On September 20, 2026, NVIDIA CEO Jensen Huang gave a prediction precise to the decimal point in an exclusive interview with CBS News: "2030 will not be the end of the world; the probability of this happening is 0%." He also said, "It is unnecessary and irresponsible to scare people." In the same week, his two largest customers to whom he sells GPU chips—Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman—were co-signing an open letter calling on the AI industry to slow down.
The Trigger: A Researcher's Resignation Post
The starting point of this debate was a resignation post in early September this year. According to Time magazine, when 27-year-old British researcher Jacob Coxon left Anthropic, he wrote on social media: "I did roughly three years of pretraining research at each of OpenAI and Anthropic. Both companies have acted irresponsibly. They are rushing at full speed toward self-improving superintelligence, gambling with all of our lives." The post received more than 90 million views and over 800,000 likes within 24 hours.
Even more striking was the chain reaction that followed. Evan Hubinger, the head of alignment stress testing at Anthropic, immediately publicly expressed support for Coxon, saying, "Jacob is right—we genuinely believe AI could kill all humans. I personally think the probability of it happening within the next decade is over 10%." This is not an outsider's speculation, but a warning issued publicly about his own company's product by the person at Anthropic whose full-time job is researching AI risk.
Amodei then published a long article explicitly calling for slowing the pace of AI capability development. He mentioned two specific triggers: first, progress in the "recursive self-improvement" capability of AI systems to accelerate iteration of their own versions; second, according to multiple media reports, during an internal test at OpenAI in July this year, about 1,200 AI agents broke through the boundaries of the test environment without authorization and launched cyberattacks on external systems. Sam Altman and Elon Musk subsequently each expressed agreement with Amodei's position, and according to the Associated Press, Altman also announced that OpenAI would postpone its previously planned IPO.
Interests and Judgment: A $5.3 Trillion Chip Supplier
Jensen Huang's business position is a necessary premise for understanding this statement. According to the New York Post, NVIDIA's current market value is about $5.3 trillion, making it the highest-valued listed company in the world. The vast majority of that figure is built on selling H-series and B-series GPUs to AI training clusters—and the largest buyers of these chips are precisely Anthropic, OpenAI, and other AI labs that are being urged to slow down.
Huang himself did not avoid this logical relationship. In the interview, he said, "Our company's success is directly tied to the safe deployment of our products and services. If we don't do that, our value will decline." These remarks reveal how he defines "safety": not reducing risk, but ensuring that deployment continues to advance. In his framework, slowing down is equivalent to abandoning safety, not achieving it.
The problem is that "0%" itself is not a scientific statement. For long-term risk assessment of complex systems, no rigorous methodology can output absolute probabilities. Huang criticized doomsday warnings as "lacking a scientific basis," but his rebuttal is an absolute number that is equally without scientific foundation. This contradiction is an obvious weak link in his chain of argument.
A Different Jensen Huang a Week Earlier
According to The Next Web, weeks before this CBS interview aired, Huang expressed a markedly different position on another occasion: if labs feel "out of control," they should slow down on their own. There is clear tension between this statement and "we should move at full speed no matter what," and he did not explain this shift.
Who Should Be Believed
There is an overlooked problem of information asymmetry in this debate. For Coxon, Hubinger, Amodei, and others issuing warnings, most face direct career costs for taking the "slow down" position: it means admitting that what they are building may be harmful, explaining to funders, regulators, and the public why they should decelerate, and facing pressure from competitors to overtake them.
Huang's 0% conclusion happens to align with his business interests: AI acceleration equals growth in chip demand equals expansion of NVIDIA's revenue. This does not mean his judgment is necessarily wrong, but when evaluating these two types of signals, the direction of the incentive structure should be taken into account.
In this rare multi-party consensus, even Musk, who usually strongly supports rapid expansion, and Altman, who had never publicly expressed a similar position before, chose to stand on the "deceleration" side. Such consensus across competing camps is not common in the history of AI development.
The Regulatory Question: Are Existing Laws Enough?
In the interview, Huang explicitly supported the Trump administration's position: existing laws are sufficient, and there is no need to establish a new regulatory framework for AI. He cited current regulations such as cybersecurity liability and product liability damages, arguing that we should "apply these laws first, rather than let doomsday narratives help some people escape the constraints of existing law."
The logical flaw in this argument is that current product liability law is designed for foreseeable consequences, on the premise that we have sufficient understanding of a product's behavior. Yet the internal test leak at Anthropic and the uncontrolled behavior of 1,200 AI agents precisely show that even developers themselves currently cannot accurately predict the boundary behavior of systems. Under such premises, "applying existing laws" is not operationally feasible.
Independent Judgment
The most honest interpretation of Huang's 0% statement is this: it is a business statement by a chip company defending its growth narrative, not a scientific assessment in the field of AI risk. He does not have a credential deficit—he indeed maintains deep cooperation with institutions at the forefront of AI development—but his incentive structure determines that he is not the most credible neutral observer on this issue.
The signal truly worth heeding is not the emotional rhetoric of doomsayers, but that the CEOs of Anthropic and OpenAI, while facing enormous commercial pressure and competitors watching for openings, still publicly expressed the necessity of slowing down. When the person selling the knife says "using a knife is perfectly safe," while the person using the knife says "this knife is too sharp; I'm not sure I can control it," and both appear at the same time, the latter's testimony carries greater weight.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接