Jensen Huang on AI Jobs, Robotics, and the Bubble Debate

TL;DR

Reasoning and grounding improved enough in 2025 that inference tokens became genuinely profitable, with companies like OpenEvidence reportedly running 90% gross margins. Jensen Huang argues AI adds jobs rather than removing them: radiology is now 100% AI-powered, yet the number of radiologists increased, because a job's purpose (diagnosing disease, doing research) outlasts the tasks AI automates.

Transcript

Hson, thanks so much for joining us today. >> So great to have you guys. What an amazing year. >> What a year. >> Happy Hanukkah, merry Christmas, >> happy new year coming up. Yep. Happy holidays. >> So, uh, with everything that's happened in 2025, um, and you know, being in the middle of the vortex with it, what do you reflect on and say like this... Read More

Key Insights

• Grounding and reasoning were the standout technical improvements of 2025, alongside connecting models to search and putting routers in front of models so that, depending on the confidence of an answer, the system goes off and does the necessary research to improve accuracy.
• Hallucination was one of the biggest skeptical responses to AI, and Huang credits the whole industry with addressing it in 2025 across every field, from language to vision to robotics to self-driving cars, through the application of reasoning and the grounding of answers.
• Inference tokens are now profitable, which Huang found genuinely surprising. He points to OpenEvidence at a reported 90% gross margins, and strong margins at Cursor, Claude, and enterprise use of OpenAI, as proof the tokens are doing valuable work people will pay for.
• AI is software, but not pre-recorded software. Excel was written by several hundred engineers, compiled, and distributed as is for years. AI takes context into account and generates every single token for the first time, every time it is used.
• AI counts as infrastructure because it affects every application, every company, every industry, and every country, placing it alongside energy and the internet rather than alongside conventional shipped software products.
• Three new plant types are needed to support AI factories: more chip plants, new supercomputer plants for machines like Grace Blackwell where an entire rack is one GPU, and the AI factories themselves. All three are being built at large scale across the United States for the first time.
• Skilled trades are a direct beneficiary of the AI buildout. Construction workers, plumbers, electricians, technicians, and network engineers are all in demand, and Huang says electricians are seeing their paychecks double and going on business trips the way executives do.
• The task versus purpose distinction explains why AI adds jobs. Geoff Hinton predicted radiologists would become unnecessary, and roughly eight years later 100% of radiology applications are AI powered, yet the number of radiologists increased instead of falling.

Questions & Answers

Q: What surprised Jensen Huang most about AI in 2025?

Huang says the scaling laws and the pace of technology advancement did not surprise him, because those were already known. What did surprise him is that token generation rates for inference, especially reasoning tokens, are growing on what looks like several exponentials at the same time, and that these tokens are now profitable. He points to OpenEvidence reportedly running 90% gross margins, along with strong margins at Cursor, Claude, and enterprise use of OpenAI, as evidence the industry is now generating tokens valuable enough that people willingly pay good money for them.

Q: How did the AI industry address the hallucination problem in 2025?

Huang credits several converging improvements. Grounding got better, reasoning got better, models were connected to search, and routers were placed in front of models so that, depending on the confidence of an answer, the system can go off and do the necessary research before responding. He describes hallucination and generating gibberish as one of the biggest skeptical responses to AI, and says the whole industry addressed it across every field, from language to vision to robotics to self-driving cars, through the application of reasoning and the grounding of answers.

Q: Why does Jensen Huang call AI data centers AI factories?

Because unlike pre-recorded software, AI has to generate output fresh on every use. Huang contrasts Excel, which several hundred engineers wrote, compiled, and distributed as is for years, with AI, which takes in what you asked of it and what is happening in the world as context and generates every single token for the first time, every time. That means computers must produce tokens continuously, so Huang calls the facilities AI factories: they are producing tokens that get used all over the world, much like a plant producing goods.

Q: Why does Huang consider AI to be infrastructure?

Huang argues AI qualifies as infrastructure because of how universally it is used. It affects every single application, it is used in every single company, in every single industry, and in every single country. That breadth puts it in the same category as energy and the internet rather than in the category of a discrete software product. He frames it as part infrastructure and part factory output, since the tokens themselves are produced continuously by computers and then consumed across the entire economy.

Q: What new types of plants are being built to support AI?

Huang identifies three new types of plants. First, more chip plants, with TSMC and SK Hynix building additional capacity. Second, new supercomputer plants, because these machines are unlike anything the world has seen before: he cites Grace Blackwell, where an entire rack functions as one GPU. Third, the AI factories themselves that generate tokens. He says all three types are currently being built in the United States at very large scale and quite broadly across the country for the very first time.

Q: Which jobs benefit most from the AI infrastructure buildout?

Skilled trades and technical labor. Huang lists construction workers, plumbers, electricians, technicians, and network engineers as the skilled labor needed to support this new industry in the near term, and says the demand will be enormous. He specifically mentions being excited to hear that electricians are seeing their paychecks double, and that they are now being paid to travel for work, going on business trips the way he and other executives do. He describes the three plant types as creating a large amount of employment.

Q: Why did the number of radiologists increase even though radiology is now AI-powered?

Huang points to the difference between a job's tasks and its purpose. Geoff Hinton predicted five to seven years ago that AI would revolutionize radiology within five years and advised people not to enter the field. Huang says Hinton was right that 100% of radiology applications are now AI powered, roughly eight years later. But a radiologist's task is studying scans, while the purpose is diagnosing disease and doing research. Because AI lets them study more scans more deeply, request more scans, and diagnose better, hospitals become more productive, see more patients, earn more, and hire more radiologists.

Q: What is the task versus purpose framework for thinking about AI and jobs?

The framework separates what you physically do in a job from why the job exists. Huang uses himself as an example: he spends most of his day typing, which is his task, but typing is obviously not his purpose. When AI automates a lot of that typing, which he appreciates and says helps a lot, it has not made him less busy. In many ways he has become more busy, because he is able to do more work. The same logic applies to radiologists, whose purpose is diagnosis and research rather than the mechanical act of studying scans.

Summary & Key Takeaways

• Huang says the scaling laws and technology advances of 2025 did not surprise him, but he was pleased by improvements in grounding, reasoning, model connection to search, and routers placed in front of models that trigger further research based on answer confidence. He credits the industry with addressing hallucination, one of the biggest skeptical responses to AI.

• What did surprise him is how fast token generation for inference, especially reasoning tokens, is growing, on several exponentials at once, and that those tokens are now profitable. He cites OpenEvidence at 90% gross margins, plus strong margins at Cursor, Claude, and enterprise use of OpenAI, as evidence people will pay good money for valuable tokens.

• On jobs, Huang frames AI as software that is not pre-recorded: Excel was compiled once and shipped for years, while AI generates every token for the first time, every time, based on context. That requires AI factories, which in turn requires three new plant types being built across the United States: chip plants, supercomputer plants, and AI factories.

• Building those plants creates near-term demand for construction workers, plumbers, electricians, technicians, and network engineers, and Huang notes electricians are seeing their paychecks double and traveling for work like executives. He separates a job's tasks from its purpose: his own task is typing, but that is not his purpose, so automating it made him more productive rather than less busy.