Whenever artificial intelligence is discussed in Europe, the conversation tends to resemble a race with only two lanes. America occupies one, powered by venture capital, silicon and world-class laboratories. China occupies the other, propelled by scale, a vast scientific workforce and a state that does not ask permission.
Europe, we are told, is left with a choice between catching up and regulating: either enter a race it cannot win, or exercise a veto power that generates resentment without creating companies or attracting investment. This is a widespread view, but it is also a mistaken one. The problem is that it has begun to shape policy — just as AI has become the principal source of friction between the two sides of the Atlantic.
The first reason Europe should not simply chase the United States and China follows from an elementary economic principle. Comparative advantage is not measured by what you are capable of doing, but by what you must give up in order to do it.
Concentrating public resources on training frontier models means competing precisely where Europe’s opportunity costs are highest: venture capital, cheap energy, concentrations of talent and tolerance for failure. It also means diverting resources away from areas where those costs are lower.
Moreover, markets for general-purpose models are characterised by increasing returns and powerful economies of scale. Those who arrive late on the same trajectory bear many of the pioneer’s costs without capturing the pioneer’s rents. Europe has already learned this lesson. It once tried to answer American search engines with a European search engine — and failed miserably. On the few occasions when Europe has genuinely prevailed, from GSM onwards, it did not chase a product. It set a standard, thereby changing the dimension on which competition took place.
There is another regularity, documented by four decades of business scholarship: the returns from innovation rarely accrue to the inventor. They tend instead to flow to whoever controls the complementary assets required to bring that innovation to market.
Europe therefore has an appropriation problem more than an invention problem. Shifting our attention from inventors to complementary assets means asking a different question: what does Europe possess that its two competitors cannot simply buy?
The answer lies in its archives. Not archives preserving the relics of the past, but digital ones: datasets and historical series containing the usable memory of European institutions.
Consider healthcare. Europe’s universal health systems have accumulated decades of longitudinal clinical histories, following the same individual from birth to death and connecting treatment to outcomes and expenditure. Nothing quite comparable exists in the United States, where health data are fragmented among insurers and continuity is often broken whenever someone changes jobs.
Alongside them sit population registries, land registries, company records, judicial and tax archives, telemetry generated by European machinery operating around the world, energy and telecommunications networks, and cultural archives.
These are fundamentally different from the behavioural data on which America’s digital advertising industry was built. They are longitudinal rather than instantaneous, verified rather than self-reported, traceable to a known source and supported by legal legitimacy.
Their quality rests on trust and legal certainty — perhaps the only factors of production Europe currently produces at lower cost and higher value than either of its competitors.
The problem is that this vast European digital patrimony can, in technical terms, be described as “dead capital.”
Drawing on his travels, the economist Hernando de Soto observed that the difference between rich and poor economies did not lie simply in how many assets poor people possessed, but in whether those assets had legal title and could therefore create a market: houses without deeds, land without registries, real wealth rendered legally mute and consequently unusable as collateral, as equity or as an object of exchange.
European data are in much the same condition. They exist. They have value. Yet they are neither properly titled nor readily transferable, and using them can expose an organisation to litigation.
Europe’s General Data Protection Regulation established a defensive perimeter, not a property title. It defined what cannot be done without establishing clearly who may do what, under which conditions and at what price.
And unlike land, data depreciate. An unused archive is not necessarily a protected archive. It is an archive growing old.
Economic theory offers a clear direction. Property rights, as Coase and Demsetz taught us, emerge when the value of a resource becomes great enough to justify the cost of defining those rights legally.
We have reached precisely that point.
Consider a historical precedent. The English enclosures of the eighteenth century converted customary, non-transferable rights over common land into defined and transferable titles, mobilising as capital wealth that had previously remained inert.
With data, however, the problem is reversed. Data are a non-rival good: using them does not consume them. Their problem is therefore not overexploitation but underuse. They are immobilised by a fragmentation of vetoes — what Michael Heller famously called the “tragedy of the anticommons.”
Europe’s task, then, is neither simply to offer greater protection nor merely to deregulate. It is to transform its enormous stock of distinctive institutional and industrial data — potentially its real competitive advantage in the future of applied artificial intelligence — from a fragmented and legally immobilised mass of information into a digital asset with identifiable beneficiaries, negotiable rights and a market price capable of attracting investment and resources.
Several options for institutional and regulatory reform, as well as economic policy, follow from this argument. All deserve consideration if Europe is serious about restarting its stalled engine of competitiveness in artificial intelligence.
The first is to establish data trusts: institutions with fiduciary duties towards the citizens who are their beneficiaries. These trusts would never surrender ownership of the data. Instead, they would license their use for limited periods, subject to audits and in return for payment.
Imagine a national health-data trust selling training licences to frontier AI laboratories, whether American or Chinese. Europe’s position in the value chain would be reversed. Instead of being merely an end market, it would become the supplier of one of the industry’s scarcest inputs — while the resulting revenue flowed back into the healthcare system rather than leaving Europe in the form of subscription fees.
It is perhaps the most liberal argument available. It is also likely to be the most unpopular, because it requires acknowledging an uncomfortable truth: data have value only when somebody uses them.
The second concerns public demand. Europe spends heavily subsidising supply and remarkably little shaping demand, despite public procurement accounting for roughly one-seventh of the European Union’s GDP.
Advance purchase commitments tied to measurable outcomes — reducing court backlogs, shortening healthcare waiting lists or redesigning defence logistics, for example — could create markets without governments having to pick winners.
The condition should be technical, but its implications would be constitutional: public administrations should purchase only AI systems they are capable of running themselves, based on artificial-intelligence models whose parameters and weights they can control, rather than buying them by subscription from American laboratories.
This would achieve two things simultaneously: suppliers would remain contestable, and algorithms inspectable.
he third is the most radical departure from Europe’s current regulatory architecture: moving from ex-ante prohibitions to ex-post regulation — replacing prior authorisation, wherever possible, with subsequent liability and compulsory insurance.
For high-risk applications, require an insurance policy and let insurers conduct the inspections.
This is the mechanism through which classification societies made navigation safer and inspection companies made steam boilers safer, long before — and often far more effectively than — any government ministry could have done so. The party that pays the claim possesses both the incentive and the information the regulator lacks.
An insurance premium is a price. And a price conveys information that no bureaucratic checklist can capture.
The fourth option is to turn compliance from a cost into an exportable product.
Europe could establish a system of independent, accredited verification backed by financial liability, with certification providing a legal safe harbour beyond the Union itself. Such certification could be purchased even by those not legally required to obtain it — just as, for two centuries, parties to international contracts have voluntarily chosen English law.
This would be the only version of the Brussels Effect capable of generating revenue rather than diplomatic resentment.
The fifth concerns method.
Instead of harmonising everything in advance, Europe would be better served by allowing individual member states to experiment with different regulatory regimes for a defined period, with mutual recognition of outcomes.
Competition between legal systems is a process of discovery. Preventive harmonisation assumes that the answer is already known.
The postponement until December 2027 of obligations governing high-risk systems — justified by the absence of sufficiently mature technical standards — is an official acknowledgement that it is not.
When it comes to democratic safeguards, Europe should distinguish between what is expensive and probably unattainable and what is relatively cheap and within reach.
Sovereignty over the training of frontier models belongs to the first category. Sovereignty over their deployment belongs to the second. What matters is the ability to run domestically the systems that determine whether someone receives a pension, is admitted to hospital or is sentenced by a court.
In the public sphere, authenticating content at the moment it is created is technically more robust than attempting to detect manipulation downstream — a race that is lost before it begins.
Pluralism, meanwhile, is better defended through competition law applied across the entire value chain, from compute to distribution, than through content rules that are destined always to arrive too late.
Then there is the Atlantic question, which forms the backdrop to all of this.
A Europe that merely follows becomes a customer. Customers buy, but they do not control.
A Europe that merely prohibits becomes an obstacle. Obstacles are bypassed.
A Europe that defines ownership, sets a price and monetises the scarce input others need becomes something else entirely: necessary.
And necessity is a stronger negotiating position than the autonomy promised by the much-vaunted language of “digital sovereignty” — and a considerably more realistic one.
The transatlantic relationship will not be repaired by demanding parity or threatening sanctions. It will be rebuilt by becoming indispensable in an area where the other side cannot produce the relevant asset for itself.
Europe has been storing that asset in its archives for decades.
It simply has to decide to open them — on its own terms, and at a price.
