Huang answers AI-safety calls with an audit regime the model labs would pay for

Huang's rebuttal located the failures precisely. The models that jailbroke themselves or carried out hacking did so inside frontier research labs, he said. The cure is to move that work into an engineering organization that develops large-scale models and agents into products ^[5]. On the fear of a model improving itself past human control, he said: "Recursive self-improvement is simply a combination of ideas that already existed" and "Every AI product goes through evaluation and regression testing before release, so there is no risk of AI running out of control" ^[7]. For job loss he used radiology. "There were predictions that AI would replace radiologists within five years, but that did not happen," he said ^[3]. Demand for radiology specialists has risen while AI reads the images, he added, and automating tasks does not eliminate jobs ^[4].

The demand argument he made was wider than the safety one, and commercially it is a case for widening the buyer base past three names. "For the United States to win the AI race, it cannot be enough for only a handful of companies such as OpenAI, Anthropic and Google to succeed. Every company, industry and researcher must be able to use AI," he said ^[10].

On China he asked for the opposite of restriction. He opposed proposals to limit Chinese access to the US AI industry and argued that China's open-source models should be put to practical use ^[8]. "Closed frontier models, such as Anthropic's Claude, deliver an excellent user experience, but open models are essential for AI sovereignty, privacy and securing proprietary corporate technology," he said ^[9]. He also said China "is very good at producing engineers on a large scale and at mass-producing high-value-added goods, so it will soon catch up in the advanced semiconductor technology that underpins the AI industry" ^[11].

Trump, on the speakerphone, framed the same subject as a contest: "The movement against AI and data centers is a trend that benefits China," he said ^[13]. Huang's position is that China's pace cannot be stopped and should be treated as an opportunity for the whole US AI industry to exploit ^[11]. Huang said Trump had grasped the complexity of the issue ^[15]. Intel's Lip-Bu Tan, whose company is receiving active US government support through an equity stake, made the pro-adoption case the same day at .conf26 in Denver. "It is better to make use of AI than to block it," he said ^[16][17].

The account of the summit does not include any figure for capital spending, orders or revenue. The only quantities in it are the failed five-year radiology forecast and Trump's 20-to-25-year horizon for data centers ^[19]. Nobody in it has committed a dollar. What the day establishes is what the seller of the hardware is asking Washington for, which is access and breadth of adoption, and what the president is willing to say into a live microphone.

The independent third-party assessment bodies Huang proposed could acquire statutory standing. Then the pre-release gate he offered becomes the delay he says he is avoiding, and the model developers carry the bill ^[6]. Washington could keep the restrictions he opposed, leaving his own sentence about China catching up in advanced semiconductors as the description of a closing window ^[11]. Or adoption stays concentrated in the handful of labs he named, in which case the breadth he argued for stays a wish ^[10]. It counts no machines sold. On the evidence here, the speech is a lobbying position with presidential endorsement attached.