Can Europe Afford to Rent Its Intelligence?
On 12 June 2026, the US government ordered Anthropic to suspend foreign-national access to its two most advanced models, Fable 5 and Mythos 5. Because Anthropic could not verify every user’s nationality in real time, it disabled both models for everyone. The European Commission contacted the company while it assessed the practical consequences for European users.
The interruption did not last. The broad export controls were lifted on 30 June, Fable 5 returned globally on 1 July, and access to the more sensitive Mythos 5 resumed for a narrower group of approved organisations while Anthropic sought permission to expand it. The episode was temporary, but that does not make it trivial. It established that access to an important AI capability could be interrupted abruptly by a supplier’s government, including for allied countries and paying customers.
For most software, a temporary outage is an operational problem. Another product may be available, data can sometimes be exported, and work can resume through a competing service. Advanced AI becomes more strategically important when institutions organise their workflows, data and expertise around a provider that they cannot readily replace. The risk is not that foreign models exist or that European organisations use them. It is that dependence becomes deep before anyone tests whether the service can be substituted.
This is the useful form of the European AI sovereignty question. Europe does not need a chatbot carrying the EU flag, nor must a European company defeat every American or Chinese model on every benchmark. It needs enough control over critical layers that a political or commercial decision elsewhere cannot disable functions on which European institutions have come to depend.
When AI Becomes a Dependency Layer
AI is not yet a single indispensable system. Its uses range from disposable consumer conveniences to functions whose failure could interrupt public services or affect consequential decisions. A company generating product descriptions faces a different risk from a hospital integrating a model into clinical documentation, an energy operator using one for cybersecurity, or a ministry building legal and administrative workflows around a proprietary platform.
The difference lies less in the sophistication of the model than in the surrounding dependency. How much institutional knowledge has been encoded into prompts, integrations and fine-tuning? Can the organisation export its data and move to another provider? Does an alternative model perform the same task well enough? Can the system continue locally during a disruption? Has the organisation retained staff who understand the underlying process, or has the model become the only practical route through it?
These questions become more urgent as AI moves from assisting individual workers to organising institutional workflows. Models can already support translation, programming, document analysis, research, cybersecurity and administrative work. They may increasingly mediate which sources are retrieved, how evidence is summarised, which languages receive strong support and which errors become normal. An institution that changes providers may therefore be changing more than software. It may have to reconstruct the assumptions, evaluations and operating procedures built around that software.
Foreign procurement does not automatically create unacceptable dependence. Europe has always relied on trade, alliances and international supply chains, and attempting to reproduce every technology domestically would make it poorer rather than safer. Dependence becomes strategically dangerous when a service is critical, concentrated among few providers, governed from outside European jurisdiction and difficult to replace within the time available during a disruption.
Sovereignty Is a Stack, Not a Nationality
The origin of a model is an incomplete test of sovereignty. A European company may train a model using American chips, host it through an American hyperscaler and depend on software maintained elsewhere. An American open-weight model may be operated entirely inside a European data centre. A supposedly sovereign system may still rely on foreign hardware, networking equipment, orchestration software and specialised engineers.
AI is a stack: semiconductor equipment, chips, data centres, electricity, cloud systems, training infrastructure, models, evaluation tools, deployment software, applications and skilled operators. Europe cannot realistically own every layer. Nor must it. Strategic resilience usually comes from controlling selected critical layers, maintaining alternatives and preventing any single dependency from becoming irreplaceable.
Open-weight models are valuable within that strategy, but the term is often asked to carry too much. Access to model weights can allow local hosting, adaptation, fine-tuning and forms of technical analysis that closed services do not permit. It can reduce reliance on a provider’s API and preserve a fallback after commercial access disappears. It does not reveal the complete training dataset, explain every development decision or make a system fully auditable. A model can be open-weight and still depend on opaque data, unavailable expertise and expensive foreign hardware.
The practical standard should therefore be operational rather than symbolic. For a critical use, can a European organisation run the model under an acceptable jurisdiction? Can it retain control of sensitive data? Can it evaluate the system against its own requirements? Can it move the workload to another model or provider? Can it continue operating if political relations, export rules or commercial terms change?
Different functions justify different answers. Consumer and routine business uses can rely heavily on international services. Sensitive research, public administration, defence, cybersecurity, healthcare and critical infrastructure need stronger fallback arrangements. Sovereignty is not a demand that every European user choose a European model. It is the capacity to choose differently when the stakes require it.
Mistral Is Evidence, Not an Answer
Mistral matters because it disproves the claim that serious European model development is impossible. Its Mistral Large 3 model was trained from scratch on 3,000 Nvidia H200 GPUs and released under the permissive Apache 2.0 licence. The model can be downloaded, customised and deployed through several different environments. Europe therefore has more than a hypothetical foothold in advanced general-purpose models.
It does not yet have a frontier ecosystem comparable in depth or capital intensity to those of the United States and China. The Stanford AI Index 2026 reports that US-based institutions produced 59 notable models in 2025 and Chinese institutions produced 35. More than 90 percent of notable models came from industry, where access to capital, compute, deployment data and large customer bases reinforces technical performance. The same report describes a global compute build-out dominated by major technology companies and a hardware supply chain concentrated around Nvidia and a single leading Taiwanese foundry.
Mistral should therefore be neither dismissed for failing to solve the whole problem nor celebrated as though one company already had. A model laboratory cannot by itself supply sovereign cloud infrastructure, abundant electricity, advanced chips, public-sector integration, safety evaluation, cybersecurity, applications and patient capital. Mistral itself distributes its models through a mixture of its own services and global cloud platforms. That is commercially sensible, but it shows why company nationality alone cannot carry a sovereignty strategy.
The useful question is where European systems are good enough to justify adoption and improvement. Public administrations, industrial firms, universities and regulated sectors do not always need the highest-scoring model available that month. They may value multilingual performance, local deployment, predictable support, control over data and the ability to adapt a model to European legal or industrial contexts. In such cases, accepting a small performance difference can be rational.
That principle also has a limit. Governments should not force institutions to use conspicuously weaker products merely to create a European customer. Protected procurement can preserve mediocrity as easily as it can build capacity. Strategic purchasing works when it gives promising technology demanding users, real feedback and a route towards competitiveness. It fails when “European” becomes an exemption from performance.
What Europe Actually Needs to Control
The European Commission has recognised compute as a strategic constraint. Its AI Continent plan envisages at least 19 AI factories built around Europe’s supercomputing network, up to five much larger AI gigafactories and a €20 billion facility intended to mobilise investment in those installations. The ambition is significant. Announced compute, however, is not the same as usable capacity.
Factories matter only when firms and institutions can obtain access without navigating a slow political allocation process. Compute needs electricity, networking, cooling, technical support and users capable of turning processing capacity into models and applications. A data centre built for strategic reasons can still become an expensive monument if access is cumbersome, workloads are distributed according to national bargaining rather than technical merit, or successful firms cannot obtain the capital needed for deployment.
Europe also needs more than training clusters. It needs several layers of replaceability: models that can be hosted and adapted locally, cloud services that can operate under European legal and security requirements, evaluation systems for critical applications, and teams able to migrate workloads when a supplier fails. Public institutions should know which AI functions are genuinely critical and should require credible exit plans before integrating them deeply.
This is closer to resilience engineering than to a race for prestige. A hospital does not need to train a frontier foundation model. It may need control over patient data, validated domain systems, continuity during an outage and the ability to change vendors. A defence ministry may require much stronger control over models, infrastructure and operators. A university may benefit most from affordable research compute and open models that can be examined and modified. Treating all three as one sovereignty problem produces either wasteful overbuilding or dangerously weak safeguards.
Specialisation also matters. Europe has substantial industrial capacity in aerospace, pharmaceuticals, energy, machinery, transport, robotics and advanced manufacturing. Applied AI in those sectors may generate more durable European advantage than a politically designed attempt to recreate every American consumer platform. Domain expertise, physical infrastructure and existing customers are assets that frontier-model leaderboards do not capture.
Regulation Needs Capacity
Europe’s regulatory instinct is not mistaken. AI systems require rules governing safety, discrimination, liability, cybersecurity, transparency and the use of personal data. The main provisions of the European AI Act apply from 2 August 2026, creating a framework that will influence how high-risk systems and general-purpose models are developed and deployed across the Single Market.
Regulatory power is nevertheless different from technical capacity. Europe can impose conditions on products entering its market, but it cannot regulate a missing European alternative into existence. If the models, chips, cloud infrastructure and deployment platforms remain concentrated elsewhere, European authorities negotiate with systems whose technical direction is largely decided outside their institutions.
The contrast between regulation and capacity should not be exaggerated. Common European rules can create scale by replacing conflicting national requirements, and safety standards can support adoption by increasing trust. Regulation becomes damaging when compliance costs fall disproportionately on smaller firms, rules remain uncertain during the period when companies need to invest, or member states implement a common framework through divergent national procedures.
The broader problem is examined in Europe’s slow competitiveness emergency. The continent often identifies a strategic weakness correctly but acts after infrastructure, capital and expertise have already accumulated elsewhere. AI compresses that timetable because model capability, hardware demand and commercial adoption are changing faster than conventional industrial programmes.
Europe therefore needs its regulatory and industrial strategies to reinforce each other. Rules should make critical systems safer and more portable. Procurement should create demanding customers for credible European suppliers. Competition policy should prevent lock-in without making scale impossible. Infrastructure funding should reward use and performance rather than geographical symbolism. Sovereignty requires institutions capable of building as well as institutions capable of supervising.
The Price of Non-Dependence
The strongest objection remains cost. Frontier AI requires chips, data centres, energy, engineering talent, deployment infrastructure and repeated investment in systems that become obsolete quickly. The largest American firms can finance this from enormous balance sheets. China can combine state direction with industrial scale. European projects pass through EU programmes, national interests, procurement law and political pressure to distribute benefits visibly.
This creates a real danger of spending heavily on symbolic capacity. Europe does not need twenty-seven national foundation models, a politically allocated GPU cathedral or public programmes that continue after their technical purpose has disappeared. Money spent on sovereignty can become protection for incumbents, consultants and projects that are too well connected to fail.
The alternative also has a cost. Persistent dependence transfers revenue, expertise, deployment experience and bargaining power elsewhere. It can make access conditional and leave institutions with workflows they cannot operate independently. These losses are difficult to see in a budget because they appear gradually—as companies acquired before they scale, researchers leaving for better compute, public systems locked into foreign platforms and European suppliers unable to secure demanding domestic customers.
A serious programme therefore has to tolerate some failed investments without tolerating permanent failure. The question of who pays when innovation fails matters because strategic capability cannot be developed through projects guaranteed to succeed. Governments must accept that some models, companies and infrastructure bets will disappoint. They must also be willing to close programmes that have stopped learning and redirect resources towards approaches that work.
The objective should remain non-dependence rather than supremacy. Europe should continue using strong American services, cooperate with allies and study competitive Chinese systems. It should also maintain enough European compute, model capacity, cloud infrastructure, domain expertise and operational skill to prevent any one foreign provider from becoming indispensable to critical functions.
Renting Intelligence
The metaphor in the title has limits. AI models do not constitute an institution’s intelligence, and using a foreign service does not mean surrendering the minds of the people who use it. The danger emerges when human expertise, data, procedures and software integrations are reorganised around a system that cannot be operated, examined sufficiently or replaced by the institution itself.
Europe does not need to own every model used by every citizen or company. It needs to know which functions must survive a supplier’s failure, a political dispute or a change in export controls. It needs domestic and open alternatives credible enough to make switching possible, infrastructure on which those alternatives can run, and procurement rules that prevent convenience from hardening silently into lock-in.
The Anthropic interruption ended quickly. That is reassuring at the level of immediate access and revealing at the level of institutional design. It showed both that allied governments can resolve such disputes and that contractual availability remains exposed to decisions outside Europe’s control.
Renting software is ordinary. Building an institution that cannot function when the rental ends is a choice. Europe’s sovereignty will depend less on whether every model is European than on whether the systems built around those models can continue when one of them is no longer available.
Comments
Post a Comment