“The Net interprets censorship as damage and routes around it.”

That line is credited to John Gilmore, and Time printed it in December 1993, back when the thing being routed around was Usenet. It became the internet’s founding article of faith: the network is a mesh, no single point matters, cut a wire and the packets find another way. A generation of engineers and activists repeated it as though it were a law of physics.

It was never a law of physics. It was a claim about one specific layer — the logical one, where routers gossip about paths and traffic flows toward whichever link is still up. And at that layer it’s still sort of true. But the sentence quietly implies something it never guaranteed: that the damage is always local, and that there’s always another way. Look at where the actual bytes travel, and both halves fall apart.

The physical internet is a few hundred cables

Strip away the abstraction and the intercontinental internet is glass on the seabed. Roughly 99% of the world’s data between continents moves through submarine cables — not satellites, not anything else. TeleGeography, which maps this for a living, counts more than 500 cable systems in service or planned, laid across something like 1.4 million kilometers of ocean floor.

That sounds like a lot of redundancy. It isn’t, because the cables don’t spread out evenly. They follow the cheapest seabed and land at the same handful of stations, so they bunch up in narrow corridors: the Red Sea and the Suez approach, the Luzon Strait between Taiwan and the Philippines, the Strait of Malacca, the Channel into northern Europe. An estimated 17% of the world’s internet traffic passes through the Red Sea corridor alone. These are chokepoints in the literal sense — geography squeezes dozens of independent cables into a strip a ship’s anchor can rake across in an afternoon.

This is the opposite of a mesh. It’s a hub-and-spoke drawn by economics, and the hubs are places, not companies.

What happens when you cut one

February 2024. Three cable systems in the Red Sea — AAE-1, Seacom, and EIG — were severed within days of each other. The suspected cause wasn’t a submarine or a special-forces team; it was the anchor of the Rubymar, a cargo ship that had been hit by a Houthi missile and abandoned, drifting for days before it sank, its anchor apparently dragging across the seabed the whole way. Roughly a quarter of the traffic between Europe and Asia had to go somewhere else.

Here’s where “routes around it” gets tested. The traffic did reroute — that’s the logical layer working. But rerouting doesn’t create capacity; it just crowds everyone onto the paths that survived, which are longer and already full. Latency jumped. Cloud services in the region degraded. And the physical fix wasn’t a config change: it needed a specialized ship to sail to the spot, grapple a cable off the bottom, haul it up, splice it, and lower it back. In the Red Sea in 2024, that took months — held up not by the splicing, which is routine, but by the permits and the plain danger of parking a repair ship in a war zone. Rerouting bought time. It didn’t route around anything; it waited.

The Baltic has become the same lesson on repeat. On Christmas Day 2024, the tanker Eagle S dragged its anchor for something like a hundred kilometers along the seabed, tearing up the Estlink 2 power link between Finland and Estonia along with several telecom cables. Estlink 2 wasn’t back until the following August — more than seven months, at a repair cost reported near €60 million. A month earlier, the cargo ship Yi Peng 3 had dragged its anchor for roughly a day and a half and cut two data cables, including C-Lion1 between Finland and Germany. Whether these were sabotage or spectacular negligence is still argued in court; from the packets’ point of view it doesn’t matter. An anchor is an anchor.

And the small places feel it worst. In early 2023, the two cables connecting Taiwan’s Matsu Islands were both cut — by a Chinese fishing vessel and a cargo ship, according to Taiwanese authorities, who stopped short of calling it deliberate. The islanders spent months on a patchy backup microwave link, unable to reliably send a text or run a card payment. According to Chunghwa Telecom, those cables had been cut 27 times in the previous five years. When your redundancy is “two cables” and both are in the same busy strait, “routes around it” is a slogan you can’t eat.

The logical layer fails too, just faster

Lest this sound like only a wet-cable problem: the layer Gilmore was actually talking about deletes itself sometimes, too. In October 2021, Facebook withdrew the BGP routes to its own network during a botched maintenance change. Because the servers that answered DNS for Facebook were inside the network whose routes had just vanished, they became unreachable — so nothing could even look up where Facebook was, let alone route to it. Facebook, Instagram, and WhatsApp were gone from the internet for about six hours. The network didn’t route around the damage. The damage was the routing.

The pattern underneath both stories is the same: we kept the mesh diagram on the whiteboard while the real thing concentrated. Traffic funnels through a few chokepoints on the seabed and a few providers in the cloud, and each act of concentration was individually rational — cheaper cable routes, better economies of scale — and collectively fragile.

The repair fleet is smaller than you think

The part that should worry you most isn’t the cutting; it’s the mending. The entire world is served by a small fleet of specialized ships that can lay and repair submarine cables, and it is aging. A 2025 report by TeleGeography and Infra-Analytics for the SubOptic Association found that by 2040 about two-thirds of the cable maintenance ships will have reached the end of their service lives, and put the bill for replacements plus a few extra ships at around $3 billion — money nobody is obviously lined up to spend. A routine repair takes weeks; a hard one takes months. There is no cloud region for boats.

So the honest version of the founding faith is narrower than the slogan. At the logical layer, on a good day, the internet still reroutes around a dead link — until the reroute has nowhere uncongested to go, or until the thing that died was the routing itself. At the physical layer, where 99% of the bytes actually live, the internet routes around nothing. It waits for a ship to show up, in weather, sometimes with a government’s permission, to pull a cable off the bottom of the sea and splice it back together by hand.

We told ourselves the network was too distributed to break. What we built is a network distributed enough to look unbreakable on a diagram, and concentrated enough that a drifting anchor can take out a quarter of a continent’s traffic and leave it that way for months. The damage is real, it’s physical, and increasingly it doesn’t route around. It just waits.