NVIDIA’s Jensen Huang on Reasoning Models, Robotics, and Refuting the “AI Bubble” Narrative
Summary
Huang’s 2025 scoreboard is the arrival of grounded, reasoning-heavy tokens that customers trust enough to buy profitably. Search-connected models and confidence-based routers materially improved accuracy; he said he had heard OpenEvidence was at a reported 90% gross margin and said Cursor, Claude, and enterprise OpenAI workloads also carry strong margins. The key transition: tokens are now “sufficiently good in value that people are willing to pay good money for.”
AI demand extends far beyond chatbots because every newly generated token requires an “AI factory” spanning chips, supercomputers, energy, and skilled labor. Huang sees three simultaneous plant buildouts—chip plants, new computer plants, and AI factories—creating enormous demand for construction workers, electricians, plumbers, technicians, and network engineers. He and Guo frame the near-term labor effect as expansion, not displacement.
The right employment unit is a job’s purpose, not whichever task AI can automate. Huang cited Geoffrey Hinton’s prediction that radiology would become AI-powered, yet said the number of radiologists increased because faster scan analysis enabled more diagnoses, research, patients, and hospital revenue. He applies the same test to lawyers, engineers, and waiters: “Oftentimes the technology addresses the task; it doesn’t address the purpose.”
Falling compute costs undermine the idea that frontier AI must consolidate permanently behind a few capital-rich labs. The hosts cited a greater-than-100× decline in GPT-4-equivalent token costs during 2024; Huang expects hardware performance to improve 5–10× annually and said a billion-fold reduction in token-generation cost over a decade would not surprise him. Because combined hardware, algorithm, and model innovation is driving costs down “well more than 10× every single year,” a competitor six months or a year behind might remain close.
Open source is strategic infrastructure for startups, science, education, and industrial AI—not merely an alternative chatbot business model. Huang called DeepSeek’s paper possibly “the single greatest contribution to American AI last year” because American labs and infrastructure companies learned from it. He rejects waiting for a monolithic “God AI,” which he places on “biblical scales” or “galactic scales,” while real industries need adaptable domain models now.
The next investable layer is verticalization across digital biology and physical AI. Huang expects multi-protein models, protein and chemical generation, reasoning vehicles, and multi-embodiment robots to produce new application markets; over the next five years, “the excitement is going to be verticalization.” General models may supply 99% capability, but industrial providers must deliver reliability approaching 99.99999%—leaving substantial value for domain specialists.
Huang’s anti-bubble case rests on capacity scarcity and a much larger addressable market than OpenAI revenue. He cited NVIDIA’s autonomous-vehicle business approaching $10 billion, billions of dollars emerging across financial services, robotics, and digital biology, and used a rough $2 trillion annual global R&D pool to illustrate the shift toward AI-enabled methods. Across startups, universities, and industry, his observed signal is emphatic: “Everybody is dying for capacity.”
Deep dive
Not yet available upstream; scheduled sync will retry.