SEVEN THINGS UNDER THE FLOOR

Article 8 of 8  ·  Series II of III  ·  Published 2 December 2026  ·  Analysis → Forecast → Recommendations

Seven articles, seven physical layers, one table. This piece closes Series II by putting the rows together — what they share, how they stack, and the four things that only became visible once all seven were measured the same way.

 What the completed table says that no single row could

1. One Building, Seven Dependencies

Somewhere in Europe or North America, a company is building a data centre. Nothing exotic. A large shed, some transformers, racks of equipment, a car park.

Follow what it needs, and you walk down every row of this series.

It needs a grid connection, and the queue for one runs four to seven years in most places, with transformer lead times above three years and a shortage of the engineers who do the work. That was the first article, and it is the only substrate in the series with no alternative path at all. You cannot buy your way to the front.

It needs chips, and the advanced packaging step that turns silicon into a working accelerator is effectively one supplier at volume. Not the chips themselves — the step after. Fifty-two to seventy-eight weeks for capacity, years for a second source.

The equipment arrives by sea, through corridors that are legally open and running well below normal traffic. Ten to fourteen extra days on an Asia–Europe voyage, a permanent freight premium, and nothing ever actually runs out.

It uses water, and here the building misleads you. The cooling towers are the small number. Around eleven twelfths of its water footprint is consumed generating its electricity, somewhere else, in a basin nobody involved in the project has looked at.

Its transformers, its fans, its motors and its generators all contain magnets, and the separation step that produces the heavy rare earths in those magnets sits at roughly ninety-one percent in one country. The same element trades at fifteen times the price depending which side of that capacity you are standing on.

Its traffic runs on submarine cable. The operator has two carriers and a diversity clause, and three of the four cable systems behind those contracts cross the same stretch of the Red Sea.

And every piece of equipment in the building that timestamps, synchronises or coordinates takes its time from a satellite. Nobody bought that. It came inside the hardware, it costs nothing, and it appears on no bill of materials.

Each article measured one layer with the same three questions: how concentrated is it, what stops if it fails, and how long until an alternative works. Asking the same questions seven times turned out to matter more than any individual answer, because the pattern in the failures of the instrument is the real finding of the series.

2. The Seven, in Plain Language

Before the patterns, the rows — stripped to what a reader should carry away.

Electricity

The hardest row and the least discussed. One grid per region, no alternative path, and connection queues that have roughly doubled in fifteen years. A transformer is a three-year order. The engineers who commission substations take a decade to train, which means the constraint in 2036 is already determined by who started an apprenticeship this year.

Nothing here is a shortage of electricity. It is a shortage of the ability to connect to it, which is a different problem with the same consequences and far less visibility. A factory that cannot connect does not appear in any statistic about energy supply. It appears, years later, as a factory somewhere else.

Computing

The chokepoint is not chip fabrication, which is where almost all the attention goes. It is the packaging step that assembles the finished accelerator — effectively one supplier at volume — and the memory that goes with it, where a single firm holds around sixty-two percent.

The useful part for a non-specialist is the shape rather than the names. A bottleneck moves. Five years ago it was fabrication capacity; now it is the step after, and in five more it will be somewhere else again. It binds absolutely for frontier systems and leaves ordinary electronics alone, which is why a washing machine is unaffected while a data centre waits.

Shipping

The good news row, and it deserves saying plainly after six articles about constraints. Three years of disruption to the busiest corridors on earth produced no shortages at all, because ships reroute in days.

What it produced instead was a permanent operating premium — ten to fourteen extra days on an Asia–Europe voyage, higher insurance, and a fixed fleet eaten by distance so that rates rise on routes nowhere near the trouble. Spread thinly enough that nobody adds it up. The article's practical finding was a vocabulary correction: open and closed is the wrong question, and what share of normal traffic is passing is the right one.

Water

Taiwan paid farmers not to plant across roughly 74,000 hectares in 2021 and trucked water to its chip plants. The chips shipped, and the episode is filed almost everywhere as a supply chain that held.

Nobody found extra water that year. A government decided who would go without. Water cannot be moved at scale, so substitution means taking it from an existing user, which is a political decision with a person on the losing side rather than an engineering timeline. The other half of the article was a measurement error: for a data centre, roughly eleven twelfths of the water footprint is consumed generating its electricity, so the number everyone publishes and argues about is the small one.

Refined materials

Rare earths are not rare. Gallium falls out of aluminium refining and graphite is mined in a dozen countries. The concentration is in the separation plants, not in the ground, and it was acquired over thirty years of decisions that each looked like discipline at the time — exiting a low-margin, chemically tedious step with difficult waste, and buying the output cheaper from somebody willing to do it.

A mine in California opened in 1952, closed by 2002, reopened on a price spike in 2008, went bankrupt in 2015 and sold for twenty million dollars. The rock never moved. The price did, and the price is set by the party any new entrant exists to replace, which is why the instrument that finally worked was a guaranteed price rather than a construction grant.

Connectivity

Fourteen thousand people on a Taiwanese island spent about fifty days on a microwave link after both their cables were cut six days apart. Not disconnected — dropped to the bandwidth of two decades earlier and held there for seven weeks, with remote classes and card payments failing.

Meanwhile three cables severed in the Red Sea produced a slower morning and not much else. Same substrate, opposite outcomes. The difference is that traffic reroutes in seconds while physical repair takes a median of forty days, and the core has enough paths to absorb that while the edge has a spare rather than redundancy. Around sixty specialised vessels do cable work worldwide, fewer than twenty of them dedicated to repair, against 150 to 200 faults a year.

Timing

A satellite was retired on schedule in January 2016. A ground software fault put thirteen microseconds of error into the time signal from several others, and fixed-line operators in Britain and Sweden failed, along with mobile carriers and broadcast systems. One company found two thirds of its timing equipment affected. Nobody had attacked anything.

Grids, databases, mobile networks and exchanges all run on precise time, and essentially none of them buy it. It arrives free, embedded in equipment somebody else specified years ago, on no bill of materials and in no supplier register. The alternatives are mature and have gone unsold for decades, because the incumbent costs nothing and the benefit belongs to everybody.

3. The Completed Table

Seven rows, measured the same way. The first three columns are the instrument’s original questions — concentration, criticality, substitution time — and the fourth is the column the series earned along the way: who actually decides whether you get an alternative.

Read down the fourth and fifth columns together and the series stops being about physics.

4. Attention Runs Opposite to Severity

Rank the seven by how hard they are to substitute, then rank them by how much public attention they get. The two lists run close to backwards.

Hardest to substitute: the grid, where there is no alternative path at any price; refined materials, where the alternative can be built and cannot be kept alive; water, where substitution means taking it from somebody. Then advanced packaging, then timing, then cable restoration, and finally shipping, which reroutes in days.

Most covered: shipping, by a wide margin. Then refined materials, which acquired a politics. Then computing. Water intermittently, during droughts. Cable cuts in brief spikes when sabotage is suspected. The grid almost never, and timing essentially not at all outside trade press.

The substrate that generates the most alarm has by far the best substitutability. The two with the worst get almost no coverage at all.

The mechanism is not mysterious. A blocked strait is footage. A tanker adrift, a canal empty, a map with a red line across it — all of it films. A grid connection queue is a spreadsheet held by a utility, and a timing dependency is a line that does not exist in any document. Attention tracks what can be photographed, and three of these seven substrates cannot be.

Which has a practical consequence for anyone deciding where to put effort. If your sense of what is dangerous comes from the volume of coverage, you are inversely calibrated on this table. The honest version of the instruction is uncomfortable: the less you have heard about a dependency, the more likely it is that nobody has looked.

5. Adaptation Destroys the Case for Repair

The same mechanism turned up in four of the seven rows, each time in a different costume, and it took until the sixth article to recognise it.

Shipping adapted. Ships took the long way round within weeks, nothing ran out, and the permanent premium — ten to fourteen days, several hundred dollars a container, war-risk insurance, a fleet eaten by distance — was never totalled because nothing failed. No board paper has ever contained a line called chokepoint.

Cables reroute. Three systems severed in the Red Sea and most people noticed a slower morning. The forty days of repair that follow, during which the network runs on less redundancy than its design assumes, are invisible from outside. The repair fleet is small and ageing precisely because the system has been so good at concealing the need for it.

Timing has holdover. Equipment carries its own oscillator and rides out a lost signal for hours or days, so an outage produces a quiet countdown rather than a failure. Aviation then adapted to deliberate interference with real competence — procedures, training, bulletins, a standing dashboard — and converted an emergency into a condition.

Water was reallocated. Taiwan protected its fabs in 2021 and the episode is remembered, when it is remembered at all, as a supply chain that held.

Anything that absorbs failure quietly will be under-resourced, because the evidence for investing in it is destroyed by the thing working.

This is the most portable finding in the series and the one least about infrastructure. Every organisation has a person, a process or a system that quietly absorbs problems, and the better it does that job the harder it is to fund. The business case arrives the week after it stops coping.

It also explains the shape of the whole table. The three substrates with the best substitution answers are the three where the adaptation is most visible and the underlying cost is least counted. Being good at absorbing failure is not the same as being resilient, and the first tends to hide the absence of the second.

6. The Alternatives Converge Upstream

Five of the seven rows produced the same structural surprise: diversity that is real at the point of purchase and absent two tiers above it.

—  Cables. Two carriers, a diversity clause, four named systems — and three of them crossing the same stretch of water. One dragging anchor took three systems in the Red Sea in February 2024 and roughly a quarter of Asia–Europe–Middle East traffic with them.

—  Magnets. Two motor suppliers on two continents, buying from three magnet manufacturers, all sourcing separated heavy rare earths from the same origin. Dual-sourced at the invoice, single-sourced at the chemistry.

—  Satellites. Four global constellations sound like four alternatives. They are four instances of one physical principle, a faint unauthenticated signal from orbit, and interference takes them together.

—  Water. Two plants on different rivers, both buying electricity from a balancing area with a high thermoelectric share, so the plant with the smaller water bill has the larger footprint once generation is counted.

—  Capacity ownership. Around seventy-one percent of global subsea capacity sits with a handful of technology firms, so two carriers selling independent routes may be reselling the same owner's decisions.

The common error is counting at the wrong tier. Suppliers are countable, so organisations count suppliers; corridors, separation trains, grid regions and physical principles are not on any invoice, so they go uncounted. Every one of these cases took exactly one question asked one tier further than usual.

7. Where the Market Stopped Supplying Resilience

The last three articles found the same economic structure, and finding it three times running makes it a pattern rather than a coincidence.

Refined materials: a separation plant takes three to seven years to build and its revenue is controlled by an incumbent with ninety percent of installed capacity. Everyone financing one knows this, which is enough to stop it being financed. The instrument that finally addressed it was a ten-year price floor with full offtake, not a construction grant — because capital was never the binding constraint. Survivable revenue was.

Cable repair: fewer than twenty dedicated repair ships worldwide, against 150 to 200 faults a year. Subsea investment has roughly doubled in a decade, to four or five billion dollars annually, and nearly all of it builds new systems. New capacity earns revenue. Repair capacity earns nothing between faults.

Timing: terrestrial timing, fibre distribution and atomic clocks are all mature and have spent decades failing to sell, because the incumbent is free, the benefit accrues to a whole economy, and the risk is invisible in ordinary operation. Britain committed £155 million in November 2025 against a stated exposure of around £1.4 billion a day — a programme costing less than three hours of the outage it exists to prevent.

The shared shape: an asset that earns nothing in normal operation and everything in a failure that has not happened yet.

Markets do not fund those, and three of our seven rows have now become public balance-sheet problems as a direct result. That is not an argument for or against public spending. It is a description of which category of thing ends up there, and it predicts where the next programme will appear: wherever resilience is a standby asset with no revenue between events.

It also sets a test worth remembering. When a resilience problem persists for decades despite mature technology and obvious need, the shortage is almost never engineering. Ask who would pay, and what they get in a year when nothing goes wrong.

8. The Instrument, Repaired Three Times

The series began with three questions and ended with five. How that happened is the methodological result, and it transfers to any dependency, not just the seven here.

Concentration and criticality held up across all seven rows without modification. They are honest questions about the world and they return honest answers.

Substitution time broke three times.

—  At water, it turned out to be political. There is no engineering answer to how long it takes to move water between users; there is only how long the decision takes and who is asked to give something up.

—  At refined materials, it turned out to be financial. Building the plant is three to seven years. Keeping it alive is indefinite, and the party setting the price is the one being replaced.

—  At connectivity, it turned out to be the wrong measurement. Traffic reroutes in seconds, which looks like the best answer in the series and tells you nothing, because nothing fails when the alternative engages. Restoration time — the median forty days — was added as a fourth question.

And then timing broke the repaired instrument too. All four questions return good answers there — alternatives are mature, restoration takes hours — and the exposure is still severe, because every one of those questions describes the world and none describes the organisation standing in front of it. Hence the fifth: who owns this dependency, in whose budget and whose job description.

The five questions, as they now stand

The first two are properties of the world. The last three are properties of a relationship — between you and the thing, or between you and whoever controls it. That split is the series' main methodological claim, and it is why physical data alone consistently understates exposure.

9. It Is a Stack, Not a List

The table presents seven rows side by side, which is how they were investigated and not how they sit. Put them in dependency order and they form roughly three tiers.

At the bottom: timing and electricity. Timing is underneath all six others — grids need it to compare phase, databases to order transactions, networks to align transmissions, exchanges to timestamp trades, ports and ships to navigate. Electricity is underneath computing, refining, water treatment, cable landing stations and every piece of equipment that keeps time.

In the middle: connectivity and refined materials. The control layer of the grid, of water systems and of logistics runs over networks. Transformers, turbines, motors and generators run on magnets.

On top: computing, water use and shipping — the layers where failures are visible, where organisations have procurement, and where most attention goes.

The two substrates at the bottom of the stack are the two with the worst answers in the table, and between them they receive the least public attention of the seven.

That is not a coincidence and it is not a conspiracy. It is the attention mechanism from section four operating on a structure: the deeper a layer sits, the more it is mediated by other layers, the less directly anyone experiences it, and the fewer people have a contract with it.

It also changes what a cascade looks like. A shock at the top stays mostly at the top — a chokepoint disruption raises costs and does not take down a grid. A shock at the bottom appears everywhere at once, in forms that each look like a local fault. January 2016 is the worked example: one timing error, and the organisations affected recorded telecommunications failures, because that is what it looked like from where they were standing.

One shock, traced

Take a sustained loss of reliable time across a region — not the twelve hours of 2016, but a week. Nothing dramatic happens on day one, because holdover covers it. What follows is a sequence that arrives in the wrong order for diagnosis.

Mobile cells begin to drift out of alignment and interfere with each other, which the operator sees as a capacity problem in specific sites. Distributed databases accumulate ordering errors that appear as data quality incidents in applications nobody connects to each other. Grid protection schemes lose the ability to locate faults precisely, which shows up as slower restoration after unrelated outages. Trading venues breach timestamp tolerances, which arrives as a compliance matter. Port and logistics automation degrades, which looks like a vendor fault.

Five organisations, five incident reports, five different root causes recorded, and not one of them says time. That is what a bottom-tier shock looks like from inside any single organisation, and it is why the 2016 event is remembered by most of the companies it hit as a telecommunications failure.

The same logic runs one tier up. A prolonged grid constraint does not present as a power shortage; it presents as projects not being built, which is invisible because it is an absence. Depth in the stack converts a single cause into a scatter of local effects, and the deeper the cause, the wider the scatter and the less likely anybody names it.

Which suggests a diagnostic habit worth having. When several unrelated systems degrade at roughly the same time with no common cause, the common cause is probably a tier below all of them.

10. Where This Lands on a Person

Seven articles about grids, refineries, cables and satellites can read as though the only people involved are utility executives and ministers. Most of the consequences land somewhere else, and in three recognisable places.

The person at the edge

Matsu is the clearest case and far from the only one. Fourteen thousand people, two cables, fifty days at a fraction of normal bandwidth. Not an outage — a downgrade, held long enough to break remote schooling and card payments, which is harder to improvise around than a blackout because nothing is obviously broken.

Islands, remote regions and countries dependent on a single transit neighbour all sit here, and so does anyone whose livelihood assumes connectivity that the place they live does not actually have twice over. The global figures say the system is robust. They are describing the core. For the edge the honest word is not redundancy but spare, and knowing which of the two you have is a ten-minute question with a decade of consequences.

The engineer who notices

In most of these rows, the person who could see the problem first sits three levels below anyone who could fund fixing it. The network engineer who knows every site takes time from one receiver. The procurement analyst who notices two suppliers naming the same plant. The facilities manager who has actually read the holdover specification.

They tend not to raise it, and the reason is structural rather than personal. A warning about something that has never failed, with no incident behind it and no budget line to attach it to, competes against projects with revenue — and the person raising it carries the cost of being wrong while the organisation carries the benefit of being right. Section five explains why they lose that argument: the system works, so the evidence is missing by construction.

Which makes the most useful thing a senior person can do here unusually cheap. Ask the question downward rather than waiting for it to come up. Somebody in the building already knows, and they are not going to volunteer it.

The small business with no leverage

Most recommendations in this series assume an organisation large enough to specify, audit and negotiate. A twelve-person firm cannot require physical route diversity from a carrier or interrogate a magnet manufacturer about separation plants.

What it can do is smaller and still worth doing: know which of its inputs cross one corridor, treat the freight premium on imported goods as permanent rather than temporary, and avoid signing long contracts priced on an index it cannot actually buy at. That last one appeared in three separate rows, and it is the mistake that quietly transfers margin for years without anyone noticing the moment it started.

11. What Is Getting Better

A series about constraints accumulates a tone, and the tone would misrepresent the evidence if it went unchallenged. Several things in these seven rows are genuinely improving, and two of them quickly.

Adaptation works, repeatedly and at scale. Three years of attacks on the busiest shipping corridors produced no shortages. A drought in the place that makes the world's most advanced chips did not interrupt chip supply. Three cables cut at once in the Red Sea produced elevated latency. Aviation absorbed a step change in deliberate interference without losing an aircraft. These are not lucky outcomes but the result of decades of engineering and procedure, and the right response is respect rather than suspicion.

Diagnosis has improved sharply. The reason this series could be written at all is that the data now exists: published connection queues, weekly maritime transit counts, aviation interference dashboards, regional price assessments, incident logs. Most of it was not available in usable form five years ago and almost all of it is free. The gap between what is knowable and what is known has narrowed, even if the gap between what is known and what is acted on has not.

And the policy instruments are better aimed. The price floor for rare earth magnets identifies the binding constraint correctly as survivable revenue rather than capital, which twenty years of grants did not. Britain's timing programme put the large majority of its money into building alternatives rather than monitoring the problem. Both may fail on execution. Neither is aimed at the wrong thing, and that is a change.

Efficiency is real too: semiconductor water recycling above seventy percent, magnet chemistries that cut heavy rare earth content, near-zero-water cooling, authenticated satellite signals shipping in volume. The honest qualification is that in several rows volume is outrunning efficiency — but the engineering is not the part that is failing.

What is not getting better. The two layers at the bottom of the stack. Grid connection queues are lengthening, and the repair and timing capacity the whole structure rests on is ageing and unowned. Everything improving sits above the two things that are not.

12. Five Years Out — to 2031

Two things we expect to be able to check.

More of this becomes a public balance-sheet problem

By the end of 2031, G7 governments have committed public funds or imposed a binding mandate for resilience capacity in at least three of these seven substrates. Probability 0.65. Confidence: Medium.

One is already done: the American price floor for rare earth magnets, a ten-year commitment with offtake and equity. A second is in progress: Britain's timing programme. Grid investment is being mandated through regulated asset bases in several markets, and cable repair capacity is the obvious next candidate because the arithmetic is simple enough to survive a briefing. The forecast is close to arithmetic rather than speculation.

The second-order effect matters more than the count. Each programme creates a procurement category where none existed, and a buyer where the market had none. That changes what gets manufactured, not just what gets funded.

What weakens it: governments prefer strategies, task forces and monitoring, all cheaper and faster to announce than a committed line. Three substrates with real money is a higher bar than three with published plans, and that distinction is where this forecast will be decided.

Nobody assesses the stack as a stack

By the end of 2031, no G7 country has published a single infrastructure dependency assessment covering grid, computing, water, refined materials, connectivity and timing together. Probability 0.60. Confidence: Medium.

Each of these substrates has a competent community, a regulator and in several cases a national strategy. They do not share a document. Energy regulators do not assess timing; telecoms regulators do not assess water; critical minerals strategies do not cover cable repair. The structure of government mirrors the structure of the table — seven rows side by side rather than three tiers — and departmental boundaries are the reason.

A reader might reasonably ask why that matters, given each is covered somewhere. It matters because the interactions are where the surprises live, and interactions fall between departments by construction. Nobody owns the question of what happens when two tiers move at once.

What weakens it: a single severe cross-cutting incident would produce exactly this document within a year, and national risk registers are already edging toward it. The forecast is a claim about institutional inertia, which is usually right and occasionally very wrong.

13. Ten Years Out — to 2036

The hardest row gets harder

Grid connection waiting times in at least one major European market are longer in 2036 than they were in 2026. Probability 0.65. Confidence: Medium.

This tests the series' most confident claim. Demand is rising from electrification and computing, transformer lead times exceed three years, and the engineering workforce takes a decade to build. Connection queues have roughly doubled over the last fifteen years, and the investment now flowing in takes most of a decade to appear as capacity. A straight-line reading says the queue grows before it shrinks.

The second-order effect is the one to watch. A long queue is not neutral: it allocates. Whoever is at the front of a seven-year line gets to build, and whoever is not, does not, which makes grid access a quiet determinant of which industries and which regions grow at all. That is an industrial policy nobody voted for, administered by a queue.

What weakens it: reform is underway in several markets, and queue management changes — clearing speculative applications, reserving capacity, charging for position — can shorten published waiting times substantially without a metre of new cable. We would resolve this on published operator figures, which that kind of reform genuinely improves.

Two of the seven produce something visible

Between 2027 and 2036, at least two of these seven substrates produce a publicly attributed disruption lasting more than one week in a G7 country. Probability 0.55. Confidence: Medium.

Only slightly above even, and deliberately so. The striking thing about seven articles of research is how little has actually broken. Three years of shipping disruption produced no shortages. Taiwan's drought did not interrupt chip supply. The 2016 timing event lasted twelve hours. These systems absorb a great deal, and the honest base rate for a week-long attributed outage in a wealthy country is low.

The word doing the work is attributed. Several of these failure modes are likely to occur and be recorded as something else — a network fault, a protection misoperation, a vendor defect — exactly as much of January 2016 was. This forecast could fail while the underlying claim holds completely.

What weakens it, in the other direction: if the state programmes in the five-year forecast arrive on time, they reduce precisely this probability. The two forecasts pull against each other, and that is recorded in the ledger below rather than smoothed over.

14. What We Are Not Carrying Forward

Synthesis has a predictable failure mode: the confident claims survive and the qualifications quietly disappear, because the qualifications are what makes a summary long. The seven articles contained caveats that would be easy to lose here, so they are listed rather than dropped.

—  The forty-day median restoration time for a submarine cable, which carried much of the connectivity article, rests on a single research citation. We found no independent series to check it against.

—  The fifteen-fold ex-China premium for dysprosium comes from thin assessments of an illiquid market rather than cleared prices. The direction is not in doubt; the multiple should be read as an order of magnitude.

—  The 2016 timing event figures — the alarms, the two-thirds estate, the timings — come from one industry case study. It is thorough and it is still one source carrying an entire opening.

—  The £1.4 billion a day figure for a British satellite navigation outage is a model derived from a study of a five-day disruption that has never occurred. It is a fact about policy, not an observed cost.

—  Aviation interference counts are counts of reports, not of events. Part of the rise between 2021 and 2024 is better reporting, and we cannot say how much.

—  Repair vessel counts disagree between sources depending on what is being counted, which is why the forecast in that article fixed a counting definition in advance.

And one correction of our own worth repeating. The maritime article found that the substrate generating the most alarm had by far the best substitutability — a finding that cut against the direction of the series, and we published it as the headline of that row rather than burying it. Section four of this piece exists because of it.

15. How to Tell If We Are Wrong

Four developments would undermine this synthesis rather than any individual row.

—  Attention and severity converge. If grid queues and timing dependency start attracting coverage proportionate to their difficulty, the mechanism in section four is weaker than we claim and the market for attention corrects itself better than we think.

—  A market supplies resilience without a mandate. One unguaranteed separation plant reaching scale, or a commercially funded expansion of the repair fleet, would falsify the public-good argument in section seven directly.

—  Organisations inventory their own dependencies unprompted. The ownership question assumes nobody asks. If timing inventories and corridor audits become ordinary practice without regulation, the fifth question matters less than the series concluded.

—  The stack proves shallower than described. Our claim is that bottom-tier failures propagate everywhere at once in disguised forms. A decade with no such cascade, in a period of rising interference and ageing repair capacity, would suggest the tiers are more isolated than the dependency map implies.

Jurisdiction and sourcing across the whole series. Seven articles, and the coverage is uneven in a way worth naming at the close. Europe and the United States appear in every row. Taiwan, South Korea, Japan and China appear substantially but almost always as objects of analysis rather than through their own institutions and data. India, the Gulf, Africa and Latin America are essentially absent, which means the series describes the dependencies of wealthy importing economies rather than of the world. Three rows — materials, timing and computing — are narrow enough on this count that a reader elsewhere should treat the recommendations as needing translation rather than application. On sourcing, the strongest rows are those resting on published regulatory and operator data (grid, aviation interference, trade flows) and the weakest are those resting on industry compilations and single case studies (repair fleet, restoration times, ex-China prices). Each figure carried its class in its own article and the weak ones are listed in section fourteen.

16. What to Actually Do

Seven articles produced a great deal of advice. Compressed, each audience has one thing that matters more than the rest, and a short list behind it.

Individuals

The one thing: recalibrate what you treat as dangerous. If your sense of risk comes from coverage volume, this table says you are inversely calibrated — worrying about straits and not about queues. That is free to fix and it changes what you pay attention to for years.

Behind it, three small things. Learn the state of your own basin and how many cables serve where you live, both of which are published and take ten minutes. If your work touches infrastructure, engineering, networks or industrial control, ask whether anything you are responsible for takes its time from a satellite, because you are probably the first person to ask. And treat satellite navigation in your car and phone as a convenience that is already intermittently unreliable across much of Europe and the Middle East, rather than as a utility.

Avoid. Reading any of this as a reason to expect shortages. Across seven substrates and several years of disruption, the recurring outcome is cost, delay and degradation — not absence.

Business

The one thing: ask what your suppliers share. Not how many you have — what they share. Corridors, separation plants, grid regions, time sources. In five of seven rows that single question, asked one tier further up than usual, found convergence nobody had written down, and it costs one email per supplier.

Behind it, four. Calculate your indirect water footprint, which is usually several times the reported one and takes an afternoon. Ask your connectivity providers which named cable systems carry your traffic and which chokepoints they cross. Inventory your timing sources and each system's holdover, then sort that list by holdover, because sorted that way it is a failure sequence. And put a named person against each of these, in a job description, because an inventory with no owner is done once and never updated.

One more, further out. Move these questions into site selection and procurement rather than compliance review. Over forty percent of semiconductor fabs announced since 2021 sit in basins projected to be highly stressed by 2030, and that figure exists because water was reviewed after the site was chosen. The same applies to grid region, corridor exposure and time source: every one of them is cheap to specify at purchase and expensive to change afterwards.

Avoid. Treating a contractual diversity clause as physical diversity, or a second supplier as a second source. Both were wrong in multiple rows, and both are the kind of wrong that holds up perfectly until the day it matters.

Capital

The one thing: watch where resilience is becoming a public purchase. Three of the last three substrates reached the same point — an asset that earns nothing in normal operation and everything in a failure that has not happened — and in each case the state became the only available buyer. That pattern says where the next procurement category appears before any company announces anything.

Behind it, three. In any exposed holding, read the guarantee rather than the announcement: duration, legal form, whether offtake is included, what ends it early. Distinguish growth capacity from standby capacity, because the first is funded well and the second is not, and the gap between them is where both the risk and the policy response sit. And treat announced capacity as unbuilt until it is operating and contracted, a discipline this series would have applied to several published figures if it had been available to apply.

Avoid. Pricing any of these on incident history. The absence of incidents reflects holdover, rerouting and reallocation working well, not a low underlying exposure — which is section five restated as an investment error.

17. The Short Version

A company builds a data centre. It depends on seven physical things, and the people building it can name two.

Seven articles asked the same three questions of each layer, and the questions turned out to be the discovery. Concentration and criticality held up everywhere. Substitution time broke three times — once political, once financial, once simply the wrong measurement — and after the repairs it broke again, because all four questions can pass while the exposure stays severe. The fifth question is the one about you: whose budget, whose contract, whose job. Four questions describe the world and one describes whether anybody is responsible for it, and that one decides whether the others ever get asked.

Four things only showed up once all seven rows were on the page.

Attention runs backwards. The substrate that produces the most alarm — shipping — reroutes in days. The two with no good answers at all, the grid and timing, get almost no coverage, because a queue and a clock cannot be photographed. If you calibrate on coverage you are calibrated upside down.

Adaptation hides the bill. Ships rerouted, cables failed over, holdover held, water was reallocated — and in each case the system worked so well that the cost stopped being counted and the repair stopped being funded. The repair fleet is small because the network is good at hiding that it needs one.

Alternatives converge above where anyone looks. Two carriers, four cables, three corridors. Two motor suppliers, one separation plant. Four constellations, one physical principle. The question that finds all of it is not how many suppliers you have but what they share.

And where resilience earns nothing between failures, the market stops supplying it. Three rows running ended with the state as the only buyer, which is not an argument about politics but a description of which category of problem ends up there.

Underneath it all, the table is a stack rather than a list. Timing and electricity sit at the bottom, under everything else, with the worst answers and the least attention. A shock at the top raises prices. A shock at the bottom appears everywhere at once wearing seven different local disguises, which is exactly what thirteen microseconds did in January 2016.

None of this predicts collapse. The striking result of seven articles is how much these systems absorb — years of disruption to the busiest shipping corridors on earth, a drought in the place that makes the world's advanced chips, three cables cut at once, and no shortages anywhere. The exposure is real and it does not look like scarcity. It looks like cost that nobody totals, margin that nobody measures, and dependencies that nobody owns.

Which is why the useful question at the end of seven articles is not what might fail. It is simpler, and most organisations cannot answer it: what is absorbing this on our behalf, and how much of it is left?

Forecast record

Four forecasts, two horizons, each with a threshold, a named source that will settle it, and a date fixed before the answer is known. These are cross-cutting; the twenty-eight substrate forecasts from the seven articles stand on their own dates.

Correlation. Not four independent observations. The two five-year forecasts share a parent cause in state attention to infrastructure dependency and move together, though the second is about institutional structure rather than money and could resolve either way regardless of spending. The first bears inversely on the fourth: resilience funded is disruption avoided, so a yes on one makes a no on the other more likely. The third is largely independent of the other three, since grid queues are driven by demand growth and build times rather than by resilience policy. So the honest reading is one positively correlated pair, one inverse link across horizons, and one standalone — and the weakest member of the set is the fourth, which sits downstream of the first.

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

Sources

This article synthesises the seven preceding pieces and introduces no new primary research. Every figure quoted here carries its epistemic class and its full source in the article where it first appeared, and the weakest of them are listed in section fourteen.

What comes next

Series II measured where the transformation materialises — seven physical layers, what each one can and cannot substitute, and who decides. Series III turns to the question those seven rows keep pointing at and never answer: what replaces the arrangement that is coming apart, and how a person, a business or a territory acts inside it rather than waiting for it to settle.

THRIVE IN CHAOS

Signal Over Noise  ·  thriveinchaos.ai

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