Inside Cisco’s Secure AI Factory with NVIDIA: Turning Enterprise Data into AI Fuel – Six Five On The Road from SC25

How can enterprises transform fragmented data into secure, AI-ready fuel while advancing from pilot projects to production-scale AI?

‍

From Supercomputing 2025, host Matt Kimball is joined by Cisco's Nicolas Sagnes, Product Marketing Leader, for a conversation on Cisco’s Secure AI Factory with NVIDIA. They explore how this security-first, full-stack solution enables organizations to turn scattered enterprise data into governed, production-ready resources for AI—highlighting real-world applications for RAG (Retrieval-Augmented Generation), agentic AI, and best practices for secure AI scale in the enterprise.

‍

Key Takeaways Include:

‍

🔹Bridging the AI Gap: The Secure AI Factory helps enterprises move beyond pilot mode by providing a production-ready, security-first framework for AI deployment.

‍

🔹Real-world AI in Action: Concrete examples illustrate how RAG and agentic AI are delivering measurable business value today.

‍

🔹Data Unification: Strategies for connecting disparate data sources to create AI-ready data—without requiring rip-and-replace architectures.

‍

🔹Secure-by-design: Daily impacts for CIOs and IT teams, with a focus on balancing governance, safety, and speed as AI scales.

‍

🔹2026 Roadmap for Progress: Where organizations should focus now to future-proof and secure their enterprise AI initiatives.

‍

Learn more at Cisco.com

‍

Watch the full video at sixfivemedia.com, and be sure to subscribe to our YouTube channel, so you never miss an episode.

‍

Or listen to the audio here:

‍

Disclaimer: Webcast is for information and entertainment purposes only. Over the course of this webcast, we may talk about companies that are publicly traded, and we may even reference that fact and their equity share price, but please do not take anything that we say as a recommendation about what you should do with your investment dollars. We are not investment advisors, and we ask that you do not treat us as such.

‍

Matt Kimball:

Hi, my name is Matt Kimball and I am here for another episode of Six Five On The Road at Supercomputing25. I have with me the famous Nicolas Sagnes from Cisco. Nicolas is a product lead or a marketing lead for all things AI infrastructure at Cisco. And today we're going to be talking about going inside Cisco's secure AI factory. Nicolas, thanks for being here. Great to have you. Thanks for having me. Yeah, this is going to be fun. So MIT, I think I've referenced this study about a million times in the last few months. MIT released a study a few months ago, talking about the challenges that enterprises face with deploying AI. I think they said something along the lines of 5% of the enterprises they had interviewed or had surveyed had failed to take that AI project out of pilot and into production. and show a measurable return on investment. Biggest culprit, no surprise, data, right? And obviously security is a huge element of that. A lot of companies are doing AI pilots, they're moving, they're doing, you know, maybe they stand up, you know, kind of one workload, but it's not scaling for some reason or another, right? What do you think is holding these companies back and what is, you know, Cisco's AI factory, AI factories in general, but Cisco's AI factory, how's it kind of got you from pilot to production?

‍

Nicolas Sagnes:

Yeah, it's a great question. I think you're spot on with the MIT report, you know, I think that puts into light a lot of the issues that customers are experiencing. And from what they're telling us and what we've observed as well, basically, they've been successful in some capacity, in some islands of AI production. Some enterprises started by just buying GPU compute, and it was the race to get the GPU compute clusters ramped up. But very quickly, as they embedded these GPUs, they discovered bottlenecks very quickly. So networking is a big thing. Data, as you said, has multiple data silos, 20 to 200 silos, sometimes even more in the enterprise with different formats, no security, no access control, no real acceleration of data everywhere. Very quickly, they got stranded in their capacity. So they had the GPU compute, but nothing to back it up and to feed these GPUs. As you said, security as well. So security was an afterthought. I think now the industry is kind of moving around into it. And that's really the stance of Cisco here as we developed a lot of solutions and also framework around security. So what is security for AI? You know, end-to-end security, not just bolted on after the fact as an afterthought, but really from the ground up when you design your cluster, when you design your infrastructure to have security in mind, that's part of the framework. And so that's what led Cisco to develop the secure AI factory with Nvidia. That was really the great collaboration there.

‍

Matt Kimball:

So, you know, one of the things we're hearing about kind of operationalizing AI, right, or activating AI, right, you kind of hit on it with, you know, they found this island of data in, you know, enterprise IT and they turned it into something, maybe it's a chatbot for HR or for, you know, support or something along those lines. But we're hearing more and more as the market goes, when it comes to practical implementations, there's agentic AI, there's RAG, this thing called RAG, right? Retrieval Augmented Generation. And it sounds really great, but what do those actually look like in terms of practice, right? How do, you know, can you talk about how organizations are actually employing AI and employing RAG and employing agentic AI? Because these really speak to that, what you talked about, those dozens to hundreds of, you know, data sources that are sitting all over the place in an enterprise.

‍

Nicolas Sagnes:

Yeah, absolutely. I think, you know, this is a great question because enterprises most of the time will not play in the training realm, you know, will not develop training clusters to develop their own models. They will use predetermined models, open source models. They will post-train them, fine-tune them, and then implement them in their application. And so that's where the economic value is for the enterprise today. So the right pipelines, the RAG implementation. So you have to have a RAG-ready fabric that feeds data. So one of the key approaches is NVIDIA AI Data Platform. So NVIDIA really understood that issue from early on when they started partnering with Vaas Data, NetApp, Pure Storage, all the ecosystem partners in the market, saying that, hey, we have an issue with data for AI. We need to bring GPU compute to the data. And so that's what the AI data platform reference design, reference architecture is. And that's what we took at Cisco as part of the Secure AI Factory. So Secure AI Factory being the overarching concept, kind of the modular blueprint to tell you exactly what you need to do for that outcome of RAG and inferencing or training, you know, regardless of what you're trying to do. But if you're an enterprise, it's going to be mostly RAG and inferencing. And you can apply the right density of compute and get the right network optimized and accelerated for your task. And that data system is underneath it, so the AI data platform At Cisco, we really pushed that concept to compute, so we embed the software piece directly in our servers, and you can really very easily configure your rack-ready fabric in a secure way using the SecurRF factory stack.

‍

Matt Kimball:

So when you talk about the kind of partnership that NVIDIA has established and Cisco has established, more importantly Cisco, with the vasts of the world that are, you know, addressing huge data set challenges, right? Or the NetApps and the peers that are maybe doing it more at the commercial enterprise level. Are you saying that Cisco has kind of created a fabric that takes data that's existing across hundreds of data sources, databases, data lakes, even just a bunch of unstructured data, and you're helping those vast, large, vast data platforms feed those GPUs consistently so that they're never starving, so to speak, and doing that in a secure fashion?

‍

Nicolas Sagnes:

Yeah, basically what we did is, first of all, we applied that science to ourselves. With Cisco and Cisco, we kind of live in our own AI data platform and our own secure AI factory because that's Cisco's approach. We have that Cisco validated design that is really the core philosophy of everything we do. We test it ourselves, we test it with customers, we make sure that it works. And so working with NVIDIA was very instrumental, and VAST as well, and other ecosystem partners as well, to get that feedback of what is the best practice, what is the reference design and architecture, and how do we operationalize it, back to your question earlier. And the way we did it is with Cisco iPods. So we developed and designed the right fabric sizes based on these compute densities of ERAs or CRAs, all the architectures. And when you want to do RAD with a specific amount of density and you size up your model, that will tell you the right fabric you need to apply. Back-end fabric and front-end fabric to optimize your GPUs, and then the right way to bring the data in. And that's the work we did. to create the storage servers with VEST.

‍

Matt Kimball:

I think when a lot of enterprises, I talk to CIOs and IT leaders and business leaders frequently, and there's this rush to, and you hit it on it earlier, right? Get as many GPUs as I can, get them deployed everywhere. And there's no like, and by the way, that kind of goes downstream into the fabric you're deploying is just as much as you can. And a lot of times it's too much, right? It's not right sized for the needs of the enterprise. When you talk about kind of stepping back and working with organizations to understand kind of what their data needs are and what their fabric needs are, and then deploying the right size front end and back end, It's nice to know that, you know, you're kind of thinking about what this looks like in the real world as it's deployed and not some, you know, I want to be like AWS or Azure and, you know, have the largest GPU farm I possibly can. That's actually really good to hear.

‍

Nicolas Sagnes:

Yeah, I think there is also something to be said about what status you are in as an enterprise? What is the size of your data science team? Are you developing your own AI? Are you taking AI out of the box and operationalizing it, creating the blueprints everywhere, the applications? So we have great partners. We partner with WWT, for example, that can help develop these applications. from that AI pod security factory foundation and help deliver that application faster. And you remove the guesswork out of the equation. How am I going to optimize my cluster, make sure that I have a correct ROI and I can distribute my GPU resources. while I feed all that data and operationalize that data. So basically, the concept of disaggregation is I break out the silos, no more tiered storage, no more tiered silos. I have a distributed environment that's shared everywhere and accelerated everywhere. And then I bring GPUs to that data plane to accelerate the data path to my RAG infrastructure. So that's really how these systems are.

‍

Matt Kimball:

You're hitting on exactly what we talked about at the beginning, which is time to first token, ROI, time to value. How quickly can I turn this really big investment into incredible value to the organization? Hey, going back to the security thing for a second, because you talk about security first, right? And what's that mean for like a CIO or an IT team on a day-to-day basis? It's like, how are organizations balancing privacy, governance, safety, security without impacting performance? And in really kind of scaling or speeding up the scale of AI, what are you hearing? What are you seeing? What do you see out there?

‍

Nicolas Sagnes:

Yeah, so and I think that's kind of the overarching question we talk about with a lot of CIOs. We had our customers in Austin, Texas recently as part of our advisory board and they're kind of all in the same boat where they want a full-stack approach basically from the ground up from the network. That's what we adopted as an approach is the hybrid mesh firewall, a reference architecture where we have the security and the DPUs embedded in our switches, and we're able to have hardware security everywhere. And then going up the stack into the Kubernetes layer, that's really where you're going to impact or help performance while embedding security. And that's all the work that Isovalent is doing, Cisco's acquisition of Isovalent that developed the CLM. and eBPF protocols where basically you can apply security guardrails and policies on every container without impacting performance and at scale. And so that's the isovaligent Kubernetes for networking offering that we announced a couple of months ago that's very helpful there. And so once you have your Kubernetes layer solved, then you can approach the application layer. And so that's where we developed AI defense. So AI Defense gives you the means to control the models you're using. If you're using open source models or if you develop your models, you're going to be able to apply guardrails and apply policies across the board on every model that you run and understand which data is being fetched, which data is being shared outside. That's what we use for ourselves in our own Cisco IT circuit. That's the name of our tool. our Gen AI tool, that's how we control this. And for a user to bring data to a third-party LLM, like a chat GPT or anything, you will get better visibility and understanding of what they're uploading while everything is watermarked in your databases. and you're able to understand the behavior of your AI models into your pipeline. So that full stack approach is really key. And then we'll see more and more tools coming up in the upcoming months and years. And that's a pretty exciting field, actually.

‍

Matt Kimball:

Because you're embedding it at such a low level and you're kind of integrating it at the foundation, you're not impacting that performance that so many organizations, especially as you get to rag and inference or You're kind of driving that near real-time responsiveness that enterprises crave. You're not impacting that by kind of building in all these guardrails and ensuring security and privacy of data.

‍

Nicolas Sagnes:

But you know, the key theme across the board is basically you have to design a system. It's not just designing a silo of compute for a specific application. with a top-down silo that is completely incoherent with the rest of your organization. You have to have that system mindset and approach where security needs to be across the board in a coherent system. Networking needs to be accelerated with the right type of fabric. Storage needs to be accelerated. And then you start getting the benefit that lifts the entire enterprise.

‍

Matt Kimball:

Yeah. It's funny you say that because to me that kind of speaks to kind of the enterprise AI story in general, right? Which is, it's got to be horizontal and span the entire enterprise. You can't stand up, you know, an application here or an AI powered application here without thinking about the impact to the enterprise application over here. That's also going to be AI-ified, right? Through agentic. It's like, you've got to take that broad view and you've got to start from the right, which is the experience and work all the way back, which impacts security performance. governance, so on and so forth.

‍

Nicolas Sagnes:

You can't do agenetic at scale if you don't have the type of infrastructure. Very quickly, these models are going to clog the entire pipeline of resources and or explode the cost. So it's either or.

‍

Matt Kimball:

Explode already, exploding costs. It's like an exponential. Yeah. Yeah. Hey, one last question for you before we go. This is a fun conversation. But, you know, you are out in the market, you know, talking to CIOs every day. You know, you had your advisory session or board down in the lovely city of Austin where we're both from. As you're talking to CIOs and IT leaders and business leaders, because this is, you know, this is touching every part of the organization, for 2026 AI strategies, 2027 and beyond, and kind of looking long term, What's the one thing you would tell them to focus on right now to be ready for that secure, production-grade AI that will not only empower the enterprise, but also secure and drive incredible results for them?

‍

Nicolas Sagnes:

I think what we're seeing is that a lot of IT leaders today, they're playing catch up on the AI roadmap and that interface with line of business and the economic value they want to prioritize, I think is very important. And they're all trying to make sense of the infrastructure they have, they deployed five years ago and the platform is like, how can I leverage my previous platforms? And so one leverage for that is, I think it's twofold. First, you can leverage new cloud types of infrastructure, like for example, NVIDIA came up with NVIDIA Workbench or DGX Cloud to help you scale quickly and iterate in the cloud and then port these workloads on-prem. So that's more the data science and AI team that will work in that field. And then when it comes on-prem, you have to have a blueprint and a framework in mind. So that's what the security factory is for. Do I have the right security policies and model? Do I have the right governance and observability across the stack? Am I able to digest and grow my GPU pool? and distribute these resources. So, some elements need to be completely overhauled, some elements need to be upgraded, and some elements just need to be tweaked. So, every enterprise has a different equation at the end of the day, but I think it's really key, again, to design the system, especially if you're at the edge, when you have distributed compute in thousands of locations, et cetera, that are all doing inference or growing into the inference, you're going to need more and more compute capacity at the edge, and so extending that framework to the edge and being able to leverage your data center and your edge deployments together.

‍

Matt Kimball:

I think the connectivity, the underlying connectivity, there are so many organizations out there that are running on a networking infrastructure, in an interconnect infrastructure that is maybe a little bit older than AI is going to demand. And I think starting the foundation up is critical. I think one of the other things, I think as AI is an uber automation or business process modernization effort, right? We did this back in the early days where we turned writing and ledgers to programs. The difference is every organization and every piece of your organization now touches the other. So this can't be Nicolas as an IT leader supporting HR goes and rewrites their HR application. He has to also speak to every other organization and make sure that they're all working together to build that kind of end-to-end fully efficient and fully operationalized

‍

Nicolas Sagnes:

Yeah, I think it's very important. You need to have that standardized infrastructure approach from the compute, so the right amount of compute for the right workloads in the right platforms that you can add on more compute and you have the fabric underneath that supports it and the data plane that supports it. But really where you'll see the more flexibility and ease of use, so to speak, is on the container level. So the container orchestration, what we're doing with Red Hat, OpenShift, OpenShift AI, for example, The ability to have all the Kubernetes value is key because you're going to embed the security at that level and you will enable teams, different teams to share resources and kind of virtualize this GPU pool everywhere. So any data science teams that want to work on a new application, your HR use case, they'll be able to leverage GPU on tap without clogging the system, and they will get the capacity they need to get their value. So you become a service model to the business as an IT professional, basically.

‍

Matt Kimball:

Yeah, you become the consultant. I love this. I think we could talk for the next hour. about how this all plays out and how Cisco really is kind of the underlying kind of enablement of this entire AI environment. Nicolas, thank you so much for taking the time. Thank you.

‍

Nicolas Sagnes:

Great conversation.

‍

Matt Kimball:

Yeah, this is fun. And thank you for, you know, I know you're busy at Supercomputing, 25, walking the floor, shaking hands and kissing babies. Thank you for coming on in and having this discussion. And the best of luck with the rest of the conference. All right, thank you. Thank you. And thank you for tuning in to this episode of Six Five on the Road at Supercomputing25. Stay tuned for more content.

‍