“Distillation – learning from AI, learning from other people, and learning from other sources of knowledge – is fundamental to intelligence. We are constantly learning from one another. AI also has to learn from something.” With these words, given in an Axios interview, Jensen Huang took a clear stance in the debate roiling the AI world: distillation between models is not intellectual property theft or a competitive threat, but a learning mechanism inherent to the very idea of intelligence.

The technique, which uses the outputs of a “teacher” model to train a “student” one, has come under scrutiny especially after lawsuits by publishers and content creators against companies training LLMs. The accusation is that distillation circumvents copyright by indirectly copying protected knowledge. Huang flips the perspective: with AI generating a growing share of online content, preventing models from learning from each other would be like forbidding a student from reading a book written by another person. Blocking this exchange, he argues, doesn’t protect anyone but slows collective progress.

The defense of distillation by Nvidia’s CEO is not just a philosophical stance. It fits seamlessly into the company’s strategy. Nvidia sells the hardware AI runs on: the more models are created, distributed, and used, the more GPUs are needed. Open-source models, often born from distillation processes, broaden the market pervasively, reaching companies and developers who couldn’t afford training from scratch. And here a decisive scenario opens up for those evaluating on-premise deployment and data sovereignty.

For enterprises that choose to keep data within their own perimeter, distillation is a formidable enabler. It allows taking a large open model, distilling it on a specific domain, and obtaining a more compact system that can run on existing company hardware – without sending sensitive data to external clouds. No massive training clusters are needed: the distillation phase reduces computational load, lowering TCO and accelerating iterations. In many cases, the student model can even operate in real-time on edge servers, preserving GDPR compliance and data confidentiality.

Huang’s stance signals a structural shift. If the industry embraces the idea that knowledge flows between models as it does between people, an ecosystem consolidates where on-premise is no longer a niche but a standard mode to refine and truly own the intelligence one uses. Those who lose out are the providers of closed models that built their advantage on exclusive access to data and weights: distillation erodes that moat, putting high-level capabilities within reach of anyone with a good vertical dataset and a handful of GPUs. For the same reason, developers, software houses, and chipmakers like Nvidia gain, seeing workloads multiply.

Of course, the legal knot remains tangled. Regulators could still tighten the screws, strengthening copyright rules. But Huang’s reasoning aims to normalize distillation, shifting the discussion from conflict to the kind of collaboration forced by the very workings of the web. After all, artificial intelligence is already shaping the corpus of public knowledge: preventing it from drawing on it would mean condemning it to ignorance, or to dependence on a few proprietary datasets. A prospect that, for anyone designing AI infrastructure today, sounds more like a brake than a protection.