How accelerating technology, weakening competence formation, and the retirement

Zuzma (OP)

Sr. Member

Activity: 552
Merit: 285

HODL

How accelerating technology, weakening competence formation, and the retirement

September 06, 2026, 02:15:45 AM
Last edit: September 06, 2026, 06:38:12 PM by Zuzma

When Technology Outruns Competence

What if the real bottleneck of the AI age is not technology — but the system required to understand, deploy, verify, and maintain it?

We are used to discussing artificial intelligence primarily as a threat to jobs.
But another problem may become at least as important:

technology may develop faster than the human and organizational capabilities required to understand, implement, verify, and maintain it.

If that happens, the main constraint on economic growth will not necessarily be computing power, software, or access to AI.
It may increasingly become complementary human and organizational capital.
The central risk is not that technology becomes useless.
It is that technological capability grows faster than society’s ability to convert that capability into reliable economic output.

The bottleneck may be shifting
AI is becoming cheaper, more capable, and more widely available.
According to the Stanford AI Index 2025, the share of surveyed organizations reporting AI use increased from 55% in 2023 to 78% in 2024. The share using generative AI in at least one business function rose from 33% to 71%.
Global corporate AI investment reached $252.3 billion, while private AI investment in the United States alone amounted to $109.1 billion.
An enormous accumulation of technological capital is taking place.
Technology itself is not disappearing as a constraint. AI systems still face problems involving reliability, hallucinations, energy consumption, data quality, cybersecurity, semiconductor capacity, and integration.
But in many sectors, the bottleneck may gradually shift from:

access to technology
toward:
the ability to deploy technology effectively.

Technology by itself does not solve a problem.
AI can write a program, perform a calculation, produce a technical diagram, analyze data, identify possible faults, or propose an engineering solution.
Someone must still determine:

AI dramatically reduces the cost of obtaining an answer.
It does not automatically eliminate the need to understand that answer.
This suggests an important relationship:

a competent specialist + AI can produce enormous productivity gains;
an incompetent specialist + AI can produce errors at enormous speed.

The relationship is not literally multiplicative in every task, but in many complex systems AI and human competence behave as complementary inputs.
A useful approximation is:

effective AI value ≈ technological capability × human competence × organizational capability

A weakness in any one of these factors can limit the result.

Technology requires complementary capital
This problem is not unique to AI.
Economists studying general-purpose technologies have long observed that major technologies often require extensive complementary investments before their productivity potential becomes visible.
New technology may require:

new methods of coordination.

Brynjolfsson, Rock, and Syverson describe this phenomenon through the idea of a Productivity J-Curve: firms may make large investments in a transformative technology while measured productivity initially improves only slowly because complementary intangible capital still has to be created.
AI may intensify this problem because the technological component is advancing extraordinarily quickly.
A company can purchase AI systems, computing capacity, robots, industrial software, digital twins, or automated engineering tools.
But between technological capability and actual productivity stand:

An economy can therefore possess extremely advanced technology while lacking the complementary capabilities required to extract its full value.
The problem is more accurately described not as:

excess technology + lack of competence
but as:
rapid growth of technological capital + slower growth of complementary human and organizational capital.

Explicit knowledge becomes cheap; tacit competence may become expensive
The economic value of professional knowledge may therefore change.
AI sharply reduces the cost of retrieving and generating explicit knowledge.
It can explain a formula.
It can generate PLC code.
It can summarize technical documentation.
It can produce a circuit diagram.
It can explain how a variable-frequency drive works.
But professional competence consists of more than explicit knowledge.
It also contains:

knowledge + experience + judgment + context + responsibility

A person may understand how a variable-frequency drive works and still be unable to commission a real industrial installation.
A person may know PLC syntax and still fail to understand the physical process controlled by the PLC.
A person may generate formally correct code while completely misunderstanding the machine on which that code will run.
Much of professional expertise is tacit.
It is accumulated through:

working under uncertainty.

As explicit knowledge becomes cheaper, tacit competence may become relatively more valuable.

The scarce skill may increasingly be not the ability to retrieve an answer, but the ability to determine whether the answer makes sense.

The important shortage may not be “AI programmers”
There will certainly be demand for specialists who build AI systems.
But a more consequential bottleneck may emerge among people working at the interface between digital intelligence and physical, regulated, or safety-critical systems.
These may include:

specialists who simultaneously understand a physical domain, software, and AI.

The causes of shortages differ between professions.
A shortage of electricians does not arise for the same reason as a shortage of cybersecurity specialists.
Labor-market conditions also vary by country and industry.
The common feature is different:

AI output must eventually interact with the real world.

For example, AI can generate a PLC program.
But someone must understand the entire chain:

sensor
↓
electrical diagram
↓
PLC
↓
program
↓
variable-frequency drive
↓
motor
↓
mechanics
↓
technological process
↓
safety

A formally correct piece of generated code may solve very little if nobody understands the complete system.
A machine may fail to start because of:

an incorrect process assumption.

The code may be perfect.
The system may still not work.
The same principle applies in power systems, medicine, chemical plants, transport infrastructure, construction, manufacturing, and other domains in which errors have physical consequences.

Foundational skills were under pressure before generative AI
Educational data do not show a universal decline across every country, subject, and population group.
They do, however, show significant deterioration in several foundational competencies that began before generative AI became widely available.
According to OECD PISA 2022 — Germany, Germany’s results were the lowest recorded for the country in mathematics, reading, and science.
Compared with 2012, the share of German 15-year-olds performing below the basic proficiency level increased by:

11 percentage points in science.

The IQB Bildungstrend 2022 / Kultusministerkonferenz shows a similar pattern.
Between 2015 and 2022, average scores among German ninth-graders fell by:

31 points in spelling.

English performance improved.
This indicates an uneven deterioration in specific competencies rather than evidence of a general decline in intelligence.
The United States shows another version of the same problem.
According to the NAEP Long-Term Trend Assessment 2023, average results among 13-year-olds declined over the previous decade by 7 points in reading and 14 points in mathematics.
More recent NAEP assessments also indicate that the broader weakness has not simply disappeared: 2024 grade-12 reading results and 2024 grade-12 mathematics results remained below their respective 2019 levels.
These assessments do not establish a single cause.
They do not prove that future engineers, doctors, or technicians will necessarily become less competent.
Professional expertise is created not only in schools but also through universities, vocational education, apprenticeships, mentoring, and years of practical work.
The evidence supports a narrower conclusion:

some foundational competencies were already under pressure before generative AI entered schools and workplaces.

That matters because AI is now entering an educational environment in which the formation of basic skills is already challenging.

AI may amplify the problem — or help solve it
Generative AI did not create these educational declines.
But it may affect what happens next.
To become an expert, a person traditionally had to perform a large amount of cognitive work:

make decisions independently.

These activities were not merely a way of producing work.
They were also a way of producing competence.
Professional intuition emerged from repeated exposure to problems.
If AI substitutes for this practice rather than augmenting it, a destructive feedback loop may emerge:

AI performs the difficult task
↓
the person practices less
↓
the person develops less independent understanding
↓
verification of AI becomes more difficult
↓
dependence on AI increases
↓
independent practice declines further

This is one possible mechanism of:

skill degradation through the automation of cognitive practice.

But the opposite outcome is equally possible.
AI can:

help students explore systems faster.

Then the cycle becomes:

AI accelerates feedback
↓
the person experiments more
↓
learning becomes faster
↓
competence grows
↓
AI becomes an increasingly effective professional tool

The critical distinction is therefore not whether AI is used.
It is how AI is used.
AI may accelerate skill formation when it supports deliberate practice.
It may cause skill atrophy when it replaces deliberate practice.
For broader evidence on the relationship between AI adoption and skills, see the OECD - AI and Skills: What We Know So Far.

The paradox of expert formation
This may be the most important long-term issue.
AI can make existing experts dramatically more productive while simultaneously weakening some of the mechanisms through which new experts are created.
An experienced engineer who already understands a process can use AI to:

explore alternative solutions.

These tools may multiply the productivity of someone who already possesses strong domain knowledge.
But junior engineers traditionally develop that domain knowledge by doing many of those same tasks themselves.
They calculate.
They search manuals.
They make design mistakes.
They write poor code.
They debug it.
They misdiagnose faults.
They correct their reasoning.
They slowly build mental models of real systems.
If these developmental tasks are increasingly delegated to AI, an important question appears:

where will the experienced engineer come from ten years later who is capable of correcting the AI itself?

Professional competence is not merely a flow of newly trained workers.
It is a stock accumulated over time.
A simplified representation might be:

H(t+1) = H(t) + training + experience − retirement − obsolescence − deskilling

where H represents the stock of human competence.
AI may simultaneously:

increase the risk of deskilling.

The central question is therefore not simply:

Does AI increase productivity today?

It is also:

Does the system continue to reproduce the human competence required to operate tomorrow?

A society could theoretically consume the professional competence accumulated by previous generations faster than it creates a new generation of experts.
That is a much more serious problem than a temporary shortage of workers.

Skills shortages are already constraining transformation
There is already evidence that companies consider skills shortages an important obstacle to technological change.
According to the World Economic Forum — Future of Jobs Report 2025:

77% of employers plan to reskill or upskill employees in response to AI.

OECD research also suggests that firms with stronger digital skills, ICT infrastructure, and complementary capabilities are more likely to adopt and benefit from AI.
See:

OECD — A Portrait of AI Adopters Across Countries
OECD — Fostering an Inclusive Digital Transformation as AI Spreads Among Firms
[/list]
These findings do not prove that a future competence crisis is inevitable.
They establish something narrower:

the ability to absorb and deploy technology already differs significantly between organizations, and skills are already one important constraint.

The open question is whether technological capability will continue to grow faster than deployment capability.

A more realistic model of AI productivity
A simplified production function might be written as:

Y = A × F(K, L, H)

where:

But the productivity effect of AI does not depend only on the existence of AI itself.
It may be more useful to treat productivity as a function of:

A = A(T, H, O)

where:

O — organizational capability.

AI sharply increases T.
Investment in data centers, machinery, energy systems, semiconductor capacity, robotics, and computing infrastructure increases technological and physical capital.
But if H and O grow more slowly, the economy encounters a complementary-capital bottleneck.
The technology remains valuable.
The problem is that additional technological investment may produce less realized output until the complementary system catches up.
This is especially important because human and organizational capital often take much longer to build than computing infrastructure.
A data center can be constructed in years.
A highly competent engineer may require a decade of education and professional experience.
An effective industrial organization may require years of accumulated routines, documentation, culture, and institutional knowledge.
The adjustment speeds are different.

High AI investment may therefore produce temporarily weak realized returns
Suppose a company spends $10 million on AI infrastructure.
Management expects a major productivity increase.
But the company may lack:

internal accountability structures.

The theoretical capability of the system may therefore be much larger than its realized economic value.
This does not necessarily mean that the technology is overvalued or ineffective.
It may mean that complementary investment is incomplete.
This distinction matters.
Weak measured returns during the early stages of adoption may represent:

or simply a delay between investment and measurable output.

The Stanford AI Index 2025 reports that AI adoption is already widespread, but the financial effects reported within individual business functions remain relatively modest in many firms.
Among companies reporting cost savings, most estimate them at less than 10%.
The most commonly reported revenue increase is below 5%.
These figures are not evidence that AI will ultimately generate small productivity gains.
They illustrate that technological capability and realized business value are not the same thing.

The United States has an important advantage
If human capital becomes a major bottleneck, the United States possesses one structural advantage: its historical ability to attract highly qualified specialists from abroad.
The American response can partly be:

insufficient domestic human capital → import human capital

Researchers, engineers, doctors, software specialists, and entrepreneurs have long moved to the United States from:

That may make a competence shortage less damaging for the United States than for economies with less attractive or more closed labor markets.
But professional competence is not equally mobile.
A software researcher may relocate comparatively easily.
A commissioning engineer, electrician, doctor, infrastructure specialist, or industrial technician may face:

The more a profession is embedded in physical infrastructure and local regulation, the harder it is to solve shortages simply through migration.

Economic concentration could increase — but AI also pushes in the opposite direction
If complementary competence becomes scarce, large firms may gain an advantage.
Large technology and industrial corporations can recruit highly qualified specialists globally.
Small firms often cannot.
This may create a reinforcing loop:

the strongest specialists
↓
the largest companies
↓
better AI deployment
↓
higher productivity
↓
more capital
↓
greater ability to recruit scarce specialists

This could widen productivity differences between large corporations and smaller companies.
But AI also creates an opposing force.
Small firms now gain inexpensive access to capabilities that once required large internal departments:

research support.

AI therefore creates two opposing tendencies:

democratization of cognitive capability
and
concentration of complementary capital.

The long-term economic outcome may depend on which effect dominates.

A useful — but limited — analogy with the late 1990s
There is a useful analogy with the internet boom, but it should not be overstated.
In the late 1990s, investors correctly recognized the internet as a transformative technology.
What many misjudged was not the importance of the technology itself, but the speed at which technological potential would become sustainable profit.
Capital flowed into internet companies and telecommunications infrastructure.
Valuations rose rapidly.
The underlying technological transformation was real.
But many investment assumptions proved too optimistic.
In his 2001 speech “What Happened to the New Economy?”, Federal Reserve Governor Laurence H. Meyer examined the productivity acceleration, investment boom, equity-price surge and subsequent retrenchment associated with the technology cycle.
The strongest analogy with AI is therefore not:

AI today = dot-com companies in 2000.

It is:

markets can correctly identify a revolutionary technology while incorrectly estimating the speed of diffusion and the timing of economic returns.

The differences are substantial.
AI software can diffuse extremely quickly.
Many AI investments support assets with real and potentially durable economic value:

robotics.

But diffusion may remain much slower in physical industries such as:

infrastructure.

These sectors require interaction with machines, regulation, safety systems, existing equipment, and skilled professionals.
A potentially unstable combination therefore exists when:

educational and training systems adapt more gradually.

If markets price companies according to expected future AI productivity, while the productivity effect takes much longer to materialize, valuations can temporarily outrun realized economic returns.
The revolutionary technology can still be real.
The timing assumptions can still be wrong.

Why the competence-bottleneck scenario may not occur
There are strong counterarguments.
AI may dramatically accelerate professional education.
Expert systems may reduce the amount of information every worker must personally remember.
Robotics may compensate for some labor shortages.
Higher wages may attract workers into technically demanding professions.
Education systems may adapt.
Companies may redesign apprenticeships and junior professional roles around AI.
International migration may partly compensate for demographic shortages.
AI systems may become much more reliable at verifying other AI systems.
Formal verification, simulation, automated testing, digital twins, standardized interfaces, and safety systems may reduce the amount of human intervention required.
The task itself may also change.
Perhaps future engineers will not need to know everything previous generations knew.
The important skill may become the ability to supervise networks of automated tools rather than perform every intermediate task manually.
In this scenario, AI would not consume human competence.
It would amplify and reproduce it more efficiently.
The outcome is therefore not predetermined.
The critical institutional question is whether education and professional training adapt quickly enough to ensure that AI:

augments the formation of expertise
rather than:
replacing the process through which expertise is formed.

The deeper problem
AI may be creating a new imbalance between different forms of capital.
Technological capability can now increase extremely quickly.
Computing capacity can be purchased.
Software can be copied.
Models can be deployed globally.
But human competence is accumulated more slowly.
A highly capable professional must often pass through:

education
↓
practice
↓
failure
↓
correction
↓
experience
↓
judgment
↓
expertise

Organizational competence develops slowly as well.
Companies accumulate effective procedures, internal standards, documentation, culture, and institutional knowledge over years.
Technology can therefore scale faster than the system required to use it.
This is the central hypothesis:

Quote

The next major economic bottleneck may not be technological capability itself, but the complementary human and organizational capital required to convert technological capability into reliable output.

The greatest long-term risk appears if several processes occur simultaneously:

technological investment continues accelerating.

Then society could begin consuming the stock of professional competence accumulated by previous generations faster than it reproduces it.
That would create a paradox.
Humanity would possess extraordinarily powerful machines.
It would possess enormous computing capacity.
It would possess advanced AI systems capable of producing answers at almost zero marginal cost.
But the scarce resource would become the person capable of asking:

Is this the correct problem?
Does this answer make physical sense?
What happens when the real system behaves differently?
What are the safety consequences?
What has the machine failed to understand?

The most important shortage of the twenty-first century may therefore not be artificial intelligence.
It may be:

Quote

a human being capable of understanding what needs to be done, why it needs to be done, how the real system works, and how to determine when the machine is wrong.

The deeper risk is that these processes may not merely develop at different speeds, but may increasingly move in opposite directions.
Technological capability is accelerating at a historically unusual rate, while the formation of complementary human competence may be slowing, stagnating, or deteriorating in some areas.
At the same time, a large part of today’s professional competence is inherited capital:
knowledge accumulated over decades by engineers, technicians, doctors, operators, and other experienced specialists who are now progressively leaving the workforce through retirement.
Their departure does not simply reduce the number of workers.
It removes tacit knowledge, judgment, troubleshooting experience, and institutional memory that cannot be reproduced quickly through software, documentation, or formal education.
This makes the adjustment process far less controllable than a conventional technology-adoption delay.
In the late-1990s technology boom, one important problem was that investment and expectations moved faster than the economy’s ability to convert new technology into sustainable returns.
The present situation may contain the same mechanism, but with an additional force acting in the opposite direction:

technological capability may be accelerating precisely while parts of the human-capital stock required to absorb it are being depleted.

If these vectors continue to diverge, the resulting mismatch could be larger, more persistent, and economically more disruptive than a simple delay in technology diffusion.
The problem would no longer be that society needs time to catch up with its machines.

It would be that the machines are accelerating while part of the accumulated human competence required to understand, deploy, verify, and maintain them is simultaneously disappearing.

The most dangerous scenario is not:
“AI replaced the human.”
It is:
“Humanity created extraordinarily powerful technologies, but failed to reproduce enough of the human competence required to understand, deploy, verify, and control them.”

Selected sources

Stanford AI Index 2025 — Economy
Brynjolfsson, Rock & Syverson — The Productivity J-Curve
OECD — PISA 2022 Germany
KMK / IQB Bildungstrend 2022
NAEP Long-Term Trend 2023
World Economic Forum — Future of Jobs Report 2025
OECD — A Portrait of AI Adopters Across Countries
OECD — AI Diffusion and Complementary Assets
OECD — AI and Skills
Federal Reserve — Laurence H. Meyer, “What Happened to the New Economy?”

[/list]

Zuzma (OP)

Sr. Member

Activity: 552
Merit: 285

HODL

Re: How accelerating technology, weakening competence formation, and the retirement

September 06, 2026, 02:16:06 PM

When Technological Capability Outruns the System Around It

The problem may not be AI itself. The problem may be the condition of the system around it.

I want to draw attention not to a single isolated problem involving artificial intelligence, skills shortages, education, or economics. These topics are usually discussed separately. But when they are examined in isolation, the larger picture can be missed.
The real problem may be that society is entering the age of artificial intelligence after already accumulating a broad set of structural weaknesses: deteriorating educational outcomes, an aging professional core, industrial weakness, debt and fiscal pressure, and institutional overload.
At the same time, public attention and investment have become heavily concentrated on accelerating one variable above all others:

technological capability.

AI is not creating this crisis.
AI is entering an already stressed system.
Capital is being invested on a massive scale into the things that can be scaled fastest:

chips → data centers → models → software → automation

Meanwhile, other parts of the system are growing slowly, stagnating, or in some cases deteriorating.
This is why the problem is not simply that technological capital is advancing faster than human competence.
The picture is broader.
Technological capability may accelerate sharply while human competence, demographic stability, fiscal capacity, industrial capacity, organizational capability, and institutional capacity all come under simultaneous pressure.

These problems began before AI
This distinction matters because most of these problems began long before AI.
The aging of skilled professionals did not begin because of ChatGPT.
Problems in school education did not begin because of generative AI.
Europe’s demographic structure did not emerge because of AI.
The problems of Germany’s industrial model did not begin because of AI.
France’s public debt did not arise because of AI.
The structural budget deficit of the United States did not begin because of AI.
Humanity is therefore not beginning the AI revolution from an ideal starting point:

healthy society + new technology

It is beginning from something closer to:

already stressed system + extraordinary new technology

That is a fundamentally different situation.
Germany is simultaneously dealing with questions of productivity, shortages of skilled workers, education, population aging, administrative burden, and structural transformation.
France faces serious fiscal and debt constraints.
The United States faces a large structural budget deficit and a continuing increase in public debt.
These are three very different economies facing three different sets of problems.
But the important point is not to analyze each of them separately.
The important point is that all of these weaknesses can limit the ability of society to convert technological potential into real economic output.

AI is an engine. The economy is the entire machine.
Imagine AI as an extraordinarily powerful new engine.
Investors are looking primarily at the engine:

But an economy is not an engine.
It is the entire machine.
It requires:

engine + transmission + chassis + driver + fuel + roads + maintenance + financing + rules

AI is only one component of that system.
Even if it is the most revolutionary one.
If the engine becomes ten times more powerful, but bridges are aging, mechanics are becoming scarce, drivers are being trained less effectively, the company is heavily indebted, the electrical grid requires modernization, the bureaucracy cannot keep up, and the industrial base is losing competitiveness, then a tenfold increase in engine power does not imply a tenfold increase in the speed of the entire machine.

Where the investment shift may begin
This is where an investment shift may eventually emerge.
In the first stage, the market sees:

AI capability is exploding.

The conclusion appears obvious:
if AI can do more and more, then the potential value of the technology must be enormous.
This produces massive investment.
In the second stage, companies begin deploying AI at scale.
But gradually a gap appears between:

theoretical capability
and
realized economic output

Investors begin asking:

Why has the model become ten times better while the company has not become ten times more productive?

In the third stage, the answer may begin to shift from:

AI is not yet good enough

to:

the problem lies outside AI.

There is not enough competence.
There is not enough energy.
There is not enough infrastructure.
There is not enough capital.
There are not enough organizational structures.
There is not enough industrial capacity.
There are not enough people.
There is not enough time.
There is not enough institutional capacity for transformation.
This is where a real moment of recognition may occur.
Investors may discover that:

the marginal value of additional technological capability is being constrained by deterioration elsewhere in the system.

In other words, the next billion dollars invested in AI may no longer produce the same realized value because the bottleneck has moved.
In simplified form:

AI₁₀₀ + Society₄₀
may not be dramatically more productive than:
AI₇₀ + Society₄₀

The limiting factor is no longer AI.
At that point, capital may be forced to return its attention to areas that have been less fashionable for decades:

Investment logic itself may eventually redirect capital back toward the real economy and complementary capital.

The problem money cannot solve quickly
But there is a more difficult layer to this problem.
By the time markets recognize the imbalance, some of these capabilities may no longer be recoverable quickly with money.
Servers can be ordered relatively quickly.
But it is impossible to order:

100,000 engineers with twenty years of experience for delivery in eighteen months.

It is impossible to instantly buy a generation of skilled technicians.
It is impossible to instantly restore institutional memory.
It is impossible to rebuild the technological culture of an industrial organization in a short period.
It is impossible to purchase decades of operational experience.
It is impossible to instantly create teachers.
It is impossible to rebuild an apprenticeship system in a year.
It is impossible to quickly purchase engineering judgment.
Society may therefore recognize the misallocation of resources only after rebuilding complementary capital already requires decades.

We may have been living off accumulated competence
This is why aging specialists occupy a much more important place in the picture than it may initially appear.
For decades, societies may have been living off an accumulated stock while failing to notice problems in the flow that replenishes it.
A 50- or 60-year-old engineer today is the product of a school system from decades ago, technical education, decades of work, mentoring from an older generation, and a vast number of real breakdowns, mistakes, commissioning projects, and corrections.
That competence is capital accumulated over 30 or 40 years.
It may now be leaving the system much faster than society can create an equivalent replacement.
AI can create a dangerous illusion in this situation:

the shortage can simply be compensated for by technology.

Technology may help replace some functions of a specialist.
It does not necessarily replace the systemic judgment on which the operation of an entire organization depends.

AI as an amplifier — or as a mask
The same applies to education.
The important point is not to use deteriorating educational indicators as direct proof of a future shortage of engineers.
The more important observation is that the system responsible for reproducing human capital was already showing signs of weakness before the arrival of AI.
AI may then produce two opposite effects.
The first scenario is AI as an amplifier.
It accelerates learning and helps repair the problem.
The second scenario is AI as a mask.
It allows a person to produce acceptable output while possessing less internal competence.
The second scenario is especially dangerous.
Because statistically it may appear that:

productivity is fine.

But the underlying stock of competence may still be declining.

In other words, AI may temporarily conceal the deterioration of human capital.

This is not simply another Productivity J-Curve
Even the analogy with the dot-com era takes on a different meaning in this framework.
In the 2000s, complementary capital largely had to be created.
Today, some forms of complementary capital may have to be:

created + preserved + rebuilt

at the same time.
The older problem looked like this:

technology grows faster than complementary capital.

The new potential problem is different:

technology grows faster while parts of complementary capital are deteriorating.

This is no longer simply a Productivity J-Curve.

The central hypothesis
The central hypothesis is therefore not merely that AI develops faster than human competence.
It is this:

Quote

The AI revolution is not arriving in a system waiting to be transformed. It is arriving in societies already carrying accumulated structural weaknesses: aging professional populations, deteriorating educational outcomes in some areas, strained public finances, industrial restructuring, demographic pressure, infrastructure constraints, and institutions that adapt far more slowly than technology.

Quote

The central risk is therefore not merely that AI develops faster than human competence. It is that technological capability accelerates precisely while several of the complementary systems required to convert that capability into economic output are simultaneously weakening.

Quote

For a time, rapidly improving AI may obscure this imbalance. Eventually, however, investors may discover that the limiting factor is no longer the intelligence of the machine, but the condition of the society and economy in which that machine must operate.

This is why the central picture should not be imagined as only two lines:
technological capability and human competence.
It should be imagined as one line rising sharply:

Technological capability ↑↑↑

and several others:

[]Human competence
[]Education system
[]Industrial capacity
[]Fiscal capacity
[]Infrastructure
[]Institutional capacity
[/list]
Some of them grow slowly.
Some stagnate.
Some come under increasing pressure.

The central question

What happens when one component of civilization becomes exponentially more capable while much of the complementary system remains slow-moving, financially constrained, demographically stressed, or begins to deteriorate?

The problem is not AI.
The problem is the condition of the system around it.

And if attention remains focused only on the one technological variable rising fastest, society may fail to notice the deterioration of the broader system required to understand, deploy, finance, maintain, control, and ultimately convert that technology into real economic output.

Zuzma (OP)

Sr. Member

Activity: 552
Merit: 285

HODL

Re: How accelerating technology, weakening competence formation, and the retirement

September 06, 2026, 06:29:24 PM

One Trade, Many Sectors: The Hidden Concentration of the AI Boom
Why Broad Market Growth Actually Depends on a Single Source of Capital
Current economic and stock-market growth is becoming increasingly dependent on a single source of demand: capital expenditure by the world’s largest technology companies on artificial intelligence infrastructure.
On the surface, the market looks broad.
It is no longer only Nvidia, Microsoft, Amazon or Alphabet that are rising. Server manufacturers, memory producers, cooling-system suppliers, transformer manufacturers, turbine producers, electrical-equipment companies, utilities, data-center builders and a wide range of industrial companies are rising as well.
This creates the impression of diversification.
In reality, a significant part of this growth represents one enormous investment cycle spreading across multiple industries.
Economically, the final source of demand remains the same.
And if that source of capital expenditure reverses, the decline will not remain confined to the technology sector. It will propagate through the entire chain.
1. The Growth Is Real. But the Engine Is One
BlackRock described the situation in 2026 with remarkable precision:
“real growth, narrow engine.”
According to BlackRock, hyperscaler capex is expected to reach approximately:
$715 billion
in 2026 — more than 80% above the previous year.
At the same time, non-residential investment, despite accounting for only around 14% of the US economy, has generated approximately 48% of real GDP growth since the end of 2024.
But the stock-market numbers are even more revealing.
A basket of just 32 AI-related companies contributed approximately +12 percentage points to S&P 500 returns year-to-date, while the entire remainder of the index combined contributed approximately:
-2.3 percentage points
Source:
BlackRock — Getting Serious in Summer Markets
The investment growth is real.
The profits are real.
The orders are real.
But the source of capital is extremely concentrated.
2. AI Has Long Since Expanded Beyond the Technology Sector
J.P. Morgan examined five AI baskets containing 148 companies, covering:

hyperscalers;
semiconductors;
memory;
data centers;
cooling;
electrification;
software;
other elements of AI infrastructure.

By July 2026:

around 70% of the companies had risen;
the median company had gained more than 20%;
8 of the 11 S&P 500 sectors were represented;
around 40% of the companies were outside the technology sector entirely.

Source:
J.P. Morgan — Is It All One Big AI Trade?
At first glance, this looks like a broadening market.
But the underlying structure of demand tells a different story.
3. One Dollar of AI Capex Travels Through the Entire Economy
The chain looks approximately like this:

Hyperscalers
↓
GPU
↓
Servers
↓
Data Centers
↓
Cooling
↓
Electricity
↓
Transformers / Grid
↓
Power Generation
↓
Construction
↓
Credit

By industry classification, these are completely different markets.
Economically, a significant share of these orders originates from one decision:

BUILD AI INFRASTRUCTURE

And this is already visible in actual order books.
Reuters reports that HD Hyundai Electric has seen its backlog rise to approximately $8.5 billion, an increase of around 23% in only six months.
At China’s Jinpan, new data-center-related orders increased by more than four times in the first half of the year, while the corresponding backlog nearly tripled.
Source:
Reuters — Not just Nvidia: these power and cooling firms are riding the trillion-dollar data centre boom
Dell raised its annual revenue forecast to approximately:
$192 billion
on the back of demand for AI servers.
Quarterly revenue increased by 58%, while backlog reached approximately:
$95 billion
Source:
Reuters — Dell shares gain after strong AI server demand boosts annual forecast
Vertiv is expanding into microgrid and power infrastructure precisely because conventional electricity grids cannot keep pace with the growth of AI data centers.
Source:
Reuters — Vertiv strikes deal for Utility Innovation Group
These are not separate growth stories.
They are one story.
4. The “Broad Market” Creates a False Sense of Security
Suppose:
Microsoft builds a data center.
Nvidia sells the GPUs.
Dell supplies the servers.
Vertiv supplies cooling and power equipment.
A turbine manufacturer receives an order for new generation capacity.
A transformer manufacturer receives an order for a substation.
A utility expands the grid.
A construction company builds the facility.
A bank finances the project.
On the screen, we see:

Technology ↑
Industrials ↑
Utilities ↑
Construction ↑
Financials ↑

This is exactly how the impression of healthy, broad-based growth is created.
But the final economic impulse is the same in every case:

AI INFRASTRUCTURE SPENDING

Different industries do not necessarily mean different sources of demand.
If five sectors depend on one customer, there is no genuine economic diversification.
5. This Cycle Is Now Entering the Credit System
The first stage of the buildout was financed mainly from the enormous cash flows of Microsoft, Alphabet, Amazon and Meta.
That is no longer sufficient.
J.P. Morgan Asset Management explicitly notes that the AI capex cycle is shifting from balance-sheet financing toward capital markets.
According to its estimates, total data-center buildout through 2030 could require approximately:
$5 trillion
of which approximately:
$2 trillion
could be financed through investment-grade credit markets.
The growth in hyperscaler bond issuance is particularly revealing:
2024: $17 billion
2025: $109 billion
First six months of 2026: $194 billion
Source:
J.P. Morgan Asset Management — Can Credit Markets Absorb the AI Buildout?
This represents a fundamental change in the structure of risk.
The AI investment cycle is moving beyond Big Tech equity and spreading through:

Tech Equity
→
Corporate Debt
→
Project Finance
→
Private Credit
→
Infrastructure

The deeper this process moves into the credit system, the greater the potential consequences of a reversal.
6. BIS Is Already Showing the Mechanism of Overinvestment
The Bank for International Settlements, in its paper The AI Investment Race, models the current situation as an investment race.
Competition forces companies to build capacity earlier and on a larger scale than would be optimal for the system as a whole.
For an individual company, this is rational.
No company wants to lose the future AI market because it failed to secure enough computing capacity.
But if everyone behaves the same way, private investment exceeds the economically efficient level.
In the BIS baseline calibration, investment reaches approximately:
1.5 times the efficient level
Under less elastic demand assumptions, the figure can approach approximately:
3 times the efficient level
BIS also identifies three systemic risk amplifiers:

increasing debt financing;
the specialized nature of AI hardware;
financial interconnections between participants in the investment chain.

Source:
Bank for International Settlements — The AI Investment Race, Working Paper 1367
The problem therefore extends far beyond the valuation of individual technology stocks.
7. Dot-Com Showed How Such a System Breaks
The internet was not fiction.
It genuinely transformed the economy.
But the investment cycle of the late 1990s built companies and infrastructure faster than the economy could generate the cash flow required to justify them.
The chain looked like this:

VC
→
Internet Companies
→
IPO
→
Telecom
→
Fiber
→
Networking

As long as capital markets continued supplying money, the system expanded.
When IPO and venture-capital markets stopped financing loss-making companies, the destruction moved backward through the chain.
Companies cut spending.
Orders disappeared.
Telecom capex collapsed.
Equipment manufacturers were left with excess capacity.
The critical point in the dot-com crisis was therefore not simply the fall of the Nasdaq.
The decisive event was the breakdown of the mechanism financing the investment machine.
8. Today’s Chain Is Longer and More Dangerous
The current structure is:

Hyperscaler Cash Flow / Debt
↓
Semiconductors
↓
Servers
↓
Data Centers
↓
Power
↓
Grid
↓
Industrials
↓
Construction
↓
Credit

As long as the top of the chain continues investing, everything below it reports strong results.
Backlog rises.
Revenue rises.
EPS rises.
Share prices rise.
Production capacity expands.
Credit continues flowing.
But this system works in reverse as well.
9. How the Reversal Begins
The real turning point has nothing to do with an ordinary correction in Nvidia or the Nasdaq.
It begins when it becomes clear that:

AI Revenue < Required Return on AI Capital

After that:

hyperscalers reduce capex guidance
↓
new data-center projects are postponed
↓
GPU and server orders slow
↓
backlog growth at cooling and electrical-equipment manufacturers stops
↓
utilities revise future load forecasts
↓
new generation and grid projects are postponed
↓
the economics of already-financed projects deteriorate
↓
credit spreads widen
↓
banks and private-credit providers tighten financing
↓
new projects become uneconomic
↓
INVESTMENT BUST

At that point, it is not simply the AI sector that falls.
The shock hits simultaneously:

semiconductors;
cloud;
servers;
data centers;
electrical equipment;
utilities;
generation;
construction;
industrials;
credit markets.

This is why dependence on a single source of capital expenditure represents a systemic risk.
10. Concentration Is Already Reaching Extreme Forms
SB Energy demonstrates how far this dependence can go.
The company is building infrastructure for AI data centers.
Its actual data-center revenue remains minimal, yet its stated backlog is approximately:
$430–439 billion
At the same time, its entire contracted capacity — around 8.8 GW — is tied to just two counterparties:
SoftBank and OpenAI.
Reuters Breakingviews notes that only around 10% of this backlog is expected to be realized during the first six years.
Source:
Reuters Breakingviews — SB Energy’s giant IPO is getting ahead of itself
This is almost a perfect illustration of what is happening across the wider system on a smaller scale:
enormous future cash flows are being assumed today because the market expects the gigantic AI investment cycle to continue.
11. Physical Bottlenecks Are Not the Main Problem
Today the discussion focuses on shortages of:

GPUs;
electricity;
transformers;
data centers;
cooling;
grid connections.

But these are intermediate constraints.
The real question lies at the end of the chain:

Who will generate the final cash flow that pays for all of this?

Several trillion dollars of AI infrastructure must ultimately be paid for by future additional economic productivity.
Powerful models and huge numbers of GPUs are not enough.
The system also requires:

skilled workers;
quality education;
organizational restructuring;
new production processes;
data;
electricity;
infrastructure;
management competence;
the ability of the real economy to absorb the technology.

This is where the AI investment cycle collides with the broader system.
Technological capability is growing extremely quickly.
But human capital, infrastructure, demographics, public finances and institutional capacity are changing much more slowly.
The result is a widening gap:

AI Capability >> Economic Absorption Capacity

Conclusion
The question ultimately comes down to one thing:
Can the economy generate cash flow fast enough to justify the amount of infrastructure being built today?
My answer is no.
Not because there is one catastrophic problem.
The problem is their combination.
Demographic pressure, shortages of skilled workers, deterioration in parts of human capital, energy constraints, weak infrastructure, debt burdens, the requirement for enormous capital expenditure and the movement of the AI buildout toward credit financing are not individually enough to create a catastrophe.
But all of these vectors are moving toward each other.
At the same time, the AI investment machine requires an ever-larger future cash flow to justify the capital already being deployed.
The more infrastructure is built today, the higher the minimum future economic return required to preserve the entire structure.
It is the combination of these factors that leaves very little margin for error.
If productivity growth and AI monetization fail to keep pace with the accumulation of capital, it will not be one technology sector that stops.
It will be the investment chain that now connects a significant share of technological, industrial, energy and financial growth.
Main Sources
Bank for International Settlements — The AI Investment Race, Working Paper 1367, July 2026
https://www.bis.org/publications/working-paper-1367-ai-investment-race
BlackRock — Getting Serious in Summer Markets, 2026
https://www.blackrock.com/us/financial-professionals/insights/summer-markets-ai-investing
J.P. Morgan — Is It All One Big AI Trade?, 2026
https://www.jpmorgan.com/insights/markets-and-economy/top-market-takeaways/tmt-is-it-all-one-big-ai-trade
J.P. Morgan Asset Management — Can Credit Markets Absorb the AI Buildout?, 2026
https://am.jpmorgan.com/ca/en/asset-management/institutional/insights/market-insights/market-updates/on-the-minds-of-investors/can-credit-markets-absorb-the-ai-buildout/
Reuters — AI data-center infrastructure and supplier backlogs, September 2026
https://www.reuters.com/business/energy/not-just-nvidia-these-power-cooling-firms-are-riding-trillion-dollar-data-centre-2026-09-01/

NotFuzzyWarm

Legendary

Activity: 4494
Merit: 3588

Evil beware: We have waffles!

Re: How accelerating technology, weakening competence formation, and the retirement

September 06, 2026, 06:29:25 PM

Merited by PowerGlove (4)

While spot-on this is all common knowledge to techs/field engineers who work "in the trenches".
Even before AI caught on this has been a problem for many years. While tech schools and universities do a decent job teaching basics it is only after hands-on working with technology can one even begin to really understand the 'why' of how things work.

With most companies compartalizing their design groups into rather specific functions it is very hard to find folks who understands the Big Picture of technology.

Zuzma (OP)

Sr. Member

Activity: 552
Merit: 285

HODL

Re: How accelerating technology, weakening competence formation, and the retirement

September 07, 2026, 10:10:38 AM

AI Infrastructure: The Cracks Are Already Visible
In my previous posts I focused on the main economic question behind the AI boom:
Who, in the real economy, is going to generate enough cash flow to justify trillions of dollars of infrastructure investment?
Now the question becomes harder:
How much of those “trillions” actually exists as real investment?
Because today the market often mixes completely different stages into one huge number:

Code:

announcement
↓
power request
↓
customer commitment
↓
financing committed
↓
financial close
↓
construction
↓
energized capacity
↓
revenue

These are not the same thing.
1. The first problem is the inflated pipeline
Cushman & Wakefield currently estimates the Americas data-center market roughly as follows:
43.4 GW operational
25.3 GW under construction
191.3 GW development pipeline
So the amount actually under construction is only a fraction of the development pipeline.
That does not mean the remaining projects will all be cancelled.
But it does mean that development pipeline cannot be treated as committed future capacity.
And above that sits an even more absurd number.
US utilities have received requests for more than:
700 GW of data-center power demand.
That is far too large to be interpreted as near-term, credible investment demand.
Utilities started checking what was behind those requests.
They began asking developers to:

pay real deposits.

And part of the demand disappeared.
Reuters already uses the term:
“ghost demand.”
This is where things become interesting.
As long as requesting capacity was cheap, demand looked enormous.
Once real money was required, part of that demand evaporated.
That is not a theory anymore.
That is the financial system beginning to filter the pipeline.
Source:
Reuters — US reckoning over data-center “ghost demand”
Cushman & Wakefield — Global Data Center Market Comparison 2026
2. An announcement is not an investment
For years the market has become accustomed to headlines like:
“Company X plans a $10 billion AI data center.”
“A new 2 GW AI campus will be built.”
“This region will receive $20 billion of investment.”
But until there is:

financial close;

it is not yet an investment.
It is an intention.
This is why headline CAPEX systematically overstates future real CAPEX.
3. Kansas showed how absurd the gap can become
The Flint Hills Digital Campus project in Emporia, Kansas involved roughly 1,000 acres and potentially billions of dollars of future investment.
There was political support.
There was land annexation.
There were public statements.
But the man publicly presented as the developer later said under oath:

Quote

“I have no interest at all in that project.”

This is almost a perfect example of how far the public narrative can move ahead of the actual financial structure.
Billions are already being discussed.
But the questions:
Who exactly is building it?
Who is financing it?
Who is the customer?
can still remain unclear.
Source:
Kansas Reflector — The developer of a Kansas data center swore under oath he wasn’t involved
4. Abilene matters more than Kansas
Kansas can be dismissed as a questionable developer story.
Abilene cannot.
There we are talking about:
OpenAI + Oracle + Crusoe
and a real, existing AI data-center campus.
Yet further expansion was abandoned after prolonged financing negotiations and changing OpenAI requirements.
That is much more important.
Because it shows that:
the need for more compute does not automatically mean the next construction phase receives capital.
Source:
Bloomberg — Oracle and OpenAI End Plans to Expand Flagship Data Center
5. Crusoe in Wyoming is another warning about “pipeline”
Project Jade was planned at:
1.8 GW
Development was paused:

Quote

“At the request of our customer.”

That is all it takes.
A 1.8 GW project can disappear from the active development pipeline because one customer changes its plans.
Therefore:
development pipeline ≠ committed demand
The larger the industry starts talking in tens or hundreds of gigawatts, the more important one question becomes:
Who has actually signed for it?
Source:
Utility Dive — Data-center cancellations and delays
6. Cancellations are already increasing
According to Baird data cited by Utility Dive:
2023 — 2 cancelled projects
2024 — 6
2025 — 25
That is a very sharp increase.
The reasons are mixed:

project economics.

It would be wrong to claim that all of those projects failed because financing disappeared.
But for industrial demand, the result is the same.
A cancelled project does not become:

So headline pipeline systematically overstates future industrial CAPEX.
Source:
Utility Dive — What’s stalling data-center projects?
7. Even Blackstone can stop a project
QTS/Blackstone terminated the Digital Gateway project in Virginia after years of preparation and regulatory work.
This destroys another convenient assumption:
If the sponsor is large enough, the project will definitely happen.
No.
Even an approved project backed by institutional capital can still become zero.
Source:
Reuters — Blackstone’s QTS terminates Digital Gateway data-center project
8. But the capital market is not closed
This distinction matters.
High-quality projects still get money.
ByteDance recently secured:
$29.6 billion
and the loan was oversubscribed.
So it would be false to say:
“Investors no longer want to finance AI.”
That is not what is happening.
Something more interesting is happening.
Capital is beginning to discriminate between projects.
A project with:

credible collateral;

can still attract billions.
A project with:

may no longer get funded.
This is the transition from:
capital abundance
to
capital discrimination.
Source:
Reuters — ByteDance secures $29.6 billion loan for AI push
9. The AI boom is becoming increasingly dependent on debt
S&P expects AI infrastructure investment to exceed:
$1.3 trillion by 2027.
At the same time, the largest hyperscalers are expected to produce negative aggregate free operating cash flow during 2026–2027.
That means further expansion increasingly depends on outside capital.
Not only profits from Big Tech.
But:

The more the system depends on external capital, the more dangerous a rise in the cost of capital becomes.
Source:
S&P Global Ratings — AI Infrastructure Investment To Exceed $1.3 Trillion By 2027
10. And the cost of capital is rising at exactly the wrong moment
The US 10-year Treasury yield is around:
4.8%
Oil has moved close to:
$100 per barrel
Inflation risk is rising again.
The Federal Reserve may have to keep rates higher for longer.
For an investment committee this is simple mathematics.
If US government debt gives nearly 5%, then a data-center project carrying:

refinancing risk;

cannot comfortably offer only 7–8%.
Investors will demand more.
And once the required return rises, part of the pipeline no longer works economically.
Sources:
Reuters — Global bond yields and US Treasury pressure
Reuters — Oil approaches $100 amid supply risk
11. The most dangerous scenario is not necessarily a recession
A normal recession can actually solve part of the financing problem:

Code:

economy slows
↓
inflation falls
↓
Fed cuts rates
↓
capital becomes cheaper

A much worse combination is:
weak growth

expensive energy

persistent inflation

Treasuries around 5%

rates stay high
Then two things happen at the same time:
end-user demand becomes more cautious
and
infrastructure financing becomes more expensive.
That is a genuine stress test for the AI investment cycle.
12. Why record backlogs prove very little
Today Eaton, Schneider, Vertiv and other suppliers still report extremely strong demand.
That is not surprising.
They are executing projects that were financed earlier.
Capital-intensive industry has long lags.
The sequence can look like this:

Code:

financial close falls today
↓
new notices to proceed fall months later
↓
construction starts fall later
↓
equipment orders weaken after that
↓
supplier backlog weakens last

So waiting until a transformer manufacturer announces:
“data-center demand is normalizing”
means you are already late.
13. Follow the money, not the factory
The earliest indicators are:

That is where the investment cycle will turn first.
Not in GDP.
Not in unemployment.
Not in PMI.
Not in Eaton’s quarterly backlog.
14. The year 2000 is a useful historical comparison
Before the dot-com crash, the US economy did not look sick.
In February 2000 the Federal Reserve wrote:

Quote

“The U.S. economy retained considerable strength.”

GDP had been growing above 4%.
High-tech investment was enormous.
Capital was widely available.
The internet was a genuine technological revolution.
And the bubble still collapsed.
Why?
Because investors were not wrong about the technology.
They were wrong about:
the speed of monetization
and
how much capital could be rationally deployed in a short period of time.
That distinction is critical.
The internet won.
Investors still lost enormous amounts of money.
Source:
Federal Reserve — Monetary Policy Report, February 2000
15. The same thing can happen with AI
AI may:

become a foundational technology.

And today's infrastructure investments can still be overpriced.
Suppose that within several years:

Then part of today’s infrastructure may become excessive or simply too expensive.
That is excellent for users.
It is not necessarily excellent for the investor who financed an expensive data center at the top of the cycle.
16. Faster AI progress can actually hurt AI infrastructure investors
The usual assumption is:
better AI = more data centers.
But the relationship does not have to be linear.
If a new model can perform the same task with:
10× less compute
and
5× less energy
the economy can consume far more AI while physical infrastructure demand grows much more slowly.
That creates a paradox:
faster technological progress can be bullish for AI and bearish for some AI infrastructure assets at the same time.
17. So $5–7 trillion is not “real demand”
When analysts write:
“AI/data-center investment may reach $5–7 trillion by 2030”
that does not mean:
$7 trillion has already been committed.
It means:
under a certain set of assumptions, that amount of capital may be required.
Between “may be required” and “money has been invested” sits the real funnel:

Code:

announcement
↓
customer
↓
power
↓
financing
↓
financial close
↓
construction
↓
energization
↓
cash flow

That gap is exactly where the first cracks are now becoming visible.
What is already visible today
We already have:

[]inflated power requests;
[]ghost demand;
[]rapidly rising project cancellations;
[]customer-requested pauses;
[]abandoned expansions;
[]growing dependence on debt;
[]more expensive capital;
[]financial filtering of weaker projects.
[/list]
These are cracks.
Not a collapse.
But cracks.
Conclusion
The main question today is no longer:
Will AI become a major technological revolution?
It probably will.
The real investment question is:
How much of the announced trillions will actually make it all the way to cash flow?
Because the market is already beginning to separate:
PowerPoint
from
money
money from
construction
and construction from
a revenue-producing asset.
That is where the first fractures in the AI infrastructure story are becoming visible.
The lesson from 2000 is simple:
the technology can be real,
the economy can still look strong,
factories can still be working through old orders,
while the marginal dollar of new capital has already stopped flowing into the next project.
If the AI infrastructure cycle turns, this is where it will show up first:
in financing, financial close, and the cost of capital — long before it appears in GDP, unemployment, or industrial backlog.

Zuzma (OP)

Sr. Member

Activity: 552
Merit: 285

HODL

⇾ Re: How accelerating technology, weakening competence formation, and the retirement

September 08, 2026, 07:24:45 AM

Everything I described is also important in the context of what phase of the market we are currently in. And the market right now is very fragile and unstable.

One of the risks worth watching separately is the US Treasury market.

China, which used to be one of the largest and most stable buyers of US government debt, has been reducing its position for many years. At the same time, the role of hedge funds has increased sharply.

According to the Federal Reserve, by September 2025 large hedge funds held around 8.5% of privately held Treasuries.

Their gross Treasury exposure reached approximately:

$4 trillion

including:

$2.4 trillion long
$1.6 trillion short

At the same time, repo borrowing was around $3 trillion.

Federal Reserve — Decomposing Hedge Funds' U.S. Treasury Exposures

But the important part is not only the size of these positions, but also their structure.

Out of $2.4 trillion in long exposure, approximately:

$830 billion was in cash-futures basis trades;
$305 billion was in swap-spread arbitrage.

In other words, roughly half of hedge funds’ long Treasury positions are concentrated in just two strategies that the Federal Reserve explicitly describes as highly leveraged arbitrage strategies.

Federal Reserve — Hedge Fund Treasury Exposures

The BIS has also pointed out that hedge funds have become important intermediaries in sovereign bond markets and actively finance their positions through short-term repo.

Zero-haircut repo is common among hedge funds, particularly among the largest funds.

This allows very high leverage to be used.

BIS — Unpacking repo haircuts and their implications for leverage

The numbers in Treasury futures are even more striking.

The BIS has estimated margin leverage in Treasury futures at roughly:

up to 175x in 5-year futures;
up to 120x in 10-year futures.

After margin requirements increased, leverage declined to approximately:

70x in 5-year futures;
50x in 10-year futures.

Important: this does not mean that an entire hedge fund operates with 50–70x balance-sheet leverage.

It means that one futures leg of the strategy can have an enormous notional exposure relative to the amount of margin actually posted.

BIS — Margin leverage and vulnerabilities in US Treasury futures

And this is where the real risk begins.

A basis trade is usually not a simple directional bet on Treasury prices.

A typical structure looks like this:

long cash Treasury + short Treasury futures

So the main problem is not necessarily that a bond falls by 2%.

The real risk is funding and liquidity.

The mechanism looks like this:

Quote

rate shock
→ volatility rises
→ margins and repo haircuts increase
→ more collateral is required
→ hedge funds reduce positions
→ Treasuries are sold
→ prices fall further
→ yields rise further

The BIS explicitly notes that higher initial margins mechanically force leveraged investors either to post more cash or to close positions.

Similar disorderly deleveraging amplified stress in fixed-income markets in September 2019 and March 2020.

BIS — Margin leverage and vulnerabilities in US Treasury futures

So this is not just a theoretical risk.

At the same time, this does not mean that hedge funds will inevitably crash the market.

Under normal conditions, they perform a useful function:

they provide liquidity;
they eliminate pricing discrepancies between cash Treasuries, futures and swaps;
they improve market efficiency.

The problem is elsewhere.

Quote

In a calm market, leverage improves efficiency.

In a stressed market, the same structure makes liquidity procyclical — it disappears exactly when it is needed most.

And the scale is no longer marginal.

According to the Federal Reserve, the basis trade alone had reached approximately $830 billion by September 2025.

There is also a concentration issue: a large share of total hedge-fund Treasury exposure is concentrated among the largest funds.

Federal Reserve — Decomposing Hedge Funds' U.S. Treasury Exposures

This becomes even more important against the background of China continuing to reduce its Treasury holdings after decades of being a large and relatively stable reserve buyer.

Reuters recently summarized this structural change very clearly: the market has gradually moved away from dependence on China and other official reserve holders toward a larger role for leveraged hedge funds and other price-sensitive intermediaries.

Reuters — Hedge funds pose greater threat to US Treasuries than China ever did

So there is a structural shift taking place:

Quote

stable reserve buyer ↓

leveraged financial intermediaries ↑

And this connects directly to what I wrote about earlier.

The AI cycle requires an enormous amount of capital:

Quote

data centers
→ energy
→ power grids
→ construction
→ equipment
→ debt

And the base cost of all this capital starts with long-term Treasury yields.

If the Treasury market becomes unstable and long-term yields rise:

Quote

Treasury yields ↑
→ corporate borrowing costs ↑
→ project hurdle rates ↑
→ part of AI CapEx stops being economically viable

So AI may run into constraints not only in electricity, infrastructure or skilled labor.

It may also run into the cost of capital.

That is why the main risk is not that “hedge funds will inevitably crash everything.”

The risk is that a meaningful share of liquidity in the world’s most important sovereign bond market is now provided by highly leveraged participants relying on short-term funding, and in a stressed environment they can very quickly turn from buyers into forced sellers.

This is why the stability of long-term Treasuries is no longer just a bond-market issue.

It is one of the lower layers supporting the entire AI CapEx cycle and the current valuation of risk assets.

Sources:

Federal Reserve

BIS — Repo haircuts and leverage

BIS — Treasury futures leverage

Reuters

Zuzma (OP)

Sr. Member

Activity: 552
Merit: 285

HODL

Re: How accelerating technology, weakening competence formation, and the retirement

September 08, 2026, 07:27:35 AM

Merited by NotFuzzyWarm (1)

Quote from: NotFuzzyWarm on September 06, 2026, 06:29:25 PM

I agree. I would only add that AI did not create this problem, it is accelerating it and making it more visible. The more work is split into narrow specialized functions, the more valuable people become who understand the whole system and can identify whether the real problem is in the code, the hardware, the process, or the original task definition.

Zuzma (OP)

Sr. Member

Activity: 552
Merit: 285

HODL

Re: How accelerating technology, weakening competence formation, and the retirement

September 10, 2026, 04:56:56 PM

All Roads Lead to One Point

I wanted to continue this line of reasoning, but I think this is a good place to close the cycle.

Today, the head of the BIS effectively confirmed several of the points I had written about earlier: AI investment is growing faster than cash flows, more financing is moving into debt and private credit, circular financing is emerging between chipmakers, hyperscalers and AI companies, and if actual returns fail to meet expectations, the investment boom could turn into a bust with consequences for the broader economy.

The BIS estimates that AI capex by the five largest technology companies could exceed $1 trillion in 2025-2026.

For me, this is not so much confirmation that “AI is a bubble” as it is confirmation of something more important:

AI has become the point where many of the current contradictions of the US economy converge.

On one side, there is the US government.

Federal debt is already around 100% of GDP, while the budget deficit is close to 6% of GDP.

For the government, a combination of moderate inflation and cheaper financing is relatively convenient because it supports nominal GDP growth, tax revenues and reduces the real burden of previously issued debt.

On the other side, there is the Federal Reserve.

Its official inflation target is:

2%

But producer inflation has already accelerated to:

PPI: 5.4% YoY

Oil is around:

$100 per barrel

and the labour market is still relatively resilient.

Because of this, the market is increasingly pricing not rate cuts, but the possibility that the Fed will have to keep rates high or even raise them.

After the latest PPI data:

2-year Treasury yield: ~4.53%

10-year Treasury yield: ~4.8-4.9%

This creates a direct conflict.

If the Fed raises rates:

cost of capital rises → investment slows → AI capex becomes less attractive → housing and equities come under pressure

If the Fed cuts rates:

inflation expectations may rise → long-term Treasury yields may rise anyway

If the Fed does nothing:

the bond market can continue tightening financial conditions by itself

So the problem is no longer simply whether the Fed should hike, cut or hold.

The problem is that all three choices have become increasingly uncomfortable.

And in almost every scenario, the most sensitive point is the AI investment cycle.

Why?

Because AI requires all of the following at the same time:

cheap capital
huge amounts of electricity
physical infrastructure
semiconductors and equipment
qualified engineers and specialists
rapid growth in final demand
very high future profits

All of this must scale simultaneously.

That is the real problem.

I am not blaming anyone here.

The cause is deeper than the policy mistakes of one institution or the decisions of one company.

It is the nature of competition itself.

If one company slows AI investment while its competitor continues, it risks losing market share.

If one country slows down while another accelerates, it risks losing technological leadership.

If the Fed protects growth too early, it risks inflation.

If it fights inflation too aggressively, it risks triggering an investment downturn.

Every participant behaves rationally from the perspective of self-preservation.

But rational behaviour at the individual level can produce an irrational result at the system level.

Companies cannot comfortably stop the AI race.

Politicians cannot comfortably allow markets to collapse.

The Fed cannot comfortably raise rates, cut rates, or even leave them unchanged.

The government cannot comfortably cut spending.

Investors do not want to leave a rising market too early.

Everyone continues moving because stopping first may be more dangerous than continuing.

This is what I now see as the main risk.

The problem is not inflation alone.

It is not federal debt alone.

It is not AI alone.

It is not war.

It is not extreme equity valuations.

It is not the shortage of human capital.

The problem is that all of these constraints have converged at the same time and have started reinforcing each other.

Technology is developing faster than the physical economy can support it.

The physical economy is changing faster than human capital can adapt.

Investment is growing faster than the demonstrated return on invested capital.

And the financial system needs more capital precisely at the moment when capital is becoming more expensive.

So if a major crisis does come, looking for one guilty party will probably be meaningless.

All roads now lead to the same point: AI is not the cause of every problem. It is the place where many of the problems meet.

And if the system breaks, it will not be because there was one unsolvable problem.

It will be because:

the solution to each individual problem made the other problems worse, until the system no longer had a cheap way out.

And given the current combination of circumstances, we will probably see that repricing very soon. I believe the next year will be decisive in determining whether my concerns and hypothesis are confirmed by what actually happens in the economy and financial markets.

Pages: [1]

Bitcoin Forum > Other > Serious discussion > How accelerating technology, weakening competence formation, and the retirement

« previous topic next topic »