HPE ProLiant Compute AI Portfolio and Solutions

So how do y'all, um, Scott Schafer, I'm the chief technologist for our computer organization. I also lead our advanced development team, and, uh, with me here is ov Yep. Uh, vibe ever og I'm responsible for compute AI solutions, And we're gonna go through a little bit of what we're doing from a compute perspective on ai.

I think as, um, uh, and hint a little bit that what our friends are doing over in Cray, I think that we kind of established earlier that, um, there's a significant portion of Hewlett Packard Enterprise working on some of the largest solutions ever built in, uh, in Cray. That's now what we're here to talk about today. But we'll point at where our, our solutions stop and theirs begin.

And with that, I'll turn it over to vi Bobb and he can go through it All. Awesome. Great.

Hey, thanks, Scott. Hey, uh, good afternoon everyone. Uh, thanks for joining us.

So let's start with, uh, some of the most common, uh, uh, and popular use cases, uh, for ai, uh, especially in the enterprise. Uh, so I think, you know, we've all seen examples of how enterprises have been able to, uh, uh, actually demonstrate business value from the AI initiatives. Whether, uh, it's a retailer implementing a combination of computer vision and inventory management to enable, uh, experiences like, uh, self-checkout, uh, cashless stores to financial institutions, uh, using AI for everything from fraud detection to, uh, credit worthiness.

Uh, manufacturers, uh, have been using AI for a number of years. Uh, everything from, uh, predictive maintenance to, uh, getting a better handle on the levels of the inventory. So, so this is great, right?

I mean, the expectations, uh, from AI are very high. I think, you know, there's emerging now proof points that sure, that there isn't, uh, too big of a disconnect between expectations and reality. But then, uh, we still, uh, come across, uh, statistics and experiences like this.

Uh, this one's, uh, from a, a study that got, uh, IDC did in 2024, uh, where they learned that, uh, uh, just about one ai POC actually makes it to production. Uh, number of reasons why, uh, I mean, concerns with the data privacy is one, uh, uh, skills gap is a major issue. Uh, just trying to keep up with the pace at which, uh, uh, the technology is evolving.

It feels like, you know, almost every week now, there's this new LLM that just closed everything, uh, else that came before it. So, I mean, these are some common challenges, uh, uh, that we all read about. Now, what I can say is, uh, this is not something that we see, you know, with most of our customers.

Our customers tend to fare better than what this statistic seems to suggest. Uh, and part of the reason why that is the case is because of our market leadership when it comes to HPC. So, if you think about our Clay super computing portfolio, uh, the ip, uh, that came to HPE Pharma, ISGI acquisition, uh, so today, seven out of 10 supercomputers are HPE.

Uh, the top three supercomputers are HP, uh, HPE. And why that's important is because, uh, you know, some of the challenges that customers face with successful cluster deployments, they're also very similar to some of the challenges enterprises are facing today when it comes to, uh, building infrastructure solutions or building complete workflows with ai. Uh, and the other thing, of course, is that most enterprises who are deploying HPC also have AI as part of that workflow.

So think about, you know, uh, a pharmaceutical, uh, somebody in, uh, who's, uh, doing like discovery. So they'll typically use AI to see how certain molecules, uh, interact with each other. Now, what's happening is that ai, they're using a combination of AI and HPC.

So AI helps them identify, uh, which molecules would make the best candidates. Uh, and then HPC, of course is, uh, you know, uh, that triggers a workflow where you simulate how a certain protein molecule interacts with others. So, you know, we think of HPC and AI as a continu, right?

And so on this slide, uh, this is, you know, something, you know, uh, for manufacturing, uh, customer. So, you know, you could be using HPC, uh, uh, if you're, let's say, building turbines. So you could use HPC to do a bunch of simulations, fluid dynamics, you know, uh, aerodynamics to design, uh, your turbines.

But then, you know, when it comes to deploying them in the field, or you've got these turbines, you know, out there, uh, these are very expensive to repair. Uh, uh, these are the kind of, uh, machines where downtime ends up being incredibly expensive. So, uh, uh, we'll talk about this example, uh, uh, a little later.

Uh, but one of our customers is using a digital twin. Uh, and then when they ship out the turbines, they have sensor data in them. And now what they're able to do is they're able to bring AI as part of that workflow to be able to do predictive maintenance.

So, so that's why, you know, we, we say that, you know, HPC and ai, uh, think of that as a continuity. Uh, and then of course, right? I mean, uh, if you look at, now, if we just focus on AI for a second, uh, it all starts with training.

So there, uh, uh, clay portfolio makes most sense. They also think about, you know, somebody who's in the business of creating their old LLMs from scratch. So they're gonna need massive scale.

So that's where, you know, we position our clay portfolio. But then, uh, uh, a lot of enterprises are starting to do model fine tuning. So if you're an oil and gas, uh, company, if you're, uh, a company in financial services, uh, the standard out of the box LLMs, uh, they don't understand the context, uh, of your particular business.

So that's where fine tuning makes sense. And there, you don't need, uh, a fairly large cluster. So that's something that you could do a smaller cluster built with alliance, but really, if you think about what comes after training, fine tuning, it's in, that's where the actual business value of all of these investments into AI is going to be unlocked.

Uh, that's all happening at the enterprise, and that's why, uh, our ProLiant portfolio, which has been the enterprise standard, uh, makes so much sense here, because now what we're able to do is we are able to, uh, deliver a complete portfolio, uh, all within HPE, uh, not just with servers, it's also with storage and networking. Uh, but we are able to address the needs of customers, whether they're training models or they're fine tuning them, or they're doing inferencing, uh, inside of your data center or inferencing at the edge. So, uh, what does our portfolio look like?

So, uh, for Gen 12, uh, based platforms, uh, this is our, and the, uh, uh, accelerated, uh, portfolio. So we've got the 3 84, it's actually on display over here. Uh, this one has, uh, eight H 200 GPUs.

Uh, uh, we also have the three 80 a Gen 12. And then of course, uh, uh, for model builders or, um, uh, uh, large GPU Cloud providers. Uh, of course we have a credit portfolio.

I don't, I don't think we have a 3 84, do we? Uh, I think, uh, right here? Yeah, no, we have a three 80 a, oh, sorry.

Yeah, yeah. Okay. But we don't, we don't have one of the gray hopper platforms here.

Okay. Um, but, um, you know, hopefully our brand renewals Grace Hopper, right? It's an ARM-based C-P-U-G-P-U combo on the same module, and it's designed to be the absolute highest performance solutions that you can put into kind of a single platform view.

Um, because of that really high speed, uh, interface between the CPU and GPU, you can, it's an envy link instead of, uh, traditional interface like PCI. And so for certain, you know, kind of getting the absolute highest performance on a single platform, that's where this, this is. Um, then when, um, you're talking about a three 80 a with up to 10 Nvidia H 200, uh, GPUs in it, um, now you're talking about a solution where you're, you're in a lot of cases, you're breaking that up.

You're apportioning individual GPUs out to a portion, like a container or a VM on that platform, or you are trying to, uh, provide a big parallel inferencing solution, right? If you're saying, Hey, I need to, was talking with, uh, our own IT organization who's providing LLM services internally, they say, I need to be able to support multiple simultaneous streams of somebody using a, a, a chat interface. That's when you, you step up to something like, um, a three 80 a, uh, gen 12 here with, with all of those, uh, Nvidia GPUs in it, for sure.

Yeah. Yeah. So How are customers reacting to Jensen's on a couple ago?

He won't be able to give away That, and that he's the biggest wealth destroyer or whatever. Is that what he said? Yeah.

Um, so first of all, I would say you, you know, look, I, I appreciate who he's talking to most of the time, right? He tends to be talking a lot to Wall Street and the, and that I appreciate that completely. Uh, he's trying to place a narrative, right?

That says, oh, no customer will want those old things. They'll all buy my new thing. You know?

Um, that's just not true. There's plenty of use cases, there's plenty of solutions that they don't have to have the absolute, you know, brand newest thing, and they don't have to, you were expressing concerns about having to upgrade everything. It's like, no, of course not, right?

The, those H one hundreds aren't instantly worthless. Um, they still do plenty of great stuff. And so, um, we're definitely seeing, you know, I would say enterprise customers be a little bit more interested in using what they have, uh, rather than have to always migrate to the latest.

Now, if you're a cloud service provider, well, you probably do have to move to the latest, uh, at least for some portion of your fleet. But, um, again, I think just like, um, we see everywhere else, like the, the cloud service providers don't swap out, swap out all of their CPUs to the new CPU U instantly, right? They're not gonna do the same with GPUs as much as Jensen might like it, that's not what's gonna happen.

Yeah. He has some pretty interesting power arguments, but I want to narrow in kind of the use case for this portfolio. Is it mainly, and it looks like you folks are targeting the typical enterprise use case of inferencing, so, absolutely.

If, you know, if you're looking for this platform, except for the, all the way to the right, if you're looking to leverage AI and put AI to work, this is the platform. You're, you're, Yeah. And, and you need these GPUs.

Mm-hmm. I think, Gina, you mentioned it, I think you mentioned it seemingly hours ago, uh, where you said, isn't there CPU only inferencing? Am I right?

Okay. Yeah. And that's absolutely, we see tons of that.

And so there's, we have plenty of great platforms that serve as really good solutions for your CPU only inferencing great use cases for that. Um, we were talking about something, well, he'll talk more about the about them in a minute, but, um, when you need GPUs, these platforms are designed for it. So, Um, do you see the new models, like LAMA just came out before with behemoth, maverick and Scout, right?

Um, and, and, and the concept of distillation, deepsea and that kind of stuff, right? Um, how is that influencing how you're designing systems? Because you've got, you know, obviously behemoth, the idea of, you know, you have to be able to get that, that entire model into memory if you can.

I don't see how behemoth you're gonna be able to do that with behemoth, but apparently you will. But the, I guess that's they're designed for the larger Blackwell, the newer ones, right? And the Vera Rubin for, for sure.

Right? So, so like, there's this continuum of, of, of where these models are gonna drive you. And as the capabilities of these, you know, of, of the Nvidia chips get, you know, bigger, um, the models that are gonna be out there that people want to use are going to, you know, it's gonna be like the hermit crab, right?

I gimme, I got a bigger shell, I'm gonna grow into it. Uh, how are you gonna be able to, to, like, do you still believe there'll be a market on the lower end for like L forties and the H 100? I see it too, but I wanna see what You a hundred percent.

Yeah, we absolutely do. I believe there's gonna be a very strong market for the platform with a single GPU unit. Um, L fours.

Sorry, what's the new thing? An RTX 6,000. 6,000.

Yep. I'm glad I remember that. Um, but RTX 6,000 class or even below, um, I definitely see plenty of customers coming in, you know, running a fours.

And So are they gonna be using like, the distilled versions of the new stuff, or are they gonna be using old stuff model Wise? Oh, wait, so, uh, my answer to you is both yes and no. And that, and I mean, yes, and that they'll be doing a, there'll be a lot of both out there, okay.

But I think we're gonna see new models that don't require a terabyte of memory to operate. And we're gonna, we can we continue to see investment in 8 billion parameter models be incredibly effective for all kinds of use cases, um, and 3 billion models, right? And as we see some of the work that's going on in research to be able to, uh, they basically call the model twice.

It's about the simplest thing you could think of, right? It's like you just call the model twice with the output from the first time and say, can you improve on this? And it dramatically improves its output.

You can do that with eights, and you get an incredibly effective, uh, response. And so some of the research that we're seeing suggests that we'll see, continue to see investment in that, and then I'm gonna be started on deep seek and some of the enhancements that they've seen. And then, I don't know, does, uh, you know, have you heard about the move down from a floating point perspective, right?

It was FP eights to fp fours, but only Blackwell's gonna support that, right? Well, and then yes, except have you heard about it going lower than that to inte integer only? And so that's some of the, the research in universities are looking at, can we get models that are integer only?

And when that happens, um, then you don't need floating point engines. And so, uh, if, if that happens, I'll say, if I don't know for sure if it's going to or not, but where, and my suspicion is it will, but it won't be great for everything. So then you'll see some models that are optimized for integer engines, and I know where the best integer engines are, right?

They're in CPUs. Mm-hmm. So those are the kind of things that I think, um, we'll see throughout this whole space.

Um, but you asked an interesting question, I think, Glen, which was, um, how are we approaching designing systems? And so we're doing a couple of things that I would characterize as part of our design process. The first is we're working very closely with our partners like Nvidia, to say, where are you going?

What do you need for your next generation system? And we're building systems around where they're headed, right? Because that is, that would make sense, right?

We built this system around Blackwell so that we would be able to accommodate it, right? Um, and so we're gonna continue to do that, but we're also gonna stay really close to our customers here as they evolve and adapt and be able to say, what do you need for the various use cases that you're, you're actually deploying? Um, vibe, Bob's gonna get into some, some specific ones, but they will, there are a lot of use cases where, as we said, you know, the big GPUs is not the answer.

So now what's the right optimal platform for that? Yeah. And I would love to hear, I'm sure you're gonna get into it, is most of like the hyperscalers, their compute is consolidated into these centralized areas.

Enterprises don't look like that. And the other big piece, to your point about the models getting bigger and bigger, they're also getting smaller and smaller. But we haven't seen the, the 70 B version of four, uh, four oh, of LAMA four oh.

And most enterprises are not deploying 400 billion, uh, parameters for typical use cases. So, excited to hear where you guys are going. Yeah.

Yeah. Thanks, Kyle. All right.

So we talked about the portfolio, but then I think, uh, you know, one of the other things, uh, uh, especially if you look at the motivations for my enterprises are actually, uh, spending a lot of money in AI initiatives. It's, uh, it's edge over their competition Yeah. To improve, uh, the customer experience.

So, uh, it's all about this. The, the underlying theme really is performance. And this is where, uh, we've actually just, uh, this is out of the press.

This is from last week. Uh, uh, we ran ML PERF benchmarks, uh, that there's a link at the bottom of the slide that gets into the details of what the individual, uh, uh, uh, ML perf benchmarks we ran. But it's, uh, uh, unstable D diffusion.

Uh, so that's, uh, uh, uh, stable diffusion was one, uh, uh, lama a couple of, uh, model, uh, one that we did there. And then of course, uh, we did one with missile. So, again, right?

So this is very, uh, a good way, uh, for us to demonstrate our, uh, commitment, uh, if you will, to be able to engineer and design these solutions, optimize for specific use cases. The performance is fundamental, uh, for ai. So now let's talk a little bit about our, uh, uh, as a service offering, our private cloud offering, uh, with private cloud ai.

One of the things that we did last year was recognize that a lot of our customers were telling us, I need to be able to acquire the, the infrastructure. I need to run my in-house ai, as you know, all at once, right? I don't want to assemble it.

The parts myself make it easy for me. And so, private cloud AI is that, and that we created a solution that includes both the hardware and software, all as one single purchasable entity that allows a customer to just acquire it and go, and it's somewhat, I mean, it is model agnostic. We don't exactly know what model.

I've talked to some of the customers, if some of them are running, uh, weather, uh, models. Some of them are running, um, medical and health related stuff. Some of them are, are running chats and building their in-house versions of chat GPT kind of approaches.

So, um, but our solution here was ready to run out of the box. It's designed to be able to be scaled. You can see, you'll see it immediately where we created some of the scale points, um, and then includes all of the software that we needed, uh, to be able to operate this.

And, and go to the next slide. What, what's our approach has been at the software layer, is to give the IT shop a DevOps solution for managing their AI infrastructure focused on inferencing. We can, we can also support people who wanna do their own training or wanna do their own fine tuning.

It absolutely supports that. You can set up Jupyter notebooks for your data scientist and, and manage it that way. But think of it as the software solution provides the, um, the DevOps model for operating your AI from a hardware perspective.

Um, we have everything from the developer system down here on the low end, which has a couple of systems, a server for AI and the storage, um, as well as, uh, the management components. And then it goes from, you know, kind of t-shirt sizes, right? I got small, medium and large based on, you know, four to eight L forties, if that's what you want to go to.

Stepping up to eight to 16 L forties, then moving up to the 16 or 32 H one hundreds. Um, and then eventually, you know, well beyond. And so, um, this setup has proven to be incredibly interesting to customers, right?

Their feedback has been, thank you. This is what, this is what we were looking for. Give me all the infrastructure.

So it includes the storage that you need. It includes the servers, obviously with their GPUs. It includes the management, uh, nodes and the management software.

And it includes all the networking, both east-west networking, uh, sort of inside the rack networking, as well as egress support for connecting into your existing enterprise, uh, data, wherever that might be. So all of that is, uh, kind of built in. Okay.

Yeah. And, um, I can, uh, I can do this. Yeah, go ahead.

Sorry. Uh, so, so now, uh, I think we introduced the portfolio. We talked about, you know, some use cases.

Now we will, uh, we'll talk about, like, you know, how we see some of the customers actually moving value, uh, with what they're doing, uh, with AI initiatives and the investments. So one of the things of course we see is that, uh, uh, ai, you know, the capabilities that's all moving to the edge, uh, edges where the data is actually generated. So if you, uh, if you're a retail store, you want to enable things like, uh, self-checkout, uh, that's all gonna be done at the edge.

Uh, if you're a manufacturer who's looking to do predictive maintenance, that's all happening at the edge. But it's not, it doesn't mean that everything's gonna happen at the edge, right? Again, it's, uh, it's gonna be a continuing.

So, uh, so if we talk about our, uh, hypothetical, uh, retail example where a customer wants to implement, let's say, self-checkout. So they're gonna use computer vision in combination with inventory management. Now, your inventory management system is gonna be part of your centralized, uh, monolith, uh, that's gonna sit in the data center.

That's not going to be sitting at the edge. But what you can do is you can have an ingen ai, uh, workflow where you can have an agent, uh, running at the edge, and anytime they notice that inventory has fallen below a certain threshold, uh, uh, using data from your centralized, uh, ERP, you can now dynamically place orders with your suppliers. And you can do things like, Hey, you know, who's giving me the best price as of right now, as of this very instant as, uh, to place an order for a particular budget.

Uh, so an example of this is, oh, Wait, hold on, wait. So, yeah, I can't help it. So wait, let me, I, the proof of this is when I've been talking to some different customer use cases, um, that really, like, oh, of course you would do it that way.

Um, I was talking to a, an airline and, uh, do you know that, you know, as you stand there and you get on the airplane and they're carrying your bags, and eventually someone says, our overhead space is full. You know, that was because a human being was tasked to count the number of bags coming on the airplane. And they don't even have a clicker.

I was doing the clicker with my hand. They don't even have that. They have to keep it in their head.

Um, and so we said, what if the camera could just count the number of bags getting on the plane, and then, then knew when it was full and declared it, cleared it full. And so they were, that makes a ton of sense. And they'd want to do that inferencing right there, right at the right there at the airport.

They don't wanna backhaul that all those streams from every gate back to their, to their central site, right? That doesn't make sense. So that's an edge use case for sure.

But I was working with a retail customer at the edge who was using it for, we'll say, inventory management. I'm gonna stay pretty vague here, but inventory management, right? Where they could take a picture of the store shelves and know what needed to be restocked.

Well, you don't have to take that picture real time, right? You don't need video of that. You only needed image every so often.

I mean, conceivably, you could say, I only need the image every hour. Um, or maybe even only every four hours if I recognize that that's the interval in which I restock anyway. So if you're just taking an image of the shelf every four hours, maybe you can backhaul that and do the infra, do the image management back, or the image, uh, AI back at a central location.

And so we're seeing customers start to make these trade-offs and say, what makes sense? What do I need? What kind of solutions can I deploy?

And it's less about ai. I think that's the thing I should get across the most. These customers do not want ai, that's seems obvious, right?

They want a solution to a problem that that's what they need. And so that's what we're seeing them start to come up with as these use cases start to become well known. I think that speaks to what I was saying before too, because obviously that's computer vision and simple, you know, counting really, and maybe drawing from, you know, airline, you know, airplane specific airline, a specific airline plane, how are they're constructed and how many bags can go on it.

But I think that's part of this too, is people don't realize what we don't think is AI anymore is ai, you know? Yeah. And AI is not just these large models being used and trained in, in an inference of different locations.

That's right. And I think that's the practical use case that we're gonna see in most enterprises. Yeah.

Okay. Okay. All right.

Now that's good. All right. So, sensei, um, is a technology company.

They're based out of Portugal. Uh, they, and the businesses building, uh, experiences for retailers where they can enable, uh, self-checkout. So it's cashless stores, uh, and they use a combination of computer vision, uh, uh, along with, uh, other algorithms that can help retailers, uh, really deploy end-to-end, uh, stores where there's no human, uh, uh, uh, interaction involved when it comes to, you know, basics like grocery shopping.

And so, uh, they're using, uh, deep line DL three 80 as the backend. It's the engine, uh, uh, and they're, uh, processing, uh, large volumes of realtime camera streams, uh, to be able to look at everything from, uh, customer dwell times in certain aisle, uh, uh, you know, when somebody picks something off of a shelf. Uh, so that's not a use case, uh, uh, that we, uh, are seeing, uh, with, uh, customers in the retail segment.

Uh, so if you take a look at another example, uh, this one's for manufacturing. And, uh, here, I mean, uh, uh, we're trying to represent a globally distributed manufacturer with factories all over the world. And so they could be using something like computer vision to look at, uh, to identify defects, uh, in products as they roll off the assembly line.

Uh, but, you know, another more interesting, I should say, example, uh, is, uh, something that's known as site balancing, uh, from a supply chain perspective. So, imagine you have factories all over the world, and a certain product ends up being like, very popular. You run out of inventory, or you have a large customer placing a very large order, you've run out of inventory, uh, at that particular location.

But you know, your sites or your factories elsewhere, uh, may still have those levels. Uh, so all of that, of course, uh, uh, you know, your supply chain systems do manage that, but AI is something that it can be used to automate that. It can, uh, uh, build in, uh, that concept of, uh, being able to predict or anticipate, uh, uh, some of these, uh, events and then take action autonomously.

Um, so, you know, certain cases, it may make sense to actually place an order with your supplier, because a particular widget has actually, uh, you know, you, you can get a very good price on it, uh, in real time. So site balancing, uh, is an option, but, uh, maybe, uh, there's, you know, it's gonna be more profitable, uh, for the manufacturer to just to, you know, acquire new inventory at like much lower prices. Uh, so that's one.

Uh, so, uh, Bosch is an example of a customer who's, uh, uh, had had success with a very successful implementation of ai. And so, uh, what they've done is they've built a digital twin of their turbines. Uh, and, uh, so they, the ultimate objective is, uh, reduce the, uh, or minimize the downtime, uh, anticipate failures when they happen.

And the way they do it is that, uh, they of course have a lot of historical data, you know, from, uh, all these machines that are out there that are deployed in the field. Uh, every new turbine that they ship out, it goes with sensors. So now what they're able to do is they're able to analyze the sensor data in real time.

And AI is part of the workflow, where now they are looking at historical data patterns and then predicting when, based on a certain threshold, or based on what your sensors are telling you, that hey, you know, you might be looking at a potential failure. So instead of waiting for that to happen, spend spares in anticipation of, uh, uh, that event. So, again, this is an excellent, I, I've been working on failure analysis for as long as I can remember, right?

As build and servers you solve, you try to solve that problem, right? I mean, um, who hears, remember a major server manufacturer saying they, they were gonna implement self-healing servers? Mm-hmm.

Right? Did they do that? I missed it.

If they did, uh, they actually don't build servers anymore, but, um, that's okay. Um, this concept's been around, and what I've looked at is various algorithms for predict, for predicting a failure for, again, as long as I can remember, um, we've had the simplest linear regression algorithms built into the hardware, even things that people don't even know anymore, but we do, they're still built in there, right? To predict memory failures and that kind of thing.

And so, uh, that's there. But what I've seen with the AI prediction stuff is unbelievable how good it can be. It's not a linear regression, right?

It is actually looking at the spikes, peaks and valleys, and when they occur, and then able to predict out what's going to happen and so much, uh, uh, uh, much better than just a linear regression saying, Hey, you're headed down and you're gonna cross the threshold in three years. It says, oh, no, you're gonna have a valley here, and that's gonna hit the threshold. And we see that happening in the nearer future.

And so let's take some preventive maintenance today in order to prevent that. It, it's really at another level. But most people wouldn't know that's ai, right?

If I just said we had a better prediction algorithm, the fact that it uses machine learning technology, sort of like irrelevant, right? Is like, they just want, I want that outcome, right? I want that solution.

And that's what we're seeing, uh, start to come in. So, really cool stuff. If you haven't seen a, there's an, an example of this that we built into comm the team, I don't think really spoke about it, but, uh, there's an AI called profit out there that does this kind of predictive analytics stuff, if you wanna look and see what kind of cool stuff it can do.

Yeah. Yeah. That's, that's excellent.

All right. And then, uh, of course, uh, what we've also done is, uh, we've, uh, partnered with, uh, ISVs, uh, who are leaders in, uh, addressing specific use cases. Uh, so Ian, now they, they're rebranded, and now they go by vaio, uh, is one that clarify StrataVision.

Uh, and, uh, I was just talking to Blanc, uh, a little earlier, uh, about, uh, uh, how vaio a family called Ian, uh, uh, and HPE, we worked together to help a school district in California, uh, implement computer vision. And so the, the challenge the customer wanted to solve was, uh, uh, you know, there was, it was a security, uh, and a safety issue. Uh, and, uh, they had about 3000 cameras.

So it's not humanly possible for, you know, somebody to be glued to like 3000 screens. And, you know, even if they're three, it's possible to build 3000 screens, um, and be able to respond in real time. Uh, so computer vision is what helped, uh, uh, address that issue.

Um, what I love about that example is the customer did not have to upgrade 3000 IP cameras. I mean, think about it, right? I mean, just to be able to do this, uh, automatically, uh, yeah.

You know? Yeah. But upgrading 3000 cameras, that's quite a bit.

So they don't have to do that. So value is not of VMS. So the customer has their own VMS, uh, audio is just analyzing the streams that are coming out of the existing IP cameras, running analytics on them.

So they can do everything from, uh, you know, uh, doing an analysis that, Hey, the scene has changed. They can do license plate recognition, they can identify certain objects like guns, uh, right? And they can lack them.

Uh, so again, right? And the way, the approach that we take with ISP solutions is that we, uh, we go through, uh, an intense validation process. We wanna make sure that everything works, uh, because ultimately, you know, the customers are enterprise customers.

There's certain expectations out of HPE. And so, you know, and that's the same experience that we try to deliver anytime we partner with ISVs to build joint solutions. Uh, and then of course, uh, uh, you know, the other thing I'd say is that what we are also starting to see is, uh, uh, enterprises are looking for complete workflows.

Uh, computer vision, for instance, is a use case. But let's say if you wanna do self-checkout, uh, your entire workflow is a lot more than computer vision. Yeah.

So it's, it's, computer vision's a part of it, I gotta take care of, okay, how are you gonna manage your inventory? Do you wanna automate that? Or is it still, you know, happening manually, right?

So, and that's really what we see as a common team across all enterprises who've been successful. They don't think of AI as use cases. To Scott's point, it's what's the ultimate, uh, uh, solution that they're looking for?

And that's usually a workflow. It's, it's, it's, you know, and that's where, uh, you know, credentials and our experience working with the enterprise, you know, gives us that experience, that insight to understand that it's not just, uh, it's not just, you know, uh, inventory management that they're trying to do. They're trying to do it to address a bigger issue.

So that's something, you know, that we've, uh, that we see, uh, as a common, uh, criteria when it comes to success of failure. So, Greg, one of the ones, um, that we used internally was our, I, I love this 'cause it, this was an technical innovation from our legal team where, um, you can imagine we were thinking about an acquisition of some sort. And so we took in a bunch of the documents associated with that and had them analyzed by an ai.

The point was to find out, um, when I, I'm trying to avoid saying we used it for search. What you're trying to look for is what kind of risk exposure perhaps comes from their contracts, or, um, this is something you do as part of due diligence, right? Or what kind of, um, risk exposure is the big one that they come up with.

But, but it could also be what kind of obligations? So you're not really doing a search, you're not searching for the word risk or obligation, right? You're saying, Hey, find all the references to where the company's obligated in some way, and then reference that back to me, or find all of the elements where these kinds of terms are used.

And then the language model will help you find ones that are similar, but aren't the exact same, right? Where a search would not get it. And so these kinds of use cases are, I think we're gonna see them just blow up.

I really do. Um, and explode. And then my, my final use case, um, so does anybody know what a low code or no code solution is?

Right? So I'm personally convinced that we're gonna see low-code, no-code solutions integrated with ais to allow individuals to get software developed for them in a much easier fashion. It's gonna, I believe, totally democratize software development, right?

There are individuals that cannot write software, I'll grant you, right? They're not capable. They just don't have that.

That's something they never learned. But they have a problem that software could solve for them. They could get an app written for them, but they just can't do it, right?

And so, if you can use a no-code solution where you just verbalize, here's my problem, and I'd like an app to do this, and it builds it for you, just think of what kinds of problems are gonna get solved. Problems that aren't solved today at all, they're simply not done, right? I don't think this is gonna take out software developers either, because I think if you want an incredibly efficient version of that, you're gonna need somebody who can write code.

But if you want the solution versus not having it at all, theis are gonna help you get that solution. So how, how, how is HPE enabling these, uh, these, you know, field solutions? I mean, uh, uh, there's first of all two categories.

'cause Egen AI is one thing. It has its own orchestration and, and, and, uh, uh, performance, you know, challenges. Uh, then there's more what we've focused on and talked about here, which is, you know, inference of, of model usage at scale, AI services.

But then outside of those, those two divisions, you, you've already highlighted quite a lot of, uh, very particular, uh, implementations of ai. Um, and, um, are you, I mean, you, you mentioned software partnerships, but there's so much activity in this space right now, so, so are you doing field validation? Yeah, totally.

So like, How are you, how are you? Yeah, yeah. We have absolutely done a couple things.

Like one is, um, we have enabled our field, uh, presales or field engineers to be able to have this conversation. I'm part of the worldwide AI ambassadors program with all the field where they are, we're teaching ourselves, right? We're conversing back and forth showing what we've learned and bringing everybody up to say, Hey, a customer wants to do this.

How can we do it? And we share that, and then we bring that back to them in very specific use case. We also have services which are designed for this, where we have people that have come in that have done it at other customers.

They come in, they actually then understand what that customer's needs are, and then build it for them and actually implement it. And then our connection back, and this is where, um, to be fair, we're just getting started on it, is that if we see that kind of same use case show up again and again, then Fabo and I here are gonna say, okay, how do we turn that into an actual kind of solution offering that we can take out more broadly, rather than having it to be onesie twosies, but you gotta have both. Right?

Now you didn't Mention the solution providers or, you know, I mean, there's, so there's three legs to the stool typically. Um, the ISV, the infrastructure and then the, the consultants or Yeah. Or SPS or whatever.

Yeah. They tend to drive a lot of the Yeah. You have partnerships there, or is, or Yeah.

Yeah. So we, first of all, we have our own services that we can take out Absolutely. And we have our own, and then yeah, we have partnerships.

So we were just talking with one of them, uh, last week. Um, one of the big, uh, consulting companies that, 'cause I think we were talking about this, the, a lot of when something's brand new, that's when the customers say, Hey, I need consulting help for this. Right?

Right. That's when they need that. And we have those partnerships to be able to work with those big consulting houses.

Yep. Totally. Yep.

So what's the, uh, magic sauce? What's a special sauce that says HPE versus somebody else, or ai? Well, I think number one, um, as vibe off led off, we've got a lot of experience with HPC Cray applications that we're bringing down to this space.

So it starts there. If you need that kind of level, you're gonna build your own models, then I believe we have expertise that nobody else has. Okay.

And how to build that infrastructure and how to operate it. Secondly, um, when you're ready for to inference, we have solution partners that we've worked with to be able to offer something that's a little more canned. Like, here's some solutions that we've vetted and the right platforms and the right infrastructure you need to deploy it, right?

And then we have the services that we mentioned before that goes out, and which is actually right here on the screen. Right? Which is, we have the services teams that could go in and help solve your individual problem.

Mm-hmm. And then that we have now thousands of service people across the world trained up on AI ready to jump in and help. Okay.

Because I'd argue, and, and I may be contrary to the rest of the, the, the group here, but I argue that, uh, AI model training is a extreme use case that is limited to the people who have their own experts. They don't need your help. They're actually with me.

They're actually people building you, right? Yep. Right.

Yeah. And it's the, the rest of the enterprise is much more going to, is much more struggling right now with trying to figure out the inferencing side of the world, or how do I, there's this thing called AI that I'm supposed to be doing. Yeah.

And how do I get that in? Yeah. And so my, you know, again, it's sort of like if we're that, that you're doing AI in the enterprise, you're, or AI at the edge, it's almost always gonna be the inferencing side of it.

Mm-hmm. Except a very extreme small use cases set use cases. So we, I mean, there was a lot of great information earlier today about, you know, what is it about the, the ProLiant platform that is great from a security, right?

Or great from a cooling power in a high density or an environment or great from the, the edge use cases, right? So if we're doing that, what I'm just trying to get a little bit more, We certainly bring all those things to the table. I try not to repeat them, but Okay.

Right. I mean, you know, we brought all that Right. Bring the security, we bring the expertise for sure we do.

Yeah. Um, we bring HPE to the table too, right. In our, in our company history and experience for sure.

And our partnerships and our relationships, you know, so we bring all that as well. Um, you know, it's interesting though. 'cause I agree with you.

I think training is real, you know, sorry. Gonna be a smaller part of the market than, right, right. Although they may represent a very large revenue opportunity, but they're absolutely.

But there won't be very, but what we're starting to see, I'll, I'll, we'll challenge this a little bit. We are starting to see people infra use inferencing to build a new model. And, um, it's not like they're building their own large language model, but they might be building a classifier, right?

Or they might be building something a little more purpose built, not a little more, a lot more purpose built. And so we are actually starting to see what I would say traditional enterprises train little, teeny, tiny models for purpose, right? But they're not gonna be the big giant, right.

Turning my own 400 billion model, right? Yeah. But also in those use cases, which I think is a very interesting use case, is that a two to five racks with a hundred kilowatts each rack type of environment?

Or is that two to three servers at four kilowatt servers? Something like that? Yeah.

More, more like that. Or, or, or maybe we'll see, and I think this is where, I mean, we fully embrace the hybrid world, right? Where you might train something in a cloud because you need actually the a hundred kilowatt of rack server set up, but you just don't need it forever.

You just need it for a month to help you build it initially, and then you bring it OnPrem to inference it, right? So we think we'll see that too. So one of the, oh, go ahead.

That have traditionally plums enterprises is the fragility of AI hardware. The NVIDIA's kind of step forward to say, Hey, here's a prescribed method for making AI infrastructure more reliable from that software layer layer, you know, changing. You're always, and I, I think part of this argument gets, uh, encapsulated as part of the big training model, uh, argument.

Those folks are trying to eke out as much performance as possible. Buying an entire network will save them, will increase efficiency five by 5%, which pays for the network. That doesn't necessarily scale down to the enterprise.

Like, we're not, if our right U clusters are at 30%, we're probably pretty happy. The, that, that's pretty high utilization for the enterprise. So, but one of the things that they do complain about is environments keeping, you know, kind of the same problem we have HPC keeping those environments stable because they're so specialized.

How's HPE helping in keeping the, uh, software layer of the, of the AI infrastructure across your portfolio, consistent across devices and stable? Yeah. So I think, you know, the approach that we are taking is that of course, right?

I mean, especially, you know, starts with the ISVs, right? And, uh, uh, to a previous comment, right? I mean, uh, right now it's too early to pick winners and losers, right?

I mean, the ecosystem is just exploding, right? And, uh, you know, the other thing that we're seeing is, uh, well, a lot of the capabilities that an ISV started off building is now are open source, are, you know, part of NVIDIA Stack, right? Or it's part of Nvidia Stack, right?

But the underlying theme that remains constant, right? I mean, you know, as we look at how this is gonna evolve, is, uh, and really, you know, uh, due to a previous discussion on, Hey, what, what's so special about HPE? I'd say it's three things.

It's expertise, it's experience, and the portfolio. Uh, right? And that comes for a lot, right?

Because I mean, when I talk about expertise, it's our credentials with HPC, uh, it's our credentials with more than three decades of building enterprise last servers with the proline pedigree in the portfolio. Uh, and it's the experience. I mean, we've, uh, we've seen a lot.

We've learned a lot, uh, working together with customers, uh, and then of course, uh, the portfolio, the breadth, the richness of the portfolio. Now, when it comes to the software stack, of course, uh, right? I mean, it depends.

It's gonna depend on, um, what's the exact workflow or the outcome the customer wants, right? So we are not trying to force that a particular stack onto a customer, which ends up being overkill. So that's why, you know, whenever we talk to customers, uh, yes, you know, the use cases are typically something that, uh, a part of the opening conversation.

But we try to stay that conversation towards, okay, what, what, what is it that we are ultimately trying to do for you? What's the exact problem we are trying to solve and take a holistic view, uh, right? Like, you know, when we were talking about edge, uh, you know, and comm uh, I mean, perfect example.

So somebody using AI at the edge, uh, you know, we talked about how, uh, it's gonna be, uh, the c right? Having a single or common way to manage your entire estate, including edge devices to come. I think that creates that, uh, common thread of like, you know, how do you go about managing Now if you're talking now about like, okay, what's the, you know, all the different libraries, the tooling, uh, well, hard to standardize that, right?

Because, well, We, well, one thing in like private cloud AI is that it is Yep. You know, 'cause it is our stack, it does include HPE AI essentials. Yeah.

Which does, in fact, one of its main jobs is to keep everything up to date. So it's nv uh, Nvidia AI enterprise software is in there, but our AI essential software keeps everything up to date, including operating systems, including the, all the infrastructure, soft software infrastructure, uh, from Nvidia as well. Um, that, that, 'cause again, it's DevOps for ai, right?

So it includes all that, The king of appliances, the, the, you're taking the three eighties and the three sixties and the portfolio and remixing and putting, adding value on top, and then selling that as an appliance, as a customer. I bought it. I found value in the fact that I'm not worried about firmware updates.

Yeah. And, and compatibility, because HPE has figured that out for me, and I just need someone to do the engineering and then roll it out. So I, I think your, uh, call out of, you know, the HPE private ai, exactly what a lot of customers are looking for.

That ability to just say, I just want, I want the infrastructure stack. I want somebody to worry about the engineering of that. And then I'll put my apps, which AI is really complic, that's still by itself, that's still enough complication for the customer to handle.

Well, And, and an important part of that, and this is sort of what I was digging at, is that outside of the training use case, customers have many different metrics of what performance is and is not always flo terra flops mega flops. Right? Right.

Yeah. Right. For sure.

In fact, if you look at got, uh, AI at the edge and inferencing at the edge, that's probably the, you can get all the performance you want in terms of flops, but that's not important. What's important are things like latency management, right? Security, all those other things.

And to be able to do that as an appliance, right? Which is from, from my perspective, AI is just another workload. It's a workload that has slightly different characteristics than OLTP, right?

But it's just a workload. And if you look at it that way, and then you say, okay, what are the characteristics, workload, what do I care about? It's the exact same thing as OTP or something else.

I still have compute requirements, memory requirements, storage requirements, networking requirements, power, heat, cooling, right? Connectivity, all of those things. And the, the HP way of building a server ecosystem rather than a server is a very strong and compelling argument.

If you can put that in the context of we do, we've been doing that for a while now. We do that for your AI workload. Got it.

Well, you, you can say AI is just a workload because AI is a workload. But, uh, the, the first of all, um, the variety of, um, AI applications in the, in high, very variable stages of development and use experimentation on the one hand versus, you know, you know, well understood trained models that, uh, at inference at scale, um, their usage patterns vary. Their ingest and Egress needs are different.

Um, the size of the models is what gets focused on a lot that that puts you, you're right, you know, you're still supplying memory, storage, you know, communications and all those, uh, resources to a workload to an app. But, um, you know, when you're thinking inference and scale, what you can store on board in, in, in, in, in memory versus, you know, uh, um, uh, what you might for a large model versus for a smaller model and how that's shared amongst the cores. And if you're using CPU or combination C-P-U-G-P-U, I mean those that really does make ai, um, uh, different and create and, and present, um, particular problems in a virtualized environment, in a container based environment.

I mean, these things all interact in funny ways to produce really funny results in production or really funny issues through the lifecycle. Or even Mobile. Even mobile like Louisville Qualcomm's doing, right?

They're doing so much work, even getting stuff onto these things. Right. Well, That, that's why there's so much offboarding from the CPU of, of, of various various, uh, uh, duties.

Uh, so I'll, I'll, I think maybe that's a good argument, uh, argument for the, the, the round table maybe, right? Yeah, fair enough. Fair enough.

I, I'm trying to get back to, sorry, I was, it was, it was, it was, to get back to this idea of how these particular proline units and the ecosystem that you guys have built around it help enable all of That. Yeah. Well, and that's, that's my point.

That's really my point. Looking at a modern data center, you have outcomes. You have engine that drives those outcomes, but you have the manageability and performance and framework that are fundamental.

HP provides the r and d budget to ensure that the stability of the foundational for these solutions are present. In my opinion. I've seen copycat system vendors go, oh, do that too.

But then when you dive down deep, it just falls apart when it scales. So, I mean, I have to tip the hat off to HP's team for this is, you know, how many generations that we have of ProLiant and the lessons, every, every generation gets better by leap and bounds, not just a little bit. And so I think it, it cuts new ground, new, new areas that we don't think about, but fundamentally we need to, we just had the discussion with power cooling, heating, you know, all that kind of stuff, and how it tee tees into this and this stuff, and how the framework working tightly within v how many, uh, how many other vendors are out there being able to do that.

And, and that's, that's what's driving the modern application containers. It's not the high, it's not the VM anymore, right. So, so definitely it's a different brave new world.

Yep. All right. Now I think, uh, we are near the end of, uh, sessions.

Uh, Scott, I'm gonna hand it off. Well, So I mean, we kind of hit on this a little bit, Jack when we were talking, which is that, you know, when, when it comes to ai, right? Why, why HPE, right?

And our kind of view is first of all, you know, we do have the specialized platforms designed for all kinds of different AI use cases. Whether they're at the edge, whether they're in the data center, whether they're training, whether they're inferencing, we do have the platform capability we bring that they're optimized as we would say. We have our, uh, management capability, highly automated environment designed with both cloud-based and on-prem management solutions based on what, what you need as the customer.

And of course, it's secured. Everything we bring to all of this is all based on our, um, you know, we consider to be world class security technology. And all of that is present.

All of that is part 100% part of the solutions all the time. I.