly. What more can you want?
We’re Back!
It’s been a minute, but we’re back and better than ever. The product had a long gestation on the storage backbone architecture and while there are still lots of great features and improvements we’ll make there, it works extremely well. The piece missing was a world class S3 compatibility layer and we’re pleased to announce that is now ready. From multi-tenancy to STS to encryption to MCP, we have built a broad S3 surface, and we publish exactly how broad — every operation, tested and listed, at xns.tech/s3-compatibility.
Done doesn’t mean “it works in the demo.” Done means the defaults your tools ship with work on the first try — the boring details that don’t matter until they break everything.
Chunked uploads. Checksum negotiation. Bulk deletes in a single transaction. Listing a 100,000-key prefix without timing out. SigV4 and SigV2. We ran S3 conformance head-to-head and we publish the number — nobody else shows you theirs.
Change one endpoint. Every tool that already speaks S3 — AWS CLI, boto3, rclone, restic, Terraform — just works.
Now it’s pointed at AI
The reason to care right now: your AI agents can read and write this storage with the same access controls as everything else.
We ship an MCP server inside the Relayer container. An AI agent that writes to XNS is scoped by the same IAM policy as any other client, and every call lands in the same durable audit log — no backdoor, no separate trust path. That’s not a gimmick. If you can’t audit what your AI systems are doing to your data, you can’t use them on anything that matters.
The MCP gets you running. The CLI moves the bytes. Add one MCP server, point your assistant at it, and your agent has a real storage backend it’s actually allowed to touch.
Who this is for
Four readers, four reasons. We didn’t build a general-purpose box and hope — we built for these:
- The team bleeding money on egress. You move data in and out constantly and the bill is the punishment for it.
- The operator who can’t put data in someone else’s jurisdiction. Residency isn’t a preference for you — it’s the law.
- The workload that listing and bulk ops keep choking. Git-style fanouts, big deletes, large objects that blow your memory ceiling.
- The reseller / MSP running storage for many customers. You need isolation that’s enforced, not promised.
The Reason We Win
Low-to-no egress
You shouldn’t pay a toll to read your own data. Our economics don’t depend on locking your bytes in — Providers earn by contributing capacity, not by holding your data hostage on the way out. Let’s be honest, this is probably an “early days” thing. Sooner or later, we’ll be forced to address egress same as other clouds, but for a decent time period, you get serious discounts on the pain point that hurts the worst. And, based on the architecture behind our network, it is clear that even when egress fees begin showing up in our pricing that we’ll continue to be able to offer a lower head to head.
Sovereignty of ownership
Your data is sharded across the network with erasure coding — 80 data sectors and 40 parity across 120 independent Providers today — not copied into one operator’s building. No single region, jurisdiction, or operator can take it offline. For regulated workloads, you bind data to approved regions yourself. You own it; you can prove where it lives. As time goes by, we’ll refine this number downward but it won’t ever go below that range where we can provide the math showing a true eleven 9s of durability. Again, ask them to show their work.
Performance — the numbers that changed
The fixes that turn a timeout into a non-event:
- High-fanout prefix listing: 1,812ms → 3.5ms.
- Bulk delete of 1,000 objects: ~30s → sub-second.
- Large-upload memory ceilings: gone — we stream to storage instead of buffering in RAM.
Build your own Virtual Private Datacenter
Why surrender control to a remote warehouse when you can assemble a Virtual Private Datacenter to your exact specification? Pick the geography, the isolation, the placement rules — the network assembles to your spec. Set it, and watch a private S3 datacenter come together.
Our posture — we’d rather show you than tell you
One thing hasn’t changed in two years: we win on transparency, not adjectives.
Eleven nines of durability, computed — not rounded up. A floor we actively hold, not a launch-day snapshot. We don’t wait for failure; repair runs in the background whether anyone’s watching or not. Every comparative or quantitative claim on this page has a citation trail behind it, and where it doesn’t yet, we say so.
The score isn’t the story. The citation trail is.
Storage is a long game — the lions’ share happens in the outyears, not launch week. So here’s the door, not the mic drop:
Get keys, change one endpoint, and store on a network whose durability you can watch — and verify. Add one MCP server, say the words, and let your AI agent work against a backend it’s actually governed to use.
Welcome back. Let’s move some bytes.
Explore the platform
The full platform
Everything you need to go from evaluation to production.
How it works →The Relayer, erasure coding, and your Virtual Private Datacenter.S3 compatibility →Head-to-head S3 feature matrix vs the field.Eleven nines →Durability, computed — not rounded up.What's new →Everything built since you last looked.Releases →Every build, newest first — what changed and when.
