NVIDIA CEO Jensen Huang said many of the most dramatic warnings about artificial intelligence were effectively invented, and argued that AI safety should be handled as an engineering discipline rather than a reason to slow innovation across the board.
Huang made the remarks on Sept. 14 at All-In Summit 2026, a live event hosted by the All-In Podcast in the United States. The discussion covered AI safety, extinction scenarios, recursive self-improvement, open-source models and AGI. The four hosts were also All-In Podcast co-founders: angel investor Jason Calacanis, former Facebook executive Chamath Palihapitiya, PayPal mafia figure David Sacks, and David Friedberg, CEO of Ohalo Genetics and The Production Board.
The conversation revolved around one central question: what kind of safety boundary AI development actually needs. Referring to recent debate triggered by Anthropic CEO Dario Amodei and calls for frontier labs to slow down, Huang said safety and innovation do not contradict each other, and incidents that have already happened should be examined through root-cause analysis, testing, process design and technical controls.
Safety and innovation should not be treated as opposites
Asked about the latest disputes around AI safety, Huang said the issue must be taken seriously, but rejected the idea that safety and technological leadership sit on opposite sides of the same tradeoff. In his view, the United States can move quickly, keep innovating, remain ahead, and still build proper safeguards.
He also commented on former Anthropic researcher Jacob Kokan, saying it took courage to speak out and that there was nothing wrong with acting as a whistleblower. But Huang drew a line between reporting concerns and making sweeping scientific claims about the future. Those claims, he said, lacked enough scientific basis because they were not built on actual scientific research.
Huang added that many frontier labs are moving from research organizations into engineering organizations, and the two are not the same. If control or management problems appear during that transition, he said, that is a separate matter. As for what Kokan specifically saw, Huang said he did not know.
On the broader wave of pessimistic AI predictions over the last several years, Huang said people should not casually assign an "extinction probability" or a "civilization destruction probability" to AI, because a lot of that material is simply made up. Giving a sensational forecast the authority of a scientist or researcher, he said, can easily push the public toward panic and is not responsible.
He cited several examples of forecasts that, in his view, did not hold up:
- Radiologists would be completely replaced by AI within five years.
- 90% of code would be generated by AI within six to 12 months.
- 50% of entry-level jobs would disappear within six to nine months.
- GPT-2 and Llama 3 were too dangerous to release.
- Half of all white-collar jobs would vanish the following year.
- An employment doomsday was close at hand.
Those forecasts should be recorded, Huang said, because looking back years later makes it easier to see which ones were right and which ones only created fear.
When asked whether AI safety should be handled more like an engineering problem, Huang said yes. If a safety incident occurs, the first step should be engineering root-cause analysis: determine what happened, identify where the failure occurred, and decide what can be put in place to stop the same issue from happening again. Those lessons then need to become policy, process and technical controls.
He said he believes companies are already doing that work, including building better sandboxes, stricter runtime environments, stronger monitoring and continuous monitoring systems. "I would even bet that these problems can be controlled and prevented," he said.
If an engineering company truly does not understand what happened or how to control it, engineers could be sent in to help, Huang said. Even so, he added that he does not think things will reach that point, because frontier labs are staffed by highly capable people who have likely already analyzed the relevant issues and taken corrective action.
On regulation, Huang said policymakers should target real problems. Right now, he said, the main issues worth watching come from frontier AI labs because they control the most compute and are working on the cutting edge. A high school student or an ordinary startup is unlikely to have enough computational power to create the same class of risks.
Those labs are trying to build companies, culture, technology and products while shifting from research into engineering, he said. That is not easy. The focus, in his view, should be on whether they are establishing mature engineering systems and whether development and testing are being carried out with adequate safety protections.
Asked whether third-party evaluators are needed, Huang said multiple assessment and audit bodies could make sense. They do not need to know everything, he said, but they do need to know which questions to ask. He compared that setup to financial auditing: one evaluator can be influenced, while several can check one another.
Recursive self-improvement is a label for existing techniques, not a mystery
Huang said the recent surge of discussion around recursive self-improvement, or RSI, has made the idea sound more exotic than it is. In his telling, RSI is mostly a new label attached to a collection of existing methods, including context handling, skills, reflection, reinforcement learning, synthetic data generation and LoRA.
Those tools can allow systems to accumulate experience while carrying out tasks and then improve further, he said. Synthetic data and reinforcement learning, for example, can strengthen a model without retraining the full foundation model, and the experience gathered in that process can later feed back into base-model training. Huang said nearly every company is already using some part of that toolkit.
He argued that the term itself makes it sound as if systems suddenly become uncontrollable, but products still need evaluation, retesting and regression testing before release. Those checks, he said, are basic engineering controls.
Frontier labs are still making the transition from research shops to engineering organizations, Huang said, but he expects them to develop better methods, knowledge, practices, tools and technologies to control, verify and evaluate these systems. Recursive self-improvement may happen inside the lab, he said, yet products that pass validation can still be released safely.
Open and closed models both matter, and China could become a major open-source contributor
On the dispute between open-source and closed-source AI, Huang said both are necessary. He said he personally uses closed frontier models too. His analogy was bottled water: water itself is free, but different product formats serve different needs.
Open models, he said, matter for sovereignty, privacy and proprietary enterprise technology. Over the last six months, roughly $400 billion in venture capital has gone into AI-native companies, and about 80% of those companies use open-source models. Without open models, he said, many startups would not be able to achieve what they are trying to build because their goals differ from those of frontier labs.
Huang said the real U.S. advantage lies in the diversity of its innovation pathways, and that winning the AI race requires open models to keep advancing. In his framing, winning does not mean just a handful of U.S. companies come out on top. It means companies, industries, researchers, teachers, students and founders across the country all get a chance to benefit from the technology. Some will use closed models, many will use open ones, and both should exist.
Asked whether China will become an important contributor to the global open-source ecosystem, Huang said China may contribute a very large portion of it. He pointed to the country’s large population of engineers and students in science and mathematics, adding that universities such as Tsinghua produce many strong graduates every year.
He cited Linux, Kubernetes and many software projects as examples where Chinese engineers already contribute. One of the defining features of open source, he said, is that once it is downloaded, it belongs to the user, who can modify it, optimize it and turn it into something of their own.
In the end, Huang said, the AI race comes down to who uses the technology best. He compared the moment to earlier industrial revolutions, where many important inventions originated in Europe but were later applied effectively by the United States. He said he hopes the next industrial revolution develops in a similar way.
Huang also said he sees China’s public discussion as more practical, with more attention on economic development and social progress than on constant debate over apocalypse or civilizational collapse. If those end-of-world predictions are true, he said, they should be addressed. But there is no need to manufacture panic. "Our job is to build it," he said.
Asked what engineers would do if AI ends up automating a large share of programming work, Huang said engineering work will still remain. He noted that engineers of his generation did not spend most of their time writing code before software became pervasive. Today’s software engineers spend a great deal of time coding, but if much of that work becomes automated, people will still do engineering. He added that his favorite keyboard key is Backspace, because the best software often means less code.
Why NVIDIA is moving deeper into the AI industrial stack
The discussion then shifted to NVIDIA’s role in the wider AI economy. Asked about jobs, reindustrialization, energy and supply chains, Huang described AI as a new industrial revolution. The sector, he said, spans models, chips, applications, data centers, buildings, electricity and power generation. His job is to watch the whole ecosystem and identify bottlenecks.
If an excellent company is constrained by one part of the chain, Huang said, that becomes something he pays attention to. The bottleneck might be supply chain, land, electricity or factories. Because NVIDIA is now operating at such scale, he said, it has to think years ahead about supply. He specifically mentioned Corning, Lumentum, TSMC and memory suppliers as companies that require long-term attention.
Asked why NVIDIA has moved into areas higher up the stack, including models and Hugging Face, even as some believe the largest long-term profits in AI will shift toward applications, Huang said NVIDIA can run nearly all of the world’s models. About a year and a half ago, he said, most discussion centered on OpenAI’s models. Now the field includes Meta’s Muse, Grok, Gemini, Anthropic and many others. A large number of frontier AI labs are built on NVIDIA’s platform.
Its strategy, Huang said, is to help everyone succeed rather than take over other companies’ businesses. NVIDIA tries to remain as low in the stack as possible and only move upward when necessary. CUDA supports a large number of frameworks, while Megatron and Megatron Core support large-scale training. The company invents the technologies that are needed, then lets the broader ecosystem develop around them and pull in more participants.
On the growing importance of regional cloud providers, Huang said they are often more flexible than hyperscalers because they can secure land, power and factories locally and have a better understanding of local demand.
He said the future will look more like a distributed enterprise network, and that more countries will treat AI as a strategic industry. Some regions, he added, are already increasing AI compute infrastructure.
When asked why NVIDIA is building its own open models in areas such as autonomous driving and biology, Huang said the answer is simple: customers need them. If NVIDIA has the ability to do the job best, he said, it will do it.
He pointed to Alpamayo, which he described as NVIDIA’s autonomous-driving model, saying it is a driving system capable of reasoning. Traditional autonomous driving has required tens of billions of hours of road data, he said, but reasoning can help systems think through similar scenarios and reduce reliance on massive amounts of real-world driving data.
Autonomous driving will not stop at passenger vehicles, Huang said. It will extend to agricultural equipment, trucks, freight vehicles and many other mobile machines. Many companies do not have a complete stack, so NVIDIA can provide one, after which customers can make the final-mile adjustments they need.
He gave the same explanation for the company’s work in biology. Huang listed ESM2, ESMFold, OpenFold, AlphaFold 2, several equivariant models and Protein Complexa, saying pharmaceutical companies genuinely need those technologies. "My starting point has always been that if someone needs the technology, we should build it," he said.
On Musk’s Terafab, AGI and domain-specific superintelligence
Asked about Elon Musk’s Terafab chip factory plan, Huang said NVIDIA knows process technology well and has spent years pushing manufacturing limits, so it can certainly discuss the subject. He added that once Musk decides to do something, it is hard to stop him from continuing, which Huang described as one of Musk’s superpowers.
When asked whether NVIDIA chips might one day be produced in Musk’s factory, Huang replied, "We can discuss it."
The conversation ended on AGI and superintelligence. If AGI is defined as reaching human-level intelligence, one host asked, has that stage already arrived? What about superintelligence? Huang’s answer was direct: "We have already reached it." He added that in some specific domains, it is fair to say the industry has already entered a phase of superintelligence.
His reason was that AI already exceeds human performance in certain tightly defined areas. Autonomous driving was one example. If a vehicle can drive better than a human, he said, then in the domain of driving it is superintelligent.
Huang said autonomous-driving systems can now reach accident rates equal to one-tenth of those of humans. Biology, he added, offers similar examples. Tasks such as synthetic protein generation and virtual screening have already reached superintelligent levels. In those specific domains, AI can carry out work beyond human capability.
Asked how it feels to stand at the edge of that technological shift, Huang said he likes it. The future, he said, is beautiful. Even if many people eventually do not need to work, he does not want to miss this era.
He closed by saying he hopes everyone can move into that future together, and that humanity will ultimately succeed together. The more important task, he said, is to encourage participation and reduce unnecessary theatrical arguments so that the entire United States can take part.