Undersea Cables: The Internet's Physical Layer

June 9, 2026 19 min read Infrastructure of Civilization

submarine-cablesnetworkinginfrastructurefiber-opticsbgpgeopolitics

Contents

Every diagram of the internet you have ever seen is a lie of abstraction. The cloud is not a cloud. When you open a video call from Denver to Frankfurt, your packets do not ascend to a satellite; they dive into the Atlantic inside a polyethylene tube about as thick as a garden hose, lying on the abyssal plain four kilometers down, and they do it because nothing else comes close on cost, capacity, or latency. Submarine fiber carries over 99% of intercontinental data traffic. The satellite constellations that dominate headlines carry a rounding error by comparison — Starlink’s entire system throughput is in the low hundreds of terabits per second globally, which is roughly what two modern cables deliver across a single ocean.

As of 2026 there are around 600 cable systems active or under construction, touching about 1,700 landing points, totaling roughly 1.4–1.5 million kilometers of cable — enough to wrap the equator 35 times. This post is about that machine: what a cable physically is, how the capacity math works, how you power amplifiers in a place no technician will ever visit, how the cables break (about 200 times a year, almost never because of sharks), who owns them now (increasingly not telcos), and what a cable cut actually looks like from inside your routing table.

What a Cable Physically Is

Strip the mystique and a modern submarine cable is a remarkably small object. In deep water — beyond about 2,000 meters, where nothing drags an anchor and no trawler reaches — the standard “lightweight” cable is 17–21 mm in diameter. Inside, from the center out:

        Deep-water lightweight cable (~17-21 mm diameter)

        ┌────────────────────────────────────────────────┐
        │  polyethylene outer sheath (insulation)        │
        │  ┌──────────────────────────────────────────┐  │
        │  │  copper tube (power conductor, ~15 kV)   │  │
        │  │  ┌────────────────────────────────────┐  │  │
        │  │  │  steel strength-member wires       │  │  │
        │  │  │  ┌──────────────────────────────┐  │  │  │
        │  │  │  │  steel tube + gel            │  │  │  │
        │  │  │  │   ~12-24 fiber PAIRS         │  │  │  │
        │  │  │  │   (each strand: 125 um glass)│  │  │  │
        │  │  │  └──────────────────────────────┘  │  │  │
        │  │  └────────────────────────────────────┘  │  │
        │  └──────────────────────────────────────────┘  │
        └────────────────────────────────────────────────┘

   Near shore, add: 1-2 layers of galvanized steel armor wire
   wrapped in tarred polypropylene -> up to ~50 mm diameter

The optical payload — a dozen or two hair-thin glass strands — occupies a few millimeters at the core. Everything else is there to keep seawater, tension, and teeth away from it. The steel strands carry the laying tension (a cable being paid out into 6,000 meters of water hangs under its own weight for kilometers); the copper tube is a power conductor, of which more below; the polyethylene is the dielectric that lets that conductor sit at thousands of volts relative to the ocean.

Closer to shore the design changes completely. In water shallower than about 1,500–2,000 meters the cable picks up one or two layers of galvanized steel armor wire and grows to double or triple the diameter, because the shallow continental shelf is where the danger lives: anchors, otter trawls, dredges. On the final approach to the beach, the cable is buried — typically one to three meters under the seabed, by a towed sea plow or a jetting ROV — and the trade-off is honest: burial and armor multiply cost per kilometer many times over, so operators armor exactly as far out as the threat statistics justify and not a meter further. In the deep ocean the cable simply lies on the bottom, unburied, trusting depth as its armor.

The Capacity Math: Pairs, Wavelengths, and the SDM Turn

A cable’s capacity is a three-term product: fiber pairs × wavelengths per pair × bits per wavelength. Each fiber pair (one strand per direction) carries on the order of 100+ DWDM wavelengths across the C-band’s roughly 4–5 THz of usable spectrum, and modern coherent transponders push 400–800 Gbps per wavelength at transoceanic distances depending on reach. That yields roughly 20–25 Tbps per fiber pair on current systems.

For two decades the industry’s instinct was to maximize that per-pair number on six or eight pairs. The interesting shift of the last few years is space-division multiplexing (SDM): accept somewhat lower spectral efficiency per pair, and spend the cable’s fixed electrical power budget on more pairs instead. Google’s Dunant (Virginia Beach to Saint-Hilaire-de-Riez, France, live February 2021) was the first long-haul SDM system: 12 fiber pairs and 250 Tbps, enabled by “pump sharing” in the repeaters — instead of dedicating pump lasers to each fiber pair, a pool of pumps amplifies several pairs, which both saves power and adds redundancy. The economics are compelling because the dominant costs (ship time, cable structure, repeater housings) barely change when you add fiber; capacity per dollar drops faster than capacity per pair.

The trajectory since then is steady: 16-pair systems are now standard for new builds, and 24-pair systems are arriving. Some notable systems for calibration:

Cable	RFS	Length	Fiber pairs	Design capacity	Ownership model
MAREA	2018	6,600 km	8	160 Tbps (200+ demonstrated)	Microsoft / Meta / Telxius
Dunant	2021	~6,400 km	12 (first SDM)	250 Tbps	Google (sole owner)
Grace Hopper	2022	7,191 km	16	352 Tbps	Google (sole owner)
Equiano	2022	~15,000 km	12	~144 Tbps	Google (sole owner)
Amitié	2023	6,783 km	16	400 Tbps	Meta-led, w/ Microsoft, Vodafone, Aqua Comms
2Africa	2023–26	45,000 km	16 (8 on trunk segs)	180 Tbps	Meta-led consortium (Vodafone, Orange, MTN, stc, Telecom Egypt, CMI, WIOCC)
Project Waterworth	~2030	50,000+ km	24	TBD (highest yet)	Meta (sole owner)

Two things to notice. First, the sheer scale of the African systems: 2Africa, completed as the world’s longest cable system, rings the entire continent with 46 landings in 33 countries and was declared fully live in early 2026. Meta’s announced Project Waterworth will exceed even that at 50,000+ km — longer than Earth’s circumference — on a 24-pair design connecting the US, Brazil, India, and South Africa while pointedly routing around the chokepoints discussed below, in deep water down to 7,000 meters. Second, look at the ownership column. We will come back to it.

Powering a Machine You Can Never Touch

Light fades. In the best deployed silica fiber, attenuation is about 0.15–0.2 dB/km, which means the signal needs re-amplification roughly every 50–80 km. A transatlantic cable therefore contains on the order of 100+ repeaters — pressure-housed erbium-doped fiber amplifiers (EDFAs) sitting on the seabed, each one a sealed cylinder engineered for a 25-year service life with zero maintenance visits, because there is no such thing as a service visit at 5,000 meters.

Those amplifiers need electricity, and the way they get it is one of the most elegant pieces of engineering in the system. The shore landing stations contain Power Feed Equipment (PFE) that drives a constant current — typically around 1 ampere — through the cable’s single copper conductor, using the ocean itself as the return path via sea-earth electrodes. Each repeater taps a few tens of volts off that series circuit with a zener arrangement, like bulbs on an old-fashioned string of Christmas lights. Because the repeaters are in series, the voltage requirement stacks up with distance: long systems need the full ~12–15 kV that the cable’s dielectric can tolerate, which is the practical ceiling — push past it and you trade capacity for insulation failures. On long routes both ends feed simultaneously at opposite polarity, placing the “virtual ground” mid-ocean and halving the voltage stress at each end:

 Shore A                                                    Shore B
 ┌──────────┐                                            ┌──────────┐
 │ PFE      │   +7.5 kV                        -7.5 kV   │ PFE      │
 │ (const.  │────■────■────■────■────■────■────■────■────│ (const.  │
 │  ~1 A)   │    R1   R2   R3  ...  virtual  ...    Rn   │  ~1 A)   │
 └────┬─────┘         repeaters    ground (mid-ocean)    └────┬─────┘
      │                ~every 60-80 km                        │
   sea earth ~~~~~~~~~~~~ ocean return path ~~~~~~~~~~~~~ sea earth

 Each repeater drops ~30-50 V off the series loop to run its
 EDFA pump lasers. Cut the copper -> the whole segment goes dark.

This design has a consequence worth internalizing: a cable fault is usually detectable and locatable from shore electrically before anyone looks at the optics, because a shunt fault (seawater contacting the conductor) shifts the voltage profile in a way that resolves the break’s distance to within a few kilometers. The PFE is also why repair crews follow strict power-safety protocols — grappling a live 15 kV conductor from a rolling deck is not a place for improvisation. SDM’s pump-sharing trick, mentioned above, exists precisely because this power budget is the binding constraint: the question for cable designers is not “how much fiber can we fit” but “how many amplifiers can we feed through one copper tube at 15 kV.”

Laying It Down, Picking It Up

Cables are laid by a small global fleet of specialized ships carrying thousands of kilometers of cable coiled in circular tanks. In deep water the ship steams along the surveyed route paying cable out over the stern, managing slack carefully — too little and the cable hangs suspended between seabed ridges, too much and it loops and kinks. Across the continental shelf the ship tows a sea plow that opens a furrow, lays the cable in it, and lets the sediment collapse back; modern plows and jetting ROVs can bury to about 3 meters and operate in over 2,000 meters of water. The shore end is floated in on buoys to the beach manhole and spliced to the land segment.

Repair is more remarkable, because it has barely changed conceptually since the telegraph era. When a cable faults mid-ocean, the repair ship locates the break (electrically, then optically with OTDR), then drags a cutting grapnel across the route to sever the cable at the fault, recovers one end with a holding grapnel, tests it, buoys it off at the surface, recovers the other end, splices in a length of spare cable from the ship’s tank — fiber splicing under a microscope, on a ship, in swell — and lowers the “final splice” back to the seabed, re-burying it by ROV if the original was buried. A routine repair takes one to two weeks of ship time once the ship is on station; getting on station is the catch, because permits, weather, and ship availability can stretch the calendar to months.

And ship availability is becoming the system’s quiet crisis. The global fleet capable of this work numbers only about 60 vessels, with an average age around 20 years;19 are over 30 years old, and roughly 65% will reach end-of-life within 15 years. TeleGeography estimates about $3 billion in new ship investment is needed just to maintain current repair service levels. The internet’s physical layer depends on a few dozen aging ships crewed by a profession that takes a decade to master, and almost nobody budgets for it.

How Cables Break: ~200 Faults a Year, and the Shark Libel

Cables fault constantly and the internet mostly doesn’t notice. The International Cable Protection Committee’s long-running statistics put global faults at roughly 150–200 per year, a number that has stayed stable even as the cable plant has grown. The cause distribution is dominated by us, not nature: fishing gear and ship anchors account for roughly 70–85% of faults, concentrated almost entirely in water shallower than 200 meters. Abrasion, component failure, and natural events (submarine landslides, earthquakes — the 2006 Hengchun earthquake off Taiwan snapped many cables at once) make up most of the rest. Nearly half of annual faults worldwide occur in Southeast Asian waters, where intense fishing overlaps the densest cable corridors.

Now, the sharks. The “sharks eat the internet” story resurfaces every few years, usually with a 2014-era video of a shark mouthing a cable. The ICPC’s actual data: fish and shark bites caused a known, documented problem in the telegraph and early coax era — at least 28 cable faults between 1901 and 1957 — and accounted for less than 1% of faults in the fiber era up to 2006. Since 2006: zero recorded shark-attributable faults. Modern cable sheathing solved it decades ago. Sharks are statistically innocent; the guilty party is a fishing trawler or a dragged anchor, roughly 150 times a year, every year.

The honest framing is that the system was engineered around this failure rate rather than against it. Operators buy capacity on multiple geographically diverse cables precisely because any individual cable has an expected several-faults-per-decade lifetime. Resilience lives in the mesh, not the strand — which works brilliantly until many strands share one strait.

Chokepoints: Where the Mesh Collapses to a Line

Geography concentrates cables ruthlessly. Three chokepoints matter most.

The Red Sea / Suez corridor is the worst. The shortest viable path between Europe and Asia funnels well over a dozen major systems through the Bab el-Mandeb strait and up the Red Sea — narrow, shallow, heavily anchored, and lately a war zone. In February 2024, the cargo ship Rubymar, abandoned after a Houthi missile strike, dragged its anchor across the seabed and severed three cables (Seacom/TGN-EA, EIG, AAE-1), disrupting an estimated 25% of Red Sea-corridor traffic; repairs were delayed for months by the question of which Yemeni authority could even issue permits. In September 2025 another round of cuts near Jeddah (SMW-4, IMEWE) had Microsoft warning Azure customers of elevated latency on Europe–Asia paths. Nothing about the geography is fixable; every alternative — around the Cape of Good Hope, or terrestrial routes across contested ground — adds thousands of kilometers or new political dependencies.

The Luzon Strait and the waters around Taiwan funnel most of East Asia’s trans-Pacific connectivity. Taiwan itself depends on roughly 14 international cables, and the vulnerability is not hypothetical: in February 2023, two Chinese vessels — a fishing boat and a freighter — severed both cables serving Taiwan’s Matsu Islands within days of each other, leaving the islands on degraded microwave backup for over 50 days while they waited for a repair ship. Subsequent incidents with Chinese-flagged vessels loitering near Taiwanese cables have turned “accidental fishing damage” into a category that security analysts no longer take at face value.

The Baltic Sea has become the test range for deniable cable interference. In November 2024 the Chinese bulk carrier Yi Peng 3 crossed the BCS East-West Interlink (Sweden–Lithuania) and C-Lion1 (Finland–Germany) with its anchor down, cutting both, AIS dark for hours. On Christmas Day 2024 the tanker Eagle S — part of Russia’s sanctions-evading shadow fleet — dragged its anchor for roughly 100 kilometers, cutting the Estlink 2 power interconnector and several Finland–Estonia telecom cables before Finnish special forces boarded the ship. Attribution in each case sits in the deniability gap between negligence and sabotage, which is precisely the point. Anchor-dragging is the perfect gray-zone weapon: it is identical, on the seabed, to the accident that happens 60 times a year anyway.

The chokepoint problem extends ashore. Cable landing stations — nondescript buildings near beaches, holding the PFE, the line terminating equipment, and the splice between sea and land plant — are single points where multiple billion-dollar systems converge into one fenced lot. Several countries’ entire connectivity terminates in a handful of such buildings, and they have moved sharply up national critical-infrastructure lists. The trust assumptions baked into the internet’s protocols were never designed for an adversary with a grapnel, a flag of convenience, and plausible deniability — a theme that runs all the way back through the network’s history, as covered in the post on internet trust from ARPANET to BGP.

Who Owns the Bottom Layer Now

For the first 150 years — telegraph, coax, early fiber — submarine cables were built by consortia of national carriers: a dozen telcos sharing cost and capacity in clubs like SEA-ME-WE. That model still exists (2Africa is structurally a consortium, albeit one Meta organized), but the center of gravity has moved decisively. Google wholly owns Dunant, Equiano, Grace Hopper, Curie, Firmina, and more. Meta is sole owner of the planned 50,000 km Waterworth. Microsoft and Amazon buy fiber pairs outright on systems like MAREA and Amitié.

The numbers are stark: content providers now account for around 90% of used capacity on the transatlantic route, and hyperscalers’ share of used international bandwidth overall has gone from negligible in 2010 to roughly three-quarters today. Practically every new long-haul cable has at least one of Google, Meta, Microsoft, or Amazon as an anchor owner or buyer.

The trade-offs cut both ways. On the positive side: hyperscaler money has funded a build-out the telco consortium model never would have — Equiano and 2Africa measurably increased Africa’s connectivity and lowered transit prices, and private cables are typically built faster and to more aggressive technical specs. On the negative side: an increasing share of intercontinental capacity exists primarily to connect one company’s data centers to themselves, is not on the open market, and concentrates infrastructure decisions — routes, landing countries, technology choices — in four boardrooms. When a hyperscaler-owned cable routes around a country, that country doesn’t just lose capacity; it loses negotiating leverage over the physical internet for a generation. Regulators have noticed; so have navies.

Latency Is Physics, and Routes Are Destiny

Light in silica fiber travels at about 200,000 km/s — the glass’s refractive index of ~1.47 costs you a third of vacuum speed, a tax levied on every photon. That single number explains most of the latency table of the global internet. A 6,400 km transatlantic cable imposes ~32 ms one-way, ~65 ms RTT, before a single router queue. Singapore to London via the Red Sea (~16,000 km of cable) is ~160 ms RTT in glass alone; force that traffic around the Cape of Good Hope after a Red Sea cut and you add several thousand kilometers, which is 30–50 ms of pure, unnegotiable propagation delay. This is why cable routes are a competitive product: a straighter great-circle path is worth real money to trading firms and real user experience to everyone, and it is why CDNs work so hard to keep content on your side of an ocean in the first place — the entire logic of anycast is about not paying the propagation tax at all.

The one genuinely interesting attack on the physics is hollow-core fiber: guide the light through air instead of glass and it propagates ~47% faster, cutting latency by about a third. Microsoft acquired hollow-core pioneer Lumenisity in 2022 and is deploying 15,000 km of HCF across Azure; in September 2025, Microsoft-backed researchers published a hollow-core design (DNANF) with 0.091 dB/km loss — below the ~0.14 dB/km floor of the best silica fiber, the first time that number has been beaten since the 1980s. Today’s HCF is a terrestrial, metro-and-backbone technology, and a transoceanic hollow-core cable still faces unresolved questions about splicing, repair at sea, and amplification at scale. But for the first time there is a credible path to a transatlantic RTT starting with a 4 instead of a 6, and the industry is quietly taking it seriously.

What a Cable Cut Looks Like From Your Routing Table

Tie this back to the layer where most of us actually work. When a cable is cut, no alarm reaches your NOC saying “anchor drag, 14.2 km off Djibouti.” What you see is BGP doing its job, abruptly. The transit providers riding the dead cable withdraw or re-advertise prefixes; routes that resolved through one AS path yesterday now resolve through a longer one; traffic shifts onto surviving cables and terrestrial detours that are simultaneously absorbing everyone else’s rerouted flows. The observable symptoms, in order: a burst of BGP churn, an RTT step-change on affected paths (tens to hundreds of milliseconds, depending on how much geography the detour adds), then congestion symptoms — loss, jitter, throughput collapse — on the alternates as they saturate. During the 2024 and 2025 Red Sea events, monitoring outfits like Kentik and Cloudflare Radar could watch the whole sequence: AS paths pivoting from Suez-corridor transits to trans-Pacific or Cape-route alternatives, with latency from South Asia to Europe stepping up by 100+ ms and staying there for weeks, because repair ships and permits move on maritime time, not internet time.

If you operate anything latency-sensitive across oceans, the practical lessons are concrete. Path diversity means cable diversity, not provider diversity — two transit contracts that both ride SMW-4 through the Red Sea are one path with two invoices, and only route-level due diligence reveals it. Watch RTT baselines per region, because a 60 ms step is a seabed event, not a software event. And understand your providers’ BGP behavior under stress — how fast they converge, where their backup capacity actually goes — because that, not the SLA document, determines your bad week. The mechanics of all of this are covered in the BGP for engineers post; the submarine layer is simply the reason those mechanics get exercised about 200 times a year.

Verdict

The internet’s intercontinental layer is a few hundred garden hoses lying on the seabed, powered like a string of Christmas lights at 15 kV, amplified every 70 km by sealed boxes no human will ever touch again, and repaired with grapnels by an aging fleet of 60 ships. It is simultaneously a triumph of engineering — 25 Tbps per fiber pair across 6,000 km of ocean, faults located electrically to the kilometer — and a system of startling concrete fragility: ~200 breaks a year, chokepoints a single dragged anchor can darken, repair logistics measured in months, and ownership consolidating into four companies. The myths get it exactly backwards. Sharks haven’t bitten a cable fault into existence in two decades; the real threats are anchors, trawlers, deniable gray-zone seamanship, and underinvestment in repair ships. For engineers, the takeaway is to treat the seabed as part of your dependency graph: latency is geography, geography is cable routes, and the day your dashboards step up 100 ms is the day the physical layer reminds everyone it exists.