Can Europe Break American Dominance Without Copying Silicon Valley?

1. Research question and methodological framework

Key ideaCan Europe develop its own ChatGPT?

Technically, a European chatbot would be possible. But it would not yet constitute an independent European AI industry.

Key ideaCan Europe build an economically viable, socially beneficial, and safely controllable AI structure that competes with American platforms in selected markets, limits critical dependencies, and retains a significant share of value creation in Europe?

This question encompasses at least five dimensions:

• Technological capability: Can Europe develop, operate, evaluate, and adapt capable models?

• Economic viability: Can these give rise to products and companies with lasting demand from paying customers?

• Industrial impact: Does the technology increase the productivity of European companies and public institutions?

• Distribution: Who receives the productivity gains—platform operators, owners, customers, employees, or the general public?

• Safety and control: Who evaluates the systems, and how can dangerous or unexpected capabilities be limited?

A strategy that pursues only the first objective remains incomplete. Europe could develop a technically capable model and still fail economically. It could build profitable AI companies while devaluing human capabilities. It could operate European servers while remaining structurally dependent on chips, cloud software, or model architectures.

1.1 Limits of the available data

The economics of leading AI companies cannot currently be assessed precisely.

Companies provide only incomplete disclosures regarding:

• inference costs per user,

• gross margins for individual products,

• internal pricing for computing power,

• utilization of their data centers,

• chip depreciation periods,

• the cost of free offerings,

• research budgets for individual model generations,

• the extent of cross-subsidization.

Technical transparency is also declining in some respects. The Stanford AI Index notes that, particularly for highly capable models, data volumes, training duration, training code, parameter counts, and computational expenditure are often not fully disclosed.

Any precise calculation of how much revenue the model industry requires or which current providers will still exist in 2030 would therefore offer a false sense of precision.

What can be analyzed robustly, by contrast, are:

• investment and infrastructure trends,

• market structure and barriers to entry,

• technological performance gaps,

• European financing shortfalls,

• energy requirements,

• governance and safety requirements,

• various industrial-policy scenarios.

2. What does it mean to win the AI competition?

The statement “America is winning AI” is meaningless without defining the market under consideration.

AI value creation comprises several layers.

2.1 Computing infrastructure

The physical foundation of AI includes:

• data centers,

• AI accelerators,

• storage,

• high-performance networks,

• cooling,

• power generation and grid connections,

• cloud platforms.

American companies possess a substantial structural advantage here. Amazon, Microsoft, and Google have global data centers, existing enterprise customers, and profitable business units from which they can co-finance the expansion of their AI infrastructure.

The IEA estimates that global data-center electricity consumption could rise from around 485 terawatt-hours in 2025 to approximately 950 terawatt-hours in 2030. Affordable and reliable energy would thus become a decisive location factor for AI.

Europe is unlikely to overtake the USA at this layer in the short term.

2.2 General-purpose foundation models

General-purpose models increasingly combine:

• language processing,

• programming,

• mathematics,

• images and video,

• research,

• tool use,

• agentic task execution.

Here, too, American companies largely lead. At the same time, the gap between the best American and Chinese systems is narrowing on certain benchmarks. The Stanford AI Index puts the gap between the respective leading models at 2,7 percentage points in March 2026, while also warning of growing problems with the reliability and susceptibility to manipulation of common benchmarks.

The USA leads. But this does not establish an insurmountable technological advantage.

2.3 Platforms and distribution

A model becomes particularly powerful economically when it is distributed through existing platforms.

Microsoft can connect AI with Azure, Microsoft 365, Windows, and GitHub. Google has search, advertising, Android, and Workspace. Amazon owns AWS and major commerce and logistics systems. Meta controls global communications platforms.

These companies do not need to build new customer relationships from scratch. They can offer models as components of existing products and spread their costs across several business areas.

An independent European model manufacturer does not initially have this option.

2.4 Enterprise integration

Companies rarely buy only a model.

They need:

• data connections,

• access management,

• security controls,

• monitoring,

• support,

• availability guarantees,

• liability,

• adaptation to internal processes.

At this layer, European companies can control a large share of value creation even if the underlying foundation model comes from the USA.

2.5 Specialized applications

In areas such as medicine, industry, energy, law, logistics, or translation, general model performance alone is not decisive.

The crucial factors are:

• domain expertise,

• proprietary data,

• certification,

• reliability,

• integration,

• liability,

• existing customer relationships.

The USA can therefore dominate the infrastructure and foundation-model layers without automatically controlling every application layer.

3. Why the USA remains the clear favorite

3.1 Capital

The American investment lead is substantial. The Stanford AI Index documents a sharp rise in global corporate investment in 2025 and a particularly dominant role for the USA.

Capital enables:

• larger training runs,

• parallel research,

• high salaries,

• long-term financing of losses,

• early acquisition of scarce computing capacity,

• global distribution.

3.2 Vertical integration

Large American corporations simultaneously control:

• infrastructure,

• models,

• platforms,

• applications,

• customer interfaces.

A standalone model provider must monetize its model directly. An integrated technology corporation can use a model strategically even when it temporarily generates no substantial standalone return.

3.3 Europe’s scale-up gap

Europe’s weakness lies not only in starting companies, but above all in scaling them.

According to analyses by the European Investment Bank, European scale-ups raise around 50 percent less capital by their tenth year than comparable companies in the San Francisco area. The shortfall persists across industries and economic cycles.

The EIB also points out that European venture-capital volume is significantly lower than in the USA and that successful companies often become dependent on foreign investors.

3.4 Speed

American technology companies frequently release products early, test them in the market, and improve them afterward.

This strategy can create substantial safety and quality problems. In the fast-moving software market, however, it also creates a head start.

Europe therefore cannot respond to American competition with political coordination processes lasting several years.

4. Why the American lead does not automatically decide the entire market

4.1 Model performance is becoming reproducible more quickly

Performance gaps between leading systems have narrowed across several evaluations. At the same time, conventional benchmarks are being saturated increasingly quickly and are sometimes vulnerable to data contamination or strategic optimization.

A technological lead can therefore be substantial without remaining stable over the long term.

4.2 The model market remains dynamic

Although the OECD sees clear risks of concentration in compute, cloud infrastructure, data, and skilled personnel, it also observes falling prices, more models, and continued new market entry.

The most likely picture is therefore not a complete monopoly, but rather:

• an oligopolistic core of very large platforms,

• several regional or state-supported providers,

• open model ecosystems,

• specialized companies.

4.3 Open models limit dependency

Open model weights enable:

• local deployment,

• proprietary adaptation,

• independent testing,

• lower switching costs.

However, the International AI Safety Report 2026 notes that safety restrictions can be removed more easily from open models and that published model weights are effectively impossible to recall. Openness promotes competition and research, but it also increases the risk of misuse in highly capable systems.

4.4 Model routing creates a multi-provider market

Companies do not have to process every task with the most capable model.

A system can:

• process simple cases locally,

• route standard tasks to inexpensive models,

• keep sensitive data in-house,

• send complex exceptional cases to frontier models.

This reduces the need to bind the organization completely to a single provider.

5. The cost problem facing independent model manufacturers

The economic tension in the model market arises from two opposing developments.

5.1 Performance already achieved is becoming cheaper

More efficient architectures, more capable hardware, smaller models, and competition reduce the cost of capabilities that are already known.

5.2 The technological frontier remains capital-intensive

Anyone seeking to remain at the model frontier must:

• secure large computing clusters,

• employ highly qualified teams,

• develop new training methods,

• finance failed attempts,

• build safety and evaluation systems.

Revenues at leading companies are growing rapidly. At the same time, infrastructure and compute spending is rising sharply, while reliable data on gross margins and free cash flow is lacking.

5.3 Inference creates ongoing costs

Every additional use requires computing power. Particularly resource-intensive tasks include:

• long contexts,

• speech and video processing,

• extensive reasoning,

• agentic systems with many tool calls.

The IEA reports that data-center electricity consumption rose by 17 percent worldwide in 2025. At the same time, energy consumption per task is falling rapidly. However, overall growth in usage and compute-intensive applications can more than offset these efficiency gains.

5.4 Advertising is one option, not the only one

A free or very inexpensive consumer assistant needs a payer. This could be:

• the user,

• an enterprise customer,

• an advertising customer,

• a platform operator,

• a commercial partner,

• a government client.

Advertising is therefore not a necessary business model for AI as a whole. It can, however, help finance free mass-market offerings.

Direct revenue models are probably more attractive to European companies:

• enterprise licenses,

• usage-based pricing,

• on-premises installation,

• integration,

• service contracts,

• outcome-based compensation.

Europe does not need to try to copy the American free-service and advertising model.

6. What Airbus actually proves

Airbus does not prove that Europe inherently builds better products.

Airbus proves:

Key ideaEstablished American dominance can be broken through coordinated industrial policy, pooled expertise, long-term financing, and corporate integration.

6.1 Pooling existing capabilities

Europe already had aerospace companies, engineers, research institutions, and suppliers. These capabilities, however, were fragmented along national lines.

6.2 Clearly defined market

Airbus developed a clearly defined product for readily identifiable customers: safe, certified, and economically viable commercial aircraft.

6.3 Committed demand

European airlines and public-sector actors formed an important home market.

6.4 Organizational integration

Airbus evolved from a consortium into a more integrated company under unified management. This integration was crucial in subordinating particular national interests to a unified product and corporate strategy.

6.5 Competitor errors

Boeing’s documented quality problems reinforced Europe’s success.

The NTSB attributed the 2024 door-plug incident to inadequate training, guidance, and oversight by Boeing. The FAA found numerous violations of the quality system and intensified its oversight.

These findings concern Boeing. They do not prove that American companies are generally incapable of delivering quality.

7. Why the Airbus organization cannot be transferred unchanged to AI

7.1 Different product cycles

An aircraft program runs for decades. An AI model can lose relative competitiveness within a few months.

7.2 Different switching costs

An airline cannot replace its fleet at short notice. Certification, maintenance, training, and spare parts create strong lock-in.

Technically, an AI model can be switched through an interface. But switching costs rise once data, agents, and processes are deeply integrated.

7.3 Greater modularity

An AI application can combine components from different sources:

• an American foundation model,

• European data hosting,

• German industry expertise,

• French model adaptation,

• an independent testing structure.

7.4 Unclear product definition

A European “AI Airbus” could mean:

• foundation model,

• cloud,

• data center,

• industrial platform,

• government assistant,

• open model,

• consumer product.

A company expected to perform all these tasks would probably be overwhelmed organizationally.

Key ideaEurope needs the Airbus principle, not necessarily a single Airbus company.

8. European quality must become measurable

Europe cannot win by declaring itself morally or technically superior.

AI quality must be assessed using specific criteria:

Factual accuracy

How often does the system produce factually incorrect results?

Reliability

Does it deliver consistent results on comparable tasks?

Calibration

Can the model recognize when it is uncertain?

Traceability

Can sources, model versions, data, and processing steps be documented?

Controllability

Can the operator limit permitted tools, data sources, and actions?

Data protection and legal control

Where is data processed, and which laws govern the operators and infrastructure?

Long-term availability

Can the system be maintained, updated, and replaced over a period of years?

Cost-effectiveness

What is the total cost of the entire process, not merely the cost of individual tokens?

Human impact

Does the system expand employees’ capabilities, or reduce their role to monitoring and correcting errors?

A European model does not have to lead on every general benchmark. It can be better if it solves a specific task with greater controllability, safety, and cost-effectiveness.

9. First condition: AI must expand human capabilities

The European debate focuses heavily on models, data centers, and company valuations. It too rarely asks what kind of work this technology is supposed to create.

9.1 Productivity gains are possible, but not guaranteed

The Stanford AI Index describes measurable productivity gains in clearly structured activities, such as software development, customer service, or marketing. At the macroeconomic level, however, the evidence remains preliminary and mixed.

Daron Acemoglu analyzes AI using a task-based model. He distinguishes among automation, the augmentation of human labor, and the creation of new tasks. A technology can replace individual activities without automatically generating large economy-wide productivity gains.

9.2 Automation is not synonymous with social progress

A European AI system would be socially weak if its primary benefit consisted of:

• monitoring employees,

• reducing staffing,

• centralizing expertise in proprietary systems,

• replacing human decisions with opaque models.

9.3 Augmentative AI as a European development path

Europe should give priority to promoting systems that:

• enable technicians to make better diagnoses,

• relieve care workers of documentation burdens,

• provide small businesses with expertise,

• support employees in their native language,

• integrate continuing vocational education into the work process,

• augment scientific work.

9.4 Distribution of productivity gains

Productivity gains can flow to owners, platform operators, customers, employees, or the state.

A European AI strategy should therefore measure more than how much working time is saved. It should also examine:

• whether wages rise,

• whether new skilled tasks are created,

• whether employees receive further training,

• whether small and medium-sized enterprises benefit,

• whether the quality of work improves.

• The more profoundly an AI system changes human work, the more extensively employees must be involved in its development, introduction, and evaluation.

10. Second condition: Europe needs a public mission

“European sovereignty” is not yet a complete mission. It describes a capability, but not a specific social outcome.

Mariana Mazzucato’s mission-oriented approach views industrial policy as more than the correction of market failures. The state and the private sector should jointly orient markets toward verifiable public goals.

10.1 A possible European mission

• By 2035, Europe develops controllable AI infrastructure that increases industrial productivity, improves public services, expands human capabilities, and limits critical dependencies on individual non-European platforms.

• lower energy and material consumption in industry,

• less documentation work in healthcare,

• shorter processing times in public administrations,

• better multilingual communication,

• access to high-quality AI for small businesses,

• open tools for science and education.

10.2 Public funding requires reciprocal commitments

If the public:

• finances data centers,

• pays for research,

• assumes development risks,

• guarantees long-term contracts,

economic success must not be fully privatized.

Mazzucato and other authors propose industrial-policy conditionalities for this purpose. These can be specified in funding, procurement, and equity-investment contracts.

Possible conditions would include:

• open interfaces,

• data portability,

• reinvestment in European research,

• commitments on training and location,

• transparent energy metrics,

• European operating options,

• fair licensing terms for public institutions,

• repayment or equity-participation mechanisms in the event of exceptional success.

The basic rule is:

Key ideaThose who receive public risk owe public value.

This does not mean that ministries should make technical decisions. The state defines missions, conditions, and boundaries. Companies compete to deliver the best implementation.

11. Third condition: Safety requires independent institutions

A European provider would not automatically be safer than an American one.

European companies would also have incentives to:

• release quickly,

• present capabilities positively,

• downplay risks,

• gain market share.

The manufacturer must therefore not be the sole judge of whether its system is safe.

11.1 Multilayered safety concepts

The International AI Safety Report 2026 concludes that capable general-purpose AI systems are developing new capabilities while technical safeguards continue to have limitations. Among other measures, the report recommends evaluations of dangerous capabilities, thresholds, and predefined safety responses.

11.2 Independent evaluation

Models for critical applications should undergo external evaluation for:

• cyber capabilities,

• biological or chemical risks,

• manipulative capabilities,

• autonomous tool use,

• data protection,

• reliability,

• the ease with which safeguards can be bypassed.

11.3 Model and compute registers

Large training projects and particularly capable systems should be documented once defined thresholds are reached.

Relevant criteria could include:

• computing power used,

• specific capabilities,

• degree of autonomy,

• deployment in critical areas.

11.4 Incident reporting

Serious incidents should be subject to mandatory reporting:

• successful circumvention of safeguards,

• dangerous erroneous decisions,

• data leaks,

• uncontrolled agent actions,

• misuse in cyber or biological contexts.

11.5 Separation of manufacturer and evaluator

A European system requires three separate roles:

• manufacturer,

• independent technical evaluation institution,

• national or European regulator.

The manufacturer must not simultaneously be the developer, evaluator, and ultimate guarantor of safety.

11.6 Risk-based openness

Small and medium-sized models can be offered largely openly. Controlled access or staged releases may be necessary for particularly powerful systems.

Openness and safety are not absolute opposites. They must be weighed according to capabilities and risks.

12. DeepL as a European case study

DeepL is strategically relevant because the company is not primarily trying to develop the world’s largest general-purpose model.

It focuses on:

• translation,

• writing assistance,

• terminology,

• documents,

• multilingual business processes.

DeepL states that it is used by more than 200.000 companies. Because the figure comes from the company itself and does not include complete information on revenue or profitability, it demonstrates market reach, not necessarily economic resilience.

12.1 Why the model is interesting

DeepL addresses an existing problem for which customers are already willing to pay. Companies already have budgets for translation, localization, and international communication.

Moreover, the product consists of more than a single model call. APIs, glossaries, and enterprise integration can build customer loyalty.

12.2 The threat from universal models

General-purpose models are becoming increasingly capable translators. Basic translation could therefore become a free add-on feature.

DeepL must expand its offering without becoming just another universal assistant.

12.3 Infrastructure dependence

DeepL uses a hybrid infrastructure comprising self-managed servers in European data centers and AWS cloud infrastructure. According to DeepL, it remains the contracting party, while AWS acts as a subprocessor.

Key ideaEurope can build an internationally relevant AI product company.

And:

Key ideaA European product company does not automatically have an entirely European infrastructure.

13. Mistral and Europe’s foundation-model capabilities

Alongside specialized companies, Europe needs at least some general-purpose model capabilities of its own.

Mistral is currently the most obvious candidate. The company offers models that can be operated on a customer’s own infrastructure.

Mistral also reports a collaboration with Airbus in civil aviation, helicopters, defense, and space. It cites local deployment and control of sensitive data. Because these claims come from the provider, they demonstrate its strategic direction at this stage, not long-term economic success.

However, Europe should avoid merely replacing dependence on an American company with dependence on a single European provider.

What is needed:

• multiple lines of technical development,

• open alternatives,

• standardized interfaces,

• genuine options for switching.

14. The architecture of a European AI industrial system

Europe does not need a single supercorporation. It needs a multilayered system.

14.1 Computing infrastructure

The EU is working on 19 AI Factories in 16 member states; most are expected to be operational by the end of 2026.

These facilities are intended to connect:

• supercomputers,

• data resources,

• start-ups,

• research,

• training

with one another.

However, they are not equivalent to the largest commercial frontier clusters operated by American hyperscalers.

14.2 AI Gigafactories

The EU intends to mobilize 20 billion euros for larger AI Gigafactories. They are to be privately managed and publicly supported, giving European companies access to large-scale computing power.

However, announcement, financing, construction, and productive operation are different stages. Europe has a plan, but not yet conclusive proof of success.

14.3 Energy

Expansion requires:

• rapid grid connections,

• a reliable electricity supply,

• competitive prices,

• cooling,

• socially accepted locations.

Without an energy policy, a European AI strategy remains incomplete.

14.4 Foundation models and open alternatives

Europe needs at least one competitive general-purpose provider and at least one credible technical alternative.

14.5 Specialized providers

Alongside DeepL, European companies should emerge in:

• medicine,

• industry,

• robotics,

• energy,

• science,

• law,

• public administration.

14.6 Industrial anchor customers

Airbus, SAP, Siemens, Bosch, and European automotive, energy, pharmaceutical, and telecommunications companies would need to:

• define use cases,

• provide data and test environments,

• invest capital,

• agree on minimum purchase commitments.

14.7 Public procurement

Public administration, healthcare, education, research, and defense can create a large domestic market.

European origin should not be the sole deciding factor. Requirements should include:

• open interfaces,

• local operating options,

• auditability,

• provider switching,

• long-term maintenance.

14.8 Growth capital

Europe’s scale-up gap requires:

• larger European venture and growth funds,

• participation by insurers and pension funds,

• EIB instruments,

• long-term industrial investors.

14.9 Safety architecture

Independent auditing bodies and technically competent oversight must be part of the infrastructure from the outset.

15. Pros and cons of the Airbus principle

High fixed costs prevent spontaneous competition

When infrastructure and research costs are high, new global competitors do not necessarily emerge through market forces alone.

Europe already has customers

European industry, public administration, and public infrastructure constitute a large domestic market.

Shared infrastructure lowers barriers to entry

Not every company has to finance its own data center.

Procurement provides planning certainty

Long-term contracts can make research and product development financially viable.

A European alternative creates bargaining power

Europe does not have to fulfill every contract itself. It must be able to switch credibly between providers.

Specialization reflects Europe’s economic structure

Industry, medicine, energy, and science do not always need the largest universal model.

Europe might be too slow

Technology is changing faster than European decision-making processes.

Response: Infrastructure and standards can be organized jointly. Product development must remain entrepreneurial and decentralized.

Support could protect mediocrity

European companies could be supported because of their origin rather than their performance.

Response: Support requires measurable performance targets and clear termination criteria.

International systems could be cheaper

A European system built from scratch may incur higher costs.

Response: Noncritical applications should continue to use the world’s best systems. Europe needs capabilities of its own where dependence becomes strategically relevant.

Open models might be sufficient

Europe could operate foreign open models locally.

Response: Open models are important, but they do not replace infrastructure, research, evaluation, and further development of Europe’s own.

Safety requirements could slow innovation

Audits can delay development.

Response: Poor bureaucracy slows innovation. Technically competent, risk-based auditing, by contrast, can build trust and create market access.

16. Measurable criteria for success

A European AI strategy should not be judged by press releases, model size, or company valuations.

Technological capacity to act

• at least two independently operable model lines,

• local operating options for critical applications,

• demonstrated ability to switch between providers.

Economic impact

• multiple European AI companies with global revenue,

• increasing private follow-on investment,

• measurable productivity gains at small and medium-sized enterprises.

Work and skills

• the proportion of systems that augment employees rather than exclusively replace them,

• trends in wages, task quality, and time for further training,

• the number of newly created skilled roles.

Public value

• improvements in public administration, healthcare, energy, or education,

• returns from publicly funded companies,

• reinvestment in research and training.

Safety

• the proportion of externally audited high-risk systems,

• the number and quality of reported incidents,

• the time between identifying and correcting a risk,

• functional shutdown and switching options.

Sovereignty

• the number of critical systems not tied to a single non-European provider,

• the degree of control over data, operations, and models,

• secure access to computing power and energy.

17. Three scenarios for Europe’s future