The GPT-NL project from TNO is easy to file under national-pride LLM and ignore. The part worth your attention is the data governance, not the model.

• 🎯 Trained from scratch on purpose — no foundation weights, so no inherited copyright or PII risk to untangle later
• ⚖️ A Content Board gives rights holders a say, and revenue flows back to data providers instead of their work being scraped for free
• 🔍 Provenance as a design constraint: IP safeguarded, personal data removed before training, confidential and harmful content excluded up front
• 📊 €13.5M and three public institutions (TNO, SURF, NFI) — a bet that “compliant by construction” is cheaper than litigating provenance after the fact
• 🌐 Dutch-first, weights under a controlled license — sovereignty here means governable, not merely locally hosted
• ⚠️ The catch: a from-scratch model on a deliberately clean dataset will almost certainly trail frontier models on raw capability

That last bullet is the real tension. Most enterprises I’ve seen don’t actually want a sovereign model — they want a frontier model with GPT-NL’s data lineage, and those two things are still in conflict. The HN discussion splits on exactly this: is provable provenance worth a capability gap?

If your compliance team had to choose between a frontier model with murky training data and a weaker one with a documented, licensed corpus, which one actually ships to production — and which one does legal sign off on?