AI Investment Paradox: If Machines Do More Work, Who Buys What They Produce?
There is a question about the AI boom that I don't think investors are asking often enough.
We keep hearing that artificial intelligence will make companies more productive. Machines will do more work, faster and at lower cost. Businesses will need fewer people for some tasks. Margins should improve. Output should rise.
Fine.
But who buys all that additional output?
That question sounds almost too simple. Yet it goes to the heart of the investment case for AI.
Productivity is not the same thing as demand
Suppose AI allows a company to produce the same output with fewer employees.
From the company's perspective, that can be wonderful. Costs fall. Productivity rises. Perhaps profits rise too.
Now repeat that across thousands of companies.
At some point we have to ask what happens to the income of the people whose work has been displaced, reduced or repriced.
Because workers are not merely costs on corporate income statements.
They are also customers.
If productivity rises substantially while purchasing power becomes increasingly concentrated, we could eventually create an unusual economic problem: enormous productive capacity without equivalent growth in broad final demand.
I am not saying this will happen.
In fact, one of the conclusions of my research is that there is not yet enough evidence to say that it is happening.
But that is precisely why I think investors should start measuring it now.
The second paradox: AI can succeed while AI investors lose money
There is another mistake I think the market can make.
A transformative technology does not automatically produce extraordinary returns for the companies that invent or build it.
The internet transformed the world. But much of the economic value eventually accrued to businesses built on top of the internet rather than simply to the companies supplying the underlying infrastructure.
AI could follow a similar path.
Today, extraordinary amounts of capital are flowing into chips, data centres, networking equipment, electricity generation, cloud infrastructure and AI models.
Somebody has to earn a return on that capital.
The important investment question therefore isn't simply:
Will AI work?
I think the more useful question is:
If AI works extremely well, where will the resulting economic surplus eventually settle?
Those are very different questions.
Follow the surplus
This is the framework I developed in my new research report, AI Investment Paradox: If Machines Do More Work, Who Buys What They Produce?
Rather than treating "AI" as one investment theme, I follow the economic chain:
Power → data centres → semiconductors → networking → cloud → models → software → businesses → consumers
At every stage, I ask two simple questions:
Who pays whom?
And:
Who gets to keep the incremental profit after competition and financing costs?
That second question matters enormously.
A company can experience spectacular revenue growth while simultaneously entering a business where competition eventually destroys excess returns.
Conversely, a company that did not invent the technology may become one of its greatest beneficiaries simply because AI dramatically lowers its operating costs.
The biggest AI winners ten years from now therefore may not necessarily be today's most obvious AI companies.
The capital-spending asymmetry is already interesting
One finding in the report deserves particular attention.
The first-wave infrastructure suppliers and the hyperscalers currently occupy very different economic positions.
Companies selling critical AI infrastructure can generate substantial cash while requiring relatively modest property and equipment spending of their own.
Meanwhile, hyperscalers are spending tens — and in some cases more than one hundred — billion dollars annually building the infrastructure required to provide AI services.
That doesn't mean the hyperscalers are making a mistake.
They may ultimately earn enormous returns.
But the return on that incremental capital has not yet been demonstrated over a complete economic cycle.
That distinction matters.
Revenue growth tells us that AI demand exists.
Capital expenditure tells us how much companies are paying to serve that demand.
Neither, by itself, tells us what shareholders will ultimately earn.
Two ledgers I'm going to keep watching
The report therefore introduces two simple monitoring frameworks.
The first is the Demand Ledger.
It asks whether AI productivity gains are eventually reaching households and end users through wages, lower prices, new employment, transfers, investment income or entirely new categories of work.
Right now, my assessment is deliberately:
Insufficient evidence.
It is too early to conclude that AI is producing a structural aggregate-demand problem.
That conclusion could change as the evidence develops. If employment remains resilient, real wages rise, consumption broadens and AI creates new categories of work and income, the concern weakens.
If the opposite occurs — productivity accelerates while labour income and broad purchasing power deteriorate — then the demand question becomes much more important for investors.
The second framework is the Scarcity Ledger.
If intelligence itself becomes abundant and cheap — particularly as open-weight models improve — then economic rents may migrate toward things that remain scarce.
Electricity.
Grid connections.
Land.
Cooling.
Network capacity.
Proprietary data.
Distribution.
Customer relationships.
Regulatory permission.
Trust.
Physical-world execution.
This is where I think the AI investment discussion becomes particularly interesting.
If the cost of intelligence continues falling, simply possessing an AI model may eventually become less economically distinctive. The valuable part of the chain could migrate elsewhere.
That could have profound consequences for portfolio construction because it means the eventual AI winners may appear in industries that investors currently do not think of as "AI companies."
This is why I'm looking beyond technology
The objective of this research is not to decide whether NVIDIA — or any particular AI company — is good or bad.
It is to develop a framework for deciding where the next dollar of investment should go as AI moves through the economy.
Industrials may benefit.
Utilities may benefit.
Healthcare may benefit.
Financial companies may capture productivity improvements.
Some consumer businesses could become significantly more profitable.
Others could suffer if labour displacement eventually weakens their customers' purchasing power.
And entirely new businesses may emerge that we cannot identify today.
That is why I don't think we should approach AI simply by asking which semiconductor company, hyperscaler or model provider will win.
The more interesting question may eventually be:
Which existing industries become structurally more profitable because AI exists?
And equally important:
Which industries appear to benefit from AI but actually find the savings competed away?
In several sectors, the correct conclusion today is simply:
We don't know yet.
I think investors should be comfortable saying that.
False precision is considerably more dangerous than uncertainty acknowledged early.
The question I want this research to answer
Over the coming months and years, I intend to keep returning to one question:
Where is AI creating genuine incremental economic surplus — and who is actually keeping it?
Not who has the most impressive model.
Not who announces the largest data centre.
Not who spends the most money on AI.
Not who uses the word "AI" most frequently on an earnings call.
But:
Who ultimately earns an attractive return on the additional capital committed?
That distinction becomes increasingly important as the investment cycle matures.
During the early stages of a technological revolution, almost everybody can appear to win. Suppliers receive orders. Infrastructure gets built. Capacity expands. Revenue grows. Investors extrapolate.
The harder test comes later.
Who has pricing power?
Who has become commoditised?
Who generated genuine incremental cash flow?
Who spent enormous amounts of capital merely to defend an existing business?
Who benefited from falling AI costs?
And who captured the productivity gains without having to finance the infrastructure that created them?
Those are the questions that ultimately matter to a portfolio.
And underneath all of them sits the original economic question that started this research:
If machines eventually do substantially more of the work, who buys what they produce?
I don't think we know the answer yet.
That is exactly why I think we should start watching now.
I have published the complete research report on Zenodo. This is not intended to be the final word on the AI investment question. It is a framework that I intend to revisit as the evidence develops — particularly employment, household demand, AI capital spending, cash generation and where the economic surplus is actually accumulating.
The advantage of publishing the research now is that the thesis is on the record.
As the data change, we can see which parts held up, which did not, and change our investment conclusions accordingly.
Read the full research:
Larry Lim Kheng Cheong
19 August 2026
This article is for research and informational purposes and does not constitute investment advice.