 Hi everybody, welcome back to HPE Discover 2023. You're watching theCUBE's continuous coverage. We go out to the events. We extract the signal from the noise. I'm Dave Vellante. I'm here with Rob Streche. Lisa Martin is also in the house. John Furrier is on a plane. He'll be at MongoDB tomorrow. So check that out. But, and we'll be back here tomorrow as well for day three. But right now we're excited to have Justin Hottard back on theCUBE. He's executive vice president and general manager for HPC and the AI business group at HPE. Hewlett Packard Enterprise and Hewlett Packard Labs, right? So exciting stuff. Justin, good to have you back. Great to be here, Dave. And we just need a supercomputer to get my title right, is that it? Yeah, GPT, we're running through it. Shorten this title. But anyway, welcome back. What kind of cool stuff you're working on these days? Yeah, we had a little announcement yesterday that I think is pretty exciting. We've announced that HP is entering the AI cloud and we're launching our first public cloud service with HP GreenLake for large language models. So we're super excited about it. Yeah, so maybe take us through the story. We had Dr. Goh on before and he was giving us the sort of 101 on how supercomputer workloads are different than the traditional cloud workloads. But maybe we could unpack that a little bit. Yeah, absolutely. And I think probably the first fundamental thing is the bigger these models get, the more that you need supercomputing. So if you just sort of step back and say, gosh, why do I run into this problem in the first place? It's that to build accurate models, whether they're large language models or they're for scientific research or climate and weather, we need big data sets. You imagine you don't want to predict the weather by looking at the last three weeks of the weather, it's not going to be very accurate. So you need big data sets and the larger those data sets get, the harder it is to fit them into one computer. In fact, the more likely you're going to need to paralyze them across a lot of computers. And that turns out to be what we've been doing in supercomputing for years. Meaning one workload versus in the cloud, you're talking about multiple workloads on a single virtual machine. Looks like you got Dr. Go's lesson. Yeah, I mean, but okay. So my next question is how hard is it going to be for the cloud guys to sort of replicate that? Is it just sort of not in their DNA? Is it, is there IP that you have that you know the old, there's no compression algorithm for experience that you've got now the kind of flipping little jiu-jitsu move on cloud? Well, first of all, I think it's complementary to what the public cloud players are doing. So I see this as an opportunity for partnership and extension. But there are things that we know how to do and they start with our core in supercomputing, which is really how we compile and optimize and distribute the code so that it runs reliably across all of those, all of those computers at once. And then the other thing is, you know, and we've all been around hardware for a long time is hardware always fails. And so it's also how you build the richness into the software layer to solve for those hardware fail failures and have the code keep running, which is very different than what you do in a traditional cloud or virtualized environment where if the hardware fails, you just move the workload to another device, right? And so solving that problem is just, it's very orthogonal to what, you know, say a typical private cloud or public cloud technology stack does. Same thing with multi-tenancy. What we're doing in a supercomputer is we're making that whole cluster available to you at once, letting you run your instance on it. It's just for you, it's completely dedicated. But then in a couple of hours when your job's done, that whole supercomputer goes away. You don't have access to it. That's very different than dividing, slicing up small parts of a computer, of a server, right, to be able to run multiple applications at once. But strategically you could have chosen to put your IP in the public cloud, run it there. You're choosing not to. So that says to me that you feel as though you've got a distinct IP advantage that you can bring to your customers that is unique. Yeah, I think there's quite a bit that we bring beyond the stack I was just talking about. There's tools that enable programming and optimization, training of large models. And those are things that we think are complementary. I mean some of that today actually runs in the public cloud so our users can get started training AI models or building out some of their simulations in the public cloud. And then we can bring them over to a much larger system to run a much bigger model, train a much bigger model. So we think we've got IP there. We think what we do is complementary. And then honestly, we've got a core customer base that comes to us for technical expertise, right? What I would almost call a technical cloud. And so that, I think that's the other thing that we bring to this market. And that model you described yesterday, I mean that's not actually that unique in the commercial world. Yeah, I'm going to do some dev in the cloud, but when I really want to, you know, get into the runtime and I want to, for whatever reason, I want to run it on-prem. Yeah, no, that's a great point, David. I think the other thing, if you think about this, we're talking about training models in a supercomputing cloud. When I go back out to deployment, right, I need to actually go run that and put the inference computing. I've got that model. Now I actually want to run in the world so I can answer questions or query images. That may be deployed back in a public cloud. That's why I see this as such a natural partnership. And so that could be a public cloud instance. It could be a private cloud instance on HPE GreenLake. But this is also where it all goes back to that one platform for customers, right? Having that unified platform with HPE GreenLake where I can see the data, I can move the data in, train my models, deploy back out and do what works for me and works for my business. And it seems like there's a strong fit with we had just had the Esmeral folks on and that whole data fabric and getting the data there and what they're doing with some of the open source stack. How did they play together, I guess, is that? Well, one thing that's kind of a principle is open source. So if you think about what they're doing, they've got this fabric, the data fabric that makes it easy to move data back and forth, helps you organize and structure it, maintains its persistence. That's really important in an AI workload because you're dependent on all that data and you want to keep capturing the data, right? As the model's running, you want to keep capturing new data so you keep it current and updated and learn from it. That fits right in. And in fact, what we're doing with our software stack with the machine learning development environment and even the work we're doing around the data management platform is they're designed to plug in right on top of Esmeral. And you guys announced LLMs as a service and you talked about kind of three different use cases right out of the gate. I'm potentially a fourth one coming soon, I believe. And I think one of the things that it made me think about was what if somebody wants to bring their own model? Is that part of it as well? Yeah, I think our real vision here is longer term as a model marketplace. We're going to contribute some things. We want customers and partners to bring their models and tools. And we also want to offer what we would say are best of breed models. And I think that the reason we started with LLMs is we see a real need for a single tenant, large language model, where someone can bring their data, bring their needs, their workload, train it on a model privately, have their data protected, and then go out and deploy that. And that's, there's a ton of demand for that. And right now that's an unmet need in the market. One of the things that the cloud brought is it sort of leveled the playing field. I could, as a small company, you can get great security. You can get at least cheap experimentation, things like that. Is there an analog in this large language model of the world where you're sort of democratizing or leveling the playing field for many more customers? Yeah, I mean, I think about what, there's a demo on the show floor here what our HP services team did. And it was really in less than a month that they started with this idea that we want to start using this HP Greenlight for LLM's platform to just start automating our support experience. And they went from that idea to within a few weeks having a functional prototype where they can actually run real cases on the HP Electra server, storage platform. So that's the kind of speed to deployment that you just can't get. You know, if we were talking about this in supercomputing, we'd still be talking about where are we going to put the facility and how are we going to get the power to run the supercomputer, right? So this is a huge acceleration from that perspective. And I think, you know, I think there's also a ton of opportunity for new innovations, new business models. I mean, that's why when I talk about this, like I always go back to web 1.0 where you think about all the businesses that would never have been able to launch their business model, you know, in web 1.0, but we're able to once, you know, once they started to get access to the cloud and cloud services, it just took the barrier down so much. And I think this is the same story. Yeah, and best practices get shared much more quickly. When mistakes are made, you fix them fast, you don't make the same mistake twice. And so everybody benefits from that. I want to ask you about sort of the example with the web. It seems like, you know, when you go all the way back, you're too young to remember, but the PC wave, it was very disruptive, obviously, you know, took out the mini computer business and so forth. The internet helped a lot of incumbent companies. Example, Dell, go and sell in PC's direct or HP Inc. as well. So you had incumbents that were able to take advantage of it. There was disruption, books, retail, of course. How do you see this AI era in terms of that wave? Is it a benefit for incumbents? Is it going to be a disruptive? Both? How do you see it? I think it's both. I think it's both. The incumbents that have taken advantage of any of these cycles did it because they recognized the trends, figured out where they had unique IP and they could add value. That's exactly what we're doing here. I think that's absolutely an opportunity, you know, a play. But there's going to be a ton of new business models, new sources of innovation. And I think that creates new opportunity for startups to come in and build and create value. And we're already seeing a ton of that with consumer, you know, consumer LLM applications and uses and workloads. But I think we're only scratching the surface because there's so much innovation in transforming the customer experience on a B2B side. Yeah, so your HPE's interest is not in helping customers, or maybe it is, but not directly helping customers disrupt search. That's not what you're doing. You're not doing consumer. But when you think about life sciences and medical those industries are so ripe. I mean, every industry has a disruption scenario that's sort of now we're starting to see the digital impact. You know, and AI is really bringing that into fine focus. Well, Dave, even if you think about some of the hurdles in the past, if you look at some of the challenges with taking some of these technologies early, like Web 1.0 and deploying them into businesses, there were so many challenges with changing the business process, changing the legacy software architectures, right? And really breaking through, I think with AI what's so compelling is you don't actually have to, you have to deal with some of that, but you don't have the same hurdles because you can bring the data, bring the insight, and you can actually automate a lot of the development. So I think it's a really interesting accelerator and that's why we're so excited to be at the ground floor of it. Help people understand sort of the whole notion of trust and explainability and reliability of the data because you think about, I mean, GPTs for amazing, takes tests, does calculus, pretty well, et cetera, passes law exams, but I wouldn't necessarily trust it to really drive medical solutions, or even legal, things like that. So help people understand your strategy in terms of driving those types of outcomes. Yeah, absolutely. So if you look at what we've announced for HV Greenlink for LLMs with Alaphalfa, they've got a foundation model that's trained on language. What it understands is language very well. What it doesn't try to do is understand everything. So then what you bring into it is your own proprietary data. So let's take a legal example. I'm going to bring my contracts into that and train a model on my contracts with a foundation model that really understands language well. By the way, understand six languages well, so I can train my contracts from France and Germany and many other countries on that foundation model. Now what I've got is an expert model that's tuned just for that application. So when I go ask it something about my legal contracts, it's trained an expert. I think it's a great point. The problem with these general purpose models, the open models, they're great for consumer applications, but they're learning from everywhere, right? And so they're learning from people that are right or wrong and they can't distinguish it. With these privately trained models, not only do you get all of that, but you get explainability, you get auditability. So you can actually see, hey, why did the model answer this question that way? What contract did it go back and find that it decided that that was the answer? That's the level of specificity. And you think, I've got a general view that for us to have broad adoption of AI in enterprise, we have to be auditable. We have to be explainable. We have to be responsible because there's no way an insurance company, for example, is going to bet replacing its actuarial tables on a model that it doesn't understand extremely well what its assumptions are and why it's underwriting the risk the way it's underwriting. So true. And then when you're trained on the internet and social media, there's going to be a lot of garbage in there. And the system has to be evolved so that it knows how to parse what's right and what's wrong. And that's really not the objective, at least today anyway, but it is your objective. Yeah, absolutely. Yeah, and I think one of the things that kind of what we've talked about previously and it was around sustainability and why it's different at the super computer level versus just the standard using GPUs in a box, at a pro line level, for instance. Yeah, that's a great point, Rob. So the supercomputers, what's happened in the history of supercomputing is we had to go down the path of liquid cooling and efficiency because what we were dealing with was so much density and so much heat coming off these chips that we had to find a different way to cool them. Fans weren't good enough. So we ventured into liquid cooling many years ago and now we're very expert at it. We also, you know, many of these centers and other things have, you know, centers and other deployers of these systems have always looked at clean power as a foundational principle because they realize the amount of power they're consuming with their systems and they want to do it in a sustainable way. When you put those two things together, what you get is something that's by de facto green if you make the investments. And what I mean by that is now I've got clean power and I'm cooling the system with water. I actually, I don't have big air conditioners. I don't have to manage all of the environmental constraints. And then the last thing that happens is if I can actually take the wastewater and use it to heat something because it comes out pretty hot, it comes out like 40 degrees C or even higher in some cases. Now I can actually reuse that. I'm actually getting carbon negative. It's a really, really powerful, really, really powerful story. And our commitment is we're not gonna, we recognize AI is gonna be a massive incremental compute demand for many enterprises. And like, as we're one of them, but many enterprises have committed to carbon neutrality, taking the pledge to be carbon neutral not only by 2050, but in some cases fully carbon neutral by 2040. There's no way we can add all this compute and not have it be at least carbon neutral because what we'll be doing is sending companies backwards while they're trying to make progress forwards, right? And so I think it's, to me, it's a foundational principle. I believe it's differentiating, but I also think it's fundamentally important. And as a citizen in the climate, as a human being, I also realize we've got to be doing, not only be doing great things for humankind with our research computers and the technology we can bring to bear, but we got to also make sure that we're doing our part to make the planet healthier. And Justin, just to refresh our memories, this is available the second half of this year? Yeah, so North America and the second half, and I think of this as availability zones. I've had a couple of questions about this. So North America second half, it doesn't mean that someone who's not in North America can't use it. It's just that if you've got privacy or GDPR requirements or other things, it'll be North America second half at 23 and then in Europe early in 24. Justin, thanks very much for coming back to theCUBE. Congratulations and best of luck doing some good work for the first society. Thanks so much guys, appreciate it. You bet. All right, keep it right there. Dave Vellante with Rob Stretche, Lisa Martin. We'll also be back. We're at HPE Discover 2023. This is day two. Up next, we're going to follow the money. Keep it right there.