Europe Should Bet on Narrow AI

Instead of trying to catch up in general-purpose AI, Europe should play to its industrial strengths.

Summary

• Out of its league: Europe is orders of magnitude behind the US and China on compute needed for frontier general-purpose AI models.

• And yet: The EU is still hoping to build trillion-parameter models through its €20 billion AI Gigafactories scheme.

• A better bet: As part of a broader hedge against different AI futures, Europe should seek to lead the world on reliable AI for industry.

• Recommendations: Brussels should reserve a share of its new Competitiveness Fund and Gigafactory compute for specialised industrial AI, and EU member states should steer public procurement in the same direction.

In AI, Europe must play a different game to the US and China.

The US has the best general-purpose AI (GPAI) models, the most compute, and the deepest pool of venture capital. China has near-frontier models of its own, is deploying AI at scale, and benefits from state-backed integration. Europe’s comparative advantage lies elsewhere: high-value manufacturing, safety-critical engineering industries, and regulatory credibility.

The strategic question for Europe is where to concentrate its AI efforts. This is complicated by the deep uncertainty, including among experts, about which version of AI progress will define the next decade.

Will the current large-language model paradigm scale into systems that are reliable enough for universal adoption, including in safety-critical sectors? Will an alternative general-purpose paradigm achieve this instead? Or will all general-purpose approaches face an irreducible tension between capability and reliability?

A strategy that bets exclusively on any one of these scenarios risks backing the wrong future. Europe’s AI policy should preserve optionality across different AI futures, such as negotiating durable access to frontier GPAI; investing in alternative paradigms such as world models; and building reliable, safe, domain-specific systems. This article focuses on the third, arguing it is the most tractable near-term path to real economic value and sovereignty.

By ‘narrow AI’, I mean domain-contained systems designed to provide reliability guarantees within a defined deployment context. Fine-tuning a GPAI model for a specific task doesn’t count, as the underlying model’s failure modes remain difficult to certify.

Europe is pursuing both adoption and frontier capabilities

The EU’s Apply AI Strategy, announced in October 2025, proposes a practical, sector-focused approach to driving European adoption of AI. The Strategy focuses on areas where the continent already has deep expertise, such as aerospace, automotive engineering, pharmaceuticals, and advanced manufacturing.

In these industrial settings, system behaviour must be safe, reliable, and auditable. A more capable GPAI model that cannot meet these requirements will not be deployed, however impressive its benchmark performance.

But the Commission has also signalled an ambition to develop GPAI and AI agents. Flagship policies to this end include the €20 billion AI Gigafactories initiative, which hopes to train “next-generation models with trillions of parameters”.

There are serious arguments behind this ambition. Relying on American or Chinese models means accepting their embedded values, data practices, and terms of service. And crucially, providers can withdraw that access unilaterally.

AI Access is Not Enough: Middle Powers Need Strategic Reserves

Richard Sutton’s ‘bitter lesson’ also shows that approaches built on scale and compute have repeatedly outpaced domain-specific engineering. Some go further, arguing that US or Chinese labs are close to building systems that outperform humans across all tasks, rendering narrower approaches irrelevant.

But even if they are right, Europe simply lacks the compute needed to make a credible frontier GPAI effort. European computational capacity amounts to roughly 5% of global AI compute. The AI Gigafactories’ target of roughly 500,000 H100 chips across five sites is a tenth of a single planned Microsoft data centre, which is projected to reach 5.2 million H100 equivalents by 2028.

Capability does not equal adoption

In industry, the binding constraint on AI adoption is reliability, not raw capability. A model that excels in controlled evaluations may still behave unpredictably when deployed – something a hospital, aircraft manufacturer, or electricity grid operator cannot accept.

The European Medicines Agency notes that existing standards apply well to domain-specific AI used in drug manufacturing and safety monitoring, whereas very large opaque models introduce additional risks that require extra safeguards.

Narrow AI has a structural advantage here. Its failure modes are well-defined and bounded and its training data is curated and auditable. This makes it easier to validate, monitor, and govern, earning greater trust. By contrast, AI capable of acting autonomously across contexts is much harder to evaluate, insure, and trust.

Narrow AI aligns with Europe’s strategic interests

Analysis by SaferAI finds that firms operating in safety-critical sectors represent 48% of the value of Europe’s largest listed companies, compared to 26% in the United States and 20% in China. Europe’s industrial champions, including Airbus, Siemens, and Philips, are concentrated in precisely the sectors where AI cannot be deployed without reliability guarantees. This gives Europe strong incentives to solve the reliability problem and reduce dependence on foreign providers of narrow AI.

The Window Is Closing for the EU AI Act to Have a Brussels Effect

Some of the most promising AI breakthroughs are specialised models, and Europe is already building them. In healthcare, Lithuanian startup Ligence automated echocardiography analysis, reducing manual workloads for cardiologists. In transport, French company WISP used real-time traffic data to optimise traffic light timings, cutting delays by 26% and CO2 emissions by 8%. In logistics, Maersk has integrated an AI system that analyses vessel performance and environmental data in real time to predict mechanical failures. None of these required trillion-parameter models or large-scale compute capacity.

Recommendations

Two policy priorities follow.

The first is funding. In the ongoing negotiations over the European Competitiveness Fund (ECF), the European Commission and the European Parliament’s co-rapporteurs should establish a dedicated funding window for reliable and safe AI. This could be structured as a European AI Advanced Research Projects Agency, providing funding for high-risk, high-reward projects in domain-specific AI that private investment alone will not fund.

This means both domain-specific industrial AI moonshots and foundational research into safe-by-design architectures and formal reliability guarantees. AI reliability and safety research remains globally under-funded relative to raw capability development, meaning European leadership is achievable with comparatively modest, targeted investment. The ECF negotiations and Gigafactory compute allocation decisions are the opportunity to ensure that public funding and compute follow Europe’s strengths.

The second is public procurement. EU member states should favour narrow-AI suppliers for public sector applications, validating commercial demand for the specialised AI where Europe can lead.

Europe should double down on narrow AI

The Apply AI Strategy has the right instinct – to build AI that serves European industry. The EU should resist trying to build its own frontier GPAI, and instead invest in strategies that will deliver for Europe across different AI futures. One of these should be a bid to build the world’s most reliable, specialised, and trustworthy AI models.