The Networking for AI Summit | Grant Shirk, Cisco

Clips
More from The Networking for AI Summit

Grant Shirk

Senior Director, Product Marketing, Campus and Branch

Cisco

play_circle_outline Cisco's perspective on evolving campus and branch networks for AI success.

play_circle_outline Real-world examples of AI-driven outcomes in diverse environments.

play_circle_outline Agentic Operations: Simplifying Network Complexities and Differentiating Networking Strategies for AI Development

play_circle_outline Importance of understanding business changes to support future networks.

play_circle_outline Future-Ready Networks: Adapting to Continuous Inference Demands in the AI Era

Info
Transcript

Grant Shirk, Cisco

Grant Shirk

Senior Director, Product Marketing, Campus and Branch Cisco

Enterprises are bracing for AI at the edge. In this Networking for AI Summit interview, theCUBE’s Bob Laliberte speaks with Grant Shirk, Senior Director of Product Marketing for Networking at Cisco, about how campus and branch networks are evolving to support agentic AI. Shirk explains why “networking for AI” and “AI for networking” must advance together, detailing capacity planning, microsecond-level latency needs and end-to-end assurance. He shares real-world examples, including a major pharmaceutical firm whose on-prem AI workloads outgrew a newly upgraded... Read more

explore Keep Exploring

What are the ways in which Cisco networking is adapting to facilitate the integration of agentic AI in campus and branch environments? add

What impact does artificial intelligence have on network capacity and performance in various industries? add

What is the distinction between AI for networking and networking for AI, and why is it important to have both in the context of future network infrastructure? add

What factors should be considered for planning the future capacity and support of networks in light of evolving customer behaviors and operational changes? add

What are the best practices that decision makers should follow to ensure their campus and branch networks are prepared for the AI era? add

bolt Powered by CUBE AI

Grant Shirk, Cisco

search

>> Hello and welcome back to the Networking for AI Summit. Joining us today is Grant Shirk, Senior Director of Product Marketing for Networking, who's going to share Cisco's perspective on how enterprises can prepare for their campus and branch networks in the AI era. Welcome, Grant.

>> Thanks, Bob. How's it going today?

>> It's going great. Glad to have you on. I think we're going to have a great conversation. To get started, as AI adoption is accelerating, it's not just about the data centers that need to evolve, right? That's getting all the hype. But clearly, the campus and branch are becoming just as critical enablers for AI, enabling AI to be successful. So whether that's connecting IoT devices and sensors to supporting dynamic workloads and machine to machine communications, these environments are really becoming the frontline in the AI era. So in this session, we're going to drill down a little bit. We're going to try and clarify some of the differences. There's been some confusion around networking for AI, AI for networking, right? The network for AI, building out those networks. And then, AI for networking, using AI to optimize and simplify operations. We're also going to touch on some best practices that Grant's seen, maybe some problem areas you want to avoid and how agentic operations can really help future-proof those distributed networks. So let's get started. Let's keep a little bit high level to start, Grant. How is Cisco networking evolving to support agentic AI across campus and branch environments?

>> Absolutely. I think you mentioned it really well at the start is that we've been talking quite a lot about the impact on the data center and where we're putting these workloads. I think Jeetu's mentioned this in the past, that's where all the digital workers live, right? But there's a collaboration between each of us, whether we're here in the office, in the studio, anywhere else, and those digital workers. And if the data center investment is the smoke, right? It is the precursor to the fire that I think is going to hit campus and branch networks. And we're seeing this across every single industry and company size that we talk to. And I think to your core question, there's a couple of things we have to prepare for. Certainly, we have to make sure that the networks themselves are built to the size, the scale, and the rigor that's required when we're 10x-ing the workloads that are running across them. That's a critical thing. You mentioned IoT devices, the complexity of the things in the applications that have identity on the network and how they interact with each other and are starting to interact autonomously at a machine speed is another critical thing. So that's going to bring in identity. MAC addresses, right? How do I handle that explosion? And even just, who can talk to whom? And I think that leads us to the third piece which is, how do we secure and protect it all? How do we ensure that every single one of those packets and transactions has a purposeful policy that we can enforce attached to it? So you really have to look at every layer. And that's not even getting to how we have to manage these networks, right? The complexity and the scale in terms of how it is impacting us and our teams, right? That's the next lever we're going to have to pull on.

>> Yeah. No, I think that was a great overview. I love the fact that you brought up security. Because now, there's going to be a lot more endpoints that are going to be connecting to the network. That's always been a challenge to... And make sure that you're enforcing the policies that you've created at that edge environment so I think that's going to be super important. You talked about a few. But I'm wondering if there are any real world examples where you've seen the campus and branch networks already enabling some of these AI-driven outcomes?

>> Oh, yeah. Absolutely. We can think about that from both the workloads themselves that our customers are driving. But also, the new applications that are coming on board and then how do we account for it? If I think quickly, we had a customer session at one of our recent Cisco lives and we're talking to customers. We asked them point-blank like, "Is AI really impacting your network?" And we had a major pharmaceutical company say, "Hey, look. We just upgraded the core of our network to 400 gig. The AI models that our researchers are running, on-prem not in the data center at the edge of the network, are already outstripping our network capacity, because of the size of the datasets and the images and everything else we're running." So it is having a today impact on the capacity that they're able to drive and their ability to get new treatments to market, right? Fascinating. I think if you also then start to look at the smaller scale. We were talking to a drive-through restaurant. And for them, the drive-through the number of vehicles you can push through your drive-through window in an hour in a day is directly comparable to your revenue. And so, they're battling staffing challenges. They're battling market challenges. And so, they're starting to deploy small language models in the drive-through speaker to do transcription, translation, quality. That is an amazing impact and, like, "How are we managing the access to that? How does it interact with my point of sale? What happens when it breaks?" Yeah. Amazing. And I'm-

>> Yeah, no. I think those are great examples at both a larger enterprise with lots of data and at a branch office being able to figure out how they're going to leverage this technology and what the use case is. Because that's really what everyone's trying to figure out right now. They all understand, "This is good." Now they're just trying to figure out, "How do I apply it to have some kind of a meaningful impact on my business?:"

>> Yeah. Absolutely. And then, "How do I measure the impact at the end of the day?"

>> So I mentioned in the intro that, obviously, there's been some confusion when people are talking about AI and networking. Is it networking for AI or AI for networking? So I'm wondering if you can, maybe from the Cisco perspective, explain the distinction and why it's not one or the other but having both of these are going to be so critical moving forward.

>> Yeah. It is both. I don't think there's a conversation that I've had, with partners or customers or folks like yourself, where those two things don't go together. So when we talk about rebuilding or re-imagining the network for this AI era, that gets to a lot of what we've started to talk about, is how am I building the infrastructure to handle the capacity, the latency, and even some of these workloads by bringing those things back and running them at the edge of my network? And ensuring that what I build today is in place for whatever we invent, five, six, seven years in the future. And so, looking at, "How am I securing the infrastructure, managing the capacity, and get the agility to grow it and address it as I grow?" And then on the flip side, we talked about the complexity of managing all of this, the need to drive, not just AI for the network, but more of an agentic operations model across all of my domains is critical when I think about, "How do I prepare for this level of scale?" And being able to take advantage of this agentic technology to scale out my team, skill up my team, teach the next generation best practices. But then, make sure that I am keeping on top of the consistent and constant change. Because I think there's one piece of this that we miss frequently and it's that all of these models are not deterministic, right? I can give it the same set of inputs. I can give it the same set of questions Day 1, Hour 1 to Hour 2, I may get a different answer back. And the same thing applies to network conditions. When we are constantly iterating, we need machine speed and machine level assessment of what's happening to then bring that back to me and present it in a digestible way so I can make more informed decisions about how I optimize.

>> Yeah. And I think that last piece that you just mentioned is so critically important, right? Because that's the human in the loop. So for a lot of people, they hear about the concern, "Oh, AI is going to take my job." AI is going to help you to do your job a lot more efficiently so you can focus on the much more important tasks, the strategic initiatives that are going on. And also, the much more complex problems that arise. But it's going to take out a lot of that mundane, routine activity and it's going to be able to provide the context that much faster so that you can solve those problems so much faster as well.

>> Yeah. That context is key. Every organization has their Tier 1 sites where the level of trust for automation is pretty low. And then, they've got their Tier 4 sites where like, "Yeah. We can trust some automation along the way. But if we are not receiving from these tools the context for why a decision is recommended or why a decision was made, we have pages and pages and stacks and stacks of business processes, compliance needs, reporting that we are still accountable to, to make sure we are operating our networks efficiently, compliantly, ethically." And being able to interact with the tools and get that push interaction for, "I need some guidance," that's kind of the nirvana that we're pushing for here.

>> No, that makes a lot of sense. We talked about IoT a little bit and the machine to machine. How are these dynamic workloads and all this machine communication really reshaping the network requirements, especially when we're thinking about the edge and highly distributed environments?

>> I think it comes down to probably two aspects of the same word which is speed. We need a speed of data transmission, we need a speed of communication, and we have agents interacting with each other. We are used to dealing with latency at a human scale of a few milliseconds here and there. You expect a response from someone. You expect a response from a tool. At machine speed, you're starting to talk about almost microseconds in terms of driving decisions. That means your gap to adjust is pretty small. And so, that has multiple impacts across the network. What latency am I delivering both on the network and across my networks? What is my margin for error if I need to do any correction of the data along the way? And so, that's the speed of the network. But then, that second aspect of speed is what we were just talking about is, what is the speed of decision making, the speed of correction when something is coming along? So how do I start building trust in my tools in order to get that level of adjustment and optimization? And I think that's what brings in the final element of what you really need to deliver a network for the AI era is that idea of end-to-end assurance, right? When I have a machine speed platform that can both accept and adjust that data at this pace. But then, start making sense of it and providing guidance, recommendations, and optimizations along the way, that is really what's going to help all of us prepare for and get ahead of this deluge of change that's coming over the next five years.

>> Yeah. I think that makes a lot of sense and I like the way you talked about the trust aspect of it. I think, for myself, I've often referred to it as, "What's the time to comfort with these technologies?" This is a pretty significant shift to go from, "I've just got an alarm storm and I have to figure it out myself," to being able to have technology that actually walks you through, can provide the context, can simplify the workflow and accelerate it. Even add in, in a lot of cases like you had mentioned, automation to it. But there's obviously a time period where organizations need to validate it, need to trust it. I can't tell you the number of times I've heard people say, "Yep, I've been using my AI and it gave me this... It told me what it was on Friday night but I didn't trust it. We worked on it all weekend. And when we finally figured it out, it was right." But again, it's going to take time to do that. So that's why I always tell people, "You need to get started now because it is going to take time. You will realize the benefits." And especially as you start adding more and more of this machine to machine languages and so forth and data, right? Everything is going to move that much faster. So you need to get started and you need to get that trust established.

>> Yeah. Absolutely. We talk a lot about the... One of the core tenants of agentic ops as we're looking at it and AI campus is this idea of multiplayer. It is one thing to have a one-on-one exchange with an individual standalone agent. It's helpful. It's interesting. Sometimes it's fun, sometimes it's quirky. We really don't want quirky in networking. But when I can start pulling you into a conversation, bringing in my security team, and we're looking at and evaluating a recommendation, a decision, a anomaly along the way, and then bringing in other players, it starts to not only build trust but start to push trust back into this system. Because that is the other power of this agentic model is it is learning along with us and we have the opportunity, and in many cases, the responsibility to teach that trusted operating model into it. Like, coach it towards our organization's standards and guidelines for how we operate.

>> Yeah. I really have been impressed with the multiplayer capability to... Instead of going to a war room, being able to invite other team members in from potentially other domains, right? Networking, security, observability, whatever that group is, applications teams, to come in and be part of that agentic workflow is something that's, I think, going to really help speed up that operation and the response. So teams aren't fighting and pointing fingers anymore, they're leveraging the contextual responses delivered through the technology to be able to accelerate the resolution.

>> I mean, we are on the cusp of being able to get to a truly blameless retro when things go on or at least blame the agent.

>> Yeah. Exactly. Exactly. Let's talk a little bit about some of the architectural and operational considerations. When you're deploying networks for AI and you've got agentic operations in the campus and branch, what can you recommend to organizations? What should they be thinking about?

>> Well, I think the first thing to think about is there is the practical reality of making this transition. And just like any of the other major technology impacts that we've seen, if we go all the way back to when we graduated from college, that whole internet thing that happened through mobility cloud and, now, on the cusp of AI, one thing is true in all of those is that there's an element of evolution and not just a forklift change to what we are doing. When you've got decades of infrastructure and process in place, you can't just magically snap your fingers and expect everything to change. Particularly if, say, one of our examples from earlier, you're a hospital or a pharmaceutical company. I have devices that are mission-critical that are 15 years old. They're not going anywhere. So I think that major consideration drives a few other things is one is, what is your long-term plan? What is the point you're trying to get to? And then, what are the key milestones along the way? And for what we're seeing in many cases is the first and most important thing to do is to get your physical networks as well as your management and your assurance into a single platform. By consolidating all of those things together and giving yourself the visibility, the reporting, and the control, that, for many people still a major architectural change, will get you to the point where you are best positioned to adapt to whatever is next. And then I think from there, you're starting to look at, how do I layer together the structure of my network? I think a fabric-based network architecture is going to be the next step for a lot of customers because it's going to help them deal with those other challenges we talked about: device proliferation, agent and identity proliferation, and policy management. And so, you're fighting for simplicity along the way. And then once you have that in place, then yes, it becomes a problem of bandwidth and latency and what do I need to prepare for?

>> Yep, yep. No, I think those are some great pointers there. And I think especially we've seen the shift towards platforms as well and the recognition that, if you are going to do this and you're going to be able to provide context, having a platform that can cover the end-to-end environment and then, obviously, also some adjacent environments are also going to be just adding more value to what you're able to eventually pull out of that system as you become more comfortable with it, as you trust it, as you understand the value it can bring for organizations.

>> That's right.

>> Very good. So I'm wondering if we could touch upon maybe a little bit about how organizations should assess both the networking for AI workloads and the AI for network operations. We talked a little bit about it just now. But obviously, they need a platform, they need to be able to pull in from other groups. Is there anything else that you would recommend for organizations to be looking at?

>> It's easy to think that this is a network only conversation, right? We've gotten very, very good as an industry at planning for what's next, preparing for what's next. But I think, now more than ever, one of the most important conversations you can have is starting to really proactively reach out to your business, understand what are the changes to customer behavior, to the applications you're going to be asked to support, and even the new systems, so you can get a sense of what capacity and support looks like. We are already seeing the convergence of traditional IT networks and those manufacturing, occupational, operational OT networks. Layer on top of that, the way we're changing office space in a highly collaborative world. There's a hotel just down the street from our San Francisco office that has fully autonomous delivery robots that are working hand-in-hand with the catering staff. Okay. Now, I have safety challenges, timing challenges, service level that's dependent on something else entirely. The network supports all of that. None of that works without the network. So if you don't have that view of your next three to five years in terms of, "What might we have to do?", you're going to miss something critical along the way.

>> Got it. Yeah. No, that makes sense. And I've been to a couple of the offices that you've re-imagined as well for the future of work and it's... I have to say, it's impressive. And like you said, the network is taking on so many more roles with PoE and so forth. So lighting, shades, et cetera. Really playing a much bigger role in infrastructure within the offices as well.

>> Absolutely.

>> So this has been great. We got to wrap up. One last question I wanted to ask you. Everyone, all the people who are watching this now, are trying to figure out, what are the best practices that decision makers should follow in order to ensure that their campus and branch networks are future ready for the AI era? Anything else that you would recommend to them?

>> What else would I add? We don't always have the luxury of planning perfectly ahead. I think the old adage is always true, "Don't bet against bandwidth." We know that surge is coming. And one of the most interesting things that we are seeing is, and I think we'll talk about this on the data center side as well, is that there's the shift from training to inference that is happening. Particularly with these AI applications. And what's interesting to me is that training loads are very spiky, right? I'm going to shove a large batch of data into a workload, it's going to process for a while, it's going to give me a result. Big, almost logarithmic spikes, in usage. But then, it comes back down again. As we get to the point that all of us are driving multiple concurrent agents across an entire, whether I'm tuning the network or trying to write my next PowerPoint presentation, the level of inference starts to... It creates these plateaus where the level of utilization across the network is going to dramatically shift and we're not going to have those peaks and valleys anymore. It's funny. I started my career way back in contact center and I still have Erlang calculations kicking around in my brain. It's like it is a new... Not port capacity necessarily. What is our network capacity when I don't have downtime anymore? And I think that's the consideration. If I'm constantly doing inferencing, I'm not sure who's driving it. It changes the nature of capacity. It starts looking at, "How am I planning planned downtime? Can I afford planned downtime anymore?" So it creates the availability of the network in a completely different light.

>> Yeah. No, that's... The other important factor there is it's not only the bandwidth for the AI inferencing. But also, keep in mind, everyone's still running all their mission-critical apps across these networks to these branch and campus environments as well. They want to make sure it's optimized as possible as well.

>> Absolutely.

>> Yeah. Awesome. Look, Grant, this has been fantastic. Thank you so much. Lots of great information. So appreciate you joining us. Shedding some light on the evolving role of Cisco networking in the campus and branch. I think we've learned a lot here. Especially the distinction between networking for AI and AI for networking, how that's really shaping strategies, why dynamic workloads and machine and machine communications are really driving a fundamental shift in the network design, what organizations should consider as they're implementing agentic operations as well. And I think for the audience, the message is clear, right? This whole networking for AI, AI for networking, it's no longer confined to the data center. Campus and branch environments have to transform, especially if enterprises want to scale AI securely and effectively. So again, Grant, thank you for joining me.

>> And thank you for having me. I'm looking forward to the rest of the event.

>> Absolutely. And thank all of you for joining us as well. Please stay tuned for more to come.