When Intelligence Becomes Capital

Something unusual happened this week.

OpenAI⁠ announced that an internal AI system had produced a solution to the Navier–Stokes existence and smoothness problem, one of mathematics’ celebrated Millennium Prize Problems.

The Economist rightly framed the development as raising “profound questions.”

I would go further.

The most profound question may not be whether AI can solve mathematics. It is what happens to capitalism when intelligence itself becomes scalable infrastructure.

For most of economic history, intelligence has been stubbornly scarce.

A brilliant mathematician cannot be copied 10,000 times. A great scientist cannot simultaneously pursue 10,000 hypotheses. A Nobel-level researcher still has 24 hours in a day.

Machines changed the productivity of human hands.

Computers changed the productivity of calculation.

The internet changed the cost of distributing information.

AI may now be beginning to change something much more fundamental:

the marginal cost of producing intellectual work.

And that could have consequences extending far beyond mathematics.

First, what actually happened?

The Navier–Stokes equations describe how fluids move. They underpin models used in fields ranging from aerodynamics and weather forecasting to blood flow.

The mathematical question that remained unresolved for roughly 90 years was deceptively simple: if a three-dimensional fluid starts in a perfectly smooth state, must its mathematical evolution remain smooth forever?

Or can the equations eventually produce a singularity — essentially a point at which the mathematical description breaks down?

OpenAI says its system has demonstrated the latter.

Its proposed solution constructs a fluid beginning smoothly at rest which, under a smooth external force, develops a singularity in finite time while maintaining finite energy. The company published both an analytical proof and a formalization using Lean, a proof-assistant system capable of mechanically checking mathematical reasoning.

The result still has to survive the scrutiny of the mathematical community. A corporate announcement, however impressive, is not the same thing as decades of mathematical consensus.

That distinction matters.

But the method used to reach the result may ultimately be almost as consequential as the proof itself.

10,000 mathematicians who never sleep

OpenAI did not simply ask a chatbot:

“Please solve Navier–Stokes.”

According to the company, it deployed roughly 10,000 concurrent AI agents.

They were divided into groups.

They explored different approaches.

They communicated.

They ran code.

They consulted cached internet resources.

Useful discoveries were circulated among other groups.

The agents collectively sent approximately 2.7 million messages and generated about 130 billion output tokens during the Navier–Stokes effort.

And they reached the proposed solution after approximately 88 hours.

Another 17 hours were used for Lean formalization and verification.

Stop here for a moment.

Because this is where the story stops being primarily about mathematics.

Imagine trying to assemble 10,000 elite human mathematicians.

Recruit them.

Pay them.

Coordinate them.

Allow them to work simultaneously.

Have them exchange millions of messages.

Let unsuccessful teams disappear while promising approaches receive additional resources.

Run the entire operation continuously, day and night.

Then complete the experiment in less than four days.

Economically, that is impossible with humans.

With AI, it has apparently become an engineering problem.

That difference is enormous.

We may have invented intellectual mass production

The Industrial Revolution did not make humans stronger.

It made mechanical power scalable.

A steam engine could perform work previously requiring enormous quantities of human or animal labour.

The computer did something similar with arithmetic.

The internet did it with information distribution.

AI agents could potentially do it with reasoning.

That does not mean machines have suddenly acquired human consciousness, wisdom or scientific intuition.

They do not need to.

Factories did not need muscles to resemble human muscles.

What mattered was output.

The economically interesting question therefore isn’t:

“Does the AI think like a mathematician?”

It is:

“How much mathematical output can a dollar of compute produce?”

That is a radically different question.

And economists should probably start paying much more attention to it.

From labour productivity to intelligence productivity

Our traditional productivity equation is roughly built around capital and labour.

Workers receive better machinery.

Output per worker rises.

Companies invest.

Productivity increases.

Wages and profits eventually respond.

AI introduces something awkward into this framework.

What if part of what we previously classified as high-skilled labour begins behaving like capital?

Software engineers.

Analysts.

Mathematicians.

Researchers.

Lawyers.

Designers.

Scientists.

Their cognitive output has historically been constrained by the number of qualified humans available.

But an AI agent can potentially be instantiated again.

And again.

And again.

The 10,001st agent does not require twenty years of education before beginning work.

It requires compute.

This means the economic constraint may migrate from:

“How many brilliant people can we find?”

toward:

“How much computational intelligence can we afford to deploy?”

That is a fundamental change in the production function.

And here comes the uncomfortable part: $1 million versus perhaps tens of millions

The Clay Mathematics Institute attached a $1 million prize to each Millennium Problem.

OpenAI says it does not intend to claim the prize.

Which is fortunate, because economically the prize almost becomes a joke.

Outside estimates cited after the announcement have put the computational expense of the experiment somewhere in the range of $10 million to $40 million, although these are estimates rather than OpenAI’s disclosed accounting costs.

In other words, humanity offered:

$1 million for the answer.

A technology company may have spent multiples of that merely to demonstrate what its machines could do.

That tells us something fascinating about the economics of frontier AI.

The immediate product was not mathematics.

The real product was capability.

OpenAI was effectively demonstrating that sufficiently large quantities of computational intelligence can be concentrated on a scientific problem.

Today: fluid dynamics.

Tomorrow?

Drug discovery.

Battery chemistry.

Materials science.

Fusion.

Climate modelling.

Genetics.

Semiconductors.

Engineering.

Financial mathematics.

The commercial prize attached to those problems is not $1 million.

It can be hundreds of billions.

Science could become a capital-intensive industry

This creates a paradox.

AI is routinely described as a democratizing technology.

And at the level of access to existing knowledge, that is largely true.

A student in Buenos Aires, Nairobi or Jakarta can increasingly obtain explanations that once required access to excellent professors, libraries or specialist consultants.

That is extraordinary.

But the economics of creating new knowledge may move in exactly the opposite direction.

Consider what OpenAI reportedly deployed:

10,000 agents.

130 billion output tokens.

Millions of inter-agent communications.

Specialized models.

Large computing infrastructure.

Verification systems.

Engineering teams.

Access to enormous quantities of accumulated knowledge.

That isn’t democratized science.

That looks remarkably like industrialized science.

And industrialized science requires capital.

A lot of it.

The new laboratory may look like a data centre

For centuries, scientific prestige accumulated around universities.

Oxford.

Cambridge.

MIT.

Harvard.

Stanford.

ETH Zurich.

The intellectual infrastructure was concentrated there because the scarce assets were professors, students, laboratories, libraries and academic networks.

But suppose frontier discovery increasingly requires:

models + compute + electricity + data + orchestration + verification.

Who owns those assets?

Increasingly, not universities.

Technology companies do.

This potentially creates one of the largest transfers of intellectual power since the emergence of the modern research university.

The best mathematician in the world may possess extraordinary insight.

But what happens when another institution can deploy the equivalent of thousands of competent mathematical agents against the same problem continuously?

The individual still matters.

But the capital surrounding intelligence begins to matter enormously.

We have seen this movie before.

A brilliant craftsman once competed with factories.

The craftsman did not suddenly become less talented.

The production system around him changed.

The productivity statistics could become absurd

Suppose 100 human researchers require five years to solve a problem.

That represents roughly:

500 researcher-years.

Now suppose an AI system reaches an equivalent result in four days.

Our conventional productivity statistics are poorly equipped to describe that transformation.

It is not a 5% productivity increase.

It isn’t 20%.

It is a different production architecture.

This is why much of today’s debate about whether AI increases annual productivity growth by 0.3%, 0.7% or 1.5% may eventually look strangely conservative.

Those estimates generally assume AI improves existing workflows.

But the more radical scenario is that AI creates new production functions for intellectual output.

That is qualitatively different.

But did the machine really invent the mathematics?

Here the story becomes considerably messier.

And more interesting.

The announcement has already produced a dispute involving mathematicians Tristan Buckmaster and Levent Alpöge, who had been working on related fluid-dynamics problems using AI systems themselves.

Questions have been raised about whether unpublished or private interactions with AI tools could indirectly influence model development.

OpenAI explicitly states that its researchers and agents did not access their specific user data while solving the problem.

But it adds an unusually important qualification: it says it cannot entirely rule out the possibility that de-identified data derived from product usage helped improve its models.

That sentence deserves far more attention than it is receiving.

Because it exposes what could become one of the most difficult intellectual-property questions of the AI era.

What if your research assistant is also your future competitor?

Imagine you are a mathematician.

You use an AI system to explore an unpublished hypothesis.

You ask questions.

Reject approaches.

Develop new ones.

Feed preliminary calculations into the system.

Over months, your interaction contains something extraordinarily valuable:

the path toward a discovery.

Now imagine the provider of that AI subsequently develops a much more powerful model.

Where exactly does your intellectual contribution end and the model’s knowledge begin?

Who owns the resulting insight?

The researcher?

The AI company?

Nobody?

Everyone?

Copyright law was designed largely around works.

Patent law around inventions.

Academic norms around authorship.

None of these systems were designed for a world in which millions of researchers continuously interact with a centralized cognitive infrastructure that may itself improve through usage.

That is a governance problem hiding inside a mathematics story.

Recent concerns from mathematicians go beyond Navier–Stokes and directly question how private mathematical interactions with AI systems might feed future model capabilities. OpenAI has denied using specific private data to generate the disputed results, while the broader debate about de-identified usage data remains unresolved.

Knowledge has always been cumulative

There is another reason to avoid simplistic accusations.

Almost no scientific discovery emerges from nowhere.

Newton stood on predecessors.

Einstein built upon Lorentz, Maxwell and others.

Modern mathematics consists of enormous chains of accumulated results.

The OpenAI proof itself cites earlier mathematical literature. Nature notes that specialists are treating the announcement seriously, although validation remains essential.

So demanding that an AI generate knowledge without relying on previous human knowledge would be absurd.

Humans don’t do that either.

The difficult question is different:

At what point does learning from humanity become appropriation from identifiable humans?

We have barely begun answering it.

Proof abundance could destroy one scarcity — and create another

Mathematician Terence Tao has discussed the possibility that AI could move mathematics toward something resembling an era of “proof abundance.”

That phrase deserves attention.

Mathematics historically suffers from scarcity.

Too many conjectures.

Too few mathematicians.

Too little time.

AI could reverse that.

Imagine millions of proofs, lemmas, conjectures and counterexamples generated continuously.

Then the scarce resource becomes something else:

attention.

Which results matter?

Which discoveries deserve human investigation?

Which proofs reveal deeper structure rather than merely establish correctness?

Which questions should machines attack in the first place?

The bottleneck could move from producing answers to choosing valuable questions.

And that might preserve a uniquely important role for humans longer than many expect.

The scientist may become the portfolio manager of intelligence

The scientist of 2040 may spend less time performing every calculation personally.

Instead, she might allocate computational intelligence.

Ten agents investigate hypothesis A.

Five hundred attack B.

Two thousand search for counterexamples to C.

Another system verifies every promising result.

A final model compares the outputs and identifies contradictions.

The human determines where resources should go.

In that world, the scientist begins to resemble something rather unexpected:

a portfolio manager.

Capital is allocated across competing intellectual hypotheses.

Bad approaches are terminated.

Promising ones receive more compute.

Returns are measured in discoveries.

Scientific research begins to acquire something resembling the economics of venture capital.

That may sound provocative today.

It could sound obvious twenty years from now.

And then there is China

There is inevitably a geopolitical dimension.

If scientific intelligence becomes partially scalable through compute, then semiconductor capacity, electricity generation, data centres and frontier AI models stop being merely technological infrastructure.

They become scientific power.

That makes export controls on advanced chips easier to understand.

The strategic question is no longer simply:

Who has the best chatbot?

It becomes:

Who possesses the largest deployable stock of machine intelligence?

The United States currently has enormous advantages through its frontier AI companies and semiconductor ecosystem.

China has scale, engineering capacity, energy ambitions, an enormous scientific workforce and powerful incentives to reduce dependence on American technology.

The AI race therefore increasingly resembles an industrial race.

And if machine intelligence begins producing genuine scientific breakthroughs, the geopolitical return on compute rises dramatically.

Universities face an uncomfortable future

There is another institution that should be paying close attention: academia.

Universities currently train researchers who frequently move into private AI laboratories.

Those laboratories possess compute resources that universities often cannot match.

If frontier research increasingly requires computational scale, universities risk becoming suppliers of talent to institutions that subsequently dominate the production of knowledge.

That would be a remarkable reversal.

For centuries, corporations commercialized discoveries originating in academia.

Tomorrow, universities may increasingly study discoveries originating inside corporations.

That shift deserves serious public-policy discussion.

Who verifies proprietary science?

Who determines research priorities?

Who receives access?

Which discoveries become public goods?

Which become trade secrets?

And what happens when knowledge with enormous societal value is generated by infrastructure owned by shareholders?

The biggest monopoly may not be data. It may be discovery.

We have spent years worrying about Big Tech monopolizing:

search,

social networks,

advertising,

cloud computing,

data.

Those concerns may eventually look relatively mundane.

The truly consequential concentration would be the ability to monopolize the machinery of discovery itself.

Imagine three or four companies possessing systems capable of running millions of scientific agents.

They wouldn’t merely own platforms.

They would own something approaching industrial-scale intellectual production.

And here lies the paradox.

AI could make intelligence cheaper than at any point in human history while simultaneously making frontier intelligence more capital-intensive than ever before.

Knowledge becomes abundant.

The machines producing the newest knowledge become concentrated.

That combination is economically explosive.

Do not fire the mathematicians yet

There is an important warning against technological hysteria.

OpenAI’s result still needs independent validation.

Its system also relied on an enormous accumulated body of human mathematics.

Human researchers developed the theories, notation, prior results and conceptual frameworks upon which modern AI systems operate.

And there are already serious debates over how much of the decisive conceptual innovation originated in previous human work. Researchers Diego Córdoba and Luis Martínez-Zoroa, for example, developed important earlier approaches involving vorticity cascades; the relationship between such human advances and subsequent AI-enabled results is already becoming part of the debate over attribution.

So the lesson is not:

“Mathematicians are obsolete.”

That would be intellectually lazy.

The more interesting possibility is:

one excellent mathematician equipped with thousands of AI agents may eventually compete with an entire traditional research institution.

That is far more disruptive.

The economics of intelligence have changed before our institutions have

For centuries, societies developed institutions around one basic assumption:

high-level intelligence is scarce because highly capable humans are scarce.

Universities.

Research laboratories.

Consultancies.

Investment banks.

Law firms.

Engineering companies.

Government bureaucracies.

Their structures all reflect that scarcity.

People are recruited.

Educated.

Organized into hierarchies.

Promoted.

Paid according to expertise.

AI introduces the possibility that part of expertise becomes reproducible capital.

Once that happens, economics changes.

Not because humans suddenly become unnecessary.

But because the relative price of human and machine cognition begins to move.

That may ultimately matter more than whether GPT can write emails, summarize PDFs or create PowerPoint presentations.

Those are appetizers.

Scientific discovery is the main course.

The question after Navier–Stokes

The Navier–Stokes result may eventually be confirmed.

It may require corrections.

Parts of the story about priority and attribution may change as the mathematical community examines the evidence.

That uncertainty should be acknowledged rather than hidden beneath AI enthusiasm.

But even if this particular proof encountered difficulties tomorrow, the economic signal would remain.

Thousands of AI agents were concentrated on a frontier scientific problem.

They collaborated.

They explored alternatives.

They exchanged millions of messages.

They generated and formally checked mathematical reasoning.

And they did so on a timescale measured in days rather than academic careers.

That capability will improve.

The important question therefore isn’t whether machines have become mathematicians.

It is whether intellectual production has begun its industrial revolution.

If the answer is yes, then we should stop discussing AI exclusively as another productivity tool.

Steam made physical power scalable.

Electricity made energy distributable.

Computers made calculation cheap.

The internet made information almost frictionless.

AI may make intelligence reproducible.

And if intelligence becomes reproducible, the decisive economic advantage may no longer belong to whoever has the smartest person in the room.

It may belong to whoever can afford to put 10,000 artificial ones beside them.

That is the profound question behind OpenAI’s mathematics breakthrough.

And unlike Navier–Stokes, we have barely begun trying to solve it.

Roberto Balbino
CEO, Fintech SCB
Independent analyst of finance, economics, geopolitics and global markets.