NVIDIA’s Jensen Huang on Reasoning Models, Robotics, and Refuting the “AI Bubble” Narrative

29:30And so I think in 2024, the cost of GPT-4 equivalent models, if you look at a million tokens, it came down over 100x. Somebody in my team did this analysis to show that. So the cost of dropping pretty dramatically and very rapidly, and part of it is all the advancements you all have been driving on the video level, but also just across the stack, people have been getting big efficiency gains. At the same time, model companies are talking about how the costs are rising, how there's enormous sort of capital moats to building these things out. How do you think about cost of training and cost of inference over time and what that means for the average end user or the average startup company trying to compete or people trying to do more in this industry?

30:08I forget the statistic, but you know, Andre Carpathi estimated the cost of building the first chat UBT, I think, versus now. I think you could do that on the PC now. Yeah, yeah. It's probably tens of thousands of dollars at this point or maybe even less. And so it costs nothing. He has an open source project that you can do in a weekend. Oh, is that right? Okay. That's incredible, right? We're talking about three years. What people said cost billions of dollars, supercomputers built, raising billions of dollars in order to do all that. now cost you know something that you can do on a weekend on a pc and so that tells you something about how quickly we're making making ai more cost effective spark sorry probably not quite a pc okay not quite a pc yeah we're improving our architecture and performance um every single year the first gbt i think was trained on voltus and then uh ampere um you know and and it wasn't I think the first breakthroughs, none of it included Hopper.

31:17And of course, Hopper last couple, two, three years, and we're off in Blackwell for the last year and a half or so. And every single one of these generations, the architecture improves. And of course, the number of transistors go up and the capacity goes up. Every single generation, very easily, every single year from a computing perspective, the combination of all that, getting 5 to 10x every single year, It's not unusual. And here comes Ruben just around the corner. And so we're seeing 5 to 10x every single year. Well, compounded, it's incredible. Moore's Law was two times every year and a half.

31:56And over the course of five years, it's 10x. Over the course of 10 years, it's 100x. In the case of AI, over the course of 10 years, it's probably 100 ,000 to a millionx. Okay? And that's just the hardware. Then the next layer is the algorithm layer and the model layer. The combination of all that, the fact that if you were to tell me that in the span of 10 years, we're going to reduce the cost of token generation by a billion times, I would not be surprised. And so that's the tokenomics of AI. On the training side, it's not quite as aggressive in cost reduction, but it's close. If you were to say that every single year we're increasing by two or three X over the course of 10 years, incredible.

32:42But the important idea is when somebody says it costs$100 million to train something or half a billion dollars to train something, well, next year it's 10 times less. Next year is 10 times less. But people just scale these things up, though, right? So the counterargument is, well, we'll just get bigger every year by 10x or 100x or, you know, we'll try and offset that decrease in cost by scale. And others can't keep up. Yeah, but really what's happening is, and this is where MOEs come in, as you know, the scale went up by a factor of 10, but the computational burden did not go up by a factor of 10.

33:17Because you're getting the compounded benefits of all three things. The hardware is going up, the algorithms of training models are going up, and of course the model architecture is going up. And we're getting the benefit of learning from each other. You know, let's face it. DeepSeq was probably the single most important paper that most Silicon Valley researchers read from in the last couple of years. It was the only thing that felt frontier that was open. That's right. In years. That's right. And that's because a lot of the labs are not. Back to the value of open source again. Yeah. Putting out these papers.

33:49Literally, DeepSeq benefited American startups and American AI labs all over. And infrastructure companies. And infrastructure company all over. probably the single greatest contribution to American AI last year. And so if you said this out loud, of course, you know, people kind of shudder that we're, American AI is actually getting learning from and benefiting from AI from other nations. But why would that be surprising? You know, AI researchers in all over America are Chinese natives and come from different countries. We benefit from every country. You benefit from every researcher. And all of the world's ideas don't have to come from the United States.

34:33And so I think back to your original question, it is the case that some of the narratives around the cost of AI is about scaring everybody out of the market. Nobody ought to do pre-training but us. Nobody should do training these frontier models but us. But because of innovation of models, algorithms, and the computing stack, the cost of AI is actually decreasing well more than 10x every single year. And so if you're just one year behind or even six months behind, you could really stay close. And I think one thing that felt very different to me about 2025 is Ilya said recently that, you know, we're in the age of research again versus an age of scaling.

35:20I think both things are happening, by the way. Everybody is also trying to scale on multiple dimensions. Yeah, exactly. Both are happening. You know, being six months behind or being at 100 versus a 200K cluster, I think matters if you are competing symmetrically. But now you have people from Frontier Labs or at the very top of the game who have very different ideas about how to progress from here or who are working on diversity of problems. That's right. And I think that felt different from 24, maybe where there was a lot of energy focused on just pre-training scale and LLMs. Yeah, and several other dynamics.

35:54As the market grows, each one of these models could choose to have verticals or segments where they want to differentiate. Somebody could decide to be a better coder. Somebody could decide to be just better at being easier to be accessible so that it could be a greater consumer product. You know, the diversity of these models. As a result, you could probably make a niche leap without having to be great at everything else and still be super valuable to the market. It's no longer necessary to boil the entire ocean. two years ago because it was called pre-training. You know, people said, well, you know, pre-training is over.

36:37First of all, pre-training is not over. But the point of pre-training is to train yourself for training. That's why it's called pre-training. To prepare yourself to do the real training. And now we call it post-training. It's kind of weird. I think it's just training. But pre-training is pre-training and therefore it's training. Training, as we all know, is where compute scaling directly translates to intelligence. Now the data necessary to train the model is actually pretty small. Maybe it's just the verifiable results. Now it's really algorithmic, very compute intensive. And you don't have to be good at everything in life, as you know.

37:21Just like all of us, we could decide because we don't have time to learn everything equally well, we decide to choose a specialty and focus all of our energy on it. And we become superhuman or incredibly good at something that other people are not. And so I think AI labors are going to start doing the same. They're going to start bifurcating into various segments. And over time, you're going to, and startups will do the same. They'll find a micro niche and they'll take something open and then be incredibly good at it. Well, I think one of the most optimistic views here is actually that these micro niches are quite valuable, right?

37:55I was talking to Andre because I've been talking to a lot of people about the predictions for next year. We'll ask you yours as well, of course. But he asked, what's an example of a prediction that would have been prescient last year? And my answer, everything's easy in retrospect, is that coding would be the first application-level business that gets to a billion of ARR as an AI-native app. And I think if you'd taken an old-world view of this, you would have believed one of two narratives. Right. One is a single model does everything and I'll just be summed into something monolithic. And two is that developer tools never get very big.

38:33Right. Well, it kind of depends on how valuable the developer tool is. Now, I think many more people understand software engineering is in a niche and there's more demand than ever for it. But I think we'll see more like that. Also interesting. We are using we use cursor here and we use cursor pervasively here. every engineer uses it and a number of engineers you just mentioned it the number of people we're hiring today is just incredible yeah right monday has come to work on nvidia day and and uh why is that uh this is now the purpose and the task the purpose of a software engineer is to solve known problems and to find new problems to solve coding is one of the tasks and so if the purpose is not coding.

39:20If your purpose literally is coding, somebody tells you what to do, you code it. All right. Maybe you're going to get replaced by the AI, but most of our software engineers, all of our software engineers, their goal is to solve problems. And it turns out we have so many problems in the company and we have so many undiscovered problems. And so the more time they have to go explore undiscovered problems, the better off we are as a company. Nothing would give me more joy than if none of them are coding at all. They're just solving problems. You see what I'm saying? And so I think that this framework of purpose versus task is really good for everybody to apply.

39:53For example, somebody who's a waiter, their job is to not to take the order. That's not their job, as it turns out. Their job is so that we have a great experience. And if somebody, if some AI is taking the order, their job, or even delivering the food, their job is still helping us have a great experience they would reshape their jobs accordingly and so so i think the um the question about about cost of compute um uh is really important let's let me come back to one the the reason why we are so dedicated to a programmable architecture versus a fixed architect remember a long time ago uh a cnn ship came along and they said nvidia's done and then and then the transformership came and video was done people are still trying that yeah and people and and the benefit of these dedicated asics of course it could perform a job really really well and transformers is a much more universal ai network but the transformers you know the species of it is growing incredibly the attention mechanism the attention mechanism how it thinks about context, diffusion versus autoregressive.

41:07These hybrid SSM transformer things. These hybrid SSMs, for example, Nemotron, we just announced a new hybrid SSM. And so the architecture of transformers is in fact changing very rapidly. And over the next several years, it's likely to change tremendously. And so we dedicate ourselves to an architecture that's flexible for this reason, so that we can, on the one hand, adapt with, Remember, because Moore's law is largely over, transistor benefit is only 10%, maybe a couple of years. And yet we would like to have hundreds of X every year. And so the benefit is actually all in algorithms. And an architecture that enables any algorithm is likely going to be the best one, right?

41:50Because the transistor, it didn't advance that much. And so I think our dedication to programmability is number one for that reason. We have so much optimism for innovation and algorithms and innovation software that we protect our programmability for that reason. The second thing is, by protecting this architecture, our install base is really large. When a software engineer wants to optimize their algorithm, they want to make sure that it doesn't run on just this one little cloud or this one little stack. They want it to run on as many computers as possible. So the fact that we protect our architecture compatibility, then flash retention runs everywhere.

42:30So SSMs run everywhere. Diffusion runs everywhere. Autoregression runs everywhere. Just depending, it doesn't matter what you want to do. CNN still runs everywhere. LSTM still runs everywhere. And so this architecture that is architecturally compatible so that we have a large install base, programmable for the future, is really important in the way that we help to advance. And as a result, all of this drives the cost down. And I'm super proud that our latest innovation, NBLink72, we're the lowest cost token generation machine in the world by enormous amounts. And the reason for that is because MOEs are really, really hard.

43:11And so, you know, people didn't expect that. That for MOEs, it's probably easier to train, but for inference, it's incredibly hard to generate tokens on. as cost drop usually you open up new applications or new verticals that become more and more accessible and we talked a little bit about coding like cursor and cognition and other companies that are really benefiting from that in this last year do you have any thoughts or predictions in terms of what the next breakthrough industries will be or new applications or areas that you're most excited about coming in 26 in particular like other one or two things that you think well because of three things i because of because of a couple two three things i i think i think several industries are going to experience their chat GPT moment.

43:52I believe that multi-modality and very long context is going to enable, of course, really, really cool chatbots. But the basic architecture, that in combination with breakthroughs in synthetic data generation is going to help create the chat GPT moment for digital biology. That moment is coming. And by digital biology, do you specifically mean other aspects of like protein folding and protein binding or do you mean diagnosis? I see. I think we're good at protein understanding. Now, multi-protein understanding is coming online. And we recently created a model called LotProtina. It's opened. it's for multi-protein understanding and and represent representation learning and generation so so i think that the protein understanding is is advancing very quickly now protein generation is going to advance very quickly chat gpt moment proteins yeah there are a lot of interesting companies working on molecule design and end-to-end way like chai exactly and then and then of course chemical understanding and chemical generation and then protein, chemical confirmation, understanding and generation.

45:07Is that right? And so that combination, the chat GPT moment, the generative AI moment, all of that stuff is coming together for digital biology. And to your point about like new industries or, you know, the way I think about it is like investing in the inputs for this AI as well. All of these things around biology and chemistry and material science, they require real world data generation and experimentation. And that's new infrastructure too. New infrastructure. Synthetic data is going to be really important because they just have such sparse, right? Sparsity of data. And they just don't have as much as human language.

45:38And there, the real breakthrough is going to be when we can train a world foundation model, a foundation model for proteins, a foundation model for cells. I'm very excited about both of those things. Once we have a foundation model, our understanding capability, our generative capability, that data flywheel is really going to take off. The second area that I'm excited about, of course, reasoning made huge breakthroughs in language. But because of reasoning, cars are going to be able to perform better. So instead of just perception cars and planning cars, they're going to be reasoning cars. So these cars are going to be thinking all the time.

46:16And when they come up to a circumstance they've never encountered before, they can break it down into circumstances they have encountered before and construct a reasoning system for how to navigate through it. And so the out-of-domain, out-of-distribution part of AI is going to very much be addressed by reasoning systems. And as a result, we could do more things than we were taught to do. Between generative AI and multimodal vision, language, action models, and reasoning systems, I think we're going to see big breakthroughs in human robots or multi-embodiment robots. What do you think is a timeframe for that?

46:59Because if you look at the self-driving analog, and obviously self-driving technologies were based on very different types of neural networks than what we're using today. in terms of, you know, there's been a big swap over the last two, three years in terms of how we do a lot there. We started too soon. Self-driving cars really had four eras. The first era was smart sensors connected into a car. The Mobileye era. The Mobileye era. And even the very earliest days of Weibo. ADAS, yeah. Yeah, even the earliest days of Weibo. You're using smart sensors, a lot of human engineered algorithms and overall education.

47:38Beer mapping. Yeah. Extreme mapping. Mapping and then different systems for planning and perception. Exactly. And so you're essentially creating a car that is driving on digital rails, right? It's no different than the rails at Disneyland, except there are digital rails. And so that's the first generation. The second generation, and during that generation, you have perception, world model, and planning. And these modules. and each one of these modules have the limits of their technology and and perception was first input was was first affected by deep learning uh first and then and then uh and then it propagated through the pipeline and so that but that system is too brittle and it only knows how to perform what you taught it and now where we are are end-to-end models and then and then where we're going to go next are end-to-end models with reasoning yeah there you go so that those are kind of the four eras.

48:31In a lot of ways, if we were to start self-driving cars probably three years ago, we'd probably be exactly the same place. All our poor friends who were working in self-driving. Yeah. And I don't mind it. I've been working on it for 10 years. NVIDIA's self-driving car stack, by the way, number one rated safety in the world today. Number one. We just got that rating today, last week. And number two is Tesla. So I'm very proud that two American companies are up on the top. So from a robotics perspective, you think, because we've already built all these sorts of technologies in the modern era, robotics won't have the same 10, 15 years log.

49:09That's right. We're more optimistic. We'll just jump straight to it. I'm much more optimistic with robotics because we've kind of advanced foundational technology. We've been through so much to get there. Now, people are thinking about human robotics. Human robotics has a lot of challenges. I mean, there's all the megatronics challenges. Like, for example, it's not helpful if the robot weighs 300 pounds. And what happens if it falls over and is interacting with kids and so on and so forth. And so you've got all kinds of challenges to deal with. I'm certain that we're going to solve those. But remember, the fundamental technology that goes into a human robot can go into a pick-and-place robot.

49:45It could be... How do you think about... One thing I've been curious about for robotics in particular is if I look at who won or who's perceived as winning and self-driving, it's largely incumbents, right? It's Waymo. It's Tesla. You mentioned the safety rating NVIDIA has gotten. And so it's people who have been working on this for a long time. It took a lot of capital. It was really intensive to get there. You have supply chain, you have hardware, you have all this extra complexity. Do you think the same thing will be true in robotics? So the winner is basically going to be Tesla with Optimus and other people who have both been in the industry for a while, but also have all those sort of incumbent effects.

50:17Do you think there's room for startups? They will be one of the leaders, one of them, and surely a major one. but everything that moves will be robotic everything that moves will be robotic and everything that moves is a very large space it's not all human or robot and yet every ai will be multi-embodiment meaning you know just like just like a human with our our multi-embodiment ai ourselves we could sit in a car and embody that we could pick up a tennis racket embody that we could pick up a chopstick embody that and so we could embody the people are general purpose right that's right all these things exactly and so ai's are going to become general purpose so you have one arm pick in place maybe it's uh two arms pick in place could be six arms pick in place you know so so i think you're going to have all kinds of different sizes and shapes it could be a caterpillar it could be you know it could be an excavator it could be all kinds of stuff and so ai will embody those just as a just as a construction worker embodies an excavator embodies a tractor.

51:23You know, they, you know, could there be a small number of companies then that do the embodiment for everything? Or are you saying more, there's going to be niche applications? I could definitely see a lot of software companies. And then those, that software company could serve a lot of, a lot of different verticals, but each one of the verticals will still have solution providers that then grounds it all, turns it into something that works perfectly. Does it make sense? Yeah. Because in the case of AI for consumers, if it works 90 % of the time, you're delighted. it you're you know you're mind blown if it works 80 percent of time you're satisfied in the case of most industrial and physical ais if it works 90 percent of the time nobody cares about that they only care about the 10 that it fails and 100 basically you know 100 dissatisfaction and so you've got to take it to 99.9999 so the core technology might be able to get get you to 99 and then that's a vertical solution provider like a caterpillar or somebody they could take that core technology and make it 99.9999 % great.

52:21Do you think that's what happens like earliest on? Because in markets that are this immature, it seems one of the fastest paths to market could be full verticalization, right? Because you just have control of iteration speed. The difficulty of verticalization for technology that is general purpose is that you don't have the R &D scale to build a general purpose technology. Now, of course, open source helps that tremendously, which is the reason why you're going to see a big surge of vertical opportunities in AI in the next several years. My prediction would be over the course of the next five years, the excitement is going to be verticalization.

53:02Notice we're excited about OpenEvidence. We're excited about Harvey. We're excited about Cursor. Cursor is a horizontal, but it's kind of a horizontal vertical. you know and so i'm i'm super excited about all the verticals you know a lot of people said yeah ai is going to get so god ai is going to get so good that all these rapper companies are going to be obsolete it's just it misses the big point you know the reason why you could talk about the reason why somebody can talk talk about somebody is creating technology could talk about the life of a surgeon is because they've never been a surgeon the reason why somebody who builds an ai and talks about the life of an accountant and a tax, you know, a tax expert because they've never been a tax expert, you know?

53:44And so I think the reason why somebody could talk about being a busboy without being a busboy is because they've never been a busboy. And so I think you've got to be a little bit more empathetic about the depth of the complexity of the work and try to truly understand the purpose of the work. Oftentimes the technology addresses the task. It doesn't address the purpose. so i guess one of the other narratives from we're looking at narratives that are true versus not true you know for 25 one other narrative that's come up has been more about energy and energy utilization and will we have enough energy to support ai how do you how do you think about that on the first week of president trump's administration he said drill baby drill he there's so much flack for that if not for this entire change in in sentiment about energy growth in our country, we can all concede now we would have handed this industrial revolution to somebody else.

54:41And we're still power constrained. We're still power constrained. Without energy, there can be no new industry. And of course, we've been energy starved now for, what, a decade? If not for the fact that President Trump reversed that narrative, we would be completely screwed. Without energy, you can't have industrial growth. Without industrial growth, the nation can't be more prosperous. Without being more prosperous, we can't take care of domestic issues. We can't take care of social issues. You know, on and on and on. And so the fact of the matter is we need energy to grow. We need every form of energy.

55:17We need, you know, natural gas. We need to be, of course, we need more energy on the grid. We need more energy behind the meter. We're going to need nuclear. Wind is not going to be enough. Solar is not going to be enough. Let's just all acknowledge that we'll take it. We'll take everything we can. But the fact that matters, I think, for the next decade, natural gas, you know, is probably the only way to go forward. What's really interesting is I agree the timeline is too far out to address people's power generation issues in 27 and 28, where large players building clusters are very concerned.

55:53But the biggest drivers of climate innovation in the U.S. have actually been as a result of this AI infrastructure problem, right? Because people look at the demand. Finally, that's right. They look at the demand. And the demand is driving people to create massive new battery companies, solar concentrators, put new energy behind new energy, like, you know, willpower behind SMRs. Building the AI industry is driving all of that sustainable energy industry. Because people see that there is going to be demand. That's right. Right. So even if and I think there is no practical answer in the small number of years time frame versus large gas.

56:34Right. it still drives climate innovation. Yeah, no question about it. I think that's exactly right, that, you know, Doomer messages causes policy, and that policy may affect the industry in some way, but there's nothing more powerful than demand. Look at all the jobs that's being created. Look at all the industries that's being formed around it. Sustainable energy, likely. And when history rewrites it, Sarah, I think you're going to be absolutely right, that if not for AI, well, AI is probably the biggest driver for sustainable energy ever. Yeah, a friend of mine has a saying that doomers are the people who sound smart at dinner parties and optimists are the people who drive humanity forward.

57:17And I think that's very true for all these things we've talked about. Yeah, it's really true. Yeah. Well, that's one of the big, big takeaways for this last year, the battle of narratives. And it's too simplistic. To say that everything that the doomers are saying are irrelevant, that's not true. A lot of very sensible things are being said. It is too simplistic to say that when somebody is optimistic, that they're just naive. It needs to be grounded in reality. Yeah, that optimistic people are just naive. And that's obviously not true. But I think we just have to be mindful of the balance of it.

57:58when 90 % of the messaging is all around the end of the world and the pessimism. And I think we're scaring people from making the investments in AI that makes it safer, more functional, more productive, and more useful to society. And so we're just more secure. All of that takes technology. Security takes technology. Safety takes technology. I appreciate that my car is safer today because it has better technology than a car 50 years ago. And so I think it takes technology to be safe, technology to be secure. And so I'm delighted to see that the advancement of technology is still accelerating and ongoing.

58:40And so we just have to make sure that the policymakers around the world, the governments, are able to, are thinking about balancing these two ideas. How do you, so I guess we've talked a lot about 25 and the narratives of 25. How do you think about 26? What are you excited about? What do you see coming? What do you think are big changes that we should be aware of? I am optimistic that, that our relationship with China will improve. That President Trump and the administration has a really, really grounded and common sense attitude about and philosophy around, around how to think about China. that they're an adversary but they're also a partner in many ways and that the idea of decoupling is naive and the idea of decoupling for whatever reason, philosophical reasons or national security reasons it's just not based on any common sense and the more deeply you look into it, the more the two countries are actually highly coupled both countries ought to invest in their own independence um i you know when you depend too much on someone the relationship becomes too emotional as you know and so it's good to have some independence and or as much independence as either either would like um but to recognize that there's a lot of coupling a lot of dependence between the two countries and and i think there's a there needs to be a nuanced strategy a nuanced attitude about how to how to how to manage this relationship in a productive way for all of the people of two countries and for all of the people around the world.

1:00:21Everybody depends on a productive, constructive relationship of the two most important nations and the single most important relationship for the next century. And so we have to find that answer. And I'm just really delighted that President Trump is looking for a constructive answer. And so I think that next year will be a much better year than the last several I'm happy with the administration was able to suggest an export control policy that is grounded on national security, recognizing that they already make so many chips themselves and they can depend on Huawei themselves for their military, for their national security.

1:01:05They got ample technology to do that. And so that American technology, although general purpose, is unlikely to be used by their military because their military is too smart, just as our military is too smart to use their technology. And so it's grounded on national security. It's grounded on technology leadership. It's grounded on national prosperity. You know, one of the things that we just always have to remember is that the world's mightiest military is supported by the world's mightiest economy. And so the wealth that we generate brings jobs home, creates prosperity in the United States, provides for tax revenues, and ultimately funds the mightiest military on the planet.

1:01:48And so that circular system, that interconnected system requires a nuanced strategy. And I'm pleased to see some of the progress in that area that allows American technology companies to keep America first and keep America ahead and to support American technology leadership on the one hand to win globally. And then China, of course, is sorting itself out. I mean, not just working, but they're sorting out the attitude about how to think about American technology. Because the historical argument there has been that if you look, for example, at the internet, there was what was known as a great firewall, right?

1:02:30China basically prevented US competition into China while the opposite wasn't as true. There's been mass expatriation of US jobs and industry to China as sort of part of the development of the 90s and 2000s. And so I think a lot of the things that people have brought up from a China-US policy perspective besides just the military-adversarial relationship or spheres of influence or all the various things like that is also just the economic imbalances that have been perceived to exist between the two countries. The way that I would think through that is go back to the first principles of technologies again.

1:03:02And let's say the internet, you have the chip industry, you have the systems industry, the software industry, you have the services industry on top. Remember, China's internet growth has been a boom for Intel and AMD selling CPUs, Micron selling DRAMs. SK Hynix and Samsung selling DRAMs. It is the second largest internet market for American technology industry. And so maybe it wasn't helpful to some layer of the stack. The Googles of the world. That's right. But don't exclude every layer of the stack. Always come back. Every single one of these things, take a step back and look at the whole stack.

1:03:42Maybe that's a theme for today as well. And it makes sense that you would send this message. But, you know, technology is actually not just the sort of internet software application layer that's been very dominant for decades. That's right. It's the whole stack. And remember, as Intel and AMD prospered with the internet industry in China, the China industry growth, don't forget, China also contributed tremendously to open source. No country in the world contributes more to open source than China. and look at all the startups here in America that were able to benefit from out of that open source to create the new startups that are here.

1:04:18And so you can't look at one area in isolation. You have to look at the whole life cycle of the technology and look at every layer of the stack. Does it make sense? When you take a look at that from that lens, China's internet industry generated enormous prosperity for America. just not at the internet company per se jensen my other investor friends will not forgive me if i don't ask you about 2026 um uh on the business side uh are we in an ai bubble ai bubble yeah there's a lot of ways to reason through that and so so again um you know when when asked that question, my mind goes to what is AI and where are we in that?

1:05:06There's AI, then there's computing. You know, as you know, NVIDIA invented accelerated computing. Accelerated computing does computer graphics and rendering. AI doesn't. Accelerated computing does data processing, SQL data processing. AI doesn't. Accelerated computing does molecular dynamics and quantum chemistry. AI doesn't. These are all things that people could say someday AI will, but it doesn't today. Accelerated computing is really essential for classical machine learning, XGBoost, recommender systems, the whole process of feature engineering, extract, load, and transform. That entire data science machine learning's lifecycle.

1:05:47Accelerated computing is used for all of that. The first thing to go to is in the context of NVIDIA. What we see is the dynamic is a shift from general purpose computing to accelerated computing because Moore's law has largely ended. You can't use CPUs for everything anymore like you used to. And so it's just no longer productive enough. It's not deflationary enough. And so we have to move towards a new computing model. And that's where accelerated comes in. If you, if generative AI, well, excuse me, if chatbots, let's just go, you know, open AI and Anthropic and Gemini, if none of that existed today, NVIDIA would be a multi hundred billion dollar company.

1:06:29And the reason for that is because, as you know, the foundation of computing is shifting to accelerated computing. That's the first thing to realize is to take a step back and ask yourself what is actually happening. Now the next layer of the question about AI now becomes, what is AI? Now we ask that, we ask the AI bubble question and we always go back to open AI's revenues. A hundred percent, don't we? You ask somebody, Hey, is there an AI bubble? Everybody goes directly to open AI's revenues. First of all, if open AI currently has twice the capacity, their revenues would double. You guys know that.

1:07:05If they have 10 times the capacity there, I really believe the revenues were 10 times. And so they need capacity. This is no different than NVIDIA needs wafers from TSMC. Just because NVIDIA exists and we're doing great doesn't mean we don't need capacity. We need capacity. We need capacity of DRAM. And so in our world, it's sensible to everybody. We need capacity. Well, in their world, they need factories. And if they don't have factory capacity, how do they generate tokens? Which is where we started our conversation today. And so they need factory capacity in order to increase their revenue growth.

1:07:36But nonetheless, we also said that AI is more than chatbots. It includes all these different fields of science. NVIDIA's AV business is coming up on$10 billion. Nobody ever talks about that. And you have to train world models. You have to train these AVs. And it's happening, robo taxis happening all over the world. Our AI work with digital biology. Our AI work in financial services. The whole industry of quants, quantitative trading. is moving towards... Yeah, exactly. They used to be classical machine learning, a whole bunch of human featured... They call quants, right? These specialized mathematicians were trying to figure out what the predictive features are.

1:08:20Now we use AI to figure it out. And so in order to have... Instead of having quants, you need a lot of supercomputers. Financial services is one of our fastest growing segments. Billions of dollars in quants, you know, in financial services. Billions of dollars in AV. Billions of dollars in... in robotics coming up. Billions of dollars in digital biology. And so how big can all that be? Well, simple logic is this, simple math. Whether you think that AI is going to replace labor shortage or workforce shortage in any kind, let's ignore that for a second. The world is at$100 trillion in GDP. Out of that, let's just say 2%, 2 % annually is R &D.

1:09:03and let's just go back in time five years ago if you were to take the largest drug discovery company in the world drug company in the world and where's all of their r &d wet labs today what are they doing building supercomputers and so there's a fundamental shift in how they think about that two trillion dollars it used to be two trillion dollars for the old way of doing things It's not going to be$2 trillion in the AI way of doing things. Well,$2 trillion is going to need,$2 trillion of R &D is going to be powered by a whole bunch of infrastructure. And that's the reason why we're building supercomputers everywhere around the world.

1:09:41And so I think if you reason about it from the outside in, either from the foundation up, from the outside in, you come to the conclusion that what we're experiencing, what all three of us are experiencing, which is the amount of computing demand is insane. Give me an example of a startup company that goes, no, we're good. They are all dying for computing capacity. Give me an example for a researcher in any university, a scientist in any company who says got plenty of capacity. Everybody is dying for capacity. And so we have a global multi-company, multi-industry shortage. It's not just about open AI, even though open AI could use a lot more capacity as well.

1:10:24So I think how we think about this, with the narrative, the narrative is not helpful. And it's a little bit too superficial to say, how do you prove there's an AI bubble? $12 billion of revenues, hundreds of billions of infrastructure being built. It's a little bit too simplistic. Yeah. The other thing people tend to point out is the MIT study. There's some study that I think came out of MIT that claimed that most enterprise deployments of AI weren't that useful. And you're like, well, did you do the change management? Did you do a reorg? Did you integrate into tooling? How long did it even take to implement it?

1:10:58If a planning cycle in an enterprise is a year, you did something in six months. And so it feels like there's a lot of these kind of, again, overstated things that get a lot of attention, but then you map it against what's actually happening and the growth of these companies using AI. And it's just a completely different world. And if you want to find out where the world's innovation is happening, I would not go find out at an enterprise. Would you guys agree? Yeah. Enterprise is like the slowest adopters of new technologies. I would go talk to all of the startups, the 30 ,000, 40 ,000 startups that are currently doing this stuff.

1:11:32I would go talk to OpenEvidence. How's it working? I would go talk to Cursor. How's coding working, by the way? You know, I would just go talk to these people. I think it's really interesting that you see that, of course, you do have companies making, you know,$100 million plus, multi-hundred million plus progress of ARR in enterprise sales, Harvey, Sierra, etc. But some of the fastest growing companies have been end user adopted, even in conservative industries, right? Like healthcare or, you know, skeptical industries like engineering. The most conservative of all. But guess what? They are so concerned about getting the right answer that the ability to have something like open evidence, to do grounded research, high quality research, and get that research as information to you.

1:12:19Nobody wants to do research. They want answers. Nobody wants to do search. They want answers. Is that right? Abridge is a great example of that too, where they're basically making it really easy to do the physician notes instead of the physician sitting there and doing it. Back to your point on task force. past versus purpose and i think a different way to think about the demand is like there are so many jobs where you're asking the the work is actually like an impossible ask right of a doctor or a radiologist keep up with the world's biomedical knowledge yeah in r &d which is accelerating you know computing and otherwise um and then just like archive papers yeah there was a time you i've been trying to read him you and i both both used to do i don't do that anymore but But now I just load it all into ChatGPT.

1:13:01Now I just load it all in with all of the ones that are interesting. And then I make it learn it. And then I make it summarize. And then another summary. And I interact with it. But the point is, we used to do search. We don't do it anymore. I don't do search. We used to do research. The goal is to get answers. The goal is to get smarter. And these AIs allow us to help us do all that. And I think all of it, all of it comes back. It's all more helpful if you come back to the framework that says AI is a multi-layer cake and that AI is not just a chatbot. AI is very, very diverse in all of the industries and modalities and information and applications that it addresses.

1:13:47When you think about wanting to win, that America should win AI, it should not just be America should have this company win AI, but we should try to win across the board. And across domains. Across domains, exactly. And when we think about open source, all of a sudden, this is a helpful framework. When we think about winning, it's a helpful framework. When we think about energy, it's a helpful framework because we need factories. Factories need energy. And without energy, we have no factory. without factories, we have no AI. That's a helpful framework. And so I think if we have a better understanding, a system, a framework for understanding what AI is, I think the narratives will become more common sense.

1:14:31The narratives will become more pragmatic, become more balanced. We want to keep people safe. But one of the best ways to keep people safe is advancing a technology quickly. And I think the industry is doing that. And I'm very proud of the industry for doing that. No one wants to drive a car from, you know, the first decade of cars. No way. I think... ABS is a really good thing. Yes. ABS is a really good thing. Lane keeping is a really good thing. There's no question FSD is a really good thing. And I think people will be excited about the, you know, third or fourth year of AI. Yeah, no doubt. And I say with great pride that the industry made tremendous strides this last year.

1:15:16All the technologies we've mentioned and that the scaling laws are so intact that we now know that more compute, more intelligence.

1:15:30and gosh, the innovations in one sector diffuses and spreads across all of the other sectors so fast. I'm so happy to see all that. And so I think the next five years, it's going to be extraordinary, no doubt about it. And I think next year is going to be incredible. Amazing. Well, we're excited to talk to you at the end of next year too. Yeah, looking forward to it. Thanks so much. Thank you guys for all the work that you guys do. Congratulations. What a great year. Wow, it's an amazing year. Yeah, a lot. Thank you. Thank you. Thanks, Johnson. Happy New Year.