THE NARROWEST POINT

Where computing actually gets stuck, and why it is not where anyone was looking

Article 2 of 8  ·  Series II of III  ·  Published 21 October 2026  ·  Analysis → Forecast → Recommendations

Series II, second row. Each article measures one physical layer with the same three questions — concentration, criticality, substitution time — and adds one row to a single comparison table. Article 1 measured electricity. This one measures computing, and finds something the first row did not: a layer whose criticality depends almost entirely on what you are using it for.

1. The Golden Screw

Picture a procurement manager at a company that assembles servers for large computing customers. She has an order for a rack of AI hardware worth around four hundred thousand dollars.

She has the chassis. She has the power supplies, the voltage regulators, the cooling assemblies, the network switches, the cables. She has the rack itself and the building to put it in.

She does not have the accelerator — the single processing unit that the whole rack exists to house. And without it, everything else she holds is inventory that cannot be shipped, cannot be invoiced, and is quietly losing value.

The industry has a name for this. It calls the missing part the golden screw: one component whose absence makes every other component worthless.

What happens next is instructive. Because she cannot ship, she holds on to everything else she has, and she orders more of the peripheral parts than she needs, because when the accelerator allocation finally arrives she cannot afford to be short of anything else. Every other assembler is doing the same.

So the shortage of one narrow component produces hoarding of dozens of abundant ones, and those abundant ones start to look scarce too.

Now follow the missing accelerator back up the supply chain to find out why it is missing, because the answer is not what most people assume.

It is not the transistors. Chips made on established manufacturing processes are widely available, with lead times of roughly four to seventeen weeks. It is not, primarily, the wafer fabrication either.

It is a step most people outside the industry have never heard of: advanced packaging — the process that places a processing chip beside stacks of high-bandwidth memory on a silicon bridge fine enough to move terabytes of data per second between them. The leading version of that process was reported fully booked in late 2025, with lead times of roughly fifty-two to seventy-eight weeks. The high-bandwidth memory that goes into it was described as effectively sold out for 2026.

That is where a four-hundred-thousand-dollar rack gets stuck. Not at the most famous step in making a chip. At the step after it.

2. Computing Is a Stack, and It Breaks at the Thinnest Layer

It is natural to think of computing as one thing that is either available or not. It is not one thing. It is a stack of separate stages, each with its own suppliers, its own lead times and its own concentration.

Each layer has to be present for the finished product to exist, and they are not interchangeable. An abundance of fabrication capacity does not compensate for a shortage of packaging. So the whole stack moves at the speed of whichever layer is thinnest at the moment.

Which is the same rule this series has applied everywhere: the stages add, and the slowest one sets the pace.

Running the instrument

And here the instrument produces a result it has not produced before in this series.

The same substrate is critical or trivial depending entirely on what you are building.

That matters, because the popular phrase — chip shortage — describes neither case accurately. For most of the economy there is no chip shortage. For the narrow frontier where the most investment is concentrated, there is a shortage so severe that access is no longer bought, it is allocated.

3. Subtheme One — The Constraint Keeps Moving Down

Article 1 of Series I set out a rule that has turned out to be the most reliable in this entire series: solving one dependency usually creates another, one layer deeper. Computing is where that rule can be watched operating in real time, over just a few years.

Where everyone was looking

For most of the last decade, public attention on semiconductors focused on two things: the lithography machines that print circuits, which come from a very concentrated supplier base, and the fabrication plants that use them, which are concentrated in a small number of locations.

Both of those remain genuine concentrations and both remain important. Governments spent heavily and legislated extensively to address them — new fabrication plants in new places, export restrictions on the most advanced equipment.

Where the constraint actually went

While attention was on fabrication, the binding constraint moved one step down, to packaging. Almost nobody outside the industry was watching packaging, because for decades it had been a routine, low-margin stage.

Then the most demanding processors started to need memory placed right beside them, on a silicon bridge, at densities ordinary packaging could not reach. Demand for that specific process grew faster than anyone had provisioned for, as recently as 2024.

The response is enormous. Advanced packaging capacity is reported growing at around 80 percent a year. The leading manufacturer is aiming to raise its monthly capacity close to fourfold by late 2026.

And the constraint is already moving again. Building a packaging line needs specialist bonding and placement equipment, which itself has twelve to eighteen month lead times. It needs a particular substrate film, reported to be in short supply. The fix for packaging runs straight into a new bottleneck one layer further down.

Why this keeps happening

It is not bad planning. It follows from the structure.

When one layer is the bottleneck, it attracts all the investment and all the attention. The layers around it are not the bottleneck, so they are not expanded. The moment the bottleneck is relieved, the next thinnest layer becomes the constraint, and it was under-invested precisely because it was not the constraint yesterday.

A stack under rapidly growing demand therefore does not have a shortage that gets solved. It has a shortage that moves, and each move takes a year or two to recognise because the new constraint is by definition somewhere nobody was looking.

What this does to policy

Public policy has a particular difficulty with a moving constraint. Legislation takes years to draft and pass, and it is aimed at the bottleneck visible when drafting began.

A great deal of recent industrial policy targeted fabrication plants and lithography, which were the visible concentration when the policy was conceived. By the time the plants are built, the binding constraint has moved to packaging and memory — layers the legislation largely did not address, because they were not the problem when it was written.

That is not a failure of the people who wrote it. It is the correction loop from Series I operating on industrial policy: two to three years to notice, two to three years to legislate, several more to build, against a constraint that relocates every year or two. The policy arrives to fix a problem that has already moved on.

The honest forecast for computing is not that the bottleneck will be solved. It is that it will be somewhere else next year, and that a lot of money will arrive there slightly too late.

4. Subtheme Two — When a Market Becomes an Allocation

The second thing computing shows is what happens to a market when the thinnest layer is sold out for years ahead.

Access stops being priced

In April 2026 it was reported that the largest accelerator company had reserved the majority of the leading packaging capacity through at least 2027. Estimates put its share of the expanded capacity at around sixty percent.

That is not a scandal and it is not unusual. It is what a rational buyer does when the thinnest layer is constrained: secure it years ahead, at whatever the terms are, because being without it is fatal and paying for it early is merely expensive.

But its consequence for everyone else is specific. When capacity is committed years in advance, a newcomer cannot buy its way in by offering a higher price, because there is nothing left to sell.

Access is decided by who committed earliest, with the largest balance sheet, through the longest relationship. That is allocation, not a market, and allocation favours incumbents in a way markets do not.

This is the two-tier pattern from Series I arriving in a new place. Article 1 of this series found large electricity users buying their way out of the grid queue with private generation. Here the largest compute buyers have reserved the thinnest layer years ahead. In both cases, the constraint is real for everyone and binding only for those without the scale to get around it.

The alternatives that exist on paper

Is there a second source? Nominally, yes. The two largest outsourced assembly firms operate advanced packaging lines, and a major chipmaker has its own packaging technology.

But their scale for the most demanding workloads is described as a fraction of the leader's, and production-volume qualification for the most complex memory-and-processor combinations has not been publicly confirmed. The chipmaker's own foundry programme has faced publicly reported delays across several earnings calls.

This is the correction this series has made in every domain: alternatives that exist when you count them and disappear when you need them. A second source that has not been qualified at volume is not a second source for the purpose that matters. It is an option on a future second source, and the option has an unknown expiry.

The hoarding that follows

Return to the procurement manager with her golden screw. Her hoarding of peripheral parts is the allocation problem leaking outwards.

When one component is allocated rather than sold, everyone waiting for it over-orders everything else to avoid being short when it arrives. Components with no underlying scarcity become temporarily scarce because they are being held in warehouses against a delivery date that keeps moving.

That is a pattern worth recognising, because it misleads anyone reading the market. A shortage of power management chips or network switches can be entirely real on the shelf and entirely artificial in its cause, and the two require completely different responses.

5. What Most Analysis Gets Wrong

That there is a chip shortage

For most of the economy there is not. Established processes are available in weeks. There is a shortage of a very narrow set of capabilities at the leading edge, concentrated in packaging and memory, and treating it as a general shortage leads to the wrong response — stockpiling components that were never scarce.

That the constraint is the famous machines

The lithography equipment and the fabrication plants are genuine concentrations and they remain strategically important. They are not the binding constraint on AI hardware supply right now. The binding constraint is further downstream, in a stage that attracted almost no public policy attention until the shortage made it unavoidable.

That expanding capacity 80 percent a year solves it

It is expanding from a small base against demand that grew faster than anyone planned for, and each expansion depends on equipment and materials with their own long lead times. An aggressive growth rate at the bottleneck is how the bottleneck moves, not how it disappears.

That the problem ends at the chip

A finished accelerator needs electricity, and Article 1 established that connection takes four to seven years. Packaging lead times of a year or so are serious; grid connection is several times longer. For anyone building a data centre, the binding constraint over the next few years is likely to move from the chip to the socket in the wall.

6. Base, Stress and Extreme

Four paths, with our probability assessment and the condition that would falsify each. Probabilities sum to one hundred.

The fourth path deserves a sentence despite its low probability. Several layers in this stack are concentrated not only by supplier but by geography, and a geographic concentration converts any regional event — natural, political or logistical — into a global supply event. Nothing in the article depends on predicting such an event; the point is that the stack has no redundancy for one.

7. Forecast — One Year, to end-2027

Leading-edge packaging stays tight

Probability 0.65  ·  Confidence: Medium-High

We expect lead times for leading-edge advanced packaging to remain at or above forty weeks at the end of 2027, despite rapid capacity growth.

The capacity is committed years ahead and demand has grown faster than provisioning. A fourfold increase absorbed by existing reservations produces no relief for anyone who was not already in line.

Second-order effect. Smaller buyers and newer entrants shift towards designs that avoid the constrained packaging entirely, accepting lower performance for available supply. That is a real innovation pressure and it is also a two-tier outcome: the frontier stays with those who reserved it.

What would weaken it. Lead times falling below forty weeks, which would mean expansion outran reservations faster than the base rate suggests.

8. Forecast — Three Years, to 2029

A second source qualifies — or does not

Probability 0.55 that one does  ·  Confidence: Medium

We put a little better than even odds on a supplier other than the current leader being publicly confirmed by a major accelerator company as qualified at production volume for memory-integrated leading-edge packaging.

This is the most important single forecast in the article, because it tests the claim about nominal alternatives directly. Qualification at volume is slow and expensive, and buyers are reluctant to move production to an unproven line when the proven one is merely scarce. But the commercial pressure to create a second source is very large.

Second-order effect. If a second source qualifies, the allocation market begins to turn back into a price market, which favours newcomers. If it does not, allocation hardens into a durable structure in which access to frontier computing is a function of relationships formed before 2026.

What would weaken it. Continued reliance on a single supplier for the most demanding workloads through 2029, with alternative lines confined to less demanding products.

9. Forecast — Five Years, to 2031

Memory concentration eases, partially

Probability 0.50  ·  Confidence: Medium

We expect the largest high-bandwidth memory supplier's share to fall below 50 percent, from around 62 percent, as competitors that have been gaining ground close the gap.

This is genuinely uncertain. Memory is a market with a small number of capable producers, each able to expand, and concentration there has historically shifted between them more readily than in packaging. We hold it at even odds because the direction is plausible and the pace is not predictable.

Second-order effect. Easing at the memory layer, if it comes, relocates the constraint again — most likely towards power and connection, where Article 1 found substitution times several times longer than anything in the chip stack.

What would weaken it. Leading memory share holding at or above 55 percent, indicating that a technological lead has translated into a durable position.

10. Forecast — Ten Years, to 2036

Geography disperses, slowly

Probability 0.45  ·  Confidence: Low

We put somewhat below even odds on more than a quarter of global leading-edge logic manufacturing capacity sitting outside its current principal location by 2036, according to industry association or equivalent accounting.

A great deal of money and legislation is aimed at exactly this outcome, and new fabrication plants in new places are being built. Against that, the leading edge moves: each new process generation starts where the expertise already is, so dispersing the current generation does not disperse the next.

We hold this at low confidence and name why. A decade covers several process generations, and the question of where the next one starts is decided by expertise and ecosystems rather than by where the last one was relocated to. That is the same shape as the bought-versus-trained finding in Series I, applied to an industry: plants can be built in a few years; the engineering depth to run the frontier takes much longer.

What would weaken it. A new leading-edge process generation launching first outside the current principal location, which would indicate the ecosystem itself has moved rather than only the plants.

11. Signals to Watch

—   Advanced packaging lead times, which are the current thinnest layer and the most direct measure of whether the constraint is easing

—   Qualification announcements for second-source packaging at production volume — confirmed by a buyer, not claimed by a supplier

—   High-bandwidth memory share by producer, and whether sold-out conditions extend into a further year

—   Lead times for the equipment and materials that packaging depends on: bonding and placement tools, substrate films. This is where the constraint moves next

—   Peripheral component shortages, and whether they are real scarcity or hoarding against delayed accelerator deliveries

—   Where each new leading-edge process generation launches first. Plants follow money; the frontier follows expertise

—   The meeting point with Article 1: whether finished hardware is waiting on chips or on grid connections. That transition is the most important shift in this whole row

12. Recommendations — Individuals

For most people this is a background factor rather than something to act on. It surfaces in two ordinary ways.

Immediate — 30 days

If you are buying equipment that depends on high-end processors — workstations for demanding work, specialist hardware — expect availability and pricing to be driven by allocation rather than demand, and buy when a suitable configuration is available rather than waiting for a better one. For ordinary consumer electronics there is no underlying shortage and no reason to change behaviour.

Build — 12 months

If your work depends on access to large-scale computing, notice that access is concentrated in a small number of providers who themselves secured their hardware years ahead. That is a dependency with a concentration you did not choose, and it is worth knowing which provider your tools actually run on.

Position — 3 years

If you are building skills around computing infrastructure, the scarce ones sit in the unglamorous layers: packaging, power delivery, cooling, and the physical integration of systems. Those have been the constraint repeatedly and are likely to remain so, precisely because they attract less attention than design.

Avoid. Reading headlines about chip shortages as a reason to stockpile ordinary electronics. The shortage is narrow and at the frontier, and general hoarding is exactly the behaviour that creates artificial scarcity in things that were never scarce.

Why this works. Knowing that the shortage is narrow tells you it mostly does not affect you, which is useful information in its own right and saves both money and worry.

13. Recommendations — Business

The practical question is which layer of the stack your own operation is actually exposed to, and it is rarely the one you assume.

Immediate — 60 days

Separate your hardware needs into two groups: anything that depends on leading-edge accelerators or high-bandwidth memory, and everything else. The first group is subject to allocation and multi-quarter delays. The second is subject to ordinary supply and is available in weeks. Treating both as one — a single chip-shortage risk — leads to over-buying the abundant group and under-planning the constrained one.

Then check whether any shortage you are currently experiencing is real or induced. A peripheral part that is suddenly scarce may be held in other people's warehouses against delayed accelerator deliveries, and will reappear when those deliveries land.

Build — 12 months

If you depend on frontier computing, secure it through a provider that has already reserved hardware rather than trying to acquire it directly. Access at the leading edge is allocated to those who committed years ahead, and for most organisations the realistic route in is through someone who did.

And design for the possibility that you will not get the leading part. Architectures that can run on available rather than frontier hardware, even at lower performance, are an insurance policy against allocation that costs relatively little to build in early and a great deal to retrofit.

Position — 3 years

Assume the binding constraint moves, and plan for where it is going rather than where it is. For most organisations building computing capacity, the next binding layer is not a component at all — it is the grid connection from Article 1, with a substitution time several times longer than anything in the chip stack.

Avoid. Counting a second supplier as redundancy before it is qualified at the volume and workload you need. A packaging line that exists but has not been proven for your product is an option on redundancy, not redundancy.

Why this works. The stack moves at the speed of its thinnest layer, and the thinnest layer changes. An organisation that knows which layer it is exposed to, and where the constraint is heading, can plan around a constraint that others experience as a surprise.

14. Recommendations — Capital

The structural observation is that the constraint on a heavily invested theme sits in a layer that is narrow, moving, and allocated rather than priced.

Immediate — this quarter

For any holding whose growth depends on frontier computing, establish where it sits in the allocation, not only in the demand forecast. A company with strong demand and no secured packaging or memory capacity faces a different outcome from one with the same demand and capacity reserved through 2027, and headline figures rarely separate them.

Build — 12 months

Watch the layer the constraint is moving to rather than the one it is leaving. Bonding and placement equipment, substrate materials, and power delivery have each been identified as the next pressure point. Attention and capital tend to arrive at a bottleneck just as it is being relieved, which is the pattern described in section 3. This is an observation about where scarcity sits, not a recommendation about any instrument.

Separate allocation risk from demand risk. Whether demand for AI hardware holds is a market question. Whether a given company can obtain the parts to meet it is an allocation question, with different causes and a different timeline.

Position — 3 years

The three-year forecast is the pivot. If a second packaging source qualifies at volume, allocation turns back towards a market and the value of early reservations falls. If none does, early reservations become a durable structural advantage. These imply opposite conclusions about incumbents and newcomers, and the event that decides between them is a qualification announcement.

Avoid. Treating rapid capacity growth at the bottleneck as a signal the constraint is ending. Eighty percent growth from a small base against faster-growing demand is how the constraint moves, and the next binding layer was under-invested because it was not binding yesterday.

Why this works. The migration of the constraint follows a visible pattern — one layer down, each time — and the next layer can usually be identified before it binds. That gives an interval in which the scarcity is foreseeable and not yet reflected in anything.

15. What Would Change Our Mind

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

—   The whole stack reaches balanced capacity at once, with no new binding layer emerging. Our central claim is that the constraint moves rather than disappears; a stack that simply catches up everywhere would falsify it.

—   A second source qualifies quickly and disperses concentration within two years. That would show the nominal alternatives were closer to real ones than we credit.

—   Access at the leading edge returns to price-based markets without any change in underlying capacity. That would mean allocation was a temporary response rather than the structural outcome we describe.

Jurisdiction, recorded on the second row. Article 1 of this series drew almost entirely on American data and named that concentration on the first row. This article's material spans Taiwan, South Korea and the United States, and its forecasts resolve against industry and company sources across those regions. That is a modest improvement in geographic spread. The data remains concentrated in industry reporting rather than official statistics, which is a different weakness and one worth naming now.

The Series II Table — two rows

The table grows by one row per article. With two rows, it already shows something a single row could not.

What the two rows show together. Computing's longest substitution time is measured in quarters to a couple of years. Electricity's is measured in years to a decade. For anything that needs both — which is every data centre — the binding constraint is whichever is longer, and it is increasingly the grid rather than the chip. The table is beginning to do what a series of essays cannot: show which layer actually governs.

The next row is the world's shipping chokepoints.

Founder's Lens

[ EDITORIAL GATE — WRITTEN BY HAND BEFORE PUBLICATION. Never generated. Replace this marker with the founder's text, or record a suspension. ]

16. Bottom Line

A four-hundred-thousand-dollar rack of AI hardware can sit unshipped for months for want of a single component, and the reason is not where most people would look. It is not the transistors, which on established processes are available in weeks. It is not primarily the famous machines or the fabrication plants. It is a stage after fabrication — packaging the processor beside its memory — that was a routine, low-margin step until it suddenly became the thinnest layer in the stack.

Computing is not one thing. It is a stack of stages, each with its own suppliers and lead times, and the whole stack moves at the speed of whichever is thinnest. That layer changes. Attention and investment pile into the current bottleneck, the layers around it go unexpanded because they were not binding, and the moment the bottleneck eases, one of them becomes the new one.

So the honest forecast is not that the constraint gets solved. It is that it will be somewhere else next year, one layer further down, and money will arrive there slightly late.

When the thinnest layer is sold out years ahead, the market stops being a market. Access goes to whoever committed earliest with the largest balance sheet, and a newcomer cannot outbid for capacity that is already spoken for. The second sources exist on paper and have not been qualified at the volume that matters — the same nominal alternative this series has found in every domain.

And for the first time in this series, the instrument splits. The same substrate is total in its criticality for anyone building at the frontier and almost irrelevant for the ordinary electronics that most of the economy runs on. There is no general chip shortage. There is a very narrow, very severe one, and confusing the two produces hoarding of things that were never scarce.

Put this row beside the first and the picture sharpens. Chips are measured in quarters. Grid connections are measured in years. For everything that needs both, the constraint is moving from the component to the socket in the wall.

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. The first and second share a parent cause in packaging concentration; if a second source qualifies early, lead times fall with it. The third and fourth are independent of those and of each other. So this set is three families rather than four observations. The geographic spread is wider than in Article 1, and the reliance on industry rather than official sources is a limitation recorded in section 15.

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

Sources

A note on the evidence base. Semiconductor supply data comes overwhelmingly from industry analysis and company statements rather than official statistics, and several figures here are estimates or reports rather than measurements. They are consistent across independent sources, which is why they are used, and they are marked by class so that no estimate is read as a measurement. The argument — that the constraint sits in a narrow, moving layer allocated rather than priced — does not depend on any single figure being exact.

In this series

—   Previous: Article 1 of Series II, The Connection Queue — electricity, the first row of the table.

—   Next: Article 3 of Series II, on the world's shipping chokepoints — the third row.

—   The method behind the Chaos Index and this series: /methodology

THRIVE IN CHAOS

Decision Intelligence for an Uncertain World

Analysis → Forecast → Recommendations  ·  Signal → Meaning → Action → Stability

Signal Over Noise  ·  thriveinchaos.ai

AI intelligence system with human editorial oversight.

Forecasts are probability-based analytical assessments, not certainties. This material supports independent judgment and does not constitute financial, legal or investment advice. Nothing here is a recommendation to acquire, hold or dispose of any instrument.