Megaport Connect Cloud Networking Summit

Keynote: Cloud Networking for AI Architectures – Full Transcript

Host: So our first speaker today is Roy Chua, and he’s the founder and principal of AvidThink, an independent research and advisory firm focused on infrastructure technology.

Roy has spent the past 20 years working across network, telecom and enterprise infrastructure. Today, he’s going to talk about how AI is changing network architecture and what it means for the future of network engineering. Please welcome Roy Chua. Here he is.

Roy Chua: Thanks. Appreciate it. Alright, good morning, everyone. How’s everyone feeling? Excellent. Great. And for all the viewers out there, hopefully you’re having a good cup of coffee, or maybe you’re just waking up. I hope you’re looking forward to learning about what’s going on with network engineering.

As you heard, I run a small research firm called AvidThink. You’ve probably never heard of it. I’d never heard of it before I started it. We used to be part of a larger company called SDxCentral. It’s a media site, which some of you may know. It used to be SDNCentral, software-defined networking, in the early days, and I co-founded that site around 2012. In 2018, we spun off the research group, and then we had to pick a new name because we were not a media site, and that’s how everything came to be.

I’ve been doing software-defined networking, virtualization and all that for the last 15 years. But in reality, I started my career back in the mid-90s. Yes, I am that old. I was at Cisco, writing ATM firmware. And ATM is not the thing you get cash out of. Obviously, that ATM failed. It was Asynchronous Transfer Mode, which we thought was the future back in the mid-90s, and it wasn’t the future. That’s what happened.

But anyhow, I’ve been in networking for a little while. One of the privileges, or one of the downsides of the job, depending on how you look at it, is that as an analyst you get to go around to all these events and conferences around the world. It’s part of the job. That’s what we do. And so you get to hear about what’s going on out there.

At the conferences and events, it’s not so much the presentations. Presentations are really important, like the ones here today, but it’s the conversations that you have with people. We are privileged to have conversations with all the different people across all the different domains of networking. I talk to carriers. I talk to enterprise users. I talk to data center network operators. I talk to the folks who run the middle mile.

One of the things is you don’t have to be super smart at it. You just have to listen and take notes. And eventually, when you talk to enough people, you get an idea of what’s really going on. So today, what I’m going to do is share with you some of the… I wouldn’t call them insights. Observations, perhaps. I think insights may be too deep, it gives me too much credit. But observations and notes I’ve collected from all these places that I’ve been, and hopefully that gives you an idea of what’s going on with networking.

Then at the end, I’ll talk a little bit about network engineering and how AI is changing that, because there are two elements of AI and networking. As you well know, there are the networks for AI, how our networks are going to help with AI data collection and AI processing, as well as carry all the actions and interactions we have with AI systems. But just as importantly, some of you are probably using AI to manage your networks. We’ll talk a little bit about the state of where that is today, and the reality from the ground truth.

So let’s start with networks. What do networks do? They connect compute, data, users, devices, and fake users, meaning AI agents. That’s what networks do. Now, we’ve gone through many eras over the years, because networks are carrying data and application traffic. What applications and workloads are we talking about? And what you see is AI is the fifth era. We’ll talk about that.

And I’ll talk about the different domains of networking, because as we all know, networks are not one thing. They’re many, many things, and it’s getting to be even more going forward. But let’s start simply with what and where networks connect. I’ll just simplify things a little bit, and I don’t know whether this fits your mental model or not. It fits mine. These are the major networking domains.

We have your campus, your branch, and your home being part of it, and homes have become a lot more important, especially during COVID and lockdown. Home was the office in many, many cases, and so the home’s up there. Then you have your access network: your broadband, your fiber, your wireless access, your mobile and so on. Your enterprise WANs, your MPLS links that still exist despite the fact that MPLS is dead. They keep telling us that. It’s not dead. It still lives.

The middle mile, which I’m sure a lot of folks here will be familiar with. The middle mile is important, and becoming more important in the near future. Your backbone, we know what the backbone is, and the data centers that run all your data, hold the majority of data, and where all your AI lives.

Fundamentally, the data center is one of the more contentious topics, might I say, especially in the US right now. So we’ll just leave that be. That’s not something I’ll talk about. But what happened during the cloud era, obviously, is that we pulled data into more and more of these data centers. And the data center itself is actually an ecosystem, and you see that. It’s gone through multiple evolutions, and I’ll talk a little bit about those evolutions going forward.

Alright, let’s move right along. So let’s talk about the five generations of workloads. You see the early internet. That was very, very straightforward. It just carried basic traffic. Some of you may be familiar with the original email systems out there. That was just very basic connectivity for educational and research institutions, governments and early enterprises.

Then along came the web, the cloud, the early SaaS. I call it early SaaS, but we didn’t call it SaaS then. It was the early SaaS era. Then along came multimedia collaboration, which drove the amount of traffic up. And by the way, that’s still the dominant traffic today, if you just count pure bits and bytes, thanks to all those vertical format videos you’re watching on TikTok. And YouTube. It still dominates.

Oh, AI dominates? Okay. Well, maybe in the near future AI video slop will dominate, but it’s still video traffic. It’s not your agentic workloads or your queries, or asking, “What should I work on today?” because I don’t know. That’s not dominant. It’s still video.

But each era has moved a lot faster. From multimedia collaboration, we moved on to mobile, social, modern SaaS, and now what everyone’s calling the age of AI. AI has been around for a long time, but this is the age of generative AI, agentic AI. And every era is shortening. That’s just life now, isn’t it? Everything gets shortened.

Client-server was 10 years, video was 7 years, mobile social, maybe 4 or 5 years, and AI is now compressed to maybe 2 or 3 years, and now we’re in the AI era. And every era had different dominant requirements for the network.

The reason why I focus on workloads, fundamentally, is that as network engineers, one of the things I learned early on was don’t just focus on the connectivity and the protocols, which we tend to do. Is that EVPN-VXLAN? How’s my BGP doing, or my OSPF doing? That’s important, but that’s not the most important thing.

The most important thing is to remember what the network is in service of. The network is in service of the users and the applications. That’s why it’s always instructive and important to remember what’s above there. What workloads are important there, and what are the requirements? What can the network do for those workloads, or what must the network do for those workloads?

Initially, it was just connectivity. Simple. Then it was reachability, data to the DC. Can I reach my data center or not? Then, when multimedia and collaboration came along, we heard things like latency and jitter. And by the way, it wasn’t actually multimedia, it was actually gaming. I didn’t put it up there, it’s a business crowd, I guess, but gaming is really important. Gaming is a form of collaboration, maybe. I don’t know. Anyway, whatever.

And then after that, it was mobile. It was now reachability on your mobile devices. Do I have a signal? Do I not have a signal? Why don’t I have a signal? What is it, five bars, four bars, three bars? Then from there, when we went into the cloud data center and modern SaaS, the focus for networking also went into the data center. What’s my east-west traffic like? What’s my latency like?

And now it’s AI. All we talk about is this word, it’s a made-up word, tokens. It’s not even a full word. Tokens are parts of words, and why are tokens important? But they are. We talk about tokens. Jensen talked about tokens, and so we should talk about tokens.

Alright, so let’s move on to a bit more detail about what the workloads are asking of the network. I talked a little bit about that. But let’s look at the early internet. It was all about reachability. It was leased lines. Do you have Ethernet in the building? Do you have routers? How do you connect that?

We went from 56k to your T1, T3 lines, if anyone here is old enough to remember that. Maybe some of us. Yeah, I see nodding heads. Thank God. I may not be the oldest here. Maybe I am. But T3 was like, “Oh, I got T3.” And you’re like, “Oh my God, I got T3.” Well, now it’s like, you got T3? What’s T3? Anyhow, Terminator, whatever.

Then from there, we went to web, cloud and SaaS, and the WAN became this hub-and-spoke for the enterprises with MPLS, and it was like, I got MPLS. I got private links. I got committed information rate from the Frame Relay days. I got fast links. I got private links. That was really wonderful. And now I got web. I got cloud. My north-south is really good. My north-south links are really, really wonderful. I got security coming in, being very important for web, cloud and early SaaS.

Then from there, we moved on to multimedia. Latency, jitter. I have symmetric flows now because my video goes both ways. Although if you have a lot of people in the party, then a lot more comes down than goes up, but still, a lot more traffic. And mobile social. You’re taking a video, so my upload speeds become important, and for the carriers, the fact that you’re uploading those TikTok videos all the time actually matters to them. Uplink, downlink matters to them.

The reason why I bring this up is that at each stage of the game, as network engineers, whether you’re in the telco, the data center or the enterprise, you’re being asked to focus on different things. You’re being asked to optimize different things and trade off different things. That’s the way it has been, and that’s the way it’s going to be.

When you look at cloud-scale data centers, you’re like, alright, that’s great, I’m going to serve all these things from my clouds. Then you’re like, okay, well, I got east-west traffic. What does the architecture look like? How do I maintain it? Do I go from a multi-tier architecture to a Clos fabric architecture? What’s the nature of what’s going on there? Can I make it easy to manage? Can I go and virtualize my network functions and manage them easily, make them more on demand, just like the cloud?

These are the questions that we ask ourselves, and that continues. And one thing that’s become important with the mobile, social, modern SaaS era is that the hyperscalers became important. I think you’ll see in the presentations later today the importance of hyperscalers, and everyone in the business world now talks about it. How do you connect? Where are the on-ramps? How do I connect to the hyperscalers? How do I connect across the hyperscalers?

And they’ve become big CapEx engines as well. In the old days, it was like, oh, well, they’re taking over subsea capacity. Of all the submarine cables, three quarters of them are built by hyperscalers. And now, with a lot of that investment for AI infrastructure, again, the word hyperscaler comes up. Thank God they have the profits to invest in it, even though some of them are starting to borrow money now. But again, the hyperscalers have become important, even for networking people. I think that’s the reality that we live with today.

Alright. So let’s talk somewhat about hyperscalers, but not necessarily, because the AI data center that we see today is not just dominated by hyperscalers. The frontier labs also run them. Meta runs them. Meta’s got huge clusters. SpaceX, SpaceX AI, SpaceX plus Tesla plus xAI, whatever it is. Whatever its name in the end, it’s Musk Inc. They run huge data centers.

What’s happening from a network perspective is we’ve gone from, alright, in the old days, I just need to know simple routing protocols. BGP, I got my Clos fabric, I got my three-tier, I got my Ethernet fabrics. I’m good. I got east-west. I may even have microsegmentation for security.

And now AI comes along and says, I like your data center from the cloud days. That’s not bad. But guess what? It’s not good enough. And so with another era, from multimedia to SaaS and cloud to AI, things change again. The workload has changed. The workload requirements have changed. And so for the lives of network engineers, we’ve got to go again one more time. Every single time. That’s the way it goes. Workload changes, network changes. And what are they demanding right now?

If you look inside the data center, it’s not a single network anymore. It’s not your east-west network anymore. It’s split into at least three parts. Three-plus-plus parts. You’ve probably heard about this, because Jensen says it on stage at GTC, at NVIDIA GTC, and everyone in the world watches NVIDIA GTC and Jensen. Because when Jensen says jump, we say how high. It seems like that, anyway, whether it’s true or not. Because it’s the most valuable company in the world.

We all wish we had bought NVIDIA stock 10 years ago and held on to it. If we had done that, we wouldn’t be sitting here. Right now, you’d be on your little island somewhere, sipping a margarita. So we didn’t. Clearly, I didn’t.

But the three domains of networking scale are, first, scale-up, which is between GPUs. Direct connection between GPUs is called scale-up, and it uses something called memory semantics. You don’t have to worry about that so much from a network engineering standpoint. It just means that it’s one memory space that the GPUs can share. Okay? That’s scale-up.

Scale-out is something that you’re mostly familiar with from the cloud days. Scale-out is connectivity across racks. If you think about scale-up, it’s scale-up within a rack, because the unit of compute is changing. It went from a server to a rack. Your NVL72, some of you use NVLs and NVLink, an NVL72 domain is a rack that’s scaled up. Across racks is scale-out.

You’re like, I know this. My top-of-rack switch, I got my little thing going. But the AI guys come and say, you know what? Your network’s not enough. You have the front-end network. We want a back-end network that’s much faster than your front-end network that you built for cloud. And so now there’s a scale-out network on the back end that connects across racks.

And then, because it’s never enough, because AI will eat whatever you give it, we know that, I need scale-across. Guess why? Because one data center is not enough. I need more data centers. It’s just never enough. It just keeps going. It’s like Little Shop of Horrors. It just keeps eating whatever you give it, nonstop. It’s not horrors, in this case it’s positive. It’s great. AI will save us all. But regardless, it will eat everything you give it.

And so now you have your scale-across network, which is called DCI, data center interconnect. We used to call it that. But now with DCI and scale-across, I need high bandwidth. I need huge throughput. The couple of pairs of fiber that you used to have between buildings, or even out to the WAN, are not sufficient. I want 1,000 pairs. I want bigger pairs. You’ve got ribbon cables. You’ve got weird configurations. You have this giant trunk between data centers, thousands of pairs of fiber across. Why? Because AI wants it and will eat it.

And so what’s happening from a network engineering standpoint is new things, new protocols to worry about. The good news is it’s mostly Ethernet. So you say, I know Ethernet. Good and bad. We almost call everything Ethernet because when you unpack it, the headers look like Ethernet, so we call it Ethernet. Fundamentally, there are a lot of things that change about it. But let’s just pretend it’s all Ethernet, because we can say, I know Ethernet, and so it’s similar.

NVLink, UALink, these are all acronyms. I won’t go into them. ESUN, you can go look up. I’m happy to talk to you about them as well, actually. It’s kind of boring, but I’m happy to talk to you about that anyway. These are new protocols that are for scale-up. And what do they need? They need low latency. They need high throughput. Lossless.

And then on the scale-out, same thing. Some of you know RoCE v2. It’s pronounced “Rocky.” Don’t ask me why, it’s R-O-C-E. It’s better than… I don’t know. RoCE v2 is a protocol for running basically a memory protocol over Ethernet. It’s RDMA over Converged Ethernet, and RDMA is remote DMA. DMA is direct memory access. It’s acronyms within acronyms within acronyms, like babushka dolls. But anyhow, that’s what it is.

And then RoCE wasn’t good enough, so we got MRC, which OpenAI came up with, with its friends at NVIDIA. Then you’ve got Meta’s RoCE, because Meta wanted its own version. Because why? It’s Meta. So Meta’s got its own version of that, and they’re competing. Then in the future, you have a consortium with 500 pages of specs called UET, Ultra Ethernet. It’s Ultra Ethernet. I’m waiting for Super Ethernet and whatever. But all these things are coming.

If you’re in a data center and you’re a network engineer, these are either coming your way or you’re dealing with them right now. That’s just the reality. And if Ethernet’s not enough, the good thing is InfiniBand has not completely gone away. That weird protocol called InfiniBand, if you’re running an HPC, or high-performance computing, cluster, that’s a sideline right now. So it’s all Ethernet. The good thing is everything you know about Ethernet is mostly reusable, so that’s great.

And then scale-across, I won’t talk too much about that. I think others will talk about scale-across and DCI for the rest of today. But there are important things like buffer management, and the increased fiber density I talked about.

And then I’ll just talk briefly, as a footnote, about the other networks. It’s not really networks, it’s connectivity, which is memory. Because of the memory wall, you’ve probably heard of that, we’re out of memory, by the way, worldwide. Can you imagine? Go find your old PCs and rip those RAM sticks out, because they’re valuable now. We’re out of memory, and so they’re trying to pool memory to make it more usable, and conserve it.

And then there is a new thing called KV cache, as part of the so-called memory management hierarchy. I won’t go into that in detail. You can see plenty of videos out there. But if you want to sound smart at the cocktail party, just use the word KV cache and they’ll be like, “My God, she’s so smart.” So just KV cache. It’s a new word.

And there’s a new thing that NVIDIA made up called scale-in. For many of us, in the old days, it’s like the DMZ. It’s a security function, a security enforcement function, but now at the NIC level. So they’ve made up another word called scale-in. I’m not sure it’s actually going to make it in terms of the naming, but given what’s happening with all the agents going rogue…

By the way, they’re not going rogue. They’re just doing what they were programmed to do. It’s not rogue. It’s a human misalignment problem. It’s not agents going rogue. But because of the enforcement requirements, you’ll see more of that scale-in type of function going forward.

Okay? Alright, let’s keep going. What are the network patterns in the new AI economy? As network engineers, again, you always have to understand the workload you’re dealing with, because fundamentally that’s what you’re doing. And if you look at the workloads for AI, it’s not one workload, by the way.

They say, oh, it’s the AI workload. There is no AI workload. It’s like the SaaS workload. There is no SaaS workload per se. You’ve got to break it down into subdomains, because as network engineers, again, you have to understand what’s actually going on with the network.

If you look at the AI workload, it can be broken down right now, and I say right now, by the way, because it changes, into pretraining, which is where you build these gigantic models and you use all the compute available to you. Within the data center, in a rack, across racks, in a row, in the data center, across multiple data centers. You are building these multi-trillion-parameter models, pretrained. Few people do it, by the way. Some people will probably do small model training, but the vast majority is done by frontier labs and hyperscalers.

Then post-training, which is where you take the model you just built and you improve it a little bit. You fine-tune it. You make it fit for task. That’s less stressful, but still somewhat stressful. And then you have the rest, which is inference. That’s the vast majority.

Inference is what happens when you take a model that’s ready and you start querying it. Why is the sky blue? What’s the answer to the meaning of life? And the answer is always 42, by the way. So that’s what inferencing is. But wait, you can’t stop there, because guess what? Inferencing is not a single workload per se.

Inside inferencing, and we’ve now learned this because of the way transformer models work, there’s a prefill step and a decode step that stress your computing differently. One stresses your compute, the other one stresses the memory, the amount of memory available to you.

So what does that mean? That means that if you decide that you have chips that are better at prefill, like the compute GPUs, and you have chips that are better at decode, guess what? There’s connectivity between them, and you do the transfer. And this is where you bring up the word KV cache. You’re like, yes, KV cache transfer. And so, again, the network plays a role, and you have to understand what you’re transferring.

So inference is not one thing, but above inferencing, you have your agentic layer. This is what everyone’s going to have. You’re going to have your personal agents running around living your life for you. Well, right. And then they’re like, what do I do? I’m not sure. That’s a harder question to answer.

But your Moltbots, your Clawdbots, or whatever new name they come up with, sound friendly. They look friendly. They look really cute. Have you seen the little mascots? Really, really cute. But just as dangerous. Remember Gremlins, the movie? Some of you are old enough. Looks really, really cute. Do not feed your agent after midnight. Guess what happens.

But fundamentally, what agentic flows do is they use the inference engine. They run it, and then they do tool use. It turns out that agents are like us. They use tools. They call into other APIs. And how do they do that? It’s the network again.

So your network patterns are going to change dramatically depending on which part of AI, because when someone comes to you, when the CIO or the CEO comes to network engineering and says, we’ve got to get our network ready for AI, what does that mean? What am I doing? For which part of AI?

But then again, fundamentally, the way you’ll be measured as a network engineer is how you enable that token economy. You’ll learn new terms, so you can spit them back to your CIO. You say to her or him, yeah, I understand the token economy. Tokens per megawatt. And they’re like, oh, yeah, you sound good. Time to first token.

And then the other one is time per output token, TPOT, or “teapot.” A teapot is not what you boil water in and make tea in. TPOT is time per output token. So you learn the words TTFT and TPOT. And again, you sound like a really smart network engineer if you know those terms, because that’s what the AI team is going to talk about, and that’s what your CEO is probably going to talk about as well, because that’s the way we’ve been trained.

Alright? Good. So that’s inside the data center. What happens on the outside? There are some very big-name vendors, who are not here today, hopefully, that will say to you, look, we have extrapolated the network traffic, and right now you have this spiky traffic because humans are using it. With agents, it’s going to be like this. And therefore you have to buy a lot more networking equipment, because it’s going to be like this. It’s not this anymore. You’re going to buy more. We’re not there yet, by the way. So that’s not entirely true.

But what’s true is the agents will be important. The agents will use network bandwidth like no one’s business. That’s the reality. If you look at Cloudflare stats, you’ll see some of that already. The agents are querying unnecessarily, too many times. So caching, not the KV cache, but our usual network caches, are going to be very important. Latency, the ability to push things out to the edge, is going to be really important.

Because when you ask an agent a query, it’s going to go out and pull all the press releases for a company, doing research, again and again and again. And it can do it a lot faster. So it’s going to stress the network. That’s the reality.

But outside the data center, in the access networks, you’ll see things that still remain the same. You’ll still have reliability issues. You’ll still have visibility issues. That doesn’t change in your life, unfortunately. With AI, it’s still the same. Determinism is still important.

Converged operations, the fact that we’re now converging networking with security, is happening across the board. That’s just a reality. And cloud management is not going to go away. In fact, there’s more and more of it. But your job is to build networks that can govern and watch agents, improve observability, and broker the on-ramps. By that, I mean that because compute is going to be spread everywhere, you’re going to have to provide good connectivity, reliable connectivity, across all the areas where your AI may live, where your models may live.

I think that’s going to be the reality going forward, but the underlying MPLS, direct internet, fixed wireless access, your low Earth orbit, that’s not going to change. The agents will use the same networks that we are using right now. That’s not going to change. The observability will have to improve. Security will have to improve. That is the reality.

And then the other part that we’re seeing very early signs of is that the interconnect is going to be very important. Why do I say that? Again, we have early signs. It’s not dominant yet, but I think the success of companies like Megaport shows that there is a desire to have more AI models available for inferencing, distributed, because that’s where the power is. That’s where there’s what they call land, power and shell available to you to build these compute clusters. And so the sites will go where the power is.

We have a constraint on power, especially in this country. And so inference will probably migrate to the metro areas for latency purposes. Some of the fundamental technologies will improve the ability to reach those places. Coherent optics, I’m not going to go into detail on that one, for sure. And scale-across will drive increased diversity of the interconnects. That’s what we will see in the near future.

Alright, common network requirements. There are things that will never change from a network engineer’s perspective. They’re always going to want more speed. Speed will keep doubling in the data center. On the campus side and the WAN side, you’re seeing it plateau.

The reality is Wi-Fi speeds or link speeds are in the 1 gig, 2 gig range. In campus, we took that 1 gig jump. The 1 gig to 10 gig jump is taking a mighty long time, because we don’t really need it. And so some of those places are slowing down in terms of that. Speed is happening, but at some point it will slow down, and then over time you switch to robustness.

Your WAN wants to be more robust. You want multiple links. Why? Because that’s critical to your business. And so increasingly we’ll see diversity. We’ll see observability, the requirement of seeing, where’s my traffic going? How’s it doing? How do I want to optimize the traffic? What’s my agentic stuff doing?

And then security is going to become very critical, especially on your access side. I’m seeing that. And they’re going to use words like PQC, post-quantum cryptography. Is your WAN PQC-ready? Is your inter-data center connectivity PQC-ready? Because God forbid someone has a quantum computer and is cracking all our public key protocols. So that’s going to be important.

And then, fundamentally, manageability. The more campus sites, the more remote sites, the more manageability you need. Those things will not change. Those five things will remain with you to the end of time as a network engineer. You’ll always be asked for one or two of those axes, and it’ll jump. The monkey will jump from one to the other to the other to the other. That’s just the way it is.

Okay. So I want to wrap up here with a couple of things. AI is supposedly taking our jobs, and they keep telling you, no, AI will not take your job. Someone who’s using AI will take your job. I’m like, okay. Yeah, sure. That’s super cliche. It may be true, but it’s super cliche.

A couple of years ago, I was telling network engineers, with software-defined networking, you’ve got to learn some programming. Remember, people told you to learn programming. As network engineers, you’ve got to learn programming because of software-defined networking and APIs and all that. I’m going to say something a little different shortly.

But the reality, thank God for network engineers, unlike software engineers, is we have a couple more years left beyond them. Software engineers are worried about their jobs. Network engineers will only worry a couple of years later. Why? Because you don’t have robots that can plug the SFP in there yet. Network engineering AI is still not as good as software engineering AI, for various reasons.

So if you’re a network engineer, know that you have more in terms of your career span, a couple more years on those software guys. So thank God you didn’t become a software engineer.

The reason I say that is the vast majority of autonomy out there, like we talk about the autonomous network, the self-driving network, is not a reality today. It’s not. And I’ve been all over the world. You would think the telcos have figured it out. They have the largest networks. They have the most to gain by that. The tier ones all over the world, with the exception of China Mobile, which is always special. Okay. Most of them have not achieved any reasonable autonomy.

If you look at the TM Forum, you will never have heard of TM Forum if you’re an enterprise guy. It’s a telco organization that borrowed the SAE autonomous driving framework and put it over on the network side, because why not copy a good thing? And so L0 is manual, L5 is fully autonomous. You can imagine what the rest are.

L2 is basically semi-autonomous, meaning it’s governed. The human is watching it. It’s assistive at this point. And most of the networks today are L2. BGP, because of its importance, guess what? Humans are still very, very involved in it. That’s just reality. Otherwise, you take down your networks, and at the backbone level, that’s not good.

Most people are still at assistive only. Very few people are at L3. Very, very few, with very few exceptions in specific domains, are at L4.

And if you look at the stats today, this is Gartner saying 80% of agent actions still require human approval, and this is across the board, all IT. And they think it’ll be 20% by 2029. Guess what? I think the network percentage will still be higher. The humans will still be involved a lot more. Reality.

And the Network Automation Forum, if you don’t know about them, you should know about them. Network Automation Forum, AutoCon. That’s a good one. I see nodding. Yes. It’s a good forum to talk about network automation, for enterprise customers, enterprise users. Last year, they ran a survey. Only 3% have operational LLMs in the network. No surprise. It’s good. For 97% of them, we have a job. So that’s good.

But having said that, one thing I would say is that as a network engineer, maybe you don’t have to learn programming now, but you will have to understand programming, because the AI will be used for automation. The AI will increasingly be used to generate automation. You have to know how to verify. You don’t need to know how to program. You need to understand it, read it and verify it, because that’s going to be your job.

And that’s going to be important in terms of understanding what the blast radius of the change is going to be. It’s going to be very important to govern and control agents in your network, and it’s going to be important to build trust in terms of automation. So again, I’m not saying you should learn programmability, but you need to understand it. You need to understand what the agent does. You’ve got to think deeper.

Alright. So this is an eye chart, but it’s the role of the network engineer depending on the different areas, the different domains, across the different aspects. Again, those five things will keep coming back. I’m not going to go into great detail on any of those. You’re welcome to read them. The slides will be available.

There’s nothing really special or brilliant about it. It’s just a capture of observations in there. And none of this will surprise you. It’s basically what’s happening in the networks right now across different domains, and as you will see, the focus is different.

But the job of the network engineer is to govern agents, just like software. Govern agents. At some point, the agents will end up doing some of the work, but the human will be important. And so it’s going to go from operation today to being able to design intent. And to design intent, you need to understand networking deeply, because you’re being asked to do design. They’re not very good at design right now, by the way.

And then eventually you’ll be governing these agents that will work on your behalf, but you have to verify it. But the good news is that I think our jobs are still safe, at least for the next 2 to 3 years. And I would say that network engineering is still going to be important. You still have to learn the basics. Otherwise, how do you know if the AI is doing the right thing or not?

And so with that, I’d like to thank you. Again, slides are available, and I’m available to chat after. Happy to go into any of those acronyms, even the ones I don’t remember the meaning of, but happy to talk about them regardless. So thank you, great audience. Appreciate your time.