Hello again, and welcome to Cisco's Inside Product series. We have some very exciting guests today. So we've got Tom Gillis who runs our data center and security business. We've got Will Eatherton who runs our engineering for the data center business. And then we've got Kevin Wollenweber who runs all of our data center product management business, which is everything from hyperscaler to neoclouds, to enterprise, to service providers, and on and on. So welcome, gentlemen. Tom, I'm going to start with you.
What is happening in the industry as it pertains to AI and infrastructure? Give us the macro picture first.
Yeah, sure. Everyone knows the power of AI. It's amazing. The nature of the applications are fundamentally different than applications that we've lived with for decades. And really, the way to think of it is they're parallel. So that means that an application doesn't have to fit in a box. And in fact, in both Kevin and Will's world, a lot of these high-end applications—we're dealing with a single AI application that can span more than just one box, more than just one rack, more than one row. Jeetu, these things run across a whole data center and multiple data centers. So in this world, the network becomes the backplane of these applications. And it's kind of akin to what the PCI bus was in the world of a server in a box, right? PCI bus is the guts of the system. Now, all of a sudden, the networks are playing that role. That puts very special requirements on the network. And I think that's one of the reasons why.
No matter how good the GPU is, if it's not network.
Dude, the network has got to be better, right?
And so that's why there's so much emphasis on the network. Now, we tend to think of this in terms of what we call frontend network, which is a network that has lived in a data center for decades. But when a GPU is talking to another GPU, that's what we call a backend network. And I think both Kevin and Will will provide perspective on backend network, frontend network. How do we turn this to make it easy for customers to be able to deploy these AI applications and not get bogged down in some of this minutia? That's, I think, the Cisco advantage.
So I'm going to go to Kevin first. Walk us through, architecturally, what changes. And what specifically do we need to make sure that we're keeping in mind as this architecture shifts in this agentic world?
Yeah. I mean, first of all, I think Tom laid it out really well. But a lot of what we build when we build out AI infrastructure are the same piece parts and components of what we've been building in traditional data centers. You have compute. And you have network. And you have storage. But now, you have multiple networks. And you have different capacities and need to access storage. And so you can think of taking the components we had, adding a ton of complexity, and then having to have it all orchestrated and managed and operating like a single system. And so there's two main areas of investment. One, we have to really focus on the piece parts and the components. And I'll talk about that in a second. But then, we also have to focus on orchestration and management and making it simple and easy to use. Because you don't want customers focused on, how does this GPU talk to that GPU? And how do I optimize the fabric to make it communicate well? You want them to focus on running AI workloads and getting value out of it.
OK, so when you go into a customer, how should the customer be thinking? Because most customers are using hyperscalers for their AI workloads. And now, what you're starting to see is these workloads, especially in inferencing, are getting pretty distributed. Some of them are getting local. You might have a desk side computer that's running agents. You might have your own enterprise data center that you're doing stuff. What is the playbook from what an enterprise should be doing to go out and keep the token costs low, the outputs high, and the complexity as low as possible so that they can actually drive this at the velocity they need?
Yeah. Well, it's kind of crazy. Although we've been talking about AI spend for what feels like decades, it's only been the last few years that we've seen this massive increase in spend. And most of that spend is with the hyperscalers, with the neoclouds, and with the model builders training the models themselves. So now, we have these amazingly powerful models. And what's evolving is this world of inference, so people actually using the models. And then when you say things like agentic workflows, it's really leveraging these models that have been trained on a certain set of functions and using them to perform tasks. And so that architecture actually looks fundamentally different. And, yeah, we've been doing all the training in these massive data centers in cloud and in neoclouds. But a lot more of this for inference is going to move either on prem in the customer. You talked about having little things on people's desktops to be able to run agents. And that makes the network even more critical in how they manage and how they operate these workflows.
So, Will, thinking about the backend network and the frontend network, you are architecting these things for customers and for our own systems. And you're starting to also see that there's a lot of semi-custom work that gets done in extremely large-scale data centers. Talk us through what your learnings are. And what should customers take away about Cisco as it pertains to the kind of things that you're doing? And why should they be as excited as you and I are?
So, first of all, just to hit that we really are seeing three distinct customer engagements today, so hyperscale for AI. We've been engaged for multiple years now. And in that context, it's a very focused, sort of high-performance networking where we're going in with our Silicon One based systems, a range of our software. And we've been working together with those hyperscalers to really tune and get the highest performance across that. And that, again, has been several years. What we're seeing in the last 12 months is obviously more engagement, both with neocloud class of customers as well as enterprises.
So neocloud would be folks who essentially— their business is selling a narrow set as a service, whether it's bare metal or it's more virtualized. And so they're a smaller sliver of what has been the IaaS cloud providers.
But that means like CoreWeave are examples of neocloud.
Yes. And sometimes it's pretty coarse as far as they carve up sets of servers that they will then lease or allocate for a year and sometimes a lot more, the higher values when they can be more flexible in how they allocate in that. So, in all of these, one of the key things Kevin mentioned is the number of potential different networks can be daunting. And so you can have what's been our traditional frontend, really stems out of what's been a traditional data center with a lot of feature and functionality on how do you connect servers to servers, servers to internet, servers to storage. That's the frontend. The backend has been InfiniBand in the early days. It's pretty clear in the last 12 months that is now ethernet. Everyone, including NVIDIA is pushing that. But we also have management networks, sometimes distinct storage networks. And so having a strategy around how you're going to operate across these is a key aspect and something we engage with customers. Also, what we're now seeing is a new trend as inference is becoming a bigger topic. And it's not as simple as the NVIDIA. They've had these reference architectures and superpods. If you've seen from GTC, that has been a repeatable pattern for the last couple years. But now, in the next 12 months, we're seeing more with the inference accelerators. That is complicating these architectures. And I think the key commonality is going to be ethernet is what we're seeing, connecting accelerators to GPUs. Also, storage. So storage to accelerators, storage to GPUs is requiring higher bandwidth. So those frontend networks are requiring increased bandwidth is what we're seeing with customers. So thinking about planning these architectures, having flexibility, and the operational simplicity is a key topic. And I'd say that's common across both. Even though there's some differences, it's common across enterprises and neocloud.
Jeetu, one of the things I think is interesting. You listen to these two guys talk. They spend a lot of time with these top of market customers that are doing crazy things. And the learnings that we, Cisco have from that experience were folding into products to make this stuff usable for the enterprise. And isn't that our opportunity? Right?
So actually, let's talk about that, Tom. So take a step back and tell me why a customer would be interested in getting something from Cisco, versus a competitor? Why are we different? What do we do that is a structural advantage in the market? And how should a customer parse the value that we provide, versus someone else provides?
Yeah. These guys can probably articulate this better. But I think it comes down to, really, simply, customers don't want different little bits and pieces of network that have to be managed differently. Historically, we've seen that. Remember when telephones were totally different networks? Ultimately, ethernet subsumes it. And customers want a single set of management, a single set of hardware that can run all this stuff. And I think that's our opportunity is these backend networks, frontend networks. With the programmability of Silicon One, we can address all of these needs with your network. It's not a different diddly bop. And I think that's going to really, really matter in the near future.
So I have an example from the last six months in that. So we were working with Quant at the beginning of this year in Cisco Live, Amsterdam time. One of the feedback we got was that InfiniBand— again, very respected technology that everyone is clear, now, won't be the winner. But they did have a feature with this partition ID that allowed jobs to have distinct allocation from a tenancy. So you could have both separation of resources as well as some security benefits of that isolation. And the feedback we were brainstorming with them— it'd be great if ethernet had that. And obviously, there should be nothing that InfiniBand has that ethernet doesn't have. So we went and architected that. And then we were actually able, with Silicon One P4 programmability, to add that feature in our Silicon One software layer, since February. We then added that into our NexusOS layer. We then added in a top level for configuration and visibility into Nexus Dashboard. And actually, at Cisco Live recently, we were able to show that feature. And that is actually— we've had interest both from neoclouds where they want to have tenancy, where the first layer of tenancy is my customer. Second layer of tenancy could be jobbed— job ID that they're working through. But there's also enterprises that want to have some layer of separation. And so this is a very specific example. But I think it really helps illustrate that full Cisco stack and also the agility that we can move at when we have all those pieces together versus, hey, we have to go to another company and work with them on the silicon side or if it's a fixed chip and it takes two years to spin that. So I like that example.
So I just want to make sure that I put some emphasis on what you just said because it's extremely important. One of the big reasons that at least I've seen customers wanting to do a great deal work with us is we are one of the only vertically integrated platforms that is a co-designed full stack. And what I mean by that is silicon and optics are the foundation. Systems hardware, systems software, platforms for security and observability and data, applications that we build, models that we build, and then agents that are built on top of that. That whole thing is a co-design full stack. And what you just articulated was the fact that you were able to deliver that was because of the fact that we actually built every single component of it. So we're almost like the Apple of the infrastructure world where we build the hardware, we build the software, we build the chip. And all of them just are designed to work with each other. And the way that they work well is so magical. And because we work with the largest of the largest customers, which are the hyperscalers, and neocloud, and sovereign clouds, and service providers, we are able to take all the learnings from there to your point, Tom, and then make sure that you can take it all the way to the enterprise. So that's at the highest level. Now, let's go into the details. What were in your mind, Kevin? Because you're the product guy that's setting the direction for our data center. The key priorities that you wanted to go drive on our entire data center business. And let's specifically talk about enterprise here. So what we doing in the enterprise? Firstly, what do you see as trends in the enterprise? And then what specifically did you say you wanted to go out and prioritize? How was that delivered and what's next?
Yeah. I think since we're kind of starting at the top and moving down, I want to talk about that orchestration, and management function, and the ability to get up and running fast. So if you think about the biggest concerns a lot of enterprises have around moving to more and more of these AI technologies and bringing them on prem is, unfortunately, right now, they're still expensive. And so nobody wants to make a decision to bring a bunch of hardware and technology in, and then have it sit idle, and not be able to run it. And so I think there's some amazing innovation that we'll talk about in silicon itself and what we're doing there. But in general, what customers want to be able to do is bring the technologies to their premise, get them up and running, and actually be able to show a real ROI for running AI workloads, right?
They're not deploying infrastructure for the sake of deploying infrastructure. They're deploying it so that they can actually bring some of those workloads on prem, deploy some of those inference and workflows that we talked about, and do it at a cost structure that's much, much lower than what they're able to do today in cloud. And so we've been prioritizing partnerships with some of our large partners like NVIDIA that have access to many and many of these AI technologies, but then bringing our simplification of the network. We've been building networks for over 40 years now. And the network started with two professors that were trying to get computers to talk together. And this is just an extrapolation of that problem space.
And so give me a set of priorities that you actually set up to say, these are the things that— so it seems like one of them was, OK, simplify the management. And you've done the simplification of the management in a multitude of different ways because you've said Nexus Dashboard and ACI got converged.
Well, think about it that way. Take the same technologies they're already building those frontends with today and extend them to AI so that if they want to start to run AI workloads, it's not bringing in another networking technology, another operational tool, and another set of experts that have to go figure out how to make that work. They can just extend the existing network infrastructure they have. And so they are experts that run the network today, can run that AI infrastructure.
And then you also made sure that was fully integrated into Cisco Cloud Control, which became our unified management plan. And, Will, you were one of the spiritual leaders internally of making that happen. Walk us through the calculus on that. Why did you do that? What was so magical and important about it?
So I think the key— and we've been working, as you mentioned, to bring the Nexus portfolio together. When we go cross domain, some of the big gaps Cisco has historically had— it's not just about a UI interface. It is about things like, I want automation to have a single gateway point, so being able to bring together that common API gateway and start to normalize the cross controller. And then the other part was a consistent way to provide an AI frontend with new interface models, with AI Canvas and not having to build that separately into each product. And so with the Cisco Cloud Control architecture that we've put together in the last year, the first pass, easy-ish, is to get the SaaS products to be able to then be more seamless with Cisco Cloud Control. The other piece we've been working on which, engineering wise, has been more work, but as a big win, is when we also now start to have the on prem controllers be able to be cloud tethered and still have that Cisco Cloud Control frontend. And so things like Nexus Dashboard being able to have— and many customers have more than one of these. They may have global deployments. We have customers that might have dozens, dozens, and dozens of controllers being able to have that then pulled together, and have everything from common inventory, topology, being able to put together the beginnings of a unified EVPN fabric that'll work across our Catalyst area, our data center area in a cohesive fashion, building upon each of those subproducts. So these are some of the key aspects. And definitely, the customer reception has been amazing. And so, now, working on our roadmap for the next six or nine months to continue to fill out these features.
So this was a very important contribution that you made. Because you were one of the big drivers of, hey, I need to have this BGP EVPN fabric that was like a data center and a campus branch network, should just be able to get configured in one click. And we showed that demo. That was a live demo on stage at Cisco Live where they were able to do that. What is the benefit for the customer? And what's the ROI for the customer when that happens? Any one of you can talk about it... Tom.
Automation. You hit the nail on the head there. Customers just want to solve a problem. I want to make sure that my telephones that are out in the office.
But why is it important for my data center and my campus branch network to be—
I've got telephones that are out in the branch. And I want those telephones to talk to the call manager. I do not want those telephones to be logging in to the source code repository or the finance system. The network is the way to do that. You want to set up a segment that goes from the data center, from the application, out to the telephones. And we make that super, super easy.
So a customer may not necessarily want to create a single fabric in the definition of a fabric. But to the point Tom is making, being able to have them connect through what we call border gateways and then be able to have policy that extends from an application running in the data center to a laptop or an endpoint that's in the campus. We stitch it together in a way that now is feasible. And we can do everything from better telemetry and reporting to better ways to insert policy on top of that. And so this is something that we've been asked from customers for a couple years. And frankly, it's taken us too long. But now that we're here, we're going to move as quickly as possible in filling out that functionality.
In a post Mythos world, this kind of— segmentation is what that is, right? This kind of segmentation is more foundational than ever. Because we have to be working on the assumption that the attackers are in. There's just so many vulnerabilities. They're finding ways to get in. Let's make it hard for them to move around. EVPN.
You're trying to prevent lateral movement.
Correct. And there's no better way to do that than in the network and putting segments in place that are almost like common sense. We make that easy to deploy.
So let's switch to Mythos. Because Mythos is a pretty important phenomenon of what's happening right now. And specifically in the data center, it gets to be—
Superbly scary when something goes awry. So you've got Mythos. Mythos has identified a bunch of vulnerabilities. Your data center infrastructure might have some of those vulnerabilities. How do we help customers? What should customers do?
Yeah. So I'm going to argue, never miss a good crisis to drive change. So let's not sugarcoat it. This Mythos thing is a crisis for our customers. Because the way that we as an industry have been working is that we would create a data center design, prove that data design, harden it, and then you don't change it.
Leave it there as long as you possibly can.
Exactly, for like ever, right? So literally, every couple years, you're like, OK, I'm going to have to update this thing only when you have to. And now in this post Mythos world, it's not unreasonable.
It's constantly evolving.
Exactly. Right? Now, these concepts are not radical here. In the cloud, that's how it works. That's how Kubernetes' apps works. Lots of little tiny changes, instead of one big, giant, gut-wrenching change. So if we can all get there, I think we're all going to be better off. But we have to adapt the way we've been thinking about managing infrastructure. Now, fortunately, Cisco— we're on our toes here, not our heels. Because we've been thinking about this for a while. How do we manage vulnerabilities in critical, in-line, high-performance systems, like a switch, a router, a firewall? And so one piece of the puzzle— and I want to point out this is only one piece of the puzzle. We have this capability we call Live Protect. And so Live Protect— you can think of it as a finger in the dike. It's an emergency control that can be applied on a system. And the intent is that it is very, very targeted, very, very specific, and lightweight. So as customers get comfortable with these capabilities, you'll be able to put them on a system without a reboot for a system. And our design intention is we want to go the— you don't even need a change control window. These things are just going to automatically go in. But, Will, I'm going to ask you this question. It's kind of a fake question. Is Live Protect a patch?
So, no, it's not a patch.
Will is actually— very seriously.
It is a compensating control. And it is something that at the next patch window, then you'll want to replace with a full patch, OK? One thing I want to add to what you said, Tom, is things like Live Protect. I'm excited about a technology that I think is going to— we are going to have— the world has changed. Things will be different for the next several years. But in the short term, I do want to really convey that we have confidence in our core network OSs— IOS XE, IOS XR, NX-OS. We've been doing a lot of hardening of these NOSs for several years. We've also been working to improve the agility, our ability to address. So while we definitely are seeing without getting any specific numbers, we are seeing—
