Europe Spends €12B on AI, but 78% of Models Are Foreign

Europe has disbursed €12 billion in artificial intelligence since 2024. It's a respectable figure, the result of a coordinated effort between Brussels, national governments, and private capital. But the practical outcome of this investment leaves a bitter taste: 78% of the language models used by European companies are still developed in the United States or China (Eurostat, 2026).

The continent faces a paradox. Never has so much been invested in its own AI. And never has it been so dependent on foreign technology.

The problem isn't a lack of money. It's the difficulty of turning capital into an ecosystem. Without competitive foundational models, Europe becomes a consumer of a technology it should be producing.

The Gap Between Investment and Outcome

The €12 billion figure includes direct public funding from the European Commission, national subsidies (such as France's €2.5 billion plan and Germany's €3 billion plan), and venture capital investments in local startups. The data comes from the latest EU Digital Strategy report, from May 2026.

But when looking at the global ranking of AI models, the picture is bleak. Among the 50 most used foundational models worldwide, listed on the Hugging Face Open LLM Leaderboard (2026), only three are European: Mistral (France), Aleph Alpha (Germany), and DeepL (Germany/Poland).

For comparison: OpenAI alone has four models in the top 20. Meta, with its Llama models, has six. China, with DeepSeek, Qwen, and Baidu, places eight.

The table below shows the distribution by origin of the 50 most popular models:

Data from the Hugging Face Rankings of June 2026. Europe, home to some of the world's largest economies, accounts for only 6% of the most used models.

Why Can't Europe Scale?

The problem isn't technical. European researchers publish cutting-edge papers at conferences like NeurIPS and ICML. The difficulty lies in turning science into a scalable product.

Three startups are trying to break this cycle. And each faces a different challenge.

Mistral AI (France) — The best-known of the trio. It has raised over €1 billion since 2023. Its Mistral Large model competes on par with GPT-4 in reasoning benchmarks. But the company struggles to secure computational capacity. European data centers are still scarce for training state-of-the-art models. The alternative was to partner with Microsoft to use the Azure cloud. In other words: to compete, Mistral depends on the infrastructure of an American competitor.

Aleph Alpha (Germany) — Focused on enterprise applications and data sovereignty. Its Luminous model is used by German public bodies and financial sector companies. The problem is scale: Aleph Alpha has fewer than 300 employees. Meanwhile, OpenAI has over 3,000. Without human capital, it's difficult to keep up with the pace of model updates.

DeepL (Germany/Poland) — Specialized in machine translation. Its language model is one of the most accurate in the world for multilingual texts. The company is profitable, which is rare in the sector. But its scope is limited. DeepL doesn't compete in the chatbot or code generation market. It has settled into a profitable niche, but one too small to be considered a general-purpose foundational model.

The European AI ecosystem resembles a constellation of bright but isolated stars. A galaxy is missing. Connections between startups, universities, and large companies that would create a virtuous cycle of innovation are lacking.

Regulation as a Double-Edged Sword

The EU AI Act comes into full effect in August 2026. It is the world's first comprehensive legislation to regulate artificial intelligence. Fines are heavy: up to €35 million or 7% of the infringing company's global turnover (EU AI Act, 2026).

The intention is good. Europe wants to be a reference in ethical and safe AI. But the side effect is worrying: regulation could further increase dependence on foreign suppliers.

European companies that need AI models for their businesses face a dilemma. Buying from American or Chinese suppliers is cheaper and faster. Developing a proprietary model requires compliance with regulations from the start, which makes the process more expensive.

The result is that many companies prefer to pay for APIs from Google, OpenAI, or Meta, even knowing that data may be processed outside Europe. Convenience trumps sovereignty.

What Does Europe Need to Turn the Tide?

The recipe is no secret. Europe needs three things it still doesn't have at scale:

1. Own computational infrastructure. Training a foundational model costs tens of millions of euros in GPUs. Most European data centers lack the installed capacity for this. Companies like France's Scaleway and Germany's Hetzner are expanding, but they are still far from the level of American hyperscalers (AWS, Azure, GCP).

2. Concentrated talent. Europe's best AI researchers are still attracted by the salaries and resources of American labs. Brain drain is a chronic problem. Initiatives like the European Lab for Learning and Intelligent Systems (ELLIS) try to retain talent, but the exodus continues.

3. Patient capital. European venture capital is still more conservative than American. Local investors want returns in 3 to 5 years. Training a foundational model takes 2 to 4 years just to reach the first viable product. The money runs out before the product is ready.

Conclusion

Europe spent €12 billion on AI and remains hostage to foreign technology. The 78% dependency figure (Eurostat, 2026) is a warning that cannot be ignored. It's not enough to just inject money. It is necessary to build an ecosystem that coherently connects capital, talent, infrastructure, and regulation.

The three startups that made it to the top 50 — Mistral, Aleph Alpha, and DeepL — show that potential exists. But they are exceptions, not the rule. As long as Europe doesn't solve the bottlenecks of computational capacity, brain drain, and lack of patient capital, the continent will continue to be a major consumer of others' AI.

The EU AI Act, coming into force in August 2026, could be a trump card or a shot in the foot. If used to create a protected market for European models, it could help. If it only serves to make local innovation more expensive, it will deepen dependency.

The money is there. The talent is too. What's missing is the right industrial policy to put the pieces together.

Related Articles

• South Korea vs. AI Chaos: The Law That Could Change the Game in Asia and Affect Brazilian Companies
• The New Chip Cold War: How the US-China Dispute is Redrawing the Global AI Map in 2026
• The AI Map of the World in May 2026: EU Retreats, Malta Innovates, Google Creates a New Mouse, and the US-China Race Reaches Boiling Point