No economy got rich by extending one line further. Farming and soldiering separated. So did medicine, law and finance. Prosperity came from depth in different places, connected well, not from one person or one system learning to do everything better than everyone else. The current race in artificial intelligence runs in the opposite direction, toward one, two or three universal models sitting behind everything from healthcare to banking to government, on the assumption that the line simply keeps extending in one direction. That is usually described as progress. I think it is better understood as an extraordinary concentration of power, and one that should worry a British reader as much as it worries anyone in Washington or Beijing.
Why no company or country should own the world’s mind
The farmer in ancient Greece might also have been a part-time warrior. He grew food in one season and took up arms when his city required it. The arrangement was necessary, sometimes effective and profoundly limited.
Modern societies work differently. Farming and soldiering are separate disciplines. So are medicine, engineering, banking, law, science and thousands of other forms of work. Each has developed its own knowledge, language, tools, standards and institutions. The modern economy became productive not because one person learnt to do everything, but because people became very good at different things and learnt how to exchange what they produced.
The same principle operates between countries. China has become exceptionally good at manufacturing electric vehicles and batteries. Britain has become exceptionally good at the financial, legal and professional services which oil the wheels of global trade. Comparative advantage allows each to contribute specialist capability without reproducing the whole world economy within its own borders.
Yet the dominant ambition in artificial intelligence is moving in the opposite direction. We are trying to place an ever larger share of human knowledge inside one, two or three universal models. Those models may eventually sit behind our businesses, public services, schools, research, communications and personal decisions.
This is usually described as technological progress. It is also an extraordinary concentration of economic and political power.
How prosperity became specialised
The division of labour is so embedded in modern life that we barely notice it. No individual manufactures their own telephone, generates all their electricity, grows all their food and provides their own medical care. The apparent vulnerability of depending upon others is more than offset by the depth of capability created through specialisation.
Specialisation does not work in isolation. It needs connections. Farmers need markets. Banks need payment systems. Manufacturers need supply chains. Countries need trade, legal certainty and shared standards. The productive unit is not the solitary expert but the network which connects different forms of expertise.
This matters because we have begun to talk about artificial intelligence as though intelligence itself were a single commodity. A model is judged against a wide collection of benchmarks and placed somewhere on an implied ladder. The model nearest the top is called the most intelligent. Greater breadth becomes evidence of superiority.
Human progress suggests a different possibility. Intelligence may be less like a ladder and more like an economy. Its value lies in deep but different capabilities, combined effectively for a purpose.
Most organisations do not need a universal artificial mind for most of their work. A bank assessing its liquidity does not require a model to understand the influence of Shakespeare upon Picasso. It needs a system which understands liquidity risk, knows the bank, retrieves reliable evidence, performs the calculations correctly and recognises when the question crosses into another discipline.
The wider knowledge may occasionally matter. A cultural shift can become a conduct risk. A war can become a liquidity event. Scientific research can alter an insurer’s liabilities. When that happens, the risk specialist should be able to call upon geopolitical, behavioural or scientific expertise.
It does not follow that every specialist must permanently contain all of it.
The universal model as an economic anomaly
Viewed in this way, the universal model is an unusual economic proposition. It attempts to internalise capabilities which an economy would normally distribute among specialists. Enormous quantities of data and computing power are used to produce a single system which can answer almost any question, even though most individual users repeatedly require only a narrow part of its capability.
This may be necessary at the research frontier. A broad foundational model can discover connections which specialists miss. It can train or support smaller models. It can address genuinely novel questions for which no established discipline exists.
But the research engine need not become the operating engine for the whole economy.
A more natural architecture would involve networks of smaller models with particular expertise. A routing system would identify what the question requires. Relevant specialists would retrieve controlled information, apply tools and exchange findings. Another component could challenge the evidence or reasoning. The universal model would remain available for exceptional cases.
This mirrors the division of labour. No component knows everything. The network knows how to assemble what is needed.
Recent research into model routing and mixtures of experts suggests that this is technically credible. A large-scale study published through the Association for Computational Linguistics found that a capable router could improve collective performance as the number of candidate models increased, including when the pool contained many relatively small open models. The important innovation may therefore be not simply a cleverer model, but a better market for machine expertise.
A model is not merely a library
The case for distributed intelligence would be strong if this were only a matter of efficiency. It becomes much stronger once we recognise that a model is not a neutral container of facts.
Every general-purpose model embodies choices. Its training material contains some languages, histories and perspectives more completely than others. Its developers decide how it should respond to contested questions, which risks justify refusal and what forms of expression are acceptable. Later adjustments can change those judgements without the user seeing how or why.
These are not all technical decisions. They concern privacy, freedom, offence, fairness, family, work, identity, the relationship between individual and state, and the boundaries of legitimate political argument. Reasonable societies answer them differently.
American social mores do not perfectly represent those of Britain, France, India, Japan, Nigeria or Greece. Chinese social and political norms differ again. Neither should become the invisible default for global intelligence merely because companies within that country assembled the capital and computing power to build the largest model.
The UNESCO Recommendation on the Ethics of Artificial Intelligence recognises the importance of cultural identity, linguistic diversity and pluralism. But pluralism cannot mean that one distant model owner agrees to imitate local difference when prompted. A culture represented at the discretion of a platform has not retained authority over its own values.
Where the power sits
At present the frontier is not quite controlled by one country. It is increasingly divided between two, with the United States holding the stronger commercial position and China possessing formidable research, industrial and state-supported capacity.
The Stanford AI Index 2026 reports that US private investment in AI reached $285.9 billion during 2025, more than 23 times the recorded Chinese private investment. That comparison understates Chinese government-directed funding, but the concentration of commercial capital is unmistakable. Stanford also finds that China leads in publication volume, citations and patents while the United States continues to lead in the development of notable models.
Other countries participate in the supply chain without necessarily controlling it. Advanced chips may be designed in America, fabricated in Taiwan using Dutch lithography equipment, installed in a data centre elsewhere and accessed through an American cloud platform. The host country can point to its physical infrastructure while remaining dependent upon foreign technology, licences and policy.
The OECD’s mapping of AI compute starts from nine cloud providers which together account for more than 70 per cent of global public-cloud expenditure. Concentration at the model level therefore sits on top of concentration in the compute and cloud layers.
Britain has begun to recognise the strategic consequence. In April 2026, the Government argued that the UK must become an AI “maker”, rather than merely an AI “taker”, if it is to retain control over its economic prosperity and national security. That is the language of sovereignty, not ordinary technology procurement.
The despotic-leader test
Good governance is not designed on the assumption that tomorrow’s leaders will be wise and benevolent. Constitutions constrain offices because their future occupants are unknown. Risk management asks what a power could do in hostile hands, not simply how its present holder says it will behave.
The same test should be applied to artificial intelligence.
Imagine that much of the world’s cognitive infrastructure is controlled by companies subject to one jurisdiction. A government does not need to nationalise them to exert power. It can use export controls, licences, procurement, energy access, taxation, security legislation and informal pressure. It can determine which countries receive advanced models or chips. It can require surveillance, restrict outputs or make continued access conditional upon political alignment.
Under a despotic leader, AI ceases to be simply an engine of economic and human development. It becomes leverage.
The risk is not confined to countries already regarded as authoritarian. Democratic institutions can erode. Governments change. Emergency powers expand. A globally important system which is tolerable only while a particular administration remains restrained is not a resilient system.
Nor must the intervention be dramatic. A subtle change to search, ranking or model behaviour can shape what millions of people see as credible. Withdrawal of an interface or cloud service can disable dependent businesses. Preferential compute access can determine which economies innovate and which merely consume.
The deepest danger is not censorship in its familiar form. It is the possibility that one country or company acquires influence over the cognitive terms on which other societies understand themselves.
The intellectual landlord
Concentrated AI power threatens competition as well as national autonomy.
If a small number of models mediate work across banking, medicine, science, law, education and government, their owners gain a view across emerging economic activity. They can see which applications are growing, which specialist services attract demand and where new knowledge is becoming valuable. They also control the price and conditions of access to the essential platform.
A specialist company may build its service upon a frontier model, contribute its own data and develop a market. The platform owner can change the terms, acquire the company or reproduce the service. The tenant improves the property and teaches the landlord where the rent can be increased.
This is not the division of labour. It is dependency upon an intellectual landlord.
An economy of specialised intelligence would work differently. Models, tools and evidence sources could interact through open protocols. Customers could replace one component without losing their data, history and operating capability. Specialists could compete on depth, reliability, service and trust rather than remaining permanently subordinate to the owner of a universal platform.
In financial services, competition should be fought on judgement, risk appetite, customer service and genuine relationships, not on the ability of one infrastructure provider to tether the market through friction. The same principle should apply to AI.
Sovereignty through substitutability
The answer is not for every country to build a national imitation of the largest American or Chinese model. That would be enormously expensive and would reproduce the concentration at a smaller scale. In authoritarian states, a sovereign national model might become an even more effective means of control.
Nor should sovereignty mean autarky. Britain does not need to manufacture every chip, host every data centre and train every model. Countries benefit from exchanging specialised capabilities, just as firms and individuals do.
The more useful objective is sovereignty through substitutability.
A country or institution has meaningful autonomy when it can change providers without losing access to its information, institutional memory or essential services. It requires portable data, open interfaces, multiple sources of compute and the ability to operate critical models locally. Dependence can be managed while suppliers remain replaceable. It becomes power when they are not.
This creates a different vision of sovereign AI. A British regulator could use models grounded in UK law, supervisory practice and institutional history. A Greek medical system could operate in Greek and reflect the organisation of Greek healthcare. Chinese manufacturing expertise and British financial expertise could interact through shared protocols without either being absorbed into a universal model controlled by the other.
General-purpose models would remain important. They would be participants in the ecosystem rather than sovereigns over it.
Specialisation has risks too
It would be foolish to treat distribution as an automatic cure.
Specialised economies create dependencies. A disruption to one critical supplier can travel through the network. A routing system can send a question to the wrong model. Several apparently independent specialists may rely upon the same flawed data or underlying technology. Responsibility can become so fragmented that no participant accepts accountability for the final result.
Local control can also protect bad practice. A national model may preserve cultural identity or enforce domestic censorship. A professional model may embody the prejudices and blind spots of its discipline. Pluralism cannot mean that every value or use is beyond challenge.
The response is governance. Systems need common evidence standards, disclosed jurisdictions and clear accountability. Users should know which models contributed, which sources were used and which behavioural rules applied. Important decisions require appeal and independent challenge. International standards should establish basic protections while permitting legitimate cultural difference.
Global commerce already works in something like this way. Countries retain different laws, institutions and social settlements, but trade depends upon standards, enforceable agreements and mechanisms for resolving disputes. Connection does not require uniformity. Sovereignty does not require isolation.
The same balance is possible for intelligence.
A different measure of progress
The race to build a universal model encourages us to measure progress by concentration. More knowledge inside one system. More users on one platform. More computing power in one estate. More industries dependent upon one interface.
We should consider the opposite measure.
An intelligent economy would contain many centres of capability. It would make them discoverable and interoperable. It would route questions effectively, preserve the origin of evidence and recognise when several disciplines are required. It would tolerate failure in one component without disabling the whole. Above all, it would prevent knowledge from silently becoming unaccountable authority.
This is not an argument against frontier research. Nor is it a nostalgic defence of human inefficiency. It is an argument that the economic principle which transformed human prosperity should also shape artificial intelligence.
The division of labour allowed us to move beyond the farmer who had to become a warrior. The division of intelligence may allow us to move beyond the idea that one machine must know everything.
No company should own the world’s mind. No country should be able to turn it off.
Richard Anderson
Founder, AndersonRisk
This essay was first published on 3 October 2026 on my Substack.