Europe is struggling to turn publicly funded AI research into lasting technology assets that can compete globally. Portugal's launch of AMALIA, the first open large language model (LLM) developed in European Portuguese, on July 1, 2026, highlights both the EU's ambition and a critical weakness in how it builds and maintains strategic AI systems.

AMALIA was developed by a consortium of Portuguese universities and research institutions with 5.5 million euros in funding from the Recovery and Resilience Plan. The model builds on EuroLLM-9B, a foundational model trained on 400 NVIDIA H100 graphics processing units (GPUs) on the MareNostrum 5 supercomputer, and was subsequently optimized for Portuguese using 256 H100 GPUs. The Portuguese government explicitly framed the launch as a step toward digital sovereignty and technological autonomy.

But the timing reveals a deeper anxiety. Just weeks earlier, the European Commission acknowledged that the EU remains structurally dependent on external suppliers for more than 80% of its digital products, services, infrastructures, and intellectual property. The Commission's Action Plan on Cybersecurity and Artificial Intelligence, released on July 7, 2026, was even more blunt: "frontier capabilities are mainly developed outside of the EU, and their availability is often determined by non-transparent, foreign-led processes".

Why Can't Europe Build AI Systems That Survive Beyond the Research Phase?

AMALIA's story illustrates a recurring problem in the European innovation ecosystem. A large language model does not become a mature, operational technology simply because its initial research objectives are met. Instead, its continued relevance depends on sustained access to computing resources, retention of specialized technical expertise, and the ability to adapt as technologies and user requirements evolve.

The challenge is organizational and financial. AMALIA was financed as a research project with a defined funding period under the Recovery and Resilience Plan. Once contractual obligations are fulfilled, the project technically concludes. But an AI system intended for continued operation requires ongoing monitoring, maintenance, and institutional support throughout its lifecycle. A further development phase extending until 2027 has been announced with additional funding, but the longer-term sustainability model remains unclear.

This gap between research completion and operational deployment is particularly acute in AI, where computational demands and technical requirements evolve rapidly. The EU's framework for supporting innovation does not naturally bridge this gap.

What Infrastructure Constraints Are Holding Europe Back?

Europe's technological dependence is rooted in physical and computational constraints that policy alone cannot quickly overcome:

• Computing Power: Council Regulation (EU) 2026/150, adopted in January 2026 to enable AI Gigafactories, states that the next generation of frontier models will require at least three to four times the number of the most advanced AI processors currently available in the most powerful AI factories. The regulation explicitly acknowledges that existing mechanisms are insufficient.
• Energy Infrastructure: Data-center electricity consumption in the EU accounts for around 3% of total demand and is projected to more than double by 2030. Construction of a transmission line takes between four and eight years, while grid-connection queues range from three to ten years.
• Investment Disparity: In 2023, venture capital invested in AI in the EU amounted to 8 billion dollars, compared to 68 billion dollars in the United States and 15 billion dollars in China. Since 2017, 73% of foundation models developed globally are US-based and 15% are Chinese-based.

These constraints are not temporary. They reflect structural differences in capital availability, energy infrastructure, and industrial capacity that cannot be resolved through a single policy initiative or funding round.

How Can Europe Strengthen Its AI Governance While Building Sovereign Systems?

The EU AI Act, which entered into force in 2024, imposes obligations on providers of high-risk AI systems, including requirements for robustness, accuracy, and cybersecurity resilience against adversarial attacks. However, the regulatory framework does not fully address the challenge of sustaining publicly developed AI systems over time.

Several practical steps could help bridge this gap:

• Long-Term Funding Mechanisms: Establish dedicated funding streams that extend beyond initial research phases to support maintenance, refinement, and operational deployment of strategically important AI systems throughout their lifecycle.
• Institutional Continuity: Create organizational structures that can persist beyond fixed research funding periods, allowing teams to retain expertise and manage ongoing technical evolution of deployed systems.
• Infrastructure Coordination: Accelerate grid-connection processes and coordinate energy infrastructure planning with AI computational capacity requirements to reduce the three to ten-year queue times currently facing data centers.
• Cross-Border Collaboration: Leverage the EuroHPC network and existing supercomputing infrastructure like MareNostrum 5 and Deucalion to distribute computational load and reduce dependence on any single national resource.

The EU AI Act also creates a regulatory framework that distinguishes between general-purpose AI models without systemic risk, like AMALIA, and high-risk systems that require more stringent oversight. This tiered approach allows smaller, regionally focused models to develop with lighter compliance burdens while maintaining safeguards for critical applications.

What Does AMALIA's Success Actually Mean for European AI Sovereignty?

AMALIA represents a significant achievement for Portugal's AI ecosystem and contributes to the EU's broader objective of strengthening technological capacity in strategically important sectors. The model was explicitly designed to preserve European Portuguese, safeguard citizens' national data from third-country jurisdictions, and be progressively integrated into public administration.

However, the project also exposes the limits of what a single national initiative can accomplish. AMALIA's development required access to world-class supercomputing infrastructure, specialized technical expertise, and sustained funding. These resources are not uniformly distributed across the EU, and smaller member states may struggle to replicate Portugal's approach independently.

The broader question is whether Europe can transform publicly supported research into technological assets capable of sustained development and operational deployment at scale. AMALIA is a proof of concept, but the EU's structural dependence on external suppliers for more than 80% of its digital infrastructure suggests that individual national projects, while valuable, are not sufficient to address the systemic challenge.

The NATO summit held in Ankara on July 7 and 8, 2026, underscored the geopolitical stakes. Although the official program scarcely mentioned AI, access to frontier AI models shaped unofficial discussions. The United States controls access to the most advanced AI capabilities and determines unilaterally which allies may obtain them, leaving European allies dependent on technology they do not govern and whose availability may be curtailed by external political decisions.

AMALIA will not solve this problem. But it demonstrates that Europe is beginning to recognize the challenge and is willing to invest in building alternatives. Whether those alternatives can be sustained, scaled, and integrated into critical infrastructure remains the defining question for European technological sovereignty in the AI era.