Jensen Huang LIVE: Nvidia's Future, Physical AI, Rise of the Agent, Inference Explosion, AI PR Crisis

Special episode, packed room, and we bumped the usual show for this. Big week, big vibes, and yes, we had to tease you about Groc and how unbearable Choff got during the close.

We laid out this strategy in daylight years ago at GTC. Dynamo is our operating system for the AI factory, built on disaggregated inference so different parts of the pipeline run on the best silicon for the job. We’ve evolved from a GPU maker into a full AI factory company, and Groc fits as another specialized processor in that mix.

The disaggregation point hit me hard. You even suggested setting aside a chunk of data center capacity for Groc to run high value inference.

We moved from single large models to agentic workloads, which pound storage, memory, and tools. That’s why we built Vera Rubin with multiple racks and expanded our TAM with storage processors like BlueField, CPUs, networking, and, I hope, a lot of Groc—putting every workload on its ideal chip.

What runs at the edge, like a talking teddy bear or other embedded applications? Do we end up with different tools and chips for each use case?

There are three computers. One trains AI. One evaluates and simulates in Omniverse with physics. One runs at the edge in robots, cars, or a tiny toy. A big one is turning telco base stations into AI edge, which turns a massive industry into part of the AI fabric.

You said inference would explode, and now it has. People worry your inference factory costs more than custom ASIC paths. Why pay a premium?

Don’t confuse factory cost with token cost. Higher throughput and efficiency win; the fastest factory makes the cheapest tokens. If you can’t keep up with tech velocity, even free chips aren’t cheap enough.

You run the most valuable company on Earth. How do you decide where to push, pull back, or enter new fields?

My job is to define vision and strategy around problems that are insanely hard, never done, and matched to our superpowers. You choose that path knowing it brings pain, and you learn to like the pain.

Give us the long-tail bets. Space data centers, ADAS, biology—what really inflects?

Physical AI is a massive category and already a multi-billion business for us, pushing toward ten billion a year and compounding. Digital biology is near its ChatGPT moment; within a few years we’ll represent genes, proteins, and cells in ways that change healthcare.

Take us from data center to desktop. Hobbyists and open source are roaring back with OpenClaw and powerful local boxes. What does this street‑level agent movement mean?

Three waves hit. Generative went mainstream with a simple UI. Reasoning grounded answers and monetization. Then Cloud Code showed agents to enterprises, and OpenClaw put agents in popular culture. OpenClaw is a new computing model with memory, scheduling, IO, and skills—basically a personal AI computer that runs everywhere—so we must build in governance and safety.

A shift like that makes a lot of AI laws feel outdated. How should policy adapt?

We need to brief policymakers regularly. This is software, not a sentient alien. Avoid doomerism, keep policy behind the state of tech, and focus on diffusion at home so we don’t fall behind nations that adopt faster.

On the Anthropic defense dust‑up, what would you advise to avoid fear and mistrust?

Their tech and safety focus are excellent. It’s good to warn, but scaring people without evidence backfires. Our words shape society now, so be balanced and humble about predicting the future.

Are we finally seeing ROI as agentic systems ramp and you talk about trillion‑dollar visibility?

AI is broader than two labs. Open models are huge. Compute jumped by roughly ten thousand times moving from generative to reasoning to agents, and people pay more for work than for chat. Agents do work, so consumption is soaring toward a million times over time.

Inside companies, token usage is exploding. How much should world‑class engineers spend?

If you pay an elite engineer top dollar and they barely use tokens, that’s a red flag. Great engineers should lean into AI heavily, shifting from coding to writing specs, architectures, and evaluation loops while orchestrating hundreds of agents.

Auto research blew our minds with a desktop breakthrough in minutes. What does that say about algorithms, hardware, and enterprise tools?

Agents will slam into the tools we already trust—SQL, Photoshop, Blender, EDA—so enterprise software grows as agents multiply. Tools remain the interface we control when work comes back to us.

Where does open source end up, especially with decentralized training efforts?

We need both. World‑class proprietary services for general intelligence, and open models so industries can own and control domain expertise. Both will thrive.

A year into new export rules, how are we doing on global diffusion, and what’s next for China?

The goal is for American tech to lead and spread. We gave up a huge share in the second‑largest market, but we’re getting licenses approved and ramping to ship again. The best outcome is the U.S. tech stack powering most of the world, not a repeat of solar, motors, or rare earths.

Conflicts and helium supply—how much risk is there, and what are you doing?

We support our teams and stay committed in Israel and the region. We’re re‑industrializing in the U.S., diversifying to allies, and urging restraint. Helium could tighten, but the supply chain has buffers.

On self‑driving, you added partners across the board. Are you the Android‑like platform, with others as the iOS?

Everything that moves will gain autonomy. We supply training, simulation, and the in‑car computer, plus a reasoning stack we call Alpomayo. Automakers can take one layer or the whole thing; we enable rather than own the car.

Your biggest customers also build chips. How do you hold share as they try to go down‑stack?

We win on system velocity and a full stack that runs in every cloud and at the edge. Forty percent of customers need turnkey AI factories, not loose chips, and that’s where we shine. Share is rising—AWS alone plans to buy a massive number of our parts.

Analysts model your growth way below your vision. Do your orders support a bigger ramp?

They underestimate the breadth of AI and what we build. We’re not just a chip vendor; we build complete AI infrastructure for far more than a handful of hyperscalers.

Help us picture data centers in space.

We’re already running CUDA in satellites for on‑orbit AI. Space has abundant energy but cooling is radiation‑only, which needs large surfaces; we’ll explore architectures while we keep scaling on Earth.

Healthcare is ripe. Where do you see real impact first?

AI physics for biology and drug discovery, agent assistants for care and documentation, and physical AI for robotic surgery. Every clinical instrument will become agentic with a safe OpenClaw‑like core.

Humanoids seemed stalled, then snapped back. How soon until robots are in our daily lives?

We were early and got tired before the brain was ready. Now we’re three to five years from widespread deployment. China’s strength in motors and magnets will accelerate things, and I expect fast movement.

Robots could unlock prosperity for everyone, like cars once did.

We’re short on labor, so robots help now. Telepresence means I can beam into a home robot, check on the dog, and do work from afar—distance starts to fade as a constraint.

Model and agent revenue could hit the hundreds of billions and beyond. Is that conservative?

It is. Enterprise software will become value‑added resellers of tokens from Anthropic, OpenAI, and others, which expands distribution and revenue sharply.

What moats matter at the application layer when models keep improving?

Deep specialization. Pair general models with proprietary sub‑agents trained on your domain, and plug them into your agentic system close to customers to spin the flywheel.

You called it years ago: people won’t lose jobs to AI, but to people using AI. Still, displacement is real. How do we navigate it?

Jobs will change more than vanish. Drivers become mobility assistants while cars drive, doing higher‑value work. For young people, become expert at using AI—there’s artistry in specifying goals without over‑prescribing.

What should students study?

Deep science, deep math, and strong language skills, since language is the new programming interface.

Even English majors could thrive if they master AI. Your radiology example proves adoption can raise demand.

Computer vision is in every radiology workflow now, and the need for radiologists grew because throughput rose and care improved. That’s the pattern: AI lifts productivity, then society invests in more service, not less.

This was a hopeful, grounded conversation. You’re the steward we need, and we get to choose progress over fear.

Let’s keep the focus on building, stay humble, and skip the scare‑mongering. We’ve done transformations like this before.

Thank you. Amazing crowd, great energy. Let’s get back to work.