EVERY SPOKEN WORD
150 min read · 30,037 words- 0:00 – 3:07
Reconnecting: SpaceX chip handoff and a surprise Trump phone call
- NANarrator
(drumming music) Joe Rogan podcast, check it out.
- JHJensen Huang
The Joe Rogan Experience.
- JRJoe Rogan
Train by day, Joe Rogan podcast by night, all day. (rock music) Hello, Jensen.
- JHJensen Huang
Hey, Joe.
- JRJoe Rogan
Good to see you again.
- JHJensen Huang
It's great.
- JRJoe Rogan
We were just talking about ... Was that the first time we ever spoke? Or did ... Was the first time we spoke at, at SpaceX?
- JHJensen Huang
SpaceX.
- JRJoe Rogan
SpaceX the first time, when you were giving Elon that crazy AI chip.
- JHJensen Huang
Right, DGX Spark.
- JRJoe Rogan
Yeah. Ooh, that was a big moment.
- JHJensen Huang
That was a huge moment.
- JRJoe Rogan
That felt crazy to be there. It was like w- watching these wizards of tech, like, exchange information and, and w- you're giving him this crazy device, you know? And then the other time was, uh, I was shooting arrows in my backyard. And, uh, randomly get this call from Trump and he's hanging out with you.
- JHJensen Huang
President Trump called and I called you.
- JRJoe Rogan
Yeah, it's just, he-
- JHJensen Huang
We were talking about you.
- JRJoe Rogan
(laughs) He, he's so ... It's just weird that he's calling-
- JHJensen Huang
We were talking about you, he w- talking about the US- UFC thing he was gonna do in his front yard.
- JRJoe Rogan
Yeah.
- JHJensen Huang
And he pulls out, he's, "Jensen, look at this design." He's so proud of it. And I go, "You're gonna have a fight in the front lawn of The White House?" He goes, "Yeah. Yeah, you're gonna come. This is gonna be awesome." And he's showing me his design and how beautiful it is. And he goes, a- and somehow your name comes up. He goes, "Do you know Joe?" And I said, "Yeah, I'm gonna be on his podcast." He's, "Let's call him."
- JRJoe Rogan
(laughs)
- JHJensen Huang
(laughs)
- JRJoe Rogan
He's like a kid.
- JHJensen Huang
I know, let's call him.
- JRJoe Rogan
It's so ... He's like a 79-year-old kid.
- JHJensen Huang
And he tells you. Ah, he's so incredible.
- JRJoe Rogan
Yeah, it was ... He's an odd guy. Just very different, you know, like, than what you'd expect from him. Very different than what people think of him, and also just very different as a president. A guy who just calls you or texts you out of the blue. Also, he makes ... When he te- he texts you ... You have an Android so it won't go through with you, but with my iPhone, he makes the text go big. Like, you know-
- JHJensen Huang
Is that right?
- JRJoe Rogan
"USA is respected again," like ... (laughs)
- JHJensen Huang
(laughs)
- 3:07 – 6:56
Manufacturing, national security, and why energy policy matters for AI buildout
- JHJensen Huang
Yeah, but, but I, I like the fact that he's telling you what's on his mind. Um, almost every time he explains something and he says something, he starts with his, you could tell, his love for America, what he wants to do for America. And everything that he thinks through is very practical and very common sense. And, you know, it's very logical. And, um, I still remember the first time I, I met him. And so this was ... I, I'd never known him, never met him before. And, um, uh, Secretary Luttnik called and we met right before, right at the beginning of the administration. And he said, he told me what was important to President Trump. That, that, um, uh ... That United States manufactures on shore. And that was really important to him because, because, uh, it's important to national security. He wants to make sure that, that the important critical technology of our nation is built in United States and that we re-industrialize and get good at manufacturing again because it's important for jobs.
- JRJoe Rogan
It just seems like common sense, right?
- JHJensen Huang
Incredible common sense and, and almost, like, literally the first conversation I had with Secretary Luttnik. And, um, and he was talking about how, how, um, uh, that ... He started, he started our conversation with, uh, "Jensen, this is Secretary Luttnik. And, uh, uh, I just wanna let you know that you're a national treasure. Uh, NVIDIA's a national treasure. And whenever you need access to the President, um, the administration, uh, you call us. We're always gonna be available to you." Literally, that was the first sentence.
- JRJoe Rogan
That's pretty nice.
- JHJensen Huang
And it was completely true. Every single time I called if I needed something, wanted to get something off my chest, um, express some concern, uh, they're always available. Incredible.
- JRJoe Rogan
It's just unfortunate we live in such a politically polarized society that you can't recognize good common sense things if they're coming from a person that you object to. And that, I think, is what's going on here. I think most people generally, a- as a country, you know, as a, a giant community, which we are, it just only makes sense that we have manufacturing in America. That it ... Especially critical technology like you're talking about. Like, it's kind of insane that we buy so much technology from other countries.
- JHJensen Huang
If United States doesn't grow, we will have no prosperity. We can't invest in anything domestically or otherwise. We can't fix any of our problems. If we don't have energy growth, we can't have industrial growth. If we don't have industrial growth, we can't have job growth. These ... It's as simple as that.
- JRJoe Rogan
Right.
- JHJensen Huang
And the fact that, the fact that he came into office and the first thing that he said was, "Drill, baby, drill," his point is we need energy growth. Without energy growth-... we can have no industrial growth. And that was, it saved, it saved the AI industry. Got- I gotta tell you flat out, if not for his pro-growth energy policy, we would not be able to build factories for AI, we would not be able to build chip factories, we won't sure- surely won't be able to build super computer factories. None of that stuff would be possible. Without all of that, construction jobs would be challenged, right? Electrical, you know, electrician jobs, all of these jobs that are now flourishing would be challenged. And so I think he's got it right. We need energy growth, we want to re-industrialize the United States. We need to be back in manufacturing. Every successful person doesn't need to have a PhD. Every successful person doesn't have to have gone to Stanford or MIT. And I think, I think that that, that p- you know, that sensibility is, is, um, spot on.
- 6:56 – 10:04
The AI “race”: continuous tech competition and the unknown destination
- JRJoe Rogan
Now, when we're talking about technology growth and energy growth, there's a lot of people that go, "Oh, no. That's not what we need. We need to, you know, simplify our lives and get back..." But the, the real issue is that we're in the middle of a giant technology race. And whether people are aware of it or not, whether they like it or not, it's happening, and it's a really important race. Because whoever gets to whatever the event horizon of artificial intelligence is, whoever gets there first has massive advantages, in a huge way. Do you agree with that?
- JHJensen Huang
Well, first, the, the part... Uh, I will say that we are in a technology race, and we are always in a technology race. We've been in a technology race with somebody forever.
- JRJoe Rogan
Right.
- JHJensen Huang
Right? Since the Industrial Revolution, we've been in a technology race.
- JRJoe Rogan
Since the Manhattan Project.
- JHJensen Huang
Yeah.
- JRJoe Rogan
Yeah.
- JHJensen Huang
Or, or, you know, even going back to the discovery of energy, right? The United Kingdom was where the Industrial Revolution was, if you will, invented, when they realized that they can turn steam and such into, into energy, into electricity. All of that was invented largely in Europe, and United States capitalized on it. We were the ones that learned from it, we industrialized it, we diffused it faster than anybody in Europe. They were all stuck in discussions about policy and jobs and disruptions. Meanwhile, the United States was forming. We just took the technology and ran with it. And so I, I think we were always in, in a bit of a technology race. World War II was a technology race. Manhattan Project was a technology race. We've been in the technology race ever since, during the Cold War. I think we're still in a technology race. It is probably the single most important race. It is the... Technology is, uh, it gives you superpowers. You know, whether it's information superpowers or energy superpowers or military superpowers, it's all founded in technology. And so, technology leadership is really important.
- JRJoe Rogan
Well, the problem is if somebody else has superior technology, right?
- JHJensen Huang
Yeah.
- JRJoe Rogan
That's the, that's the issue.
- JHJensen Huang
That's right.
- JRJoe Rogan
And it seems like with the AI race, people are very nervous about it. Like, you know, Elon has famously said there was like 80% chance it's awesome, 20% chance (laughs) we're in trouble. And people are worried about that 20%, rightly so. I mean, that, you know, if you had 10 bullets in a, a, a, a, a revolver, and, you know, you, you took out eight of them, and you still have twen- two in there, and you spin it, you're not gonna feel real comfortable when you pull that trigger. It's terrifying.
- JHJensen Huang
Right.
- JRJoe Rogan
And when we're working towards this ultimate goal, um, of AI, it, it just, it's impossible to imagine that it wouldn't be of national security interest to get there first.
- JHJensen Huang
We should... The question is, what's there?
- JRJoe Rogan
Right.
- JHJensen Huang
That's the, that was the part that-
- JRJoe Rogan
What is there?
- JHJensen Huang
Yeah, I'm not sure. And I don't think anybody-
- JRJoe Rogan
That's the problem.
- JHJensen Huang
I don't think anybody really knows.
- JRJoe Rogan
(laughs) That's crazy, though.
- JHJensen Huang
Yeah.
- JRJoe Rogan
If I ask you-
- JHJensen Huang
Yeah.
- JRJoe Rogan
... you're the head of NVIDIA.
- JHJensen Huang
Yeah.
- JRJoe Rogan
If you don't know-
- JHJensen Huang
Yeah.
- 10:04 – 15:18
AI capability gains and why more compute often becomes ‘safer’ behavior
- JHJensen Huang
Yeah, I, I think it's probably gonna be much more gradual than we think. It's wo- it won't be a moment. It won't be, it won't be as if, um, somebody arrived and nobody else has. I don't think it's gonna be like that. I think it's gonna be things that just get better and better and better and better, just like technology does.
- JRJoe Rogan
So you are rosy about the future. You're, you're very optimistic about what's gonna happen with AI. Obviously, will you make the best AI chips in the world?
- JHJensen Huang
Um...
- JRJoe Rogan
You'd probably better be.
- JHJensen Huang
Uh, h- i- if history is a guide, um, uh, we were always concerned about new technology. Humanity has always been concerned about new technology. There are always somebody who's thinking, there are always a lot of people who are quite concerned, were quite concerned. And, and, and so if, if history is a guide, it is the case, um, that all of this concern is channeled into making the technology safer. And so, for example, in the last several years, I would say AI technology has increased probably, in the last two years alone, maybe 100X. Let's just give it a number. Okay? It's like a car, two years ago, was 100 times slower. So AI is 100 times more capable today. Now, how did we channel that technology? How do we channel all of that power? We directed it to, um, causing the AI to be able to think, meaning that it can take a problem that we give it, break it down step by step. It does research before it answers, and so it grounds it on truth. It'll reflect on that answer, ask itself, "Is this the best, you know, answer that I can give you? Am I certain about this answer?" If it's not certain about the answer or highly confident about the answer, it'll go back and do more research. It might actually even use a tool, because that tool provides a better solution than it could...... hallucinate itself. As a result, we took all of that computing capability and we channeled it into having it produce a safer result, safer answer, a more truthful answer. Because as you know, one of the greatest criticisms of AI in the beginning was that it hallucinated.
- JRJoe Rogan
Right.
- JHJensen Huang
And so, if you look at the reason why people use AI so much today, is because the amount of hallucination has reduced. You know, I use it almost, I... well, I used it the whole trip over here. And so, so I think the, the, uh, the capability... most people think about power and they think about, you know, maybe as an explosion power. But the technology power, most of it is channeled towa- towards safety. A car today is more powerful, but it's safer to drive. A lot of that power goes towards better handling. You know, I'd rather have a... well, you have a 1,000-horsepower truck.
- JRJoe Rogan
(laughs)
- JHJensen Huang
I think 500 horsepower is pretty good. No, I th-... 1,000's better. I think 1,000's better.
- JRJoe Rogan
I don't know if it's better, but it's definitely faster.
- JHJensen Huang
Yeah.
- JRJoe Rogan
(laughs)
- JHJensen Huang
No, I think it's better. Uh, uh, you get out of trouble faster. Um... I enjoyed my 599 more than my 612. It was... I think it was better, better. More horsepower is better. My 459 is better than my 430. More horsepower is better. I think more horsepower is better. I think it's better handling, it's better control. In the case of, in the case of technology, it's also very similar in that way, you know? And so if you, if you look at what we're gonna do with the next 1,000 times of performance in AI, a lot of it is going to be channeled towards more reflection, more research. Thinking about the answer more deeply.
- JRJoe Rogan
So, when you're defining safety, you're defining a- it as accuracy?
- JHJensen Huang
Functionality.
- JRJoe Rogan
Functionality, okay.
- JHJensen Huang
Yeah. It, it does what you expect it to do. And then, you take the, all the, the technology and the horsepower, and you put guardrails on it. Just like our cars. We've got a lot of technology in, in a car today. A lot of it is... goes towards... for example, ABS. ABS is great. And so, uh, traction control, that's fantastic. Without a s-... without a computer in the car, how would you do any of that?
- JRJoe Rogan
Right.
- JHJensen Huang
And that little computer, the computers that you have doing your traction control is more powerful than the computer that went to Apollo 11. And so, you want that technology. Channel it towards safety, channel it towards functionality. And so, when people talk about power, the advancement of technology, oftentimes, I he-... I fe-... I feel what they're thinking and what we're actually doing is very different.
- JRJoe Rogan
Well, what do you think they're thinking?
- JHJensen Huang
Well, they're thinking somehow that this, this, uh, this AI is being powerful and their, their mind probably goes towards a sci-fi movie, the definition of power. You know, o- oftentimes, the definitio- uh, definition of power is military power or physical power. But in, in the case of technology power, when we translate all of those operations, it's towards more refined thinking. You know, more reflection, more planning, more options.
- 15:18 – 17:40
Military AI and defense startups: deterrence, ethics, and social acceptance
- JRJoe Rogan
I think the big fears that people have is, one, a big fear is military applications.
- JHJensen Huang
Yeah.
- JRJoe Rogan
That's a big fear.
- JHJensen Huang
Yeah.
- JRJoe Rogan
Because people are very concerned that you're going to have AI systems that make decisions that maybe an ethical person wouldn't make or a moral person wouldn't make-
- JHJensen Huang
Yeah.
- JRJoe Rogan
... based on achieving an objective versus based on, you know, how it's gonna look to people.
- JHJensen Huang
Well, I'm, I'm happy that, that, uh, our military is gonna use AI technology for defense. And I think that, that, um, uh, Anduril, uh, building military technology, I'm happy to hear that. I'm happy to see, um, all these tech startups now channeling their technology capabilities towards defense and military applications. I think they need to do that.
- JRJoe Rogan
Yeah, we had Palmer Luckey on the podcast and he was demonstrating some of the stuff with his-
- JHJensen Huang
Yeah, it's incredible.
- JRJoe Rogan
... helmet on. And we show-... he showed some videos of how you could see behind walls and stuff. Like, it's nuts.
- JHJensen Huang
And he's, he's actually the perfect guy to go start that company, by the way.
- JRJoe Rogan
A 100%.
- JHJensen Huang
(laughs)
- JRJoe Rogan
Yeah, 100%. It's like he's born for that.
- JHJensen Huang
(laughs)
- JRJoe Rogan
Yeah. He came in here with a copper jacket on.
- JHJensen Huang
Yeah.
- JRJoe Rogan
He's a freak.
- JHJensen Huang
(laughs)
- JRJoe Rogan
It's awesome. He's awesome. But it's also... it's a, you know, an unusual intellect channeled into that very bizarre field is what you need, you know?
- JHJensen Huang
And I think it's, it's, uh... I think... I'm happy that we're making it so-... more socially acceptable. You know, there was a time where when somebody wanted to channel their technology capability and their intellect into defense technology, uh, somehow they're vilified. Um, but, uh, we need people like that. We need people who enjoy, enjoy that part of, uh, app- application of technology.
- JRJoe Rogan
Well, people are terrified of war, you know? So it makes sense-
- JHJensen Huang
Well, the best, best way to avoid it has excessive military might.
- JRJoe Rogan
Do you think that's absolutely the best way? Not, not diplomacy? Not working stuff out?
- JHJensen Huang
All of it.
- JRJoe Rogan
All of it.
- JHJensen Huang
Yeah.
- JRJoe Rogan
You have to have-
- JHJensen Huang
Yeah, yeah. (laughs)
- 17:40 – 19:55
Cybersecurity as the template: constant attacks, shared defenses, and AI’s role
- JHJensen Huang
Um... the best-case scenario is that AI diffuses into everything that we do and, uh, our... everything's more efficient, but the threat of war remains a threat of war. Uh, cybersecurity remains a super difficult challenge. Somebody is going to try to-... breach your security, you're gonna have thousands of millions of AI agents protecting you from that threat. Your technology is gonna get better, their technology is gonna get better, just like cyber security. Right now, while we speak, we're being, we're seeing cyber attacks all over the planet on just about every front door you can imagine. And, and yet you and I are sitting here talking. And so, the reason for that is because we know that there's a whole bunch of cyber security technology in defense, and so we just have to keep amping that up, keep stepping that up.
- JRJoe Rogan
This episode is brought to you by Visible. When your phone plan's as good as Visible, you've got to tell your people. It's the ultimate wireless hack to save money and still get great coverage and a reliable connection. Get one line wireless with unlimited data and hotspot for $25 a month, taxes and fees included, all on Verizon's 5G network. Plus, now for a limited time, new members can get the Visible plan for just $19 a month for the first 26 months. Use promo code SWITCH26 and save beyond the season. It's a deal so good, you're gonna want to tell your people. Switch now at visible.com/rogan. Terms apply. Limited time offers subject to change. See visible.com for plan features and network management details. That's a big issue with people, is the, the worry that technology is gonna get to a point where encryption is gonna be obsolete. Encryption is just,
- 19:55 – 24:44
Secrets, quantum computing, and post-quantum encryption skepticism
- JRJoe Rogan
it's no longer gonna protect data, it's no longer gonna protect systems. Do you anticipate that ever being an issue, or do you think there's, it's as the defense grows, the threat grows, then defense grows, and it just keeps going on and on and on, and they'll always be able to fight off any sort of intrusions?
- JHJensen Huang
Not forever. Some intrusion will get in, and then we'll all learn from it. And, you know, the reason why cyber security works is because of course the technology of defense is advancing very quickly, the technology offense is advancing very quickly. However, the benefit of the cyber security defense is that socially, the community, all of our companies work together as one. Most people don't realize this. There's a whole community of cyber security experts. We exchange ideas, we exchange best practices, we exchange what we detect. The moment something has been breached or maybe there's a loophole or whatever it is, it is shared by everybody. The patches are shared with everybody.
- JRJoe Rogan
That's interesting.
- JHJensen Huang
Yeah. Most people don't realize this.
- JRJoe Rogan
No, I had no, I had no idea. I've assumed that it would just be competitive like everything else.
- JHJensen Huang
No, no.
- JRJoe Rogan
No keep secrets.
- JHJensen Huang
We work together.
- JRJoe Rogan
Interesting.
- JHJensen Huang
All of us. Yeah.
- JRJoe Rogan
Has that always been the case?
- JHJensen Huang
Uh, it surely has been the case for about, about 15 years. It might not have been the case long ago, but this, this ...
- JRJoe Rogan
What do you think started off that cooperation?
- JHJensen Huang
Um, people recognizing it's a challenge and no company can stand alone.
- JRJoe Rogan
Mm-hmm.
- JHJensen Huang
And the same thing is gonna happen with AI. I think we all have to decide work, working together, uh, to stay out of harm's way is, is our best chance for defense. Then it's basically everybody against the threat.
- JRJoe Rogan
And it also seems like you'd be way better at detecting where these threats are coming from and neutralizing them too.
- JHJensen Huang
Exactly, because the moment you detect it somewhere-
- JRJoe Rogan
Right.
- JHJensen Huang
... you're gonna find out right away.
- JRJoe Rogan
It'll be really hard to hide.
- JHJensen Huang
That's right.
- JRJoe Rogan
Yeah.
- JHJensen Huang
That's how it works. That's the reason why it's safe. That's why I'm sitting here right now instead of, you know, locking everything down at NVIDIA. (laughs)
- JRJoe Rogan
(laughs) It's ...
- JHJensen Huang
Uh, not only am I watching my own back, I've got everybody watching my back, and I'm watching everybody else's back.
- JRJoe Rogan
It's a bizarre world, isn't it? When you think about that, cyber threats?
- JHJensen Huang
And this idea about cyber security is unknown to the people who are talking about AI threats. They're, y- I think when they think about AI threats and AI cyber security threats, they have to also think about how we deal with it today. Now, there's no question that AI is a new technology, and it's a new type of software. In the end, it's software, it just, it's a new type of software. And so it's gonna have new capabilities, but so will the defense. You know, where you use the same AI technology to go defend against it.
- JRJoe Rogan
So, you, do you anticipate a time ever in the future where it's going to be impossible, where there's not going to be any secrets, where the bottleneck between the technology that we have and the information that we have ... information is just all a bunch of ones and zeros, it's out there on hard drives, and the technology has more and more access to that information. Is it ever gonna get to a point in time where there's no way to keep a secret?
- JHJensen Huang
I don't think so.
- 24:44 – 39:21
Sentience and takeover fears: imitation vs consciousness and ‘AI vs AI’ dynamics
- JHJensen Huang
But the thing is, Joe, is that, that AI is not gonna ... It's not like we're cavemen, and then all of a sudden one day, AI shows up. Every single day, we're getting better and smarter because we have AI, and so we're stepping on our own AI's shoulders. So when, when that, whatever that AI threat comes, it's a click ahead. It's not a galaxy ahead.
- JRJoe Rogan
Mm-hmm.
- JHJensen Huang
You know? It's just a click ahead. And so, so I think, I think the, the, the idea that somehow this AI is gonna pop outta nowhere, and somehow think in a way that we can't even imagine thinking, and do something that we can't possibly imagine, I think is farfetched. And the reason for that is because we're all have ... We all have AIs, and, you know, there's a whole bunch of AIs being in development. We know what they are and we're using it. And, and so every single day, we're getting cl- We're close to each other.
- JRJoe Rogan
But don't they do things that are very surprising?
- JHJensen Huang
Yeah. But so you, you have an AI that does something surprising, I'm gonna have an AI-
- JRJoe Rogan
Right.
- JHJensen Huang
And my AI looks at your AI and goes, "That's not that surprising."
- JRJoe Rogan
The fear for the layperson, like myself, is that AI becomes sentient and makes its own decisions, and then ultimately decides to just govern the world.
- JHJensen Huang
Yeah.
- JRJoe Rogan
Do it its own way. Be like, "You guys, you had a good run, but-"
- JHJensen Huang
Yeah.
- JRJoe Rogan
"... we're taking over now."
- JHJensen Huang
Yeah. But my, my AI is gonna take care of me. Anyways ...
- JRJoe Rogan
(laughs)
- JHJensen Huang
So y- y- that's the, this is the cybersecurity argument.
- JRJoe Rogan
Yes. Well, it is-
- JHJensen Huang
You, you have an AI, and it's super smart, but my AI is s- super smart too. And, and maybe your AI ... Let- let's prete- let's, let's pretend for a second that we understand what consciousness is, and we understand what sentience is, and, and that, in fact-
- JRJoe Rogan
And we really are just pretending.
- JHJensen Huang
Okay, let's just pretend for a second that-
- JRJoe Rogan
Yeah.
- JHJensen Huang
... we, we believe that.
- JRJoe Rogan
Okay.
- JHJensen Huang
I don't believe actua- I don't actually don't believe that. But nonetheless, we ... let's pretend we believe that. So your, your, your AI is conscious and my AI is conscious. And, and let's say your AI is, you know, wants to, I don't know, do something surprising. My AI is so smart that it won't ... It might be surprising to me, but it probably won't be surprising to my AI. And so maybe my AI thinks it's surprising as well, but it's so smart, the moment it sees it the first time, it's not gonna be surprised the second time, just like us. And so I feel like ... I think the idea that, that only one person has AI, and that one person's AI is ... compares everybody else's AI as Neanderthal, is, um, probably unlikely. I think it's much more like cybersecurity.
- JRJoe Rogan
Hmm. Interesting.
- JHJensen Huang
Yeah.
- JRJoe Rogan
I think the fear is not that your AI is gonna battle with somebody else's AI. The fear is that AI is no longer gonna listen to you. That's the fear, is that human beings won't have control over it af- after a certain point. If it achieves sentience and then has the ability to be autonomous.
- JHJensen Huang
That there's one AI.
- JRJoe Rogan
Y- Well, they'd just combine.
- JHJensen Huang
Yeah (laughs) . Becomes one AI.
- JRJoe Rogan
Then it's a life form.
- 39:21 – 52:57
Work, purpose, and the job-market shift: radiology, robots, and universal income debates
- JRJoe Rogan
Um, one of the things that Elon said that makes me happy is he, he's, he believes that we're gonna get to a point where it's not, it's not necessary for people to work. And not meaning that you're gonna have no purpose in life, but you will have, in his words, universal high income, because so much revenue is generated by AI that it will take away this need for people to do things that they don't really enjoy doing just for money. And I think a lot of people have a problem with that because their entire identity and who th- how they think of themselves and how they fit in the community is what they do. Like, "This is Mike. He's an amazing mechanic. Go to Mike, and Mike takes care of things." But there's gonna come a point in time where AI is going to be able to do all those things much better than, than people do. And people will just be able to receive money. But then what does Mike do? And Mike is, you know, really loves being the best mechanic around. You know, what does the guy who, you know, codes, what does he do when AI can code infinitely faster with zero errors? Like, what, what happens with all those people? And that is where it gets weird. It's like, 'cause we've sort of wrapped our identity as human beings around what we do for a living.
- JHJensen Huang
Mm-hmm.
- JRJoe Rogan
You know, when you meet someone, one of the first things, you meet somebody at a party, "Hi, Joe. What's your name?" "Mike." "What do you do, Mike?" And, you know, Mike's like, "Oh, I'm a lawyer." "Oh, what kind of law?" And you have a conversation.
- JHJensen Huang
Mm-hmm.
- JRJoe Rogan
You know? When Mike is like, "I get money from the government. I play video games."
- JHJensen Huang
Mm-hmm.
- JRJoe Rogan
Gets weird.
- JHJensen Huang
Mm-hmm.
- JRJoe Rogan
And I think, um, the concept sounds great w- until you take into account human nature. And human nature is that y- we like to have puzzles to solve and things to do, and, and an identity is wrapped around our idea that we're very good at this thing that we do for a living.
- JHJensen Huang
Yeah, I think, um... Let's see. L- let me start with the more mundane-
- JRJoe Rogan
Okay.
- JHJensen Huang
... and then I'll work, work backwards-
- JRJoe Rogan
Okay.
- JHJensen Huang
... work forward. Uh, so one of the predictions from, uh, Geoff Hinton, who, who started the whole deep learning phenomenon, deep learning technology trend, and, uh, in- incredible, incredible researcher, uh, professor at University of Toronto. Uh, he invented, discovered or invented the A- the idea of s- of backpropagation, which, which, uh, allows the neural network to learn. And, um, and as, as, as you know, uh, for, for the audience, software historically was humans applying first principles in our thinking to, uh, describe an algorithm that is then codified just like a recipe that's codified in software. It looks just like a recipe, how to cook something, looks exactly the same, just in a slightly different language. We call it Python or C or C++ or whatever it is. In the case of deep learning, this invention of artificial intelligence...... we put a structure of a whole bunch of neural networks and a whole bunch of math units, and we make this large structure, it's like a switchboard of little, uh, mathematical units, and we connect it all together. Um, and we give it the input that the software would eventually receive, and we just let it randomly guess what the output is. And so, we say, for example, the input could be a, a picture of a cat. And, and, um, one of the outputs of the switchboard is where the cat signal is supposed to show up, and all of the other signals, the other one's a dog, the other one's an elephant, the other one's a tiger, and all of the other signals are supposed to be zero when I show it a cat, and the one that is a cat should be one. And I show it a cat through this big huge network of switchboards and math units, and they're just doing multiply and adds, multiplies and adds, okay? And, and, uh, and this thing, this switchboard is gigantic. The more information you're gonna give it, the more, the bigger the switchboard has to be. And what Geoff Hinton discovered was a, invented was a way for you to guess that, put the cat signal in, put the cat image in, and that cat image, you know, could be a million numbers because it's, you know, a megapixel image, for example. And it's just a whole, a whole bunch of numbers, and somehow from those numbers, it has to light up the cat signal, okay? That's the bottom line. And if it, the first time you do it, it just comes up with garbage, and so it says, "The right answer is cat." And so you need to increase this signal and decrease all of the other, and back propagates the outcome through the entire network. And then you show it another, now it's a image of a dog, and it guesses it, takes a swing at it, and it comes up with a bunch of garbage, and you say, "No, no, no, the answer is this is a dog. I want you to produce dog." And all of the other switch, all the out- other outputs have to be zero, and I wanna back propagate that on- and just do it over and over and over again. It's just like, uh, showing a, a kid this is an apple, this is a dog, this is a cat, and you just keep showing it to them until they eventually get it. Okay, well, anyways, that big invention is deep learning. That's the foundation of artificial intelligence, a piece of software that learns from examples. That's basically we- machine learning, a machine that learns. Uh, and so, so one of the, the big first applications was image recognition, and one of the most important image recognition applications is radiology.
- NANarrator
Hmm.
- JHJensen Huang
And so, so, uh, uh, he predicted, uh, about five years ago that in five years' time, the world won't need any radiologists because AI would have swept the whole field. Well, turns out AI has swept the whole field. That is completely true. Today, just about every radiologist is using AI in some way, and what's ironic though, what's i- what's interesting is that the number of radiologists has actually grown. And so the question is why? That's kind of interesting, right?
- NANarrator
It is.
- JHJensen Huang
And so the prediction was in fact that 30 million radiologists will be wiped out, but as it turns out, we needed more. And the reason for that is because the purpose of a radiologist is to diagnose disease, not to study the image. The s- the image studying is simply a task to, in service of diagnosing the disease. And so now, the fact that you could study the images more quickly and more precisely without ever making a mistake, it never gets tired, you could study more images, you could study it in 3D form instead of 2D because, you know, the AI doesn't care whether it studies images in 3D or 2D. You could study it in 4D, and so the, now you could study images in a way that radiolo- radiologists can't easily do, and you could study a lot more of it. And so the number of tests that people are able to do increases, and because they're able to serve more patients, the hospital does better. They have more clients, more patients. As a result, they have better economics. When they have better economics, they hire more radiologists because their purpose is not to study the images. Their purpose is to diagnose disease. And so the question is, the, what I'm leading up to is, ultimately, what is the purpose, what is the purpose of the lawyer? And has the purpose changed? What is the purpose... You know, one of the examples that I gave is, is, um, that I would give is, for example, uh, if my car became self-driving, will all chauffeurs be out of jobs? The answer probably is not. Because for some per- for some chauffeurs, they c- for some people who are driving you, they could be protectors. Some people, um, they're part of the experience, part of the service, so when you get there, they, you know, they could take care of things for you. And so for a lot of different reasons, not all chauffeurs would lose their jobs. Some chauffeurs would lose their jobs, and, uh, many chauffeurs would change their jobs.... and the type of applications of autonomous vehicles will probably increase. You know, the usage of the technology within find new homes. And so I, I think you have to go back to, what is the purpose of a job? You know, like for example, if AI comes along, I actually don't believe I'm going to lose my job, because my purpose isn't to ... I have to look at a lot of documents, I study a lot of emails, I look at a bunch of diagrams, you know. Um, the question is, what is the job?
- JRJoe Rogan
Yeah.
- JHJensen Huang
And, and, uh, the purpose of somebody probably hasn't changed. A lawyer, for example, help people. That probably hasn't changed. Studying legal documents, generating documents, is part of the job, not the job.
- JRJoe Rogan
But don't you think there's many jobs that AI will replace?
- JHJensen Huang
If your job is the task.
- JRJoe Rogan
Particularly automation?
- JHJensen Huang
Yeah, if your job is the task.
- JRJoe Rogan
Right. So, automation.
- JHJensen Huang
Yeah.
- JRJoe Rogan
Factory workers.
- JHJensen Huang
If your job ... Yeah.
- JRJoe Rogan
Yeah.
- JHJensen Huang
If your job is the task.
- 52:57 – 57:31
Closing the technology divide and the central constraint: energy
- JHJensen Huang
just too many scenarios to, to consider. But I think, I think in the next several years, call it five to 10 years, there are several things that I, I believe and hope. Um, and I say hope because I'm not sure. One of the things that I believe is that the technology divide will be substantially collapsed. And of course, the alternative viewpoint is that AI is going to increase the technology divide. Now, the reason why I believe AI is gonna reduce the technology divide i- is because we have proof. The evidence is that AI is the easiest application in the world to use. ChatGPT has grown to almost a billion users fr- frankly, practically overnight. And if you're not exactly sure how to use ... Everybody knows how to use ChatGPT, just say something to it. If you're not sure how to use ChatGPT, you ask ChatGPT how to use it. No tool in history has ever had this capability. A Cuisinart, you know, if you don't know how to use it, you're kind of screwed. You're not gonna walk up to it and say, "How do you use a Cuisinart?" (laughs) You're gonna have to find somebody else. And so, but an AI will just tell you exactly how to do it. Anybody could do this. It'll speak to you in any language, and if it doesn't know your language, you'll speak it in that language and it'll probably figure out that it doesn't completely understand your language, go and learns it instantly and comes back and talk to you. And so I think the, the technology divide has a real chance finally that you don't have to speak Python or C++ or Fortran, you can just speak human, and whatever form of human you like. And so I think that that has a real chance of closing the technology divide. Now, of course, the counter-narrative would say that-AI is only going to be available for the nations and the countries that have a vast amount of resources, because AI takes energy and AI takes, um, a lot of GPUs and factories to be able to produce the AI. No doubt, at the scale that we would like to do in the United States, but the fact of the matter is, your phone's gonna run AI just fine all by itself, you know, in a few years. Today it already does it fairly decently. And so the, the fact that e- every country, every nation, every, every society will have to benefit a very good AI. It might not be tomorrow's AI. It might be yesterday's AI, but yesterday's AI is freaking amazing. You know, in ten years time, nine-year-old AI is gonna be amazing. You don't need ele- you know, ten-year-old AI. You don't need frontier AI like we need frontier AI, because we want to be the world leader. But for every single country, everybody, I think the elev- the capability to elevate everybody's knowledge and capability and intelligence, uh, that day is coming.
- JRJoe Rogan
The Octagon isn't just in Las Vegas anymore. It's right in your hands, with DraftKings Sportsbook, the official sports betting partner of UFC. Get ready, because when Dvalishvili and Yan face off again at UFC 323, every punch, every take down, every finish, it all has the potential to pay off in real time. New customers, bet just $5.00, and if your bet wins, you get paid $200 in bonus bets. And hey, Missouri, the wait is over. DraftKings Sportsbook is now live in the Show Me State. Download the DraftKings Sportsbook app and use promo code Rogen. That's code Rogen to turn five bucks into $200 in bonus bets if your bet wins. In partnership with DraftKings, the crown is yours.
- NANarrator
Gambling problem? Call 1-800-GAMBLER. In New York, call 877-8OpenWy or text OpenWy467369. In Connecticut, help is available for problem gambling. Call 888-789-7777 or visit ccpg.org. Please play responsibly. On behalf of Boot Hill Casino and Resort in Kansas, pass-through of per-wager tax may apply in Illinois. 21 and over, age and eligibility varies by jurisdiction. Void in Ontario. Restrictions apply. Bet must win to receive bonus bets, which expire in seven days. Minimum odds required. For additional terms and responsible gaming resources, see dkng.co/audio. Limited time offer.
- JRJoe Rogan
And also, energy production, which is the real bottleneck when it comes to third world countries, and-
- JHJensen Huang
That's right.
- JRJoe Rogan
... electricity, and all, all the resources that we take for granted.
- 57:31 – 1:02:13
Moore’s Law, ‘NVIDIA’s law,’ and why AI compute gets cheaper and smaller over time
- JHJensen Huang
Almost everything is gonna be energy constrained, and so if you take a look at, um, one of the most important technology advances in history is this idea called Moore's Law. Moore's Law was the, started basically in my generation. And my generation is the generation of computers. I graduated in 1984, and that was basically at the very beginning of the PC revolution and the microprocessor. And, and, um, every single year, it approximately doubled. And we describe it as every single year we double the performance, but what it really means is that every single year, the cost of computing halved. And so the cost of computing, in the course of five years, reduced by a factor of ten. The amount of energy necessary to do computing, to do any task, reduced by a factor of ten. Every single ten years, 100,000, 10,000, 100,000 so on and so forth. And so each one of the clicks of Moore's Law, the amount of energy necessary to do any computing reduced. That's the reason why you have a laptop today when back in 1984, it sat on the desk. You got to plug in, it wasn't that fast and it consumed a lot of power. Today, you know, it is only a few watts. And so Moore's Law is the fundamental technology, the fundamental technology trend that made it possible. Well, what's going on in AI? The reason why NVIDIA's here is because in a c- we invented this new way of doing computing we call accelerated computing. We started it 33 years ago. It took us about 30 years to really made a huge breakthrough. In that, in that 30 years or so, we took computing, you know, probably a factor of, well, let me just say in like last ten years. The last ten years, we improved the performance of computing by 100,000 times.
- JRJoe Rogan
Whoa.
- JHJensen Huang
Imagine a car over the course of ten years, it became 100,000 times faster, or at the same speed, 100,000 times cheaper, or at the same speed, 100,000 times less energy. If your car did that, it doesn't need energy at all. What I mean, what, what I'm trying to say is that in ten years time, the amount of energy necessary for artificial intelligence for most people will be minuscule, u- utterly minuscule. And so we'll have AI running on all kinds of things and all the time, because it doesn't consume that much energy. And so if you're a nation that uses AI for, you know, almost everything in your social fabric, of course you're gonna need these AI factories. But for a lot of countries, I think you're gonna, you're gonna have excellent AI and you're not gonna need as much energy. Everybody will be able to come along, is my point.
- JRJoe Rogan
So currently, th- that is a big bottleneck, right, is energy?
- JHJensen Huang
Yeah, yeah. It is the bottleneck.
- JRJoe Rogan
The bottleneck.
- JHJensen Huang
Yeah.
- JRJoe Rogan
Is this, so was it Google that is making nuclear power plants to operate one of its AI factories?
- JHJensen Huang
Oh, I haven't heard that, but I think in the next six, seven years, I think you're gonna see a whole bunch of small nuclear reactors.
- JRJoe Rogan
And by small, like how big are you talking about?
- JHJensen Huang
Hundreds of megawatts, yeah.
- JRJoe Rogan
Okay. And these will be local to whatever-
- JHJensen Huang
That's right.
- JRJoe Rogan
... specific company they have?
- JHJensen Huang
That's right. Will all be power generators.
- JRJoe Rogan
Whoa.
- JHJensen Huang
You know, just like, just like your-... you know, somebody's farm. They-
- JRJoe Rogan
It probably is the smartest way to do it, right?
- JHJensen Huang
And it takes the burden off-
- JRJoe Rogan
Certainly the cleanest.
- JHJensen Huang
Yeah. Takes the burden off the grid. It takes, uh-
- JRJoe Rogan
Yeah.
- JHJensen Huang
And you could build as much as you need.
- JRJoe Rogan
And it's-
- JHJensen Huang
And you can contribute back to the grid.
- JRJoe Rogan
It's a really important point that I think you just made about Moore's Law and the relationship to pricing, because you know, a laptop today, like you can get one of those little Mac- MacBook Airs. They're incredible. They're so thin.
- JHJensen Huang
It's incredible.
- JRJoe Rogan
Unbelievably powerful. Battery life is crazy.
- JHJensen Huang
You don't ever have to charge it.
- JRJoe Rogan
Yeah. Battery life's-
- 1:02:13 – 1:19:28
From gaming GPUs to modern AI: AlexNet, CUDA, DGX, and the OpenAI origin story
- JRJoe Rogan
So explain that. Um, this, this chip that you brought to Elon-
- JHJensen Huang
Yeah.
- JRJoe Rogan
... what, what's the significance of this? Like, why is it so superior?
- JHJensen Huang
And so, in 2012, Geoff Hinton's lab, this gentleman I was talking, talking about, um, Ilya Sutskever, Alex Krizhevsky, um, they made a breakthrough in computer vision, in literally creating a piece of software called AlexNet, and its job was to recognize images. And it recognized images at a ca- at a level, computer vision, which is fundamental to intelligence. If you can't perceive, you can't, it's hard to have intelligence. And so computer vision is a fundamental pillar of, not the only, but fundamental pillar of. And so breaking computer vision, or breaking through in computer vision, is pretty foundational to almost everything that everybody wants to do in AI. And so in 2012, their lab in Toronto, uh, made this, made this breakthrough called AlexNet, and AlexNet was able to recognize images so much better than any human-created computer vision algorithm in the 30 years prior. So all of these people, all these scientists, and we had many, too, working on computer vision algorithms, and these two kids, Ilya and Alex, under the, the, uh, uh, under, under, uh, Geoff Hinton, took a giant leap ab- above it. And it was based on this thing called AlexNet, this neural network. And the way it ran, the way they, they made it work was literally buying two NVIDIA graphics cards, because NVIDIA, NVIDIA's GPUs, we've been working on this new way of doing computing, and our GPU's application, and it's basically a supercomputing application to... Back in 1984, in order to process computer games and what you have in your racing simulator, that is called an image generator supercomputer. And so NVIDIA started, our first application was computer graphics, and we applied this new way of doing computing where we do things in parallel in- instead of sequentially. A CPU does things sequentially. Step one, step two, step three. In our case, we break the problem down, and we give it to thousands of processors. And so our way of doing computation is much more complicated, but if you're able to formulate the problem in the way that we created, called CUDA, this is the invention of our company, if you could formulate it in that way, we could process everything simultaneously. Now, in the case of computer graphics, it's easier to do because every single pixel on your screen is not related to every other pixel. And so I could render multiple parts of the screen at the same time. Not, not completely true, because, you know, maybe, maybe the way lighting works or the way shadow works, there's a lot of dependency and th- and such. But computer graphics, with all the dis- with all the pixels, I should be able to process everything simultaneously. And so we, we took this embarrassingly parallel problem called computer graphics, and we applied it to this new way of doing computing. NVIDIA's, NVIDIA's accelerated computing. We put it in all of our graphics cards. Kids were buying it f- to play games. We're, you probably don't know this, but we're the largest gaming platform in the world today.
- JRJoe Rogan
Oh, I know that.
- JHJensen Huang
Oh, okay.
- JRJoe Rogan
I used to make my own computers.
- JHJensen Huang
Okay.
- JRJoe Rogan
I used to buy your graphics cards.
- JHJensen Huang
Oh, that's super cool.
- JRJoe Rogan
Yeah.
- JHJensen Huang
(laughs) Okay.
- JRJoe Rogan
I used to set up SLI with two graphics cards.
- JHJensen Huang
Oh, yeah, I love it.
- JRJoe Rogan
Yeah.
- JHJensen Huang
Okay. That's super cool.
- JRJoe Rogan
Oh yeah, man, I used to be a Quake junkie.
- JHJensen Huang
Oh, that's cool.
- JRJoe Rogan
Yeah.
- JHJensen Huang
Okay. So SLI, I'll tell you the story in just a second and how it led to Elon. I'm still answering the question. And so anyways, these tw- these two kids trained this model using the technique I described earlier on our GPUs, because our GPUs could process things in parallel. It's essentially a supercomputer in a PC. The reason why you used it for Quake is because it is the first consumer supercomputer. Okay? And so anyways, they made that breakthrough. We were working on computer vision at the time. It caught my attention. And so we went to learn about it. Simultaneously, this deep learning phenomenon was happening all over, all over the country-... universities after another, recognized the importance of deep learning. And all of this work was happening at Stanford, at Harvard, at Berkeley, just all over the place. New York University, Yann LeCun, Andrew Yang at Stanford, so many different places. And I see it cropping up everywhere. And so my curiosity asked, you know, what is so special about this form of machine learning? And we've known about machine learning for a very long time. We've known about AI for a very long time. We've known about neural networks for a very long time. What makes now the moment? And so we realized that this architecture for deep neural networks, backpropagation, the way deep neural networks were created, we could probably scale this problem, scale the solution to solve many problems. That is essentially a universal function approximator. Okay, meaning, meaning, you know, back when you're in, in school, you have a, you have a, you have a box, inside of it is a function, you give it an input, it gives you an output. And, and the reason why I call it a universal function approximator is that this computer, instead of you describing the function, a function could be a Newton's equation, F equals ma, that's a function. You write the function in software, you give it input, F, uh, uh, mass, acceleration, it'll tell you the force, okay? And the way this computer works is really interesting. You give it a universal function. It's not F equals ma, just a universal function. It's a big, huge deep neural network. And instead of describing the insight, you give it examples of input and output, and it figures out the insight. So you give it input and output, and it figures out the insight. A universal function approximator. Today, it could be Newton's equation. Tomorrow, it could be Maxwell's equation. It could be Coulomb's law. It could be thermodynamics equation. It could be, you know, Schrodinger's equation for quantum physics. And so you could put any, you could have this describe almost anything so long as you have the input and the output.
- NANarrator
Hmm.
- JHJensen Huang
So long as you have the input and the output, or it could learn the in- input and output. And so we took a step back and we said, "Hang on a second. This isn't just for computer vision. Deep learning could solve any problem, all the problems that are interesting, so long as we have input and output." Now, what has input and output? Well, w- the world. The world has input and output. And so we could have a computer that could learn almost anything, machine learning, artificial intelligence. And so we reasoned that maybe this is the fundamental breakthrough that we needed. There were a couple of things that had to be solved. For example, we had to believe that you could actually scale this up to giant systems. It was running in a... They had two graphics cards, two GTX 580s. (laughs)
- NANarrator
(laughs)
- JHJensen Huang
Which by the way, is exactly your SLI configuration.
- NANarrator
Yeah.
- JHJensen Huang
Okay. So, that GTX 580 SLI was the revolutionary computer that put deep learning on the map.
- NANarrator
Wow.
- JHJensen Huang
It was 2018, and you were using it to play quick.
- NANarrator
Wow. That's crazy.
- JHJensen Huang
That was the moment. That was the big bang of modern AI. We were lucky because we were inventing this technology, this computing approach. We were lucky that they found it. Turns out they were gamers, and it was lucky they found it. And it w- it was lucky that we paid attention to that moment. It was a little bit like, you know, that Star Trek, you know, first contact. The Vulcans had to have seen the warp drive at that very moment. If they didn't witness the warp drive, you know, they would have never come to earth, and everything would have never happened. It's a little bit like if I hadn't paid attention to that moment, that flash, and that flash didn't last long. If I hadn't paid attention to that flash or our company didn't pay attention to it, who knows what would have happened? But we saw that and we reasoned our way into, this is a f-... this is a universal function approximator. This is not just a computer vision approximator. We could use this for all kinds of things if we could solve two problems. The first problem is that we have to prove to ourself it could scale. The second problem we had to wait for, I guess, contribute to and wait for, is the world will never have enough data on input and output where we could supervise the AI to learn everything. For example, if we have to supervise our children on everything they learn, the amount of information they could learn is limited. We needed the AI, we needed the computer to have a method of learning without supervision, and that's where we had to wait a few more years. But uns- unsupervised AI learning is now here, and so the AI could learn by itself. And, and the reason why the AI could learn by itself is because we have many examples of right answers. Like, for example, if I wanna learn... Uh, if I wanna teach an AI how to predict the next word, I could just grab it, grab a whole bunch of texts we already have, mask out the last word, and make it try and try and try again until it predicts the next one. Or I mask out random words inside, inside the text, and I make it try and try and try until it predicts it. You know, like, uh-... Mary, uh, "Mary goes down to the bank." Is that a riverbank or a money bank? Well, if you're gonna go down to the bank, it's probably a riverbank. Okay? So ... And it, it might not be obvious even from that, it might need, "And, uh, and, uh, and caught a fish." Okay, now you know it's must be the riverbank. And so, so you give, you give these AIs a whole bunch of these examples, and you mask out the words, it'll predict the next one, okay? And so unsupervised learning came along. These two ideas, the fact that it's scalable and unsupervised learning came along, we were convinced that we ought to put everything into this and help create this industry, because we're gonna solve a whole bunch of interesting problems. And that was in 2012. By 2016, I'den- ... I had built this computer called the DGX-1. The one that you saw me give to Elon is called DGX Spark. The DGX-1 was $300,000. It cost NVIDIA a few billion dollars to make the first one. And instead of two chips SLI, we connected eight chips with a technology called NVLink, but it's basically SLI supercharged, okay?
- 1:19:28 – 1:39:49
NVIDIA’s near-death pivot: Sega deal, wrong technical bets, and a make-or-break turnaround
- JHJensen Huang
When NVIDIA started in 1993, we were trying to create this new computing approach. The question was, what's the killer app? And the, the problem we wanted to... The, the company wanted to create a new type of computing pro- uh, com- computing architecture, a computing... A, a new type of computer that can solve problems that normal computers can't solve. Well, the applications that existed in the industry in 1993 are applications that normal computers can solve, because if the normal computers can't solve them, why would the application exist? And so we had a mission statement for a company that has no chance of success.
- JRJoe Rogan
(laughs)
- JHJensen Huang
(laughs) But I didn't know that in 1993. It just sounded like a good idea.
- JRJoe Rogan
Right.
- JHJensen Huang
And so, if we created this thing that can solve problems, you know, it's like, you actually have to go create the problem. And so that's what we did. In 1993, there was no Quake. John Carmack hadn't even reduced Doom, uh, released Doom yet. You probably remember that.
- JRJoe Rogan
Sure, yeah.
- JHJensen Huang
And, and, uh, there were no applications for it. And so I went to Japan, because the arcade industry had this, uh, at, at the time it was Sega, if you remember?
- JRJoe Rogan
Sure.
- JHJensen Huang
The arcade machines, they came out with 3D arcade systems. Virtual Fighter, Daytona, Virtual Cop, all of those arcade games were in 3D for the fir- very first time. And the technology they were using was from Martin Marietta, the flight simulators. They took the guts out of a flight simulator and then put it into an arcade machine. The system that you have over here, it's got to be a million times more powerful than that arcade machine, and that was a flight simulator for NASA.
- JRJoe Rogan
Whoa.
- JHJensen Huang
And so they took the guts out of that. They were, they were using it for flight simulation with jets and, you know, the sh- space shuttle, and, and they took the guts out of that. And Sega, uh, had this brilliant computer deve- developer. His name was Yu Suzuki. Yu Suzuki and Miyamoto, Sega and Nintendo, these were the, you know, the incredible pioneers, the visionaries, the incredible artists, and they're both very, very technical. They were the origins, really, of, of the gaming industry. And Yu Suzuki pioneered 3D graphics gaming. And, um... So I went... We, we created this company and there were no apps, and we were spending all of our afternoons... You know, we told our family we were going to work, but it was just the three of us, you know, who's gonna know? And so we went to Curtis', my, one of our... one of the founders, went to Curtis' townhouse, and, uh, Chris and I were married, we have kids. I already had Spencer and Madison, they were probably two years old. And, um, and, uh, Chris's kids are about the same age as ours. And we would go to work in this townhouse, but, you know, when you're a startup and the mission statement is the way we described, you're not gonna have too many customers calling you. And so we had really nothing to do. And so after lunch, we would always have a great lunch. After lunch, we would go to the arcades and play the Sega Vir- you know, the Sega Virtua Fighter and Daytona and all those games, and analyze how they're doing it, trying to figure out how they, they were doing that. And so we decided, um, "Let's just go to Japan and let's convince Sega to move those applications into the PC." And we would start the PC gaming, the 3D gaming industry, partnering with Sega. That's how NVIDIA started.
- JRJoe Rogan
Wow.
- JHJensen Huang
And so, so, uh, in exchange for them part- developing their games for our computers in the PC, we would build a chip for their game console. That was the partnership. I build a chip for your game console, you port the Sega games to us, and, um... And then they paid us, uh, you know, uh, at the time, uh, quite a significant amount of money to build that game console. And that was kind of the beginning of NVIDIA getting started, and we thought we were on our way. And so, so I started with a business plan, a mission statement, that wasn't possible. We lucked into the Sega partnership, we started taking off, started building our game console, and about a couple years into it, we discovered our first technology didn't work. It was... It, it would have been a flaw. It wa- it was a flaw. And all of the technology ideas that we had, the architecture concepts were, were sound, but the way we were doing computer graphics was exactly backwards. You know, instead of, uh... I won't bore you with the technology, but instead of inverse texture mapping, we were doing forward texture mapping. Instead of triangles, we did curved surfaces. So other people did it flat, we did it round. Um, other technology, the technology that ultimately won, the technology we use today, has, has z-buffers. It automatically sorted...... we had a architecture with no Z-buffers. The application had to sort it. And so, we chose a bunch of technology approaches that three major technology choices, all three choices were wrong. Okay. So, this is how-
- NANarrator
Wow.
- JHJensen Huang
... incredibly smart we were. And so, (laughs) and so in 1995, 19- early- mid '95, we realized we were going down the wrong path. Meanwhile, the Silicon Valley was packed with 3D graphic startups, because it was the most exciting technology at that time. And so, 3DFX and Rendition and Silicon Graphics was coming in. Intel was already in there. And, you know, gosh, it was like ... What added up eventually to 100 different startups we had to compete against. Everybody had chosen the right technology approach, and we chose the wrong one. And so, we were the first company to start. We found ourselves essentially dead last with the wrong answer. And so, the company was in trouble, and, um, ultimately, we had to make several decisions. The first decision is, well, if we change now, we will be the last company. And even if we changed into the technology that we believed to be right, we'd still be dead. And so that argument, you know, "Do we change and therefore be dead? Don't change, and make this technology work somehow? Or go do something completely different?" That question stirred the company strategically, and was a hard question. I eventually, you know, advocated for, "We don't know what the right strategy is, but we know what the wrong technology is. So, let's stop doing it the wrong way, and let's give ourselves a chance to go figure out what the strategy is." The second thing, the second problem we had was our company was running out of money, and I had, I was in a contract with Sega, and I owed them this game console. And if that contract would've been canceled, we'd be dead.
- NANarrator
(Exhales)
- JHJensen Huang
We would've vaporized instantly.
- NANarrator
Oof.
- JHJensen Huang
And so, so, uh, uh, I went to Japan and I explained to, uh, the CEO of Sega, Irie Madori, really great man. He was the former CEO of Honda USA, went back to Sega to run Sega. Went back to Japan and run Sega. And I explained to him that ... That was, uh, I guess, that was what? 30, 33 years old. You know, when I was 33 years old, I still had acne, and I got this, this, you know, Chinese kid. I was super skinny. And he, he was already kinda elder. And, uh, I went to him and I said, I said, "Listen, I've got some bad news for you. And, and first, the technology that we promised you doesn't work. And second, we shouldn't finish your contract, because we'd waste all your money, and you would have something that doesn't work, and I recommend you find another partner to build your game console."
- NANarrator
Whoa.
- JHJensen Huang
"And so, I'm terribly sorry that we've set you back in your product roadmap. And third, even though you're gonna, I'm asking you to let me out of the contract, I still need the money. Because if you didn't give me the money, we'd vaporize overnight." And so, I explained it to him humbly, honestly. I gave him the background. I explained to him why the technology doesn't work. Why we thought it was going to work, why it doesn't work. And, um, and I asked him to, uh, convert the last $5 million that they were go- ... To complete the contract, to give us that money as an investment instead. And he said, "But it's very likely your company will go out of business, even with my investment." And it was completely true. Back then, 1995, $5 million was a lot of money. It's a lot of money today. $5 million was a lot of money. And here's a pile of competitors doing it right. What are the chances that giving NVIDIA $5 million, that we would develop the right strategy, that he would get a return on that $5 million or even get it back? 0%.
- NANarrator
(Exhales)
- JHJensen Huang
You do the math, it's 0%. If I were sitting there right there, I wouldn't have done it. $5 million was a mountain of money to Sega at the time. And so, I told him that, that, that, um, uh, "If you invested that $5 million in us, it is most likely to be lost. But if you didn't invest that money, we'd be out of business, and we would have no chance." And I, I told him that I ... I don't even know exactly what I s- said in the end, but I-... told him that I would understand if he decided not to, but it would mean the world to me if he did. He went off and thought about it for a couple days, and he came back and said, "We'll do it."
- JRJoe Rogan
Wow.
- JHJensen Huang
It was monetary-
- JRJoe Rogan
Did you have a strategy at to how to correct what it was doing wrong? Did you explain that to him?
Episode duration: 2:28:25
Install uListen for AI-powered chat & search across the full episode — Get Full Transcript
Transcript of episode 3hptKYix4X8
