Sovereign AI has become the rallying cry for governments seeking independence from Big Tech, but a closer look at Switzerland's strategy exposes a fundamental gap between the headline and the reality. While the country successfully built Apertus, an open-source AI model trained on Swiss soil, it simultaneously renewed its Microsoft licensing agreement for 140 million Swiss francs (CHF), keeping government operations dependent on U.S. technology. The contradiction highlights a critical blind spot in how nations are approaching AI sovereignty: they're winning the model layer while losing the infrastructure war underneath.

What Does Sovereign AI Actually Mean?

Sovereign AI typically refers to a nation building its own artificial intelligence systems to reduce dependence on foreign technology companies and maintain control over sensitive data. The concept gained momentum as governments worldwide grew concerned about relying on U.S. tech giants for critical infrastructure and citizen information. Switzerland's approach seemed to embody this ideal. In December 2023, the Swiss AI Initiative launched with over 800 researchers across a dozen institutions and a mandate to train a national model rather than simply fine-tune someone else's.

The Swiss government invested in Alps, a supercomputer housed in Lugano with roughly 10,000 Nvidia graphics processing units (GPUs), delivering 270 petaflops of computing power and ranking as the eighth-fastest supercomputer in the world. The machine runs on carbon-neutral power, cooled with water drawn directly from the lake. By September 2025, Switzerland shipped Apertus, a model available in two sizes: 8 billion and 70 billion parameters, trained on 15 trillion tokens across more than 1,000 languages. The model was released under an open Apache 2.0 license, meaning the weights, training recipe, and data were all published, not just a downloadable checkpoint.

Why Is Switzerland's Approach Different From Other Countries?

Switzerland's sovereignty anxiety runs deeper than policy fashion. The country has a documented history of data protection concerns. A national e-ID vote nearly failed because voters didn't trust who would hold their data, and the chief of the armed forces formally raised concerns about Microsoft dependence to the government. This context explains why Switzerland invested in building rather than buying. The legal and territorial achievements are genuine. Apertus cleared the room in advance on data transparency, European Union AI Act alignment, and respect for Swiss copyright and privacy law, addressing concerns that closed labs face regarding training data.

However, the model layer represents only one piece of a much larger infrastructure puzzle. When you examine what actually runs government operations, the picture becomes more complicated. The Swiss federal government renewed its Microsoft licenses for CHF 140 million on the reasoning that walking away from Big Tech had become too risky and too costly. In other words, Switzerland built a sovereign model while keeping the actual government desktop, email systems, and document management on Microsoft platforms.

Where Does Sovereignty Actually Live in the AI Stack?

The AI infrastructure stack consists of multiple layers, and sovereignty at one layer doesn't guarantee independence at another. Consider the components that matter for real-world deployment:

• Model Layer: Switzerland achieved genuine sovereignty here with Apertus, an open model trained on Swiss soil with published weights and training data.
• Compute and Hardware Layer: Apertus was trained on Nvidia silicon and Hewlett Packard Enterprise (HPE) hardware, both U.S. companies subject to U.S. export controls.
• Inference and Distribution Layer: The model is distributed through Amazon's SageMaker, a U.S. cloud service, meaning deployment still depends on foreign infrastructure.
• Operational Layer: Government civil servants continue using Microsoft for email, documents, and daily operations, the systems that actually run the state.

This fragmentation reveals the core problem. Switzerland planted its flag on the most photogenic peak, the model layer, while leaving the supply chain, runtime, and its own civil service in foreign hands. A single U.S. export control decision could reach the entire stack. If the U.S. restricted Nvidia chip sales or Amazon cloud access, Switzerland's sovereignty would evaporate despite owning the model weights.

How Should Governments Approach True Sovereign AI?

Real sovereignty requires thinking beyond the model and addressing the entire operational stack. Here are the critical components governments must consider:

• Compute Infrastructure: Secure domestic or allied GPU capacity that isn't subject to foreign export controls, ensuring training and inference can happen without external dependencies.
• Connectivity and Data Residency: Private networks and local data centers that keep information within national borders and prevent data egress to foreign cloud providers.
• Operational Integration: Replacing foreign software in government systems with domestic or trusted alternatives, not just building a model that sits alongside existing dependencies.
• Supply Chain Resilience: Developing or securing access to semiconductors, networking equipment, and infrastructure components that won't be subject to geopolitical restrictions.

The UAE is taking a different approach. Core42, a sovereign cloud and AI infrastructure company backed by G42, partnered with e& UAE, the country's dominant telecommunications group, to launch a GPU Sovereign AI Compute platform giving enterprise and government customers immediate access to in-country accelerators paired with carrier-grade connectivity. The platform requires no capital expenditure from customers, keeps data within the UAE, and applies zero egress fees. Beyond compute, Core42 is developing the Signature National Disaster Recovery Cloud, sovereign infrastructure designed to keep government agencies and critical-sector operators running through national-scale disruptions.

"AI deployment at scale demands infrastructure that can operate securely, compliantly, and with the resilience required for real-world execution. This partnership with e& UAE extends sovereign AI from foundational infrastructure into an operational capability, integrating compute, connectivity, and implementation into a single framework built for deployment," stated Jaafar Al Hashmi, VP of Government Relations and Market Development at Core42.

Jaafar Al Hashmi, VP of Government Relations and Market Development, Core42

The UAE's strategy bundles GPU access with premium connectivity in a single commercial package, targeting government, healthcare, financial services, energy, education, manufacturing, and industrial operations. This approach addresses the operational layer directly, not just the model layer.

What's the Real Cost of Incomplete Sovereignty?

Switzerland's experience demonstrates that incomplete sovereignty creates a false sense of independence while maintaining actual dependence. The country funded a sovereign model and kept running the actual government on Microsoft. The flagship doesn't replace the incumbent; it sits beside it while Microsoft handles email, documents, and everything that matters operationally. Renewing for another two years doesn't retire the risk of Microsoft dependence; it books it as a recurring line item.

When the chief of the armed forces is formally raising Microsoft concerns to the government, the dependency has stopped being a panel discussion abstraction and turned into a defense question. Yet Switzerland's response was to commit to another two years of Microsoft licensing while announcing its independence through Apertus. The model is impressive as open research, but as a sovereignty strategy, it leaves the leverage sitting underneath in the compute, inference, supply chain, and machines on civil servants' desks, all still owned abroad.

The broader lesson is clear: governments pursuing sovereign AI must resist the temptation to declare victory at the model layer and call it strategy. True sovereignty requires unglamorous work across the entire stack, from securing domestic compute capacity to replacing operational systems. Switzerland did the hard work most governments skip and earned the legal and territorial marks honestly. But it planted its flag on the most visible peak and left the actual infrastructure in foreign hands, then paid Microsoft CHF 140 million to keep it there.