A router in El Dorado, Arkansas, may have grounded every airplane in America.
That is the FAA’s preliminary account of Friday, November 13. A contractor had been installing a new router at a radio site there when old equipment across the network began to drown in data, about a thousand connection attempts every four or five seconds, apparently from El Dorado and from a radar site in Amarillo. At about 3:30 that afternoon, seventeen of the country’s twenty-two en route control centers went unstable. The FAA stopped every departure in the nation, only the second time since September 11. Some 3,500 airplanes were in the sky.
Every one of them landed safely. Nobody will write a headline about that, but it’s the story. Some radio frequencies and radar feeds stayed alive, and behind them stood controllers, pilots, and a book of procedures written on the assumption that the machinery would sometimes quit. The airspace system’s real redundancy was never in the wires. It was in the people.
The work was being done at midday, not at night. By 8:15 the airplanes were moving again, though more than twelve thousand flights had been canceled. A router, a flood, a Friday afternoon.
The cloud has been having the same Friday all year, and I’ve stopped being surprised by it. A configuration change. A latent race condition in a DNS system. A script that deleted what it was never meant to touch. One tracker counts nineteen major outages at the four biggest providers so far in 2026, and November isn’t over. Almost none were attacks, which I find more unsettling than if they had been. An attack has a villain, and a villain can be hunted. A routine change, made by competent people and traveling through a tightly coupled system faster than anyone can stop it, has only us.
At Visa we had a name for the worst case: a worldwide outage. I’ve always liked how flat that sounds. A named thing can be discussed, and a discussed thing can be designed against. So we designed against it. Multiple datacenters, each able to handle the entire world’s traffic. Redundant networking. Cost mattered, but it was never the barrier, because everyone understood what the network was for. Continuous availability wasn’t a feature. It was the premise.
The hardest enemy was software. You can duplicate hardware. But the copies run the same code, and a bug is the one failure that’s perfectly redundant: it arrives everywhere at once, which is exactly how those cloud outages unfolded. So we tested, and tested again before committing a rollout. Nothing magical. Just an institution that had decided it would rather be slow than wrong.
That word, decided, is the whole difference. Resilience isn’t what a system has when it’s well connected. It’s what a system has when someone chose to pay for it, and kept paying through the years when nothing went wrong and the spreadsheet said the money could be spent better elsewhere. Efficiency is what you optimize for when you believe the bad day isn’t coming. And for twenty years we’ve believed it: a handful of clouds, shared edge networks, a single contractor, copper quietly pulled from the walls. Each was a sensible bet. Each narrowed the places where a bad Friday could begin.
The cruel part is that we knew how. My friend Paul Baran drew the blueprint for a network that could take a punch, one in which no node mattered enough to be worth destroying. We built it. Then, over the decades, we gathered it back into a few buildings, a few vendors, a few routers in a few rooms.
I don’t know what the controllers were thinking that afternoon. I suspect it was the airplane in front of them. At Visa, nobody was ever supposed to need to be that good.