Jensen Huang Makes the Case for Open AI Models… and for Our Sovereignty

Jensen Huang had never tweeted. The head of NVIDIA, the one man who by himself accounts for a meaningful slice of the S&P 500, had spent years leaving it to his counterparts at Intel and AMD to work the platform. And for his very first post, published this Friday, July 24, he talks about neither GeForce nor data centers. He shares an open letter addressed to American policymakers, titled “Open Weights and American AI Leadership,” signed by roughly twenty-five organizations, and he sums it up in a single line: “The world needs both frontier closed models and frontier open models.”

Let me be blunt: full support. Jensen is right, the letter is right, and the fact that it was written to defend American leadership changes nothing about that. Quite the opposite. That is precisely what makes it interesting for us Europeans. Let me explain.

What the letter says, in substance

The text opens on a parallel I have been drawing for years: open-source software. In the 1980s, the prevailing belief held that software would advance only if the companies that wrote it kept tight control over their code. The open-source movement proved the opposite, to the point that the infrastructure of the internet, the systems of the largest technology companies, and even the critical missions of U.S. federal agencies now run on code that anyone can read.

The letter carries that reasoning over to AI. Open-weight models, the ones anyone can download, inspect, modify, and run on their own infrastructure, widen access to the AI economy: a small business, a university, a hospital need not train a frontier model or pay frontier-model prices for every task. They keep competition alive at every layer, from the silicon up to the application. And they hand organizations back control over their data, their deployments, and the value they create, with no lock-in to a single provider.

The most courageous passage in the text is the one on security. The signatories turn the usual argument on its head: no, concentrating advanced capabilities behind a handful of closed models is not inherently safe. It actually manufactures single points of failure, opaque and impossible to verify from the outside. A few months ago I documented the seven families of structural flaws that every large closed model shares, vulnerabilities that let an attacker siphon off conversations without the user’s knowledge and that are not specific to any one provider but common to the whole industry: the industry knows it, and this letter has the merit of stating it plainly. The transparency that proved itself in software applies just as well to a model’s weights. You do not audit a black box, you trust it. And in security, trust is not a method.

One last point, technical but crucial: distillation. The letter does not merely defend it, it draws a line. On one side, distillation as a legitimate technique, the practice of using one model’s outputs to improve, evaluate, or validate another, heir to a long tradition of learning from what exists in order to go beyond it. On the other, the unlawful extraction of value from a closed model, which raises real concerns but belongs to ordinary law, to be handled through targeted legal and commercial frameworks rather than a blanket ban on the technique itself. This distinction is no matter of vocabulary. At the very moment the U.S. Treasury is brandishing the threat of sanctions against “industrial-scale distillation attacks” attributed to Chinese labs, the signatories refuse to let the method be conflated with the abuse. They ask that the theft be targeted, not the tool. Put another way: you do not ban the knife because stabbings exist.

The absentees say more than the signatories

Look at the list: NVIDIA, Meta, Microsoft, IBM, Mozilla, the Linux Foundation, Andreessen Horowitz, Palantir, Perplexity. And two sources of French pride, Mistral and Hugging Face, which does not surprise me coming from players whose strategic choices I have praised before.

Now look at who is missing: OpenAI and Anthropic. The two labs whose entire business model rests on the rent of the closed model, reachable only through an API, billed by the token, hosted on their own servers. I am not imputing motives to them; their safety arguments exist and deserve debate. But let us simply note that the line between signatories and absentees maps perfectly onto the line between those who make their living from openness and those who make theirs from the tollbooth. When a position “of principle” overlaps this neatly with a commercial interest, the principle at least deserves to be questioned.

Jensen is no philanthropist, and so what?

Let us be honest in the other direction too: NVIDIA is not signing out of idealism either. A world of open models is a world where everyone buys their own GPUs to run inference at home rather than renting OpenAI’s servers. Every set of weights downloaded from Hugging Face is a sales pitch for an accelerator card. Jensen is preaching to his own choir, obviously.

But that is exactly what makes the alignment interesting. For once, the seller’s commercial interest coincides with the user’s sovereign interest. I laid out the true cost of local AI, its three bills: the machine, the watts, and the jurisdiction. Open weights are the necessary condition for that third bill, the only one not denominated in euros, to stop being paid in silence to a provider subject to the Cloud Act. That an American industrialist has a financial stake in making this possible does not keep me up at night. Objective alliances make the best policy.

An American letter that argues, in spite of itself, for everyone else

Here is the delicious paradox. The title speaks of American leadership, and the conclusion calls on the United States to build this open ecosystem. But reread Jensen’s tweet: AI will be “built by every country,” and open models “enable sovereignty.” That word, sovereignty, in the very first tweet of NVIDIA’s chief, is no accident.

Because an open weight, by definition, knows no borders. The same file that spreads AI through the factories of Ohio lets a consultant in the Vercors run his inference on his own machine, lets a French ministry deploy a model within its own walls, lets a European company build its automation loops without handing its keys to a proprietary product. I wrote back in April that local open source has become the only serious path to independence: this letter confirms it, and the confirmation comes from the very heart of the American ecosystem.

America is debating for itself, and that is its right. But every argument in this letter counts double for us. If open weights protect Americans from concentration among three American providers, what does that say about what they offer to those who, on top of the concentration, also bear the jurisdiction?

My take

One worry remains, one I have voiced before: an open ecosystem lives only if people keep reading, understanding, and maintaining what they run. Downloading weights is no more sufficient than forking a repository. Sovereignty is a practice, not a file.

But the precondition for that practice is that the file exists and stays legal. That is exactly what is at stake in Washington right now, and it is why this letter matters. The day open weights are restricted in the name of American national security, our technological autonomy would revert to what it was in 2022: a monthly subscription to Silicon Valley, cancelable by them, never by us.

So yes, full support. When the first tweet from the most powerful man in global tech defends the right to run artificial intelligence at home, you do not spoil the moment.