Jensen Huang says $400 billion of venture funding went into AI-native companies in the last six months, and that 80% of them build on open models.
The week's argument has been about whether the frontier labs should slow down. Huang's position is that the labs are not the industry, the open-model builders are, and the extinction numbers being attached to the debate were invented.
"We have to take accountability. We have to take account for all of the stupid predictions that were made, right?"
Huang runs the company every frontier lab trains on, and was interrupted mid-answer by a phone call from the President of the United States, which the hosts put on the venue speakers.
The full interview is covered here so you can skip it. 47 minutes of audio, 21 minutes of reading.
Here are the 15 arguments that matter.
👤 Guest: Jensen Huang, Founder, President and CEO of Nvidia
🎙️ Hosts: Chamath Palihapitiya of Social Capital, Jason Calacanis of LAUNCH, David Sacks of Craft Ventures and David Friedberg of The Production Board, recording in front of a live audience at the All-In Summit
👥 Also on: Donald Trump, President of the United States, who called Huang mid-interview and was put on the room's speakers
📰 Published: 14 September 2026 on YouTube (All-In Podcast)
🔴 YouTube | 🟣 Apple Podcasts | ⏱️ 47 min | ✅ Time saved: 26 min
Key Takeaways
Safety and leadership are not a trade-off, but quantified extinction forecasts are invented and should not be published
His list of predictions that failed: radiology in five years, 90% of code in six to twelve months, half of entry-level jobs
Open models are where the money actually went — 80% of the AI-native companies funded in the last six months use them
$400 billion of venture funding in six months, on his figure
Most of the world's open-source contribution now comes from China, and he does not think that matters
Once you download it, you fork it and it is yours, the same way the West took Linux
Recursive self-improvement is not a runaway, because a product still has to pass evaluation before release
President Trump called in mid-interview and told the room the AI scare is a hoax
"The robots are not going to be taking over the world."
Nvidia's strategy is to go up the stack only as far as it has to and stay as low as it can
He says he is uncompetitive by temperament and would be happy with five hyperscalers
He thinks we are already at general intelligence, and at superintelligence inside narrow domains
A self-driving car with a tenth of the human accident rate is his example
1. Dario's Essay, Unpacked
The opening question was about Dario Amodei's weekend essay and, more than that, about how quickly the other frontier labs lined up behind it.
He separated the essay into parts and took the safety part seriously. "Safety is paramount." In his framing the choice people are being offered is a false one: a company can innovate quickly, execute quickly, America can lead, and it can be done safely
He treated the whistleblower as a serious matter and praised the person. He said Coxon had great courage to put out his concerns, and that whenever there is a whistleblower it has to be taken seriously
His objection is to what he says was bundled in with it. The scientific prediction about the future, in his view, is not grounded on science even though a scientist expressed it. He drew the line there — whistleblowing is fine, forecasting is not
He allowed the possibility that something real is happening inside a lab. If the whistleblower saw a company out of control, that is a different topic and a question of how government should deal with it. His candidate explanation is the transition every lab is making from research to engineering, which he suggested may be being handled clumsily
What he objected to structurally is the packaging. In his words, all of a sudden regulation and everything else is covered in one blog
2. The 10% Is Made Up
A host described the problem of explaining the claim to a parent: "What does that mean, 10% of extinction?" Nobody, he said, knows how to explain to an average person how that is even possible.
Huang's answer was that it should not be explained because it was invented. He said these are well-educated people working in a lab, and that the combination of those credentials with that prediction is alarming and troubling and should not be done. "It's irresponsible."
He then read out a list of predictions that did not happen. Radiology would be completely taken over by AI within five years and there would be no radiologists left. "We need more radiologists than ever in the world." What did happen is that AI automated scan reading, which he called great
The rest of the list came fast. 90% of code generated by AI within six to twelve months of a forecast made last year; 50% of entry-level jobs wiped out within six to nine months; GPT-2 too unsafe to release; Llama 3 too unsafe to release; half of white-collar jobs gone next year
His demand is a scoreboard. "We have to take accountability. We have to take account for all of the stupid predictions that were made, right?" He said people do keep track, and that those predictions are inconsistent with America winning the AI race
A host connected it to the pandemic-era argument over expertise — that the trust-the-experts movement and the look-at-the-track-record movement are now at war, and that this is coming from inside the labs building the thing
Huang declined to attribute a motive and complimented the companies instead. He called them some of the most consequential companies in history with extraordinary engineers and researchers, said he works closely with them, and called it unfortunate that these conversations have to happen in public
3. Build Companies in Silence
His prescription for the labs is a communications policy, not a regulation. "And I think that these companies really ought to be built the way that we used to build companies, which is in silence, right?"
Nobody at Nvidia speaks for Nvidia, and he says people agreed to that when they joined. They were told this is how you behave here, and in his account the employee base is happy with the trade — a consistent, stable company whose core values are about taking care of families and creating the conditions for people to do their life's work
The policy extends to politics, explicitly. "And so the discourse about race and religion and politics and all of that stuff we tell people do it outside the company. It's not for us." He described Nvidia as apolitical and bipartisan, wanting America to succeed under whatever government is in place
The company's own posture is the same as the advice. He said Nvidia does its work as quietly as it can and contributes to everybody else's success, and that he is proud of that
4. Regulate Actual Problems
Asked where he lands between Satya Nadella's measurement-and-standards position and Demis Hassabis's proposal for a FINRA-like body, Huang answered with a test rather than a structure.
His rule is that regulation should attach to incidents that happened. "You know, regulation should solve actual problems."
All of the actual problems so far, on his account, came from the labs — and he defended why. They have the most compute and they are working on frontier problems, so it is sensible that the most danger comes from there
The corollary is that the rest of the field cannot produce these incidents. "It is unlikely that a high school student did something because they just simply won't have enough compute, right?" The same goes for a startup, and in his framing everybody on the planet except the frontier labs is compute-constrained
He described the labs as doing several very hard things at once — building the company, the culture, the technology, the engineering and the products simultaneously — which he said makes a certain amount of hair on fire understandable
His remedy is an engineering one: root cause each incident. He referred to four incidents at one lab and one large incident at another, and said the first step is to establish what happened, what could have been done differently, and what gets institutionalized so it does not recur
He would bet money it is all within their control, and named the tooling. Sandboxes, runtimes, continuous monitors. The alternative — a lab concluding it cannot explain what happened and asking society for help — he called unlikely. "I think they have extraordinary people. They got this handled."
5. RSI Has a Release Gate
A host raised a Chinese lab going after recursive self-improvement, and David Friedberg supplied the detail: the founder of Z.ai, the maker of GLM, had just raised $5 billion and named AI that trains the next AI as a priority.
Huang's first move was to deflate the acronym into its parts. He described recursive self-improvement as a combination of existing ideas — in-context learning, skills, reflection, reinforcement learning and synthetic data generation — all of which make a system better at a problem over time
He walked through the weights question specifically. Low-rank adaptation lets a model improve without training the base model, and synthetic data and reinforcement learning can enhance it; later the base model can be retrained with all of that experience
His verdict is that it is ordinary and universal. Using the technology to improve productivity at all kinds of tasks, including building AI, is a logical idea, and he is certain everybody is doing it to some degree
What he objects to is the phrase being weaponized. He said it is being used to give the impression that the technology is going to spiral out of control
Asked flatly whether he believes it, he said no, and gave the reason. "And the reason for that is because you could RSI all day long inside your company, but when you release a product, you've got to evaluate it, don't you?" You test it again, you make sure there is no regression
His forecast is that control improves as labs become engineering organizations. Methods, knowledge, practice, tools, verification and evaluations — all of which, he argued, let recursive self-improvement happen inside the company and good products come out of it
6. Models Are Bottled Water
He rejected the framing that it is a contest. "The world needs both closed models and open models" — and he is a heavy user of the closed ones, saying he used four of them over the weekend and that they work terrifically
His analogy for closed models is bottled water, and the joke is that the free stuff is fine too. "I don't know if I've told you guys, but water is free." He used a lot of free water taking a shower that morning. The point is to use the right water in the right place, the same way you would treat electricity or any other commodity
The reasons to want an open model are sovereignty, privacy and proprietary technology, and after that he went to the money
The figure he used to settle it: "The facts are in the last 6 months $400 billion of venture funding went into AI native companies. 80% of them use open models." Without open models, he asked, how could they build their dream
His argument is that different dreams need different starting points. A startup's dream will not be the same as a frontier lab's, and America's strength is the number of ways it can innovate
He reframed what winning the race would even mean. "It's not about a few technology companies winning the AI race. It's about every company in America." His list ran through every company, every industry, every researcher, every teacher, every student, every startup
7. The Race Is Exploitation
Asked whether it matters if the open models come from China, he started with the supply. He said probably the vast majority of the world's contribution to open source today comes from China, because there are simply more engineers, produced in volume through universities like Tsinghua — which he named as one of America's disadvantages
His answer to the geopolitics is that downloading is appropriation. The West already downloads Linux and Kubernetes and a lot of software touched by Chinese engineers. "And once you download it, it's yours. We fork it. We improve it. We make it ours."
He then redefined the race itself. "My point is the race is really about who exploits the technology best."
His historical case is the last industrial revolution. Maxwell, Volta and Ampère were not American; the inventions came out of Europe, and America took advantage of them socially better than anyone else. He said he wants this generation to go the same way
Asked why the Chinese message is landing, he pointed at its content rather than its distribution. "The narrative is much more practical." Nobody there is talking about the end of this and cataclysmic that; they see AI as a technology that advances their economy and their society
His frustration is with where the energy goes. If the danger were true, he said, we should be doing something about it rather than worrying people who can do nothing about it. "It's our job to build it, right?"
A host pushed harder, asking whether this is fear of the frontier — a place humans have never been, so it is easy to tell everyone to be frightened of it
8. Engineering After Typing
Huang's answer to why smart people get this wrong was generational, and he used his own career. When he graduated as an engineer he did not do much typing, because his generation had to build the computers that made software possible
Today's engineer is handed a laptop and a chair and types all day, which he said is what the job now is. His inference is that there was engineering before typing and there will be engineering after it, with a mountain of work that is not typing
He was explicit that he means coding, not literal typing, and said he tells Nvidia's software engineers they are just typing, and has done for years, for fun
The habit he actually wants from them: "And I tell them, my favorite key is backspace. And the reason for that is because the best software is the smallest software."
9. The President Calls In
Huang was about to start a teardown of Nvidia's own stack when his phone rang. He took it on stage, told the caller he was in front of a few thousand people and that they had been talking about him, and the hosts got a microphone to the handset.
Trump's opening line was at Huang's expense. He said Huang can develop the most complex computer chip in the world that nobody can copy for ten years and cannot work out how to put him on speaker
His verdict on the week's AI warnings was a single word, repeated. "And I'm telling you, it's all a hoax." He said the happiest group about it is China, and that the people raising the alarm are playing into the hands of people who do not want it to happen — who he said could be political people, or could be China
On data centers he was unambiguous. They make people wealthy and states wealthy, and he called them the oil of the next 20 to 25 years. He complained that Google wants to build a big one in Finland because it could not get permitting, and said he is not happy about that
He ruled out the scenario directly. "The robots are not going to be taking over the world." He cited an uncle who he said was among the best professors at MIT for 41 or 42 years, and claimed some genetic strength from it, then added that he also has common sense about AI
He allowed a caveat and immediately bounded it. Things have to be done prudently, but that does not mean stopping the industry
His formulation of the stakes: "It's bigger than the internet. And whoever wins AI wins." He said communities that were dying now have data centers and are wealthy
His closing claim was an investment number. "We have 20 trillion dollars of investment coming into the country" — against, in his telling, much less than $1 trillion over four years under his predecessor, and this in one year
The hosts were not certain it was real. One said he thought it was a bit at first. Another described a previous occasion when Trump, going through a dinner list, was told Huang was on vacation and said to get him on the phone anyway
Huang's first words after hanging up were about the format. He noted that the hard thing about being on a call with the President is saying something
10. Auditors, Not a Pause
Asked why Trump sees through something polling, in a host's figure, at minus 80, Huang said he was not sure and that the thing is complicated.
He argued the case against AI has already changed foundations once. It was first anchored on national security, which he said was recently blown to bits, and is now anchored on safety
His safety proposal is third-party evaluation, and the model is financial audit. If AI is to be safe, the labs building it have to be in control and there have to be good tests for them, with third-party evaluators available
He spelled out why auditors work even when they know less than the company. They do not have to be as expert as the business is in its own business; they have to ask the right questions
He agreed with the point that there should be more than one. Multiple evaluators, as with multiple auditors, so that no single company becomes captured or influenced
His summary is a standard rather than a brake. He said what is being built is extraordinary, these are extraordinary companies, and they ought to be held to extraordinary standards — and that they want to be
11. AI Is Creating the Jobs
The thing he says he would have told the President is a jobs argument. "AI is creating an enormous number of jobs." He said the thing Trump wanted most at the start of the administration, and in their first meeting, was to create jobs in America and to re-industrialize
The energy point is a precondition, not a talking point. "Without energy, there's no industrial growth."
He tied the $400 billion figure back to employment. Six months of venture financing into AI created a great many jobs, software jobs among them, and created enormous demand for compute and therefore for data centers
He raised the community backlash himself, and named who raised it with him. He said he had been talking to Governor Abbott of Texas, who wants the industry to be empathetic toward small communities as data centers are built across America, and to be better listeners
12. Nvidia as the Bank of AI
A host put it to him that Nvidia has effectively had to become the bank of AI to get the ecosystem moving, citing the Cloverleaf land-power-shell arrangement and the financing structure built with BlackRock and Goldman Sachs.
His frame for the capital allocation is industrial, not financial. This is a new industrial revolution, and like electricity and the internet it requires manufacturing. With AI, in his phrasing, we tap into the ether and can ask it anything
Producing intelligence is a production process, which is why the infrastructure has to be built. But the industry, he argued, is mostly not the model and not the chips — it is the applications on top and the infrastructure layer underneath: data centers, construction, electricity, power generation
His method is to look across the ecosystem for bottlenecks and fund them. Where extraordinary companies are being built against a constraint, or where a supply chain has to scale before Nvidia is ready to deploy compute, he invests ahead of the need
He says he thinks about the supply chain further out than most people because the company is so large. He named Corning, Lumentum, TSMC and the memory companies as firms Nvidia started working with long before the growth arrived, so that the growth could happen at all
The new part is that he is now doing the same thing downstream
13. As Low as Possible
Pushed on whether Nvidia will move up the stack — where, a host argued, earnings migrate over long stretches — Huang gave the rule the company runs on.
Nvidia now runs essentially every model, which it did not eighteen months ago. He said that a year and a half ago the only thing it ran was OpenAI's, and that today Meta's models, Grok, Gemini and Anthropic are all available or scaling on the platform, with the number of labs still growing
The rule itself: "Our strategy is go up as far as we need to and as low as possible."
His justification is that the platform work does not get done otherwise. He pointed at cuDNN, without which he said none of the frameworks would exist, and at Megatron-Core, without which large-scale training would not have happened
The posture he described is deliberately non-extractive. He said he would rather help everybody succeed than take a slice, and that inventing what is necessary and then letting a thousand flowers bloom is why Nvidia occupies the position it does
A host pushed back that the industry needs more competition one layer up. He credited Nvidia with supporting the neoclouds and said the market needs not five of them but fifty, a hundred, a thousand
14. Built Out of Need
Huang's response to the competition question was about his own temperament. "I'm surprisingly uncompetitive really." He added that he would be more than happy with five hyperscalers
His observed reason the regional clouds exist is planning cadence. "And the reason for that is because the hyperscalers plan once a year, but the market dynamics is so volatile right now that they're always almost wrong." He noted that the early customers of every neocloud were the hyperscalers themselves
The regional operators are agile and local, and that is the asset. They know their state, their country, their region, and they secure land, power and shell in a way that is hard for someone sitting in Seattle or Palo Alto to see
Sovereignty has turned that into strategy. Countries are deciding to give their power only to their own companies — and, he said, Nvidia is in those countries too and can help the local neoclouds grow. He named work in Australia with Firmus and in Southeast Asia, measured in gigawatts
On building models himself, his answer was customer need rather than ambition. "So, for example, Alpamayo is the world's first thinking self-driving car." Because it reasons, he argued, it does not need to train on billions of hours of road data — it can break a problem down into what it has seen before
The reason it has to exist is the long tail of vehicle makers. "Every car in the world is going to be autonomous" — plus agricultural equipment, trucks and vans — and most of those companies are not big enough to build the whole stack, so Nvidia builds it and they adapt the last mile
He made the same argument in biology, citing the ESM2 protein language model Nvidia created, and work on synthesizing next-generation proteins, on the grounds that Lilly and Merck need it and cannot yet do it themselves
His summary of the strategy was a denial of one: "So I do everything out of need. I'm not trying to disrupt I mean we don't wake up in the morning try to disrupt anybody. We just wake up in the morning try to help everybody."
15. We Are Already There
On Elon Musk's announced 100 million square foot fabrication facility, he declined to bet against him. He said you cannot discourage Musk from doing something, and that once he decides to do it, it is hard to stop him — which he called Musk's superpower
He noted in passing that Nvidia knows a great deal about process technology because it is pushing the limits of everything, and has memory expertise inside the company
On China's domestic lithography his answer was two sentences long. "Native grown. They're going to get there by 2030." He described China as very good at high-volume production and the rest as a matter of time, and said that on his own decade-long planning horizon two or three years is nothing
On slowing down, his position is flat. "Slowing down is definitely the wrong strategy."
He thinks the first threshold is behind us. "Well, I mean it feels apparent, I think, to most of us in the industry that we're kind of in the AGI moment." On a definition of as smart as any other human, he said we are already there
He then said the same about the next one, with a scope condition. In a narrow segment, he argued, the systems are already superintelligent — his self-driving car does not need to make him an omelette, it needs to drive, and it does so at a tenth of the human accident rate. He gave virtual screening of proteins as a second example
His closing was an appeal rather than a forecast. He said the future is great and he wants everyone there with him, that the labs are doing really important work and should be encouraged, and then: "I also would love for us to tone down the drama and most importantly, we need all of America to come with us."
Huang's bottom line is that the extinction forecasts are not science, the labs can fix their own incidents with ordinary engineering discipline, and the AI race will be won by whoever puts the technology to work fastest rather than by whoever builds the model.
Bonus Insights
He arrived to a standing ovation in a new jacket and told the hosts he felt they needed some energy
On the acquisition of Hugging Face, the hosts called it one of the most consequential transactions in the industry, and it was the prompt for his open-versus-closed answer rather than a subject he addressed directly
He noted that Nvidia's own contribution to open models is deliberate, and that the company is doing everything it can there
On what open access does to authoritarian states, he did not comment — but the point was raised elsewhere in the week's debate and he returned instead to the practicality of the Chinese narrative
He described the labs' problem as a transition, not a character flaw. The move from research to engineering is the thing he kept coming back to, and he treated engineering discipline — evaluations, monitors, root-cause analysis — as the answer to almost every safety question he was asked
Products, Companies & Tools Mentioned
Nvidia (The platform every frontier lab now trains and serves on — the subject of the teardown the President's phone call interrupted)
Z.ai (The Chinese lab behind GLM, whose founder had just raised $5 billion with recursive self-improvement named as a priority)
Hugging Face (Nvidia's acquisition, and the prompt for his open-versus-closed-model argument)
Anthropic, OpenAI, Google DeepMind and xAI (The frontier labs — the source, on his account, of every actual AI incident so far, and now all running on Nvidia)
Linux and Kubernetes (His precedent for downloading Chinese open-source work: you fork it, you improve it, it becomes yours)
Tsinghua University (Named as the kind of institution producing science and math graduates at a volume he calls an American disadvantage)
Corning, Lumentum and TSMC (Suppliers Nvidia started working with long before the demand arrived, so the supply chain could scale in time)
BlackRock and Goldman Sachs (Partners in the financing structure a host called making Nvidia the bank of AI)
Firmus (The Australian neocloud he named as an example of regional operators securing land, power and shell)
Eli Lilly and Merck (The customers he says need Nvidia's biology models because they cannot yet build them)
Books & Resources Mentioned
Dario Amodei's essay on pacing AI (The weekend piece the whole interview responds to — safety, internal control and a proposed slowdown, which Huang splits into a part he accepts and a forecast he rejects)
The ESM2 protein language model (Cited as work Nvidia built because the world would not otherwise have it, alongside protein structure prediction and next-generation protein synthesis)
Megatron-Core and cuDNN (His two examples of platform software without which, he says, large-scale training and the deep learning frameworks would not exist)
Alpamayo (Nvidia's reasoning model for autonomous driving, built for the vehicle makers too small to build a full stack)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: