Jensen Huang on Reasoning Breakthroughs, the AI Layer Cake, and Why There's No Bubble

The CEO of the world’s most valuable semiconductor company doesn’t think you’re asking the right question when you ask “Is AI a bubble?” The real question, he insists, is whether the world can build enough factories fast enough.

Setting the Stage

Jensen Huang joins Sarah Guo and Elad Gil on the No Priors podcast just after CES 2026, reflecting on a year that, in his telling, validated nearly every bet NVIDIA placed. The conversation spans 76 minutes and covers reasoning models, robotics, open source strategy, energy constraints, US-China dynamics, and why enterprise adoption speed is the wrong metric for measuring AI’s impact. What emerges is less a CEO defending his company than an industrialist laying out a worldview where computing is the new commodity input for civilization.

The Year Reasoning Became Real

Jensen’s biggest takeaway from 2025 isn’t any single model but a convergence: reasoning, grounding, search integration, and model routing all matured simultaneously. Hallucination, which he calls the industry’s biggest credibility problem, was meaningfully addressed across every modality, from language to vision to robotics to self-driving.

He’s especially struck by the economics of reasoning. Test-time compute, where models spend more cycles thinking through a problem, turned out to have its own scaling law: more compute at inference time reliably produces better answers. This made inference tokens profitable, which Jensen calls a “wonderful surprise” because it means NVIDIA’s GPUs aren’t just needed for training anymore. The installed base becomes a revenue-generating asset during inference too.

“The thing that surprised me probably the most is this notion of test-time compute and reasoning. For the very first time we realized that inference could be very compute-intensive, and because it’s very valuable, the tokens are worth more.”

The implication for NVIDIA’s business model is significant: training demand creates the initial GPU purchase, but reasoning-intensive inference creates sustained demand. It’s not a one-time buildout; it’s ongoing infrastructure consumption.

Task vs. Purpose: Reframing AI and Jobs

Jensen introduces a distinction between tasks and purpose that he uses to deflect the “AI will take all the jobs” narrative. Farming’s purpose is feeding humanity, but the tasks of farming have been almost entirely automated. The result wasn’t mass unemployment for farmers; it was a 300x increase in productivity and an explosion of new roles in the food supply chain.

He applies this framework directly to software engineering: the purpose of software engineering (building useful tools) will persist, but the tasks (writing individual lines of code) will be increasingly automated. His prediction: coding will become more accessible, the amount of software will increase dramatically, and “IT budgets are going to go through the roof” as companies that couldn’t afford custom software can now generate it.

The more provocative claim: labor shortages, not job displacement, will be the dominant problem. Jensen points to nursing, teaching, and manufacturing, fields where there simply aren’t enough humans willing to do the work, and argues robotics will fill gaps that immigration policy alone cannot solve.

The Layer Cake: AI Is Not a Chatbot

This is Jensen’s central metaphor and arguably his most important conceptual contribution in the conversation. He describes AI as a “multi-layer cake” of technology:

Foundation layer: Chip design, accelerated computing hardware Training layer: Large-scale model pre-training Post-training layer: RLHF, fine-tuning, grounding Inference layer: Serving models, routing, test-time compute Application layer: Domain-specific tools (medical, legal, engineering, robotics)

His point: most public discourse treats AI as synonymous with chatbots (the visible tip), which leads to distorted narratives about bubbles, moats, and competition. In reality, there’s a vast industrial stack underneath, and NVIDIA operates at multiple layers.

“When you think about wanting to win, that America should win AI, it should not just be America should have this company win AI, but we should try to win across the board, and across domains.”

This framework also shapes his view on open source. He’s a vocal proponent of open-source AI because in a multi-layer cake, the foundation layers become commodity infrastructure. If the US wants to maintain AI leadership, it needs open-source models that any American company, startup, or researcher can build on, rather than concentrating capability in a few closed providers.

Why Open Source Is a National Security Strategy

Jensen frames open source as fundamentally about competitiveness, not altruism. His argument:

1. There are 30,000-40,000 AI startups building applications right now
2. Most of them can’t afford to train foundation models from scratch
3. Open-source models (like Llama) give these startups a competitive base
4. The alternative is every American company depending on one or two closed API providers, which is a fragility risk
5. China has its own closed models; America needs a broad base of capability, not a narrow one

He pushes back hard on the idea that open-sourcing models helps adversaries more than it helps the US: “If you believe that the open-sourcing of technology somehow only helps adversaries, that is that is a silly concept.” The US’s advantage, he argues, is in the breadth and dynamism of its ecosystem, which open source fuels.

The “God AI” Myth and Why Monolithic Models Won’t Win

Jensen makes a pointed argument against what he calls the “God AI” framing: the idea that there will be one superintelligent model that does everything. His counter-thesis:

• Different domains require different types of reasoning (medical diagnosis vs. chip design vs. legal analysis)
• The future is ensembles of specialized models with routers directing queries to the right system
• Agentic AI, where models decompose tasks, verify work, and iterate, is more promising than single-pass answers
• Safety is better achieved through system-level design (guardrails, routing, grounding) than through restricting the models themselves

This connects to his regulation stance: he argues that regulating AI at the model level is like regulating steel or electricity. You regulate the applications (cars, medical devices, financial products), not the underlying technology.

The Plummeting Cost of Compute and Tokenomics

Jensen presents a detailed economic argument about compute costs:

• NVIDIA’s GPUs have improved in performance by roughly 1,000x over the past decade
• This means the cost per token has dropped by approximately 1,000x
• At the same time, the revenue from AI infrastructure has grown from near zero to hundreds of billions
• This apparent paradox (price drops 1,000x but revenue grows massively) is explained by demand elasticity: cheaper tokens create entirely new use cases

He uses this to attack the “AI bubble” narrative directly. Bubbles, he argues, are characterized by price inflation without corresponding value creation. In AI, the opposite is happening: prices are collapsing while utility is expanding. The comparison he draws is to Moore’s Law-era semiconductors, where every generation of chips cost roughly the same to manufacture but delivered vastly more capability.

“We reduced the cost of computing for inference and training by a factor of 1,000 in the last 10 years. During this time, the revenue of this industry went from essentially nothing to hundreds of billions of dollars.”

The Return to Research

One of the more nuanced points Jensen makes is about the nature of current AI development. He describes a shift from pure scaling (just make the model bigger) to genuine research and engineering:

• Model architecture innovations (mixture of experts, sparse attention)
• Training methodology improvements (curriculum learning, synthetic data)
• Post-training techniques (RLHF, DPO, constitutional AI)
• Inference optimization (speculative decoding, quantization, routing)

His frame: the industry went through a phase where “scaling was easy money,” just add more compute and watch metrics go up. Now, the gains require real computer science and engineering innovation. He sees this as healthy and sustainable, unlike the scaling-only paradigm which had obvious physical limits.

Industries Due for Their “ChatGPT Moment”

Jensen identifies several sectors he thinks are about to experience transformative AI adoption:

Healthcare: Already happening with tools like OpenEvidence. Doctors don’t want to do research; they want answers. AI that grounds medical queries in current literature is “irresistible” to an industry that struggles to keep up with accelerating biomedical knowledge.

Robotics and physical AI: Jensen sees the convergence of foundation models with simulation (NVIDIA’s Omniverse/Cosmos) as enabling a breakthrough in robotic capabilities. He describes a pipeline: simulate a robot’s behavior in a digital twin, train the model on synthetic data, then deploy to physical hardware.

Self-driving: He draws an interesting distinction between self-driving cars (which face “long-tail” safety challenges because they share roads with unpredictable humans) and warehouse/factory robotics (which operate in controlled environments and are therefore easier to deploy safely). He predicts autonomous trucking and warehouse logistics will scale faster than urban self-driving.

Software engineering itself: He sees a future where “everybody in the company is a programmer” via natural language interfaces, which he believes will grow IT budgets rather than shrink them.

Energy: The Real Bottleneck

Jensen is blunt about energy being the binding constraint on AI growth:

• Every AI factory needs power
• The US needs to build “hundreds” of new data centers
• Nuclear, natural gas, and renewables all need to scale simultaneously
• Permitting and grid infrastructure are the actual bottlenecks, not technology

He frames this as an economic opportunity, not just a cost. The countries that solve energy infrastructure fastest will lead in AI, which he sees as the core productive technology of the next century.

US-China: Optimism from a Surprising Source

On geopolitics, Jensen is notably more optimistic than most tech executives. He argues:

• China is NVIDIA’s third-largest market and a vital customer
• Complete decoupling would hurt American companies and accelerate Chinese self-sufficiency
• The right approach is “small yard, high fence” export controls on the most advanced chips, while maintaining commercial relationships on everything else
• He sees the new administration’s approach as more pragmatic and business-friendly

The most striking claim: Jensen believes both sides are “tired” of escalation and that 2026 will see a de-escalation in trade tensions. Whether this is genuine analysis or hopeful positioning for NVIDIA’s China business is left to the listener.

Is There an AI Bubble?

Jensen’s full answer to the bubble question synthesizes his earlier arguments:

1. Price dynamics are wrong for a bubble: Compute costs are falling 1,000x while utility is rising. Bubbles show the opposite pattern.
2. The installed base generates ongoing revenue: Unlike the dot-com era, where infrastructure was built for demand that didn’t materialize, AI GPUs generate inference revenue from day one.
3. Enterprise slowness is misleading: The fastest adoption is happening in startups and end-user tools, not enterprise sales cycles. Measuring by enterprise adoption is “going to the slowest adopters of new technologies” for your signal.
4. The TAM is infrastructure-level: Jensen compares AI compute to electricity: the total addressable market isn’t “chatbot subscriptions,” it’s “every industry that uses information to make decisions,” which is all of them.

His parting shot: “No one wants to drive a car from the first decade of cars.” AI in 2025 is the equivalent of early automobiles. The technology is real, the improvements are compounding, and the doubts will look quaint within a few years.

Some Thoughts

Jensen Huang is, of course, the world’s most motivated narrator for the AI infrastructure buildout. Every GPU sold validates his thesis. But the “layer cake” framework is genuinely useful for cutting through oversimplified narratives about AI winners and losers.

A few observations worth sitting with:

• The task vs. purpose distinction is elegant but potentially too reassuring. History does show that automation creates net new jobs at the macro level, but the transition period can be brutal for specific communities and professions. Jensen glosses over this.
• His open-source argument is the strongest case for why a chip company should advocate for model commoditization: if models are cheap and abundant, the bottleneck (and the margin) shifts to hardware. Open source is good strategy, not just good citizenship.
• The energy framing may be the most important takeaway. If AI compute demand really follows the trajectory Jensen describes, the countries that win the energy infrastructure race will win the AI race. Everything else is downstream.
• His China optimism is notable given NVIDIA’s direct financial exposure. Take it as data about incentives, if not as dispassionate analysis.