SWITCHED IN SECONDS, MENDED IN WEEKS

Why the substrate with the best substitution time is not the safest one

Article 6 of 8  ·  Series II of III  ·  Published 18 November 2026  ·  Analysis → Forecast → Recommendations

Series II, sixth row. Each article measures one physical layer with the same three questions — concentration, criticality, substitution time — and adds a row to one comparison table. This row produces the best answer the instrument has given, and the answer is misleading. Working out why is the most useful thing in the article.

1. Fifty Days on Matsu

In February 2023 the two submarine cables serving the Matsu Islands, a small Taiwanese archipelago close to the Chinese coast, were cut six days apart. Around fourteen thousand residents lost normal internet access for roughly fifty days.

Service did not stop entirely. A microwave backup link carried a trickle, enough for text messages and intermittent calls and nothing resembling ordinary use. Students could not attend remote classes. Card payments failed. The islands were not disconnected; they were reduced to a bandwidth from two decades earlier, and held there for seven weeks.

Matsu's cables have been cut more than twenty times in five years. Taiwan–Matsu No. 3 was damaged twice between March and April 2026. The pattern is now routine enough that local authorities plan around it.

Set that against a larger event. On 6 September 2025, three systems — SEA-ME-WE 4, IMEWE and FALCON GCX — were cut near Jeddah in the Red Sea. India, Pakistan and the United Arab Emirates were affected, and Microsoft acknowledged increased latency on its Azure platform. Services were substantially restored by the following day, through rerouting.

Physical repair was expected to take weeks or months.

Median restoration time for a submarine cable fault is around forty days. For those forty days the network is not broken and it is not whole: it is running on its own redundancy, with less of it left.

2. What a Connectivity Failure Actually Does

The intuitive model of a cable cut is a blackout: the link fails, the service stops, someone fixes it. That almost never happens, and the reason it almost never happens is also the reason the real constraint goes unnoticed.

The scale of the thing

There are around 574 active and planned commercial subsea cable systems, spanning roughly 1.42 million kilometres, and they carry about ninety-nine percent of all intercontinental internet traffic. Satellites carry a rounding error of it, and will for the foreseeable future.

Faults are routine rather than exceptional: 150 to 200 occur every year, and roughly two thirds to four fifths of them are caused by fishing gear and ship anchors. Sabotage is the exception, not the baseline, and treating every fault as hostile obscures how ordinary the failure rate is.

The routing layer handles this well. Traffic shifts to other paths within seconds, usually without any user noticing, which is why a network carrying almost all intercontinental data can lose several cables in a week and produce no headline beyond a latency complaint.

Running the instrument

The question the instrument should be asking

Three questions have served five rows. Here they produce a flattering answer that would lead a reader to the wrong conclusion, so a fourth is required for this substrate and probably for others.

Restoration time: how long does the system run on its alternative before the original is repaired — and how much redundancy remains during that period?

For a grid connection, substitution time and restoration time are the same number, because there is no alternative to run on meanwhile. For shipping, they nearly coincide. Here they diverge by four orders of magnitude: seconds against forty days. The gap is the whole risk, and no amount of attention to the substitution figure will reveal it.

What happens inside that gap is the depletion of margin. A route designed with three independent paths and running on two is operating as designed. Running on one, it is a single point of failure that no diagram shows and no alarm reports, because from the outside the service is working normally.

3. Subtheme One — The Constraint Is the Repair Fleet

If restoration time is the variable that matters, then the capacity that sets it is the real chokepoint of this substrate. That capacity is a fleet of ships, and it is small, old and thinly spread.

The numbers

Roughly sixty specialised vessels exist worldwide for cable work of any kind, and fewer than twenty are dedicated repair ships. Against 150 to 200 faults a year on 1.42 million kilometres of cable, that is a maintenance fleet in the low tens servicing the physical layer of the entire global internet.

The fleet is also ageing out. About half the vessels in the global cable fleet, and nearly two thirds of the maintenance vessels specifically, reach the end of their service life by 2040. Separately, around a quarter of installed cable kilometres require retirement by 2030. Both curves point the same way at the same time.

Investment has risen — roughly $4 to $5 billion a year now, about double the figure of a decade ago — but the great majority of it goes into new systems rather than into the ships that keep existing systems alive. New capacity is commercially attractive and generates revenue. Repair capacity is a cost centre that sits idle between faults, which is precisely the economic profile that markets under-provide.

Why a repair takes forty days

The engineering is not the slow part. Locating a fault, grappling a cable from the seabed, splicing and testing it is measured in days once the ship is on station with the right cable type aboard.

The delays are elsewhere, and they are mostly administrative or geographic.

—  Ship availability. A vessel may be on another fault, in dry dock, or several thousand kilometres away. A fleet in the low tens has no spare capacity by design, and simultaneous faults queue.

—  Permits. Government approval to enter territorial waters can take a month or more. This is frequently the single largest component of the elapsed time, and it is purely procedural.

—  Security. In contested water a crew must hold position for days, sometimes near a conflict zone, and insurers and operators both have views about that.

—  Weather and depth. Seasonal windows and deep-water recovery add time that cannot be compressed.

Which reframes the problem usefully. The repair fleet is a capital constraint with a long lead time and a bad commercial case. The permitting layer is a diplomatic constraint that could be fixed in a year by people who are already in the room for other reasons, and it is where the attention currently is not.

4. Subtheme Two — Redundancy That Is Not Independent

The second thing the substitution figure conceals is that the many paths are far less separate than their number suggests. They converge in three different ways, and only the first is widely understood.

Geographic convergence

Cables follow the same narrow water that ships do, for the same reasons: the Red Sea and Bab el-Mandeb, the Luzon Strait, the Strait of Malacca, the approaches to Egypt. In February 2024, three systems — AAE-1, Seacom and EIG — were damaged in the Red Sea by a dragging anchor, disrupting roughly a quarter of internet traffic between Asia, Europe and the Middle East in a single incident.

Three systems is a meaningful fraction of what passes through that corridor, and one anchor found all three. Formally independent cables sharing a seabed corridor are correlated assets, and they are almost never modelled that way in a resilience plan.

The same holds at landing stations, where multiple systems come ashore within a few kilometres of each other because the suitable coastline is limited and the terrestrial backhaul already exists.

Ownership convergence

A newer and less discussed form. Hyperscale technology companies now control about seventy-one percent of global subsea capacity and account for roughly eighty-four percent of trans-Atlantic bandwidth demand. One holds direct ownership stakes in thirty-three systems; another in nineteen.

For a buyer this changes what diversification means. Two carriers offering independent routes may both be reselling capacity on systems owned by the same handful of firms, and a commercial decision taken by one of those owners — a route retirement, a pricing change, a reallocation of capacity to internal use — propagates to both. This is the same structure as the shared separation plant in the previous article: diversity at the invoice, convergence two tiers up.

Core and edge are different substrates

The most consequential convergence is the one between the two halves of the network, and it is the reason the Red Sea and Matsu episodes look so different.

At the core — the dense mesh between major economies — redundancy is genuine and a cut produces latency. At the edge, a territory may be served by one or two cables, and the honest description is not redundancy but a spare. Matsu lost both and spent fifty days on a microwave link. Island states, landlocked countries dependent on a single transit neighbour and remote territories all sit in this category, and for them criticality is total while the global figures say the system is robust.

A single aggregate for a network that behaves this differently at its two ends is the same error the previous article found in a commodity price index: an average across two populations that no longer resemble each other.

5. What Most Analysis Gets Wrong

Four errors recur, and each changes a decision.

—  Measuring substitution instead of restoration. Traffic reroutes in seconds, so the event looks handled. The exposure lives in the forty days that follow, during which the system runs with less margin than its design assumes and nothing reports that fact.

—  Treating every cut as sabotage. Two thirds to four fifths of faults come from fishing gear and anchors, and several high-profile Baltic incidents have been formally investigated and attributed to accident or dismissed. Reading an ordinary failure rate as an attack pattern produces the wrong defences, and the right ones — more ships, faster permits — are the same either way.

—  Counting cables as independent paths. Systems that share a chokepoint, a landing cluster or an owner fail together. A resilience plan listing route diversity without naming the corridors and the owners is counting paths that are correlated and calling them alternatives.

—  Generalising from the core to the edge. The same network that shrugs off three simultaneous cuts in the Red Sea leaves fourteen thousand people on a microwave link for seven weeks. Both are accurate descriptions of global connectivity and only one of them describes any particular place.

6. Base, Stress and Extreme

Three ways the next decade runs, so the forecasts below have a frame.

Base — routine faults, slow fleet, rising attention

Faults continue at 150 to 200 a year. Restoration times stay near their current median and worsen slightly in contested water. Governments issue strategies and fund studies; a small number of repair vessels are ordered, arriving late in the decade. Edge territories continue to suffer disproportionately and the core absorbs everything without a visible event. This is the most probable path and describes 2026 well.

Stress — simultaneous faults exhaust the queue

Several faults occur close together in a region whose repair vessels are already committed, and the queue becomes the binding constraint rather than any individual break. Restoration stretches to several months, a corridor runs on thin redundancy long enough for a second failure to matter, and a major economy experiences sustained degradation rather than a brief latency change. The consequence is not a blackout; it is a period in which the margin everyone assumed existed is demonstrably absent.

Extreme — a sustained national disconnection

A territory dependent on one or two systems loses them during a period when repair access is blocked by permitting, security or weather, and runs degraded for months rather than weeks. The economic damage is real but bounded; the more durable effect is that connectivity stops being treated as infrastructure that simply works and starts being treated as a dependency with a named restoration time — which, on the evidence of this article, would be an improvement.

7. Forecast — One Year, to end-2027

Direction. At least four publicly reported submarine cable damage incidents occur in the Baltic Sea and around Taiwan combined during 2027. Probability 0.65. Confidence: Medium.

Both areas have produced repeated incidents every year since 2023: the Baltic across 2023, 2024 and January 2026, and Taiwan continuously, with Matsu alone cut more than twenty times in five years and Taiwan–Matsu No. 3 damaged twice in a two-month window in 2026. Four across two active areas in a year is close to the established run rate rather than an escalation claim.

Second-order effect. A steady incident rate with inconclusive attribution is corrosive in a specific way: it produces neither a response nor a return to normal. Several Baltic investigations have ended in dismissal or a finding of accident, which leaves states carrying the cost of suspicion without the standing to act on it.

What weakens it. Reporting is the weak link rather than the events. Minor faults are repaired quietly and never announced, so the count depends on what operators and authorities choose to disclose — which has been rising for reasons unrelated to the underlying rate.

8. Forecast — Three Years, to 2029

Direction. At least one new government-funded or government-mandated cable repair vessel programme is announced in Europe, Japan or the United States. Probability 0.70. Confidence: Medium.

The diagnosis is now public and the arithmetic is simple enough to survive a briefing: fewer than twenty dedicated repair ships, two thirds of maintenance vessels retiring by 2040, and a commercial case that will not fund replacement because the asset earns nothing between faults. That is the standard profile of a capability that gets publicly procured, and the same logic that produced the price floor in the previous article applies here in a different form.

Second-order effect. A publicly funded repair vessel raises a question nobody has had to answer: whose cables does it prioritise when several are broken at once? Repair priority is currently a commercial matter settled by maintenance agreements. Public ownership makes it a political one, and the ordering will not be neutral.

What weakens it. Governments can respond with strategies, task forces and surveillance patrols, all of which are cheaper and faster to announce than a vessel programme and none of which shortens a repair. Monitoring the seabed is a different capability from mending it, and the first is far easier to fund.

9. Forecast — Five Years, to 2031

Direction. The global count of dedicated cable repair vessels remains below thirty at the end of 2031. Probability 0.70. Confidence: Medium.

From fewer than twenty today, exceeding thirty would require more than a fifty percent expansion of a specialised fleet within five years, against a replacement curve that is simultaneously retiring vessels. Cable ships are purpose-built with long order books and few yards. Even if several programmes are announced on the three-year horizon above, the ships largely arrive after this one.

Second-order effect. If fleet capacity is flat while cable kilometres and fault counts grow, restoration times lengthen without any single event causing it. That is the quietest way this substrate deteriorates: no incident, no headline, simply a median that drifts from forty days toward sixty.

What weakens it. Fleet counts are reported inconsistently — one source gives roughly sixty vessels for all cable work, another sixty-two cable-laying ships, and the dedicated-repair subset is defined differently again. A generous definition could cross thirty without a single new hull.

10. Forecast — Ten Years, to 2036

Direction. At least one sovereign state experiences national-scale internet degradation lasting more than seven consecutive days, caused by submarine cable damage, between 2027 and 2036. Probability 0.55. Confidence: Medium.

The mechanism already exists at sub-national scale — Matsu lost fifty days in 2023 — and the question is whether it reaches an entire state. Several island and single-cable states have the exposure, fault rates are not falling, and repair capacity is not growing. The probability sits only slightly above even because the states with this exposure are mostly small, and because satellite capacity, while a poor substitute for a cable, is improving fast enough to blunt the worst case over a ten-year horizon.

Second-order effect. A week of national degradation would do more for subsea policy than a decade of reports. It would also, on the evidence of past incidents, be attributed within hours and investigated for years — and the response would likely fund surveillance rather than repair, because surveillance is what the event appears to call for.

What weakens it. Low-earth-orbit satellite capacity is the genuine wildcard in this forecast. It cannot replace a cable's bulk capacity and will not within the horizon, but it can hold a small state above the threshold of national degradation, which is exactly where this forecast is set.

11. Signals to Watch

Five observable items, each published, each moving before the outcomes above.

—  Median and maximum time from fault to restoration, rather than fault counts. Faults are routine; the restoration figure is the one that measures whether the substrate is deteriorating.

—  Repair vessel orders and retirements, counted as hulls rather than as announced programmes. A strategy is not a ship, and the gap between the two is usually several years.

—  Bilateral or regional agreements on expedited repair permits. The cheapest available improvement, and the one that would show up in restoration times within a year of being signed.

—  Incident counts in the Baltic and around Taiwan, read alongside their attribution outcomes. The pattern of dismissals and accident findings matters as much as the raw count.

—  New systems that avoid existing chokepoints — routes around the Cape, across the Arctic, or overland alternatives. These change the correlation structure rather than adding capacity to it.

Read together: the first three measure the capacity to restore, the last two measure the exposure that creates demand for it. This substrate improves only when restoration capacity grows faster than the kilometres and chokepoints it has to cover, and neither figure is in the headline numbers.

12. Recommendations — Individuals

For most people in well-connected places this constraint is invisible and will stay that way. For anyone on an island, in a remote region, or dependent on a single transit route, the exposure is real and the preparation is simple.

Immediate — 30 days

Find out how many cables serve where you live, and whether the backup is another cable or a radio link. For most of Europe and North America the answer is reassuring and you can stop there. For an island, a remote territory or a single-cable country it is the most useful piece of infrastructure knowledge available to you, and it takes ten minutes to look up.

Build — 12 months

If you live somewhere with thin connectivity, make sure the things that must keep working do not all depend on the same link. Offline copies of essential documents, a payment method that functions without connectivity, and a plan for how your household communicates if bandwidth returns to a trickle rather than to zero. The Matsu failure mode is degradation, not darkness, and degradation is the harder one to improvise around.

Position — 3 years

If connectivity is load-bearing for your livelihood — remote work, an online business, anything that cannot be done on paper — treat the number of independent routes into your location as a real factor in where you choose to live. This is not a reason to avoid islands. It is a reason to know the number before you commit, and to know it for the place rather than for the country.

Avoid. Treating satellite service as a full substitute. It is a genuine improvement over a microwave link and it is not a cable, and the difference becomes obvious at exactly the moment you need it to be.

Why this works. The failure mode here is weeks of degraded service rather than a brief outage, and almost all the practical preparation is about functioning at low bandwidth for a long period rather than about having no service for a short one.

13. Recommendations — Business

This is the substrate where stated redundancy and actual redundancy diverge most sharply, and the audit that closes the gap is cheap.

Immediate — 60 days

Ask your connectivity providers which physical cable systems carry your traffic, and which chokepoints and landing stations those systems pass through. Most contracts promise diverse routing and very few specify physical diversity. Two providers whose routes share the Red Sea corridor are one route with two invoices.

Then establish your restoration assumption rather than your availability assumption. Your service agreement describes uptime. It almost certainly does not describe what happens when the network is running on reduced redundancy for forty days, which is the condition that actually precedes an outage.

Build — 12 months

Specify physical route diversity in contracts, by corridor and landing station, and accept that it costs more. This is the one correction in this article that requires money, and it is the only one that changes the correlation structure of your exposure rather than merely documenting it.

And test the degraded case rather than the failed case. Most continuity planning exercises a total outage, which the network makes rare. The realistic scenario is sustained high latency and reduced bandwidth across a quarter, and systems that are fine when the link is dead or healthy often behave badly in between — timeouts, retry storms, partial synchronisation.

Position — 3 years

For any operation in an edge location, treat connectivity as a site-selection variable with the same weight as power. The difference between a two-cable island and a six-cable mainland location is not a service-level difference; it is a difference in how long you can be degraded with no recourse.

And when you review a resilience plan, check whether its diversity is diversity of supplier or diversity of cable. The first is a procurement outcome and the second is a physical fact, and only one of them survives an anchor.

Avoid. Reading a contractual uptime figure as a measure of this exposure. Uptime is measured on your access link. The forty-day condition this article describes happens upstream of anything your agreement covers and shows up as latency, not as downtime.

Why this works. The two corrections — asking which physical systems carry your traffic, and testing the degraded case rather than the dead one — address the two ways this substrate actually fails, and neither is covered by a standard service agreement.

14. Recommendations — Capital

The structural point is that the market finances new capacity well and maintenance capacity badly, and the gap between those two is where both the risk and the policy response sit.

Immediate — this quarter

For holdings whose operations depend on specific corridors — data centres, financial infrastructure, anything latency-sensitive between regions — establish which cable systems and chokepoints they depend on. This is knowable from public route maps and is almost never in a disclosure, because the exposure sits with a supplier's supplier.

Build — 12 months

Watch repair and maintenance capacity as a sector distinct from cable laying. They use different vessels, different economics and different contracts: laying is a project business with a visible order book, repair is a standby business with an unattractive profile and a retiring fleet. Public funding, if it arrives, lands on the second.

And note the asymmetry in how capital currently flows. Of roughly $4 to $5 billion a year in subsea investment, now about double the figure of a decade ago, the overwhelming majority goes to new systems. Growth is funded and resilience is not, which is the condition that precedes public intervention in every substrate this series has examined.

Position — 3 years

The durable observation is that ownership of this layer has concentrated into a small number of technology firms holding around seventy-one percent of capacity, which changes what this infrastructure is. It is no longer a shared carrier utility with many owners; it is increasingly a private asset class whose routing, retirement and pricing decisions are made inside a few companies with their own priorities. That is an observation about how a layer of global infrastructure has changed hands, not a recommendation about any instrument.

Avoid. Treating cable count as a measure of resilience. Five hundred and seventy-four systems sounds abundant and tells you nothing about how many pass through the corridor your traffic uses, or how many ships are available to mend them.

Why this works. Route maps, fleet registers and maintenance agreements are public, slow-moving and read almost exclusively by specialists, which makes this one of the better-documented and least-priced infrastructure exposures available.

15. What Would Change Our Mind

Each forecast carries its own weakening condition. Three developments would undermine the article as a whole.

—  Restoration times fall materially without fleet growth. If permitting reform or regional repair agreements cut the median from forty days to something like fifteen, the central constraint we identify is administrative rather than physical, and considerably easier to fix than we suggest.

—  Satellite capacity becomes a genuine substitute for bulk transit. We treat it as a thin backup. If low-earth-orbit capacity reaches a point where a mid-sized economy can run on it for weeks without serious degradation, the restoration-time gap stops mattering.

—  The core proves less meshed than it appears. Our claim is that core redundancy is real and edge redundancy is not. A cut sequence that produced sustained degradation between major economies would mean the margin at the core is thinner than the fault record has so far shown.

Jurisdiction and sourcing, recorded on the sixth row. This article is the best distributed in Series II so far — Taiwan, the Baltic, the Red Sea, Australia and the trans-Atlantic all appear, and the incidents cited span five jurisdictions and four years. The sourcing weakness is different. Fleet counts are inconsistent between sources, with roughly sixty vessels for all cable work in one account and sixty-two cable-laying ships in another, and the dedicated-repair subset defined differently again. The forty-day median restoration figure comes from a single research citation and we have not found an independent series to check it against. Capacity-share figures for hyperscalers come from industry compilations rather than audited disclosure. Each is labelled accordingly below.

The Series II Table — six rows

What six rows show that five could not. The instrument has now failed three times, each differently. For water, substitution time was political. For refined materials, financial. Here it is simply the wrong measurement — the answer is accurate, instant and irrelevant, because the system does not fail when the alternative engages. It fails if the original is not restored before the alternative is also needed elsewhere.

Taken together these three failures say something about the instrument that five rows of agreement could not. Concentration and criticality have held up across every substrate. Substitution time only works where the alternative is a thing you build and then own. Where it belongs to someone else, or engages by itself, that single number hides more than it reveals — and this row suggests the missing companion is restoration: how long you run on the alternative, and what is left while you do.

The final row is the one that has been implied by all six. Series II closes with the thing that sits underneath them.

16. Bottom Line

Fourteen thousand people on a small Taiwanese archipelago spent about fifty days in 2023 on a microwave link after both their cables were cut six days apart. Not disconnected — reduced to the bandwidth of twenty years earlier, and held there for seven weeks. Their cables have been cut more than twenty times in five years.

In September 2025 three systems were cut near Jeddah. India, Pakistan and the UAE felt it, Azure latency rose, and service was substantially back the next day through rerouting. Physical repair was expected to take weeks or months.

Those two events are the same substrate behaving completely differently, and the difference is the article.

Traffic reroutes in seconds, automatically, with no decision and no construction. By the measure this series has used for five rows, connectivity is the most substitutable thing yet examined. And that answer is useless, because nothing fails when the alternative engages. What matters is the median forty days before the original is mended, during which the network runs on its own redundancy with less of it left — a condition no diagram shows and no alarm reports, because from outside everything looks normal.

Which makes the real chokepoint a fleet of ships. Around sixty vessels worldwide do cable work of any kind and fewer than twenty are dedicated to repair, against 150 to 200 faults a year across 1.42 million kilometres. Two thirds of the maintenance fleet reaches the end of its life by 2040. Investment has doubled in a decade, to $4 or $5 billion a year, and nearly all of it builds new systems rather than keeping old ones alive — because new capacity earns revenue and repair capacity earns nothing between faults.

And the delays are mostly not engineering. Splicing a cable takes days. Waiting for permission to enter territorial waters can take a month, which means the cheapest available fix in this entire substrate is a diplomatic agreement signed before a fault rather than negotiated after one. Nobody is working on it.

Meanwhile the redundancy is less independent than its count suggests. Cables share the same narrow water as ships — one anchor took three systems in the Red Sea in 2024 and about a quarter of Asia-Europe-Middle East traffic with them. They share landing stations. And about seventy-one percent of global capacity now sits with a handful of technology firms, so two carriers selling you independent routes may be reselling the same owner.

Above all, the core and the edge are not the same substrate. The network that shrugs off three simultaneous cuts is the network that left Matsu on a microwave link for seven weeks. Both describe global connectivity and only one describes any particular place.

For anyone with a decision in front of them: ask which physical cables carry your traffic and which corridors they cross, not how many providers you have. Find out what your restoration assumption is, not your uptime figure. And test what your systems do at high latency for a quarter, because that is the failure you will actually get.

 Forecast record

Four forecasts, one per horizon, each with a threshold, a named source that will settle it, and a date fixed before the answer is known.

Correlation. Three relationships, and only one of them is a straightforward family. The first and fourth share a parent cause in incident frequency: a year with more damage makes both more likely. The second and third share the policy response, but inversely — programmes announced by 2029 make a fleet above thirty by 2031 slightly more likely, though shipbuilding lead times mean the effect is weak within the window. And the second bears inversely on the fourth: more repair capacity reduces the chance of a sustained national outage. So the set is not four independent observations, and it is not two clean families either. It is one positively correlated pair, one weakly inverse pair, and a cross-link between them, recorded as such rather than tidied.

Directional statements elsewhere in this article carry no threshold and are excluded from the record.

Sources

THRIVE IN CHAOS

Signal Over Noise  ·  thriveinchaos.ai

AI intelligence system with human editorial oversight. Analytical material, not investment, legal or tax advice.