In-depth articles on technology shaping what comes next.

Matrix at Government Scale: 5 Myths, Tested

The European Commission rolled out Matrix to 40,000 staff. We test five myths about federated messaging, E2EE, and sovereign chat at government scale.

Isometric illustration of separate fortified buildings connected by a mesh of communication lines
Federation means every organization keeps its own fortress — but the lines between them are the hard part.

When a government body with 40,000 employees decides its internal chat can't live on a US vendor's servers, the options get thin fast. The European Commission's answer was Matrix — a federated, end-to-end encrypted protocol — and the rollout is one of the largest real-world tests of Matrix government scale deployment we've ever seen. It worked. Mostly. And the 'mostly' is where the interesting engineering lives.

I've spent years running messaging and collaboration infrastructure for organizations that actually care where their data sits, so I read the details of this rollout with a jaundiced eye. A lot of what people believe about federated chat — both the cheerleading and the doom-posting — doesn't survive contact with a deployment this size. Here are five myths, tested against what the Commission's rollout and the broader Matrix ecosystem actually show.

Myth 1: End-to-End Encryption Means You're Done With Security

This is the most dangerous belief in the whole space, and Matrix's critics on this point aren't wrong — they're just often imprecise. Enabling E2EE is the beginning of your security story, not the end. Matrix encrypts room content with Megolm, a ratcheting group encryption scheme, and that part is fine cryptographically. The hard part is everything around it: device verification, key backup, cross-signing, and what happens when 40,000 humans are involved.

Here's the operational reality. In an encrypted room, every device of every participant is a potential recipient of the group keys. A user with a phone, a laptop, and a web session is three devices. Multiply by thousands of users in a large room and key distribution becomes a genuinely hard systems problem — this is why big encrypted rooms have historically been painful on Matrix, and why the community still swaps horror stories about 'unable to decrypt message' errors. The cryptography didn't fail. The key management UX did.

Device verification is the classic casualty. If users don't verify each other's devices, you're trusting the server's key directory — which means your 'end-to-end' encryption is really 'end-to-end modulo the homeserver.' Most real users never verify anyone. They click through the shield icons because the modal appeared at a bad time. At a national-security-adjacent organization, that's not a UX nitpick; it's the difference between the threat model you wrote down and the one you're actually running. The long-promised migration to MLS (Messaging Layer Security), the IETF group messaging standard designed to handle exactly these large-group cases more gracefully, has been slow — slow enough that a tracking site exists just to answer whether it's done yet. It isn't.

The verdict: E2EE in Matrix is real, but at 40,000 users the weakest link isn't the cipher — it's verification fatigue, key recovery flows, and the gap between what the spec promises and what implementations ship. If your security review ends at 'messages are encrypted,' you haven't done a security review.

Myth 2: Federation Is Just Harder Centralization

The Slack-brain take goes like this: federation is email's mistake repeated — every server an island, inconsistent features, spam and abuse with no central bouncer. Why not just run one big server and be done with it? The Commission's rollout shows why that framing misses the point entirely. For them, federation wasn't an architectural burden they tolerated. It was the entire requirement.

The European Commission doesn't exist in a vacuum. It needs to talk to other EU institutions, member state governments, and eventually bodies like UNICC — all of whom have their own sovereignty requirements and none of whom will accept 'just get an account on our server.' That's the multilateral problem: how do governments collaborate digitally without forcing every organization onto the same platform? Centralized platforms answer it with 'they can't, pick a winner.' Federation answers it with 'each keeps their own infrastructure and the protocol handles the boundaries.' For a geopolitical entity built out of sovereign members, the federated answer isn't just nicer — it's the only one that's politically deployable. Germany's public sector, Dutch government projects, and healthcare systems are all running similar experiments for the same reason.

That said, the costs are real and you should budget for them honestly. Federation means room state is replicated across participating servers, and resolving that state — especially in rooms with a long history of servers joining, leaving, and getting disconnected — is computationally expensive. State resolution and event authorization are the parts of Matrix that make server operators sweat. Cross-server search and discovery are awkward. Abuse tooling is per-server, so a problem user on a federated server isn't your problem until they very much are. The Commission's deployment reportedly had to fix federated search behavior to make the thing usable — a detail that tells you the protocol-level happy path wasn't enough for production.

So no, federation isn't harder centralization. It's a different product for a different buyer. If you're a single company with one legal jurisdiction, the complexity buys you nothing. If you're the EU, it buys you the only deployment model your lawyers and your member states will sign.

Myth 3: If It Scales to 40,000 Users, It'll Scale for You

This myth runs in both directions, and both directions are wrong. Skeptics assume a government rollout proves nothing because 'they threw money at it.' Boosters assume 40,000 government users proves Matrix scales, period. The truth is more useful: scale on Matrix is not a single number, and the Commission's rollout demonstrates exactly which axes break first.

My favorite detail from the whole deployment: they switched off live presence. As user counts climbed, the load from the real-time online/away indicator started hurting performance, so the team disabled it. Think about that. Presence — the feature every chat user assumes is free — was one of the first casualties of scale. Presence is deceptively expensive in federated systems because it's fan-out by nature: your status change is interesting to every server that hosts someone sharing a room with you. In a 40,000-person organization with lots of shared rooms, that's a lot of gossip traffic for a green dot. Killing it was the right call, and it's a lesson I wish more teams learned earlier: features with O(everyone) fan-out are the first thing you should have a kill switch for.

A control panel of analog gauges with one large lever being switched off
Presence was one of the first features switched off when load climbed — fan-out features need kill switches.

The other scaling axis is sync, and here's where the small-deployment horror stories matter as much as the big one. Plenty of self-hosters report the same arc: works great for the first weeks, then rooms with history take hours or days to join, and resource usage dwarfs what a comparable centralized system would need. I've seen this pattern before in other eventually-consistent systems — it's the same class of surprise people hit in distributed databases where replication metadata grows faster than the data. In Matrix's case, joining a large federated room means backfilling and validating event history, and the cost is borne by the joining server at the worst possible moment: when a new user is forming their first impression.

Synapse, the reference homeserver, has gotten dramatically better — worker processes, better caching, sharded writers — and lighter implementations exist. But '40,000 users works' came with a commercial support contract, dedicated operations effort, and features deliberately disabled. That's not a knock on the deployment; it's just the honest bill. If your plan is 'apt install and walk away,' your plan ends in the 'it fell over after six months' genre of forum post.

Myth 4: Self-Hosting Means You've Achieved Sovereignty

Digital sovereignty is the reason this deployment exists, so it's worth being precise about what the word buys you. Running your own homeserver on your own hardware in your own jurisdiction means a foreign cloud provider can't be compelled to hand over your message metadata, can't deprecate the product out from under you, and can't unilaterally change the terms. That's real. But sovereignty is a property of your whole supply chain, not your data center's postal code.

Ask the uncomfortable questions. Who writes the server software you run, and could they walk away from it? Who builds the clients your staff actually touch, and what happens to your deployment if that vendor's roadmap pivots? Who pushes the push notifications — because on mobile, the answer is almost always Apple and Google's push infrastructure, which is a metadata channel straight through two American companies no matter where your homeserver sits. The Commission's rollout leans on Element's server suite, which is a sensible procurement decision — but 'sensible procurement' and 'sovereign' are different claims, and conflating them is how you end up surprised later.

There's also the operational sovereignty gap: do you have the in-house skill to run, patch, and incident-respond this thing at 3 AM, or have you just moved your dependency from a SaaS contract to a support contract? Neither is wrong! Support contracts are how adults run critical infrastructure. But a dependency you can see and negotiate is what sovereignty actually means in practice — not the absence of dependencies. The open protocol helps enormously here: if the vendor relationship sours, the server implementation, the spec, and your data are all portable in a way that a Slack export never will be. That exit option is the concrete, enforceable part of sovereignty. Guard it.

Myth 5: An Open Protocol Kills Vendor Lock-In

The final myth is the most seductive because it's half true. Matrix is an open standard with an open spec process, multiple server implementations, and a nonprofit foundation stewarding it. In theory, you're never locked in. In practice, look at the client landscape. Element — in its various incarnations — is effectively the only full-featured, polished, non-developer client. The ecosystem of alternative clients is real but skews heavily toward terminal UIs and passion projects. For a 40,000-person government workforce, the realistic client shortlist has one entry.

That matters because users experience the client, not the protocol. When the dominant client ships a redesign that hides information power users relied on — and the Matrix community has been loudly unhappy about exactly that class of change — 'it's an open protocol' is cold comfort. Your users can't swap in a comparable alternative because there isn't one at feature parity. This is a subtler lock-in than proprietary SaaS: you can always leave, but there's nowhere else to go. Open protocols prevent data hostage-taking; they don't guarantee a competitive market.

The hopeful counterargument is that deployments like the Commission's change the economics. Forty thousand paying seats and a government stakeholder with opinions about usability is exactly the kind of pressure that funds development — better verification UX, faster sync, MLS progress, more viable clients. Several commenters watching this rollout made the same point: the ecosystem's problems are real, but a customer this size adds money and forcing functions, not just users. Whether that materializes depends on whether the contract money flows into the shared protocol and reference implementations or into proprietary differentiators. Watch that, because it's the difference between a rising tide and a bigger moat.

One open doorway among a long wall of closed doors with light spilling through
An open protocol means you can always leave — but it doesn't guarantee there's anywhere else to go.

What This Rollout Actually Proves

Strip away the hype and the hostility and the Commission's deployment proves something narrower and more valuable than either side claims: federated, encrypted messaging is deployable at institutional scale if you treat it as critical infrastructure — funded, operated, and with features you're willing to shed under load. It does not prove that Matrix is easy, that its encryption UX is ready for adversarial threat models, or that a five-person self-hosted setup will have the same experience as a professionally run 40,000-seat one. The gap between those two experiences is most of what people argue about online, and both experiences are real.

If you're evaluating this path for your own organization, my advice is unglamorous. Pilot with the rooms and user populations that stress the known weak points: large encrypted rooms, mobile users on push notifications, new joiners syncing history. Decide in advance which features you'll sacrifice first when load shows up — presence is the canonical candidate. Treat client choice as a strategic risk, not an aesthetic preference. And write down whether you're buying sovereignty, federation, or encryption, because they're three different products wearing the same trench coat. The Commission bought the first two and is still working on the third. You probably will be too.