Jensen Huang: The AI Doom Narrative Is Made Up, and Superintelligence Is Already Here
Jensen Huang calls the frontier labs' extinction predictions "made up" and irresponsible, and says superintelligence is not future tense — in narrow domains like autonomous driving and protein synthesis, it already does better than humans.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The doom predictions should be settled one by one
Huang does not rebut AI risk in general. He lays out specific predictions from the past few years and audits them one at a time: someone predicted radiologists would be fully replaced by AI within five years, and the result is that the world needs more radiologists, while AI did take over reading the scans; someone predicted 90% of code would be generated by AI within six to twelve months, wrong; someone predicted 50% of entry-level jobs would disappear within six to nine months, also wrong. His conclusion is that these predictions are "not based in science," and that the people making them should be held to account. The strength of the argument is not the position — it is that he drags the debate from "are you afraid" back to "were you right last time."
— Jensen HuangDanger can only come from the few with the most compute
Huang draws a compute-based boundary for regulation: every incident that has actually gone wrong so far has come from frontier labs, and the reason is not that they are less moral but that they have the most compute and are solving the most frontier problems. A high schooler cannot cause trouble because they do not have enough compute; a startup will not be the source of risk either, for the same reason. So regulation should watch these few, and what these few should do is root-cause analysis in the engineering sense — when something goes wrong, find out what happened, what could have been done, and what technology and process to institutionalise. He adds one judgment: if these labs finish their analysis and say "we don't know what happened, we don't know how to control it, please help us, society," that is when other engineers should be let in to look.
— Jensen HuangRSI is an engineering process, not a runaway spiral
Confronted with news that Chinese labs are investing in recursive self-improvement, Huang breaks it down into a concrete set of techniques: in-context learning, skills, reflection, reinforcement learning, synthetic data generation, plus methods like LoRA that improve a model without changing the base weights, and finally retraining the base on the accumulated experience. His key point is that the control step cannot be skipped — you can RSI inside the company all day, but before a product ships it must be evaluated, must be tested, must be confirmed to have no regressions. So RSI does not lead to loss of control; rather, as labs shift from research to engineering and the controls get stronger, it gets boxed inside the company, and good products ship as usual.
— Jensen HuangOpen-source models are the default for startups
Huang gives a set of numbers: in the last six months, $400 billion of venture money went into AI-native companies, and 80% of them use open-source models. His inference is that without open-source models these companies simply could not build what they want to build, because their dreams are not the same as the frontier labs' dreams. He compares closed models to bottled water — water is free, you use the right water in the right place — and both are needed. As for whether open-source models come from China or the US, his answer is that once you download it, it is yours: Linux and Kubernetes have been touched heavily by Chinese people, and you fork it, improve it, make it your own. From this he redefines the AI race as "who is best at using the technology," not who invented it first.
— Jensen HuangThe doom narrative's anchor moved from national security to safety
Huang notices the narrative has swapped its fulcrum: at first it was anchored in national security, that argument was recently demolished, and so now it is anchored in safety. He does not dismiss safety because of this; instead he offers something concrete — labs themselves must be controllable and have good testing, third-party evaluators must exist, and there must be several of them, for the same reason as financial audits: an auditor does not need to understand the business better than the company, only to ask the right questions; and multiple evaluators coexisting prevents any one of them from being steered by a single institution. The value of this passage is that it turns "should there be regulation" into "what should regulation look like."
— Jensen HuangNvidia's strategy is to go up only as far as necessary
Asked why it went from chips all the way to Hugging Face, autonomous driving and biological models, Huang gives the principle: "go up as far as necessary, go down as low as possible." His examples: without Nvidia building cuDNN, the various frameworks would not exist; without Megatron Core, large-scale training would not happen. So Nvidia builds the necessary technology first and then lets a thousand flowers bloom. He admits he is "strangely uncompetitive" and is happy to see five hyperscalers coexist. He also explains why NeoClouds have a place: hyperscalers plan once a year, and the market is too volatile, so they are almost always planning wrong; regional clouds are more agile, know the locality better, and see land, power and shell more clearly than people sitting in Seattle or Palo Alto.
— Jensen HuangIt builds open models because customers need them, not to disrupt anyone
Huang explains that Nvidia's logic for doing open models is demand-driven: Alpamo is the world's first autonomous driving that thinks, relying on reasoning rather than billions of hours of road-test data, because many carmakers, truck and freight companies are not big enough to build the whole stack themselves — Nvidia builds it and they do the last-mile adaptation. Biological models are the same — ESM2, ESMFold, OpenFold, AlphaFold 2, and Proteina Complexa for next-generation protein synthesis all exist because customers like Lilly and Merck need them and cannot yet do it themselves. His words: "We don't wake up thinking about disrupting someone, we wake up thinking about helping everyone."
— Jensen HuangSuperintelligence has already happened, just in narrow domains
At the end of the episode Huang gives the most radical and most easily misread judgment of the hour: he does not think superintelligence is future tense. His way of defining it is by narrow domain — my autonomous driving does not need to make an omelette, it just needs to drive, and it drives better than a person, with an accident rate one tenth of a human's; synthesising proteins, doing virtual protein screening, in these domains it already exceeds humans too. So his conclusion is that AGI is already here, superintelligence is the next waypoint, but superintelligence already exists in narrow domains. This claim and "AI will destroy humanity" are two ends of the same argument: he concedes neither loss of control nor that the capability has not arrived.
— Jensen HuangIn their own words · checked verbatim
Well, first of all, we shouldn't uh because it's made up.
Jensen Huang4:05
The facts are uh there was a prediction that in 5 years time radiology will be completely taken over by artificial intelligence and there'll be no radiologists in the world. That has proven to be exactly the opposite. We need more radiologists than ever in the world. However, AI has taken over radiology completely which is great is automated scan reading which is great.
Jensen Huang5:05
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 have to test it again, don't you? You have to make sure that there's no regression, right?
Jensen Huang15:12
The facts are in the last 6 months $400 billion of venture funding went into AI native companies. 80% of them use open models. If not for open models, how could they build their dream, right?
Jensen Huang17:13
My point is the race is really about who exploits the technology best.
Jensen Huang19:16
Nobody's in China is saying that there's end of this and end of that and you know cataclysmic this and you know doom or that doom or that. They're much more pragmatic about it.
Jensen Huang20:16
When you when you when you take a narrow segment a narrow segment I mean my my self-driving car I don't want you to make me an omelette I just want you to drive the car right that is super intelligent super it's better it's better than a human yeah yeah onetenth the the accident rate exactly
Jensen Huang44:31
Figures
| Venture money into AI-native companies in the last six months | $400 billion | 17:13 |
| Share of those companies using open-source models | 80% | 17:13 |
| Autonomous driving accident rate relative to humans | one tenth | 44:31 |
| Investment figure Trump says the US is currently attracting | $20 trillion | 27:21 |
| Expected timeline for China to achieve advanced lithography self-sufficiency | 2030 | 43:30 |
Glossary
- RSI / recursive self-improvement
- Using AI to improve AI itself, including synthetic data, reinforcement learning, LoRA fine-tuning and retraining the base model.
- LoRA / low-rank adaptation
- A method that improves a model by training only a small set of low-rank parameters without changing the base weights.
- NeoCloud
- A regional compute provider more agile than a hyperscaler, strong at local land, power and shell.
- land power shell
- The three prerequisite resources before a data centre can be built; Huang uses it to name the downstream infrastructure bottleneck.
- hyperscaler
- The top cloud providers that plan once a year; Huang thinks their planning is often out of step with the market.
How to listen
Founders and investors watching where AI regulation is heading, the open-source model ecosystem and compute capex — especially anyone who wants Huang's own full account of the doom narrative and RSI.
The Trump call from 22:58 to 28:21 — nothing new beyond one line, "this is a scam."