No Priors: Artificial Intelligence | Technology | Startups

NVIDIA’s Jensen Huang on Reasoning Models, Robotics, and Refuting the “AI Bubble” Narrative

1/8/2026

Reflections on 2025

The Rise of the AI Factory

Reflecting on a year of "several exponentials," Jensen Huang notes that while scaling laws were expected, the surge in reasoning capabilities and the emergence of "profitable tokens" surprised even the architects of the revolution. We have moved from pre-recorded software to intelligence generated fresh, every single time.

Three New Plants for a New Era

AI is no longer just software; it is heavy infrastructure. To support the generation of tokens—which Jensen deems "AI Factories"—the US is witnessing the construction of three distinct types of industrial facilities, driving a massive boom for skilled trades like electricians and plumbers.

  • 1. Chip Plants TSMC, SK Hynix expanding fabrication capacity.
  • 2. Supercomputer Plants Facilities for Grace Blackwell and liquid-cooled racks.
  • 3. AI Factories Data centers dedicated to token generation (inference).
"Every time you use the software... AI is being generated for the first time ever, just like intelligence. I call them AI factories because it's producing tokens that will be used all over the world."
— Jensen Huang

The Radiology Paradox

Geoffrey Hinton famously predicted AI would make radiologists obsolete. The reality? AI pervades 100% of radiology, yet the demand for human radiologists has increased. This illustrates the counter-intuitive nature of AI's impact on labor.

Analysis Prediction vs. Reality

If radiologists are thriving despite AI ubiquity, how do we explain the labor panic? The answer lies in distinguishing between the task and the purpose of a job.

Continuing from the infrastructure boom: If we are building the physical and digital foundation for AI, how does this reshape the definition of work itself? We move from raw labor demand to a fundamental re-evaluation of value.

The Task is not the Purpose.

Productivity paradox resolved: Why does automation lead to more hiring, not less? The answer lies in distinguishing the mechanical actions of a job from its ultimate human goal.

Core Framework
Latent Demand
Infinite capacity for better outcomes.

Case Study: Radiology

THE TASK Study Scans
THE PURPOSE Diagnose Disease & Research

"The more productive we make them, the more demand there will be. We haven't reached the mountaintop of what global healthcare could be."

Case Study: Law

THE TASK Reading Contracts
THE PURPOSE Resolve Conflict & Protect

"The purpose is to protect you. That's more than reading a contract. Automation allows lawyers to focus on the human element of conflict resolution."

"If the world's problems were already specified and there were no other problems to solve, then productivity would reduce the economy. But it's clearly going to increase it."

The "Sucking Sound" in the Economy

We are not facing a surplus of labor; we are facing a critical deficit. Factory work, trucking, nursing—sectors limited by human availability are begging for automation to maintain society.

The New Industry Prediction

Just as cars created the mechanic industry (a 10-year lag), a billion robots will create the largest repair and maintenance industry on the planet.

Critical Labor Shortage Areas (Contextual Severity)

The definition of work is evolving... Next: The Layer Cake of AI Technology

Connecting the Dots

We’ve explored how robotics can solve the physical labor crisis by distinguishing between "task" and "purpose." But for a robot—or any system—to perform a task, it relies on a deep, hidden infrastructure. We now peel back the chassis to reveal the functional anatomy of intelligence itself.

The Five-Layer
Technology Cake.

AI is often misunderstood as just a chatbot. In reality, it is a complex "goop of a mesh of problems" structured into a definitive industrial stack.

Layer 5 (Top)

Applications

Biology, Harvey, Full Self-Driving, Mechanical Humans.

Layer 4

The Models

Understanding information (Bio, Physics, Language).

Layer 3

Infrastructure

Hardware & Software orchestration.

Layer 2

Chips

Compute power.

Layer 1 (Base)

Energy

The raw input transformed into intelligence.

Beyond Human Language

We tend to anthropomorphize AI, thinking it only understands English. But molecules don't speak English. Proteins don't speak English. The "Model Layer" is actually a diverse spectrum of information modalities.

Figure 3.1: The Model Layer must interpret varied information structures, not just text.

The Oxygen of Innovation

Without open source, the AI ecosystem suffocates. While frontier labs may choose closed models for ROI, the rest of the world—startups, universities, industrial manufacturing—relies on pre-trained open foundations to build specialized tools.

"Whatever you decide to do with policy, do not damage that innovation flywheel. Don't forget open source. Don't forget biology."

Analysis Complete

The structure is diverse, layered, and dependent on open exchange.

Coming Up Next

The Myth of "God AI" & The Doomer Narrative →

Continuity

Having established that the modern AI stack is a "layer cake" built on the foundation of open source, we must now address the theoretical specter that threatens to stifle this practical innovation: the hyperbole of the "God AI."

"God AI just
doesn't exist."

The notion of a monolithic model understanding human language, genome language, physics, and molecular biology all supremely well is a fantasy on "biblical scales."

We cannot wait for a deity to advance the economy. The world needs to move forward next week, not next century. AI is simply the next computer industry—and there isn't a nation on earth that doesn't need computers.

0% Probability of "God AI" appearing next week

The Purpose of the "Doomer" Narrative

The Fiction

Science Fiction as Policy

Extremely respectable figures have painted an "end-of-the-world" scenario. While great for cinema, bringing science fiction to government hearings causes real damage. It suffocates startups and confuses policymakers who aren't native to the technology.

The Reality

Regulatory Capture

Why do CEOs advocate for regulations on their own industry? The intention is often to pull up the ladder. It is a classic move to prevent new startups from competing effectively.

"I don't think companies ought to go to governments to advocate for regulation on other companies... their intentions are clearly deeply conflicted."

The Safety Paradox

Why falling costs make AI safer, not more dangerous.

Policing the Agent

The fear is that cheap AI leads to dangerous, unmonitored agents. The reality is the inverse. If the marginal cost of AI drops, we can afford to surround every "Action AI" with millions of "Monitor AIs."

"It's no different than if the marginal cost of keeping society safe was lower. We have police in every corner."

The mechanism of this safety is economic.

To achieve this ecosystem of "digital guardians," the price of intelligence must collapse. Next, we examine the tokenomics driving this plummeting cost and the industry's return to fundamental research.

Up Next: Tokenomics

Continuing from the debate on regulation and the "God AI" narrative...

While the world worries about controlling monolithic power, the real revolution is happening in the economics. The cost of intelligence is collapsing, shifting power from the few to the many.

TOKENOMICS
OF ABUNDANCE

"If you told me that in 10 years we reduce the cost of token generation by a billion times, I would not be surprised."

The 2024 Benchmark

100x Drop

In a single year, the cost of GPT-4 class models per million tokens plummeted by over 100x.

"What used to cost billions and required supercomputers can now be approximated on a high-end PC."

Jensen's Observation

The Compound Acceleration

Vital Insight

The "Deep Seek" Paradox

"Deep Seek was probably the single most important paper Silicon Valley researchers read in the last two years."

Contrary to isolationist narratives, American AI labs are heavily benefiting from global open research. Innovation isn't zero-sum; the cost reduction is driven by a collective, global feedback loop of architecture improvements.

From Generalist to Specialist

We are moving from an era of pure scaling to an era of research and specialization. Models don't need to "boil the ocean" anymore.

  • Micro-niche startups finding efficiency.
  • Bifurcation into vertical expertise (e.g., Coding).
  • Pre-training is just "learning to learn."

What happens when intelligence is free and specialized?

The plummeting cost of compute isn't just an infrastructure story; it's about to fundamentally rewrite the job description of the software engineer.

Coming Next The Future of Coding
Connecting the threads: We've established that the plummeting cost of compute is democratizing research. Now, we examine the direct consequence of this abundance: the fundamental redefinition of the "programmer" and the unlocking of the most complex dataset in existence—biological life itself.

The Death of Coding,
The Birth of Engineering.

Why the "AI-Native" billion-user app isn't a new social network—it's the collapse of the barrier between human intent and machine execution.

Segment: 37:49 — 46:00
"
Nothing would give me more joy than if none of our engineers are coding at all. They're just solving problems.
Jensen Huang

The Paradigm Shift

The Task (Coding)

Translating instructions into syntax. Vulnerable to AI replacement (Cursor, Copilot).

The Purpose (Engineering)

Discovering unknown problems and designing solutions. Demand is infinite.

"If the purpose is not coding, but solving problems, we suddenly have infinite work to do."

The Programmability Moat

Why does NVIDIA resist fixed ASICs? Because algorithms evolve faster than silicon. Moore's Law has slowed to a crawl, but algorithmic innovation demands 100x leaps annually.

  • Fixed Architecture: Great for yesterday's CNNs, obsolete for today's SSMs.
  • Programmable Architecture: Adaptable to Transformers, Mamba, diffusion, and whatever comes next.

Annual Performance Contribution

The Next Frontier

The "ChatGPT Moment"
for Biology

Just as language models learned to predict the next token, biological models are learning to predict the next protein structure. The convergence of multimodality, synthetic data, and long-context compute is digitizing the physical substrate of life.

1

Protein Understanding

Decoding the language of biological structures.

2

Generative Biology

Designing new proteins and molecules (e.g., Lot Protein).

3

World Foundation Models

A unified model for cells, chemicals, and physical interactions.

"We are building the infrastructure for the heavy industries of the physical world."

We have digitized the coder. We are currently digitizing biology. But to truly complete the cycle, AI must leave the screen and enter the physical environment. The final barrier isn't language or protein—it is physics.

Up Next: The Evolution of Self-Driving Cars & Robotics

Continuity

Having explored how software engineering itself is being rewritten by AI, the conversation now leaps from the screen to the street. Intelligence is no longer just generating code; it is beginning to navigate physical space.

The Era of Embodied Reasoning

Self-driving cars were just the warm-up. We are moving from "Perception Cars" to "Reasoning Cars"—machines that don't just follow digital rails, but think through novel scenarios in real-time. This foundational shift is accelerating the timeline for general-purpose robotics.

02
Physical AI

The 4 Eras of Autonomy

Era 1: Digital Rails

Smart sensors, extreme mapping, human-engineered algorithms. (Early Waymo/Mobileye).

Era 2: Modular Deep Learning

Perception, World Model, and Planning as separate modules. Brittle execution.

Era 3: End-to-End Models

Current state. Neural networks handle input to output directly. (Tesla FSD/Modern stacks).

Era 4: Reasoning Systems

Cars that "think" through out-of-distribution events using reasoning chains. The future.

The Verticalization Paradox

Why general models aren't enough for industrial robotics. Satisfaction vs. Failure tolerance.

"Nobody cares about the 90%. They care about the 10% failure."

"The reason why somebody could talk about being a busboy without being a busboy is because they've never been a busboy."

— On the necessity of Vertical Integration

Vertical "wrapper" companies are not obsolete. Deep domain empathy is required to bridge the gap from 99% accuracy to the 99.9999% required for physical work.

Robotics: Faster than Cars?

Yes. Foundational models now exist. We aren't starting from scratch like Waymo did 15 years ago.
Scale. "Everything that moves will be robotic." From excavators to humanoids.
Current Safety Leader NVIDIA (#1) & Tesla (#2)

Coming Next

Scaling robotics and reasoning cars requires immense physical infrastructure. This surge in embodied intelligence leads to a critical bottleneck...

Energy Demand & 2026 Outlook

Continuity

We have established the physical evolution of the machine—from self-driving cars to autonomous robotics. But a machine, no matter how intelligent, creates an insatiable hunger. As we pivot from the "body" of AI to its "blood," the conversation turns to the one resource that governs all growth: Power.

The Thermodynamics
of Intelligence

"Without energy, there is no industrial growth. And without growth, there is no prosperity."

The Energy Pragmatism

The narrative has shifted. The idealistic view of purely renewable timelines has collided with the sheer physics of AI demand. Jensen argues that while wind and solar are vital, they are mathematically insufficient for the immediate explosion of data centers.

  • Natural Gas: The necessary bridge for the next decade.
  • Nuclear (SMRs): The long-term inevitable solution.
  • Demand drives Green Tech: The massive power needs of AI are actually accelerating battery and solar innovation, not hindering it.

The "All of the Above" Strategy

AI INFRASTRUCTURE POWER SOURCES (PROJECTED NECESSITY)

"Doomers are the people who sound smart at dinner parties. Optimists are the people who drive humanity forward."

2026: The Geopolitical Stack

Moving into 2026, the discussion turns to the US-China relationship. Jensen posits that "decoupling" is naive. Instead of viewing technology as a monolith, we must view it as a Stack.

The Software Layer China blocks Google/Meta (The "Firewall").
The Hardware Layer China is a massive consumer of US Chips (Intel, AMD, Nvidia).

US-China Tech Coupling

The hidden prosperity engine beneath the firewall

"The world's mightiest military is supported by the world's mightiest economy."

We have secured the energy to run the machines. We have analyzed the geopolitical chessboard that allows the hardware to flow. But massive infrastructure and global trade require massive capital.

Coming Up Next Is There An AI Bubble?
Context: We've established the geopolitical stakes and the massive energy infrastructure required. But with billions being poured into hardware, the economic question looms: Is this demand real, or is it a speculative frenzy?

The Bubble Myth.

To understand the economics of AI, you must look past the chatbots. We aren't witnessing a speculative bubble; we are witnessing a fundamental re-platforming of the world's $2 trillion R&D sector.

$100T
Global GDP Impact

The Fundamental Shift

"The first thing to realize is to take a step back and ask yourself what is actually happening. Moore's Law has largely ended. You can't use CPUs for everything anymore."

If OpenAI, Anthropic, and Gemini didn't exist today, NVIDIA would still be a multi-hundred billion dollar company.

The demand isn't driven solely by generative text. It is driven by the physics of computing. We are moving from General Purpose Computing to Accelerated Computing. From classical machine learning and data processing (SQL, XGBoost) to molecular dynamics—every layer of the stack requires acceleration to remain deflationary.

The Hidden Iceberg: Beyond Chatbots

Autonomous Vehicles

NVIDIA's AV business is approaching $10 billion. Training world models for robo-taxis requires massive compute capacity.

Digital Biology

Pharmaceutical giants are converting wet labs into supercomputers. Simulation is replacing trial-and-error.

Financial Quants

The entire industry of quantitative trading is shifting from human-featured mathematics to AI-discovered predictive features.

The $2 Trillion Reallocation

Legacy

Traditional R&D spend was allocated to physical experimentation, wet labs, and human-driven analysis.

Shift

Companies are realizing that to compete, their R&D budget must convert into infrastructure budget.

Future

That $2 Trillion is now buying supercomputers. It's not "new" money printed for a bubble; it's repurposed capital.

"Give me an example of a startup company that goes, 'No, we're good.' They are all dying for computing capacity. Give me an example of a researcher who says, 'Got plenty of capacity.' Everybody is dying for capacity."

The skepticism often comes from looking at slow-moving enterprises. But innovation happens at the edge—in the 40,000 startups, in digital biology, in coding assistants like Cursor. We are moving from "Search" (retrieving links) to "Answers" (synthesizing knowledge). This isn't a bubble; it's the industrialization of intelligence.

THE INFRASTRUCTURE IS SET
Coming Up Next Conclusion: The Future of Intelligence

From Speculation to Strategy

Having dissected the anatomy of the hype cycle and questioned the sustainability of current valuations in the previous segment...

The Verdict.

While the market exhibits classic signs of overheating, the underlying utility suggests this is less of a bubble and more of a correction of expectations. The "AI Bubble" isn't about the technology failing—it's about the timeline sobering up.

Market Sentiment

Short-term volatility is inevitable, but long-term capex remains bullish.

The Real Risk

Not adoption, but execution. The winners will be determined by implementation, not announcements.

Related Episodes