France and Germany have partnered with Mistral AI and SAP SE to launch “Sovereign AI for Public Administration,” yet the initiative rests on a hardware layer dominated by US hyperscalers and NVIDIA GPUs.

The Public Administration Partnership

The “Sovereign AI for Public Administration” framework, spanning 2026 to 2030, integrates European AI solutions into state machinery to reverse a specific dependency: over 80% of the digital tools used by the French state are designed or controlled outside of Europe BMDS. Roland Lescure, the French Minister of the Economy, describes this as “reclaiming control over our technological future.”

The ambition is clear, but the physical reality is contradictory. Europe may own the application layer of these models, but the compute required to run them is largely rented via “GPU-as-a-Service” (GPUaaS) from providers like AWS, Azure, and Google, who control approximately 70% of European cloud revenue The Next Web.

This creates a structural paradox. The state claims sovereignty over the intelligence it employs while the infrastructure sustaining that intelligence remains subject to external pricing and non-European jurisdictional control. Ownership of the model is not ownership of the capability.

The technical capacity to build sovereign AI now collides with a legal constraint. The European Court of Justice (ECJ) has ruled that EU-mandated data retention and mass surveillance entail a “wide-ranging and particularly serious interference” with fundamental rights to private life and personal data protection Digital Rights Ireland. By rejecting the principle of untargeted monitoring without suspicion, the court has drawn a hard boundary around the operational deployment of AI in state security.

For the France-Germany partnership, this implies that “sovereignty” cannot be defined simply as the ability to deploy a tool. Any sovereign AI system designed for automated mass surveillance is legally untenable under the Charter of Fundamental Rights. The technical capacity to monitor the population exists. The legal authority to do so does not.

The Infrastructure Shift

The failure of the “data-sharing framework” approach, seen in Gaia-X, has forced a pivot toward physical infrastructure. Germany’s National Data Center Strategy, adopted in March 2026, formally elevates data centers to “critical infrastructure” for economic security Morgan Lewis. The government aims to double national data center capacity and quadruple AI-specific High-Performance Computing (HPC) capacity by 2030.

This shift from software frameworks to concrete and silicon reveals new bottlenecks. Limited grid connection capacity, high electricity costs, and complex permitting processes constrain the speed of deployment. Ownership of the software stack is useless without the power to run it. The physical layer is the new political frontier.

The Case for Federal Compute

The scale of frontier AI development has moved beyond the capacity of individual member states. The European Frontier AI Initiative, a joint effort by France, Germany, and the European Commission, acknowledges a steep cost barrier: by 2027, only organizations with billion-dollar budgets will have the resources for frontier AI training CFG. To circumvent this, the initiative relies on “AI Gigafactories” and a “Data Union Strategy.”

The structural logic points toward a federalized compute approach. The astronomical costs of GPU clusters and the necessity of coordinated energy infrastructure exceed the fiscal and physical limits of national strategies. Individual state efforts provide the building blocks, but they cannot reach the scale required to compete with US or Chinese labs. Integration is no longer a political preference. It is a technical requirement.

The Compute Ceiling

The transition from software frameworks to physical infrastructure marks the end of the “policy-first” era of digital sovereignty. While Germany and France can optimize their national data strategies, the billion-dollar cost of frontier AI training remains a barrier that no single member state can overcome alone. A federalized compute authority would require a level of fiscal and energy coordination that currently exceeds the EU’s treaty-based architecture.

The gap between political ambition and physical capacity is now a concrete resource problem. Scale is the only currency that matters. Acting as twenty-seven national strategies, Europe can coordinate. Acting as one, it can govern. The choice between those two is now a technical necessity, not a political debate.

Sources