The Physical Internet Runs on the Seafloor

The intercontinental internet is a physical object lying on the ocean floor. Roughly 99 percent of international data traffic travels through submarine fiber-optic cables, not through satellites. Satellite constellations matter at the edges and for specific coverage gaps, but the trunk of global connectivity, the paths that carry data between continents, is a network of a few hundred subsea cables spanning well over a million kilometers of seabed. Any serious model of where AI compute can be delivered globally has to start from this physical fact, because it means intercontinental reach is not an abstract service but a set of specific glass threads with specific owners and specific routes.

Consider the anatomy of a submarine cable. At its core it is a small number of fiber pairs, historically a handful, in newer systems a dozen or more, wrapped in protective layers of steel wire, copper for power, insulation, and waterproofing. The whole assembly is often no thicker than a garden hose in the deep ocean and armored more heavily near shore where hazards are greater. The cable carries not only the fibers but electrical power, delivered from shore, to run repeaters. Repeaters are optical amplifiers spaced along the cable, roughly every several tens of kilometers, that boost the signal so it can survive a journey of thousands of kilometers. A single trans-Pacific cable is a continuous engineered system tens of thousands of kilometers long with power and amplification integrated along its entire length, a single object of remarkable fragility given what depends on it.

The cable terminates at a landing station, a specific building on a specific stretch of coast where the submarine cable comes ashore, connects to terrestrial fiber, and joins the land network. The landing station is the hinge between the subsea and terrestrial worlds, and it is a critically important and often overlooked point. A continent's connectivity is only as good as its landing stations and the terrestrial fiber that carries traffic inland from them. A cable landing far from where compute or users sit still requires a long domestic haul, with its own latency and diversity questions, exactly the terrestrial analysis of the earlier modules applied to the stretch between the beach and the data center. The ocean crossing is the glamorous part; the last stretch onshore is where connectivity is often made or lost.

Latency across oceans is set by distance and route, and it is large enough to shape architecture. Light in subsea fiber obeys the same roughly five-microseconds-per-kilometer floor as terrestrial fiber, so a trans-Pacific path of over ten thousand kilometers imposes a one-way latency on the order of tens of milliseconds before equipment overhead, and a round trip is more. That is why cable routing is not a detail. A more direct path between two regions is a durable latency advantage that a faster computer on either end cannot recover, and it is why operators care intensely about which cities a new cable directly connects. Shortest viable path wins, and it wins permanently, because the geography does not move.

Capacity, by contrast, has grown spectacularly. Modern systems carry hundreds of terabits per second across a single cable, and each new generation raises the ceiling, again largely through better endpoint and repeater technology on new builds. This abundance of raw capacity is real, but it can mislead. Capacity is plentiful on the busy, well-served routes between major markets and thin on the routes that fewer cables serve. Global connectivity is not uniform; it is a map of thick corridors and thin ones, and the thin ones constrain what workloads a region can realistically host regardless of how much aggregate capacity exists somewhere on the planet.

My read: the single most important consequence of the seafloor reality is that a nation's or region's ability to participate in AI compute is partly determined by cables it does not control and may not even host. A country with many diverse cable landings and short domestic haul to its data centers has first-class access to global compute and can credibly host workloads that serve other regions. A country served by one or two cables, landing at one or two points, with long domestic backhaul, is structurally disadvantaged no matter how much power or land it offers. Connectivity geography, set by where cables were laid and where they land, becomes a determinant of digital economic opportunity, and it is set by decisions made largely offshore and often years earlier.

The durable framework: to assess any region's suitability for AI compute from the network side, ask how many independent subsea systems serve it, how diverse their landing points are, how direct their routes are to the markets that matter, and how far the landing stations sit from the compute. These are the oceanic versions of the five site questions from the first module. The physical internet on the seafloor is not background infrastructure to be assumed; it is a first-order variable, and the next lesson examines who now owns it and the chokepoints that make some of these routes dangerously fragile.