📺 Today’s recommended deep-dive video: https://www.youtube.com/watch?v=k-xtmISBCNE

Jensen Huang on the Industrial Revolution of Intelligence: Beyond the AI Bubble

NVIDIA CEO Jensen Huang explores the fundamental shift from general-purpose computing to “AI factories,” arguing that the current technological surge is not a bubble but a permanent re-architecting of the global economy. He breaks down why 2025 was the year grounding and reasoning finally conquered hallucinations, paving the way for profitable “tokconomics” and specialized industrial applications.

Core Question: Is the current AI explosion a speculative frenzy, or the birth of a new $100 trillion industrial stack?

Highlights

• The Five-Layer Stack: AI is a “layer cake” starting with energy and chips, moving through infrastructure and models to specialized vertical applications.
• Task vs. Purpose: Automation replaces repetitive tasks (like scanning a slide) but preserves the human purpose (diagnosing disease), leading to higher job demand.
• The Death of Hallucinations: Significant leaps in grounding and reasoning have turned AI from a creative toy into a trusted expert tool with 90% gross margins.
• 2026 Predictions: The next “ChatGPT moment” will occur in digital biology, specifically in multi-protein understanding and generative molecule design.

⏱️ Reading time: approx. 8 minutes · Saves you about 68 minutes vs. watching.

Want to take notes while watching? Click the image below and let AI Notebook capture the key points for you 👇

The Rise of the AI Factory

From Software to Real-Time Generation

Artificial Intelligence represents a departure from the “pre-recorded” software era typified by tools like Excel, where code is written once and distributed as a static product. In the AI era, software is generated for the first time, every time, relying on live context and reasoning to produce unique tokens for every query. This shift demands a constant stream of computational power, turning data centers into modern “AI factories” that produce intelligence as a commodity.

This transition has triggered a massive expansion of physical infrastructure across three distinct categories: chip plants, supercomputer assembly lines, and the AI factories themselves.

Because this movement requires immense physical resources, it is creating a massive demand for skilled labor, including electricians, network engineers, and construction workers. It is not uncommon to see technicians in these fields receiving double pay and traveling across the country like executives to meet the urgent needs of the grid. This surge proves that the “intelligence industry” is as much about heavy industrial capacity as it is about neural network architecture.

The profitability of this endeavor is now grounded in reality rather than just hype. Companies specializing in legal research or medical evidence are seeing gross margins as high as 90%, because the tokens they generate provide high-value, accurate results that professionals are willing to pay for. This “tokconomics” model suggests that intelligence is becoming the most valuable raw material in the global market.

We are essentially building the power grid for the next century of productivity.

💡 Digging Deeper

Q: Why call it a “factory” instead of a data center?
A: A data center stores data; a factory uses raw materials (energy and data) to manufacture a valuable output (tokens/intelligence) that never existed before.

Q: Are the high margins sustainable?
A: Yes, because as the cost of token generation drops, the value of the “answers” provided by specialized models remains high, particularly in regulated industries like law and medicine.

The Paradox of Employment and Automation

Task vs. Purpose

The common fear that AI will result in mass unemployment stems from a failure to distinguish between the specific tasks a worker performs and the ultimate purpose of their profession. While a radiologist’s task might involve studying a scan, their purpose is to diagnose disease and save lives. When AI automates the task of scanning, the radiologist doesn’t lose their job; instead, they can diagnose more deeply and treat more patients.

Productivity gains in an economy with infinite problems to solve do not lead to layoffs; they lead to expansion and the discovery of new problems.

Consider the history of the “paperless office” or the introduction of the PC. These technologies automated rote tasks but made professionals so much more effective that demand for their services exploded. In fields like nursing, accounting, and trucking, the primary challenge is actually a massive labor shortage that AI is uniquely positioned to fill.

When we create billion-robot fleets, we aren’t just replacing human labor; we are creating the world’s largest repair and maintenance industry. Every autonomous taxi or warehouse robot will require a human-led infrastructure to keep it operational. The “sucking sound” of the economy is currently pulling resources toward these new support roles that didn’t exist a decade ago.

The purpose of a lawyer is to resolve conflict and protect the client, not just to draft a contract.

Tokconomics and the Open Source Flywheel

The Collapsing Cost of Intelligence

The cost of generating AI tokens is currently dropping at an exponential rate, far outstripping the historical pace of Moore’s Law. While traditional computing improved 2x every 18 months, AI performance is compounding at 5x to 10x every year through a combination of hardware, architectural, and algorithmic breakthroughs. This suggests a billion-fold reduction in the cost of intelligence over a ten-year horizon.

This deflationary pressure on intelligence makes it accessible to startups that don’t need to “boil the ocean” or build a monolithic “God AI.”

NVIDIA maintains a programmable architecture because the “species” of AI, such as Transformers and State Space Models (SSMs), is evolving too fast for fixed-function chips to keep up. By protecting programmability, the industry ensures that a researcher’s latest innovation can run on the largest installed base of hardware immediately. This flexibility is what allows the “innovation flywheel” to spin faster every month.

The role of open source is critical here, as it provides a foundation that prevents the industry from being suffocated by closed-source monopolies. Open-source models like DeepSeek have significantly benefited the American ecosystem by providing frontier-level research that everyone can adapt. It ensures that 100-year-old industrial companies can integrate AI into their specific domains without being dependent on a single provider.

Winning at AI means winning across the entire stack, not just possessing one dominant model.

💡 Digging Deeper

Q: Is the AI bubble going to burst?
A: Unlikely. Even without chatbots, the shift from CPU-based computing to GPU-accelerated computing for data science and physics is a multi-hundred-billion-dollar transition.

Q: How does NVIDIA compete with dedicated ASICs?
A: ASICs are “brittle.” AI algorithms change so fast that by the time a fixed-function chip is made, the researchers have moved on to a new architecture like SSMs or Hybrid-Transformers.

The 2026 Frontier: Biology and Reasoning

The Next ChatGPT Moment

The most exciting breakthrough on the immediate horizon is the “ChatGPT moment” for digital biology. We are moving from simply understanding protein folding to generative protein design and multi-protein interaction modeling. This will transform drug discovery from a “wet lab” trial-and-error process into a predictive supercomputing challenge, shifting billions in R&D spending toward AI infrastructure.

Beyond biology, the next wave of physical AI will be defined by “reasoning cars” and multi-embodiment robots.

Self-driving technology is evolving from brittle, human-engineered perception systems into end-to-end reasoning models that can handle “out-of-distribution” scenarios. If a car encounters a situation it has never seen, a reasoning system allows it to break the problem down into known variables to find a safe path. This same foundational technology will eventually allow a single AI to embody different forms—from a surgical arm to a giant excavator.

Energy remains the final constraint on this industrial revolution. We need to embrace every form of energy—from natural gas to nuclear—to power the factories that will drive national prosperity. Demand for AI is actually the single greatest driver for sustainable energy innovation, as companies race to build “behind-the-meter” power solutions to keep their clusters humming.

National security and economic prosperity are now two sides of the same silicon coin.

Key Takeaways

AI is transitioning from a “science fiction” narrative into a pragmatic industrial reality. The most successful applications in 2025 have focused on “grounding” models in real-world data, effectively solving the hallucination problem and allowing experts in law, medicine, and engineering to trust these tools. As token costs continue to collapse, the focus is shifting toward vertical specialization where companies solve deep, niche problems.

The global economy is currently energy-starved and labor-short, making AI a necessary deflationary force rather than a threat. By automating rote tasks, we allow the human workforce to focus on high-level purpose and problem-solving, which historically leads to economic expansion. We are not just building better software; we are building a new type of factory that manufactures the most precious commodity in the world: intelligence.

Q&A

Q1: Is the cost of training models becoming a barrier to entry?
A: While frontier training costs are high, the cost of achieving a specific level of intelligence is actually dropping by 10x every year. This means startups can stay competitive by specializing in niches or using improved training algorithms.

Q2: Why is “reasoning” such a big deal for 2025?
A: Reasoning allows models to think before they speak, using “thinking tokens” to verify their own answers. This has drastically improved the accuracy of AI in fields where “mostly right” isn’t good enough.

Q3: Will AI replace radiologists and lawyers?
A: No. It replaces the “tasks” of scanning images or drafting contracts, but it increases the demand for their “purpose”—diagnosing disease and resolving legal conflicts. History shows that making a professional more productive increases the total demand for their services.

Q4: What is the biggest constraint on AI growth right now?
A: Energy and “factory” capacity. The world is currently facing a global shortage of the supercomputing infrastructure needed to generate tokens at the scale the market demands.

Q5: How does China fit into the AI ecosystem?
A: China is a massive contributor to open-source AI and a major market for hardware. While they are adversaries in some areas, the technological stack is highly coupled, and open-source contributions from China benefit American startups.

Q6: What does “multi-embodiment” mean for robots?
A: It means a single “brain” or foundation model can be used to control many different types of hardware—a car, a robot arm, or a factory line—rather than having to build a new AI for every specific machine.

Q7: Is the $2 trillion R&D market shifting?
A: Yes. Historically, R&D for things like drug discovery happened in wet labs. Today, that budget is shifting toward building supercomputers to simulate biology and chemistry, which is far faster and more scalable.