The Netherlands is building its own large language model from scratch. Not as a vanity project, but as a deliberate bet on a different kind of AI: one where control over data, training, and deployment stays within national borders.
The project, called GPT-NL, is led by TNO, the Dutch applied research organization, in partnership with SURF, the national IT cooperation for education and research, and the Netherlands Forensic Institute (NFI). The Dutch government has allocated €13.5 million through the Netherlands Enterprise Agency (RVO), on behalf of the Ministry of Economic Affairs and Climate Policy. That is a rounding error in the budgets of OpenAI or Anthropic. Yet the project’s ambition is not to match GPT-5 or Claude 4 on benchmarks. It is to prove that a public-interest AI can exist at all.
GPT-NL is trained entirely from scratch. This is the project’s most consequential design choice. By not fine-tuning an existing model like Llama 3 or Mistral, the team avoids inheriting opaque data provenance, unresolved copyright claims, or embedded personal data. TNO says it applies strict criteria: safeguarding intellectual property, removing and anonymizing personal data before training, excluding confidential information, excluding harmful content, and avoiding dataset duplication. The model’s source code is open source. Model weights are released under a controlled license, allowing the team to track usage and inform users of updates or data opt-outs.
This is not how frontier labs operate. OpenAI and Anthropic train on vast, poorly documented web crawls, then retroactively negotiate with publishers. GPT-NL inverts the model. It builds a “clean and lawful data supply chain” from the start, collaborating directly with data providers and rights holders through a Content Board. Part of any future revenue flows back to creators. The project describes this as “reciprocal” — value is shared, not extracted.
The sovereign framing is explicit. TNO states that GPT-NL gives the Netherlands “full control over the model, the data and the choices we make” and avoids dependency on non-European providers. That language echoes a growing sentiment in European capitals: that reliance on US-based AI infrastructure is a strategic vulnerability. The EU AI Act creates a regulatory framework, but it does not build models. GPT-NL is a concrete attempt to fill that gap with a model aligned to Dutch and European laws, values, and societal goals.
Product Manager Saskia Lensink and R&D Manager Frank Brinkkemper are leading the effort. A progress report is available in Dutch. The project is still in development, but the architecture is already defined: a transparent, sovereign, publicly funded model that treats data governance as a first-class design constraint rather than a compliance afterthought.
The obvious question is whether a €13.5 million model can be useful at all. Frontier models now cost billions to train. GPT-NL will not compete on general knowledge, creative writing, or code generation. It does not need to. The project’s value proposition is narrower: Dutch-language applications in government, education, healthcare, and legal services, where accuracy, provenance, and compliance with Dutch privacy law matter more than raw capability. A model that can summarize parliamentary records, assist with forensic analysis at the NFI, or power a chatbot for the Dutch tax authority does not need to beat GPT-5 on MMLU. It needs to be auditable, controllable, and legally defensible.
The controlled license on model weights is a notable departure from the pure open-source ethos that many European AI projects have adopted. TNO says this allows the team to “know who uses the model and to inform users about updates or changes, for example following a data opt-out.” It is a pragmatic compromise: transparency without losing the ability to enforce data governance terms. Whether that approach gains developer adoption, or whether it creates friction that pushes users toward unrestricted models, is an open question.
The energy and water consumption of training is also on the project’s radar. TNO says it is optimizing model size and training process with “explicit attention to energy and water consumption,” based on scientific research. This is another point of differentiation from frontier labs, which have been criticized for opaque environmental reporting.
GPT-NL is not the only sovereign LLM project in Europe. France has Mistral, Germany has Aleph Alpha, and the EU has funded OpenGPT-X. But those are private companies or research consortia with different governance structures. GPT-NL is unusual in being a direct public-sector build, funded by a ministry, with a governance model that includes a Content Board for data providers. It is a test of whether a government can act as an AI developer rather than just a regulator.
The model’s ultimate impact will depend on adoption. If Dutch public institutions, universities, and companies actually use GPT-NL for real workloads, it will validate the thesis that sovereign AI can be a practical alternative. If it sits on a shelf as a reference implementation, it will be a well-documented experiment that changed nothing.
For now, GPT-NL is the most detailed blueprint available for what a public-interest language model looks like when built from scratch with public money. The report from Lensink and Brinkkemper will be worth reading when it arrives. The question it will need to answer is not whether the model works, but whether the governance model works better than the alternatives.