Good morning everyone and welcome to the Cube's live coverage of NetApp Insight 2026 here in Las Vegas. I'm your host Rebecca Knight and I'm sitting alongside Kristoff Bertrand who is of course an analyst in residence at the Cube Research. I'm excited to be doing this show with you.
Great to be here. Me too.
And our first guest needs no introduction. I would like to welcome back to our show George Kurian who is of course the CEO of NetApp. Welcome, George.
Good morning and thank you for being here with us. Good to see you. It is always a delight.
Let's start talking about AI factories because we're hearing a lot about them. They're almost the oil refineries of the digital economy. If to strip away the buzz for a second, can you talk a little bit about what problems customers are solving with AI factories?
AI factories are accelerated computing clouds which have large amounts of GPUs that are being used today to train these large language models, frontier models, but over time will be used by every organization to derive knowledge and insight from their data. They have unique demands. You know, GPUs are extraordinarily expensive, so you don't want them to sit idle. They require, you know, solutions that are power efficient because of the constraints on the energy infrastructure, and they need to have operational efficiency because you've got thousands and thousands of GPU servers. And increasingly, data and data infrastructure is the bottleneck because to use the GPUs efficiently you need to serve them huge amounts of data that can be processed.
Right. And we could be talking about trillions of files. So definitely there's a scale conversation. So on Friday you made an acquisition announcement, PKIO, and you just introduced yesterday Novice, which is a brand new architecture designed for AI factories. So tell us more about this. What is sort of behind, what's the inspiration behind it, and maybe talk to us about AI factories and the market, the operational efficiencies a little more, because clearly these two in combination are changing the game for NetApp.
Yeah, you know, GPUs have unique data requirements. They process data at much higher rates than CPUs. Typically CPUs are around 40 gigabytes per second, and the GPU can start anywhere from 200 GB per second and now reach one terabyte per second. And to train models faster, you want to aggregate more GPUs so that you can process data faster, and you need very, very large data sets that can be unified under one construct called a namespace. PKIO is a company that we have been working with that had very unique technology for the metadata layer. And what we introduced together with them at our conference yesterday was the combination of a very scalable, proven data service with ONTAP together with a disaggregated, independently scalable metadata service and a very, very, very large namespace. You know, we are the first storage solution to reach 100 terabytes per second in performance and to be able to aggregate multiple zettabytes of data into a single namespace. And this allows everyone from the largest frontier models in the world to the largest sovereign clouds as well as, you know, high performance computing environments to be operated way more efficiently, both in terms of using the GPUs and to be able to be operating efficiently as well as power efficiently.
Just hearing multiple zettabytes of data is making my head spin. I want to talk about AI experimentation because it seems as though there's a real shift in what's going on right now, where so many companies were all about the pilots and trying to get, frankly, employees to embrace and adopt AI, but this year there seems to be a shift in the conversation. What are you seeing as CEO?
Yeah, I think the original promise that large language models and reinforcement learning created was the idea that you could move intelligence from just a small part of your data to all of your data, and to move the tools to access that intelligence from a small part of your employee base to everybody. Right? So if you look at an organization historically, about 15% of the data is structured and it's fairly kind of, you know, sanitized and doesn't have a lot of context and sentiment associated with it, and it was represented to a small part of the organization, typically about 11 to 15%, through what's called BI tools. 85% of the data in an organization was essentially untapped unstructured data, and with conversational and natural language capabilities, pretty much anybody in the organization can have access. What we've seen is two things, I think. One is a real recognition that data is the foundation for the intellectual property and the strategic moat of an organization. Everyone can have access to all the, you know, the latest models. And the second, beyond the fact that data is sort of the intellectual foundation, is the idea that you need to build kind of solutions that allow you to scale from a curated one-time pilot into production. And so we are stepping up to help solve those challenges.
Well, let's talk about the one-time pilot because we're past that point now. Clearly, there has been a shift in the market where it's not just projects; it's becoming more a question of existential implementations from many enterprises. At the same time, we are also hearing about some risks with agentic AI becoming a bit crazy, candidly — agents making decisions without people or sometimes with people, models escaping, etc. So I think there's a lot of hype around this. From your standpoint, as you think about the market and what you're seeing from customers, are we now in phase two of what we could call AI adoption in the enterprise?
That's right. I think we see clients moving from very constrained pilots to business process and work transformation using AI, and there are important new considerations. Right? I think 93% of organizations that are trying to embark on that journey cite data as the number one challenge. You know, it's easy to curate a small subset of data in batch for a one-time use case. When you want to make AI scale across your enterprise, you need to integrate it into the operational systems and the workflows and how work gets done, and that creates, alongside the risks you mentioned, new requirements on the infrastructure — right? Governance, security, control, and so on. I think to deal with the risks around AI safety and security, you know, every part of the technology provider and the operators and service providers have a role to play. You know, the model builders have to step up to make their models safer and to put harnesses and guardrails around the models. The defense-in-depth security vendors, including people like us, have to build more, you know, modern architectures that keep pace with the fact that agentic attacks can happen in an instant, and the enterprises and operators of their service need to make sure they're adopting the best practices so that the users of their service can trust that.
Right, which is why I think the Novice architecture, the upcoming acquisition, and other innovations around intelligent data make a lot of sense. Because for AI to succeed, whether it's for one process or for all processes, you need clean data at scale. So everything's fun and games until you add 'at scale' at the end of the sentence, as I like to often say. And it feels to me that the architecture is now sort of meeting that promise of, you know, originally what people wanted to engage on with projects, just they couldn't scale. So scale across data in terms of just storage — billions of files — but scale also in terms of consistency and compliance of data. So definitely a very interesting time for NetApp because you've been doing a lot of the other components you mentioned, like security, replication, etc., snapshotting, for a long time. So it feels like it's a perfect convergence for you at this point.
