Jensen Huang: The Doomer Hoax, Superintelligence is Here, and The Future of AI (ft. President Trump)

Summary

• Jensen Huang attacks the Dario essay’s unsupported doomsday predictions, while separating safety from leadership and defending the Kokotajlo whistleblower. Radiology was supposed to be fully automated in five years — instead, “we need more radiologists than ever”; predictions of 90% of code being AI-generated within 6–12 months and half of entry-level jobs disappearing in 6–9 months also proved wrong. “We have to take accountability for all of the stupid predictions that were made.”
• President Trump phoned in live mid-episode and called the AI-takeover narrative “a hoax,” saying China was “the happiest group” benefiting from opposition to data centers. Data centers are “the oil of the next 20–25 years… bigger than the internet,” and “whoever wins AI wins.” He also claimed $20 trillion of investment was coming into the country in one year versus much less than $1 trillion under President Biden over four years, while adding that “we have to be a little bit careful” and act prudently.
• Jensen says all actual AI problems so far have come from frontier labs because they have the most compute; the fix is engineering root-cause analysis, better controls and multiple independent evaluators, not sweeping regulation. He would “bet you money” that the “four incidents from one lab” and “one giant incident from the other lab” were within the labs’ future control.
• Recursive self-improvement will not spiral out of control because products still have to be tested and evaluated. Responding to a host report that Zhipu AI, the makers of GLM, would put $3 billion toward an RSI effort—and David’s note that its founder had raised $5 billion—Jensen calls RSI a sensible combination of in-context learning, skills, reflection, reinforcement learning, synthetic data and LoRA. He is “certain that everybody is using it to some degree.”
• Open models are a foundation of the boom: 80% of the AI-native companies receiving $400 billion of venture funding in the last six months use them. Closed models are “bottled water,” while “water is free”; once a Chinese model is downloaded, “it’s yours… we fork it, improve it, and make it ours.” The real race is “who exploits the technology best,” just as America exploited a European-invented industrial revolution.
• NVIDIA’s posture is to go “as far as we need to, and as low as possible”: build the enabling stack and infrastructure, help customers succeed, and avoid taking a slice of every layer. It says it now runs every model in the world versus only OpenAI 18 months ago, supports regional neoclouds because hyperscalers plan once a year and are “almost always wrong,” and Jensen calls himself “surprisingly uncompetitive.”
• Jensen predicts China will reach homegrown advanced lithography by 2030, with mainland-fab production following “almost immediately,” and says China is “already there” from its own perspective. He also says AGI is already here and that narrow superintelligence is already here in domains such as self-driving—where he cites one-tenth the accident rate—and protein synthesis and virtual screening.

Deep dive

1. The doomer essay is unsupported — and Jensen keeps a ledger of the misses

• Jensen’s read of Dario’s essay: safety and leadership are not a false choice; companies can innovate and execute quickly while leading safely. He treats the Kokotajlo whistleblower matter as serious and says Kokotajlo showed “great courage,” while cautioning that the essay conflated separate issues. He does not know what Kokotajlo saw, and says pausing or slowing down are voluntary choices the labs could make if they believe they are out of control.
• The forecast ledger: radiology was predicted to be fully automated in five years — instead AI took over scan reading while “we need more radiologists than ever”; 90% of code was supposedly about to be AI-generated within 6–12 months; and 50% of entry-level jobs were supposedly going to disappear within 6–9 months. The hosts add predictions that GPT-2 and Llama 3 were too unsafe to release. “We have to take accountability for all of the stupid predictions that were made.”
• Chamath’s framing—his mother asking about “this whole civilizational-death thing,” and how anyone quantifies a 10% chance of extinction—draws Jensen’s flat answer: “we shouldn’t, because it’s made up… it’s irresponsible.” He says the scientific prediction is not grounded in science and argues that, if the danger were real, people should spend more time addressing it than alarming people who cannot act on it.
• On public lab discourse, Jensen says these consequential companies should be built “the way that we used to build companies, which is in silence.” At NVIDIA, employees are told how to behave on the company’s behalf, and political discourse is kept outside the company: “take it home… the company is apolitical, we’re bipartisan.”

2. Regulation should solve actual problems — and the problems so far have come from frontier labs

• Jensen’s compute-concentration logic is probabilistic, not absolute: the problems so far have come from frontier labs because they have the most compute and are attempting the hardest work. A high-school student or startup is unlikely to cause the same kind of problem because they would not have enough compute. The labs are also transitioning from research to engineering, and “maybe that transition is clumsy.”
• His remedy is engineering discipline: root-cause each incident, then institutionalize sandboxes, runtimes, monitors and continuous monitoring. Jensen would “bet you money” that the four incidents from one lab and the one major incident from another were within the labs’ future control. He doubts the alternative—that the labs would conclude they have no idea how to control what they built and must ask society to solve it.
• His governance model is independent evaluation: multiple third-party auditors or evaluators, “no different from financial controls,” with multiple evaluators reducing the risk that one company or evaluator becomes captured or influenced.
• David reported that Zhipu AI, the makers of GLM, had announced a $3 billion effort toward recursive self-improvement; he also said Zhipu’s founder had just raised $5 billion. Jensen dedramatizes RSI: it combines in-context work, skills, reflection, reinforcement learning, synthetic-data generation and low-rank adaptation (LoRA). He calls it logical and is certain everybody uses it to some degree. It cannot simply spiral past control because “when you release a product, you’ve got to evaluate it,” test for regression and verify it again.

3. Trump calls in live: “the whole thing is a hoax”

• Mid-conversation, Jensen’s phone rings and President Trump goes on speaker to the crowd. Trump says the opposition to data centers is “almost a conspiracy,” with China “the happiest group,” and calls data centers “the oil of the next 20–25 years… bigger than the internet.” He says robots will not take over the world, AI will not take over the rest of the world, and “the whole thing is a hoax”—while also saying the country must act prudently rather than stop industry.
• Trump’s claims and complaints include $20 trillion of investment coming into the country in one year versus “much less than $1 trillion” under President Biden over four years; Google wanting to build a large data center in Finland after being unable to get permitting; and the slogan “whoever wins AI wins.” He also says data centers have revived communities that were previously dying.
• Asked afterward how Trump sees through something that Jensen calls “polling minus 80,” Jensen says he is not sure because many people are falling for it. He says the narrative was initially anchored on national security and is now anchored on safety: the immediate need is to ensure the labs are in control and that reliable testing and independent evaluation exist.
• After the call, Jensen says what he had wanted to tell Trump and the audience was that AI is creating jobs, including software jobs. He points to roughly $400 billion in recent AI venture financing and the resulting demand for compute and data centers; he separately notes Governor Abbott’s request that the industry listen more carefully to small communities.

4. Open models are how America wins — even when they are Chinese

• Jensen’s bottled-water framing: closed models are like bottled water—useful in the right places even though “water is free”—and he used four closed frontier models over the weekend. But 80% of the AI-native companies receiving $400 billion of venture funding in the last six months use open models: “if not for open models, how could they build their dream?”
• Does Chinese origin matter? Jensen says that once a model is downloaded, “it’s yours”: the user can fork it, improve it and make it their own. He compares this with Linux and Kubernetes, much of whose software has been touched by Chinese engineers. He attributes China’s contribution partly to its ability to produce science and math students at scale through universities such as Tsinghua.
• The race, in his view, is “who exploits the technology best.” Maxwell, Volta and Ampère were European rather than American, but America exploited the last Industrial Revolution more effectively than anyone else. Jensen says China’s messaging is more pragmatic because it emphasizes economic and social advancement rather than civilization-ending predictions.

5. NVIDIA’s posture: building the bottlenecks, “as far as we need to and as low as possible”

• Chamath’s capital-stack setup points to Cloverleaf’s land-power-shell work and financing efforts involving BlackRock, Goldman and others. Jensen describes a bottleneck strategy: NVIDIA worked with Corning, Lumentum, TSMC and memory suppliers long before demand surged, and is now working downstream on land, power, shell, construction and generation capacity.
• Why not take over the application layer? Jensen says NVIDIA’s strategy is to go “as far as we need to, and as low as possible.” It builds enabling technologies such as cuDNN and Megatron Core, then lets “a thousand flowers bloom.” He says NVIDIA now runs every model in the world, compared with only OpenAI roughly a year and a half earlier, because he would rather help everyone succeed than take a slice out.
• The neocloud logic: hyperscalers plan once a year, but volatile market conditions mean they are “almost always wrong.” Regional NCPs can move faster and secure land, power and shell locally. Jensen cites Firmus in Australia and IOH and others in Southeast Asia, with additional gigawatts being brought online as countries increasingly treat compute infrastructure as strategic.
• The host cites Hugging Face, Poolside, Llama, Nemotron and NVIDIA’s open self-driving stack while asking whether NVIDIA is pursuing an open-source “gold medal.” Jensen answers that NVIDIA builds out of need: “we are the frontier model in five domains.” Alpamayo, which he calls the world’s first thinking self-driving car, is intended for car, truck, van and ag-tech makers that cannot build the full stack themselves. He also cites ESM-2, ESMFold, OpenFold, AlphaFold 2 and Proteina-Complexa as biology work NVIDIA built because companies such as Lilly and Merck need it. “I do everything out of need,” not to disrupt competitors.

6. China is “already there,” Elon cannot be stopped, and narrow superintelligence has arrived

• On Elon Musk’s announced 100-million-square-foot Terafab: “if anybody could do it, he can.” Jensen says they discussed it at length on a shared flight. He adds that NVIDIA knows a great deal about process technology, has substantial memory expertise and is a systems company, without saying that NVIDIA itself will fabricate chips there.
• On China’s homegrown advanced lithography, Jensen predicts, “They’re going to get there by 2030.” Asked whether production would move into mainland fabs almost immediately once the capability exists, he answers, “Almost immediately,” and says China is already there from its own perspective because high-volume production is ultimately a matter of time. The host concludes that slowing down would be the wrong strategy.
• When the host defines AGI as being as smart as a human, Jensen says, “I think we’re already there.” When asked about superintelligence, he agrees but narrows the claim to specific domains: a self-driving car that does not need to make omelets but drives at one-tenth the accident rate is “superintelligent,” and he says protein synthesis and virtual protein screening are already there as well.
• His closing message is to tone down the drama, encourage the frontier labs and bring all of America into the transition: “The future is great,” and humanity can be enormously successful together.