Accelerating the AI-Driven Future Together

Accelerating the AI-Driven Future Together: Jensen Huang and Anirudh Devgan in Conversation

Estimated read time: 1:20

Summary

In this engaging conversation at Cadence Live, Jensen Huang and Anirudh Devgan explored the rapid evolution of artificial intelligence and its transformative impact on industries and technology. They delved into the transition from generative AI to reasoning AI, and the emerging concept of AI factories. The dialogue also touched on the strategic collaboration between Nvidia and Cadence in driving innovation in chip design and system automation, along with the role of digital twins in revolutionizing industries. The discussion was both insightful and forward-thinking, highlighting a future where AI Agents and accelerated computing will become foundational elements of technological infrastructure.

Highlights

The AI industry evolves rapidly, with generative AI moving towards reasoning AI 📈
AI factories are being defined as new types of data centers focused on processing AI tasks efficiently 🏭
Agentic systems to involve AI agents as assistants to human designers and engineers 🛠️
Nvidia's innovations in chip and system design fueled by partnerships with companies like Cadence 🔗
Digital twins are revolutionizing how industries simulate and plan their operations in AI and robotics 🌐

Key Takeaways

Jensen Huang and Anirudh Devgan discuss AI's transformational impact across industries 🚀
Introduction of AI factories as a new paradigm in data centers 🏭
Nursemaid AI to Agentic AI - Evolution in AI’s ability to reason and assist 🤖
Partnership between Nvidia and Cadence pushing the limits of chip design and system automation 🤝
The holistic vision for AI extending its reach into enterprises, sovereign AI, and industrial AI 🏢

Overview

The conversation between Jensen Huang and Anirudh Devgan offered a deep dive into where AI technology is heading and its potential to reshape industries. The duo explored the transition from perception and generative AI to an era of reasoning AI that eerily mimics human problem-solving. This change heralds a new wave of applications and responsibilities for AI across various sectors.

Huang introduced the concept of AI factories, which are set to transform data centers into entities focused solely on maximizing AI task processing. This move represents a shift in how computational power is structured and utilized globally, leading to discussions on industrial priorities and economic infrastructures designed around AI capabilities.

As the conversation concluded, the importance of partnerships was underscored, notably the collaboration between Nvidia and Cadence in enhancing chip design automation. This relationship exemplifies how companies are combining strengths to drive the AI revolution forward, preparing for a future where AI and human expertise work in tandem to push technological boundaries.

Chapters

00:00 - 00:30: Introduction and Welcome The chapter introduces the event with music and applause, welcoming the audience to Cadence Live. Dr. Anarude, the President and CEO of Cadence, greets the attendees and promises an exciting day ahead.
00:30 - 03:00: The AI Transformation: Current State and Future Vision The chapter begins with an introduction of Jensen, a well-known business leader and technologist, highlighting his global recognition and influence.
03:00 - 04:00: Collaboration between Companies and AI-driven Innovations This chapter discusses the collaboration between companies and innovations driven by artificial intelligence. It begins with a positive note on the importance and impact of collaboration with Cadence, a company that seems to have played a significant role in the speaker's professional journey. The chapter includes an appreciation for drawing benefits from partnerships, emphasizing friendship and cooperation between the two parties involved. The conversation hints at various ongoing projects, suggesting a dynamic and evolving partnership aimed at achieving common goals through joint efforts.
04:00 - 07:00: The Technology and Industrial Arc of AI The chapter begins with a conversation about exciting news regarding AI advancements. The speaker emphasizes the transformative impact AI is having on various industries and regions globally. An assessment of the current state of AI is given, highlighting the rapid pace of change where a single year can feel like a decade in this fast-evolving field. The discussion is structured around two perspectives to provide a comprehensive view of AI's present situation.
07:00 - 11:00: AI Factories and the Infrastructure for the Future This chapter discusses the evolution of AI technologies, highlighting the transition from perception AI to generative AI, which includes capabilities like translating between different modalities such as text and images. The focus is on the advancements into reasoning AI, indicating a progression towards more sophisticated AI systems.
11:00 - 15:00: Innovating at Scale: NVIDIA's Approach The chapter discusses NVIDIA's approach to AI, focusing on reasoning step by step rather than providing one-shot answers. It highlights the capability of AI to solve new problems through understanding rules, principles, and reading to gather information. This method enables AI to tackle questions it has never encountered before by processing information in a structured manner, reading relevant content, and thinking through each step before concluding.
15:00 - 18:00: Digital Twin and Simulation Technologies In this chapter, the discussion revolves around the evolution of artificial intelligence (AI) technologies, primarily focusing on AI's ability to understand and reason about the physical world. The concept of 'physical AI' is introduced as a next-generation technology that comprehends the laws of physics and common-sense elements of the environment including friction, inertia, cause and effect, object permanence, and gravity. The development of reasoning AI is highlighted, suggesting that it opens up new opportunities and categories in technology advancement.
18:00 - 21:00: Power Efficiency and AI Factory Economics The chapter discusses the future of AI in chip design, where AI agents, referred to as 'agentic systems', will play a pivotal role. The vision is for every chip designer to have AI-based assistant designers, similar to how software engineers currently work with coding AIs. This future will involve hiring numerous AI-assisted designers, advancing the work collaboratively with initiatives like Jedi.
21:00 - 26:00: The Rise of Enterprise and Sovereign AI The chapter discusses the concept of 'agentic AI,' referring to advanced AI that can perform tasks typically done by humans. It highlights how this AI could revolutionize fields such as biological engineering by amplifying the capabilities of human engineers. The chapter also touches on the future of AI in robotics, indicating a significant technological evolution from software to physical applications.
26:00 - 31:00: The Future of Industrial and Physical AI The chapter explores the transformative role of AI in industrial sectors. It discusses how AI technology impacts various dimensions including operations, product development, and the very structure of workplaces. The focus is on the multifaceted changes driven by AI, highlighting its potential to redefine industrial processes and outputs.
31:00 - 36:00: The Role of AI in Intelligent Systems and Robotics The chapter discusses the evolving role of AI in the creation of intelligent systems and robotics. It emphasizes the necessity for new types of manufacturing systems known as 'AI factories,' which differ from traditional data centers. Unlike data centers that store files and support enterprise resource planning systems, AI factories are specialized for the development and deployment of AI technologies, indicating a shift in infrastructure to support emerging technological needs.
36:00 - 43:00: Digital Biology and the Future of Drug Design The chapter titled 'Digital Biology and the Future of Drug Design' discusses the transformation in data center functions with the advent of AI factories. These AI supercomputers, unlike traditional data centers, are designed to produce tokens continually, drawing a parallel with the dynamo systems of the past that generated electricity. This introduction marks a significant shift towards AI-driven processes and implications for future technological advancements.
43:00 - 43:30: Conclusion and Acknowledgements The chapter explains the concept of AI factories and their economic model, where electrical input is measured in terms of cost per kilowatt hour and output is measured in tokens per dollar. It highlights the scale of these AI factories and introduces the idea of emerging industries and companies that form the next technological and industrial layer.

Accelerating the AI-Driven Future Together: Jensen Huang and Anirudh Devgan in Conversation Transcription

00:00 - 00:30 [Music] Please welcome President and Chief Executive Officer for Cadence, Dr. Anarude [Music] [Applause] [Music] [Applause] [Music] [Applause] Devga. Hello everyone. Welcome to Cadence Live. We have an exciting day
00:30 - 01:00 for you and we'll kick it off with with a truly amazing guest, you know, person who needs, you know, virtually no introduction, known by first name throughout the world, the most uh premier business leader and technologist. Please welcome Jensen. [Music] [Applause] [Music]
01:00 - 01:30 Everybody, I love Cadence. What would I do without Cadence? I'd still be a bus boy at Denny's. That's what's going to be. Well, Jensen, welcome. Thank you. And thank you for the collaboration between the two companies. And thank you for your friendship and partnership. Yeah. So much is going on and we're doing so much together. Yeah. So much is going on. We're doing so much together. We
01:30 - 02:00 have some pretty exciting news today. Absolutely. So before we get to the exciting news, one thing you know of course you are you are driving this amazing transformation. It's affecting all industries affecting all kind of regions of the world. So what's your assessment of the state of AI? How is AI now? There's lot of you know one year seems like 10 years in this space. Yeah. So what's your take on current AI state? Um let's let's let's let me answer that in in in through two lens. The first
02:00 - 02:30 lens is what's what's happening to the technology and the second lens is what's happening to the industry. The so the technology as you know the last time we spoke we went from uh perception AI to generative AI and we can understand all kinds of modalities now. we can translate from from text to text and image to text, but even text to images. And so we now have generative AI. Um, we're now well into well into generative AI into the world of reasoning AI. And so now you could give
02:30 - 03:00 it a prompt, you can give it a problem. Instead of a oneshot answer, the AI will reason about it step by step before it takes a one-shot guess. It might reason about it so that uh it can even answer questions that that uh it's never seen before. uh just through rules and principles and uh maybe through reading and right and so we now have read it would go off and read get some information read about it uh go to the next step think about it uh and after it's done it might even
03:00 - 03:30 reflect about the quality of the answer and if it's not good enough it reasons about to see that sometimes tells you what it's doing right yeah exactly you know it's not very confident about the answer and then and then of course the next next generation beyond that is uh physical AI AI that understands the laws of physics, the common sense of the world, uh understands friction, inertia, cause and effect, you know, object permanence, gravity, those kind of things. Uh and and uh uh a so reasoning AI uh gave gave us a whole new category
03:30 - 04:00 of products called agentic systems. A right AI agents and this is going to be very important in the work that we do together with Jedi. Mhm. You know, this is our vision that someday all of our chip designers would have a whole bunch of assistant chip designers. Just as every software engineer today has already coding AIS, um, every chip designer in the future should have assistant chip designers and every system designer would have assistant, right, system designers. And we're going to, you know, hire a whole bunch of them
04:00 - 04:30 from Cadence, right? And so I hire one biological engineer, you know, and then I rent a thousand from an uh and and then and then you know every every uh biological engineer is a super biological engineer. Right. They have a whole team. They have a whole team. Right. Exactly. Yes. And and so so now that's the agentic AI. The next generation of physical where physical AI comes into the world is robotics. And so we're seeing all that. So that's that's kind of the technology arc. The
04:30 - 05:00 industrial arc is is every bit as interesting. You know, we uh we talk about what uh AI can do, the technology aspect of it, what its applications are, um how it's changing SAS, how it's changing the way we we do work, how it changes the the type of products that we could build, right? So everything from the way companies are run, the way factories are run, the way products are built and what they can do completely changed. That's the kind of the technology layer. The the second layer
05:00 - 05:30 is is now recognizing that this form of technology requires manufacturing. You know, and and and and the work that we're doing together recognizes that that there's a whole new type of factory that's necessary and we call it AI factories. you know, instead of calling it AI data centers because data centers had a purpose and and it holds files and uh supports all of our employees and runs our ERP system, you know, that's
05:30 - 06:00 not going to change. Data centers are going to continue to run uh software that we wrote and um but there's going to be a new type of data center and I call it AI factory because it has one purpose and one purpose only. You have this big machinery inside instead of 300 years ago the dynamo system that produces electricity. Now, we have these AI supercomputers and they produce tokens and every single day they're producing millions and millions of tokens. And so, you you pay for
06:00 - 06:30 electricity coming into the AI factory and that's dollars per kilowatt hour and what goes out is uh millions of tokens per dollar, right? And so so between those two ratios, you know, that's your revenues, that's your profits and and that these AI factories are fairly gigantic. The next layer, the next layer, um uh technology, industrial, so this is a whole bunch of new industries, new companies. And uh the next layer
06:30 - 07:00 above that is is infrastructure. People are starting to realize that every country, every company will need this infrastructure because it affects everything we do. just as profound as electricity was. Um we had the energy infrastructure and then we had the the we are in the generation that invented this new infrastructure called information infrastructure. Right. Nobody understood what we were talking about turned into the internet the information superighway. Right. And so
07:00 - 07:30 now we have a new new superighway the intelligence superighway. Exactly. And the world doesn't understand it yet. But clearly in you know in 20 years time when you look back we'll be sitting here you know for the first time talking about this new infrastructure called intelligence and every every country will have it every you know everybody will use it every company will build built on top of it. Every industry will be affected by it. So that's kind of the two lens I would say. Absolutely. Thank you. You know one thing I'm amazed by is you know of course Hopper was a huge
07:30 - 08:00 jump. I mean Nvidia has transformed continuously, right? And then and then Blackwell is even more amazing than Hogin. We have all these exciting results in Blackwell and then at GTC you showed this really impressive road map going out several years. So I think what I'm curious about is how do you you know like it's very difficult to have this kind of such a big and important company but at the same time innovating at a very fast pace. So how do you do that? What is the
08:00 - 08:30 culture? How do you make sure that such a big organization so critical is also out innovating, out executing? It's truly remarkable to see. You're running a master class in how to how to run these L. So what is there some thoughts on you know for the audience? Well, there's there are two things that and in fact this is this is perfectly aligned with with the vision that you described for Cadence when you and I first met and got together when you became CEO. Um if you look at look at look at um at
08:30 - 09:00 the chip level I would say that the chips are continuing to advance uh because uh we could you know we could obviously make them large larger we could stack it in a whole lot of different ways um but those are those are kind of incremental you know they're they're 50% 63% 24% that kind of incremental type stuff and it's it's important we'll take all of it and and um uh we also reformulate the way computation is done maybe using things like tensor cores. So instead of instead
09:00 - 09:30 of uh processing in vectors uh maybe you process it in blocks uh maybe maybe uh we change the precision of the way we do mathematics because because u obviously neural networks is not precise uh it's it's statistical and so therefore maybe maybe uh each one of the layers could be done in different level levels of precision. maybe sparity could be taken advantage advantage of. Um and so so we'll we'll do all that kind of stuff architecturally, but the big idea really
09:30 - 10:00 is that that um we're no longer just a chip, we're a whole system. And so and in fact, we're not just a whole system, we're a whole data center. And if we could innovate the architecture at the whole data center level and so Blackwell of course did all the things that I just mentioned, used all the latest technology and you know all of the tricks that I mentioned. Um but we also had MVLink 72 that scaled up um the compute nodes and by doing that we have to create new certities we had to go to
10:00 - 10:30 look at cooling the system has to be redesigned so the way the system looks is completely different than the past uh we refactor the way that systems are interconnected and so so it looks different the cabling is different the architecture is different but you know if that the if that whole thing is a is a fabric for you you know for innovation is a medium for innovation, then all of a sudden your creative ideas can come. And then beyond that, how we scaled it out was different. And so we could change the switches, we could change the
10:30 - 11:00 networking protocols. Uh and then on top of that, we rewrote the operating system of how uh software is deployed and we call it Dynamo. And so basically at every single layer we changed it as a result you got, you know, instead of 2x you got 30, 40, 50x. It's kind of crazy. Exactly. Yeah. You know, one thing we have talked for years and Jensen has Oh, but I was going to say, yeah, but that's the reason why Andrew went from EDA to SDA. Exactly. There you go. Yeah. Exactly. System level design. So, we
11:00 - 11:30 have discussed this for a while. You know, not just chip design, system design. Also, Jensen asked me even few. He always asked me questions where he already knows the answer. You know what I'm saying? No, it's because you know, people want to hear from you, Jensen. They want Yeah. Yeah. Yeah. Whatever. Yeah. Yeah. He knows. He look this that is the new vision of cadence to move the the design automation move the design medium from a chip to an entire system.
11:30 - 12:00 Yeah exactly. Yeah. Now one thing you also asked me and told me for years if you remember so you know first of all as you know we have a great partnership on on palladium right and palladium is like a is like a boolean supercomputer so it can design chips especially like logic verification emulation which is critical for for design but Jensen always asked me like what about rest of the stuff you know there's lot of numerical computation there's lot of other system you know other disciplines for engineers
12:00 - 12:30 right So can we accelerate sciences in general you know design design activities and we also have wanted to do this for a very long time actually and actually you were saying that you were like why has it not happened in EDA and SDA and the and what I told Jensen at that time was that you know like if you go back like 10 15 years we did lot of what we would call parallel computing but we did it on the CPU because we didn't have a good architecture
12:30 - 13:00 But now especially with with hopper and then with blackwell especially not only you know everybody knows GPUs are fast but I think what people don't realize it that GPUs have become lot more journal purpose at the same time right you know have become a lot more and then you have now grace and CPU to go with the you know closely and the envy link and all that so finally I think you know a couple of years ago it was clear to us and of course you have seen this for a while that this can be a computing ing
13:00 - 13:30 platform uh for much bigger things more general purpose application okay and that is what we want to announce today right this is years in the making and it's a combination of advancement in the on on the hardware and system side by Nvidia and then of course we have to ret rewrite our software to take advantage of that and you know you always talked about this hardware and software together so let's roll the video Okay. Yeah.
13:30 - 14:00 Wow.
14:00 - 14:30 I know. They're in shock. They're in complete shock. So, Jensen, you know, that's beautiful. trying to learn from your beautiful videos. Yeah. Well, hey, hey, so this is what happened. You know, as as you know, uh, Palladium was was uh going back to the the origins of our company. We deeply believed in the concept of of chip system and
14:30 - 15:00 software codeesign and if you don't codees uh then you can't innovate in that loop um at the same time obviously and and ultimately accelerated computing is not general purpose computing. Accelerated computing is about accelerating the software and and there only type certain algorithms that are conducive to being accelerated and that's why code design is so important. You have to change the algorithm, change the architecture, change the algorithm, change the architecture. And so, so that palladium was so so deeply ingrained in
15:00 - 15:30 our company in a lot of ways. Don't forget what Palladium is. Palladium is Nvidia's digital twin. It is the digital twin of the most important thing we do. Long before we taped out Blackwell, I knew Blackwell was going to be perfect because it lived in that digital. It lived inside Palladium for a long time. Isn't that right? It was it was alive to me for a year, you know. I was only wonder when it was going to pop out, right? You know, and and so like it's been alive in palladium for
15:30 - 16:00 like nine months. You know, and and if I bought if I bought twice as many palladiums, can you do it in four and a half months? Yes. Now that now that now that I said that out loud, it kind of makes sense. And so and so uh the thing that's really incredible is is um beside our our chip digital twin, we've never been able to digital twin almost anything else. And we want a digital twin everything we built. And that's the reason why Millennium M2000
16:00 - 16:30 is such a fantastic vision. Now we can accelerate uh the rest of EDA. We can accelerate SDA and as and now we we're going to we're going to have digital twins of everything that we do. And I I didn't tell any of this, but but uh today uh I would like to uh place a purchase order right here on stage. Right here on stage. I haven't
16:30 - 17:00 told anybody yet. Okay. So, so uh right here on stage, 10 Millennium M2000s, maximum configuration. Thank you, James. Whoever the salesperson is is calling on on me, just wow, home run. Yeah, he's there doing the math and he's going, I like this. This this is going to change. And and so this is a big deal for us. Uh we started uh building our
17:00 - 17:30 data center to get ready for it. Uh we're obviously working on accelerating the EDA software, the SDA software. uh with millennium will speed it up 50 60 100 time and as a result of that some of these numbers so that you can see this is interesting see one box of course has bunch of black wells and optimized software depending on the application is equivalent to tens and thousands of CPU sometimes 80,000
17:30 - 18:00 sometimes 100 thousand depends on the application but that's huge and And you know, listen everybody, I just got to pat myself on the back. Yeah. Yeah. Amazing. Good job. And the thing is that you know like we have done because a lot of times the customers will run these applications on multiple CPUs like we have done. But there is a limit to how many even if you have a lot of CPUs for
18:00 - 18:30 one job there is a limit of how much we can do. And I did this for you know from 2005 onwards we ran on more CPU more and like not just on a single chip you know like eight or 32 CPUs but there's a limit to few hundred after few hundred CPUs the application will not scale so typically you know a lot of the use cases are 300 500 kind of CPUs for one scale anymore because the overhead of communicating with each other you know this is the MDOS law problem and then you know with with Blackwell not only there is more computer in the chip but
18:30 - 19:00 the interconnection between the chips and So the thing is that yes it is equivalent to 10,000 or 50,000 CPUs but you can never run a job. So it is actually more impressive than it looks because you can never even if you had 10,000 CPUs it would not scale that well. That's right. This is for single applications and if there are multiple you know batch runs you can always tag them together. So and then you know what is exciting is is the breadth of the application you know silicon system and
19:00 - 19:30 and of course bio bio also requires lot so so thank you Jensen for this collaboration thank you for appreciate it for and so so let me let me just let me just say this and and you you know I'll just bring it back down to earth for us we're going to design the entire system completely inside a digital twin and we'll simulate it as if we put it into you know a heat chamber as if we literally turned it on and see exactly how it's going to perform thermally. We're going to put the entire
19:30 - 20:00 data center, the entire AI factory, and we've put that together right between us. We've built an entire AI factory and simulated its thermal performance and found the heat, the hot spots, and how to ventilate, how to cool that entire AI factory completely inside a digital twin. Exactly. With a cadence reality digital twin with Omniverse. Yeah. Exactly. See, that's the other cool thing that that I think sometimes people don't appreciate the power of omniverse. Of course, everybody knows this one, you know, and and Jensen and I talk about
20:00 - 20:30 this three layer cake. Of course, he's he has cookies and I have cake, but he does lot more. By the way, I've seen your videos. You like you you like taking, you know, in your kitchen, right, with co you had like out of the oven, you take out like Yeah, that's that's impressive. That's how you cook an oven. Yeah. Yeah. But you know last so in now I make them and I put them into ovens. Exactly. Yeah. They came out of oven now I got to put it back in. So in that u uh in the in the digital twin
20:30 - 21:00 with the cadence reality you know we talk about this uh three layer cake talking about oven. So you talk about the accelerated compute at the bottom and then the you know principle simulation in the middle and then AI orchestration and agents at the top. That's right. So you know everybody knows that you know Nvidia is of course doing a lot in AI of course doing a lot in accelerated computing and in principle simulation of course we have a lot of history but Nvidia has done a great job with omniverse so not only
21:00 - 21:30 built cadence reality it's all plugged into omniverse and then you can simulate the so now you have a virtual wind tunnel virtual AI factory a virtual you know entire manufacturing plant and everything is running with software in the lube the simulations are either principally based fully accelerated on Millennium and so you know or it's AI infused AI infused yeah right AI fizz ML and all you know because you just like in graphics you know and you talked about at GTC you don't have to render
21:30 - 22:00 each pixel physically you can do with AI same thing in simulation if there are multiple time points or multiple variation not only can the principal simulator be accelerated by accelerated computing it can be further accelerated by AI that's right by not doing each point and do like this model. You ground every you know you might ground every every few frames or every few pixels and then you you know predict everything else now on the on the dig AI factories. Okay, that's that's a big thing and you
22:00 - 22:30 know, thanks for this like cadence reality collaboration, but what do you see with this? You know, this thermal is a big problem, right? And so I'm going to be asking for a delivery schedule here pretty soon. Just Okay. Okay. He You could tell he's trying to change the subject. No, no, I want I placed I placed the PO. I I'll drive the truck myself. Okay. Yeah, I'll drive the truck myself. Okay. It's I'm coming. I'm coming. So, you know, it's one and a half tons each. Yes. I'm coming. I We'll We'll get it. We'll get few trucks, J. Yeah. Yeah. Yeah, I just bought 15 tons
22:30 - 23:00 of computing. Thank you. Yeah. Isn't it exciting? No, but that's actually that's that's a very good point that that Nvidia and of course we collaborate on the EDA and the SDA also uses the we're working on Jedi. Yes. Agentic AI, right? We're working on uh Millennium accelerated computing and and AI emulation and for digital twins for simulation of everything. Emulation
23:00 - 23:30 of everything. Digital twins of everything. And then we're also working on Orion, our digital biology platform. Yeah, exactly. I'm on top of my stuff. Yeah, you are. Yeah. No, I'm coming. I'm coming with the truck. I'll bring few trucks and we'll make sure it works. You know, like I want to have his same photo like you had with when you delivered a box to OpenAI. That's I'll deliver a Millennium Box. Okay. I'll be ready. Yeah. Yeah. I'll be ready. So So Jensen, what what one thing that
23:30 - 24:00 you know on the digital factory side or AI factory side is this whole you know power consumption, right? It's huge. It's a huge problem, right? And then that's why we have this, you know, we're building we're building AI factories that are one gawatt large one gawatt AI factory the the capital necessary is about $60 billion $60 billion just so you know what $60 billion is $60 billion is like like Boeing not the plane the
24:00 - 24:30 whole company the revenues of the whole company is one AI factory for a year okay and so that kind of puts it in perspective and we have we have tens of gigawatts of of buildout in the coming years and then of Nvidia also has DJX cloud. Okay, that's that's exciting. And this whole building of factories and all your partners throughout the world. What do you see in terms of you know like power consumption of course is big thing now
24:30 - 25:00 but one thing people don't realize even for millennium you know if you do it for the same you know ISO performance it's 20 times lower power. That's right. Exactly. That's the whole point because a lot of time people talk about how much pro power these things are taking. Yeah. But for the same you know for to do that computation is still much more power effic. That's right. It's never it's never ever about power. It's about power efficiency. Always about power efficiency. So let's let's put that into money. And so let's say let's say you built a a data center. It's an AI
25:00 - 25:30 factory. Uh everybody's AI factory. you know, some people have more capital than others, but the the power that's available to that AI factory is the power that's available to that AI factory. That's it. That is the only laws of physics. And so, whatever you provision, um, that's all you're going to have. Well, if you're limited by power and your goal is to maximize the million tokens per dollar, okay, then you're trying to maximize the
25:30 - 26:00 million tokens per watt. Mhm. And that if you maximize million tokens per watt per per watt then your revenues will be the highest. And so if the architecture has 20 times using your example 20 times the performance throughput per unit energy. Yes. Then the revenues of that data center is 20 times higher. That's the amazing thing. That's why it's an AI factory. Exactly. It's ultimately about revenues generated. And this is the
26:00 - 26:30 future. We have a new, you know, the last factory was, you know, electric power plants and the more energy efficient they are, the, you know, the more the more kilowatt hours they generate and therefore the more dollars they generate here is the more tokens you generate from the data center and more tokens you generate, the more dollars you generate. It's a very very big deal. The difference in architecture per perf per watt being 2x versus 3x versus 4x just translate that to revenues. Would you like to be a $10 billion company this year or would you
26:30 - 27:00 like to be a $20 billion company this year? That's it. One thing on power consumption, I saw all these reports that you know people are putting generators right next to the Yeah. and some are on the grid, some are like local power generated. So power generation is going to be a big deal. The reason for that is because the power grids aren't going to be enough to sustain the growth of this industry. And this this industry likes to be built we like to build this industry on shore. And if you want to do that, then we're going to we're going to see a lot of, you know, we're going to see um of
27:00 - 27:30 course a lot of diesel power generators and um you're going to see some SMRs and you know, all kinds of stuff. And then the other thing you have talked about and is this whole you know there are these big of course there are these big cloud providers which have uh you know of course footprint all over the world and then there is these sovereign AI you know people you know all these countries are building their own uh models and they need to right so and
27:30 - 28:00 then there are these enterprise AI so how do you see these different parts of the market right we're in the first stage of AI if you look at all of our economic success everything that that's been done, everything that's been talked about, it's basically cloud AI. Yeah, it's basically AI for the AI natives. You know, these are startups, brand new companies. OpenAI is a new company. Um, right, Perplexity is a new company, really fantastic. You know, all these all these are new companies. Uh, Runway ML is fantastic. Um, Black Forest is
28:00 - 28:30 fantastic. Um, there's a lot of really cool companies and but they're AI natives. Um but we you and I know that AI is going to infuse into every single aspect of everything we do. Every company will will be uh run better because of AI. We're going to build better products because of AI. And so the next phase, the next chapter is enterprise AI. The architecture of enterprise AI is very different. Like for example, uh a lot of companies still run off of their mainframes, IBM mainframes and Fujitsu
28:30 - 29:00 mainframes. So the question is how do we infuse AI into those systems? You can't you can't just, you know, cleave it off and say we are going to brand brand new enterprise IT. You're going to have to add to it. And so you and I both know that nothing no IT systems ever passed away, you know, and and so the standards just keep getting added on top of each other. And so we're going to have to augment um AI into the world's enterprise AI system. And and that's one of the one of the initiatives of our company is enterprise AI. Whole new
29:00 - 29:30 systems, whole new networking, whole new storage stack. Everything is all brand new. Um yet integrated into uh existing software infrastructure. Could be VMware, could be Red Hat, it could be, you know, even even um classical uh Java. And you know, you know, one of the things that's really going to be really cool is AI understanding cobalt, right? And so and um and so that's enterprise AI. The next click after that uh uh simultaneously with enterprise AI is
29:30 - 30:00 exactly what you say sovereign AI. Every country is going to build their own infrastructure. They have you know your your AI your intelligence embeds your uh history, knowledge, culture, values, you know that's in in collection is intelligence. It's not just knowledge, right? It's not just, you know, things you read in in a document. And so so intelligence has to be something that is codified and built in each one of the countries in each one of the regions. And so we're going to see
30:00 - 30:30 this being built out everywhere. We've built out uh we're building out United States that's first but but countries all over the world. Japan is getting built out, UK is getting built out just you know all the countries are getting built. India is building out and so there's there's a whole bunch of initiatives going on. And so that's simultaneous going on. And the next click after that the big idea you know this is the really big stuff is is uh industrial AI you know that's the physical AI the robotics omniverse is designed for that all the things that we were just talking about um you know the
30:30 - 31:00 number of plants and factories being built around the world is incredible for all of the geopolitical reasons and otherwise reasons uh people realize that re-industrializing their country is really important for economic security national security not to meion a stable stable society because you have a distribution of of uh different types of workers and different type of labor um that needs to be available in different societies to balance it out to
31:00 - 31:30 to properly shape that society and and so so re re-industrialization is going to cause plants and factories to get built all over the place. We're building a whole bunch here and we want to build it first in a digital twin simulate it as a digital twin and then build it just like we build right just just the way I use palladium build black well that's right is that right millennium can do that you know but the physical physical itself I mean there's report so palladium was a home run millennium is going to be a giant home run exactly
31:30 - 32:00 giant home run you heard it first yeah giant home run new product category ategory for this entire industry. Isn't that right? I love your quote. I don't know if you see new product category. New product category, right? With this physical AI the I mean there's so many reports even yes two days ago there was like Morgan Stan just so you guys know we didn't rehearse. I don't know when I'm supposed to leave. See the thing with Jensen is that he's
32:00 - 32:30 he's always ready. You don't need to rehearse. When do you need to rehearse? You're always, you know, always I'm your wingman. He's Batman. I'm Robin. I'm just he's 100% charged 24 hours. Right. Yeah. Right. So, uh I saw a report two days ago. There's so many reports that physical robotics robots will be 5 trillion. Yeah. Market. That's huge. Yeah. Well, on first principles, just go
32:30 - 33:00 back. Uh it turns out the world has a severe labor shortage. While we're sitting here talking about AI, you know, changing jobs, taking jobs, um, first of all, AI is not going to take your job. Somebody who uses AI is going to take your job. And, and so, so you have to engage AI. Uh, but obviously AI is going to change every job. And that's that's for certain. It's changed my job. It's changed every job. And, and a software programmer at Nvidia job completely changed. Uh, but the
33:00 - 33:30 world has a short a severe shortage of labor. uh the the the people that that um society is changing its its uh preference for the type of work that they do uh population decline all over the world right we know this and so we know we need AI and robotics to come and because otherwise the the world's GDP is not sustainable right and we want the GDP to grow and we want inflation to come down and so the only answer really
33:30 - 34:00 is is um we need to get to robotics as fast as possible and that is a completely different stack, right? Yeah. So, the hardware, you know, the chips are different. You're designing special chips for robotics. They have to be much more power efficient than the cloud. Uh the the simulation digital twin is different. Has to be like we built the world's first robotics processor was called Xavier. Exactly. And when I first built it and I showed it to to everybody, everybody, I don't know what that is. Don't want it. And and then I built and and because I was so so deeply
34:00 - 34:30 hurt by the reaction from everybody, I built the second one. And so obviously I didn't build that one well enough, you know, and or I didn't tell you hard enough. And so so I built a second one. It's called Orin. Home run. Bam. You know, home run. Exactly. Self-driving cars. Yeah. Robotics. Yeah. And then autonomous system. Autonomous system has lot of similar. So the car multi-billion dollar business for us now. Exactly. Yeah. Crazy. And so we got another one coming. is called Thor and so yeah
34:30 - 35:00 so the chip is different of course and the system is you know the hardware system is different uh this digital twin has to be different you know more numerical so something like millennial omniverse is the software version of digital twin I'm going to plug millennium into it exactly that omniverse that's right that's the digital twin and then the AI model is different you want to talk about you have a new AI model like physical you talked about it at GTC right I love that you have is new. It's not an LLM model. It's like a word model, right? So all
35:00 - 35:30 these three things. So this is this is an incredible thing. Yeah. If you could have models that reason, text models that reason, you could have physical model, physical AI models that reason. So So you're driving down the road. You're driving down a road and the AI model that's in the car, of course, is a physical AI model. It understands the physical world. Okay? It understands you can't go through this, you can't go through that, and understand how to reason about traffic and roads and road signs, and it reasons about all that. And so you come you come up to a stop you could actually tell your car ask your car what are you going to do next
35:30 - 36:00 and the car will explain to you I see this I see that and therefore I've decided I'm going to pause for a little longer before I take a right turn. And so so uh and then and if it does something wrong you could actually reprimand it. What did you do wrong? And and the car say well I think I might have missed that. You know I'll never do that again. Good. You know, reasoning AI socredible. So the models are different, right? Car for the robots
36:00 - 36:30 for the it understands. It understands that if it has to go from here to there, it can't go through this table. It's got to go around, right? Yeah. Kids know that. Dogs know that. Yeah. But AI doesn't know that. So we got to teach an AI to know that. Absolutely. Absolutely. Pretty cool. Yeah. That's why I really like that the new kind of word, you call it a word model, right? Rather than MLM model. It's a world model, right? So the all the three layers of the stack they all get transformed but but there's automotive in physical there's automotive autonomous systems there is like humanoid robots there's industrial
36:30 - 37:00 robots tens of millions of drones needs to get built so is like I know you're working in all all of these areas but is one like more you know in the physical AI one will happen faster than others or how do you see the different the one that's going to happen fastest largest is self-driving cars you know that's that's that's already a 5 billion dollar business for us. Mhm. Yeah. Actually, I went to to China a few months ago and they're all these AV companies. Yeah. And all using Nvidia chips and software.
37:00 - 37:30 It's remarkable. You know, VYD, all these, you know, you talked about that. You know, it's quite remarkable how how well because people don't realize that, you know, they realize the cloud part, but not the automotive part of of of Nvidia. We're a three computer company. Right. Right. We're the the one that trains the AI models and the digital twin Millennium and then the edge computers. Yes, the three computer company. So you think the self the cars will happen first? first because it's the easiest in a lot of ways. It's the
37:30 - 38:00 easiest, you know. You know, if if you know our parents could drive, you know, how hard could it be? And so, so, you know, and it was remember we we shaped the world for our limitation of our systems. Uh, but for human or robots, you can also shape the world. We could put them into specific domains. That's why I'm quite quite enthusiastic about humanoid robots. You you don't have to build a human or robot that can do that could operate in every in any part of the world. Yeah, you know, you could give it a particular domain, you could, you know, adapt them. They could they
38:00 - 38:30 could do quite quite well. Um, but these two these two will likely be simultaneously large and quite successful. And the reason for that is because technology needs a technology flywheel. And if the volume is not there, then the innovation cycle can't be there. And and so what are the two autonomous systems that are highly generalizable? cars and humano robots both high volume right and it could be
38:30 - 39:00 like billions of human over time right that's right yeah and then now the human robot fly will be quite high and then its innovation cycle will be very fast and therefore it'll get fast it'll get good fast but if you if you change the shape of it I I think a a quadriped could be quite good too because you know we've created a world where dogs could be quite successful and so so um I think a quadriped could be good but otherwise Especially off off off track, right? Off track. Right. Exactly. In jungles or,
39:00 - 39:30 you know, Exactly. Yeah. And and there's there's even ones that the one that I'm super excited about is the one from Kawasaki. Basically a robotic horse, you know, and so you could take it through the, you know, through all the different terrains and it's you're riding it. That's makes a lot of sense to me. Yeah. Yeah. Makes a lot of super exciting. That's a horse I might enjoy. That's a photo I want to see. You want to reward a horse? That would be great. That's it. And then you know we have only a few minutes left. I know you are a very busy
39:30 - 40:00 schedule. I mean the other thing we I had nothing else to do today. I was this was it. Yeah. Thank you. You know one thing I do want to touch on is is uh you know we talked about of course agentic AI we talked about you know physical AI and all digital to accelerate. But one other thing I know both of us are passionate or you talked about it for a while is is bio you know we always believe computer science and math has to be applied to bio and life sciences right that's why you know
40:00 - 40:30 cadence we started this effort few years ago and I know you have a big effort so how do you see that part in this you know life sciences bio you know with AI and in a lot of ways it's the journey is very similar to our journey with EDA you know the idea that there would be an EDA company for biology makes perfect sense today. It's called life science and and we call it drug discovery. Drug discovery is funny. It's a funny term to me and it makes a lot of sense. It's
40:30 - 41:00 kind of like truffle discovery. You know, you go out and then one day you come back, you go, "Hey, look honey, look what I brought home." And then someday, most of the days you come home, you don't have anything. Okay, but engineering is not like that. you I you know Nvidia engineers don't come home and hey hey hey you know Jensen listen this year was pretty much a bust didn't find anything you know you can't have chip discovery and we have right we have chip design and so in the future of course it's going to be drug design yes and and in order for you to
41:00 - 41:30 drug design uh to design exactly the right drug you have to your your your insitue your environment has to be well represented well understood and that that that environment is is um cells and right and the human body and you know and so we have to be able to represent that environment that systems environment really well you have to understand how to represent it the way we represent transistors and gates and functional units and right C models and ultimately system models that that
41:30 - 42:00 representation that stack of representation has to be has to be created for digital biology and then once we do that then there's manipulation technology you know and probably probably um simulation based, some of it is uh AI based and some of it could even be uh you know uh hybrid, you know, quantum classical, you know, we might be able to use quantum QPUs connected to your Millennium processors. Um and and so I
42:00 - 42:30 could imagine a a quantum millennium someday, you know, and so our GPU sitting next to QPUs and and use the QPUs for uh first principle simulations of um you know of of uh of your of your of the of the biology and and we use that to train our AM models on Millennium and you know so so I could see all that and so that that's kind of the future but but the exciting thing is because of your vision you've you really given cadence now several new growth factor, growth vectors. One, you're in
42:30 - 43:00 system design. Two, you have digital just as you have palladium for chip design, you now have Millennium for SDA, you know, and so you have this brand new growth driver and hopefully together uh we have uh the uh uh chip design or engineer a Gentic system that that will you'll provide for us the core business, you know, the EDA, the the hardware. Oh, Jens it's it's a fabulous collaboration. And I want to thank you for for working with Kaden so closely for for being such
43:00 - 43:30 a teaching customer and driver in the industry and and for the personal friendship and partnership. Yeah, thank you for the friendship and thank you all all of the cadence engineers in the audience. Thank you very much. Thank you. I love Cadence. And if not for Cadence, I just so you know, if not for Cadence, I'm still a bus boy. Just [Music]