 Welcome back everyone to theCUBE's coverage of NVIDIA's GTC, the conference for the era of AI. I'm John Furrier with Dave Vellante. We're here for three days getting all the action with our run and gun pop up cube. We'll do whatever it takes to get that story where there's dropping in one camera, three cameras, full stage, whatever it takes theCUBE is there. We've got you covered and we have here our CUBE contributor, CUBE collective member, Sarbjeet Joel, analyst, friend of theCUBE, out there doing the analyst notebook. I call it because he doesn't have a notebook. He's got an iPhone and he's got probably a notebook but Sarbjeet, great to see you. One of the things I love about how you attack events is you really kind of go out the keynote. By the way, your keynote tweet when viral, just getting that data out there really fast, really does the community great service. I want to say thanks for that. And number two is you always like to go and check out all the sessions, check out the hallway conversation. So I love, I call it the analyst notebook section because you're kind of notebook and it's a good summary. So thanks for coming on. Thanks John. Actually, I saw your coverage from yesterday with Zeus and you and David. I think that segment summed up the first two days after you guys had a session with Jensen Q&A, right? So that was awesome. Yep. I'm happy to- So let's take, so you had a bunch of dinners that you had a bunch of people that they're brainstorming those conversations like what is the conversations in the hallways what's the conversations at the events and the parties that you're seeing. Also we saw the keynote, Jensen was well documented. It's a new era. It's not a pre-recorded era. It's a generative era. Everything's changed. How software's built changed. The computer has changed. It's now a system. You got the AI Foundry, which is a bunch of NIMS, which is an API to NVIDIA. And you got Neo Retriever, which is such a rag which is retrieval augmentation generation. And then you got robotics and the thing you didn't talk about is the telecom side. Again, tell us what you're hearing in the hallways. Hallways discussions are like, my sort of interpretation is that things don't change as fast as what Jensen wants to change. So it's a new paradigm shift, right? So going from CP to GPU, the way we program our systems, especially the systems of record, we're not going to change that to be programmed by GNAI, for example, right? Because there's a lot of hallucination problems and like we're trying to figure that stuff out if you will, right? Over time, these things will improve. Of course there's code generation in place. And I'm a coder at heart and I've coded many systems and I have played with the GNAI code generation. It's not that good. It's just rudimentary, you know, crud, you know? It's just, it does the basic stuff, right? It does the coding, but it doesn't do programming, what I call it. So, long story short, this is just the beginning. It will take a while for AI to give the fruit what we are promising through these narratives. So I think it's going to be hurry up and wait, but definitely they are killing it with their, you know, compute speed and having the memory sitting next to the computer and the multi-terabyte, you know. I want to get your, first of all, great analysis. I want to get your take on the show itself because, you know, you got this as a stake in the ground as an inflection point. You got people here that shouldn't be here, right? I don't say that they shouldn't be here. They shouldn't be here in the sense of in the classical GDC fashion. It's a fricking developers conference. Yet you got hedge fund guys, you got VCs, you can start up popping up. A lot of money to be made. This is clearly the next wave, so a lot of people are here. So this represents kind of like that cultural moment in time. But you got to look at this and say, is it an enterprise show or is it a consumer show? Because AI today, if you look at all the hyperscale investments on the AI side, it's essentially an investment in the consumer hyperscalers. Meta, for instance. ByteDance, Tencent. You name all those companies. Those are consumer-oriented social media networking companies. Yet you got a bunch of hosters coming in here with DGS Cloud. Where's the demand and where's the supply going to go? Because remember, one of the things that's coming out and the idea is that there's still a huge demand for the product. Okay, software, no one's got a chance against CUDA. Clearly, they got an edge there. But NVIDIA doesn't sell software. They sell GPUs. They're a chip company. No, but they want to sell software now. They are going to sell software. They're selling a license for CUDA. For CUDA. For GPU. For GPU per year. So they are going to make money from that. They're selling GPUs. So actually they are- Are they selling GPUs or are they selling a license to the software on the GPU? They are trying to do both. Actually that messaging is confusing in many ways because it also collides with their customers. It collides with what AWS is doing. It collides with what Google Cloud is doing. They are trying to give that as a service to the builders, if you will, right? They are the platforms. They are past platform as a service. It happens, historically speaking, it happens to many companies when they are a technology provider and they want to be a service provider, they get lost in that transition. Most of the companies get lost and they don't. Explain that. So in what way that they're too focused on the platform, not on the ecosystem, or where they get lost? No, okay. Technology provider versus service provider. So when you are providing the technology, you are giving it to every service provider so they can build stuff and give it to the builders of these applications, right? That's the technology provider. You are working with many service providers. You're enabling people to do stuff. Exactly. So if you try to be the service provider, now you are competing with other service providers and then it muddies the waters. VMware wanted to do their own public cloud and they scripted that, sort of canceled that, right? That was the main reason because then nobody will pick up VMware stock to put in. It's the classic strategy problem. You're stuck in the middle between two strategies. Are you saying that Nvidia is trying to be a service provider or are they still leaning to technology provider or are they telegraphing a little bit of a service provider? I think they are telegraphing some, they have inclination to be service provider in some sense. And I hope I am wrong, reading that wrong because if they are trying to do that, they're going to stumble. That's what my- Jensen said on his speech, we are building a data center, just selling it part by part. And that's what the GPUs, they do in the allocations. So they are providing a service. I mean, they're clearly the DGX cloud is a service or is that technology? So it's again, the lines are blurring. It's a great point. And again, it's a critical question for Nvidia because I was saying earlier on the show floor when I was wandering around and talking to some influencers and some executives, they asked me and I said, hey, you can't have an ecosystem play with AI Foundry if you're going to be a closed garden. So if you're a closed garden, ecosystems don't thrive because the control points to closed garden, meaning I'm a monolithic system and I'm going to basically sell you as a service and charge on the software license per GPU, which is basically a GPU license. So, you know, that's going to be a tough call. Will AI Foundry, will that flower up and grow? Or is it a red herring? Is it a stalking horse? But on the other hand, playing Davos advocate here, like John Ford does on CNBC, on the other hand, he has a segment, right? The two hand- On the other hand. On one hand, on the other hand. On the other hand, what Jensen and Company, what they have done is they have put a lot of energy behind all these researchers in different domains to use the computing to describe their industries digitally. So the whole idea of digital twin is very revolutionary in my point of view because you can create these products using digital twins before even you muck around with the physicality of these systems. So that's huge actually. It saves a lot of time and money and agony, you know? So mRNA vaccine is one example, you know? Like building cars. There are cars on the floor here of this show, right? There's auto industry heavily represented here and there's telco. There's like financials, all kind of industries are like. Okay, so let me ask you a question. So as you walk the show floor here, which by the way, it could be bigger, it is pretty tight, small space, this place is kind of bustling out of the seams here. If you had to look forward next year, okay, we heard this question earlier, what's gonna be different next year at GTC? What's gonna change in the next 12 months in the industry? What's gonna the show floor gonna look like in your opinion? If you had to throw a dart at the board, what would you say? I think this show will be bigger, definitely. There will be more use cases. There will be some more like a practicality attached to the use cases. Right now it's all like experimentation. So there might be fewer use cases which are in sort of pre-production stage, if you will. So that's what I'm expecting. What's your note in your notebook that you've been collecting all the observations here in the hallway sessions? What's the big takeaway this year from your perspective as you compile all the data? What's it telling you? I think the main thing is that you can't avoid NVIDIA and their stack if you are serious about gen AI and AI in general and gen AI in particular. So that message is loud and clear. Other people are trying to catch up, AMD's of the world and Intel's of the world, right? So the programmability of these chips is super important. The one thing actually I tweeted a few days back I think maybe after the keynote that we need standards to lower the cost of doing AI and also not only the cost but having lower power consumption. So we need specialized chips so we can use lower power because the whole idea is that when you have a computer if it is idle like 70% of your GPUs or CPUs are idle you're wasting a lot of energy and energy is... Power is a new envelope. That's the new constraint that design around. I think one of the things I think is going to be different and this is interesting that you brought that up, power is I think you're going to see a lot more competition. I think NVIDIA has raised the bar. It's a high bar on the gen AI systems approach. My big takeaway from this year is two things. One, the accelerated computing is a great narrative. I think Jensen hit a home run on that piece. I think this idea of a monolithic system that's not like a mainframe but it's like it's more distributed computing is going to pull forward a bunch of use cases for AI to be accelerated faster. So when you'll see more successes on use cases that are prime for what they're building with their hardware and their performance. The second thing that jumps out at me is I think this sets the standard now for what we've been calling clustered systems. A way has now been shown how to organize servers and multiple configurations. You mentioned power. We were just talking to the Broadcom executives this morning about power, ethernet, that whole spine thing. Infiniband is way more expensive on a port basis than you say ethernet, wide open, open standard and connects with PCIe is just as good of a switch in those configurations and with Broadcom's retimer. You're not going to go distance within the rack and ethernet lowers the power envelope in the rack and lowers your cooling requirements. So what you're seeing is a complete reconstruction of what a system is. So I think this idea of what the rack used to be. Member stack and rack those blades, throw the top of rack switch up there and connect the next rack. That's what the IT was for what, two decades? Yeah, more than that. I think now we're on an IT platform that's going to be clustered systems, my word. Maybe it's a better word, AI systems, that are going to have to work to support multiple workloads in a distributed computing paradigm. That's public cloud on premise edge and the software will have to be compatible to run in that super configuration. It's like a super cloud configuration and you're going to see the rise of NVIDIA clearly leading the way, but I think competition will come in and that's going to come down to who can design the best low power, high speed networking interconnects for these new systems and AMD's not going to lie down. All these other chips aren't going to lie down and let NVIDIA run to the end zone. We need the vendor diversity. That's the need of our actually. We have gone through this, like we want competition. We want competition. Two more observations. One is that can you do AI in colors, right? Or on-prem versus doing that in the cloud? So that was the discussion. And the answer observation is yes or no? The answer is yes, yeah. You can do it both places, okay? You can do it over both places, but I think the cloud guys are saying, no, no, no, the colos don't have enough power, what AI needs to give you the performance and do AI in a sort of pro manner. You can do rudimentary testing and all that stuff at colos, but I mean, you do want to do production. But if you could do distributed computing, I mean, Amazon's outpost was theoretically a good idea. You put an edge device, different hardware, but you don't need to use outposts, for instance, with a colo that uses all their power. Again, you optimize the racks in that facility and you build around the constraints of that power and then you find another colo. So I think the idea of having a diverse colo environment is just the nature of distributed computing. If the speeds and the interconnects, whether it's optical across campus, across the internet, then you have a whole nother arc, again, back to the systems design. Yes, systems thinking. So CNBC wrote an article today that Equinex is selling the pipe dream of AI. You know, like, I'm going to... I don't think there's a pipe dream. I think it's reality. I am going to their dinner, so I can't say much about that right now. CNBC is there or? No, no, Equinex. Okay. It's not a pipe dream. It's not a pipe dream. I think CNBC got that wrong. But somebody just wrote this. Pipe isn't like the fat pipe internet, but like pipe dream isn't not gettable. I think it's a pun there. Okay. So the other observation was, I'm kind of forgetting that, the second same thing now. The first observation was the... Your first observation was power. Yeah. The other thing was the session with all these scientists or these experts from different domains, Jensen did that today. So that was, I think, I opener from many angles, because these researchers also questioned the need for that much GPU, actually, during that talk to say, if you want to add two plus two, do you need a model with two trillion parameters to do that math, or you can just go to CPU and say, do this for me. So I think the switching of the computing needs on the fly between CPU and GPU, that is going to be the biggest sort of work in progress in next two to five years or so. I think the paradigm, that observation about that is a good one. I think in this place of what Jensen was saying about the word discovery versus design. Drug discovery, you're kind of fumbling around to discover something and then lock it in with the idea of the omniverse which they've been pulling as their digital twin platform. You can actually do all that in simulation and then design something. So it's drug design, not drug discovery, an example of finding cures to things. And that brings up my point about custom silicon, because if you're looking at what Broadcoms and these players are doing is that with automation and foundation open source, open standard foundation models and technologies, you can actually do all that automation in the front end and get chips to the market within three to seven months. Okay, and then a three-month production ramp. So within a year, you can have chips, custom chips designed for workloads. So I think you're going to start to see this notion of custom chips at scale, mass scale, designed for all workloads. So every workload that's in production, if it's a good workload, will be designed specifically for the workload. So I think AI, we're going into the systems thinking mindset with the notion of designing something, use case purpose built is the future. And I think AI is showing that you can do that and the benefits are multi-fold because now you have intelligence built in to the applications that's never seen before. So I think watch that custom silicon market, that's going to change the infrastructure game with clustered systems. Yeah, I'm with you on that. I think because having open standards, it makes economics better on all fronts, power, cooling, and diversity of vendors. Because if there's only one player in the game, they're going to charge anything. My final observation that I'll pass it off to you to get a reaction and then close out is the idea of software defined versus physical. And we had this conversation in context of storage. So for example, Pure Storage is a company that has flash arrays. You got Dell makes blade servers. So you got flash array. This is a physical where compute is in Pure's case, compute is on the flash array. And blades, the compute is on the blade. When you separate compute and storage, you're starting to think if you're not software defined, the physical fabric has to be, the physical materials has to be built for software defined. So I think if you're not looking at a software defined environment going forward, you might be in a tough hurt. What's your reaction to that? Am I overreacting on that or what? I think there might be some overreaction there, but you have a point there. Because if you are separating the compute from the physical things, it has to be software defined. If it's not, then it's like more like embedded software into that hardware. And that's old stack sort of paradigm. And the beauty of having software defined physical assets is that we can get the most juice are out of these things. And because most of these things will need power to run and upkeep and we want to shut these things down when we don't need them. So again, the switching of the computing from GPU to CPU or any other mechanism which emerges. Sorry guys, there's a lot of sound coming behind us. So I think that software defined everything is the future and we see that from happening. I see that happening. Everything's connected. Sharpe, G. Chowal, thanks for coming on theCUBE. Real quick vlog, what are you working on right now? What's getting your fancy right now? What's got your attention and what are you working on? What I'm working on is like I'm just extending my work with the soft economics of systems creation versus economics of systems operations. I am spending a lot of time on that. So because creation versus operation are two different things and at some point we switch to operations. In creation versus ops. Yeah, like dev versus ops, right? And I have said that many times dev ops has done a lot of damage to our industry. I was giving examples why, because the, you know. Sure, not sure I agree with that, but okay. That's another podcast, that was a whole nother one. We might have to debate that one, arm wrestle. Yes. Cool, well good to have you on theCUBE. Appreciate the commentary as always. Sargiccio all here, breaking it down with the analyst angle here. I'm John Furrier with Dave Vellante. As the music plays, we are going to exit out of day three of GTC. Thanks for watching. Thank you guys.