AI Wants the Keys to Your Network

0:00Genuinely put an AI native network to test in a small clinic, in a small school, in a small building. One building is all it takes. Our track record is 95 % of the time when somebody goes into a production pilot with us, they never look back. That's why we only ship new equipment to all new customers for pilots, because our track record is 95 % of the time. It does not come back because people can't unsee what they see with the self-driving network. It is that fundamentally different. For the last few years, the enterprise AI conversation has largely revolved around models, data, and compute. But as AI shifts from something we all use to agents that operate constantly and on their own, another part of the AI stack is becoming impossible to ignore, the network.

0:52Today, we're asking what happens when the network itself becomes intelligent, able to observe, optimize, and eventually act on our behalf. We'll be talking with HPE Senior Vice President of Products, Sadir Mata, and WWT Senior Director of Networking, Christine Fierro. They'll explain why AI is making the network more critical to the business and what it takes to make that network intelligent, trusted, and increasingly self-driving. This is the AI Proving Ground podcast from Worldwide Technology. Let's jump in.

1:29We're talking about the self-driving network here as it relates to AI. I mean, enterprise AI discussions have tended to focus on models, data, compute, things like that for years now. Sadeer, just to start with you, tell me, build me a case for why the network should be treated as, you know, an equally important component of all this and not just, you know, something that comes after the fact. The challenge has been, historically, you know, you would say, hey, is the network relevant for your business? And every CIO would say, of course it's relevant. And, you know, what has happened in just the last 24 months is with the advent of AI and agentic AI, just the traffic workloads are going up, the criticality of AI in the enterprise from an operating system perspective is going up.

2:15And so, yeah, we're seeing the network has never been more relevant. And again, think about the types of places we all visit. You go to a grocery store, you don't want to be an extra nanosecond waiting in line. You know, if you're a student taking an exam, if you're a doctor walking out of surgery, the network has never been more critical than it is. And AI is pushing the boundaries of it. So it's a great conversation to have on the relevance of network. Christine, how are you seeing it? Even just thinking about, like you said, where we work. Right. So not just if you're a student or something, but I'd say post pandemic.

2:51Right. We've even a little bit before, but definitely after we work from coffee shops, we work from airports, we work from airplanes. That one's a little iffy sometimes, but it's where we work. But then also, I think, too, just the traffic itself and the pattern has changed. So bursty human traffic that's a set kind of during time of day to constant AI agent workflow, whatever traffic is very different today versus five years ago. So the way we are thinking about it is there is basically two types of networks you're building in an enterprise. These are, there's networks for AI, for the AI workloads, you know, in the data center with security, you know, built in and not as a bolt-on.

3:34And then AI for networks in the enterprise workspace in the campus and branch space, right? So, you know, networks for AI and AI for networks. And so if you think of the AI for networks, which is where Christine and I spend a lot of time on, genuinely, the traffic pattern has changed. The criticality has changed. I'll give you an example of a retailer. They used to have, you know, two backup van links. Now they have five van links. They literally will not operate their network, their store without the network. And so they just have to figure out a way to get the network to be constantly available.

4:11The number of connected devices in any one place. I mean, think of what is happening as we sit here. The number of users, devices, applications, and bandwidth consumption is growing exponentially. Now to that, you add agents constantly, 24-7 trying to leverage the network. So it's the continuity, the frequency of how often the network is used and how much the network is used. All of that is changing. And consequently, for us, in this growing world of more users, more agents, more applications, more bandwidth, more consumption, more, more, and more, what is the one thing that is not growing? Nobody is getting an exponential growth in their IT team, right?

4:58The IT staff is not growing. So how do you handle more and more and more of happening every day with maybe the same size team or even fewer people in many cases? that's where AI makes a difference. If AI materially does not improve the operational realm of your infrastructure and the team's functions, then you're not doing it right, right? And so that's where we are seeing demand for doing more with less every day. It's not just having AI and unleashing it and doing those things. You have to have that trust there, and there has to be a timeline or step level or whatever to implementing it and utilizing it and having it do more.

5:42So like you gave great examples, but nobody's going out there today and just unleashing it on their network for full control. It's a trust thing that builds over time. So maybe let's put this into a real world setting here. I was reading an article recently about a retailer. It happened to be one of the world's largest retailers that everybody would be familiar with, but they have this vision or they're putting into action where AI is identifying what's soon to go out of stock so that they can anticipate ahead of time what they need to stock. So my question is going to be, as AI becomes part of keeping the shelves or whatever the business's value to their end customers might be, as it becomes keeping those shelves stocked and orders moving, what needs to change about the network to actually make that reality happen?

6:30Yeah. So let's actually lay out a blueprint of how should, if I was a CIO today and an enterprise IT leader, what's the blueprint for getting from here to there to a network that's always on, that's able to take in this new agentic workload and the business demands? So it starts with, for us, fundamentally, we've believed in, you can't improve what you cannot measure, right? You've got to be able to measure. So we started this company, Mist, on the fundamental premise that let's measure user experience. So first, always the question we would ask when we get to an existing legacy customer and say, we usually say, hey, how's your network doing?

7:16They'll say, yeah, it's actually okay. And the very next question we would ask is, how do you know? How do you know users are having a good experience just because they're not opening tickets? Think of yourself. How many times when you went to college, you actually cared enough to open an IT ticket, right? Were you happy, right? So step one is measure user experience. If your new network cannot measure user experience, you don't have the right network. Step two, once you can measure user experience every user, every minute, then, you know, are you able to actually improve it, right? You've got a foundation.

7:51So we have an interesting model that we launched. It's called Shapley. and what Shapley does is it'll tell you real time, you know, if you're having a good Teams or Zoom experience, right? So we launched the model for our worldwide customers and candidly, I mean, our own, HP's own network, our own corporate network looked blood red on people having bad user experience on Teams Zoom. First, we were like, oh my God, that cannot be true. There's something wrong. No, it turned out to be true. And so once you know, then you say, okay, what do I need to do factually do to improve it? Right. And so that's where AI comes in and say, hey, you know, maybe it's your user going on VPN when they shouldn't be going.

8:32Teams call shouldn't be happening or VPN. Maybe there's a better RF design, something else. Right. So measure the experience, then know how to improve it, then make the changes. Right. So today humans making changes. But where the true north is, is if your network genuinely cannot self-drive for lots of, for major parts of it, then you still have to throw humans at it. And you don't have that kind of, in this ever-growing demand workspace, you don't have the team to throw at it. So measure user experience, know what to make changes on, make some changes. Humans are needed still in the loop and some changes are self-driving.

9:12That's been our blueprint. And so as you think of a network that is for the future for your AI workloads, start with knowing if you're actually making a difference. Sadeer, you had just mentioned it. I've mentioned a number of times now, self-driving network. It occurs to me that that's a very compelling phrase. It can also feel like marketing. I've certainly seen some of the HBE commercials out while watching TV on the weekends and whatnot. But level set up for us on exactly what is being self-driven in the network and why that's important. So what should be first? What does it mean? Like you said, what does a self-driving network mean?

9:49And what are the outcomes does it drive? What will it actually give to me? What is a self-driving network? I'll give you a great example. One of our customers is actually a large university in Atlanta area, very reputed technology institution. And so the engineer there, Nathan said, Sudhir, I don't believe I've been hand chiseling this network for these, you know, 800 person classrooms for the last 20 years. I do not believe that AI can magically do what I was doing because it takes a lot of understanding of when students come in, when students leave, the size and scope and scale of this network and everything.

10:27And so I said, look, Nathan, you know, let's do this. just replace without pulling an extra cable, without adding any extra access points, just replace the same number of APs with the same number of APs of MIST and step away from the keyboard. So now let me tell you what should a self-driving network do. A self-driving network should first self-organize, meaning it should pick the right and best parameters, the most optimized parameters to operate the network in. Historically, in this university's case, Nathan was doing it. In our case, humans should not be doing it. AI can see packets on the air better than humans can.

11:06So step first, self-organize. Then self-optimize because the network is a living, breathing thing. So continually make it better based on this notion that I know what is the measure of user experience. If I make it better, I should know that the experience is becoming better. That's one of the most foundational technologies of our generation is reinforcement learning. And so we've adopted reinforcement learning nine years ago to say, not just self-organize, then self-optimize. Now, when you self-optimize, there are things that can go wrong in a network, right? It could be because somebody changed config or it is because a software process failed or a hardware process failed.

11:46That's where I want the network to self-heal. Self-organize, self-optimize, self-heal. And today, all these three parts of the stack, you take any missed network in the world, Those are built in, you know, no extra cost into the base product. That's the entry level to the ticket. And then comes, okay, maybe there are some things, you know, will a hospital network let me change their configuration because, you know, just because I think that their config should be different. That's where what Christine is talking about comes in is trust. So we've created this notion of these are the things that we could actually self-drive for you, but we'll let you give us permission to do it.

12:29And so, example, imagine you have four devices connected to an Ethernet switch. Three of them are happily working. One of them, their VLAN wasn't properly plumbed to that switchboard. I see that that device is trying to communicate. The packets are getting dropped and black holed. I could go add that VLAN into that switch config right now and self-heal and make the device work. Well, what if the network engineer designed it that way and there was nobody supposed to be passing packets on that VLAN, right? So that's the piece where trust comes in and say, you know, yeah. So self-driving networks, these are self-optimized, self-organized, self-optimized, self-heal.

13:07And then self-heal has a second element where we could actually change configs and optimize. Maybe somebody made a mistake and revert a change. That's where trust comes in. That's sort of the anatomy of it. And then I'll pause. You know, one of these days I'll stop talking. This is what Christine was warning you about when you bring me on a podcast. But how does this actually manifest to an end user, right? How should, what should you expect? Let me give you three examples. Literally, you know, one of the world's largest retailers, one of the world's largest financial institutions and one of the world's largest health systems in the world are all three of them are wwt and hpe customers right i wish they would allow us to sell their names but you know set that aside for a moment every single one of them you know by as measured in by their own internal systems when they went from a legacy network to this new mist network every one of them said we had seen anywhere from a 70 % to 90 % reduction in incoming user complaints on the network.

14:10That's it, right? If I just need happier users, the network should just be invisible and disappear in the background if self-driving networks are actually working. That's our goal. That's how we measure it. This episode is supported by Equinix, where your AI runs determines what your AI delivers. And Equinix Distributed AI is built for exactly that. Equinix brings together your data, clouds, and AI ecosystem so you can run AI wherever your data lives, securely, at scale, and without lock-in. From inference at the metro edge to data sovereignty and high-density GPU infrastructure, Equinix gives enterprises the freedom to run AI everywhere it matters.

14:48Is there a practical progression here from what customers have right now to AI in the networks that observes, to AI that recommends, and then ultimately to self-driving networks that actually act on our behalf? Yeah, I mean, there's been a lot of innovation, right, over the last few years, and just that continual improvement of, okay, what can AI do in the network? And so it's starting small. And it's just taking, you know, one or two things that maybe are an easy thing for you to ask AI to do, or see if they can do, right, and verify that it's correct, and then trust, you know, build that trust that it can do it, and then just keep building from there, right?

15:30So none of this was done overnight. None of this happened just instantly. And everybody was like, here, I have the network. Let's go for it. It's come over time. But trusting that you can hand things off and not just being stuck in the mud. Like me, sometimes I'm like, I don't like change. I don't want to do this. But being willing to change is where it's key. And so we have seen that, right? Just find out maybe one or two things where you can start small, something that won't take down your network, right, but is a change that can be meaningful, build that trust, and then continue to build that over time to ultimately get to that, you know, self-driving network phase.

16:10And so, Brian, you know, if anybody listening is saying, well, this sounds all great, but, you know, this is not my reality, right? My reality is my business is demanding more from the network. My CISO is demanding, you know, in the world of, in the age of mythos demanding more security from the network, where do I even start? What's the first logical step to take? And again, this is where, don't take our word for it, genuinely put an AI native network to test in a small clinic, in a small school, in a small building. One building is all it takes. Our track record is 95 % of the time when somebody goes into a production pilot with us, they never look back.

16:57That's why we only ship new equipment to all new customers for pilots because our track record is 95 % of the time. It does not come back because people can't unsee what they see with the self-driving network, right? It is that fundamentally different. And so it starts with a pilot. That really starts simple. Where do we go from here? How do we continue to improve the idea, the concept and the potential of the self-driving network? What is it not capable of doing today that it could potentially do down the road? Yeah, I would say, to be honest, again, we are in the early innings of the possibilities here, right?

17:39So the question is, you said, so what can't we do today that we will be able to do in four months, in six months, in 12 months? An example is, let's say I push an approved change in a maintenance window, okay? Today, by design, we don't go and say, hey, you made a mistake. This one application in these four locations is not working. I'm just going to revert the change for you. But we're very, very close to basically telling them, look, we have digital twins on the network that are constantly validating the network. We call them Marvis minis. The moment you push a change, Marvis minis take some and say, hey, you've tested 99 percent of your applications, they're working great nationwide.

18:21But in this one particular state for this one particular application, that's failing. We just alert them. But in reality, if I were a hospital network and do I want somebody to have a change that has broken an application out there for four hours, can a self-driving network revert that? We need to gain the trust of customers. So it really, it will come down to config changes. There's more and more config changes we're going to take off of customers' hands. I'll give you an absolute real-world example. I wish we could mention names of all of these things. An EV manufacturer had a four-hour manufacturing site outage.

19:06True story of what happened. What happened is an engineer used AI to write a model for some business application. Everything was okay. All his agent did was it was spewing 3 ,000 MDNS queries on the network per second. It was literally blasting multicast traffic on the network, basically took down some of their wireless infrastructure and had a four-hour outage. That was AI-generated code for the most part did okay, but this one thing got snuck into the program. This is a real customer, a real outage that we just lived with last week. This is where security and AI and networks have to be all in one thing.

19:52Trust is not just trust on AI, but security has to be built into all of the networks of the future. I was reading a recent report. I'm looking at it here. Agents could account for more than 90 % of AI token traffic by 2035. I mean, whatever that exact number turns out to be, how should network strategy shift and change and alter when one employee can initiate that kind of chain of events that can lead to adverse outcomes? Yeah. I mean, I think it's similar to how we treat human traffic today, right? Like you have to verify, you have the identity piece, you have maybe the segmentation piece, you have like, what can you access?

20:31What can you not? Where can you go? Where can you not go? So I think we have to evolve to have that same thing for the non-human aspect, right? Like for that agent piece and, you know, just letting an agent have free reign throughout the entire network and access, you know, to all the applications and all of the things and all of the data, like, that's not great. Like, that's probably not the best thing we can do, but applying those policies and things that the network doesn't apply or doesn't create the policy, they'll apply the policy, right? So to Sadir's point, we still need to work with the security counterparts and the other parts of the business to create and define what those are and then implement them to ensure we have those guardrails for the agent so that they're treated, quote, like a human and they can't access things they're not supposed to or things like that.

21:18And I think that's the, I think part of maybe, I love your point, but we're not quite there yet, but we're figuring it all out because like Sadir said, we're still so early in kind of all of this. To use a football analogy, I feel like we're at the 20-yard line. There's so much more we could do. And, you know, so two things we have done. First, you know, everybody gets it. Everybody says, I want to build a zero-trust network. I want to be like a coffee shop. I want anybody to be able to plug in anything and do no harm to my network, right? Historically, that meant you've got to redo your basic, complete network fabric and deploy a whole EVP and VXLAN kind of fabric and zero trust and segmentation and tags and this, that, and the other.

22:03And that hill was too steep to climb for many, many, many, many customers. One of the things we launched as in preparation for this agentic native network is something we call zero trust inline segmentation, meaning without re-IPing your network, without deploying fabrics in every retail store, in a medical clinic and every school district, without having to reinvent the way you've always deployed the network, can we get this inline segmentation seamlessly, simply delivered through the cloud? It's truly one of, I feel, one of the best innovations we've put out there. We call it zero trust inline segmentation.

22:44In concert with this agentic network, I think it's game-changing. So segmentation is a non-negotiable, right? If you read about how somebody got through a fishbowl at a casino, the fishbowl had a thermostat, and the thermostat was connected to the network. Through the thermostat, they were able to get to the high roller database and hack their entire network through a fishbowl. So micro-segmentation needs to be ubiquitous. Zero-trust needs to be ubiquitous. Historical solutions have been hard. That's one of the things we've done a good job of simplifying it. Right? Yeah. And then the second part of that is how do you actually measure?

23:24I mean, cyber is certainly a leading conversation amongst pretty much every stakeholder and customer that we encounter. But increasingly, cost is becoming an equal conversation as well. I am curious here, I mean, if we're talking about token limits and tighter boundaries on what AI tools can access, I mean, in networking, how can you ensure that you're giving AI access to all the right areas and tools and whatnot, but also making sure that you have the ability to rein it in in case it just runs and runs and runs. And all of a sudden is landing you a massive token bill. MCPs have become the new buzzword.

24:01I mean, literally, I was using Cloud Cowork the other day. In like a two-hour window, I spent$45 in tokens. And I'm like, damn, that was unnecessary. So anyway, so what are we doing is what we've done is over a 10-year lifecycle, we've built a lot of models. And so if you look at the tech stack of what goes into a self-driving network, here's the tech stack. So what we've done is we first measure every user every minute data. So we have the data. Then over 10 years, we've built models to say, hey, is a user having a good experience or a bad experience? Is that happening every Tuesday at five o 'clock that happens and that's normal or no, this is an unusual Tuesday and somebody is having an anomaly and we got to alert and wake people up.

24:52So models are important as part of the text tag. And then on top of this, basically think of the models as the models have become tools in the text tag. So when an LLM wakes up and an agent wakes up, it is now, you know, basically the conversation is happening in text. That means it's happening in tokens. And so you've got to be very careful is you can spin a lot of tokens very quickly. So what we have done is we basically, there's an agent gateway and an agent orchestrator. And to that, you give the right skills and say, let's say someone says, hey, the Wi-Fi sucks in this building. That's good enough.

25:28That's a great start. will say, we'll take that. How would a human troubleshoot a Wi-Fi sucks problem? Because it could be the wireless, it could be the device, it could be the wired, it could be the van, it could be the application, it could be any of it. So do you throw tools at all of them and go send LLMs across all of them? So we have written skills for each of these types of major problems. And then the skills have defined tools and the tools, basically you do a lot of prep work before you dip into an LLM and tokens. And so today, what I can tell you, you know, factual, actual, yeah, bring it on.

26:06A customer can connect an MCP service that we've launched on our cloud and use their own tokens or use Marvis natively, our AI engine. Marvis is consuming less than 50 % of the tokens we've seen for the same question to get better efficacy. We're seeing, in many cases, more than 60%, 70%, 80 % better efficacy than any foundational model out there. We use foundational models. Nothing wrong with them. But having the right skills, the tools, and leveraging the right agentic gateways to shape the answer has really helped us essentially not pass a raw token cost to our customers and really fine-tune this.

26:53So this is very thoughtful, very methodical in how we're approaching it. And that's the only way because the token cost is only going to continue to grow. Yeah, that's interesting. Christine, are customers that you're interacting with, are they connecting the dots there that this could be kind of a leading edge for how to control costs or make sure that you're making the most efficient use of your AI? I'm not sure that that's like a true dotted line or a true solid line versus a dotted line, right? I think it's thought about and it's part of the conversation of especially some things when you introduce it and it's like, okay, if I do this, but am I going to have to pay a license fee or something that is now going to be token based, right?

27:37Because I mean, like you said, we all use Claude or any other kind of AI and we start to think now of, oh, when I do this, it's going to cost me money. That same question is coming up in conversations around, okay, well, if I utilize this, how is that going to impact the license or recurring or whatever price or fee that I pay? And is it going to be additional or is it just part of what I already own when you purchase the equipment? So some of that I think is still kind of being figured out and determined and some of it is already and doesn't have an additional cost. Well, as with everything with AI, Sadeer, like you mentioned, we're on the 20-yard line.

28:13We're in the red zone, but a lot of yards left to be played before we see any actual progress here. Sadeer, Christine, thank you so much for taking the time. This is an enlightening conversation, one of my favorites for sure. We'll have you on again soon, I hope. Okay, thanks to Sadeer and Christine, AI is putting more pressure on the network than ever. More traffic, more agents, more decisions, and more risk. The opportunity is to make the network smarter at the same time so that it can measure, optimize, and eventually act on its own. The destination may be self-driving, but as Sadir and Christine made clear, the path there is built on trust.

28:49This episode of the AI Proving Ground podcast was co-produced by Nas Baker and Kara Koon. Our audio video engineer is John Knobloch. My name is Brian Felt. Thanks for listening, and we'll see you next time.

29:06you