Why Belgium needs a national, sovereign artificial intelligence centre

Digital sovereignty is sometimes described as a choice between making everything in Belgium and depending entirely on other countries. That choice does not exist. No country the size of Belgium will produce on its own the processors, models, software and energy that modern artificial intelligence requires.

The real question is more concrete: when AI enters a hospital, an administration, an energy company or a justice department, do we retain the ability to understand the system, check its results, protect the data and change supplier if necessary?

A national, sovereign AI centre should be designed to answer that question, not to add a public logo to a building full of servers.

What "sovereign" should mean

A sovereign capability makes it possible to choose where data is processed, who can access the models, which dependencies are acceptable and how the service continues if there is a disruption. It also gives a real right to say no: no to a use that has not been sufficiently audited, no to data that cannot be protected, no to a contract that locks the administration into a single supplier.

This requires four forms of autonomy:

• technical autonomy, to deploy, evaluate and move systems;
• legal autonomy, to apply Belgian and European law without depending on vague commercial commitments;
• operational autonomy, to keep a service running and respond to an incident;
• intellectual autonomy, to have experts able to challenge results and architectural choices.

Sovereignty is therefore not isolation. It is built with European partners, open standards and private suppliers, but without giving up the skills that make it possible to stay in control.

Belgium's AI Factory Antenna is already changing the situation

Belgium is developing an AI Factory Antenna connected to the European EuroHPC network and is becoming a partner of LUMI AI, with additional access to the capacity of JUPITER in Germany. In 2026, the federal government announced an investment of 4.3 million euros in this Antenna, as part of funding that also involves the federated entities and the European Union.

This step matters. It prevents access to high-end computing from being reserved for very large groups or teams already well connected to European networks. A Belgian SME, university or public service will be able to be steered towards suitable infrastructure without having to understand the entire high-performance computing ecosystem on its own.

But an Antenna remains an access point. It does not automatically solve data preparation, recruitment, cybersecurity, public procurement or the move from a prototype to a reliable service.

The gap between a demonstration and a public service

Building an AI demonstration sometimes takes a few weeks. Running it for five years is another story.

You have to track model versions, document the data, measure errors, manage rights, handle complaints, monitor drift and plan for what happens when the supplier changes its terms. A small administration or SME cannot build a complete team for every project on its own.

This is where a national centre is useful: pooling what should not be rebuilt twelve times over.

A hospital testing a medical summarisation tool and a municipality automating the sorting of its mail do not do the same job. Yet they need common building blocks: a secure environment, logging, evaluation methods, retention rules, access control and an incident procedure.

The concrete missions of a national centre

### Providing access to compute without reserving AI for insiders

The centre should support teams from defining the need through to choosing the infrastructure. Some projects need a supercomputer; many do not. Good guidance avoids both under-sizing and the spectacular purchase of unused capacity.

### Creating secure environments for sensitive data

Health, mobility or justice data should not be copied into a central warehouse by default. The centre can provide isolated spaces, pseudonymisation techniques and governance models while leaving responsibility for the data with its legitimate holder.

### Pooling evaluation

A model that performs well in English may fail in administrative French, legal Dutch or the German phrasing used in Belgium. Shared test benches would make it possible to compare systems on Belgian cases, including bias, hallucinations, accessibility and resistance to manipulation.

### Supporting open models and interoperability

The centre should not impose a single national model. It should make it possible to test open, European and commercial models against transparent criteria, with the ability to transfer data and processes from one environment to another.

### Organising the response to AI incidents

What should be done if an automated agent discloses data? If a dependency is compromised? If a new version abruptly changes the results? A national centre can maintain procedures, forensic expertise and coordination capacity accessible to organisations that do not have them.

Belgian data is a working asset, not free fuel

Belgium holds highly valuable data in health, mobility, energy, research and multilingual administration. Its usefulness does not justify centralising it without limits.

The centre should favour approaches in which the analysis comes to the data rather than the other way round, with temporary, traceable and proportionate access. Citizens must know for what purpose their data is used, how to challenge a decision and how long the results are kept.

Multilingualism is an asset here. By seriously evaluating models in the three national languages, Belgium can produce methods useful to other European countries facing comparable linguistic realities.

A procurement policy that avoids lock-in

Dependence often begins in the contract. An administration adopts a convenient interface, then discovers that it cannot export the data, reproduce the results or change model without rewriting the entire service.

A national centre can provide common clauses: portability, right of audit, data location, incident notification, log retention, documentation of updates and exit conditions. These requirements carry more weight when they are shared by several public bodies.

Governance matters more than the address

The centre can be distributed across universities, computing infrastructure, research centres and public services. Belgium already has strong expertise; the challenge is to make it accessible and to decide on priorities together.

Governance should bring together the federal level, the regions and communities, universities, SMEs, public services and civil society. It must publish its access criteria, budgets, results and conflicts of interest.

A scientific board is not enough. It also needs representatives of the professions that will actually use the systems, fundamental rights specialists and people able to express the experience of the citizens affected.

Measuring usefulness rather than the number of projects

A centre could easily announce dozens of pilots. Success is measured elsewhere:

• how many services have made the move from prototype to safe operation;
• how many small organisations have actually gained access;
• how many proprietary dependencies have been avoided;
• how many incidents have been detected or contained;
• how many evaluations have been published and reused;
• what concrete benefits have been observed for citizens and public servants.

Official sources

• Federal government — investment in the Belgian AI Antenna (FR)
• European Commission — network of AI Factory Antennas
• European Commission — AI Factories and BE-AIFA
• BOSA — national convergence plan for artificial intelligence