Progress has reportedly stalled on an executive order draft circulating within the Trump administration about a public-private organization to supervise the release of AI models. While the White House is still figuring out its approach to regulating AI, it is also considering a ban on Chinese open models despite outcry from Nvidia and other tech companies.
The draft has been passed around internally within the administration in recent weeks, according to The Information. The order would need support from US President Donald Trump to pass, although the timeline and proposed structure for the self-regulating body remain unclear.
The idea originates from an essay published in July on X by Google DeepMind chair Demis Hassabis calling for a self-regulating body modeled on the Financial Industry Regulatory Authority (FINRA), which oversees broker-dealers and securities brokers, an idea that has reportedly gained widespread support.
Within the administration, Treasury Secretary Scott Bessent is said to have proposed the creation of the new AI group. However, the administration's former AI and crypto czar, David Sacks, strongly opposes a FINRA-like organization, and has called for one that does not report to the government.
In his essay, Hassabis proposed that a FINRA-like body could work with federal agencies and develop assessment protocols, set thresholds for defining frontier models, and encourage model makers to use best practices such as stronger security measures and publish model cards. He said that the arrangement could have labs share their models with the organization up to 30 days before their release.
Trump originally took a more hands-off approach to AI regulation earlier in this term, but the development of powerful models like Anthropic's Mythos has shifted the administration's mindset. The US president released an executive order in June ordering the creation of a voluntary AI evaluation framework. While work on the framework is complete, many questions still surround it since many details are not open to the public.
States, meanwhile, are moving faster with AI regulations, including in California, New York, Illinois, Colorado, and Texas. However, this scattered approach leads to high legal costs and complexity, according to Forbes, while major frontier labs prefer comprehensive national standards for simpler compliance. Various laws are also in the works within Congress covering evaluations, incident reporting, risk assessments, and proposed kill switches.
The open model challenge
Open models are excluded from the already completed voluntary framework. Such tools present another challenge to the Trump administration, as Chinese open models are quickly catching up to leading US closed-source ones and are gaining widespread traction for their good-enough capabilities and lower cost. Officials are considering whether to ban their use in the US, which would follow its recent restrictions on other Chinese technologies including robots and some power equipment.
Tech leaders like Nvidia CEO Jensen Huang have advocated against a ban on Chinese models, arguing that it would harm US enterprises. Nvidia has deepened its involvement in open-weight AI, with the idea that it promotes AI proliferation more broadly across the economy, which is good for its chips business. It recently announced that it is creating optimized hardware for Chinese models such as DeepSeek and Alibaba, according to CNBC.
The chip maker is also developing open AI models itself. It agreed to acquire Hugging Face, a widely used open-source platform, for US$12.9 billion. Nvidia also entered a deal to license technology from Poolside, a startup that builds open-weight foundation models, in a move likely intended to bolster its own open AI Nemotron models.
This puts it in the awkward position of competing with some of its largest closed-source customers, including OpenAI and Anthropic, though they are also developing in-house chips of their own.
Article edited by Jack Wu
