Imagine a procurement manager responsible for assembling a rack of AI hardware worth around $400,000.

Almost everything is ready.

The chassis.

Power supplies.

Cooling assemblies.

Network switches.

Cables.

The rack itself.

Even the building that will house it.

One component is missing.

The accelerator.

Without it, the entire rack becomes inventory that cannot be shipped, cannot be invoiced and slowly loses value while sitting in a warehouse.

The industry has a useful expression for this:

the golden screw.

One missing component makes everything around it temporarily worthless.

But follow that missing component backwards through the supply chain and something more interesting appears.

The shortage is not where most people have been looking.

Public discussion tends to compress the entire semiconductor industry into one word:

chips.

That makes the problem sound simple.

Either chips are available or they aren't.

But computing is actually a stack.

There is processor design.

Lithography equipment.

Wafer fabrication.

High-bandwidth memory.

Advanced packaging.

Packaging materials and equipment.

And finally the electricity and physical connection required to run the finished hardware.

Every layer has different suppliers.

Different geography.

Different lead times.

Different levels of concentration.

And every necessary layer has to work before the finished product exists.

An abundance of fabrication capacity cannot compensate for missing packaging.

An abundance of processors cannot compensate for missing memory.

And a warehouse full of finished accelerators cannot operate without electricity.

The entire stack therefore moves at the speed of its thinnest layer.

For frontier AI hardware, a major constraint now sits after fabrication.

Advanced packaging places the processor beside stacks of high-bandwidth memory and connects them through an extraordinarily dense substrate capable of moving huge quantities of data between them.

This used to be a relatively unglamorous part of semiconductor manufacturing.

Then AI workloads changed what packaging had to do.

Demand grew faster than capacity had been designed for.

Reported lead times for leading advanced packaging reached roughly 52–78 weeks.

High-bandwidth memory has been described as effectively sold out for 2026, with one supplier accounting for roughly 62% of the market in the industry estimates used in our analysis.

Compare that with chips made on established manufacturing processes, which can be available within roughly 4–17 weeks.

This gives us the first important correction:

There is not one general chip shortage.

For much of ordinary electronics, substitution is available and relatively fast.

For frontier AI computing, the shortage is narrow, concentrated and severe.

Those are completely different markets.

If you believe everything is scarce, the rational response is to secure everything.

That is exactly where the next distortion begins.

Return to our procurement manager.

She doesn't know when the accelerator will arrive.

But when it finally does, she cannot afford to discover that she is missing a power-management component, cable or network switch.

So she buys those parts early.

And she buys more than she immediately needs.

Every other assembler facing the same uncertainty does the same.

Soon, components with no fundamental production shortage begin disappearing from available inventory.

The shortage is real on the shelf.

But its cause is not a lack of manufacturing capacity.

It is hoarding around the actual golden screw.

One narrow bottleneck has created apparent scarcity throughout the surrounding system.

That is why diagnosing the exact layer matters.

There is a deeper mechanism here.

For most of the last decade, strategic attention focused on semiconductor fabrication and lithography.

That made sense.

Both were genuinely concentrated.

Governments responded.

Industrial policy shifted.

New fabrication plants were funded.

Capacity began expanding geographically.

But while public attention was focused on fabrication, the binding constraint moved downstream.

To packaging.

Now packaging capacity itself is expanding extremely quickly.

That sounds like the end of the story.

It isn't.

A new packaging line requires specialised bonding and placement equipment.

Those machines have reported lead times of around 12–18 months.

The process requires specialised substrate materials.

Some have already been reported in short supply.

And the finished package requires high-bandwidth memory, which is itself constrained.

So the sequence begins to look like this:

Fabrication → Packaging → Memory → Equipment → Materials → Power

The constraint moves.

Why?

Because yesterday's bottleneck receives today's investment.

Everything around it receives less because it isn't yet binding.

Once enough capacity arrives at the bottleneck, the next thinnest layer becomes visible.

And that layer was under-expanded precisely because it was not the problem yesterday.

This leads to one of the most important rules in this series:

A rapidly growing stack does not necessarily have a shortage that gets solved.

It can have a shortage that migrates.

This creates an uncomfortable problem for industrial policy.

Government response is slow for understandable reasons.

A concentration has to become visible.

The problem has to be studied.

Legislation is drafted.

Budgets are approved.

Projects are designed.

Factories are built.

By the time physical capacity arrives, several years may have passed.

But a bottleneck inside a rapidly expanding technology stack can move within one or two years.

So policy can be completely rational and still arrive late.

It solves the bottleneck that was visible when the legislation was written.

Meanwhile, the system has moved on.

This is the same correction-loop problem we identified in Series I, now appearing in industrial policy.

The honest forecast is therefore not:

the semiconductor bottleneck will be solved.

It is:

the bottleneck will probably be somewhere else next year.

And a lot of money will arrive at the old location slightly too late.

There is another consequence of extreme scarcity.

When critical capacity is sold out years ahead, price loses some of its normal function.

Suppose a newcomer wants access to the leading packaging process.

Normally, a buyer facing scarcity can offer a higher price.

But what happens when the capacity through 2027 has already been reserved?

There is nothing available to bid for.

Access is increasingly determined by:

who committed first;

who reserved the largest volumes;

who has the strongest balance sheet;

and who already has the deepest supplier relationships.

That is not quite an ordinary price market.

It is allocation.

And allocation has a different competitive structure.

It favours incumbents.

Nominally, alternative suppliers exist.

That sounds reassuring.

But this series uses a stricter definition of redundancy.

A second source counts only if it can actually replace the first source for the workload, quality and volume required.

A packaging line that exists but has not been qualified at production scale for your product is not redundancy.

It is a possible future source.

That distinction matters because companies often count suppliers rather than capabilities.

Two supplier names on a spreadsheet do not necessarily mean two functioning alternatives.

Our base path, with a 50% probability, is the moving bottleneck.

Advanced packaging expands aggressively, but the binding constraint shifts into memory, specialised tools, substrate materials or electricity.

The shortage changes location rather than disappearing.

We assign 25% to genuine second sources emerging.

If another supplier qualifies at production volume for the most demanding workloads, concentration begins to fall and allocation can start turning back into a normal price market.

Another 15% goes to demand cooling enough for capacity to catch up.

And 10% goes to a concentration shock — a disruption affecting one of the highly concentrated suppliers or locations and interrupting leading-edge supply for months.

The point of the last scenario is not to predict a particular event.

It is to recognise that several layers of the stack have very little practical redundancy.

First, stop treating all hardware as one procurement category.

Separate:

frontier accelerators and high-bandwidth memory

from

ordinary computing hardware.

The first operates under allocation and multi-quarter delays.

The second often operates under relatively normal supply conditions.

Combining them under one “chip shortage” assumption encourages over-buying the abundant category while failing to plan properly for the constrained one.

Second, distinguish real scarcity from induced scarcity.

A peripheral component may be difficult to find not because production is insufficient, but because other companies are holding inventory while waiting for their accelerator allocations.

Third, design for hardware flexibility where possible.

An architecture that can operate on available rather than frontier hardware — even at lower performance — creates an option when access becomes more important than maximum performance.

And finally:

plan for where the constraint is going, not only where it is today.

Demand and allocation are different risks.

A company can have extraordinary demand for its product and still be unable to obtain the capacity required to supply it.

So for any business whose growth depends on frontier computing, the useful question is not only:

How much demand exists?

It is:

Where does this company sit in the allocation?

Has packaging capacity been secured?

Has memory been secured?

For how long?

Then look one layer beyond today's bottleneck.

Specialised packaging equipment.

Substrate materials.

Power delivery.

Grid connections.

Capital tends to notice scarcity after scarcity has become obvious.

But by then the system may already be moving toward the next constraint.

This is where Article 2 connects directly to Article 1 of Series II.

For frontier computing, the current substitution times are mostly measured in quarters or a few years.

Advanced packaging: roughly 52–78 weeks.

Qualifying a genuine second source: potentially years.

Mature chips: roughly 4–17 weeks.

Now compare that with electricity.

Large grid connections can take four to seven years.

Transformers can take more than three.

The engineering pipeline behind the system is longer still.

A data centre needs both computing hardware and electricity.

So whichever takes longer governs.

And as semiconductor capacity expands, the binding constraint can migrate beyond semiconductors entirely.

The AI infrastructure story may therefore move through a sequence that looks something like:

Chip → Packaging → Memory → Equipment → Power

Which brings us to the most important conclusion from this article.

The problem with complex systems is not simply that they contain bottlenecks.

It is that the bottleneck moves after you invest in it.

The companies, governments and investors that keep looking at yesterday's narrowest point will repeatedly arrive late.

The useful question is different:

Once today's bottleneck is relieved, what becomes the narrowest point tomorrow?

That is THE NARROWEST POINT.

Next in Series II:

The world's shipping chokepoints.

THRIVE IN CHAOS
Decision Intelligence for an Uncertain World

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

Forecasts are probability-based analytical assessments, not certainties. This material supports independent judgment and does not constitute financial, legal or investment advice.