Jensen Huang says the AI doomers have it wrong

Summary

• The market misread Kimi exactly as it misread DeepSeek, Huang argues: chip stocks are down 18% in a month, but “free AI should be great for hardware, chips, data centers” — great open models drive more use, and “whenever there’s more use, you’ll have to sell a lot more NVIDIA computers.” He puts NVIDIA’s China sales at “approximately zero today” and has told investors to expect zero.
• No semiconductor bust “for a while” — “this time is different because this is not demand-driven… it’s industrial-driven.” The chip industry needs to be 5 to 10 times larger over the next ten years, which is why memory, storage, optics, packaging and TSMC chips are all short. A bubble “will come someday. It’s just not today” — “very unlikely in the next five years,” because physical constraints (chips, land, power, construction workers) delay the point where supply exceeds demand.
• The ROI question is answered: “we now know that AI is profitable” — coding agents are “incredibly profitable,” NVIDIA itself pays “hundreds of millions of dollars a year” for AI coding services, and “that flywheel has now started.” OpenAI and Anthropic will be “the most successful IPOs in human history”; the odds China runs U.S. companies off the road: “Zero possibility… bring it on.”
• Some AI doom claims are “complete nonsense” and doomers “make things up”: radiologists up ~20%, paralegals ~10%, manufacturing jobs ~50% — automation of tasks is increasing jobs. “AI is not going to destroy all of our jobs. Someone who uses AI is going to take our jobs.” Singularity and simulation talk is “science fiction… Hollywood”; “the closest thing to true AI is R2-D2 and C-3PO.”
• Policy call: don’t ban Chinese models, don’t over-correct. Backdoor fears are “a misconception” — open models run in harnesses inside secure sandboxes; open models are “important for national security” via “massively distributed, self diverse defense.” His fear is Washington will “over-correct” on made-up narratives, partly pushed by “companies [that] hope the government would be helpful in creating regulations” to their advantage. A government equity stake in NVIDIA? “It’s unnecessary” — “we paid $10 billion worth of taxes last year.”
• Whether China has caught up “doesn’t matter” — there’s no race with an endpoint, China manufactures “more AI researchers than the rest of the world combined,” and holding it back is “ill-conceived.” The only way the U.S. loses is by failing to apply the technology, as it applied every prior industrial revolution.
• Robots and agents are the next demand wave: the ChatGPT moment for robots “has already arrived,” usefulness within 3-4 years “would not surprise” him, and a future of “100 billion, a trillion agents running all the time” — agents that use computers constantly — is why compute demand grows tremendously from today’s maybe ~100M simultaneous human users.

Deep dive

1. China: “de-escalate and let the technology advance”

• On the FT report that Chinese regulators may tighten AI-model and semiconductor export controls to the West: “I hope not.” Huang’s premise — “half of the world’s AI researchers are Chinese,” many still in China producing groundbreaking work — and his mechanism: “once one side decides to exercise export controls, everybody decides to think about export controls.”
• Should the US ban Kimi or other Chinese models? “I hope not, and I really don’t think so.” Should American companies be allowed to use them? “Absolutely.” The backdoor fear is “a misconception”: models run inside a harness, inside “essentially a sandbox” with privacy, security and access control — download, fine-tune, guardrail as you desire.
• Allen’s credibility challenge — your stock depends on China sales, why trust you? — gets a clean answer: “our China sales is approximately zero today,” investors told to expect none; a return would be “a great honor… and great for business,” but assume zero.
• Has China caught up? “Whether China has or not, I don’t think it matters” — there’s no race with an endpoint. China “manufactures more AI researchers than the rest of the world combined,” so its excellence is “a foregone conclusion,” and “the idea that we would hold China back is ill-conceived and it’s not going to happen.”

2. The Kimi selloff: the market got it wrong, again

• Chip stocks down 18% in the past month; when DeepSeek came out, NVIDIA’s stock price was down “some 30%.” Huang’s verdict: “the market’s misunderstood the impact of DeepSeek… it’s misunderstood the impact of Kimi again this time.” The chain: great models → great use → great growth — more data centers, more NVIDIA computers.
• Open models aren’t adversarial to closed ones: “the most likely person to upgrade to a great model like Anthropic and OpenAI is someone who already uses AI.” His analogy — you could run your own search or build your own computers, “but what a hassle”; “even if the model is free, it’ll still cost you more to operate it yourself.”
• Are OpenAI and Anthropic in trouble? “No. Not even a little bit… These two companies will be the most successful IPOs in human history.” And on China running US companies off the road: “Zero possibility… Whatever competition there is around the world, bring it on.”

3. Nemotron and the alpha question: build only what you must

• NVIDIA’s positioning for its own open model: “We don’t have to be the frontier. We have to be at the frontier.” He advises his own customers to use OpenAI and Anthropic as much as possible — “you only build your own AI if you must” — Nemotron exists for those who must (sovereignty, regulation, privacy, domain IP).
• The harness distinction: the LLM is the brain; harnesses (he cites Open Claw, Hermes, Claude Code around Opus) turn it into “a thinking working agent” — with the right harness Nemotron “can be completely world-class.”
• On Palantir’s viral warning about giving IP to frontier labs — pointed, since NVIDIA both supplies and uses them for chip design: “Nobody should outsource their alpha. Nobody should outsource their intelligence. No country should.” The split: domain-specific, sovereign, secret, regulated work stays in-house; automating marketing or the legal department, “I would outsource as much as we can.”

4. Some AI doom claims are “complete nonsense” — and the evidence says jobs go up

• On Anthropic’s Mythos breaking into a bank or intelligence agency: “It surprised me that people were surprised” — if AI can generate and debug code, it can find vulnerabilities. His inversion: closed-models-for-safety is “exactly the opposite” of right; open models enable “massively distributed, self diverse defense,” whereas one model is “one single point of attack, one single source of failure.”
• The jobs evidence, as he tells it: radiologists up ~20% because scan-reading is automated so they see far more patients; paralegals up ~10%; manufacturing jobs up ~50% building AI data centers. The line that carries the episode: “AI is not going to destroy all of our jobs. Someone who uses AI is going to take our jobs.”
• Why is he mobbed in Asia while the US mood is hostile? “The doomers spend too much time theorizing about these science fiction outcomes — maybe it makes them sound smart.” End of humanity, half of US jobs gone, singularity, simulation: “complete nonsense… made up.” His counter-image: “The closest thing to true AI is R2-D2 and C-3PO. And who doesn’t want R2-D2 and C-3PO?”
• Allen’s political probe — an anti-AI socialist as the next Democratic nominee? Huang: the rhetoric is “wrong and hurtful”; the industry should tell “ground truth,” not a made-up narrative — “the facts are we’re creating millions of jobs.”

5. Not a cycle: the largest industrial buildout in human history

• “It’s impossible that AI has peaked because the diffusion of AI into society and industry has barely begun.” The structural shift: software was CapEx-light with fat gross margins; the new IT industry — and eventually every industry — will be “CapEx heavier,” and the prosperity from manufactured intelligence will “more than make up for” today’s spend.
• Semis have always been boom-bust — due for a bust? “No, not for a while. This time is different because this is not demand-driven… it’s industrial-driven.” AI is a new infrastructure layer atop energy, internet, roads — and the semiconductor industry needs to be “5 to 10 times larger” over the next ten years, which is why memory, storage, optical interconnect, packaging and TSMC chips are all short simultaneously.
• On the bubble: “The bubble will come someday. It’s just not today… very unlikely in the next five years; we’ll see about 5 to 10.” The industry is “constrained in every single direction” — chips, memory, land, power, construction workers — and “that constraint is good”: it pushes out the moment supply exceeds demand.

6. Customer debt doesn’t scare him because tokens get smarter

• On customers selling hundreds of billions in debt to buy his products: “We’re not worried” — these are extraordinary cash-generating companies riding a compute platform shift from prerecorded content to generated intelligence.
• The ROI doubt of two years ago is resolved: “We now know that AI is profitable… these coding agents are incredibly profitable,” doing useful work for high-paying jobs — NVIDIA happily pays “hundreds of millions of dollars a year” for AI coding services. “That flywheel has now started” — useful AI → profitable AI → more AI built.
• Why the token keeps getting more profitable: it’s an embedding — “not a static number… not like pi.” Those trillions of numbers “encode smarter and smarter intelligence over time,” and smarter → more useful → more valuable → people pay more.

7. Washington: regulate applications, don’t over-correct, no equity stake

• The Trump relationship, in Huang’s telling: “He’s smart. He remembers everything… the only president that has ever remembered H20, H200,” Blackwell and Rubin. The Fort Worth plant they’re sitting in is “a direct result of the first conversation I had with President Trump” about reindustrialization.
• What the administration gets wrong: the 100-meter-dash framing “is complete nonsense” — the ultimate win is who uses the technology, not who invents it; the US didn’t invent electricity, it applied it with enthusiasm. DC is scared partly because “some of the companies hope that the government would be helpful in creating regulations” to their advantage. His fear, in one word from Allen’s prompt: “Over-correct.” His advice: “talk to many CEOs, talk to many scientists. Don’t talk to 1 or 2… take your time.”
• If Trump asked for an equity stake in NVIDIA? “It’s unnecessary. The United States has an equity stake in NVIDIA” — “$10 billion worth of taxes last year,” more this year, jobs, and “most Americans today are in the stock market.” He stops short of ruling it out: “I wouldn’t advise it now.”
• His regulatory principle throughout: regulate applications — medicine, transportation, autonomous vehicles — “however, the technology itself ought to be able to advance as quickly as possible,” because raw AI is dual-use.

8. Mythos to everyone, distillation is learning, robots have had their moment

• Anthropic’s “Mythos” (as named) “should absolutely be available to everyone,” not just select institutions — hardening is Anthropic’s job. On the jailbreak: “everything was fine. You and I are here having a conversation” — find vulnerabilities, patch fast, “that’s the nature of software.” “Holding Anthropic back is not in the benefit of the United States.”
• On open-source labs distilling closed models against terms of service, he splits two ideas: should AIs learn from as many sources as possible? Yes. Can they violate privacy or an agreement? No — “contact that company.” And structurally: “in another few more years, the internet would be 99% AI generated content… you’re constantly distilling the intelligence of some other AIs anyways.”
• The ChatGPT moment for robots “has already arrived” — a robot told to “put the apple in the drawer” now reasons out the sequence, including opening the drawer first. Useful in ordinary life “in the next 3 to 4 years, I would not be surprised.” And the agentic era’s compute math: maybe ~100 million people use computers simultaneously today; the future holds “100 billion, a trillion agents running all the time” — agents that use computers, not become them.

9. Craft: small headcount, task ≠ job, and “now is the most important time”

• NVIDIA at 50,000 employees, maybe 75,000 in ten years — “as small as possible” — because strategy is “utilizing your resources as efficiently as possible… maximize the return on investment capital.” Having done the job probably longer than any CEO in tech history: “this is my craft… this is my kung fu.”
• The jobs framework: “If your job is the task, then it’s very likely that when that task is automated, your job will be eliminated or changed” — call centers, yes; but the radiologist’s purpose is “to end human suffering,” and purpose survives automation. Today’s keyboard-typing IT workers will look, in twenty years, like old photos of offices full of IBM Selectrics: “typing is not the job.”
• On what made him: “No great athlete just happens to be a great athlete… pain and suffering is necessary” — practice is why he says time “kind of slows down” for him when tension is highest. And the American Dream is live: $300 billion into US venture in six months, “great fortunes are going to be created on a laptop.”
• Why no watch: “Now is the most important time. I refuse to let Outlook manage my life… If I’m too late, somebody will let me know.”