Jensen Huang: NVIDIA - The $4 Trillion Company & the AI Revolution | Lex Fridman Podcast #494

· Source ↗

Jensen Huang explains NVIDIA’s four scaling laws, $1.5B near-death moment, and why install base beats technology as a moat

  • CUDA adoption dropped NVIDIA market cap to $1.5B after launch — it added 50% cost to GeForce GPUs and consumed all gross profit of a then-35% margin company; recovery took a decade.
  • NVIDIA’s #1 moat is CUDA install base, not chip technology — Jensen argues install base defines an architecture, citing x86 surviving despite being “barely aesthetic” while elegant RISC architectures failed.
  • Four scaling laws: pre-training, post-training (synthetic data), test-time compute (inference = thinking, not easy), and agentic (spawn agents like hiring employees) — each requires different hardware anticipation 2-3 years out.
  • Token cost drops an order of magnitude per year while chip prices rise — NVIDIA scaled compute 1 million times in 10 years vs Moore’s law’s 100x over the same period.
  • 50% of world’s AI researchers are Chinese; Jensen calls China “the fastest innovating country in the world” — attributes it to open-source culture, provincial competition (same reason 100+ EV companies exist), and engineering social status.
  • Power grid runs at ~60% capacity 99% of the time — Jensen’s pitch: data centers should contractually accept graceful degradation so utilities can offer excess idle power immediately instead of 5-year grid expansion timelines.
  • TSMC offered Jensen the CEO role in 2013; he declined — they’ve done hundreds of billions in business across three decades with no formal contract.
  • Vera Rubin pod: 7 chip types, 5 rack types, 40 racks, 60 exaflops, 1.2 quadrillion transistors — NVIDIA ships ~200 pods/week; NVL72 rack alone is 1.3M components and ships pre-assembled at 2-3 tons because on-site assembly is now impossible.
  • Jensen predicted OpenClaw’s exact architecture (file access, tool use, sub-agents, IO subsystem) two years before its release at GTC — used as proof of first-principles reasoning over roadmap leaks.
  • Nemotron 3 Super (120B params, hybrid transformer+SSM) fully open-sourced including weights, data, and training methodology — strategic: lets NVIDIA understand model evolution to co-design future hardware.

Guests: Jensen Huang (NVIDIA CEO & co-founder) · 2026-03-23 · Watch on YouTube