The AI Burger Model: How to Manage AI Sovereignty Layer by Layer

Artificial intelligence is often presented as a single technology. Yet behind a conversational assistant or an intelligent agent lies a stack of components, ranging from hardware to the user interface, each raising different sovereignty challenges. This is why, during the Nexus event in Luxembourg, I presented an original framework to better understand where dependencies truly lie—and what room for manoeuvre organisations have today.

When an executive looks at an AI solution, they see a single application. In reality, it is a combination of five layers working together. When discussing sovereignty issues, we should imagine AI as a hamburger made up of several stacked layers. Today, management teams can draw inspiration from the “AI Burger” model to map and manage the technological dependencies of their AI projects and contribute to Europe’s technological sovereignty.

The five layers of AI

The “AI Burger” model consists of five layers, each relying on the one beneath it:

• “Harness”: the user-facing layer, which includes interfaces, conversational assistants and AI agents that enable interaction.
• “Middleware”: the layer responsible for managing workflows, access, resource consumption and observability.
• “Models”: the models themselves—the core of the system.
• “Inference Software”: the inference layer that runs AI models on the infrastructure.
• “Hardware”: processors, GPUs, memory and accelerators, all of which are essential for running the models.

The two dimensions of sovereignty

In my view, the sovereignty of the different layers that make up AI can no longer be addressed as a single, overarching issue. It is based on two complementary concepts:

• Sovereignty of Origin - This dimension examines the origin of the technologies:

  • Who developed them?

  • Where is the intellectual property located, and who owns it?

  • How is the supply chain structured?
     

• Operational Sovereignty - This dimension concerns the day-to-day operation of the solutions:

  • Where is the data processed?

  • Who operates the solution and can access the systems?

  • Who could, in the future, interrupt a service that has become critical?

In light of these distinctions, it is worth asking whether a foreign administration or a supplier could decide, overnight, to restrict the operation of a component that is essential to an AI solution. Recent restrictions imposed by the United States on access to certain sensitive technologies are a reminder that such scenarios are no longer merely hypothetical.

Not all layers present the same levels of dependency

Is it therefore possible today to adopt genuinely sovereign artificial intelligence solutions? This framework helps identify the main areas requiring attention more precisely.

The “Hardware” layer remains largely dominated by a small number of US companies. The GPUs required to train and run AI models are primarily manufactured outside Europe. This dependency currently represents our greatest challenge in terms of sovereignty of origin.

By contrast, the “Inference Software” layer can now rely on mature European open-source technologies, which we deploy within sovereign cloud environments such as the one we offer at DEEP, without creating any particular dependency.

Prioritising open models

Artificial intelligence models—the “Models” layer—represent another major challenge. While proprietary frontier models remain among the most powerful, open-weight models are advancing rapidly. For many use cases, their performance is now approaching that of closed models.

This development enables organisations to deploy models directly within their own infrastructure or in the cloud environment of their choice, while retaining control over their data. In this area, I believe it is now essential to see more European players emerge with the ability to develop their own models.

Although it is often overlooked in discussions about AI sovereignty, the “Middleware” layer plays a strategic role. It orchestrates exchanges between users, models and applications. In particular, it manages access rights, API keys, data flows, the use of different models and resource consumption.

Everything passes through this layer. If it is poorly secured or relies on unsuitable technology, it can become a major point of vulnerability. Here again, open solutions, combined with the right implementation and hosting partner, now make it possible to build controlled architectures within European environments.

Finally, the “Harness” layer—the application layer encompassing conversational assistants, AI agents and user interfaces—remains dominated by US companies. However, it is also developing rapidly in Europe, with the emergence of solutions capable of meeting a wide range of business requirements.

Strengthening European sovereignty

The good news is that operational sovereignty is now within reach. By relying on controlled infrastructure, open technologies and partners capable of integrating these different building blocks, organisations can deploy AI solutions while retaining control over their data and environments.

The approach we are developing at DEEP together with OVHcloud is specifically designed to address these concerns. We offer a sovereign cloud infrastructure hosting artificial intelligence solutions based on open and European technologies, enabling organisations to maintain operational control over their environments.

Sovereignty of origin, however, remains a more complex challenge, particularly at the “Hardware” layer. Europe continues to depend heavily on foreign suppliers for certain essential components, especially GPUs and, to a lesser extent, foundation models.

Achieving sovereignty of origin will require more time, investment and the emergence of new European players. Nevertheless, the foundations are already in place. The challenge now is to strengthen our ecosystem so that Europe can control the entire artificial intelligence value chain.