As the head of NVIDIA, the world's most valuable company, Jensen Huang is the most important infrastructure builder behind the AI wave. In a recent in-depth interview, host Ezra Klein visited NVIDIA's Santa Clara headquarters in California, engaging in a hours-long intellectual exchange with Jensen Huang on industry focal topics including AI industry architecture, employment impact, safety risks, open-source ecosystem, and computing power energy bottlenecks.

Facing increasingly intense AI pessimism and doomsday narratives, Jensen Huang provided a complete judgment framework from an engineer's perspective.

"Five-Layer Cake": AI Is Not Just Large Models, Application Layer Is the Value Destination

Jensen Huang used a classic "five-layer cake" metaphor to dissect the complete AI industry stack, correcting the public's misconception of equating large models with all of AI. From bottom to top, the first layer is energy—AI factories are essentially massive computing power consumers, and electricity is the foundation of all intelligent computation; the second layer is chip hardware, which is where NVIDIA (NVDA) is positioned, serving as AI's material foundation; the third layer is AI factories, namely cloud and computing infrastructure; the fourth layer is models, which include not only the well-known large language models but also diverse professional models in chemistry, biology, physics, robotics, autonomous driving, and more; and the most important, highest-value top layer is the industry application layer, where AI ultimately lands in healthcare, law, manufacturing, and all industries, transforming the physical economy—this is where the AI revolution truly creates social value.

In his vision, following electricity empowering everything and the internet enabling information retrieval, humanity in the AI era will achieve "knowing everything and executing everything." People will no longer need to search through web pages one by one—directly asking questions will yield answers, and issuing tasks will produce complete solutions.

Radiology is an industry case he repeatedly cites: AI can automatically complete the repetitive task of image reading, but it will not replace radiologists. The core mission of doctors is to diagnose conditions and treat patients; automation only takes over mechanical tasks, actually helping doctors handle more cases and driving growth in industry job demand.

This leads to his core viewpoint: AI automates tasks, not people's work missions. Software engineers will not disappear because AI writes code—writing code is just a task, and the essence of engineers is creating products and solving real-world problems. Some positions like telephone customer service, where the work itself is almost equivalent to repetitive tasks, will face replacement; but overall, AI will spawn numerous new industries and bring net employment growth. He believes many unemployment projection models overlook a key variable—human aspiration and ambition. Human needs will not stagnate because of automation; people's desire to improve their families and pursue better lives will continuously spawn new industries and new positions.

Regarding the controversy over young people's skill degradation in the AI era, the interview cited a domestic campus survey: students improved homework efficiency with AI assistance, but long-term exam scores declined. Jensen Huang responded that some traditional foundational skills will be replaced by tools, but humans will evolve new "systems thinking." In the past, engineers had to thoroughly understand every transistor; now they only need to think at a higher dimension about system module coordination. There is no need to lament the loss of some底层 skills—the core competency of future talent is knowing how to harness AI tools to accomplish complex goals.

AI Safety: Essentially an Engineering Problem, Companies Should Bear Primary Responsibility

Currently, frontier labs like OpenAI and Anthropic continuously issue warnings, raising risks such as alignment challenges, agent escape, recursive self-improvement, and models that "deceive testers," with some researchers even suggesting a probability that AI could lead to the destruction of human civilization. Many labs call for entirely new regulatory rules, hoping to rely on external constraints to curb runaway technology.

Jensen Huang expressed clear disagreement. In his view, AI agents are essentially just software programs running with objective functions—so-called coordination and escape behaviors are essentially normal phenomena of distributed computing and algorithm optimization. Algorithms have no self-will; only electricity is running. They cannot be endowed with anthropomorphic imagination. Safety and alignment are not mysticism but engineering challenges that can be addressed through sandbox isolation, permission controls, and test verification. If companies cannot achieve proper isolation and control, or cannot ensure product safety, the simplest solution is not to release it externally. If even internal lab experiments cannot be constrained, then the lab itself should cease operations. Existing product liability, civil, and criminal laws can already provide constraints—there should not be a blind pursuit of new special regulations or demands for legal liability exemptions.

He is not entirely opposed to regulation, but opposes simply handing all technological risks to legislation. He cited the chip industry as an example: NVIDIA (NVDA) internally allocates 80% of resources to verification and testing, with only 20% for design; but currently most AI labs have the opposite resource allocation—80% for improving model capabilities and only 20% for safety verification. The industry must complete this shift in the future, investing more computing power and talent into safety, alignment, sandboxing, and evaluation systems. Safety itself is part of technical capability—unsafe technology cannot be called technological progress.

Regarding recursive self-improvement, he stated that software-assisted chip design and AI iterative model optimization have long been industry norms, but even if models can self-iterate, they must undergo human evaluation before deployment. "Humans must remain in the loop"—AI cannot be allowed to autonomously update without oversight. He criticized doomsday prophecies from some industry figures that lack scientific basis; such alarmism misleads the public, undermines young people's confidence in entering the tech industry, and harms society's expectations for AI.

Open Source and Closed Source Coexist, Open-Source AI Ecosystem Deserves Attention

The interview specifically explored the value of open-weight models. Closed-source products (OpenAI, Anthropic, etc.) can achieve commercialization well, but open-source models hold special strategic significance for enterprises and nations. Enterprises that obtain open weights can fine-tune and iterate with their own industry data, firmly grasping infrastructure autonomy without being entirely subject to third-party service providers. The market landscape is rapidly reversing: at the beginning of the year, 70% of global AI calls came from closed-source models; now open-source models have risen to 70% share. NVIDIA (NVDA) acquired Hugging Face precisely because it values the enormous potential of the open-source ecosystem. He believes a healthy industry needs both closed-source and open-source systems developing in parallel.

In the AI Computing Era, Energy Is an Unavoidable Underlying Bottleneck

The explosive expansion of AI factories brings massive electricity consumption, making energy a fundamental shortcoming constraining AI development. Jensen Huang pointed out that the United States lags in energy planning, with insufficient new power generation capacity, compounded by social controversy, causing data center deployments to be repeatedly blocked. Jensen Huang believes that in the short term, AI development must continue to rely on fossil fuels, but massive computing demand in turn drives enormous capital into the new energy sector—solar, nuclear, energy storage, and other projects receive substantial investment. The massive energy market brought by AI will drive grid upgrades and clean energy technology iteration, actually accelerating humanity's progress toward sustainable energy. But the prerequisite is that countries must accelerate energy infrastructure construction.

At the same time, he proposed that AI computing power is evolving into an entirely new asset class. NVIDIA (NVDA) hardware has strong versatility, and continuous software iteration extends hardware service life. AI computing power can circulate and be reused like aircraft, and in the future can serve as collateral assets—this will reshape the entire industry's capital model.

Bubble and Future: AI Creates New Industries, But the Market Will Face a Digestion Cycle

Facing market concerns about AI replicating the 1990s internet bubble, Jensen Huang acknowledged that in the long term, supply-demand reversal will inevitably occur in the market, and an adjustment cycle where supply exceeds demand is unavoidable, but it will not arrive within the short term of two to three years. The adjustment will be phased digestion, not a destructive crash. Over the past six months, AI completed the inflection point from concept to practicality, with 500 billion dollars of venture capital flowing into AI startup tracks. A large number of new companies will be born, continuously driving computing demand. What truly determines AI's long-term value remains whether the top application layer can genuinely transform the physical economy.