Mar 23, 2026 · 2h 32m
NVIDIA CEO Jensen Huang details the hardware engineering driving the AI revolution
#494 – Jensen Huang: NVIDIA – The $4 Trillion Company & the AI Revolution
As AI models demand unprecedented computational power, understanding NVIDIA's hardware strategy reveals where the limits of machine intelligence will be pushed next.
- 1NVIDIA relies on extreme co-design to build hardware and software in lockstep for maximum computing efficiency.
- 2The frontier of AI is shifting toward agentic scaling laws that require massive GPU clusters to support autonomous reasoning.
- 3Proprietary interconnect technologies like NVLink are critical to overcoming communication bottlenecks between modern GPUs.
Don't miss
Jensen Huang explains how the transition to agentic scaling laws will reshape the future of computational workloads.
The brief
NVIDIA CEO Jensen Huang sits down with Lex Fridman to trace the company's evolution from a specialized graphics accelerator pioneer into the multi-trillion-dollar engine driving the global artificial intelligence revolution.
Huang attributes the company's dominance to extreme co-design, a philosophy where hardware and software are built in lockstep. This unified approach allowed proprietary technologies like CUDA and NVLink to become industry standards.
The discussion details how the scaling laws of large language models are shifting toward agentic scaling, where autonomous AI systems require massive, highly interconnected GPU clusters to execute complex reasoning tasks.
Beyond raw computing power, Huang shares his unconventional leadership philosophy, emphasizing flat organizational structures and a first-principles approach to solving seemingly impossible engineering bottlenecks.
What was said on this episode
41 statements · 31 positive · 5 negative · 2 mixed · 3 neutral
Extreme co-design is necessary because modern AI problems exceed one computer’s capacity.
“the reason why extreme code design is necessary is because the problem no longer fits inside one computer to be accelerated by one gpu”
Listen at 7:15
Becoming more general-purpose creates tension with specialized computing optimization.
“The better computing company we become, the worse we became as a specialist.”
Listen at 14:41
An architecture’s installed base is its most important strategic asset.
“the install base is in fact the single most important part of an architecture”
Listen at 18:10
Nvidia should distribute CUDA through GeForce PCs to build an installed base.
“we ought to put CUDA on GeForce and put it into every single PC”
Listen at 19:36
Nvidia’s market capitalization fell to approximately $1.5 billion after CUDA launched.
“our market cap went down to like one and a half billion dollars”
Listen at 20:52
Adding CUDA increased Nvidia’s GPU costs by approximately 50%.
“we increased our cost by 50%”
Listen at 22:17
AI training data will continue scaling, increasingly using synthetic data.
“we're going to keep on scaling the amount of data that we have to train with. A lot of that data is probably going to be synthetic.”
Listen at 29:59
AI training will become compute-limited rather than data-limited.
“Training is no longer limited by data, is now limited by compute.”
Listen at 31:08
Test-time scaling and inference require substantial computational resources.
“test time scaling is intensely compute intensive”
Listen at 32:58
Agentic systems introduce a further scaling law by multiplying AI agents.
“the next scaling law is the agentic scaling law”
Listen at 33:46
CUDA’s flexibility has allowed it to remain resilient while adapting to new algorithms.
“CUDA has been so resilient on the one hand, and yet we continue to enhance it”
Listen at 36:40
NVLink 72 can place models with four to ten trillion parameters in one computing domain.
“We could now take an entire 4 trillion, 10 trillion parameter model and put it in one computing domain as if it's running on one gpu.”
Listen at 37:05
Agentic AI systems will use tools, access files, and conduct research.
“it's going to use tools, that it's going to access files, it's going to be able to do research”
Listen at 40:37
Agentic systems can be made safer by restricting simultaneous access to three capabilities.
“We could keep things safe if we gave you two out of those three capabilities at any time, but not all three.”
Listen at 43:15
Nvidia claims computing scale increased one million-fold over the past decade.
“We progressed and scaled up computing by a million times in the last 10 years.”
Listen at 44:25
AI token costs will decline by roughly an order of magnitude annually.
“that token cost is coming down, it's coming down an order of magnitude every year”
Listen at 45:00
HBM memory will become mainstream in data centers.
“this was going to be a mainstream memory for data centers in the future”
Listen at 47:45
The power grid has excess capacity almost all the time because it is sized for rare peaks.
“99% of the time our power grid has excess power”
Listen at 54:10
Data centers should receive less power during periods of peak societal demand.
“the data centers would get less”
Listen at 54:48
Data centers should be engineered to reduce performance without losing data during power constraints.
“we have to build data Centers that gracefully degrade”
Listen at 57:11
Elon Musk uses systems thinking to minimize processes while preserving necessary capabilities.
“he has the ability to question everything to the point where everything is down to its minimal amount that's necessary”
Listen at 59:50
Nvidia evaluates engineering constraints against physical limits.
“everything that we do is compared against the speed of light”
Listen at 1:02:56
Huang prefers first-principles redesign over incremental continuous improvement.
“I don't love the other methods, which is continuous improvement”
Listen at 1:04:15
China is currently the world’s fastest-innovating country.
“this is the fastest innovating country in the world today”
Listen at 1:10:43
Open-source AI is necessary for broad industrial adoption and innovation.
“open source is fundamentally necessary for many industries to join the AI revolution”
Listen at 1:13:49
Nvidia’s computing-platform installed base is its most important company property.
“Our single most important property as a company is the install base of our computing platform.”
Listen at 1:21:21
Huang expects Nvidia’s computing model to expand toward planetary scale.
“I'm hoping my next click is when I'm thinking about building computers. It's, you know, planetary scale.”
Listen at 1:26:36
Satellite imagery should be processed by AI at the edge rather than transmitted wholesale.
“AI ought to be done at the edge”
Listen at 1:28:08
Nvidia’s growth is extremely likely and effectively inevitable.
“Nvidia's growth is extremely likely and in my mind inevitable”
Listen at 1:30:48
Customers will soon pay approximately $1,000 per million AI tokens.
“The idea that somebody's willing to pay $1,000 per million tokens, it's just around the corner.”
Listen at 1:34:24
AI-driven productivity will accelerate global GDP growth.
“I am absolutely certain that the world's GDP is going to accelerate in growth.”
Listen at 1:35:17
Computing’s share of global GDP will become approximately one hundred times larger.
“the percentage of that GDP that will be used for computation will be a hundred times more than the past”
Listen at 1:35:25
OpenClaw is the fastest-growing application ever.
“It is the fastest growing application in history.”
Listen at 1:39:23
Huang believes artificial general intelligence has already been achieved.
“I think it's now, I think we've achieved AGI.”
Listen at 2:02:33
Nvidia’s software-engineer headcount will increase despite AI automation.
“The number of software engineers at Nvidia is going to grow, not decline.”
Listen at 2:07:04
AI may expand the number of people able to code from 30 million to one billion.
“I think we just went from 30 million to probably 1 billion.”
Listen at 2:08:19
Students and workers should become proficient AI users to improve their careers.
“I would advise that every college student, every teacher should encourage their student to go use AI.”
Listen at 2:13:58
AI chips will not experience human anxiety, excitement, or other emotions.
“I don't think my chips will feel those.”
Listen at 2:17:39
Intelligence is a functional commodity distinct from humanity.
“I actually think intelligence is a commodity”
Listen at 2:20:28
Ending disease is a reasonable future expectation.
“the end of disease”
Listen at 2:29:02
Human biology will be substantially understood within approximately five years.
“Understanding the biological machine is right around the corner. It's not 10 years, it's five years probably.”
Listen at 2:30:16
Statements are attributed to the speaker as said on the episode and reflect their view at the time, not PodLume's. They are not advice.
Books & mentions
Lex Fridman Podcast
Explore more deep-dive conversations with the world's leading AI researchers and tech executives