NVIDIA CEO Jensen Huang recently participated in an extensive two-hour discussion on the Lex Fridman Podcast, covering core topics such as AI scaling laws, computational power and energy constraints, AI factories, the company's future outlook, and the societal impact of artificial intelligence.
Tokens have emerged as a new commodity class, and the proportion of global GDP dedicated to computing is set to increase a hundredfold. A central thesis presented by Huang is that the fundamental nature of computation has undergone a paradigm shift—transitioning from a "storage system," where humans pre-recorded information for computers to retrieve, to a "generative system" capable of contextual awareness.
More importantly, the economic role of computers has transformed. Historically, computers functioned as retrieval systems, primarily for file storage. Huang likened this to a "warehouse," which itself does not generate substantial profits directly. In contrast, modern AI computers have become "factories" directly linked to corporate revenue generation. He posits that AI foundries are manufacturing a commodity called "Token," which is already being categorized and priced.
"We are not only witnessing these foundries producing goods that people want to consume, but the Tokens they generate hold immense value for diverse audiences. They are even beginning to stratify, much like the iPhone: there are free Tokens, premium Tokens, and various tiers in between," Huang stated. He added, "The idea that someone would pay $1,000 for every million Tokens is on the near horizon. It's not a question of if, but when."
Through this "Token factory" model, computing infrastructure has completed the transition from a cost center to a profit center. Huang confidently extrapolated this macro trend: "If productivity increases substantially, global GDP will experience accelerated growth. I am completely convinced that the future share of GDP allocated to computing will be 100 times greater than in the past."
Addressing the prospect of NVIDIA reaching a $10 trillion market capitalization, Huang responded based on his "Token" economic theory. He remarked, "That figure is just a number," but clarified, "NVIDIA's growth is highly probable—inevitable, in my view." He suggested that achieving a revenue scale of $3 trillion in the future is not an impossibility.
Regarding the bottleneck of AI expansion, particularly energy, Huang acknowledged, "Power is a concern, but not the only one." He proposed two parallel solutions: continuing to push energy efficiency higher and finding ways to access more electricity. On efficiency, he emphasized the metric of "tokens per second per watt," noting that through "extreme co-design," efficiency is improving such that while computer costs are rising, token generation efficiency is increasing even faster, leading to a declining cost per token—by an order of magnitude annually.
On sourcing more power, Huang offered a specific insight: power grids are designed for extreme peak demand but operate well below capacity most of the time. He explained that grids are built to handle worst-case scenarios with a safety margin, meaning "99% of the time, our grid is not operating at its worst-case capacity," often running at around 60% of peak. To utilize this idle energy, he advocated for renegotiating the stringent power supply contracts between cloud providers and utility companies, moving away from a blind pursuit of "six nines" (99.9999%) availability.
"We need to build data centers capable of 'graceful degradation'," Huang elaborated. "When the grid says, 'We need to reduce your power to 80%,' the data center can shift critical loads or reduce computational rates. The quality of service might slightly decrease, but we consume less energy."
Concerning potential supply chain constraints, such as those related to ASML's lithography machines or TSMC's advanced CoWoS packaging, Huang expressed no anxiety. "I tell them what I need, they tell me what they plan to do, and I trust them," he said. On systems engineering, Huang revealed that NVIDIA is profoundly altering infrastructure manufacturing. A single Vera Rubin rack contains 1.3 to 1.5 million components, involving technologies from 200 suppliers. To accommodate the extreme interconnect density, the traditional model of assembling components on-site at data centers has become obsolete. NVIDIA has shifted "supercomputer integration" upstream into the supply chain's "supercomputer manufacturing" phase. Racks are now built entirely within the supply chain and shipped as complete units weighing two to three tons each, requiring gigawatt-scale power reserves for testing before shipment.
In the memory sector, a potential choke point, Huang disclosed that about three years ago, High Bandwidth Memory usage was minimal, confined mostly to a few supercomputers. However, he successfully persuaded several memory CEOs that HBM would become mainstream for future data centers, prompting significant investment in production capacity. He also broke convention by encouraging suppliers to adapt low-power memory, typically used in mobile phones, for supercomputing applications.
Regarding AI scaling laws, Huang delineated four phases of expansion: pre-training, post-training, test-time scaling, and agentic scaling. Addressing concerns about "data exhaustion," he stated, "We will continue to expand training data... much of it will be synthetic data." He concluded, "Training is no longer limited by data; data will be limited by compute power." On the computational intensity of inference, he was direct: "Inference is thinking, and I think thinking is hard... How could it possibly be computationally light?" He believes test-time scaling, encompassing reasoning, planning, and searching, will drive significant demand for inference compute.
When asked about NVIDIA's strongest competitive moat, Huang pointed unequivocally to the vast installed base and trust ecosystem surrounding CUDA. "CUDA wasn't built by three people; it was driven by 43,000 employees," he emphasized, noting that this moat is built on the trust of millions of developers in NVIDIA's ongoing optimization of the underlying platform, combined with its extensive horizontal integration into various clouds, OEMs, and edge devices worldwide.
On frontier explorations like space-based data centers to address energy distribution, Huang confirmed that NVIDIA GPUs are already in space, currently used primarily for edge-side filtering of high-resolution satellite imagery. However, for large-scale orbital data centers, he highlighted a fundamental physical challenge: "In space, there's no conduction, no convection—only radiation for cooling. While there's 24/7 solar energy at the poles, we would need massive radiators." The most pragmatic approach for now, he suggested, remains maximizing the use of idle electricity on Earth.
Praising Elon Musk's execution with xAI's Colossus supercomputer—built in just four months with 100,000 GPUs—Huang attributed the success to first-principles thinking and minimalism. "He questions everything: Is this necessary? Does it have to be done this way? Does it need to take this long?... He appears on the front lines of action. When you act with such intense urgency personally, it compels everyone else to act with urgency too."
Addressing global workforce anxieties about AI, Huang offered a pragmatic standard: if choosing between two new graduates today, he would invariably hire the "AI expert" over someone with no AI knowledge. This principle applies broadly across professions—accountants, lawyers, salespeople, supply chain managers, pharmacists, electricians, and carpenters. He clarified the boundary for job displacement: if a role consists essentially of a series of "tasks," where the task itself constitutes the entire value contributed, disruption by AI is almost inevitable. However, if the work involves a deeper "purpose," individuals can use AI to automate routine tasks, transitioning from mere "executors" to "innovators" within their fields.
For those just starting, Huang gave reassuring advice: "If you don't know how to use AI, you can simply ask AI, 'How do I use you?' It will guide you through the entire process from scratch." He believes the barrier to entry is now zero, with the only obstacle being the decision to start, as the cost of waiting increases daily with AI's evolution.
For programmers, Huang presented a striking view: "The number of NVIDIA software engineers will grow, not shrink... If programming is defined as 'describing specifications for a computer to build,' then the number of people who can do that has just jumped from 30 million to potentially 1 billion. In the future, every carpenter will be a programmer; every plumber will go crazy with the possibilities."
On the timeline for Artificial General Intelligence, if defined as a system capable of autonomously developing applications and generating profit, Huang stated, "I think it's now. I think we have achieved AGI." He envisioned a scenario where an AI-created web service or digital influencer app suddenly gains billions of users and generates profit, a possibility already technically feasible today.