NVIDIA GTC Washington, D.C. 2025 | Gilad Shainer, NVIDIA & Will Eatherton, Cisco

Clips
News
More from NVIDIA GTC Washington, D.C. 2025

Gilad Shainer

SVP of Networking

NVIDIA

Will Eatherton

SVP, Data Center, Internet & Cloud Infrastructure Engineering

Cisco

Three insights you might have missed from theCUBE’s coverage from Nvidia’s GTC Washington, D.C. event

Enterprise infrastructure is no longer measured in servers or racks but in the architectural decisions that determine what those systems can accomplish at scale. As organizations race to deploy artificial intelligence workloads, the question is less about acquiring technology and more about integrating it through reference architectures that support rapid iteration without forcing teams to start from scratch.That integration challenge took center stage at the recent Nvidia GTC Washington, D.C. event, where partnerships between silicon providers, networking vendors and systems integrators demonstrated how reference architectures and validated ecosystems are reshaping deployment timelines

Building the AI data center: Why reference architecture matters now

The next wave of artificial intelligence infrastructure is taking shape through reference architecture, connecting innovation from silicon to the data center floor in ways that promise scale, trust and operational simplicity.Enterprise computing is shifting from hardware-centric design to systems built around integrated architectures that unify networking, compute and management. The collaboration between Cisco and Nvidia highlights how reference models can accelerate AI adoption without forcing organizations to rebuild everything from scratch

play_circle_outline NVIDIA's GTC special event highlights advancements in networking and AI partnerships.

play_circle_outline New Nexus 9,100 switch integrated with Cisco's technology and AI capabilities.

play_circle_outline Importance of security in AI clusters and enterprise environments addressed.

play_circle_outline Collaboration enables integration of AI into existing IT infrastructure without major changes.

Info
Transcript

Gilad Shainer, NVIDIA & Will Eatherton, Cisco

Gilad Shainer

SVP of Networking NVIDIA

Will Eatherton

SVP, Data Center, Internet & Cloud Infrastructure Engineering Cisco

In this NVIDIA GTC Washington, D.C. segment, theCUBE’s John Furrier sits down with Gilad Shainer, senior vice president of marketing at NVIDIA, and Will Eatherton, senior vice president of network engineering at Cisco, to unpack how the companies are bringing AI to the enterprise through a “better together” approach. Eatherton outlines three announcements: a cloud reference architecture using Cisco Nexus and SONiC with NVIDIA Spectrum-X enablement to scale to thousands of GPUs; the new Nexus 9100, a Spectrum ASIC–enabled Cisco switch that fits into NCP progra... Read more

explore Keep Exploring

What is the core news being announced at NVIDIA's GTC special regarding the partnership with Cisco? add

What are the announcements regarding new products and solutions related to Nexus and AI? add

What is the significance of NVIDIA's transition from traditional data centers to AI factories in collaboration with Cisco, particularly regarding security and network integration? add

What is your perspective on the partnership between Cisco and NVIDIA and its impact on customer experiences and technology deployment? add

bolt Powered by CUBE AI

Gilad Shainer, NVIDIA & Will Eatherton, Cisco

search

John Furrier

>> Hello, I'm John Furrier with theCUBE. I'm here in Washington D.C for NVIDIA's GTC special, kind of a pop-up mini GTC. With so much action happening this year. Of course the main show in San Jose will be even bigger, but so much important news happening at the networking level and the partnerships powering the next generation of AI. We got two of the leaders from Cisco and NVIDIA. Gilad, thanks for coming on. He's the head of networking at NVIDIA, SVP and Will, SVP of network engineering at Cisco with big partnership news. Thanks for taking the time today. I know you're super busy and thanks for spending the time. As first, let's talk about the partnership with Cisco. You got a lot of threads here. What's the core news?

Will Eatherton

>> So I'll just hit the three key things we're announcing this week. The first one is cloud reference architecture, which is based on currently shipping technology. So that's Cisco Nexus and Sonic running with Spectrum X enablement so that our customers can get the management and the switches that they've been using, but with a reference architecture that allows them to scale under the thousands of GPUs. So that was the first announcement. Second announcement is a new Nexus 9,100, and this is going to be an NCPRA, which is for especially NeoCloud, but other customers a very key blanket program. And so the 9,100 is a spectrum ASIC enabled Cisco switch that's running Nexus or Sonic and will be part of that NCP. And then the third one is the orderability and soon to be GA of hyper fabric AI. And this is an innovative new solution for full cluster management because management of these clusters, especially for high-end enterprise, is something that is a really big deal and they don't want to be managing super computers.

John Furrier

>> On the NVIDIA side, you guys, the market's obviously growing the demand for tokens in the AI factory, we covered that. The traffic demands on the networking's high. The systems are optimized for networking as kind of an OS, and the ecosystem is developing with customers kind of rapidly forcing in and with the open source developer community booming. This kind of puts you guys in a good position. What does this mean for that market? Because you're now enabling more, I guess, compute, programmability, manageability, essentially network support, which people want to adapt to their existing but also build in new AI capabilities. What's this mean for that?

Gilad Shainer

>> Right. Inside NVIDIA, we're essentially focusing on building AI data centers. So building AI super computers, and you all know that data center has become the unit of computing. We also know that the way that you connect GPU ASICs will determine what that data center can do. And we also know that we're an annual cadence of technology because there's a fast pace of progress. There is a demand from the workloads on inferencing, trainings, tokens and so forth. So we are on annual cadence of releasing of new technology, and we want that AI technology to be accessible by everyone. And therefore working with the ecosystem, it's an important thing for NVIDIA to make sure that, again, AI is accessible for everyone. We are focusing on a lot of open source and bringing the flexibility in the options. The collaboration with Cisco actually help us to drive the AI technology that we develop into the enterprise class, into the cloud customers that they're already using the great Cisco software infrastructure within their data centers. They would like to continue using the same things because this is how their IT infrastructure is built for, this is what they invested so far and they want to continue to reuse their investment or from an investment. And our collaboration enables us to actually bring that AI technology, those AI supercomputers into the enterprise, into the clouds, and they will continue to use the greatness of Cisco. So for them, it actually will be very, very easier to consume AI and to new generations of AI as we move forward.

John Furrier

>> This is the story I think everyone needs to hear because it's very nuanced what you just said. You're bringing AI to the networking, but you already have networking inside NVIDIA. KV Cache and Jensen was on stage last year and he said, KV Cache is the operating system of AI factory. And I'm like, oh my god. Boom. bells went off. Bringing AI into an existing network is not a bug, it's a feature, Will. This is what I think what's happening is that the worlds are coming together. Networking is not one thing, it's many things. Explain to the person out there who's trying to figure this out from an investment standpoint, staffing, deployment, it's a generational investment. So bringing networking NVIDIA AI, AI to networking, is a feature. Why is that so important? What should people be thinking about? It's not replacing anything, it's only additive. What's your view on that?

Will Eatherton

>> So very much it's additive and I think that's where, first of all, trust is a big factor for a lot of the customers we were talking to where they want to have those reference architectures, are a key aspect. They want to know that, and this is where my team is building up these systems using the Cisco Silicon one spectrum, the combined technology management. We're running all the benchmarks and testing and then we're walking through with Glatt and his team. And so our customers want to trust that that's something that we've done. So they don't have to do, they don't want to be deploying those teams doing all that work. And so that's one key aspect. And then following the reference architecture, the next big topic comes in is operational management. And so for us being able to bring in Nexus, Nexus Dashboard, the new hyper fabric AI, these are ways that our customers with small teams and expertise that they already have can add that. So for them, they should think of it as instead of a DGX or a single box, it's now that cluster. And maybe it's starting with 64, 128 GPUs is very common, but they shouldn't have to be experts on Rocky and RDMA and all those aspects of troubleshooting. That's something that the tooling and the management software should do for them, and they can trust us to put that together and validate.

John Furrier

>> As I say, supercomputing brings superpowers. So this is, you're adding AI to the networking. What's the consideration, Gilad, on the engineering side because the reference architecture speaks to, okay, that's that environment. But as we move forward, the configuration of the software and the hardware is huge. The network fabric is one of the most discussed items around these large-scale systems. Where do you put it, what is it? The role of HBM and the networking amongst Blackwells, connecting together the role of ethernet as that connects into other networks, you have kind of a multi-network perspective. Multiple interconnects. It sounds more complicated, but this is the design. So as you look at the reference architectures, what's the consideration for someone who's investing in this? What should they be thinking about when they start deploying?

Gilad Shainer

>> Yeah. First, you're completely right. You build a supercomputer. And when you build a supercomputer, it's not just one network that covers everything and it's not even a network. It's a full infrastructure and it's not just an infrastructure, it's actually a computing infrastructure. And you have a computing infrastructure that we call scale up that actually creates that GPU. There is many GPU silicon or many GPU devices, scale up, create one large GPU, and then you have a scale-out computing infrastructure that connects all those racks together and scale up the data center and making the data center a single unit of computing. You have the access network, you have the storage network. There is multiple infrastructure that actually build that data center. And when you think about it, it's very complicated. It's actually a complicated task to try and build that supercomputer. And you want to build it quickly because when you build an AI data center, there is a good amount of investment. You cannot afford the old concepts of building data centers that took years or even months. You need to get this stuff working in weeks because every day is expensive. And that's the importance of the reference architecture. That's why when we build a new, or when we design a new platform, it's a full platform. We build a data center, we create a reference architecture, we test everything, making sure the performance is there, making sure that everything runs, all the software frameworks are there, everything is good to go. And after we build that reference architecture, we provide it to our customers and partners. And if you're following that, you can build that new AI data center in weeks, and you don't need to have huge amount or a huge group of ITs to cover all those elements because it's no more, let me buy component from A and component from B and try to assemble that. You actually go reference architecture. You can build it in weeks and should go to production very, very quickly and be able to enjoy all the software frameworks and all the applications running on top. And partnership with Cisco bring also the ability of actually utilizing Cisco, the Cisco management, everything that they built around. So now there is a reference architecture from Cisco for people that actually want to use or use to using Cisco.

Will Eatherton

>> Going beyond the base cluster. This is where the solution level comes in because how you plug that in to your wide area network, how do you storage site and then security. So for us, the security AI factor has been a key component. And so how do we, everything from LLM protection to East-West firewall to putting all those pieces together and growing out from the cloud reference architecture is key for us.

John Furrier

>> This is why I like this, because what you just basically said was NVIDIA is standing up AI factories now, not data centers. Data centers have been around bounded by power. You guys have built those racks. Now, when you get into the partnership with Cisco, I think what this shows the market is now when you have customers that have been doing networks. You're the networking, okay, he's got the factories. They have networking inside the factories, which is designed and engineered. Complicated. They've done that. Then the customers then deploy. Take us through that piece because now this is kind of like a scale across, but not the way you guys talk about it at NVIDIA, but more of, okay, you're going to commercialize and create value from the factories. It could be a smaller data center, it could be an edge node, it could be a regional city, sovereign cloud. Who knows what could it be. This is where the value creation comes in. So talk about the importance of how that value's extracted. Because if the AI factory does its job and you got the networks you connect, now I can code on top of it. I can bring apps in, I can do manageability. Talk about the importance of how you guys are enabled to create the value, and then just your vision on how it's extracted.

Will Eatherton

>> So the industry talks a lot about Neoclouds that are standing up Greenfield. So as you branch out and you talk about the wider enterprises, it does start with existing infrastructure. And for them it is about bringing in and being able. So how am I going to deploy this in and connect it to the ethernet that I have today? And so what does that look like from a proto? So these are things that in our reference architecture and as we work with customers, we're making it very easy to stitch that in essentially. And then things like security are big topics. While you may not necessarily need fine grain security on the east west network, that cluster and the accompanying data, and many times you're pulling together some of the most critical data for your enterprise. So securing that, being able to have policy wrap around that. So for instance, for us, Isovalent which was an acquisition we did a year or so ago, is a key component of a lot of these AI clusters. And that, being able to bring that in and then even accelerate that on top of things like Bluefield three, moving to Bluefield four in the future, is a key part of you can only extract value if it's secure and you trust, then it makes sense. And that's where a lot of customers are going to want to have this infrastructure on-prem where they can control it versus an API in the cloud.

John Furrier

>> It's interesting. During the SAAS generation, the cloud enabled a lot of values, mostly extracted by Apple and Google in their stores. Now with NVIDIA, the value extracted by the network enablement, which is the owners of that network, it could be Neoclouds, even clouds, but also enterprises. So the enterprises and the Neoclouds seem to be the ones that eye in on this. You got to talk about that value creation because now you have new participants who are engineering solutions for doing that. So if I'm an enterprise, I'm like, okay, I'm going to have an AI. In fact, I might even work with a NeoCloud. What does that mean? What does the Cisco relationship mean , because you've got the routing, you've got the factories. What is the key thing to take away from that? Because that's where the market seems to be shifting the most right now in terms of the GPU deployments and then the software that's running on them to enable just apps to run AI native or retrofitted legacy, modern legacy. That's coming. What's your take on that piece? What's your reaction to that.

Gilad Shainer

>> Yeah, Will, what do you think?

Will Eatherton

>> Well, I think from the partnership standpoint, the key things are pre-validation, then running, for instance, our technology, Cisco Silicon one with Spectrum and from a customer standpoint, they don't have to understand all these nuances. And so it really comes back to that trust that Cisco and NVIDIA together are offering these solutions, that there's choice in there. It's an ecosystem between the two companies as well as embracing the broader industry open standards. And so a lot of times customers are confused. There's just so much changing every few months. And in this standpoint, they can work with Cisco and NVIDIA, excuse me, and get these deployments up and running quickly.

Gilad Shainer

>> The key thing is that we want to get those networking AI engines inside, and we want to make it in the easiest way for enterprises to consume that inside their data centers. Now, they have built enterprise data centers for years. They know how to run that. They're using Cisco, they have the full Cisco management across that data center and so forth. For running AI, you need to bring that infrastructure built for AI. So those are the elements of, this is the SpectrumX Ethernet technology that we're bringing in, and it goes into the Cisco infrastructure. So for those enterprise customers, they see no change from the way that they operate the data center, it's the same thing. But now the data center is not the data center. It's actually an AI supercomputer.

John Furrier

>> Yeah, I think this is the final point where I've seen the harvesting of the term AI for networking or AI for the data center. Because what you just said is NVIDIA's bringing AI to the data center. And that, to me, I think is categorically the new thing. It's not changing it, it's only making it better. So it's a better together story on the partnership. But it also changes the game on what actually runs on it. So if you've got the keys to the kingdom, which is networking. Networking, at end of the day, we know it's the OS of things as you bring things together because NVIDIA, with the connecting Blackwells together made it happen. So the interconnects become a topic everyone's talking about, but you got networks, you're connecting. So the word, I mean, interconnect is networking, it's just focused. What you're saying is with Cisco, you're bringing AI to the enterprise and all networks.

Gilad Shainer

>> Correct.

John Furrier

>> What does that mean? What's the outcome? What happens next from this?

Will Eatherton

>> Well, first of all, it's key. This isn't a one and done. So this has been a... as we've put this together, this is going to be multi-generations. We're already working on the next chips, the next software, the evolution of the algorithms between the NICs and the switch chips and how that all worked together. So it's a many years. So as our joint customers buy into these architectures, we will support them and keep going. And so that's a key aspect. But I think a couple key things, the internal algorithms will keep getting more complex, but the other part is operationally making it simple. And so we see that's the area that needs a lot of attention right now is that these can be quite complex. There might be four or five different types of networks, all that come together at the cluster, and our customers don't want to have to have an army of folks. So operations is where I'd say there's a lot of focus.

John Furrier

>> And the hyper-fabric AI is the modern approach.

Will Eatherton

>> That is a new take. We're disrupting ourselves at Cisco. So we have our existing tools and we continue to modernize. But the hyper-fabric AI is a radical new approach.

John Furrier

>> And that's essentially just optimized for the AI cluster to power what we were just talking about.

Will Eatherton

>> It is, yes. And having a view not only of network, but of DPUs and things like across all the DPUs, being able to do a mesh of what are the latencies and the connectivity, and then bringing in storage and being able to view it all. So yeah, it's a fun new take on infrastructure.

John Furrier

>> And it's relevant. Cool and relevant. So thanks for coming on guys. Appreciate it. Thanks for the time. I'm John Furrier with theCUBE here at NVIDIA's GTC in DC, getting all the stories and action and the data and sharing that with you. Thanks for watching.

>> And we're clear.

John Furrier

>> You want to do a 30 minute, 30 second clip for social? All right, so you want to? Okay. Here's the thirty-minute clip. This is for social media.

Will Eatherton

>> 30 seconds, okay.

John Furrier

>> What does the partnership mean between NVIDIA and Cisco? What should people take away?

Will Eatherton

>> I'll start with trust. Cisco, NVIDIA working together, we're making these reference architectures cohesive and we're allowing our customers to bring the supercomputer into the data center.

Gilad Shainer

>> Bringing AI to the enterprise. That's the meaning of the collaboration between NVIDIA and Cisco.

John Furrier

>> Okay. All the action here in DC. Thanks for watching.