When Economic Power Becomes Political Power
The AI monopoly problem is not market share. It is institutional power.
There is a lesson buried inside 25 years of institutional economics that should make governments much more uncomfortable about the AI boom.
Economic power has a habit of becoming political power.
And once political power becomes embedded in institutions, it becomes extremely difficult to reverse.
That is one of the central insights running through the work of Daron Acemoglu, Simon Johnson and James Robinson.
The paper I have been reading is a retrospective on their contribution to economics. It is not an AI paper. It is about something much older: why some societies build institutions that distribute opportunity broadly, while others develop institutions that protect the interests of relatively small elites.
But reading it in 2026, it is difficult not to see AI everywhere.
Because we may be watching the creation of an entirely new economic elite.
Like oil, through compute.
First, forget the chatbot
Most AI policy debates are still stuck at the product layer.
Copyright.
Bias.
Hallucinations.
Deepfakes.
Job displacement.
Model safety.
All important.
But underneath those debates, another process is taking place.
AI is concentrating four things simultaneously:
capital, compute, infrastructure and information.
And they are being concentrated inside a remarkably small number of companies.
The companies building frontier AI increasingly need semiconductor supply, enormous data centres, access to electricity, cooling infrastructure, global networks, proprietary datasets, engineers, financing and distribution.
This is not a normal software market.
You cannot reproduce a frontier AI company in a garage.
The barriers are increasingly physical.
Gigawatts.
GPUs.
Transformers.
Land.
Fibre.
Power contracts.
Billions of dollars of capital.
The AI economy is starting to resemble heavy industrial infrastructure more than the software economy that preceded it.
And that distinction matters enormously.
Because infrastructure creates dependency.
The institutional economics lesson
Acemoglu and Robinson’s political-economy framework starts with a conflict that sounds almost embarrassingly simple.
Societies contain groups with different economic interests.
Those groups compete over institutions.
Institutions determine how resources are distributed.
And groups with greater political power can shape those institutions in ways that protect their economic position.
The paper describes this as the struggle between elites and broader society, including the relationship between inequality, democracy and political power. Where wealth and asset ownership become highly concentrated, elites have stronger incentives and greater capacity to resist redistribution and institutional change.
The important distinction is between de jure political power and de facto political power.
De jure power comes from formal political institutions: elections, constitutions, parliaments, courts.
De facto power comes from resources.
Money.
Organisation.
Networks.
Control over strategically important assets.
The paper explicitly discusses how economic resources translate into de facto political power, and how political and economic power can persist across generations even after formal institutions change.
This is where the AI story becomes interesting.
Because the most important political power in the AI economy may not come from lobbying.
It may come from indispensability.
Imagine negotiating with a company you cannot replace
Suppose a government decides that a particular AI company has become too powerful.
Fine.
Regulate it.
But what happens when the same company operates models used across government agencies?
What if its cloud infrastructure hosts critical national services?
What if domestic companies depend on its APIs?
What if universities depend on its models?
What if defence agencies depend on its compute?
What if its data centres are among the largest new electricity customers in the country?
What if pension funds own billions of dollars of its equity?
What if the company is building infrastructure the state itself cannot build quickly?
Now regulation becomes more complicated.
The government is no longer regulating a normal corporation.
It is negotiating with part of the country’s technological infrastructure.
That changes the balance of power.
The question stops being:
What rules should this company follow?
And becomes:
What rules can we impose without disrupting infrastructure we increasingly depend on?
Those are very different political environments.
Dependency is political power
There is a tendency to think corporate political influence means campaign donations or lobbyists walking around parliament.
That is the old model.
The more interesting form of political power is structural.
Consider a country trying to regulate a dominant energy producer during an energy crisis.
Or a government trying to regulate its largest bank during a financial panic.
Or a city negotiating with its largest employer.
The formal authority still belongs to government.
But bargaining power becomes asymmetric.
AI could produce the same phenomenon.
Except the dependency may be deeper because the technology could sit inside decision-making itself.
Governments may increasingly depend on private AI infrastructure for:
administration, intelligence analysis, cybersecurity, healthcare, defence, education, taxation, research and public services.
At that point, AI companies are no longer simply vendors.
They become institutional counterparties.
And eventually, perhaps, institutional participants.
The trap is institutional persistence
One of the deepest arguments in institutional economics is persistence.
Institutions do not reset every election.
Historical arrangements can survive for decades or centuries because the groups benefiting from them have both the incentive and the power to preserve them.
The paper repeatedly returns to this idea, showing how colonial institutions, land ownership, political elites and historical distributions of power continued influencing economic outcomes long after their original conditions disappeared.
It also reviews evidence that political and economic power can survive dramatic formal changes: abolition, independence, democratisation and constitutional reform do not necessarily eliminate the influence of incumbent elites.
That should change how we think about AI competition.
Everyone is asking:
Will today’s AI leaders still dominate in five years?
Perhaps that is the wrong question.
The more important question is:
Will the institutional architecture being built today survive even if the companies change?
Long-term power contracts.
Data-centre clusters.
Cloud dependencies.
Model ecosystems.
Government procurement frameworks.
Technical standards.
Chip supply chains.
Training-data agreements.
National AI partnerships.
Once these become embedded, they create path dependency.
The company may change.
The institutional structure can remain.
Markets can become politics without anyone conspiring
This does not require corruption.
It does not require a smoke-filled room.
Nobody needs to secretly capture the government.
That is what makes the problem more interesting.
Imagine a government trying to build sovereign AI capability.
It needs GPUs.
There are only a handful of suppliers.
It needs cloud infrastructure.
Again, a handful of providers.
It needs frontier models.
Another small group.
It needs billions of dollars.
Now large financial institutions enter.
It needs electricity.
Utilities and data-centre developers enter.
Eventually an entire policy ecosystem forms around making these projects possible.
Permitting gets accelerated.
Grid connections become national priorities.
Tax incentives appear.
Planning rules change.
Energy policy adapts.
Training programs are funded.
Universities reorganise research priorities.
None of these decisions individually looks like political capture.
They may all be perfectly rational.
But together they produce something much more consequential:
the state begins reorganising itself around the requirements of the AI economy.
That is institutional power.
The railroad analogy is incomplete
People compare AI to electricity, the internet, railroads and oil.
All of those analogies contain something useful.
But there is an important difference.
AI infrastructure may eventually influence the decisions made through every other infrastructure.
The electricity company supplies electricity.
The railroad moves freight.
The telecom company transmits information.
AI systems may help decide:
where electricity goes,
which infrastructure receives investment,
who receives credit,
which companies are audited,
which patients receive treatment,
which military targets receive attention,
which research receives funding,
which laws get enforced,
and which information reaches policymakers.
That makes concentrated AI infrastructure unusually powerful.
It is both an economic input and potentially a decision-making layer.
Acemoglu and Johnson have already warned about the direction
The paper eventually moves from historical institutions to technological change.
It discusses Power and Progress, where Acemoglu and Johnson connect technological development with questions of who captures the gains from innovation.
Their basic argument is uncomfortable for Silicon Valley mythology.
Technological progress does not automatically produce broadly shared prosperity.
Who benefits depends on institutions, bargaining power and the direction technology takes.
The paper describes their concern that technological progress can coexist with increasing inequality and political power concentrated among technological elites.
That idea deserves much more attention in the AI debate.
Because productivity is only one variable.
The distribution of power created by productivity may matter more.
The AI companies may become quasi-sovereign actors
We are used to thinking about sovereignty geographically.
States control territory.
But the AI economy is creating another form of sovereignty.
Compute sovereignty.
Model sovereignty.
Data sovereignty.
Infrastructure sovereignty.
If governments lack these capabilities domestically, they become dependent on companies that possess them.
And dependencies constrain policy.
The uncomfortable scenario is not that AI companies overthrow governments.
That is science fiction.
The realistic scenario is much more boring.
Governments retain formal authority.
They pass laws.
They hold elections.
They appoint regulators.
But increasingly important parts of the economy rely on infrastructure controlled by a small number of firms.
Those firms therefore become impossible to ignore when writing policy.
Not because they control politicians.
Because they control capabilities the state needs.
That is a much stronger position.
The second-order effect of AI concentration
Economists will measure AI concentration using familiar tools.
Market share.
HHI.
Margins.
Return on capital.
Cloud concentration.
GPU ownership.
Model usage.
Those metrics matter.
But they may miss the larger institutional transformation.
The second-order effect of AI concentration is not simply monopoly pricing.
It is institutional influence.
Capital becomes infrastructure.
Infrastructure becomes dependency.
Dependency becomes bargaining power.
Bargaining power becomes political influence.
Political influence shapes institutions.
Institutions then reinforce the economic structure that created the influence in the first place.
That is the feedback loop.
And institutional economics tells us that once these loops become established, they can persist for a very long time.
This is why sovereign AI matters
I have argued before that sovereign AI cannot simply mean putting a national flag on a data centre.
The deeper question is whether governments retain meaningful strategic alternatives.
Can the state switch providers?
Can domestic firms access compute?
Can universities train models independently?
Can regulators inspect critical systems?
Can national infrastructure operate without permission from a handful of foreign companies?
Can electricity markets absorb AI demand without subordinating other industries?
Can competition policy intervene before infrastructure becomes impossible to replicate?
Those questions sound technical.
They are actually institutional.
The goal should not necessarily be to weaken successful AI companies.
That would be stupid.
The goal should be to prevent technological success from becoming irreversible institutional dependence.
The problem governments should solve now
The institutional economics literature contains a fairly brutal lesson.
It is much easier to prevent extreme concentrations of power than to dismantle them after they become embedded.
Once groups control valuable resources, they can use that position to influence the institutions governing those resources.
The paper’s review of Acemoglu, Johnson and Robinson repeatedly returns to this interaction between the distribution of economic resources, political power and institutional persistence.
AI policy therefore cannot only ask whether models are safe.
It needs another question:
What political economy are we building around AI?
Who owns the compute?
Who finances the infrastructure?
Who controls access?
Who sets the standards?
Who owns the energy contracts?
Who has the ability to exit one provider and move to another?
Who becomes indispensable?
Those questions may ultimately matter more than whether the next model scores 10 percent higher on a benchmark.
Because models will change.
Companies will rise and fall.
But institutions have a nasty habit of sticking around.
And the history of economic development suggests something else.
When extraordinary economic power accumulates somewhere, we should not assume it will remain merely economic.
Eventually, it starts asking for a seat at the political table.
The AI companies may not need to ask.
We may build the table around them.