Jensen Huang — NVIDIA & the AI Revolution

TL;DR

The four ideas that hold the whole conversation together

The problem no longer fits inside one computer. So NVIDIA stopped designing chips and started designing racks, pods and data centers as one system.
Install base defines an architecture. That is the whole reason CUDA went onto every GeForce card — even when it nearly killed the company.
Scaling is not over. Data stopped being the limit; compute is the limit now. Beyond test-time comes agentic scaling.
Computers are no longer warehouses. They are factories that generate tokens, and tokens are the product.

~50%

of the world's AI researchers are Chinese, mostly still in China

1.3M

components in a single NVIDIA rack, from 200 suppliers

1M×

compute scaling in ten years, versus ~100× from Moore's law

~60%

of peak the grid runs at, about 99% of the time

Extreme co-design: what it is, why it exists

The old game — build the best GPU you can — is over. The new game is co-designing GPU, CPU, memory, networking, storage, power, cooling, software, the rack, the pod and the data center as a single system.

Why it is necessary: you want to go faster than the number of computers you add. Add ten thousand machines, aim for a million times faster — that is no longer one computer's problem.
So the algorithm gets broken up, refactored, sharded across pipeline, data and model.
Amdahl's law bites: if computation is only 50% of the workload, speeding it up a million times still only doubles the total.
Now everything is in the way at once — CPU, GPU, switching, networking, power delivery, cooling.

You have to bring every technology to bear — otherwise you scale up linearly, on a Moore's law that has largely slowed.

The company is designed like the product

NVIDIA does extreme co-design internally too. Jensen has 60 direct reports and holds no one-on-ones.

One problem is presented, and everyone attacks it together — memory, CPU, optics, algorithms, architecture in the same room.
Anyone who tunes out is called back in: "You could have contributed. Let's get in here."
Most of the staff keeps a foot in engineering.

The architecture of the company should reflect the environment in which it exists. A company is the machinery that produces the output.

The CUDA bet that nearly ended NVIDIA

CUDA turned out to be the foundation of the deep learning era. It almost didn't survive to see it.

Install base is everything. x86 is not an elegant architecture, and it still became the defining one. Beautifully designed rivals died.
So CUDA was put on every GeForce card — whether or not anyone used it — as the starting point for an install base.
It added roughly 50% to the cost of a consumer GPU, inside a company running about 35% gross margins.
NVIDIA's market cap fell to about $1.5 billion. It clawed its way back over roughly a decade.
Developers found CUDA on GeForce in university labs, built their own clusters, and taught themselves.

NVIDIA is the house that GeForce built.

Four scaling laws, not one

Pre-training. More model plus more data equals a smarter AI. Data is no longer the ceiling — compute is.
Post-training. Most of what humans teach each other is already synthetic: created, consumed, modified, regenerated. AI can now augment ground truth at enormous scale.
Test-time. Inference is thinking, and thinking is hard — reasoning, planning, search, decomposing new experience into solvable pieces.
Agentic. One agent spawns sub-agents, uses tools, bangs on databases. You are multiplying AI the way you multiply a workforce.

And the loop closes: agents generate experience, experience becomes data, data feeds pre-training, which feeds post-training, which feeds the agents again.

Pre-training is just memorization and generalization. Inference is thinking — how could that possibly be compute-light?

From warehouse to factory

Computing used to be retrieval: pre-write it, store it, then recommend it. It is now generation: context-aware, situational, produced in real time.
That is why compute demand is not going back — the old model stored, the new model computes.
The unit of computing moved: GPU → computer → cluster → entire AI factory. Jensen's next mental model is planetary scale.
Tokens are the product, so power efficiency is literally revenue: tokens per second per watt.
Token tiers are emerging like phones — free, premium, and everything between. He expects a market for intelligence at around $1,000 per million tokens to be close.

Warehouses don't make much money. Factories directly correlate with revenue.

The power bottleneck is mostly waste

The grid is built for the worst case — a few brutal days a year — and sits far below that the rest of the time.

Three fixes, all reasoned from first principles:
Customers: many CEOs don't realise their data centers are contractually demanding perfection. Reopen that conversation.
Data centers: build them to degrade gracefully. If the utility says 80%, shift workload, slow the rate, never lose data.
Utilities: offer segmented power promises at different prices and different guarantees. Excess capacity is already sitting there.

So long as it obeys the laws of physics on first principles, I think we're good.

A supply chain built on belief, not contracts

A single Vera Rubin rack carries 200 suppliers and about 1.3 million components. One pod: 40 racks, 1.2 quadrillion transistors, 60 exaflops, 10 petabytes per second of scale bandwidth.

NVIDIA moved supercomputer assembly out of the data center and into the supply chain — racks now ship two to three tons at a time.
Memory CEOs were persuaded to invest billions in HBM and low-power phone memory for data centers, years before it was mainstream.
He tells the whole industry what he needs, then trusts their word. That is why he does not lose sleep over the backlog.

I spend my time informing and inspiring the supply chain — informing them like I inform my own employees.

"Speed of light" thinking

Every design is measured against the physical limit — memory speed, math speed, power, manufacturing cycle time, even headcount.

Low latency and high throughput require fundamentally different architectures. Know the physical limit of each, then trade.
He rejects continuous improvement as a starting point: "Don't tell me 74 days becomes 72. Why is it 74 days at all?"
Strip it back to zero and rebuild. It might come out at six days — and now everyone can see the compromises.
Elon's approach is the same instinct: question everything until only the necessary remains, and be present at the point of action.

As complex as necessary, but as simple as possible.

China, and the TSMC relationship

China: roughly half the world's AI researchers. A tech industry that arrived exactly at the mobile-cloud era. Provinces and mayors competing furiously, producing brutal domestic competition and world-class survivors.
A culture of family first, friends second, company third — so knowledge moves fast and open source makes sense. "What are we protecting?"
TSMC: not just transistors. It is the ability to orchestrate the wildly shifting demand of hundreds of companies at once, with high yield and kept promises.
Technology plus customer service plus trust — very few companies hold all three at world class.
Three decades, hundreds of billions of dollars of business, no contract.
Morris Chang once offered him the TSMC CEO job. He declined: the work at NVIDIA was the work he had to do.

Why NVIDIA gives its models away

To be a great AI computing company you must understand how models are evolving — that knowledge shapes the hardware.
Proprietary frontier models and open models can coexist. If everything is proprietary, most industries simply cannot join.
AI is not just language: biology, chemistry, physics, fluids, weather. Someone has to push those frontiers too.
NVIDIA open-sourced the models, the weights, the data, and how it was made.

We don't build cars, but every car company should have great models. We don't discover drugs, but the best biology AI should be available to those who do.

AGI, jobs and the radiologist lesson

By his own narrow definition — an agent that can start, grow and run a company worth more than a billion dollars — Jensen says we have achieved AGI. Millions of small apps and viral experiments are already within reach. One hundred thousand of those agents building NVIDIA: zero percent.

The radiologist story: computer vision became superhuman around 2019–2020. Radiologists were supposed to disappear. Instead their numbers grew — and the world now has a shortage.
The purpose of the job and the tasks inside it are related, not the same. Faster scans meant more scans, more diagnoses, more patients.
Coding is specification. If that is what coding means, the number of people who can do it goes from roughly 30 million to perhaps a billion.
Carpenters become architects of their own trade. Accountants become financial advisors. The tools amplify the artistry.
If your job is the task, you are exposed. Automate your own tasks, and the anxiety has somewhere to go.

If I had two graduates in front of me, I'd hire the one who is an expert at using AI.

The resilience method

Decompose. Break the pressure into pieces small enough to act on.
Share the load. Whatever worries him, he tells someone about it immediately — pass it on, don't carry it.
Forget. Systematic forgetting, the same as in machine learning. You cannot hold everything.
Refuse to panic. Make the list, act on what you can, then sleep.
Tolerate embarrassment. He reasons out loud in public, so everyone can intercept and disagree with a step — not just the conclusion.
Keep the mind of a child. "How hard could it be?" — the question that starts the impossible project.

If you didn't do it, and you didn't get anybody else to do it, then stop crying about it.

Intelligence is a commodity. Humanity is not.

He sits at the centre of 60 people who, in their own fields, are more educated and more capable than he is. His point: intelligence is a functional thing — perception, understanding, reasoning, planning. It is not the same word as being human.

What he wants to elevate: character, compassion, generosity, tolerance for pain, determination.
On mortality: "I really don't want to die." But he does not believe in succession planning — he believes in transferring knowledge continuously, every single day.
He wants to die on the job, and quickly.
What gives him hope: a standing belief that people want to do good — and that he is proven right by it constantly.

Lines worth keeping

"Inference is thinking. And I think thinking is hard."

"Install base defines an architecture. Everything else is secondary."

"The purpose of your job and the tasks you use to do your job are related, not the same."

"I actually think intelligence is a commodity."

"What we should really elevate is humanity, character."

"How hard could it be?"

Five things to take into next week

Become the person in your field who is genuinely expert at using AI. That is now a hiring criterion, not a bonus.

List the tasks inside your job. Start automating those tasks yourself, before someone else's tool does it around you.

Take your biggest worry, cut it into pieces small enough to act on, and hand each piece to someone who can move it.

Redesign one thing from first principles instead of optimising it. Ask what the physical limit even is.

Pass on what you learn the moment you learn it. Nothing should sit on your desk.

The arc of the conversation is surprisingly simple: NVIDIA became central because it kept widening what "the computer" means — from a chip, to a rack, to a pod, to an entire factory that turns electricity into tokens.

Almost everything Jensen says about scaling comes back to one habit: reason from first principles until the outcome is obvious, then spend years making everyone around you believe it too.

And his closing note, borrowed from Alan Kay: the best way to predict the future is to invent it.

Jensen Huang — NVIDIA &amp; the AI Revolution

Jensen Huang — NVIDIA & the AI Revolution