#494 – Jensen Huang: NVIDIA – The $4 Trillion Company & the AI Revolution

Today I sit down with Jensen Huang, CEO of NVIDIA, the company driving much of the modern AI wave; dear friends, let’s dive into extreme co‑design and the era of AI factories.

You’ve moved from single‑GPU excellence to full rack and data center design; what makes end‑to‑end co‑design so hard?

Problems no longer fit on one machine, so we split algorithms across thousands of computers where Amdahl’s law, networking, switching, CPUs, and GPUs all become bottlenecks; only deep, cross‑layer co‑design beats linear scaling as Moore’s law slows.

How do all the specialists—memory, networking, power, cooling—actually work together on one rack?

We optimize from apps down to chips and power, and we structure the company around the product we want; my staff is large and technical, we solve problems in the room together, no rigid one‑on‑ones, and people jump in when their discipline matters.

When did you shift from gaming GPUs to thinking of NVIDIA as an AI factory?

We started as an accelerator, then walked a tightrope toward accelerated computing: programmable shaders, IEEE FP32, CG, then CUDA; the pivotal bet was putting CUDA on GeForce to build a vast install base, even though it crushed margins for years.

How did you make that existential call?

Curiosity and first‑principles reasoning gave me conviction, and I spent years shaping belief—board, team, partners—brick by brick so that when we declared deep learning or a big acquisition, people already felt, what took you so long.

You’ve long believed in scaling laws; what still scales and what blocks the future?

Pre‑training keeps growing with synthetic data, inference is hard because thinking and planning are compute heavy, and the next curve is agentic scaling where we multiply coordinated agents; it all feeds back into training, with compute as the ultimate limiter.

Models change fast, but hardware takes years; how do you anticipate the future?

We run basic and applied research, listen across labs, and keep CUDA specialized yet flexible; that’s why we built NVLink‑72 and evolved from LLM inference racks to agent‑ready racks with storage accelerators, new CPUs, and GROC—because agents must use tools, read files, and search.

OpenClaw exploded into public attention and raised security worries; how are you handling safety?

We added OpenShell and enterprise controls with a two‑of‑three permission model across sensitive data, code execution, and external communication, plus policy integration, so agent systems are powerful but constrained.

What might slow scaling now that agents are inevitable?

Power is the concern, so we chase tokens‑per‑second‑per‑watt and drive token cost down with extreme co‑design; I also align the entire supply chain—upstream and downstream—so capacity grows with demand.

You even nudged memory vendors; what did that take?

I laid out why HBM and even adapted low‑power mobile memory would become mainstream in data centers, and several CEOs invested early; today each rack has over a million parts sourced from hundreds of suppliers.

Are you worried about tools like EUV and advanced packaging keeping up?

No; I told partners what we need, they told me how they’ll deliver, and I believe them.

On energy, how do we grow compute without overloading the grid?

Most of the time the grid has headroom, so data centers should gracefully throttle or shift when society needs peak power; utilities can offer graded guarantees sooner if customers accept flexible SLAs, and we’ll engineer for controlled degradation.

Elon built a huge cluster fast; what’s instructive there?

He’s a minimalist systems thinker, shows up at the point of action, questions everything, and radiates urgency so suppliers prioritize him; that breaks the usual waiting games.

What echoes of that in NVIDIA’s approach?

We design from the speed‑of‑light limit—latency, throughput, power, cost—and rebuild from first principles before we iterate, instead of nibbling at legacy processes.

Your pods are jaw‑dropping; do you still chase simplicity?

We keep only the complexity that’s necessary and discard the rest; even so, these are the most complex computers ever made, and we manufacture them at scale.

Why is China’s tech scene so fast?

A huge talent base hit the mobile‑cloud era perfectly, intense home‑market competition prunes the weak, and a culture of sharing and open source accelerates iteration; many leaders are engineers, and it shows.

Why open‑source advanced models and recipes?

It informs our co‑design for future models, brings every industry and student into AI, and pushes beyond language into biology and physics; we open weights, data, and how we built them.

What makes TSMC unique?

Beyond devices, they orchestrate global demand with yield, cost, and service, blend technology with customer obsession, and build rare trust over decades—even without a formal contract.

What’s NVIDIA’s moat?

CUDA’s install base and developer trust, plus our execution velocity and one architecture that runs everywhere—from clouds to cars to space—make developers choose us first and most.

How has your mental model of compute changed?

We went from chips to clusters to full AI factories; I think in racks, pods, and gigawatts now, and GPUs already run AI at the edge in orbit where data is born.

Space compute soon, or later?

We’re learning radiation tolerance and graceful slowdown, but near term I want to harvest idle grid power first.

Could NVIDIA reach ten trillion; what would that world look like?

Computing shifted from retrieval to generation, turning warehouses into factories that mint valuable tokens; as productivity jumps, GDP accelerates and compute spend soars, and we can scale with partners and energy because no hard physical limit stops us.

Agents feel like the iPhone moment for tokens.

Yes—OpenClaw showed explosive demand for agentic work.

How do you handle pressure and low points?

I decompose problems, share the load, act, and forget what I should not carry; stay childlike about new challenges, re‑evaluate assumptions, and keep moving toward the future.

On gaming, some say DLSS 5 makes things look like AI slop; your take?

I understand the worry, but DLSS 5 is 3D‑conditioned and faithful to geometry and artist intent—it enhances rather than overwrites, lets creators steer style, and pairs with tech like skin shaders; tools like RTX mod help communities refresh classics without losing soul.

People are anxious about jobs, but tools aren’t the job’s purpose. Computer vision beat humans years ago, yet we now have more radiologists because faster, better reading created more demand for diagnosis and care; it’ll be similar for software—headcount grows, not shrinks.

So coding might expand, not contract. Current engineers still have an edge in shaping systems with language, and the craft of programming and design principles remains valuable.

Coding is becoming high‑level specification—telling computers what to build—so the pool of people who can do it explodes from tens of millions to potentially a billion. Trades like carpentry or plumbing level up too, because AI turns them into designer‑builders who deliver far more value.

There’s an artistry to how tightly you specify. Sometimes you set a precise blueprint; other times you under‑specify to explore with the AI and push creativity. Choosing your spot on that spectrum is the future of coding.

People in white‑collar roles feel real fear, and we should meet it with compassion. I hope AI removes drudgery, amplifies creativity, and brings back the joy of building—I’m having more fun programming than ever.

When I face anxiety, I separate what I can act on and move. If I’m hiring an accountant, lawyer, marketer, or engineer, I choose the one fluent with AI; if your job is mostly tasks, you’re at risk, so use AI to automate the tasks and elevate your purpose.

A chatbot can help you break big worries into steps. It’s the best beginner’s companion I’ve seen—reduces friction, teaches the first moves, even plans a trip to China or Taiwan in seconds.

Ask it for my favorite spots in Taiwan and it’ll find them. You’ll eat well.

Do you think parts of consciousness are beyond computation—like nerves before a big moment, love, grief, the fear of death?

AI can read and model the patterns behind emotions, but my chips don’t feel. Two people can face the same situation and perform wildly differently; I don’t see our machines reproducing that lived, subjective swing.

Still, scaling surprises me. Intelligence keeps blooming in unexpected ways, so I’m open to being amazed.

Intelligence is perception‑understanding‑reasoning‑planning—a functional loop, not the entirety of being human. As intelligence becomes a cheap utility, what matters most is character, compassion, grit; let that lift you instead of scare you.

I think AI will push us to celebrate humans even more. Do you think about mortality?

I don’t want to die—I love my family and the work; this is a once‑for‑our‑species moment. I don’t do a formal succession plan; I constantly share context, decisions, and insights so knowledge spreads, and I hope to pass while working, quickly, without a long decline.

Nvidia’s pace is thrilling; as a fan of engineering, it’s a joy to watch. What gives you hope for the next decades?

I trust human kindness and our capacity to build; near‑term progress makes curing major diseases and cleaning the planet feel plausible. I even plan to launch a humanoid and beam an AI trained on my digital life to meet it—likely within about five years.

Cracking the mind and physics would be incredible, and it feels within reach. Thank you for everything you do; hope to see you in Taiwan.

Thank you, Lex. Your long‑form interviews, depth, and care reveal what makes builders tick—I’m grateful for the platform you’ve created.

I’ll close with a nod to Alan Kay: the surest way to meet the future is to help invent it.