NetApp Insight 2025: AI-Driven Data Management with Arun Gururajan

We are here at NetApp Insight 2025, and we are learning about some of the new capabilities NetApp just announced. Now, one of the things that's been very important in the storage industry is moving beyond just storing data and into data management, data classification, data protection. That doesn't mean necessarily that, uh, NetApp is, you know, going to be wiping out the market for these other products.

What it means is that NetApp is going to use these capabilities to make their products better. So I am here with Arun, who is the, uh, chief, uh, VP of, uh, data science, right? Yes.

And so, Arun, talk to me a little bit about how NetApp is bringing data science and AI into the storage platform. Awesome. So I'm gonna talk from a data scientist perspective, because that's what I've been doing for the past couple of decades.

Uh, one of the things about, uh, data science is that data scientists spend about 80% of their time wrangling the data, right? Like, only the remaining 20% is all about the model building and deployment. And so making sure that we are able to effectively discover the data, understand the data, classify the data, is paramount to good data science.

And this is exactly like what we at NetApp want to help, uh, solve for the data science personas. And specifically, uh, you know, NetApp is in a wonderful position because this is a platform that is at the heart of many enterprises already. Essentially, uh, NetApp's products are already, uh, housing most of the unstructured data at many of these customers, and yet it's unstructured and the customers don't know much about it.

How is NetApp helping with that? So, You're correct. I mean, most of the data that's housed in NetApp is unstructured data.

And the biggest challenge with unstructured data is understanding because there is no schema for unstructured data. So what It's part of the definition. It's Part of the definition E exactly.

Like, so what we want is to extract semantic meaning from the unstructured data and put some structure over it so that we can use it in upstream AI or downstream AI applications. So what we do is we take the data, data and then we essentially extract embeddings. Embeddings are nothing but a stream of, uh, uh, numbers, we call it as, uh, embedding vectors.

And what they do is they essentially provide, uh, meaning and context to, to unstructured data. And this is provided out of the box with NetApp's AI data engine, so that AI applications can be rapidly built downstream. Now, for a long time, storage systems have been able to do indexing, essentially.

Uh, we read all the text in this file, and we can show you which file contains this string or that string, or a combination of these two strings. But you're talking about multimodal data. You're talking about data that is not necessarily text files, and you're talking about using it in a way that is not just a traditional index, right?

Correct. Uh, so this is essentially extracting a stream of numbers. So think of it as a vector.

And this vector actually has some meaning, uh, that essentially relates the underlying data to a set of like other entities around. So for example, uh, you can take the word, I mean, this is a very classic example. So you take the word king, the word king will have a certain embedding vector, and the, the word queen will have a certain embedding vector.

And in a multidimensional space, these vectors will be close to each other because both the word king and queen represent royalty. Right? So the semantic meaning is extracted with the, with the help of these vectors.

So yes, uh, when whether we talk about text or whether we talk about images or video, like these stream of numbers help provide semantic context to the underlying data. And NetApp is doing this, uh, autonomously, basically by examining the data, um, using software, using AI to extract this meaning from the data, again, not just keywords, meaning from the data. And then that can be then presented to an AI application that can use that to, uh, provide some, uh, end user benefit.

But you're not running the client's application. What you're doing is you're, you're, you're preparing and presenting the data to that application. That's exactly right.

And that's why we call ourselves intelligent data infrastructure company because you have all this raw data underneath. And what we wanna do is we wanna extract the intelligence from this data and provide it in a layer above. This could be embeddings, this could be context, this could be like knowledge graph that con that essentially connects all of these different entities together.

So in some sense, we are kind of like providing all of the intelligent data attributes so that they can be readily consumed by an AI application that runs on top. Now, I think anybody listening to this is gonna say, wait a second, you're gonna take all my corporate data and present it to an AI application. Are you crazy?

You're not crazy. That's right. Uh, so this isn't, this shouldn't be an afterthought.

Uh, that's why as we heard in the keynote today from George, uh, this is not something that was bolted on, but rather, uh, these kind of like security guardrails are built into the AI data engine. And this is possible through our classification technology, uh, which we can speak more about, which essentially extracts, uh, metadata content-based metadata and provides the right set of guardrails so that your AI application never exposes anything that is sensitive or that is IP for your organization. And that, that's one of the things that we've been hearing a lot about, that essentially there are, there's really no way to truly control what a generative AI application is going to do.

People are constantly trying to, to jailbreak these things. If it has access to that data, it could theoretically, um, present that data. And so what you're saying is NetApp is proactively preventing a, you know, applications from seeing data that they shouldn't see based on intrinsic aspects of that data.

And also based on, uh, sort of an, uh, an intelligent, uh, statistical formula that, that allows you to detect information about that data that may not even be obvious. So it doesn't mean, oh, well, none of the, you know, credit card data is gonna be exposed. It means no data with credit cards or no data with financials even is gonna be exposed, right?

That's right. So, uh, just to kind of like paraphrase what you said, uh, once the data goes from storage into a vector database, it's only a matter of time before the right malicious prompt extracts that, uh, sensitive data out. So what we are doing is we are tagging, and, you know, like enterprises have terabytes, hundreds of terabytes of data sets, uh, which could contain sensitive information, and they have no way of figuring out which document has what entity in it.

What we are doing is we are built in like, uh, advanced, uh, AI classifiers that essentially tag your sensitive data and essentially build guardrails on top of them so that a data compliance, uh, uh, officer or a data sec, uh, protection officer can come and set the right policies for their organization, which can limit what is the, uh, type of data that is ingested into the AI application. For example, if they say that, Hey, no data containing credit card should ever go into the vector database, then so be it. That won't happen.

Right? Even If somebody made a copy of something and stored it in the wrong place, correct. Or Whatever.

Correct. Exactly. So, because all the, all of the protections that we provide, they go along with the data.

So it doesn't matter whether you copy the data, all of the protections carry along with the data. So you, you can be sure that once you set the guardrails, the only the right data goes into your AI application. And this is even true for multimodal data.

Again, so we are talking about video and images and things like that. You're also classifying those, right? Correct.

So this is just, just, this is not just for text. And again, like when our a ID uh, releases, like, you'll see that this essentially is covered for multiple modalities, right? Like, because we can essentially extract, intelligently extract information from all of the different modalities and make sure that we provide appropriate guardrails for them.

Yeah. Which is really important because I, I mean, essentially if we're trying to build, uh, intelligent applications or artificially intelligent applications, we also need to have intelligence embedded in the supporting systems that allow us to be flexible and responsive. And of course, this same technology is useful in terms of protecting the company.

Uh, cybersecurity is a big issue. Uh, attackers are trying to get in. Um, they are absolutely attacking storage systems these days.

How does this protect from cyber risks? Correct. So the biggest threat for enterprises is ransomware attacks, right?

And ransomware attacks come in two forms. Uh, they could be like read based random, uh, ransomware attacks where an attacker comes in and tries to read the data and exfiltrated out of the organization. There could also be right based transmor attacks.

What I mean by this is that once the data attacker has exfiltrated the data out, they would then go ahead and encrypt the data so the data becomes unusable for the organization. At NetApp, we want to essentially be the last line of defense for cyber attacks, because once the attacker has reached the data, it means that they have bypassed all of the security defenses in place. So we want to be the last line of defense where we provide the, an AI based, uh, real time, uh, threat protection against both these read based ransomware attacks, exfiltration, which we call as data breach, as well as write based ransomware attacks, which is data encryption.

And all of this is, is absolutely critical because again, these are things that are actively being exploited right now because ai, uh, potentially opens up new avenues, especially, uh, agentic and, um, rag applications that actually have access to the underlying data ne potentially. And so if the storage system can't do the things that we're talking about, if it can't, um, intelligently classify data and intelligently apply, uh, standards and protections to that data, then that is really going to be a, a, a risk to the company. Absolutely.

AI essentially increases the attack surface for your organization because, uh, as you said, agentic AI can access not only the wrong, uh, data sets, but it can also access the wrong set of tools that are right now as we speak. There are thousands of MCP servers that have been published, and some of these MCP servers have tools that are malicious, right? Like, how do we know, uh, that the agent AI can access the right set of tools?

So it's very easy for a malware to come inside the organization. So it becomes paramount that, uh, uh, uh, CISO for any organization thinks about the different layers that, uh, protect an organization against security attacks. And most importantly, they need to focus on the data layer, the data security layer.

And this is where NetApp provides its ransomware solution for preventing, uh, against the, these types of harms. So I'm sure that people are listening to this and they're thinking, this sounds great. Um, NetApp must be charging extra for this, or it must be something that is still in the lab.

Uh, but that's not true, right? That's right. So it, it has been public preview today, so you can absolutely try it out in our ransomware resilience offering.

And, um, uh, is this included with the standard NetApp, uh, ONTAP license, or is part of this, uh, optional? So the autonomous ransomware protection with AI is included as part of the standard on top offering, but the ransomware resilience is essentially an add-on offering. And in terms of, uh, the support for, uh, AI applications with MCP and RAG and all that sort of thing that they talked about today, is that also an add-on or is that part of the standard?

So That's, uh, included as part of the ransomware resilience, uh, package because our, uh, scanning, uh, malware scanning and all of that is part of the ransomware resilience offering. Okay. Great.

Well, thank you so much. It's, it's great to learn a little bit more about what adapp is doing to make a truly intelligent, uh, data platform. And I can't wait to hear more about this.

We'll be having some more details at our Tech Field Day presentations. We'll put publish on, uh, Thursday, uh, as well as of course on YouTube if you're watching this after insight. And, uh, we will have a lot more reactions from our delegates and staff here at NetApp Insight as well.

com.