A Datacenter Burned Down in France. Some Data Never Came Back.

Shortly after midnight on March 10, 2021, a fire alarm went off at a datacenter campus in Strasbourg, France. Within hours, one of the four buildings on the site — SBG2, a five-story concrete hall full of servers — was burning out of control. Firefighters from across the region fought it through the night. By dawn, SBG2 was a ruin.

This was not some fly-by-night host. It was OVHcloud, one of Europe's largest cloud providers, with millions of customers. Octave Klaba, the company's founder, spent the night posting updates on Twitter because there was no press release fast enough for what was happening. He warned customers plainly: activate your disaster recovery plans. By morning, millions of websites were dark — government services, game servers, small businesses, personal projects. And for a meaningful number of customers, something worse: their data was simply gone. Not temporarily unreachable. Gone.

The cloud, it turns out, is not a cloud. It's a building. And buildings burn.

The cloud is a building

We talk about "the cloud" as if it were weather — diffuse, everywhere, immune to local disaster. The metaphor does a lot of quiet work for the industry, because it suggests something you can't quite touch also can't quite break. But every file you "put in the cloud" sits on a spinning disk or a flash chip in a specific room, in a specific building, on a specific street. That building has a roof, a power feed, a battery room, and — as Strasbourg demonstrated — a fire load.

Datacenters are engineered against fire with extraordinary care: gas suppression systems, compartmentalization, redundant power paths, batteries that can carry the load if the grid fails. The OVH campus was no exception; it was a modern, well-run facility. And still the fire started, still it spread, still a whole building was lost. The investigators spent months on the cause, focusing on the building's electrical systems — the same systems that keep the servers alive are, in a bad hour, what can kill them. That irony is not unique to OVH. Every datacenter concentrates enormous electrical energy in one place, because that is what computing at scale requires. Concentration is the business model. Concentration is also the risk.

What burned, and what didn't come back

Here's the part people misunderstand: the customers who lost data permanently didn't lose it because OVH was careless. Many lost it because their backups were in the same building.

A backup that lives in the same facility as the original is not a backup. It is a second copy waiting for the same fire. When SBG2 burned, customers whose redundancy was "another rack in SBG2" discovered that redundancy and duplication are not the same thing. Real duplication means separation — different building, different power grid, different failure mode. A meaningful number of OVH customers had never made that separation, because the provider made it easy not to think about. Everything was in one dashboard, one bill, one place. Convenience had quietly replaced resilience.

The uncomfortable truth of the incident is that the technology worked exactly as designed for most people. SBG3 and SBG4, on the same campus, were shut down preventively and later restored. Customers with offsite copies rebuilt. The ones who lost everything lost it because of an assumption, not a bug: the assumption that a professional facility is, by itself, a guarantee.

It never is. A guarantee is a process, not a place.

The 3-2-1 rule, and the one nobody does

The storage industry has a mantra for exactly this scenario. It is called 3-2-1: keep three copies of your data, on two different media types, with one copy offsite. Everyone in the business can recite it. Almost no private individual actually follows it.

The first two are easy. Three copies, two media types — that's a hard drive and a cloud sync, or a NAS and an external disk. People manage that much. It's the final "1" — the offsite copy, physically separated, ideally offline — that quietly gets skipped. It gets skipped because it's the only part of the rule that requires ongoing effort: a second location, a routine, someone who tends the copy when you're not thinking about it.

That skipped "1" is precisely what Strasbourg punished. The customers who had it rebuilt in days. The customers who didn't, didn't. The rule doesn't care whether the threat is fire, flood, ransomware, or a backhoe through a fiber trench. It only cares whether your copies share a fate.

"But my provider handles that"

The most common objection we hear goes like this: surely a serious provider replicates across regions, so I don't need to think about it. Sometimes that's true — if you paid for it, configured it, and verified it. But many people are on the default tier, where replication means "within the campus," and the campus is, as we've established, a building.

There is a deeper problem with outsourcing the thinking entirely. When the provider handles everything, you stop asking questions — which version is where, how old is the offsite copy, has anyone tested a restore this year. The provider's dashboard is green, so everything must be fine. Dashboards are green until they aren't. After the fire, some OVH customers discovered their "geo-redundant" setup had never been what they assumed. The gap between what people believe their setup does and what it actually does is where data goes to die.

This is why old hands in the storage business are boring about verification. Trust the provider's building; verify your own copies. The craftsman's version of the rule: if you haven't personally confirmed that a copy exists somewhere a single fire can't reach, you don't have a backup. You have a hope.

Someone has to be the one who remembers

Seneca wrote that we should expect misfortune and plan for it — not out of pessimism, but because planning is what turns disaster from an ending into an interruption. The OVH fire is a case study in that distinction. For the customers with a real offsite copy, March 10 was a terrible week and then it was over. For the customers without one, it was an ending.

The lesson isn't "don't use the cloud." The cloud is a fine tool for access and collaboration. The lesson is that anything you can't afford to lose needs a copy that shares no fate with the original — different building, different grid, different threat model — and a routine that keeps that copy alive: powered on periodically, verified, replaced before the media ages out.

Almost nobody does that routine themselves, because the whole point of a backup is to stop thinking about it. That's the paradox, and it's why custody exists as a job. Not a dashboard, not a sync client — a person, in a building you've never visited, whose entire job is the unglamorous tending: numbering, sealing, photographing, inspecting seals twice a year, logging — and verifying the media itself only when its owner asks, in writing. The "1" in 3-2-1, done by someone who does nothing else.

Fire taught Strasbourg's customers what the old hands already knew. The cloud is a building. Buildings burn. Plan like it.

Everkeep provides offline custody for encrypted drives you own and encrypt yourself. Every slot includes twice-yearly condition checks — seals inspected, photographed, and logged; your drive is never powered on unless you ask in writing. Two copies, two separate locations. $199/year per device slot.