Lex Fridman

Lex Fridman: Jensen Huang on NVIDIA’s Existential CUDA Gamble, Agentic AI, and the Real Limits of Scaling

NVIDIA’s CEO explains why risking the company on CUDA was necessary, how collective engineering drives their edge, and why the next AI revolution will force data centers—and the grid—to adapt.

If you only read one thing

Jensen Huang didn’t just bet on GPUs—he nearly bankrupted NVIDIA by putting CUDA on every GeForce card, sacrificing profits to build the install base that would make AI possible.

Jensen Huang’s talk with Lex Fridman is a blunt look at what it takes to win in computing. The heart of the story is CUDA: Huang put it on every GeForce GPU, knowing it would wipe out profits and tank NVIDIA’s market cap from $8 billion to $1. 5 billion. He did it anyway, convinced that only a massive install base would attract developers and make NVIDIA indispensable for AI. Technical elegance, he argues, is irrelevant if nobody builds on your platform—a lesson learned watching x86 outlast more elegant rivals.

Huang also details how NVIDIA’s structure—60 experts working as a group, never in one-on-ones—lets them move fast and anticipate AI workloads before they exist. This “extreme co-design” approach is why NVIDIA can build for agentic AI: systems that spawn sub-agents and use real-world tools, demanding new hardware and software. Huang’s final challenge is to the entire infrastructure world: data centers and power grids are built for rare peak loads, wasting energy and money. He says dynamic power allocation—letting data centers slow down during grid peaks—could fix this, but only if customers and CEOs stop demanding perfection. The real lesson: NVIDIA’s edge comes from betting the company on the right abstractions, enduring the pain, and building for where the world is going, not where it is.

Why it lands

Huang’s story is a playbook for platform dominance: win developers by any means, even if it means risking everything. His approach to team structure and power grid efficiency is a model for anyone facing scaling bottlenecks. The message: bold bets and collective problem-solving beat incremental tweaks and lone geniuses.

The CUDA Gamble That Nearly Sank NVIDIA

Huang describes the decision to put CUDA on GeForce GPUs—a move that gutted profits and market cap, but seeded the install base that made NVIDIA the backbone of AI.

  • CUDA on GeForce wiped out NVIDIA’s gross profit, dropping market cap from $8B to $1.5B.
  • The move was about reaching every PC user, not just workstations.
  • Huang calls it 'as close to an existential threat' as NVIDIA ever faced.

Extreme Co-Design: How NVIDIA Actually Works

NVIDIA’s 60-person expert team solves problems together, not in silos. No one-on-ones—just collective engineering.

  • All meetings are group-based to drive extreme co-design.
  • Team includes experts in memory, CPUs, GPUs, algorithms, and more.
  • This structure lets NVIDIA anticipate and build for future AI workloads.

Agentic AI: The Next Scaling Challenge

Huang explains how agentic systems—AIs that spawn sub-agents and use tools—demand new hardware and software, and why NVIDIA is already building for them.

  • OpenClaw is to agentic systems what ChatGPT was to generative AI.
  • NVIDIA’s Vera Rubin racks are designed for these agentic workloads.
  • Agentic scaling multiplies AI capabilities and generates new data for training.

Why Data Centers and Power Grids Need to Change

Huang argues that data centers and power grids are built for rare peaks but run at 60% capacity most of the time, wasting energy and money.

  • Dynamic power allocation could let data centers reduce consumption during grid peaks.
  • Customer demand for 'six nines' uptime locks in inefficiency.
  • Huang says this is an engineering problem, not a law-of-physics limit.

Worth stealing

  • Install base, not technical elegance, decides platform winners.
  • Extreme co-design lets NVIDIA adapt to new AI workloads before they arrive.
  • Agentic AI will force a rethink of hardware and infrastructure.
  • Power grid and data center inefficiency is an engineering problem, not a physical law.

Lines worth repeating

  • That was the first strategic decision that is as close to an existential threat.

    Jensen Huang

  • Install base is everything.

    Jensen Huang

  • No conversation is ever one person. That’s why I don’t do one-on-ones.

    Jensen Huang

  • OpenClaw did for agentic systems what ChatGPT did for generative systems.

    Jensen Huang

Lex Fridman: Jensen Huang on NVIDIA’s Existential CUDA Gamble, Agentic AI, and the Real Limits of Scaling | Briefly Heard