Jensen Huang doesn't think AI is going to end civilization — and he's not shy about saying so. Appearing on the All-In Podcast in front of a live audience, the Nvidia founder and CEO delivered a thorough, direct takedown of AI existential risk predictions, calling them irresponsible, unscientific, and measurably wrong. On the question of what Jensen Huang thinks about AI safety and existential risk, his answer is clear: the fears are overblown, the predictions have a terrible track record, and the people making them should be held accountable. What followed was one of the most wide-ranging conversations about AI, Nvidia's strategy, open source, AGI, China, and the future of work that Jensen has ever given publicly.

Is AI an Existential Threat? Jensen Huang Says No

Jensen opened with a methodical dismantling of the AI doom narrative. His approach was simple: look at the actual predictions and check if they came true.

  • Prediction: Radiologists would be replaced by AI within 5 years. Reality: We need more radiologists than ever, though AI has helpfully automated scan reading.
  • Prediction: 90% of code would be AI-generated within 6–12 months. Reality: Wrong.
  • Prediction: 50% of entry-level jobs wiped out within 6–9 months. Reality: Also wrong.
  • Prediction: GPT-2 and Llama 3 would be too dangerous to release. Reality: Both released, neither caused civilizational collapse.

Jensen's conclusion was blunt: "We have to take accountability for all of the stupid predictions that were made." He argued that when researchers attach precise-sounding numbers — like a "10% chance of extinction" — to claims that have no scientific grounding, it's not just wrong, it's irresponsible. The words are alarming. The credentials are real. But the actual content is made up.

He also pushed back on the idea that doom-saying and safety are the same thing. Real safety, in Jensen's view, means engineering discipline: root-cause analysis when something goes wrong, better sandboxes and runtime monitors, third-party evaluators similar to financial auditors. That's solvable. Predicting the end of humanity is not safety work — it's theater.

Will AI Destroy Jobs? Jensen Calls Out the Bad Predictions

Jensen made an analogy that stuck: before software became dominant, engineers barely typed. The entire first generation of computing engineers had to build the computers before software was even possible. Did typing destroy engineering? No — it transformed it.

"I've been saying that my favorite key is backspace," Jensen joked, explaining that the best software is the smallest software. Coding, in his framing, is just typing. And there was extraordinary engineering before typing. There will be extraordinary engineering after it too.

His broader point: AI is creating jobs, not destroying them. The $400 billion in venture funding that flowed into AI-native companies in just six months created enormous demand for compute, data centers, construction, power infrastructure, and software talent. Every layer of the stack needs people. The narrative of a jobs apocalypse doesn't match what's actually happening on the ground.

How Does Nvidia Stay Ahead of Every Competitor?

Jensen described Nvidia's strategy with a phrase that sounds deceptively simple: "Go up as far as we need to and as low as possible." The goal is never to compete with customers — it's to solve the problems that nobody else can solve, build the infrastructure that makes everything else possible, and then step back.

He gave concrete examples. If Nvidia hadn't built cuDNN, the deep learning frameworks wouldn't exist. If they hadn't built Megatron Core, large-scale model training wouldn't have happened. They invented the tools, let others build on top, and maintained their position as the only computing platform that runs every major AI model in the world.

A year and a half ago, he noted, Nvidia's platform ran essentially one model: OpenAI. Now it runs Meta's models, Grok, Gemini, Anthropic, and a growing list of frontier labs. That's not an accident — it's the result of a deliberate strategy to be indispensable to everyone rather than competitive with anyone.

On the question of moving up the stack, Jensen was honest: Nvidia is building frontier models in five specific domains — self-driving (Alpamo), biology, and others — but only because customers need it and no one else can do it yet. "I do everything out of need. We don't wake up in the morning trying to disrupt anybody. We wake up trying to help everybody."

Open Source vs Closed Source AI: What Jensen Actually Uses

Jensen's take on the open vs. closed source debate was refreshingly practical. He uses both. This past weekend alone, he used four closed models. He compared closed models to bottled water: "Water is free, you guys. But sometimes you buy bottled water." The right tool for the right job.

But open source, he argued, is essential for a different reason: it's how most of America actually builds AI companies. Of the $400 billion in VC funding that went into AI-native companies in the last six months, 80% of those companies use open models. Without open source, those startups couldn't build their products. The frontier labs' closed models are too expensive, too restricted, or simply not the right fit for every application.

On the question of Chinese open source models — and whether it matters that they come from China — Jensen was characteristically pragmatic. He pointed out that a huge percentage of the world's open source software, from Linux to Kubernetes, has already been touched by Chinese engineers. Once you download it, it's yours. You fork it, improve it, make it your own. The same logic applies to AI models.

Is Recursive Self-Improvement AI Actually Dangerous?

The podcast touched on a Chinese AI lab announcing a $3 billion investment specifically targeting recursive self-improvement (RSI) — AI that trains the next version of AI. Jensen demystified the term completely.

RSI, he explained, is really a combination of existing, well-understood techniques: in-context learning, skill acquisition, reflection, reinforcement learning, synthetic data generation, and low-rank adaptation (LoRA). These are sensible, practical ideas for making AI systems better over time. "It's a very logical idea and I'm certain everybody is using it in some degree."

The phrase "recursive self-improvement" is being weaponized, Jensen argued, to create the impression of a system that could spiral out of control. But the reality is mundane: any model trained internally still has to be evaluated, tested for regression, and released through normal product engineering processes. The move from research lab to engineering company — which all the frontier labs are undergoing — naturally produces better controls, not less.

How Close Is China to Winning the AI Race?

Jensen's answer here was sobering and worth taking seriously. On advanced lithography — the chip-making technology currently controlled by ASML — he estimates China will get there by 2030. And 2030, he noted, is just around the corner.

His broader framing of the AI race itself was even more interesting. He pointed to the last industrial revolution: Maxwell, Volta, Ampere — none of them were American. Europe invented the technology. America exploited it better than anyone else and won decisively. His argument is that the same dynamic could play out with AI: the race isn't about who invents the models, it's about who exploits the technology best.

China's pragmatic narrative — AI as an economic accelerant, not a civilizational threat — is actually helping them in Jensen's view. While American discourse gets consumed by doom and safety theater, China is building. The absence of self-inflicted drama is a competitive advantage.

Is AGI Already Here? Jensen Huang Thinks So

The conversation ended with a moment that deserves more attention than it got. Jensen was asked whether we're in the AGI moment. His response: "I think we're already there."

His definition is narrow-domain focused. A self-driving car that operates at one-tenth the accident rate of a human driver isn't just competent — it's superhuman in that domain. An AI system that synthesizes proteins and runs virtual screening better than any human researcher is superintelligent in biology. By that framing, we haven't just reached AGI — we've already surpassed human performance in multiple critical domains.

The definition debate matters because it shapes policy, investment, and public perception. Jensen's view suggests the revolution isn't coming — it's already here, distributed across specific high-value domains, and the opportunity now is to build on top of it rather than debate whether it exists.

His closing message to the audience was characteristically optimistic: "The future is great and we want to get there. We're going to be enormously successful together as a humanity." Tone down the drama, bring all of America along, and keep building. That, in Jensen's worldview, is the entire strategy.