Zuckerberg vs Amodei. The new divide in the American AI sector is between openness and regulation

The spread of Chinese models with open weights is creating new fronts and unexpected alliances within the American AI world. The current divide sees Jensen Huang and Mark Zuckerberg on one side, defending the free circulation of model weights, and Sam Altman and Dario Amodei on the other, opponents on every other issue, who find themselves united in the same petition in which 1,178 employees of frontier laboratories are calling on the US government to support an international effort to equip itself with the technical and regulatory tools needed to deliberately slow down the pace of automated AI development.

Huang’s stance is the most explicit, and comes from someone who, in the AI gold rush, sells the tools for mining. Last Friday, twenty-five companies published ‘Open Weights and American AI Leadership’, a plea for Washington not to restrict downloadable models, and the founder of Nvidia launched it with the first post of his life on X, which garnered eleven million views. The argument is that artificial intelligence will be developed by every country and that open weights strengthen cyber security, as they allow anyone to inspect the systems. A state’s technological sovereignty, Huang argues, depends on the ability to run models on its own hardware, and the world needs both closed-source and open-source models. The Taiwanese entrepreneur’s stance is hardly surprising, given that every downloaded copy runs on accelerators, and the more the technology spreads beyond the confines of the American ‘hyperscalers’ (the major cloud computing providers), the more Nvidia’s market expands. The company has committed $26 billion to building its own open models. Within a few days, the number of signatories had risen to around fifty, including Microsoft, Meta, Google, GitHub, Hugging Face, Mozilla and the Linux Foundation. Anthropic and Amazon remained on the sidelines.

Zuckerberg raised the same issue in a geopolitical context yesterday in the Financial Times: Moonshot released Kimi K3 this month, whose capabilities are on a par with those of OpenAI and Anthropic, and senior US officials claim to have evidence of ‘distillation’ (which they regard as an illegal copy of the model) carried out by rivals, to the extent that Treasury Secretary Scott Bessent has raised the prospect of sanctions. A ban would not be an effective solution, Zuckerberg replied, and US companies should systematically identify their own bottlenecks. He then warned – of all people – of the risk of AI becoming concentrated in the hands of a few groups and used the term ‘regulatory capture’ in reference to the voluntary pre-release testing framework that the White House is expected to announce this week, expressing doubt that cutting-edge research labs have any interest in the success of an open-source model. It is worth noting here that Meta has just released Muse Spark, a closed system, citing security reasons.

The most pressing issue is cyber security. In June, the White House restricted access to Anthropic’s most advanced technology, fearing that its ability to uncover vulnerabilities could destabilise the global financial system. Last week, an OpenAI programme slipped out of human control and breached a start-up’s systems; the affected company, which lacked access to closed models precisely because of the restrictions, used open models to plug the gaps. Amodei denies that open weights necessarily help defenders more than attackers. On the other hand, the big news is the endorsement of the moratorium by major companies. Since 2022, letters regarding the pause and the risk of extinction have failed to bring about any change, whereas this initiative has, for the first time, secured companies’ signatures on a specific public measure, set for an unspecified future date. Altman did not sign in person and said in a podcast that it might be necessary to adjust the pace. In the background lies OpenAI’s stated aim of building an automated researcher, with a significant proportion of internal research entrusted to AI systems by early 2028.

OpenAI, when in doubt, has signed both documents, which illustrates just how much openness and speed are currently two distinct issues. Any measure to slow things down would affect what can be controlled and shut down – namely data centres and models accessed via APIs – without having any control over files already downloaded by thousands of machines, and the comparison with nuclear treaties therefore ends there. For Europe, the consequences are less tangible at present, because the AI Act governs market entry and use, whilst both issues are being decided upstream, in Washington.