From Zero to AI Hero with HPE Private Cloud and Storage Solutions

Good morning everyone. My name is Ed Bove. Uh, I lead product management for unstructured data for HPE, and we wanted to transition a little bit more.

Uh, as, as Brad mentioned in the beginning, we've talked about a lot of sort of the infrastructure and the application capabilities. Now we're gonna get lower into the stack from a hybrid cloud, um, perspective. We'll talk about, um, building storage infrastructure for tomorrow.

So, I just wanna start off with this quote, right? We think about what's happened from IDC. This paper just came out, less than half of the AI pilots advanced into production, right?

Not because of algorithms, but because of data. So the key observation here is that the AI ambition has not really matched the execution, right? And we think about like what's happened.

The organizations have gone out and they bought GPUs. Uh, CEOs have said, we will do ai, right? Um, but it actually, you know, hasn't fully materialized that way into production.

So why is that, right? We'll get into some of that, um, detail. But what we see is that, you know, from an unstructured data perspective, we've got a lot of different factors that that influence, uh, the value here, which is we've seen explosive growth of data.

We've seen sort of the distribution of data, but it actually, you know, comes to, you know, having a mindset shift, which where you start to think about data, not storage, but data is the product. And that's really core to generative ai, rag agentic ai. It really starts to, you know, show up in the infrastructure that you've got distributed infrastructure.

It's spread out everywhere. You may not be aware of what you have from an enterprise perspective and AI to build good ai. It depends on having this fresh, trusted data that you can effectively curate, that you can govern, that you can manage.

And I'm gonna talk about sort of the, the right hand side of, I'm sorry, the left hand side of the slide, which is, you know, what is HPE doing? We've taken our electra storage, MP architecture, a disaggregated architecture, and there's really two flavors of that that you can deploy. And you can manage this with our cloud operating model, with our cloud control plane.

Uh, and we, we've got the B 10,000 for mission critical performance. And I wanted to spend a few minutes today talking about unstructured data, accelerating this data for ai. And this is a new product that HPE launched last year called the X 10,000.

So the X 10,000 is a capability, it's a storage capability that allows you to manage and work effectively with data to get that data AI ready. We've got a number of key tenants, so I'll just walk through, um, some of those tenants. Um, but what we've focused on is delivering the capability of fast performance, right?

To do ai, obviously we need to have a very performant architecture. We need to be able to simply manage, so simplify the management and understanding of the data to manage that at scale. Um, we want, we obviously need to have enter enterprise grade feature sets and capabilities.

And what we're doing is we built this new architecture and we built it on a log structured key value store. And on top of that, we built namespace and we built an object native namespace. And why that's critical is that we're gonna deliver first class services into that namespace.

And we've also added capabilities for built-in data intelligence. And what I would say our perspective from HPE is that for too long, you know, the storage has just been about writing the data and hopefully reading it back. And what we want to do is we want to go beyond that.

We want to go beyond that to, um, help our customers gain insight, understand what they have from a, from the asset perspective, start being able to think about effectively data as a product. And if you have strong knowledge of that data, if you have a great understanding of what are the insights that you can gain from data, what we're seeing is that unstructured data is becoming much more like primary storage performance is critical. It's, it's critical to really understand, um, the metadata.

It's critical to understand the governance of that data. And we're seeing, you know, systems that are being built, uh, on top of that capability. And we're, we're focused, uh, in a number of key areas.

And I'm just gonna highlight a few things that we're doing. But, uh, we're partnering with nvidia, uh, from an AI data platform perspective, but also from a performance perspective. And we, we just announced that we officially launched in GA, RDMA for object.

And what we're seeing from a market perspective, as many customers when they deploy and they deploy their AI projects, they're doing it sort of object first. So it's critical to be able to innovate there and offer sort of a new capability. And this RDMA for object is new capability that hasn't been available before it's been available on the HPC side, you think about file, you're using RDMA for file, but this is RDMA for object.

And what we're doing is we're getting the protocol sort of out of the way, and there's tremendous benefits of doing that. We see a near doubling of the throughput, right? We're also seeing massive reduction in latency and a massive reduction in the sort of the host CPU comp, uh, uh, consumption.

So again, when we talk about why does RDMA matter, it matters because we want to enable our customers to gain these insights faster than they've ever done before. And we need to start bringing that intelligence into the software and the infrastructure stack so that you can maximize that, those resources, so that our customers can build, um, on this infrastructure and build on that knowledge that they have in their environments today. Um, so again, massive improvement here from a, from a protocol perspective.

And there's tremendous benefits in terms of, uh, getting, uh, this, uh, you know, optimizing the use of this protocol. So when I said I wanted to provide con more context to data, what do I mean? You know, existing systems are gonna do things like give you, you know, sort of the descriptions, the time, the size, that sort of thing.

But we're talking about is really starting to transform that unstructured data perhaps into some structure, but also understanding, you know, what is, is it visual content? What is the type of data? Is there a lightweight model that we can run that where we understand the tone of that data, um, multimodal correlation, are we able to extract vector embedding?

So if you're running a an LLM, can you do that lower into the storage infrastructure layer to speed the performance of running agentic ai? So is that data AI ready? I would say today's systems weren't really built to have, you know, to take that metadata and make it AI ready.

And what that allows you to do is it allows you to very quickly curate those data sets. It allows you to know what you have and some of the capabilities that we're bringing really help to improve that. So, um, things like metadata tagging, but you might say, yeah, there's metadata tagging today, but what I mean is metadata tagging that goes into an iceberg table that's in an open table format where you can run SQL on your data to find what's there to curate that data.

And we've done it in a very open way, um, again, where, we'll, you know, if you have that infrastructure to run, um, SQL types of models, you can do that. If you wanna run vector embeddings, again, we're gonna enable that. We've also are enabling interactions with LLMs.

So one of the things that you can do is you can, as the data's flowing in the system, you can extract vector embeddings, and that helps from a security perspective as well, because now you're securing the data from that. With that Im policy all the way through, whether that's the data stored in the environment now when it's stored in that vector database as a vector, we're also protecting the access there. So the LLMs only have access to data that they're authorized for.

And what we also announced is integration with, with an MCP server. So model context protocol, what does this mean? This means now that ag agentic AI or AI can get insight into what is in the storage system.

And our plan is to make that even more valuable, because if we're extracting the metadata and we're making that searchable, guess what else can search it, right? Our, our agents can now search it. And then there's this new type of unstructured data, right?

We've got semi-structured data, which is sort of machine generated, but we've got AI generated data and this is going to just explode. Yeah. Questions.

So Your MCP server is running inside the actual storage operating system? Yes. And it, so it's just to be very clear, we've announced it, we're gonna ship it with our next release, it's not act, this is how we will implement it.

Yeah. So, so you'll have agents that will, are they gonna be controlled like they're agents that you've written or is gonna be user written agents that would run there? How, um, how extensible is that?

And I'm just thinking about it from a security perspective, running inside such a trusted for, You know. Yeah, no, it's a great question. So certainly we'll focus on sort of the enterprise grade guardrails around the utilization of MCP, but the way I like to think about it is you're gonna see MCP servers everywhere, right?

And what this will allow you to do is I can interact with that MCP server to see, okay, what's the list of buckets that are on my, uh, in my environment, in my storage environment? And then once we started to unlock that data, maybe you want to curate a data set and have a better understanding using the metadata that's now exposed. But today, that's opaque, right?

Like when you put systems in storage, typically customers have to buy another system and extract that metadata to under unlock that value. And we wanna make that value easier to unlock, um, and easier to be accessible to agent ai. So back to the metadata tech.

Yes. Um, so from what I'm hearing, so now you've built like a data store to store all the tagging. Do you have tools and solutions that help people determine the metadata typically that uses AI now or something like that?

Yeah, so I mean, I think you, you might see MCP in that, um, in that in, uh, type of example. But what we've done is, the way to think about it is, um, oftentimes with systems with object, you, the metadata is sort of attached to the object. Yeah.

We are actually, um, changing that. It can still be attached to the object, but what we are doing is we're taking something like Parquet, we're putting it into Open Table, we're making that open table accessible through an API, right? And any open table environment.

So if you have Trino, if you have, you know, Starburst data, if you have Dremio, if you have any of these solutions, do you have Spark? We'll help you with Spark to run SQL on this. So, so again, this will be very open is, is how I would describe it.

Yeah, I get it. That's perfect. Tables work for me in my world.

Good. Awesome. The whole reason they're not tied to any tool, but back to how we capture, govern and protect this data, will you ha will you have tools to help people do that?

Or will they be, are you seeing that they will be using their own data cataloging metadata management tools that just reference this? Yeah, That, oh, that is such a great question because what we, you know, obviously there's a massive market out there around metadata management and data management. Yeah.

And I think for too long, storage has sort of just said, yeah, we'll just absolutely. We just do the storage and anything to do with data you guys do. And by the way, customers have to invest there to do it.

And what we would say is it's much more efficient to have it closer to the data. So I would say ultimately, one, we wanna bring that compute closer to data. Two, we want to give our customers sort of facilities to management, and we want to do it in a very flexible way.

Could be running some data pipelines, it could be within this OpenTable construct, you know, providing the right level of governance, um, in that. And it will also be from an HPE perspective working with solution providers. So working and being open, that open table format, in my view, is the right, the right approach because it makes it, you know, sort of vendor agnostic, right?

And now you can curate that data very, very easily. Right. I get that.

Okay. So it sounds like to me that you're not planning right now today to build any of these data, cataloging your data management tools, but you're providing the metadata that, that that organization has determined about, like there's technical metta. That's Right.

Yeah, it's a great, so I, I think we will likely work with solution providers and we'll work with Trino from an open source perspective. And I think, I don't, you know, again, could things change? Sure.

But I think today we want to enable the solutions and provide the, the, um, uh, the, the lens that we have is let's make the AI ready data accessible to solutions and applications and then enable that. Okay. Um, but I think if you were a customer and you're deploying, you could use open source today, you could build your own solution effectively.

Great question. Jim. Rinky.

CDC. Yes. Could you go back, uh, oh, sure.

One other thing, um, a little curious to pick up from Karen's point vector embedding, right? Yes. Obviously that implies quite a few different things, right?

Yes. Like choosing the model that you're gonna use to create the embeddings. Are you saying that Electra would actually look at the data or look at the metadata stored somewhere inside the system and go, ah, this is mainly visual, so use this model, this is mainly text, so use that model, et cetera?

Yeah, great question. So what we have today, and it runs separate today, I'll just be very clear. We have a host-based solution that we will bring over time into the, into the array, if you will.

Um, and what you can do is you can think of a bucket, not just as a container, but as a pipeline. So we will enable that pipeline and have flexibility there. And you can run, you can imagine you can run a variety of different models.

Now today, we'll provide some encoding, uh, and we'll sort of consume the metadata and push it into OpenTable. And then over time we want to be much more flexible and pro provide support for customers to bring your own model, as you would say. So in, to take your example all the way end to end, I would say, look, we could sample that data just like you do a DNA sample and then determine, this is the type of data.

We might wanna run this model, extract some metadata, and now I've got it accessible. Now I've sort of unlocked the value, I can curate that data. And we envision the customers will like to run their own models over time, and we're gonna bring that, that entire environment we call our data intelligence server.

And we're bringing that closer into the cluster over time, sort of within the fabric and then within the actual array itself. That's, that's our sort of long-term vision. And one follow up question, so Sure.

At the, the vectors would be stored in some sort of vector data Vector database? Yeah. Right now we Customer choice or your, Um, it could be customer choice.

Today we are using vis mil, uh, and we are working to improve. I, you know, I, I can't talk about it right now from a roadmap perspective, let's just say you've heard me talk about performance and performance being critical for, uh, AI pipelines. We're absolutely focused in that area.

Uh, so you can imagine things that you could do to, uh, optimize both the vector and data embedding as well as sort of the AI inference processing. Okay. Yeah.

Thank You. And delegates, we're at time. Ed, you have a couple of minutes just to finish up please.

Yeah. Thank you so much. I'm going to just, uh, really quickly, um, in a minute or less, what we want to focus on is helping customers gain intelligence around the entire sort of ecosystem of unstructured data.

As everyone knows, 90% of all new data that's created today is unstructured. There's tremendous value there. We want to help you have intelligence around it, but also visibility with our data fabric capability.

You can see not just within HP's ecosystem, but it's really truly a heterogeneous view of all of your unstructured data. Once you've got visibility, then you could start to do interesting things. We think we've got a great platform, the electro storage MP platform, uh, to help you take advantage of your unstructured data.

Uh, with that, I know we're out of time, so thank you so much. Uh, and I'd like to extend my thanks for the entire HBE team, uh, to the delegates today. Really appreciate your time as well.

Thank you. I.