Jensen Huang opens his 𝕏 account with a plea for open models

NVIDIA signs a pledge with twenty-four tech giants to protect open models. The stakes of security and independence decrypted.

For his first post on 𝕏, the head of NVIDIA chose to share a letter that his company has just co-signed. His comment boils down to one stance: open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and make sovereignty possible, as the world needs both closed and frontier-level open models.

The document, titled "Open Weights and American AI Leadership," gathers twenty-five signatures. Alongside NVIDIA are Meta, Microsoft, IBM, Mistral, Hugging Face, Mozilla, the Linux Foundation, Black Forest Labs, CrowdStrike, Palantir, Perplexity, Replit, Andreessen Horowitz, and Y Combinator. The most prominent closed-model laboratories are not included.

The text opens with a parallel to the open-source pioneers of the 1980s and posits that American leadership will be judged less by a single frontier model than by the ability to diffuse AI across all sectors. Open weights, defined as models that anyone can download, inspect, modify, and run on their own infrastructure, serve as the foundation: broader access to the AI economy, competition, and control retained by the customer, with the idea that no organization wants to find itself locked into a single vendor.

The most debatable part concerns risk. The signatories concede that an open model escapes its creator once published and that its modified versions are difficult to track or recall, before drawing the opposite conclusion of a ban: faced with attackers equipped with advanced AI, defenders would need models of comparable capability. The reasoning goes further, arguing that closed models are not secure by default, as the concentration of capabilities creates a small number of single points of failure, invisible from the outside.

The demands follow: broader access to compute for startups and researchers, investment in shared training resources (datasets, tools, evaluation frameworks), and no premature restrictions on open models. A final paragraph defends distillation as a legitimate technique, not to be confused with the illicit extraction of value from a closed model, which should be addressed by targeted legal frameworks.