 Hey, everyone. Welcome back to SuperCloud 5, the battle for AI supremacy. I'm Lisa Martin, along with Savannah Peterson. We're live in Palo Alto and our esteemed colleagues are in Las Vegas creating and producing some fab live content on AI, the battle for supremacy. Guess who's back? Yes, he's back. Back again. Sorry, I couldn't resist that. 12-timer von Steuer, VP of Systems Engineering at Vast. And of von, great to see you again. It's only been a matter of weeks. Yeah, it's just been days, it feels like. It does. Hopefully, everybody had great Thanksgiving and got an AWS re-invent going on, and I appreciate you having me here for SuperCloud 5. So, SuperCloud 5, this is our fifth edition. Savannah and I, this is our first one, so we're very excited to be here. We think it's the sexiest one, by the way. But I did a little asking of the CUBE AI to find SuperCloud for us. And it talked about what the CUBE is doing, gathering this community of experts, hyperscale computing experts, technologists like von, investors, thought leaders, exploring gen AI, its impact on the major cloud titans. Talk a little bit about some of the things that, when you hear the term SuperCloud, the concept, what does it mean to you? That's a great question. I think the way that we, at Vast, need to look at SuperCloud is really in the retooling of enterprise infrastructures to make them AI-ready, to allow the democratization of their IP and assets and to be consumed by AI tools, large language models, multimodal, large language models, accelerated computing applications, as well as this birth of this new class of hyperscale cloud providers. The AI or the GPU hyperscale providers like the CoreWeave or Lambda or Core42, they're dramatically different than AWS and Azure and Google. Sure, all of them have some form of GPU or TPU or IP, whatever accelerated platform they have, but their ability to actually service the needs of large-scale customers who are deep in this gen AI space, it's like orders of magnitude difference between these hyperscale GPU cloud providers. I don't want to say legacy, but traditional. Traditional, maybe that's a better term. Venerable. Venerable, right. Yeah, I think, I want to talk, you just said a phrase that we haven't heard on the show, but I think I want you to break it down a little bit. We hear a lot about the democratization of AI. We've got a lot of collaborators in the space, but you talked about democratizing IP to give people access. Now when we think about the big old players in the Silicon Valley or home here in Palo Alto, you wanted to keep your secrets. In fact, there have been oodles and oodles of lawsuits and a lot of expenditure around secret keeping. When you say democratize IP for these enterprises, what does that journey actually look like for them? Great question. The best way that I think I could communicate what we're doing and how we're democratizing the data for AI is vast data provides a data platform for enterprises and cloud service providers. Our platform is really unifying data storage with databases and compute engine services that's all packaged together in a scalable platform that allows customers to have their data accessible globally on-prem or in the cloud. The reason why we've got all these capabilities within our platform is because the way that data is being used today, you've got to be able to support the security, availability, the governance aspects of enterprise data, where your assets live today, with the performance scalability that you need to power GPUs so that you can do your research. You have to make it accessible because data management is as important as model management and model security. Some of the things that we're seeing customers do with our data platform is deploy on-prem, have some Gen AI models, LLM work going on internally, but also extend that into the cloud. They don't have to move their data, but they can extend their vast instance into AWS and now they can have access to more models and more tools for their data set. With the cost of GPUs today and the strategic investments that are being made, there's a lot of pressure on validating your thesis of expanding your portfolio or getting to ROI if you will. The ability to tap into the current IP to accelerate it with the new tools to make it available through different models, we're the only ones that are powering that capability today and we're seeing it just across the board, all types of industries, use cases, whether we're talking about natural language processing, VFX and rendering forms, life science, like brain research and analysis, on and on and on. Off camera you ask me, what's the most interesting thing happening in AI? My comment was, it's this whole leapfrog or evolution within what we can do inside of our data centers on-premise or in the cloud and the outcome is like profound, it's exponential acceleration of computing power. Speaking of that acceleration, I'd love to understand what you're seeing from a budget perspective for AI adoption versus traditional IT budgets. Where are customers in trying to understand kind of the financials and how to make the best decisions with the resources that they have today? We're seeing from every industry significant investment in AI, it's the number one budget line item inside of IT departments right now, in fact even taking budget away from more kind of run of the mill or keep the business running initiatives. I think beyond AI, maybe cyber protection is probably like the second highest right now and then maybe third is like cloud costs. Again, that investment is driving a lot of pressure on getting your initiatives up and running. What we're seeing is kind of bifurcated a little bit. We're seeing the on-prem infrastructure folks really struggling on making a decision about what type of infrastructure purchase do I make? They know Cerebrus and they know NVIDIA and they know AMD. They know they're going to get some form of accelerated computing platform but what do they do for fabric? What do they do for storage? There's a lot of focus from that perspective and conversely in the same equation we're seeing customers say I just need to move faster and so I'm going to go reach out to a core weave or a lambda or a core 42 because I can move faster. Give you one example, we talked about the traditional hyperscalers. They all have GPUs and they've got a lot of customers using them. This is not meant to be a negative comment on them but there was a gap in the market about the performance and scale at which they deliver these services. If you look at something like inflection AI, they leverage 22,000 NVIDIA H100 GPUs from core weave. Core weave's business is growing like mad. They just built a data center in Texas. It's valued at 1.6 billion dollars and they're building 13 more. You flip over and you look at something like core 42 and core 42 who's partnering with Cerebrus, they're working on supercomputers and that are going to release next year that provide 36 exaflops of AI computing power. That's not even possible for me to comprehend. What do we do with that? It's like an infinite universe of stars and galaxies of potential power. Bring it into the west coaster. Look at Lambda Labs. The Lambda Cloud, they're making everything super simple for their customers. They've got their whole one click, you download or you type the line, I should say, but you download their toolkit and whether you're on-prem, in the cloud, on your laptop, you're enabled. It's a one-click refresh of the tool sets and they will rent you today an NVIDIA H100 GPU for just over a dollar an hour. That's anywhere from 50 to 75% off the price of the traditional hyperscalers. So all three of these, again, these AI or GPU hyperscalers, they're just, their business is going gangbusters. So then what's your forecast for the traditional hyperscalers? What are some of the things that they need to do to be able to be as competitive with the GPU hyperscalers, to enable customers to go as fast as they're expecting they can? Yeah, that's a good question and I've got to watch what I share here. We're for all the secrets. Yeah, yeah. Obviously, all the hyperscalers are investing in chips, right? I just recently read about the Azure, Microsoft Azure, as you know, I think it's pronounced Maya chip. I want to make sure I don't mispronounce it, but so they're all getting into expanding the number of accelerated compute platforms that they have. Obviously, they've got all the three major hyperscalers have relationships with NVIDIA. When you look at the rest of their infrastructure, the speed of the fabric, the storage platform to be able to feed the GPUs, GPU utilization right now is probably the biggest challenge that they have. I think they're going to have to look at some infrastructure investments and maybe even partnering opportunities to try to help solve that if they want to get to a level where the current GPU hyperscalers are. I think there's going to be a middle market there that we're not talking about. We're talking about very large enterprise applications. There are people who need to run smaller batches that require less compute to do that. I think what you're talking about there, that optimization, is actually going to be one of the maybe less sexier conversations, but more lucrative and important margins to optimize as people build out their stacks. I agree, absolutely. I want to be clear here. There's a lot of AI work going on in all the traditional hyperscalers, but when we talk to customers and we talk to a lot who've gone from a journey from a traditional to one of the more modern GPU hyperscalers, they talk about challenges of scale of some of their use cases. I talked about it before, like multimodal large language models, rendering forms. These are areas where pixel streaming, for example, they're not able to scale on the traditional infrastructure, not because of the GPU, but because of what's behind the GPU inside of the traditional, I should say infrastructure, but the traditional clouds, if you will. What's happening inside the server? I'm sure I'm not saying anything that the hyperscalers aren't well aware of, but maybe customers who are looking for enterprises who are looking to figure out, how do I accelerate my adoption of AI today? What cloud choices should they be looking at? What's that journey like for customers that you've talked to that are going from traditional to GPU powered hyperscalers? What is that journey like for them from an infrastructure perspective, but also from a culture and a leadership perspective? Yeah, I wish it was a simple sentiment. I can share with you a couple larger themes. The first is organizations that are in regulated industries. Not all, I don't want to make an absolute statement, but a large portion of those types of companies or enterprises are still kind of investing on-prem. It's IP leakage, data protection of their intellectual property. They're forcing significant investments on-prem, but with that said, they have teams that are out testing the different models and tools on various clouds so that they can build up their skill sets, start to accelerate the pace at which they're going to actually be able to utilize this new form of science inside of their infrastructure. We're also seeing with those that are whether they're on-prem or in the cloud, this challenge between how long can I keep my legacy on-prem infrastructure in place. It's tried and true. It's trusted. It's big name brands with banners in logos in blue or orange, but they can't feed their GPUs. There's this realization that the on-prem infrastructure, just like we talked about it with the traditional hyperscalers, the on-prem infrastructure needs to be modernized and either do brittle and fragile HPC for performance or you do slow and reliable and easy to manage enterprise storage or you look for an alternative. That's really the gap that we've filled within the market today. Our growth is being driven by the ag growth. From the cloud provider side, I think we touched base on it from the customers or the consumers of that cloud. It's a race to evolve their products and their product offerings to be able to do so without having to make a capital investment. I think speed, time to market, whether these customers stay on these cloud platforms indefinitely, I think will probably years from now kind of see just like what we've seen with the traditional hyperscalers where you launch your product, you get to a certain scale, now you got to start looking at cost economics. There's a lot of stories about organizations born in the cloud who eventually repatriated to some other type of platform. Maybe we'll see that over time. Who knows? We're talking a lot about traditional versus some modern applications in the GPU space and a lot of spaces. Something that we haven't defined very well but I think it's actually quite relevant to this conversation is the difference between traditional and generative AI. Can you break that down for the audience? Generative AI really kind of burst under the scene about a year ago with chat GPT-3. It really moved everyone's mind share. I think you could look at traditional machine learning. It still was kind of like this black art science, if you will, adjacent relative to HPC, but gen AI I think really got everyone's imaginations moving forward. What can I build with this generative AI? For example, if you look at like multimodal large language models or MLLMs, they have the ability to take in a diverse set of inputs, audio, visual, whether still or video, text inputs, and they're able to take that diverse set of input and actually learn from it, understand, and then actually apply learnings from that. You look at that type of technology, a great use case right now is like self-driving cars. It always was this deep technical stack kind of built from the bottom up but was always with strict rules, if you will, and logic. It would really struggle if like the emergency vehicle or road repairs right where lanes got, yeah, where you got to switch over, cross over lanes, scroll dog, crossover lanes, man's holding a sign. And now when you look at like MLLMs in the autonomous driving space and the opportunity that they present to now be able to react to tens of thousands, if not millions of unique scenarios that maybe can solve some of the issues that we've seen here even in San Francisco in recent months, right, with some of the self-driving vehicles having some challenges. A lot of drama, almost as much drama as open AI on the autonomous vehicle space in the city the last few weeks in the Silicon Valley. It's been a spicy one. It has been spicy. You talked about some really cool use cases, VFX rendering, autonomous driving, life sciences. As we round out 2023, we're heading into 2024, which is hard to believe that Y2K was that long ago and we cannot remember it, but as we do, what- Thanks for that lovely, useful reminder this morning, Lisa. I just had an eight spot, the cure of my hand thing. I'm so sorry. But as we do that, as my young friends have read historical analysis of Y2K, what do you think being AI-ready in 2024 is actually going to mean to organizations and how is VAS going to help them achieve that status? That's a really good question. I think of it in two ways. One that we touched on a little bit already, which is customers making AI investments quickly learn that they can't keep their GPUs running efficiently if they have to keep shuffling data from their legacy silo to the AI silo and then evict it and move it back and forth. In fact, we've met with some large enterprises that there's kind of like a consistent theme like what's your GPU utilization? 20%, but we're not moving things fast enough, so we're going to buy more GPUs. That's common. Yeah. It's your fabric in your storage platform that's your bottleneck, not your GPUs. In relative dollars, the fabric in the storage is a fraction of what enterprises spend on the GPUs. I think there's a little bit of a maturation of understanding that AI is no longer an HPC science experiment often in the back room. It now has to become pervasive through the enterprise and so retooling architecture is part of it. The second aspect of it is if you look at how important it is to AI, particularly around the accuracy of your models, the output of the models, the accuracy of your inference, for example, it's really a byproduct of how much data you can process, real data, not synthetic data, data that's been enriched with lots of metadata tags and is always being expanded upon in terms of the model base. If you look at what we're doing in terms of collapsing the different tools in the infrastructure so that we can support querying of all your data and the tags so you know which data that you need to be looking at, the ability for us to be able to take streams, process the data, and get it better prepared for AI, we're taking time out of that model. That's just native within our platform where today, if you're going to do it, you're dealing with lots of different bespoke systems, lots of multiple points of failure, different teams with subject matter expertise, we're trying to collapse it off, simplify, and at the end result, accelerate what customers can do with their data. And that's what they want. They want to be able to accelerate what they can do with data, make business, impact drive, new revenue streams, new products, new services, delight customers because it's the one thing that keeps going up as consumers is our expectations that everything is going to be real time, it's going to be personalized, and you're going to deliver me exactly what I want. And they want to do it wherever they are too, and I think that location matters. I'm curious, you're in a really sweet spot in the market, I can tell you're having a bit of a moment and have that wonderful grin on your face. What are the conversations around AI at the edge? Because this just makes that problem even harder. You know, I would love to pontificate on the challenges for AI at the edge. We don't make a product that fits every use case today. Today, our scale really limits us to service providers and enterprise. Our starting footprint is 100 terabytes, and so you don't see a lot of 100 terabyte at the edge. If you do, I would like to see what that edge device is doing, quite frankly. And so I'd love to speak to the edge, but to be honest with you right now, so much of our focus is helping the infrastructure for these cloud providers and these enterprises that I haven't had time to take a step back and look at the edge. Do you find that your customers are more excited about the future or feeling FOMO that they might get to the future and be in the wrong position? I think there's excitement FOMO and Fofu. What's both FOMO? Fing up? Yes. I'm great with my half-acronyms, y'all. You really are. John doesn't let me say them right off from here on the queue, but thank you for that opportunity. Yeah, so there's excitement, right? And it's coming from the CTO, the Office of the CTO, and the CIO's office about how do we evolve our product line, right? These are the kind of, if you will, the glass half full visionary track folks, right? There's everything we might be able to do with gen AI. There's the FOMO, which are the folks who are like, I'm not sure what I can do with it, but I'm worried about being leapfrogged, so let's get an order process right now and figure this out. And then there's the poor infrastructure folks. And those are the folks with the Fofu. They're like, I'm being tasked. I don't know what to do. I've got a lot of brands and vendors that I trust historically, and now I'm getting pushed into a new wave of infrastructure support, and I don't want to be the one that the finger gets pointed to if for some reason this falls down. You don't want to be the Fofu culprit? Right. I'm so glad I learned a new acronym today. Thank you both. Here to help you with your acronym anytime you want, darling. Here to help you. Always great to have you back. We need to get like a 12-timers jacket, but next time is... He needs a pan or a bag. Do I get a gold jacket at like 12 times? I feel like a jacket with a gold lapel is an order. I'm going to put it on. We're taking orders. We're taking orders. Yeah, we'll take measurements later. Yeah, it'll be great. 40 regular. All right. Cool. All right. I think that might be the... Okay, sweetheart. Sorry, we're both so excited to talk to you. They didn't tell us who was closing this interview, so I was just giving them a little bit of limbo. I want to give you one last closing note since we just had that moment. We talked about aliens in our last segment, and I am going to make it a theme of the day because why not if we're talking about a future that is yet to be realized in a lot of aspects. I don't see why we want to be talking about the futures of other things going on in the universe or other universes or who knows. Do you believe in aliens, Vaughn? Absolutely, but not in the bipedal sci-fi fashion. The universe is so vast, and there are so many planets that have the potential to support life. I absolutely believe in life throughout the universe. Now, do I believe in little green men in spaceships? The bipedals. Yeah, come into earth and landing and maybe collecting some of us for probes? No, I don't believe in that. All right, Vaughn, well, on that note, we look forward to appearance number 13 here on theCUBE. Lisa, I look forward to the next four days getting the co-host next to you. Thank you. And I look forward to sharing this entire experience with all of you live here from theCUBE Studio in Palo Alto, as well as our fantastic editorial coverage from Las Vegas. That's actually where we'll be going next. We have John Furrier and Dave Vellanti with their keynote analysis, editorial take from AWS re-invent down there in Sin City. In the meantime, my name's Savannah Peterson and you're watching theCUBE, the leading source for cloud and generative AI coverage.