Picture the internet and you probably picture something in the air. The cloud. Wireless. A signal bouncing off a satellite. The truth sits at the bottom of the ocean. Almost all of the world's intercontinental data crosses roughly five hundred and seventy fiber-optic cables lying on the seabed, each about the width of a garden hose, carrying somewhere between ninety-five and ninety-nine percent of the traffic between continents. Satellites handle a sliver of the rest. The network that feels weightless and everywhere is, in physical fact, a few hundred strands of glass that a ship can sever with its anchor.
The Mesh That Was Not
The internet was designed to survive damage. Its founding idea, drawn from Paul Baran's work on distributed communication in the 1960s, was a network with no center, one that would route around a break and keep running. That design is real, and it lives in the routing software. The problem is the layer underneath it. Logical resilience rides on physical cables, and the cables do not branch the way the protocols pretend. They funnel. More than ninety percent of all communication between Europe and Asia passes through the Red Sea. A comparable share of Taiwan's connection to the world runs through the Luzon Strait. The cables that cross the Red Sea cannot even reach open water without first coming ashore in Egypt and running overland between the Mediterranean and the Gulf of Suez, where the operators pay rent to a single national carrier for the crossing. Whole continents hang on pinch points a few miles wide.
The repair system is thinner than the cables it protects. The entire planet is served by barely sixty cable ships, fewer than twenty of them dedicated to repair, and the fleet is old, much of it converted from oil-and-gas and ferry work and nearing the end of its service life. Around two hundred faults happen every year, more than three a week. The average repair took about forty days in 2023, and the gap between how fast cables are laid and how fast they can be fixed is widening. The internet absorbs a routine break without anyone noticing. It has very little slack for several at once. On current trends about half of that fleet will reach the end of its working life by 2040, and the industry estimates it needs billions of dollars in new ships simply to keep pace.
When The Strings Are Cut
The self-healing network does not always heal. In 2022 an undersea volcano severed Tonga's single cable to the outside world, and the country went dark for weeks until a batch of satellite terminals restored a thin connection. One cable was all it had, so one break was all it took.
Taiwan has fifteen, and that is most of what stands between the island and a near-total blackout. In 2023 two of them, serving the Matsu islands, were cut by passing Chinese vessels, and fourteen thousand residents spent more than fifty days with internet barely usable while the island waited for a repair ship. Between 2022 and 2025 Taiwan recorded twenty-eight separate cable incidents, several of them suspected deliberate.
The pattern is not confined to contested waters. In February 2024 a cargo ship struck off Yemen drifted with its anchor scraping the seabed and severed three cables in the Red Sea, knocking out about a quarter of the traffic through one of the most concentrated corridors on Earth, and the repair stalled for weeks waiting on permits to work in a war zone. That November, two cables were cut in the Baltic in a single day, and European governments called it sabotage by ships dragging anchors across the seabed.
A few cables, a few chokepoints, a few aging repair ships. And the new lines are increasingly owned outright by the same handful of cloud companies that already own the data centers. A decade ago the big content providers used about a tenth of the world's international capacity. Today Google, Meta, Amazon, and Microsoft account for more than seventy percent of it, and Google alone holds stakes in more than thirty cables, several of which it owns by itself. Building a major line without one of them has become difficult. Concentration stacked on concentration.
The Architecture That Needs The Backbone Less
There are real efforts to fix this, and they climb a ladder from patching the wires to rethinking what depends on them. At the bottom, the industry is laying more diverse routes. The Far North Fiber project will run a cable through the Arctic to link Europe and Asia without touching the Red Sea or the South China Sea, a billion-euro line backed by the European Union, roughly four times the cost of an ordinary Atlantic crossing, built because melting ice finally opened the route. Governments are funding sovereign repair capacity, from the United States Cable Security Fleet to a British push for a dedicated repair ship by 2030. This helps, but only states and the largest companies can lay a cable, and the repair fleet keeps shrinking against demand.
One step up is a parallel path. Low-orbit satellites are the obvious backup, and they have already proven it, restoring Tonga and standing by for Taiwan, which is deliberately courting several constellations rather than betting the island on one. But satellites carry only around five percent of what the cables move, and every constellation is owned by a single company. That trades a chokepoint on the seabed for a monopoly in orbit.
The deepest answer is the one that does not try to fix the cables at all. You cannot decentralize the ocean. What you can change is how much depends on crossing it. When computation and data sit close to the people using them, a severed cable or a hostile owner takes less down with it. Content that already lives near you does not need to cross a strait to reach you. This is the same argument that runs through why the cloud keeps failing and how ordinary laptops can outperform a data center: the centralized path is the fragile one, and resilience comes from spreading the work out.
The internet's decentralization was always a story about its protocols, not its wires. The mesh is real in the routing tables and a fiction on the seafloor, where the whole of it narrows to a few hundred strands through a few stretches of water owned by a few companies. Resilience will not arrive as a better cable map. It arrives when we need the map less.