Imagine that extraordinary intelligence becomes cheap enough for almost anyone to use.
A student, a small business, a scientist and a government agency can all consult systems more capable than today's best experts across most cognitive tasks. One enormous inequality appears to have disappeared: access to intelligence itself.
It is tempting to take the next step and conclude that power has been democratised too.
That step is not automatic.
Intelligence can tell you how to design a semiconductor fab. It does not give you a semiconductor fab. It can devise an energy strategy. It does not grant you a power station, a grid connection or permission to build one. It can draft a legal argument. It does not make you a judge. It can propose a political programme. It does not give you a vote in parliament, command of an army or ownership of the assets required to carry the programme out.
Broad access to superintelligence could still be one of the most democratising technological changes in history. But access to intelligence and access to power are not synonyms.
The question is what stands between an idea and its effect on the world.
1. The strongest case for abundance
There is a serious argument that widely available advanced AI could flatten differences in human capability.
We can already see a limited version of that mechanism. In a large field study of customer-support workers, Erik Brynjolfsson, Danielle Li and Lindsey Raymond found that access to a generative-AI assistant increased productivity on average, with the largest gains among less-experienced and lower-skilled workers.1 The result should not be stretched into a forecast about the whole economy, and the authors explicitly caution against doing so. But it demonstrates something important: a capable system can transfer useful patterns of expertise toward people who do not already possess them.
Extend that mechanism far enough and the implications become extraordinary. A small company could obtain analytical capabilities once requiring whole departments. A student could have an expert tutor available continuously. A researcher with modest funding could interrogate a literature larger than any individual could read. An entrepreneur could call on coding, legal, design and strategic capabilities that today require teams of specialists.
So the equalising case should not be treated as naive. Intelligence is an unusually powerful input. If it becomes cheap, replicable and widely accessible, some capability gaps really could collapse.
The mistake is assuming that every other gap collapses with them.
2. Capability is not the same as power
Consider two people using the same superintelligent system.
One has a laptop, a salary and ordinary legal rights. The other controls billions in capital, factories, data centres, energy contracts, distribution networks, political relationships and fleets of machines.
Their cognitive tool may be identical. Their capacity to alter the physical and institutional world is not.
A useful way to frame the problem is to separate four interacting dimensions: cognitive capability, effective agency, ownership and control, and institutional authority.
They influence one another, but they need not move together. Better reasoning can increase agency. Agency can help acquire assets. Assets can buy influence. Authority can determine which actions are legal. Yet none of those relationships is automatic.
This is why “everyone has superintelligence” is not yet a complete political or economic proposition. It tells us something profound about the distribution of cognition. It does not tell us who owns the complementary resources, who may deploy them, or which institutions decide whose plans become reality.
3. The substitutability test
The strongest objection to this argument is also the most important one.
Superintelligence may not behave like an ordinary input sitting beside capital, energy and institutions. It may be a meta-capability: something that helps its user obtain, improve or route around the other things they lack.
An advanced system could design more efficient chips, discover cheaper materials, optimise energy use, automate a company, negotiate contracts, coordinate thousands of people, find financing, write software, design robots and identify legal routes around institutional barriers. If it can repeatedly lower the cost of its own complements, today's bottlenecks may prove temporary.
That possibility is why the simple claim that “scarcity moves elsewhere” is too strong.
Instead, every proposed bottleneck should face a substitutability test:
Can intelligence replace the complement directly? Can it make the complement dramatically cheaper? Can it help acquire it? Can it coordinate shared access to it? Can it design around it? Or does the complement remain rivalrous, physical or institutionally allocated?
Knowledge and some forms of expertise may be highly substitutable. Organisation and coordination may become far cheaper. Compute may be partly substitutable through algorithmic efficiency, but computation still has a physical substrate. Energy can be used more efficiently and new generation can be designed, but electricity still has to be produced and delivered. Factories can become more automated, but land, materials, machines and construction remain physical.
Institutional complements are different again. Intelligence can help someone understand a law, argue a case or organise a political movement. It cannot simply declare itself the legal owner of somebody else's property or confer legitimate public authority on its user.
The long-run outcome therefore depends on which complements intelligence can dissolve and which survive it.
4. Scarcity is not concentration
There is another distinction that matters.
Energy is scarce in the ordinary economic sense, yet electricity can be available to millions of households through regulated grids. Land is rivalrous, but ownership can be distributed across many actors. Conversely, a resource can become cheaper while control over a strategically important layer remains concentrated.
Today's AI infrastructure illustrates why the distinction is useful. Advanced systems depend on chips, fabrication capacity, data centres, networks, cooling, electricity and large amounts of capital. OECD analysis describes high concentration and significant barriers to entry in several layers of the AI infrastructure stack.2 That is evidence about the present transition. It is not proof that the same concentration must survive a future of superintelligence.
Efficiency improvements, competition, open models, new hardware, regulation or technological substitution could broaden effective access. Equally, growing demand could keep the frontier capital-intensive even while yesterday's capabilities become cheap.
So the relevant question is not merely, “What remains scarce?” It is also, “How is control over what remains scarce distributed?”
5. Three different kinds of democratisation
Much confusion disappears if we stop using democratisation as though it described one outcome.
Capability diffusion asks who can perform sophisticated cognitive tasks.
Economic distribution asks who receives income, wealth and ownership claims generated by the technology.
Political decentralisation asks who possesses legitimate authority, influence over collective decisions and the capacity to enforce them.
These can move in different directions.
AI could make a less-experienced worker far more productive without giving that worker ownership of the company. Open weights could allow thousands of developers to adapt powerful models while the most advanced chips remain concentrated. A country could provide universal access to capable assistants while political authority remains exactly where it was.
The reverse combinations are possible too. Institutions could deliberately distribute ownership or access to infrastructure even if the underlying resources remain expensive.
This is why evidence that AI narrows one productivity gap cannot, by itself, establish that AI will narrow wealth inequality or decentralise political power.
6. What history can—and cannot—tell us
Previous general-purpose technologies are useful here, but only if we resist turning analogy into prophecy.
Technology can diffuse widely while differences in intensity of use persist. Diego Comin and Martí Mestieri's work on technology diffusion shows why formal adoption and effective use should be treated separately.3 Historical work on electrical technologies likewise illustrates the importance of complementary human and institutional capabilities for turning a general technology into downstream innovation.4
But history also provides evidence against a simple incumbent-power story. Research on transistor licensing and knowledge transfer finds that lowering access barriers helped broaden follow-on innovation, including among younger and smaller firms.5
The lesson is not that superintelligence will repeat electricity, computing or the transistor. It almost certainly differs from all of them in important ways. Software can replicate at extraordinary speed. Cognitive capability may help create its own complements. Autonomous agents may reproduce parts of organisations themselves.
History is therefore most useful as a catalogue of mechanisms: diffusion, complementary assets, learning, ownership, infrastructure and institutional rules. It tells us what questions to ask, not what answer the future must give.
7. Transition is not equilibrium
Today's AI economy begins from a world in which frontier training is expensive, specialised chips are difficult to manufacture and large-scale infrastructure takes time to build. It would be easy to project those conditions indefinitely and conclude that advanced AI must remain concentrated.
That would be a mistake.
It would be equally premature to observe falling inference costs, algorithmic improvements or open-weight models and conclude that effective power must become evenly distributed.
The transition and the eventual equilibrium may look very different.
One path is concentration. The frontier becomes more capital-intensive; large actors compound their advantages through compute, energy, data, robotics and ownership; intelligence increases the returns to assets they already control.
Another path is diffusion. Models become locally runnable; efficiency gains outpace demand; open ecosystems narrow performance gaps; AI helps small organisations acquire capital, coordinate labour and design around incumbent infrastructure.
Reality could contain both processes at once. Yesterday's frontier may diffuse broadly while tomorrow's frontier remains expensive. Capability could decentralise at one layer while ownership reconcentrates at another.
That is not a rhetorical escape hatch. It is the central empirical question.
8. A better Abundance Paradox
The earlier version of the Abundance Paradox suggested that when one important resource becomes abundant, scarcity simply moves elsewhere.
That intuition is useful, but stated as a law it is too deterministic.
A stronger formulation is:
When a powerful capability becomes abundant, the distribution of power depends on the remaining complements required to turn that capability into outcomes — and on whether the abundant capability can itself substitute for those complements.
Making intelligence cheap changes the power equation, but it does not solve it in advance.
This formulation leaves room for genuinely radical abundance. If superintelligence can make compute, energy, manufacturing, coordination and capital dramatically cheaper—and if institutions permit broad use of those capabilities—then access to intelligence could translate into an extraordinary decentralisation of effective power.
It also leaves room for a different result. If important complements remain physically rivalrous, legally allocated or institutionally controlled, then intelligence may increase the value of controlling them.
Neither outcome follows from intelligence alone.
9. What would change our mind?
A framework is more useful when it can lose.
The case for persistent complementary constraints would weaken if frontier-quality systems became cheap enough to run locally; if small organisations rapidly converged with large ones in real-world capability; if AI-enabled actors routinely substituted for major capital and institutional disadvantages; and if automated physical production made access to machines, energy and infrastructure broadly affordable.
The case would strengthen if the opposite occurred: frontier capability remained tied to enormous compute and energy investment; physical deployment required concentrated robotics and infrastructure; ownership of critical assets became more concentrated; or legal and political authority remained difficult for intelligence alone to route around.
Most importantly, we should measure the outcomes separately. Are cognitive capabilities converging? Are ownership and wealth converging? Is political influence becoming less concentrated? Those are different tests.
10. Access is necessary, not sufficient
There is a powerful moral and practical case for broad access to advanced intelligence. If such systems can expand education, scientific capability, entrepreneurship and individual problem-solving, restricting them to a small elite could preserve inequalities that technology is capable of reducing.
But universal access should not be confused with a completed theory of power.
Power is exercised through a world of physical resources, property claims, institutions, networks, legitimacy and collective rules. Superintelligence may transform every one of those things. It may make some of them less important. It may make others more valuable. It may even allow people to build substitutes that are difficult to imagine from today's technological position.
That uncertainty is precisely why neither utopian nor oligarchic certainty is warranted.
Giving everyone extraordinary intelligence would change civilisation. It could remove one of the deepest capability inequalities humanity has ever known.
What happens next depends on what intelligence can do to the constraints around it.
So the question is not whether everyone will have intelligence. It is what intelligence will still need in order to matter — and who gets to control it.
References & further reading
The central framework is an argument rather than an empirical finding. The references below support the historical and present-day claims used to test it.
- Brynjolfsson, Li & Raymond — Generative AI at Work, Quarterly Journal of Economics
- OECD — Competition in Artificial Intelligence Infrastructure
- Comin & Mestieri — If Technology Has Arrived Everywhere, Why Has Income Diverged?
- Lo & Sutthiphisal — Crossover Inventions and Knowledge Diffusion of General Purpose Technologies
- Nagler, Schnitzer & Watzinger — Transistor licensing, knowledge transfer and follow-on innovation
- Sastry et al. — Computing Power and the Governance of Artificial Intelligence
- OECD — Competition in the Age of AI
Final editorial edition, August 2026. Think Deeper welcomes substantive criticism, corrections and stronger counterarguments. Where the evidence changes, our conclusions should change with it.