NVIDIA GTC 2024 | John Furrier & Dave Vellante, Mini Pod

Clips
More from NVIDIA GTC 2024

play_circle_outline Emphasis on the new industrial revolution driven by AI and generative content

play_circle_outline Importance of tokens in AI and their role as the new currency

play_circle_outline Focus on NIMs and their role in the enterprise market

play_circle_outline Exploring the Interplay between Robotics, Climate Change, and Quantum Computing: A Futuristic Outlook

play_circle_outline The Transformative Influence of AI on Industries and the Prominence of NVIDIA in HPC: Unveiling 6G's AI-Powered Future

Info

John Furrier & Dave Vellante, Mini Pod

search

(soft pulsating music)

>> Hello, I'm John Furrier. Dave Vellante. The Cube is here in San Jose for NVIDIA's GTC24. We've been here for two days. Wall-to-wall coverage.

Dave, great to see you. >> Hey John.

>> I know we've been popping around. We just had the chance to get with Jensen, hear him speak privately to all the analysts who were there, as well as a few other industry analysts. We have Broadcom tomorrow. Big time meetings with those guys. Broadcom, NVIDIA, the big one. But this is kind of an abbreviated CUBE interview, John and Dave, emergency pod analysis. Because I really want to dig dive deep and unpack the keynote yesterday. And then the private comments that we heard from Jensen Wong, CEO of NVIDIA, who's an awesome legend. I just love the guy. He's got a great personality. He's kind of snarky in a good way. He's clever. He is certainly got a computer science degree, and obviously he's... Impeccable degree. And he's the story of him being, you know, in college when he was like 15 or 12 years old. Forgot the number. But obviously really strong. But just worked his way through the system in Silicon Valley. It's a great American story, I love it. One of my favorites.

>> He's got a great sense of humor too, doesn't he? Very disarming.

>> He's just a cool dude, and he's common sense, and he's an engineer and computer scientist. And so he just speaks in that kind of, like, parlance of like, "Yeah, let's solve some problems." But he really has got a great future. Let's unpack this. You know, my first takeaway, obviously the Blackwell super chip, it's the successor to Hopper. Which, it's not going away. It becomes more of a specific thing, NVLink Switch, DGX system. But I think the big thing that jumped out at me right out of the gate was his whole industry revolution. And why I like this is because we've been saying on theCUBE, you know, I've been saying we're going to have another revolution like the '60s, you know, revolution, you know, free love, summer of love, AI love, you know? This generational shift with AI. And I think he points to this as the new industrial revolution, meaning AI will, and its systems, of AI, will power new sets of experience. And specifically he went on to talk about how the world that we live in is prerecorded. It kind of reminded me when I used to bust your child spot, having the Wall Street Journal on the podcast, and say, "That was yesterday's news," 'cause it was programmed for you. Pre-recorded, pre-written down. It's not generative. So, like, our whole world is pre-recorded, if you think about it. So he's kind of right on that. And then it's moving to a generative world. And retrieval, generative AI, generative content. Obviously they've done graphics, seeds of innovation, that's the role of data. But he honed in on tokens, and that jumped out at me because I think the whole embedding experience, we're seeing vector databases. The new way to do the math in AI is going to create new experiences. And the generative part of generating stuff with AI requires tokens. And tokens will be the new currency of AI. And I think you're going to start to see another step function change with token economics, token value. And I think the whole business model question of how they price the software is going to be pale in comparison to the monetization opportunities with tokens. So this whole new industrial revolution's coming, we call it a systems revolution with clustered systems, the intelligent data platform, six data platform you've been doing. So, you know, theCUBE Research has got their fingers on the pulse here, but Jensen's on point, Dave, he's on point and there's so much more. Let's get into it. So, share your thoughts.

>> I was writing furiously because he's so compelling when he speaks. And you're right, he talked about the industrial evolution that we're entering a new era, and he used the analogy of electricity. He said, "An entire industry is going to be producing tokens." And that's what we're doing. The world of generative content is here. And he said the computer sees tokens. It doesn't know the difference between tokens, animations, whether the tokens are for image, whether it's a robotic arm. And his point was, if you can generalize, then you can talk to it, you know, with natural language. So he's basically saying that the entire industry is going to be about tokens, and he's saying none of this existed prior to ChatGPT, you know? And so he said, "We at NVIDIA generate a ton of content and Blackwell is built for this Gen AI moment."

>> It's interesting.

>> And then he went through, you know, a litany of sequences that I want to dig into a little bit with you.

>> Let's dig into it because I think this token model... He also said something like, "Tokens don't know about language," but yet we're seeing language. They didn't know if it's an arm movement. So what he was saying was, is that what we know about the prerecorded world is completely different with generative world, the generative world we're living in, where things generate on your behalf, prompt response, or reasoning, or, you know, running a digital twin to simulate a factory. So these new things are here, and their accelerated computing platform will be powering that. And you know... And I remember when Sergey and Larry launched Chrome, I said, "That's an operating system." 'Cause they, you know, the Android was yet to be launched. What Jensen's doing here, in my opinion, is setting the table for a new AI infrastructure and an operating system, Dave.

>> That's built for tokens, and-

>> Token operative systems. >> And generative AI. And it's like... So when other people talk... I mean, look, Gelsinger, as you know, is amazing. He's a legend. He'll talk in terms of how AI is going to affect society and the amazing things that we're going to be able to do with it. And Jensen talks about that as well, but he goes into this layer, which I wasn't expecting, about this world of tokens, and he said, "We're going to have models talking to each other." Why would we do that? Remember this comment that he made? He goes, "Well, think about it. We go into a room, we debate, we argue, we dissect. We're going to have models do that.

They're going to learn- >> And they're

going to come out of- >> From each other.

>> Smarter."

>> And they'll come out smarter.

>> Not to pat ourselves on the back, but we've been saying model integration is going to be a big part of that. He validates that. So, you know... So again, what we've been conjecturing on theCUBE, now becomes validated as fact with the industry, with Jensen essentially saying model integration, model synergies. Whether they build them or whether they enable someone else to build their own LLMs. He said, "We're going to be in the business of manufacturing LLMs."

>> Yeah. I mean, this is the LLM economy. And I was like, "Okay, interesting." And the token economy, so... And then he talked about, you know, simulation. Yesterday, he said, "This is not animation that you're going to see. Everything you see today is going to be simulation." And the simulation was amazing. And then he talked about reinforcement learning, and that learning loop, and training. He said, "You do it enough, it becomes reality." And then he talked about, you know, the critical tech being NVLink. They had to create this in order to allow all these different parts, these GPUs, these components, these chips, these 35 or 50,000 components, to talk to each other and share that same memory. And that is-

>> Which by the way, no one else yet has, and they're on Gen5.

>> Yeah. Nobody has that. >> Fifth generation. This on the fifth generation.

>> This is their moat.

>> He made a point to say "It's our fifth generation and nobody's even in the market."

>> Now, the interesting thing is, and we're going to hear this, I'm sure, tomorrow, at Broadcom's AI Day, Broadcom has a completely different philosophy. They're like, "Look, they're proprietary. That's cool. That's how Charlie Kawwas talked about this. That's how markets get started. And he's right about that. We're going after open, you know, we're going after scale, we're going after lower power, and we're going to do this with a different approach. We're going to use this chiplet approach." And I think there's a massive market for that. But it's a different philosophy. Now, by the way, we heard today, NVIDIA's also doing chiplets. So the difference is NVIDIA can play in that big monolithic marketplace. The biggest, most-

>> He said, "Don't call us chiplets."

>> No, no. He said... He sad, "We're not chiplets." But we heard that they're doing chiplets for media tech.

>> Got it, okay. >> Which was new information to me. I got to dig into that. But they're a systems company. They're a systems and software company, right? And yeah, they make chips, but they take those chips and they build systems. I guess the point I'm making is, and this is really where Floyer had sort of educated me, is those chiplets are very specialized, they give you a lot of flexibility. But they're connecting through relatively slower than what NVIDIA has, a communications network, and it's asynchronous. And so these big monolithic chips have a place. Who was building monolithic chips? NVIDIA. Guess who else is? Apple.

>> Yeah. >> Apple has a big shared SRAM chip.

Okay. >> He said the one thing they wanted to do is build the largest chip possible for all, that they put inside them. You know, it's not a-

>> They're betting on big.

>> It's not a chiplet, it's a monolith.

>> They're betting on big. Right. But they do, it was news to me. But they're doing chiplets evidently. And Ben Bajarin told me this on a breaking analysis, said, "There's a rumor that they're going to be doing, you know, their own chiplet design," fine. 'Cause it's probably a different market. I got to dig into it and learn more about it. But essentially he's saying, "We're building this AI foundry, and AI," he said, "Needs to understand the laws of physics." He said, "It doesn't really." He goes, "Sora kind of does." You know, it knows how snow behaves, when you're driving on gravel. You're driving, you know, above the gravel, but the gravel moves. So it's starting to train these models about the laws of physics. And then he basically said Omniverse will ground AI to the laws of physics. And so they're really thinking about this differently than any other company I've heard speak

about this. >> Well, the thing about the laws of physics and the grounding, he really talked about the reinforced learning from grounding. So you can... If you don't ground, you can basically go insane, that's where hallucinations come from. So once you get some training going, you start reinforcing it after it's been grounded in truth. So that was interesting. So this whole NVIDIA, you know, when we want to be content for everyone, the AI factory. Tokens generation will be a key for NVIDIA. He said, direct quote, "Blackwell obviously is the center of that design." And Blackwell was designed for the LLM future. NVLink was... Switch was a critical technology to make that work. And obviously NVLink creates the giant CPUs to work together. Manufacturing LLMs at scale. But the AI foundry was something interesting with these. And N-I-Ms, okay? The NVIDIA Inference Microservice. This is a key technology, and he talked about this as a path to the enterprise.

>> Yep. >> He said NIMs are like the API for AI, and that they're going to connect language models. To your point, he says there's two ways that NVIDIA's going to get to the enterprise. One, NIM, create these NIMs and let people run their new AI applications to replace their business applications.

You talked about- >> App devs.

>> Doing it to app developers integrating their own NIMs, okay? The second path was through enterprise AI platforms that are sitting on the gold mine of data and that they're going to use the... They have all their own tools, so they're going to create co-pilots. So the combination of new apps using NIMs, okay? Talking to each other, and then co-pilots leveraging tool and data on the enterprise platforms, IT platforms, to create copilots. So IT departments, IT platforms, and system integrators will be the key different distributions of enabling that market. So, very clever strategy. That's going to be the landgrab in the enterprise. So you're going to see an NVIDIA aggressively going into the enterprise to seed the base, get them hooked on the heroin, you know, first hit. You know, the old drug dealer business model, first hit's free, you know? Get 'em a NIM, you know,

get them NIMed up. >> NIMed up.

>> Get 'em NIMed up,

and then go into the IT- >> Low cost

of entry. >> Department where there's actually value creation. And I think the copilot is a great way to get in there. So, build your apps using NIMs. That's going to be a harder sell than leveraging, in my opinion, the tool data. And he's right. There's a fricking gold mine in the enterprise when it comes to data exhaust, tool data, data sitting around, data has potential. Potential exhaust that could be net new business model opportunities. He's absolutely right on the co-pilots, you'll see agents come out of that. That's still chat bots today, agents tomorrow. But those two ways, pretty good strong strategy, I'm not going to lie-

>> You know. >> I like that strategy.

>> The other thing that we saw yesterday on stage was the robots, right? He was surrounded by all these different robots. And he was asked the question about robots, and where do you see that going? He said, you know, generalized robotics will very, very, very likely... We're going to see that in 10 years. And then he went on, he said, think about, you know, extrapolating from LLMs, we learned to emulate words that we wrote, and we learned how to reason what their meaning is, what the structure of those words are. He goes, "At the end of the day, it's just a bunch of tokens." What if we tokenize movement? So he's saying that's why he's highly confident that we'll have generalized robotics withinside of 10 years. Because essentially he's saying we're going to tokenize movements, that we're going to... We have tons of data on how humans walk, how they sprint. Oh, that's skipping. Oh, that's hugging, that's throwing, And he said there's a finite number of these, the vocabulary is not infinite. And so we are going to actually recreate that and simulate that. And so all these things, he said, can be tokenized and made generative. And he's going to ground Omniverse. Again, I'll say it, is going to allow them to ground this AI with the physical world.

>> And- >> And the laws of physics.

>> And then the question came up, "What about mobile and Edge?

Smaller footprint?" >> Yeah.

>> He said OpenGL for the phone and the PC is pretty much the same thing. The size, similar size for the platforms. They'll have similar, same memory. He then said things like this, that I thought was really cool. "LLMs will spawn SLMs, small language models," highly-specialized SLMs where the fine tuning is different for each, and customized. So you really can't generally, have a general purpose SLM. SLM, small language models, will be highly specialized, therefore it's going to be harder to do fine tuning. He kind of implied that, he didn't say it's harder to do fine tuning. I'm saying that, but I'm kind of implying that because when he says fine tuning's different... He was unsure of how that was going to be solved, but he was sure that SLMs will be around and are important. That again, that validates our power law research that theCUBE Research put out. But he's totally... He doesn't know. We don't yet know how SLMs will operate specifically. His bet, in my opinion, is the SIM, is the NIM integration. The NIM, the NVIDIA Inference Microservice. So he then said, when you do these things, you got to constrain the use cases. So he laid out very eloquently the six areas: fine tuning, this is for LLMs and SLMs. Fine tuning, context windows, RAGs: retrieval, augmentation, generation. Prompting, and agents, okay? That was critical. And so what he's doing there, he is just telegraphing the initial constraints how these apps are going to get built. The language models are going to come out of the gate for RAG, and it's going to come down to large language models, either spawning SLMs, or companies creating their own SLMs through their NIMs. So every company will, they want every company to go to NVIDIA AI Foundry, and create their own NIM, and then create NIMS to integrate with their NIMs. And I think that's where it's going to be very interesting to see. And I think that's a play. If they can pull that off, that is a land grab, that is a net new opportunity for NVIDIA.

>> I think the other thing... Again, I'm fascinated by this sort of future of robots. And he said the process is, we're going to imitate, you know, walking, sprinting, skipping, hugging, throwing, we're going to learn, we're going to adapt. You talk about Isaacson, which I think is Omniverse, and then we're going to generalize and then we're going to do reinforcement learning, and then we're going to ground that in physics. And that's again, what Omniverse allows you to do. So I thought that was really interesting, that sort of sequence. And then he was asked a question about, you know, "Will AI solve climate change?" And he said, "I don't think so because humans," you know, "Aren't going to change their behavior. We're going to keep eating steaks." So our kids are on their own. And then he talked about quantum, I thought this was interesting. I said this earlier when we had Zeus on, NVIDIA's the world's largest quantum computing company that doesn't build quantum computers. He kind of, like you said, threw quantum a little bit under the bus, saying, "Yeah, AI's going to solve a lot of these problems,"

you know, don't- >> He's kind of right. He's kind of right on that.

I'm not- >> He did say-

>> I kind of like what he's saying.

>> It's not going to solve quantum encryption. You need quantum for that. And there's national defense issues where you need quantum, but-

>> Encryption, no-brainer.

>> He said most quantum computing problems, you know, are, you know, basically going to be handled by AI. And AI keeps punting it down the road.

>> Yeah.

>> And so delaying

the adoption of quantum. >> It's doing better. That was my argument about crypto the other day in the podcast, when you asked me about crypto developers versus AI, you kidding me? The act is in AI.

>> He said AI was getting us a long way. And by the way, he was asked about crypto

and he goes, "Meh." >> Yeah.

>> Kind of poo-pooed it.

>> He definitely doesn't want that.

>> You know? We've moved on.

>> He also said, you know, around Blackwell, three years with 25,000 people working together for Blackwell. He also said, "We are not TMSC."

>> TSM. >> I mean, TSMC.

>> Yeah. >> We spent $10 billion. Yesterday, the keynote, he said, "AI foundry is the TSMC for NIMs." They want to be the reference implementation for that. And then he went on to say, "We built a data center. We built data centers, just we build them in parts, we sell them as parts. We build the data center and we sell them in parts."

>> Yeah, what he means by that is they build a data center. So they say, "Okay, we got a data center, it works, it's got all these cables, it's got liquid cooling, all this advanced packaging. Okay, now split it up into a, whatever, 27,000 SKUs, and sell it and get POs for each, however your customers want to buy."

>> Sounds like a mainframe time sharing to me, Dave.

>> Well, except... >> Token sharing.

>> Never unbundled the thermal, you know, conductor module, or whatever that thing was called. 208 billion transistors, he said, I think, he said running at 4 GHz And so pretty amazing. You know, he wants to build the largest chip you can, you know, with the lowest energy possible, and MVLink allowed him to do that.

>> He was also obviously kind of implying through his gestures and his posture, was he had a lot to get information to get in on the keynote, he felt rushed. He kind of made a couple comments that way, but he said he deleted a few slides. One of them was, I found fascinating. I'll share here, because it's really important for our industry to know, that he had a telecom slide in there, where they were going to use AI to reduce the 5G energy levels while keeping the transmission speeds up. Essentially managing spectrum and radio frequency, which is not really done with the MIMO, which is a technique for managing antenna, you know, antenna action. So with RTX and Omniverse, they can do digital twins. And then he talked about initiative that I wasn't aware of, but I'm going to be fascinated to research more, the 6G research cloud. He's excited about it. But basically this could make our wireless world so much better, because the output of spectrum in these radios, AI can come in and reduce the energy level significantly. So you're starting to see NVIDIA, their supercomputer mindset, and the super pods, and the super cloud, and all the super data levels. They could come in and change climate change, change biology healthcare, drug research, these verticals that require a lot of HPC, but like just... But they're not really HPC candidates. I mean, HPC has some workloads that are just like, obvious old school HPC use cases. But as compute is in more demand for high-performance gear iron, big iron, you know, they're going to need some horsepower and NVIDIA can bring it to the table.

>> Yep. And he did say on the 6G, said, "It will be software-defined and it's going to be AI-powered." But then he, you know, talking about the moat, people want to know what's their moat, what's it look like? He said, "Look, yeah, we build GPUs, we build NICs. We had to build two switches. We have 600,000 parts. We got to manage 600,000 parts." Imagine the vendor management there. "Then we write the software, we got to make it perform, then we got to get apps to run at it. So we have to go train people how to run the apps. So we got to create demand, and when we build this," like you said, "data center, where we're done, we disaggregate it and sell it in parts. We are market makers, not share takers." So it was really a masterclass in how this guy thinks.

>> I found it was a great interaction. I'm glad we attended. And the keynote... Watching the keynote, then having that one-on-one, the industry stands was phenomenal. This is a systems revolution we've been talking about in theCUBE. If you've been watching theCUBE, and listening to our... And watching our podcast, you'll know that we've been talking about this. And this is just the beginning. I heard a quote earlier today in the floor here. It's not even the first thing, it's the first pitch, okay? And it's right down the middle. It's a strike. It's a good one. So, great stuff. That's our review of NVIDIA's GTC24, Jensen keynote analysis. Deep dive. Check out my Quick Take yesterday with Zeus Kerravala, and then the analyst angle we just did earlier today. Zeus, Dave, and I, this is a kind of an emergency pod, Dave. Great to dig into it.

>> Yeah. >> There's so much more. We'll keep pounding this content all day long. It's going to keep coming back. This is the systems revolution and it's happening here in theCUBE, and with NVIDIA. (soft pulsating music)

John Furrier & Dave Vellante, Mini Pod

search

(soft pulsating music)

>> Hello, I'm John Furrier. Dave Vellante. The Cube is here in San Jose for NVIDIA's GTC24. We've been here for two days. Wall-to-wall coverage.

Dave, great to see you. >> Hey John.

>> I know we've been popping around. We just had the chance to get with Jensen, hear him speak privately to all the analysts who were there, as well as a few other industry analysts. We have Broadcom tomorrow. Big time meetings with those guys. Broadcom, NVIDIA, the big one. But this is kind of an abbreviated CUBE interview, John and Dave, emergency pod analysis. Because I really want to dig dive deep and unpack the keynote yesterday. And then the private comments that we heard from Jensen Wong, CEO of NVIDIA, who's an awesome legend. I just love the guy. He's got a great personality. He's kind of snarky in a good way. He's clever. He is certainly got a computer science degree, and obviously he's... Impeccable degree. And he's the story of him being, you know, in college when he was like 15 or 12 years old. Forgot the number. But obviously really strong. But just worked his way through the system in Silicon Valley. It's a great American story, I love it. One of my favorites.

>> He's got a great sense of humor too, doesn't he? Very disarming.

>> He's just a cool dude, and he's common sense, and he's an engineer and computer scientist. And so he just speaks in that kind of, like, parlance of like, "Yeah, let's solve some problems." But he really has got a great future. Let's unpack this. You know, my first takeaway, obviously the Blackwell super chip, it's the successor to Hopper. Which, it's not going away. It becomes more of a specific thing, NVLink Switch, DGX system. But I think the big thing that jumped out at me right out of the gate was his whole industry revolution. And why I like this is because we've been saying on theCUBE, you know, I've been saying we're going to have another revolution like the '60s, you know, revolution, you know, free love, summer of love, AI love, you know? This generational shift with AI. And I think he points to this as the new industrial revolution, meaning AI will, and its systems, of AI, will power new sets of experience. And specifically he went on to talk about how the world that we live in is prerecorded. It kind of reminded me when I used to bust your child spot, having the Wall Street Journal on the podcast, and say, "That was yesterday's news," 'cause it was programmed for you. Pre-recorded, pre-written down. It's not generative. So, like, our whole world is pre-recorded, if you think about it. So he's kind of right on that. And then it's moving to a generative world. And retrieval, generative AI, generative content. Obviously they've done graphics, seeds of innovation, that's the role of data. But he honed in on tokens, and that jumped out at me because I think the whole embedding experience, we're seeing vector databases. The new way to do the math in AI is going to create new experiences. And the generative part of generating stuff with AI requires tokens. And tokens will be the new currency of AI. And I think you're going to start to see another step function change with token economics, token value. And I think the whole business model question of how they price the software is going to be pale in comparison to the monetization opportunities with tokens. So this whole new industrial revolution's coming, we call it a systems revolution with clustered systems, the intelligent data platform, six data platform you've been doing. So, you know, theCUBE Research has got their fingers on the pulse here, but Jensen's on point, Dave, he's on point and there's so much more. Let's get into it. So, share your thoughts.

>> I was writing furiously because he's so compelling when he speaks. And you're right, he talked about the industrial evolution that we're entering a new era, and he used the analogy of electricity. He said, "An entire industry is going to be producing tokens." And that's what we're doing. The world of generative content is here. And he said the computer sees tokens. It doesn't know the difference between tokens, animations, whether the tokens are for image, whether it's a robotic arm. And his point was, if you can generalize, then you can talk to it, you know, with natural language. So he's basically saying that the entire industry is going to be about tokens, and he's saying none of this existed prior to ChatGPT, you know? And so he said, "We at NVIDIA generate a ton of content and Blackwell is built for this Gen AI moment."

>> It's interesting.

>> And then he went through, you know, a litany of sequences that I want to dig into a little bit with you.

>> Let's dig into it because I think this token model... He also said something like, "Tokens don't know about language," but yet we're seeing language. They didn't know if it's an arm movement. So what he was saying was, is that what we know about the prerecorded world is completely different with generative world, the generative world we're living in, where things generate on your behalf, prompt response, or reasoning, or, you know, running a digital twin to simulate a factory. So these new things are here, and their accelerated computing platform will be powering that. And you know... And I remember when Sergey and Larry launched Chrome, I said, "That's an operating system." 'Cause they, you know, the Android was yet to be launched. What Jensen's doing here, in my opinion, is setting the table for a new AI infrastructure and an operating system, Dave.

>> That's built for tokens, and-

>> Token operative systems. >> And generative AI. And it's like... So when other people talk... I mean, look, Gelsinger, as you know, is amazing. He's a legend. He'll talk in terms of how AI is going to affect society and the amazing things that we're going to be able to do with it. And Jensen talks about that as well, but he goes into this layer, which I wasn't expecting, about this world of tokens, and he said, "We're going to have models talking to each other." Why would we do that? Remember this comment that he made? He goes, "Well, think about it. We go into a room, we debate, we argue, we dissect. We're going to have models do that.

They're going to learn- >> And they're

going to come out of- >> From each other.

>> Smarter."

>> And they'll come out smarter.

>> Not to pat ourselves on the back, but we've been saying model integration is going to be a big part of that. He validates that. So, you know... So again, what we've been conjecturing on theCUBE, now becomes validated as fact with the industry, with Jensen essentially saying model integration, model synergies. Whether they build them or whether they enable someone else to build their own LLMs. He said, "We're going to be in the business of manufacturing LLMs."

>> Yeah. I mean, this is the LLM economy. And I was like, "Okay, interesting." And the token economy, so... And then he talked about, you know, simulation. Yesterday, he said, "This is not animation that you're going to see. Everything you see today is going to be simulation." And the simulation was amazing. And then he talked about reinforcement learning, and that learning loop, and training. He said, "You do it enough, it becomes reality." And then he talked about, you know, the critical tech being NVLink. They had to create this in order to allow all these different parts, these GPUs, these components, these chips, these 35 or 50,000 components, to talk to each other and share that same memory. And that is-

>> Which by the way, no one else yet has, and they're on Gen5.

>> Yeah. Nobody has that. >> Fifth generation. This on the fifth generation.

>> This is their moat.

>> He made a point to say "It's our fifth generation and nobody's even in the market."

>> Now, the interesting thing is, and we're going to hear this, I'm sure, tomorrow, at Broadcom's AI Day, Broadcom has a completely different philosophy. They're like, "Look, they're proprietary. That's cool. That's how Charlie Kawwas talked about this. That's how markets get started. And he's right about that. We're going after open, you know, we're going after scale, we're going after lower power, and we're going to do this with a different approach. We're going to use this chiplet approach." And I think there's a massive market for that. But it's a different philosophy. Now, by the way, we heard today, NVIDIA's also doing chiplets. So the difference is NVIDIA can play in that big monolithic marketplace. The biggest, most-

>> He said, "Don't call us chiplets."

>> No, no. He said... He sad, "We're not chiplets." But we heard that they're doing chiplets for media tech.

>> Got it, okay. >> Which was new information to me. I got to dig into that. But they're a systems company. They're a systems and software company, right? And yeah, they make chips, but they take those chips and they build systems. I guess the point I'm making is, and this is really where Floyer had sort of educated me, is those chiplets are very specialized, they give you a lot of flexibility. But they're connecting through relatively slower than what NVIDIA has, a communications network, and it's asynchronous. And so these big monolithic chips have a place. Who was building monolithic chips? NVIDIA. Guess who else is? Apple.

>> Yeah. >> Apple has a big shared SRAM chip.

Okay. >> He said the one thing they wanted to do is build the largest chip possible for all, that they put inside them. You know, it's not a-

>> They're betting on big.

>> It's not a chiplet, it's a monolith.

>> They're betting on big. Right. But they do, it was news to me. But they're doing chiplets evidently. And Ben Bajarin told me this on a breaking analysis, said, "There's a rumor that they're going to be doing, you know, their own chiplet design," fine. 'Cause it's probably a different market. I got to dig into it and learn more about it. But essentially he's saying, "We're building this AI foundry, and AI," he said, "Needs to understand the laws of physics." He said, "It doesn't really." He goes, "Sora kind of does." You know, it knows how snow behaves, when you're driving on gravel. You're driving, you know, above the gravel, but the gravel moves. So it's starting to train these models about the laws of physics. And then he basically said Omniverse will ground AI to the laws of physics. And so they're really thinking about this differently than any other company I've heard speak

about this. >> Well, the thing about the laws of physics and the grounding, he really talked about the reinforced learning from grounding. So you can... If you don't ground, you can basically go insane, that's where hallucinations come from. So once you get some training going, you start reinforcing it after it's been grounded in truth. So that was interesting. So this whole NVIDIA, you know, when we want to be content for everyone, the AI factory. Tokens generation will be a key for NVIDIA. He said, direct quote, "Blackwell obviously is the center of that design." And Blackwell was designed for the LLM future. NVLink was... Switch was a critical technology to make that work. And obviously NVLink creates the giant CPUs to work together. Manufacturing LLMs at scale. But the AI foundry was something interesting with these. And N-I-Ms, okay? The NVIDIA Inference Microservice. This is a key technology, and he talked about this as a path to the enterprise.

>> Yep. >> He said NIMs are like the API for AI, and that they're going to connect language models. To your point, he says there's two ways that NVIDIA's going to get to the enterprise. One, NIM, create these NIMs and let people run their new AI applications to replace their business applications.

You talked about- >> App devs.

>> Doing it to app developers integrating their own NIMs, okay? The second path was through enterprise AI platforms that are sitting on the gold mine of data and that they're going to use the... They have all their own tools, so they're going to create co-pilots. So the combination of new apps using NIMs, okay? Talking to each other, and then co-pilots leveraging tool and data on the enterprise platforms, IT platforms, to create copilots. So IT departments, IT platforms, and system integrators will be the key different distributions of enabling that market. So, very clever strategy. That's going to be the landgrab in the enterprise. So you're going to see an NVIDIA aggressively going into the enterprise to seed the base, get them hooked on the heroin, you know, first hit. You know, the old drug dealer business model, first hit's free, you know? Get 'em a NIM, you know,

get them NIMed up. >> NIMed up.

>> Get 'em NIMed up,

and then go into the IT- >> Low cost

of entry. >> Department where there's actually value creation. And I think the copilot is a great way to get in there. So, build your apps using NIMs. That's going to be a harder sell than leveraging, in my opinion, the tool data. And he's right. There's a fricking gold mine in the enterprise when it comes to data exhaust, tool data, data sitting around, data has potential. Potential exhaust that could be net new business model opportunities. He's absolutely right on the co-pilots, you'll see agents come out of that. That's still chat bots today, agents tomorrow. But those two ways, pretty good strong strategy, I'm not going to lie-

>> You know. >> I like that strategy.

>> The other thing that we saw yesterday on stage was the robots, right? He was surrounded by all these different robots. And he was asked the question about robots, and where do you see that going? He said, you know, generalized robotics will very, very, very likely... We're going to see that in 10 years. And then he went on, he said, think about, you know, extrapolating from LLMs, we learned to emulate words that we wrote, and we learned how to reason what their meaning is, what the structure of those words are. He goes, "At the end of the day, it's just a bunch of tokens." What if we tokenize movement? So he's saying that's why he's highly confident that we'll have generalized robotics withinside of 10 years. Because essentially he's saying we're going to tokenize movements, that we're going to... We have tons of data on how humans walk, how they sprint. Oh, that's skipping. Oh, that's hugging, that's throwing, And he said there's a finite number of these, the vocabulary is not infinite. And so we are going to actually recreate that and simulate that. And so all these things, he said, can be tokenized and made generative. And he's going to ground Omniverse. Again, I'll say it, is going to allow them to ground this AI with the physical world.

>> And- >> And the laws of physics.

>> And then the question came up, "What about mobile and Edge?

Smaller footprint?" >> Yeah.

>> He said OpenGL for the phone and the PC is pretty much the same thing. The size, similar size for the platforms. They'll have similar, same memory. He then said things like this, that I thought was really cool. "LLMs will spawn SLMs, small language models," highly-specialized SLMs where the fine tuning is different for each, and customized. So you really can't generally, have a general purpose SLM. SLM, small language models, will be highly specialized, therefore it's going to be harder to do fine tuning. He kind of implied that, he didn't say it's harder to do fine tuning. I'm saying that, but I'm kind of implying that because when he says fine tuning's different... He was unsure of how that was going to be solved, but he was sure that SLMs will be around and are important. That again, that validates our power law research that theCUBE Research put out. But he's totally... He doesn't know. We don't yet know how SLMs will operate specifically. His bet, in my opinion, is the SIM, is the NIM integration. The NIM, the NVIDIA Inference Microservice. So he then said, when you do these things, you got to constrain the use cases. So he laid out very eloquently the six areas: fine tuning, this is for LLMs and SLMs. Fine tuning, context windows, RAGs: retrieval, augmentation, generation. Prompting, and agents, okay? That was critical. And so what he's doing there, he is just telegraphing the initial constraints how these apps are going to get built. The language models are going to come out of the gate for RAG, and it's going to come down to large language models, either spawning SLMs, or companies creating their own SLMs through their NIMs. So every company will, they want every company to go to NVIDIA AI Foundry, and create their own NIM, and then create NIMS to integrate with their NIMs. And I think that's where it's going to be very interesting to see. And I think that's a play. If they can pull that off, that is a land grab, that is a net new opportunity for NVIDIA.

>> I think the other thing... Again, I'm fascinated by this sort of future of robots. And he said the process is, we're going to imitate, you know, walking, sprinting, skipping, hugging, throwing, we're going to learn, we're going to adapt. You talk about Isaacson, which I think is Omniverse, and then we're going to generalize and then we're going to do reinforcement learning, and then we're going to ground that in physics. And that's again, what Omniverse allows you to do. So I thought that was really interesting, that sort of sequence. And then he was asked a question about, you know, "Will AI solve climate change?" And he said, "I don't think so because humans," you know, "Aren't going to change their behavior. We're going to keep eating steaks." So our kids are on their own. And then he talked about quantum, I thought this was interesting. I said this earlier when we had Zeus on, NVIDIA's the world's largest quantum computing company that doesn't build quantum computers. He kind of, like you said, threw quantum a little bit under the bus, saying, "Yeah, AI's going to solve a lot of these problems,"

you know, don't- >> He's kind of right. He's kind of right on that.

I'm not- >> He did say-

>> I kind of like what he's saying.

>> It's not going to solve quantum encryption. You need quantum for that. And there's national defense issues where you need quantum, but-

>> Encryption, no-brainer.

>> He said most quantum computing problems, you know, are, you know, basically going to be handled by AI. And AI keeps punting it down the road.

>> Yeah.

>> And so delaying

the adoption of quantum. >> It's doing better. That was my argument about crypto the other day in the podcast, when you asked me about crypto developers versus AI, you kidding me? The act is in AI.

>> He said AI was getting us a long way. And by the way, he was asked about crypto

and he goes, "Meh." >> Yeah.

>> Kind of poo-pooed it.

>> He definitely doesn't want that.

>> You know? We've moved on.

>> He also said, you know, around Blackwell, three years with 25,000 people working together for Blackwell. He also said, "We are not TMSC."

>> TSM. >> I mean, TSMC.

>> Yeah. >> We spent $10 billion. Yesterday, the keynote, he said, "AI foundry is the TSMC for NIMs." They want to be the reference implementation for that. And then he went on to say, "We built a data center. We built data centers, just we build them in parts, we sell them as parts. We build the data center and we sell them in parts."

>> Yeah, what he means by that is they build a data center. So they say, "Okay, we got a data center, it works, it's got all these cables, it's got liquid cooling, all this advanced packaging. Okay, now split it up into a, whatever, 27,000 SKUs, and sell it and get POs for each, however your customers want to buy."

>> Sounds like a mainframe time sharing to me, Dave.

>> Well, except... >> Token sharing.

>> Never unbundled the thermal, you know, conductor module, or whatever that thing was called. 208 billion transistors, he said, I think, he said running at 4 GHz And so pretty amazing. You know, he wants to build the largest chip you can, you know, with the lowest energy possible, and MVLink allowed him to do that.

>> He was also obviously kind of implying through his gestures and his posture, was he had a lot to get information to get in on the keynote, he felt rushed. He kind of made a couple comments that way, but he said he deleted a few slides. One of them was, I found fascinating. I'll share here, because it's really important for our industry to know, that he had a telecom slide in there, where they were going to use AI to reduce the 5G energy levels while keeping the transmission speeds up. Essentially managing spectrum and radio frequency, which is not really done with the MIMO, which is a technique for managing antenna, you know, antenna action. So with RTX and Omniverse, they can do digital twins. And then he talked about initiative that I wasn't aware of, but I'm going to be fascinated to research more, the 6G research cloud. He's excited about it. But basically this could make our wireless world so much better, because the output of spectrum in these radios, AI can come in and reduce the energy level significantly. So you're starting to see NVIDIA, their supercomputer mindset, and the super pods, and the super cloud, and all the super data levels. They could come in and change climate change, change biology healthcare, drug research, these verticals that require a lot of HPC, but like just... But they're not really HPC candidates. I mean, HPC has some workloads that are just like, obvious old school HPC use cases. But as compute is in more demand for high-performance gear iron, big iron, you know, they're going to need some horsepower and NVIDIA can bring it to the table.

>> Yep. And he did say on the 6G, said, "It will be software-defined and it's going to be AI-powered." But then he, you know, talking about the moat, people want to know what's their moat, what's it look like? He said, "Look, yeah, we build GPUs, we build NICs. We had to build two switches. We have 600,000 parts. We got to manage 600,000 parts." Imagine the vendor management there. "Then we write the software, we got to make it perform, then we got to get apps to run at it. So we have to go train people how to run the apps. So we got to create demand, and when we build this," like you said, "data center, where we're done, we disaggregate it and sell it in parts. We are market makers, not share takers." So it was really a masterclass in how this guy thinks.

>> I found it was a great interaction. I'm glad we attended. And the keynote... Watching the keynote, then having that one-on-one, the industry stands was phenomenal. This is a systems revolution we've been talking about in theCUBE. If you've been watching theCUBE, and listening to our... And watching our podcast, you'll know that we've been talking about this. And this is just the beginning. I heard a quote earlier today in the floor here. It's not even the first thing, it's the first pitch, okay? And it's right down the middle. It's a strike. It's a good one. So, great stuff. That's our review of NVIDIA's GTC24, Jensen keynote analysis. Deep dive. Check out my Quick Take yesterday with Zeus Kerravala, and then the analyst angle we just did earlier today. Zeus, Dave, and I, this is a kind of an emergency pod, Dave. Great to dig into it.

>> Yeah. >> There's so much more. We'll keep pounding this content all day long. It's going to keep coming back. This is the systems revolution and it's happening here in theCUBE, and with NVIDIA. (soft pulsating music)