 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. We were there, as well as a few other industry analysts. We had Broadcom tomorrow, big time meetings with those guys. Broadcom NVIDIA is a big one, but this is kind of an abbreviated Cube interview, John and Dave, emergency pod analysis, because I really want to dive deep and unpack the keynote yesterday, and then the private comments that we heard from Jensen Wong, CEO of NVIDIA, who's 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's certainly got a computer science degree, and obviously he's impeccable degree, and the story of him being in college when he was like 15 or 12 years old. I forget 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? Phenomenal. 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 parlance of let's solve some problems, but he really has got a great future. Let's unpack this. My first takeaway, obviously, the Blackwell Super Chip, it's the successor to Hopper, which is not going away. It's become more of a specific thing. NVLink switch, DGX system, but I think the big thing that jumped out of 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 revolution, free love, summer of love, AI love, this generational shift with AI, and I think he points to this as the new industrial revolution, meaning AI 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. Kind of reminded me when I used to bust your chops about having the Wall Street Journal on the podcast and say that was yesterday's news, because it was programmed for you, prerecorded, pre-written down, it's not generative, so our whole world is prerecorded, if you think about it. I've got a 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 of me because I think the whole embedding experience we're seeing, vector databases, the new way to do the math and 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 the CUBE 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 what's your, share your thoughts. I was writing furiously because he's so compelling when he speaks, and you're right, he talked about the industrial revolution 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 generated 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, 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 the GPT, you know, and so he said we at NVIDIA generate a ton of content and Blackwell's built for this GNAI moment. It's interesting. And then he went through 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 pre-recorded world is completely different with generative. The generative world we're living in, where things generate on your behalf, prompt response, or reasoning, or running a digital twin to simulate a factory. So these new things are here, and their accelerated computing platform will be powering that. And everyone's Sergey and Larry launched Chrome. I said, that's an operating system, because 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 that's built for tokens and generative AI. So when other people talk, I mean, look at Gelsinger, as you know, he's amazing, he's a legend, but he'll talk in terms of how it's how it's 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 from each other. And they come out of that room smarter. And they come out smarter. Not to pat ourselves in the back, but we've been saying model integration is going to be a big part of that. He validates that. 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. I mean, this is the LLM economy. And I was like, okay, interesting. And the token economy. So, and then he talked about 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, if you do it enough, it becomes reality. And he talked about 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- By the way, no one else yet has on there on gen five. Yeah, nobody has. Fifth generation. This is their mode. 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 Cowell has 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 is also doing chiplets. So the difference is NVIDIA can play in that big monolithic marketplace. The biggest, most honkinest GPUs. He said, don't call us chiplets. No, no. He said, we're not chiplets. But we heard that they're doing chiplets for media tech, which was new information to me. I got to dig into that. But they're a systems company. They're 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. Apple has a big shared SRAM chip. He said the one thing they wanted to do is build the largest chip possible for all the benefits 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 Baharan told me this on a breaking analysis. He said, there's a rumor that they're gonna be doing their own chiplet design. Fine, because 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, doesn't really. He goes, Sora kinda does. You know, it knows how snow behaves when you're driving on gravel. You're driving 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, you 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, when we want to be content for everyone in the AI factory, tokens generation will be a key for NVIDIA, he said. Direct quote. Blackwell, obviously the center of that. 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 that's with these NINIMS, okay? The NVIDIA Inference Microservice. This is a key technology. And he talked about this as a path to the enterprise. 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 is going to get to the enterprise. One, create these NIMS and let people run their new AI applications to replace their business applications. App devs doing it. App developers integrating their own NIMS, okay? The second path was through enterprise AI platforms that are sitting on the goldmine of data and that they're going to use, 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 co-pilots. So IT departments, IT platforms and system integrators where they'll be the key different distributions of enabling that market. So very clever strategy. That's going to be the land grab and the enterprise. So you're going to see NVIDIA aggressively going into the enterprise to seed the base, get them hooked on the heroin. First hits the old drug dealer business model. First hits free. Get them a NIM, get them NIMed up. NIMed up. Get them NIMed up. And then go into the IT departments where there's actually value creation. And I think the co-pilot is a great way to get in there. So builds 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 goldmine in the enterprise when it comes to data exhaust, tool data, data sitting around, data as 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 chatbots today, agents tomorrow. But those two ways, pretty good, strong strategy. I'm not going to lie, 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 a 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 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 be have generalized robotics with inside 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. 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, Omdiverse, again, I'll say it's going to allow them to ground this AI with the physical world and the laws of physics. And then the question came up, what about mobile and edge, smaller footprint? He said, open GL for the phone and the PC is pretty much the same thing. The size, similar size for the platforms. They all 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. SLMs, 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 is 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 the key 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 NIM integration. The NIM, the NVIDIA inference microservice. So he then said, when you do these things, you've 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. That was critical. And so what he's doing there, he's 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 the future of robots. And he said, the process is we're going to imitate walking, sprinting, skipping, hugging, throwing. We're going to learn, we're going to adapt. He talked 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 the question about, will AI solve climate change? And he said, I don't think so, because humans aren't going to change the 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 Xeas on, NVIDIA is 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, AI is going to solve a lot of these problems. He's kind of right. He's had a right on that. I'm not, I'm kind of like what he's saying. He's not going to solve quantum encryption. You need quantum for that. And there's national defense issues that 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. And so delaying the adoption of quantum. Doing better. That was my argument about crypto the other day in the podcast. When you asked me about crypto developers versus AI, are you kidding me? The actor's in AI. He said AI was getting us a long way. And by the way, he was asked about crypto and he goes, man, kind of poo pooed it. He definitely doesn't want that. We've moved on. He said, 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. We spent $10 billion. Yesterday at 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. Let's sell it and get POs for each. However, your customers want to buy. Sounds like a mainframe time sharing to me, Dave. Well, except. Except IBM never unbundled the thermal conductor module or whatever that thing was called. 208 billion transistors. He said, I think he said running at four gigahertz. And so pretty amazing. You know, he wants to build the largest chip you can. You know, with the lowest energy possible and that was NVLink allowed him to do that. He was also obviously kind of implying through his gestures and his posture was he wanted, he had a lot to get information to get in on the keynote. He felt rushed, kind of made a couple of comments that way. But he said he deleted a few slides. One of them was I found fascinating. I'll share. 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 action. So with RTX and Omniverse, they can do digital twins. And then he talked about an 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 and these radios, AI can come in and reduce the energy level significantly. So you're starting to see Nvidia with 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 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 more demand for high performance gear, big iron, they're going to need some horsepower and Nvidia can bring it to the table. Yeah, and he did say on the 6G, he said it will be software defined and it's going to be AI powered. But then he's talking about the mode, people want to know what's their mode, 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, manage 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 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 understanding was phenomenal. This is a systems revolution we've been talking about on 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 on the floor here. It's not even the first thing. It's the first pitch. Okay. And it's right down the middle of the strike. It's a good one. So great stuff. That's our review. Nvidia's GTC24 Jensen keynote analysis deep dive. Check out my quick take yesterday with Zia's Carvella. And then the analyst angle we just did earlier today, Zia's Dave and I. This is kind of an emergency pod, Dave. Great to dig into it. There's so much more. We can keep having this content all day long. It's going to keep coming back. This is a systems revolution and it's happening here in theCUBE with Nvidia.