Some people join X by posting a photograph of their breakfast. Others announce that they are “excited to connect.” Jensen Huang, founder and CEO of NVIDIA, chose a different entrance: a three-page manifesto about open AI models, American technological leadership, cybersecurity, national sovereignty and the proper legal treatment of model distillation.
It was less “Hello, world” than “Hello, policymakers.”
That choice matters. Huang is not merely the head of a successful semiconductor company. He has become something closer to the chief architect of the physical economy underlying artificial intelligence. OpenAI, Anthropic, Google, Meta and xAI may disagree about how intelligence should be built, controlled and sold, but their arguments generally take place on NVIDIA hardware. Huang supplies the picks and shovels, certainly—but also the railway, the power tools and, increasingly, the municipal planning department.
His first post argues that the world needs both closed frontier models and open ones: models whose weights can be downloaded, modified and run on infrastructure controlled by their users. The accompanying letter, signed by 25 organisations, presents open weights as a foundation for competition, security, innovation and sovereignty. The signatories range from Meta, Microsoft and Mistral to Hugging Face, Mozilla, Palantir, CrowdStrike, IBM, ServiceNow, Andreessen Horowitz and Y Combinator.
It is an impressively varied dinner party. The Linux Foundation is seated near the venture capitalists; cybersecurity companies are passing the potatoes to model developers; Palantir and Mozilla have somehow found themselves at the same table. What unites them is not a shared ideology but a shared concern: that the AI economy could become dependent on a handful of closed models operated by a handful of companies through a handful of cloud APIs.
That would be an unusually concentrated arrangement, even by the standards of modern technology.
Open weights offer a different architecture. A hospital could operate a model within its own secure environment. A manufacturer could adapt one to its machinery. A government could build a national-language model without surrendering sensitive data to a foreign provider. A small company could use a specialised model for routine work rather than renting the cognitive equivalent of a particle accelerator every time it needed to classify an invoice.
The point is not that everybody should train a frontier model. Almost nobody can afford to. The point is that more people should be able to possess, alter and operate the resulting technology.
This is where Huang’s philosophical argument and NVIDIA’s commercial interests perform an elegant little dance. NVIDIA benefits from a world containing many models, many providers and many forms of deployment. If one closed model were to become the universal intelligence utility, much of the industry’s power would accumulate inside that company and its chosen cloud. But if every bank, factory, university and country becomes an AI operator, demand spreads across data centres, workstations, edge devices and sovereign infrastructure.
In other words, open models turn AI from a service one visits into machinery one owns. NVIDIA happens to be exceptionally good at selling machinery.
This does not make Huang’s argument insincere. Commercial interest and public benefit are not mutually exclusive; civilisation has occasionally advanced because someone found a profitable way to sell useful things. But it does mean we should read the letter as both a vision and a lobbying document. When it calls for broader access to computing capacity, NVIDIA presumably does not object to the possibility that much of that capacity will contain NVIDIA chips.
The most politically revealing section concerns distillation: the process of using one model’s outputs to train or improve another. The letter warns policymakers not to confuse this general technique with unlawful extraction from proprietary systems. Misappropriation should be addressed through targeted legal and commercial remedies, it argues, rather than broad restrictions on model-development methods.
That is a deceptively technical paragraph with enormous consequences. Whoever controls what models may learn from other models can influence whether AI develops like an open scientific ecosystem or like a collection of fortified estates. Too little protection may reward imitation and theft; too much may allow incumbents to declare intellectual ownership over machine-generated knowledge. The law must somehow distinguish education from burglary when both involve asking a computer millions of questions.
The safety argument is equally provocative. Closed-model companies often suggest that restricting access makes powerful AI safer. Huang’s coalition replies that concentration creates its own dangers: opaque failures, single points of control and security vulnerabilities outsiders cannot investigate. Open models allow independent researchers and defenders to examine behaviour, conduct red-team testing and develop safeguards.
There is truth on both sides. Openness enables scrutiny, but it also enables modification. Once model weights are released, they cannot be recalled, and safety mechanisms may be removed. Moreover, comparing open-weight AI to open-source software is slightly too convenient. A programmer can read source code; nobody meaningfully “reads” several hundred billion floating-point numbers over coffee. Transparency of possession is not necessarily transparency of understanding.
The deeper issue, however, is sovereignty. Huang writes that AI will be built by every country. This is not simply a prediction that national chatbots will proliferate. It implies that artificial intelligence is becoming infrastructure on the order of energy, telecommunications and money. Nations will be reluctant to outsource such infrastructure completely, especially when it mediates education, administration, industrial production, defence and scientific research.
Open models therefore become instruments of geopolitical influence. If the models adopted around the world are American, they pull countries toward an American ecosystem of chips, software, standards and institutions. If they are Chinese, the gravitational field points elsewhere. “Open” does not mean geopolitically neutral. Sometimes the most effective way to extend an empire is to give away the map.
And then there is X itself.
Huang’s account was created in June but remained silent until this post. The first message was neither a product advertisement nor an inspirational anecdote about washing dishes at Denny’s. It was a coordinated intervention in public policy. That suggests the account is not primarily a new hobby. It is infrastructure too: a direct channel to politicians, founders, researchers, markets and the increasingly excitable court of technological opinion.
The obvious temptation is to interpret his arrival as a gesture toward Elon Musk. Musk owns X, and xAI is an important NVIDIA customer. But Huang neither mentioned Musk nor included xAI among the letter’s signatories. Joining the platform is good diplomacy; it is not necessarily a declaration of allegiance.
The larger signal is more consequential. NVIDIA has grown from a component manufacturer into an institution powerful enough to shape the rules of the industry it supplies. Huang is no longer content merely to describe the future from a keynote stage. He intends to participate in the political struggle over who may build it, who may own it and who will provide the machines.
The leather jacket has logged on. It appears to have brought a policy agenda.
No comments yet