The
Garden Hose

At the centre are pairs of silica fibres, each one 125 microns of glass in a plastic coat, floating in a water-blocking gel inside a hermetically welded steel tube. Around that tube sits a seam-welded copper conductor. Around that, five to eight millimetres of extruded polyethylene. That is the whole thing.

The glass is passive. Light fades as it travels, at about 0.16 decibels per kilometre, which sounds gentle and is not: across a six-thousand-kilometre Atlantic crossing the cumulative loss reaches roughly 900 decibels, which is to say the signal would vanish several hundred times over. So every seventy to eighty-five kilometres the cable is interrupted by a repeater, a pressure housing the size of a large fire extinguisher containing ten to thirty metres of erbium-doped fibre and a set of pump lasers. Fire the pumps into the doped fibre and the arriving photons trigger the excited erbium atoms to emit identical ones. The signal is not read or decoded. It is made brighter, and sent on.

Powering those repeaters is where the design turns strange. The copper conductor carries a constant direct current of between one and 1.6 amps, fed from power equipment in the landing stations at each end. One station runs positive, the other negative, so the electrical midpoint sits somewhere in the ocean. At each end a sea earth, a grid of titanium or platinised niobium buried in coastal sediment, closes a direct-current loop several thousand kilometres wide, using the Atlantic as the second wire.

Getting one onto the seabed takes two to three years, and very little of that is spent at sea. It begins with months of someone reading charts: bathymetry, fishing intensity, wind-farm leases, pipeline registers, wreck records, munitions dumps. Then comes a marine route survey, at $2,000 to $5,000 a kilometre, with a ship driving the corridor at three to five knots, firing sonar at the seabed and pushing probes two or three metres into the sediment to learn whether a plough will pass through it. Every existing cable and pipeline needs a crossing agreement with its owner, and each agreement can take three to nine months.

Four companies build the wet plant, meaning everything that goes in the water: Alcatel Submarine Networks, which the French state took into public ownership from Nokia in 2024; SubCom in New Hampshire; NEC in Japan; and HMN Technologies in Tianjin. ASN's repeaters are still assembled at Enderby's Wharf in Greenwich, on a site that has been making submarine telegraph equipment since the 1850s.

Loading a ship is done by hand. Cable cannot be spooled onto a drum without twisting, so it is coiled, or “flaked”, into circular tanks fifteen to twenty metres across, walked round in flat rings, each layer dusted with talc. The repeaters are too rigid to follow the tank radius, so they are carried on overhead monorails and lashed into padded cradles. A full cargo of five to seven thousand kilometres takes three to four weeks of round-the-clock work to load.

Then the lay, which is the fast part: six to eight knots and 150 to 220 kilometres a day in deep water, slowing to walking pace on the continental shelf, where a twenty- to thirty-five-tonne plough opens a trench and drops the cable one to three metres down. The ship pays out one to three per cent more cable than the distance it sails, so the line drapes into the seabed's contours instead of hanging across them like a tripwire. Below about 1,500 metres, burial stops. Nothing down there can reach it, and no tow wire could hold a plough at that depth anyway.

The industry's account of the result is confident and, on the record, correct. Around six hundred systems are in service. The network takes roughly two hundred repairs a year, mostly in shallow water, mostly from fishing gear and anchors, and almost none is visible to users. Traffic reroutes across surviving paths. The network is not strong. It is designed on the assumption that it will keep failing, and it does, and it is fine.

The price of a wire

In 1999 a 155-megabit circuit between London and New York leased for around $208,000 a month. By 2003 the same circuit cost about $4,400. A hundred-gigabit wavelength was about $7,500 a month in 2021 and roughly $5,400 by 2024. Demand did not fall. The price fell because supply arrived faster than traffic, and because the marginal cost of sending one more packet down a fibre that is already lit is effectively zero. Discounting surplus capacity is rational for every operator individually and ruinous for all of them together.

For anyone whose business is selling capacity, that compounds into something fatal. Payback horizons on merchant systems have stretched from five to seven years out to twelve to eighteen. Digital 9 Infrastructure, a London-listed fund that bought subsea assets without locking in long-term buyers first, was pushed into managed liquidation and sold its fibre at a discount.

The buyers who are not selling

In 2010 traditional telecoms carriers accounted for around ninety per cent of used international subsea capacity and content providers for under ten. By 2025 that had inverted. Content and cloud companies accounted for three quarters of used international bandwidth, participated in more than two thirds of planned systems, and financed effectively all announced new transatlantic capacity.

Where they co-build with carriers, the carrier partners still light their own fibre pairs and sell them wholesale. But none of the four is primarily in the business of selling you a wavelength. Google owns Dunant, Grace Hopper, Equiano, Curie and Firmina. Meta prefers majority stakes in co-builds. Microsoft buys fibre pairs before construction starts. Amazon buys slices of optical spectrum. They are converting transit they would otherwise buy forever into a depreciating asset, recorded inside network equipment, fibre and infrastructure and written off over fifteen to twenty-five years. What they get is control: capacity lit by changing a card in a landing station, and predictable latency between data-centre regions for synchronising AI training runs across hundreds of thousands of processors.

The old model survives on thinner routes. Across the Indian Ocean, West Africa and Oceania, carrier consortia, development banks and national operators still do the work. On the Atlantic, the market for capacity has become a by-product of four companies' internal cost accounting.

Which leaves Britain holding a receipt

Eleven cables run direct from Britain to North America with a combined potential design capacity of 850.8 terabits per second, according to evidence submitted to Parliament by Jeremy Steventon-Barnes, a former chief technology officer of EXA Infrastructure. Grace Hopper and the UK branch of Amitié account for 628 terabits, or 73.8% of that total, and both land at Bude. The arithmetic describes design capacity, not traffic: how much capacity is lit, available and contracted is not published.

Not one of the eleven is owned by the British state. The government's residual golden share in BT was redeemed and cancelled in 1997. Britain does have commercial UK-flagged repair vessels. What it does not have is a sovereignly controlled cable-repair ship the state can command in a crisis. C.S. Sovereign sits at Portland in Dorset, but she is a commercial asset tasked through a private maintenance agreement. RFA Proteus can survey a cable, film it and hold station over it. She cannot splice it.

On its own terms the arrangement has been a triumph. Britain got a highly connected coastline without paying the public capital bill, watched the price of moving a gigabit across the Atlantic collapse, and let American companies underwrite the next generation because they needed it for themselves. The bill arrives as decisions Britain no longer takes: where the next cable lands, which routes get lit, and whose ship comes when one breaks.

When somebody may be trying

On Christmas Day 2024 the tanker Eagle S, flying the Cook Islands flag, dragged its anchor through the Gulf of Finland and damaged the EstLink 2 power interconnector and telecoms cables. In October 2025 the Helsinki District Court held that Finland lacked jurisdiction. On 27 August 2026 the Court of Appeal reversed that ruling and sent the case back, finding that the alleged conduct occurred in Finland and that the key events were not an “incident of navigation” under the law of the sea. It convicted nobody. The decision was not final, and whether the damage was deliberate or grossly negligent remains contested.

Five weeks before Eagle S, two Baltic data cables were damaged as the Chinese-flagged Yi Peng 3 passed through the area. Swedish authorities treated the ship as relevant to criminal investigations but said they had no jurisdiction to investigate a foreign vessel in international waters. No indictment followed. That is the limit of what the official record then established.

Most of the record is not malice. In March 2024 the abandoned MV Rubymar drifted through the Red Sea after a Houthi missile strike and damaged three cable systems, disrupting roughly a quarter of the data capacity between Asia and Europe: collateral harm from a war. That same month an underwater landslide in the Trou Sans Fond canyon off Abidjan severed four cables and cut connectivity across West Africa: geology. In January 2022 a volcano cut Tonga's only international cable, and the country went dark for thirty-eight days. Government estimates say fishing and anchors cause up to 97% of faults. The shadow fleet is also full of ageing tankers with failed anchor brakes and crews who could not recover the ground tackle if they wanted to. Treating every drag as an attack buys warships to solve a problem that is mostly bad seamanship.

The delay is paperwork

Dedicated maintenance vessels spend much of the year alongside. The physical repair, grappling the seabed, recovering both ends, stripping the armour, fusing the fibres in a clean container, moulding polyethylene over the joint and lowering the loop, takes five to twelve days. A single joint is sixteen to twenty-four hours of continuous work by four people. The operation costs roughly $1 million to $3.5 million.

In Indonesian waters, cabotage rules can strand a foreign repair ship for thirty to ninety days waiting on a ministerial exemption. In the Indian exclusive economic zone, multiple clearances can take forty to sixty days. A permit can take longer than the ocean does.

The question Britain has not answered

The Joint Committee on the National Security Strategy reported in September 2025 that the government's posture relied too heavily on the assumption that having many cables is the same as being safe. Landing stations are often unfortified industrial sites in remote coastal locations. The overland fibre from the beach to data centres converges on Slough and the Docklands along road and rail corridors. The committee asked for a sovereign repair capability by 2030. The government said it was exploring options.

In May 2026 the government proposed a future white paper: a new offence of reckless damage, security duties on landing operators, emergency powers and market engagement on sovereign repair capability. Those were proposals, not enacted law. No purchase or ownership model had been chosen.

The reassuring counter-argument is real. Britain has more than eight thousand terabits of potential cross-Channel design capacity, while continental Europe has substantial transatlantic landings in France, Spain and Portugal. Cut Britain's direct Atlantic paths and, on paper, traffic goes via Europe. The words doing the work are “on paper”. Capacity that is not lit cannot absorb traffic at short notice, and no one has published what share is powered, spare and contractually available to British operators in a crisis. The number that would show whether the system is resilient is the same unpublished number in both directions.