Jensen Huang on building NVIDIA from near-death to AI dominance

Executive overview

NVIDIA survived multiple near-death moments by betting everything on a single, fully-simulated chip rather than iterating through failure. The company's 30-year edge comes not from luck but from systematically positioning near emerging opportunities decades before markets materialise.

Jensen's core framework: simulate everything you can before committing, target zero-billion-dollar markets a decade early, and wire the organisation like a neural network rather than a command hierarchy.

When you bet the farm, you've already run the simulation — the bet is the confirmation, not the gamble.

The Riva 128 moment: one shot, no prototype

• In 1997, NVIDIA had six months of cash and couldn't afford a failed tape-out
• Bought a chip emulator from a company shutting down; wrote the entire software stack against it
• Ran full QA — games, drivers, VGA apps — before a single physical chip was produced
• When the chip came back, they went straight to production and marketing simultaneously
• Lesson: prefetching the future — doing everything risky in advance — is now a permanent operating principle

How NVIDIA reasoned its way to AI

• AlexNet (2012) forced Jensen to return to first principles: why did it leap 30 years of computer vision work?
• Key insight: a sufficiently large neural network is a universal function approximator — causality is irrelevant, only prediction matters
• Applicable to commerce, weather, drug discovery, social feeds — essentially every industry
• NVIDIA was already embedded with researchers via CUDA, making it natural to back every major AI lab early
• Jensen personally tracked arXiv; paper velocity from quarterly to daily confirmed the exponential
• First DGX was delivered personally to OpenAI — framed as helping researchers, not chasing a market

CUDA: the non-negotiable platform decision

• Before CUDA there was CG, before that DirectNV — NVIDIA has always been a developer-platform company
• The one unnegotiable rule: every NVIDIA chip must be architecturally compatible with every other
• ~250–300 million active CUDA GPUs form an installed base no competitor can replicate
• CUDA turned gaming revenue into an install base for scientific and then AI computing — each paid for the next

The data centre: a 17-year setup

• NVIDIA saw that tying compute to a display would cap the market; separating them explodes opportunity
• GeForce Now (cloud gaming) was NVIDIA's first data centre product — started ~18 years before this interview
• Progression: cloud gaming → remote enterprise graphics → CUDA supercomputing → AI data centres
• Each step built operational knowledge required for the next; they were ready when LLMs arrived

Mellanox and why networking defines a data centre

• A data centre isn't defined by its processors — it's defined by its networking and infrastructure
• Hyperscale (cloud) = many users sharing one computer; AI = one job distributed across millions of processors
• Commodity Ethernet is fine for Hadoop and search; sharding a model across racks requires high-performance fabric
• Mellanox was the world leader in that fabric; the acquisition is widely considered one of the best in tech history

Zero-billion-dollar markets: NVIDIA's competitive strategy

• Enter markets before they exist — automotive, supercomputing, AI — so competitors aren't yet shaped for them
• By the time the market arrives, NVIDIA has a platform, an ecosystem, and skills competitors can't quickly acquire
• Moat framing is wrong: don't build walls around a castle, build a network where others succeed alongside you
• Omniverse cited as the current zero-billion-dollar bet (40 customers at time of recording)

Org structure: computing stack, not command hierarchy

• 40+ direct reports; information disseminated to everyone simultaneously, including new grads
• "Mission is the boss" — teams wire up across the organisation to a specific deliverable, not a reporting line
• Authority comes from reasoning ability, not proximity to information
• The org chart should mirror the architecture of what you're building — not look like every other company's

Founder advice and the emotional reality

• Starting a company is "a million times harder than expected" — the superpower of founders is not knowing this
• Luck matters: Carmack adopting OpenGL for Quake created the killer app that validated consumer 3D
• You can do all the right things and fail; you can do many wrong things and succeed
• What sustains you: unwavering support from co-founders, long-tenure colleagues, family, and board members who never quit
• Jensen's greatest fear, then and now: letting employees down — they adopted your hopes and dreams as their own
• On AI and jobs: productivity creates prosperity; prosperous companies hire more people to pursue more ideas; net employment rises, but job types shift — learn AI now

On Jensen's leadership philosophy

• Read widely, but always ask "what does this mean to me in my context?" — never imitate directly
• On time: "Don't let Outlook control your day" — prioritise ruthlessly, there is enough time
• On NVIDIA's scale: once you stop being a chip company and become an AI company, the addressable market grows by ~1,000x
• "We're in the manufacturing of intelligence" — measured in trillions, not billions