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

Special episode energy today—we bumped the usual show because this week’s run has been wild, the room is packed, and it feels like every industry and AI team showed up.

We’ve been open about our roadmap for years, including Dynamo, the OS for AI factories, and the shift to disaggregated inference that spreads pipelines across heterogeneous chips—moving us from a GPU maker to a full AI factory company.

You pushed high‑value inference and even suggested carving out data center space for Grok LPUs; how do you see the industry reacting to this disaggregated prefill and decode approach?

Agents changed the workload mix, so we built systems like Vera Rubin to run diverse models at scale and added racks and processors to put the right job on the right silicon.

What happens at the edge—like toys, robots, or other embedded devices—do we need custom chips there?

I think in threes: one computer to train, one to evaluate in a physics‑true virtual world, and one at the edge for robotics, with telco base stations turning into AI endpoints over time.

Your inference factory looks bleeding edge, but skeptics say it costs more than custom ASICs; why pay the premium?

Don’t confuse capex with token cost—the highest‑throughput factory drives the lowest cost per token, and if you can’t keep pace, even free chips aren’t cheap enough.

You’re steering the world’s most valuable company; how do you choose what to pursue?

I pick insanely hard, never‑done problems that fit our superpowers and accept the pain that comes with building what matters.

How do your long‑tail bets like autos, biology, or space mature from here?

Physical AI is now a multibillion‑dollar business on a path to transform a massive industrial base, and digital biology is approaching its breakout within a few years.

On the desktop, open agents like OpenClaw and powerful local workstations are igniting a maker wave; what does that groundswell mean?

OpenClaw crystallized agentic computing and functions like a personal AI computer with memory, scheduling, IO, and skills, so we’re helping harden it with strong governance and safety.

Does this rapid shift outpace current AI laws and proposed rules?

We need to keep policymakers anchored that AI is software we understand, avoid doom narratives, and focus on fast U.S. diffusion as a security priority.

What would you have advised Anthropic during the DoD dust‑up?

Their tech and safety culture are outstanding, but warnings should inform without scaring; we need moderation and humility because our words now move policy.

Are revenues now scaling with the intelligence curve of these models and agents?

Compute rose approximately ten thousand times going from generative to reasoning to agents, and because people pay for work, agentic systems are driving consumption toward a million‑times expansion.

Inside companies, token use is exploding; what’s a healthy budget per top engineer?

If a high‑paid engineer barely spends on tokens, that’s a red flag; we expect heavy usage as they orchestrate scores of agents by writing specs, tests, and goals instead of just code.

AutoResearch and zero‑shot genomics are collapsing timelines; what does that signal?

Agents won’t kill enterprise tools—they’ll flood them, driving more work through SQL, CAD, creative suites, and simulation so results land in systems people already control.

Where do you land on open weights versus proprietary models and even decentralized training?

We need both: frontier products you subscribe to and open models for domain control, with routers giving you the best model day one and specialization over time.

How’s global diffusion under the new policy regime?

The goal is for U.S. tech to win and spread; we’re reentering China under licenses and pushing for a world that largely runs on an American AI stack for security and prosperity.

Do current conflicts and supply risks like helium change your manufacturing posture?

We support our teams and remain committed to Israel, while at home we reindustrialize, partner closely with Taiwan, diversify to allies, and manage supply risks that likely have buffers.

What’s your auto strategy in a world with Tesla or Waymo building their own stacks?

We enable everyone with training, simulation, evaluation, and in‑car compute plus a reasoning AV stack, and automakers can pick the mix they want.

Clouds are building their own chips; what happens to your share?

Customers buy full AI factories that run across clouds and on‑prem, and we’re gaining share as open models surge, edge grows, and orders like AWS’s scale up.

Analysts still model you like a CPU vendor with modest growth.

They underappreciate AI’s breadth and that we build end‑to‑end infrastructure, not just parts.

How should we think about data centers in space?

Space cooling needs radiative designs with large surfaces, and while we already run CUDA AI on satellites for on‑orbit processing, we’ll explore architectures as we keep building on Earth.

Where’s healthcare headed with AI?

AI for biology, agent assistants for care, and physical AI in instruments and surgery will reshape workflows, with safe agentic interfaces embedded in devices.

Humanoid robots—how soon, and how real?

The brain is here, so expect useful products in roughly three to five years, with China moving fast on motors and materials and robots spreading across factories and homes with telepresence.

And jobs—how do we handle the shift?

Roles evolve rather than vanish, like drivers becoming mobility assistants, and the new craft is mastering AI to direct outcomes.

What should teens study to thrive?

Go deep in science, math, and language—the coding medium of AI—and become fluent with AI tools, noting how radiology grew after computer vision by shifting tasks while keeping purpose.

Huge thanks for an upbeat, clear-eyed conversation; we choose how this unfolds.

Let’s stay optimistic and humble—it’s software we build, not sci‑fi.

Thank you; what a crowd—this was a blast.