From SD-WAN to AI Traffic: How Enterprise Networks Are Evolving

[00:00:00] A big thank you to Denodo for helping me make more than 60 monthly interviews possible across the Tech Talks network. And as businesses move from Gen.AI to Agentic.AI, Trusted Data Becomes Everything. Everything from Gen.AI to Agentic.AI, Denodo is helping organisations build intelligent, secure and scalable AI solutions with data access, governance and explainable results.

[00:00:28] So, build AI that you can trust and do it with Denodo. And you can learn more by simply visiting Denodo.com. What happens when AI ambition runs straight into the limits of the network? Well today I'm joined by Jamie Pugh, CTO of Global Gig. We're going to enjoy a conversation that looks at the part of the AI story that many businesses still overlook.

[00:00:54] Yeah, this year I've spoken lots about agents, models, data, governance and ROI. But Jamie is going to shine a light on the connectivity layer that makes all of this possible. Now, Global Gig helps multinational enterprises build composite global networks across carriers, countries, clouds and security environments.

[00:01:18] But he will describe this today as creating a single programmable one rather than leaving organisations with a patchwork of regional contracts, providers and blind spots. Because we're going to talk about how AI workloads are changing network traffic patterns. Because for years, enterprise connectivity was largely built around people accessing SaaS tools and cloud applications.

[00:01:44] But as agents begin talking to tools and models and data lakes and inference engines, not to mention more agents, traffic is becoming more distributed, bursty and system to system. So he will explain today why cloud on ramps, single carrier strategies, asymmetric broadband links, poorly coordinated network designs, how all these things can quickly cause hidden bottlenecks.

[00:02:12] And he'll share why resilience now requires diversity, better observability, network as a service fabrics, SD-WAN and security policies that can adapt as traffic moves across regions and clouds. And we'll also discuss why milliseconds now matter to productivity, AI performance and business continuity.

[00:02:32] Because if people increasingly rely on AI tools to research, analyse, automate and support decision making, even the smallest connectivity problem can have a much, much bigger impact than they did in the old world of simple web access. So as enterprises race to scale AI, are they paying enough attention to the networks carrying the workload? I'd love to hear your thoughts on this one. But enough for me.

[00:03:00] Let me officially introduce you to the guests right now. So a massive warm welcome to the show. Thank you for joining me today. Can you tell everyone listening a little about who you are and what you do? My name is Jamie Pugh. I'm the CTO for Global Gig. You can think of Global Gig as building composite global networks for multinational enterprises.

[00:03:23] We assemble underlays from around 200 or so global ISPs across 180 countries. We layer on cloud and DC on ramps through network as a service fabrics and wrap it with SD-WAN, SASE providing, you know, and SASE providing secure access and observability. So the result there is a single programmable WAN instead of a patchwork of regional contracts.

[00:03:52] And my role spans the technology platform, the SE organization, sales engineering organization, the AI direction and automation. Well, it's a pleasure to have you join me today. And when we're talking around things like network traffic, it is one of those things that I think we all take for granted when we're downloading, I don't know, a 4K video off Netflix or a 100 gig game download on Xbox. You look inside the average house, there's so much bandwidth being consumed.

[00:04:22] And then now we're throwing in AI workloads into the mix. So I'm curious though, more recently, where are you seeing global networks struggling under modern AI workloads? Is there a big difference here? What are you seeing from your side? Yeah, I am. There's definitely a change. You know, the traffic used to be really kind of designed. The architecture was designed around users consuming.

[00:04:45] In the last several years, we've seen that's been instead of, you know, the data center hub spoke over MPLS model, it's been SD-WAN, you know, and out to SaaS applications. And that's an interesting, you know, download pattern where it's very download centric. But now traffic is actually being, you know, sent between systems. So it's systems talking to systems. And when I talk about that, it's more like, you know, think about agents.

[00:05:13] So AI workloads are creating massive traffic now between the model to the data store, the agent to the tools, inference engines to data lakes. And that often spans across regions and clouds. Most architectures that we've built in the last 10 years weren't designed for that pattern. And the cloud on ramp is now what we see as the bottleneck.

[00:05:42] Let me expand on that a little bit. So the SD-WAN economics, you know, we built a lot of networks, you know, saying, hey, we're going to replace these expensive MPLS and give you some diversity with some internet circuits. So typically that was a DIA at a location that had symmetric bandwidths. And then we built, you know, some, we added in some broadband connectivity. And I think broadband was perfect for SaaS and it was a low cost option to put some resiliency in there.

[00:06:09] But what we're seeing now is that, you know, these agent to agent or agent to MPLS, because the, right now on my workstation, for instance, I've got several agents that are doing research for me. Right. And when I go to research something like something that I want to purchase personally, it's now going out there and it's looking at about a thousand websites. I'd only look at five or six. Right. But it's like looking at a thousand. So that's an interesting, you know, bandwidth characteristics that's changing.

[00:06:39] And as I start to do that kind of research and infer back to, you know, models that might be hosted in the cloud or hosted in my data center infrastructure, there's a lot of upload traffic going as well. And those broadband circuits, I think what we're seeing is they're not suited for that type of traffic at all. They're asymmetric and their upload bandwidth characteristics are our best effort.

[00:07:04] So, you know, we're seeing where, you know, the SD-WAN is struggling with those circuits. So, you know, we're typically having to re-engineer them to think of like download only links because they're causing some kind of degradation.

[00:07:19] Listening to your one example there, your personal use case of how your bandwidth has changed is a great example, especially when you think about scaling that across an entire workforce and everyone working the same, an entire state, even or even country and continents around the world.

[00:07:38] So if we look under the hood at the IT infrastructure that makes this possible, that start in the field, the strain, when enterprises begin distributing AI workloads across regions, clouds and edge locations, what is it that breaks first in that network layer? And why are so many teams only just discovering this after deployment? Yeah, again, I think that goes back to that cloud on-ramp. You know, that's where the apps are. That's where the inference is likely going to be.

[00:08:05] Or also the data center on-ramp, you know, wherever the inference is going to be or the models are hosted, that's where it's going to be. So most enterprises still have that single on-ramp regionally. And sometimes the single on-ramp to the cloud, period, and that's it.

[00:08:24] I think that worked great for human-paced application traffic, but AI workloads will create a, you know, sustained bursty traffic pattern that will overwhelm that single on-ramp.

[00:08:37] So the way we design and the way I think the fix is, is to build multi-cloud network as a service fabrics with this elastic capacity component on both ends that allows teams to, you know, it allows the network to be more dynamic as far as, you know, hey, I'm starting to see traffic patterns. I need to adjust and accommodate.

[00:09:01] You can do that easily across a NAS fabric where you can't do it with these static, you know, circuits that you have today. And before talking to you today, I was reading that you've warned previously about the risks of relying on single carrier or fragmented network models. So I also suspect that many people listening, we're waking them up at the moment to some of the risks around those bottlenecks.

[00:09:25] So can you walk me through how a connectivity failure in one region could cascade into wider disruption across AI systems just to bring to life the problem we're talking about here? Yeah, I think, you know, there's a couple of things that often get conflated here. You know, fragmented is an interesting word because if it's poorly coordinated, that's bad.

[00:09:48] But carrier diversity, you know, is also when you have, you know, you need fragmented networks in order to actually build that carrier diversity, right? But what you need is to strategically onboard them and orchestrate them through a single control plane or have a provider that can do that for you. And that's resilience, not fragmentation.

[00:10:07] So, you know, what we've seen is in a lot of instances where we take over somebody's network, they've bought, you know, what they thought were, you know, fragmented or diverse solutions. But in the end, there was some sort of, you know, component that was shared across them. So that could have been a, you know, a last mile that was actually the same carrier.

[00:10:32] They just didn't know it because they bought from two different ISPs or, you know, they share the same path or the same peering relationships upstream. And, you know, when a fiber cut happens somewhere in the network, all of a sudden they've got an additional 100 to 200 milliseconds of latency to get somewhere. And, you know, that characteristic shared across either surviving provider.

[00:10:56] So, you know, that's where, you know, you really have to get in and understand, you know, that these carriers nationwide, understand their peering relationships, understand, you know, their fiber paths so that you can design something that's more resilient. And latency and reliability are also increasingly becoming tightly linked to business outcomes.

[00:11:17] So on that side of things, how should infrastructure teams be maybe rethinking network architecture when milliseconds could directly impact not only AI performance, but the user experience as well? Should they be doing things differently or what should they be doing differently? Yeah, I think that, you know, there's some architectural shifts there, designing for predictable paths, not just available ones. You know, that's what we were kind of talking about a little bit before.

[00:11:46] So because the AI workloads are path aware, you know, the routing needs to account for not just the current performance, but also the policies that you're going to be putting in place, you know, to support those AI workloads, right? I mean, to me, I mean, we've always done this, right? We've designed kind of these application performance profiles.

[00:12:15] But the problem now is that AI is bringing a whole new performance characteristic and the inference is all over the map. So I don't think that all the tools that are out there today have all the profiles in order to be able to do that. So if I think about this personally, let me just get to a personal example. I am a big user of AI tools, right?

[00:12:43] And, you know, I've got subscriptions to Perplexity, to ChatGPT, you know, to Claude. And I put a lot of my research and, you know, I guess a lot of my Excel work now. So a lot of my busy work, it actually goes into tools like Claude.

[00:13:04] And it's made me so much more efficient to the point where when there's a network issue that's keeping me from, you know, using those tools, I feel like I've lost my best employee. And my productivity is just like shot. Like it's incredible now. I get frustrated so quickly because of this. And typically in the past, I've been able to kind of say, you know what? The Internet's down.

[00:13:34] I'm going to go find some busy work, you know, so I'll go grab an Excel workbook and I'll work on some formulas and I'll go drive in it or I'll take some downloaded content that I downloaded and I'll start reading it from a research perspective. But now that is so inefficient because I know as soon as the network comes back up that the Excel work will be faster. So it's just wasted time for me to go in there and actually call that busy work.

[00:14:00] And then the downloaded material that I might have read is probably already aged out because, you know, everything is happening so fast. So my mind wants real-time data.

[00:14:11] So, you know, I feel like this is going to be a new frustration point across the enterprise as, you know, as people have poorly architected networks and they're sitting in an office or maybe they're on ramps through their, you know, their endpoint agents that are taking them into the VPN infrastructure, you know, and they're having issues with that.

[00:14:33] I think the frustration levels are going to be much higher now going forward than they have been in the past just because people get so reliant on having such a good set of tools that can help them accomplish goals. And they're used to going faster. And when they have to slow down, it's going to be quite disruptive. Yeah, 100% with you there. And we are seeing a shift from reactive network management to predictive and self-healing systems.

[00:15:00] And saying that out loud reminds me of the Christine movie, the Stephen King film all those years ago where the car heals itself and he just says, show me. It would be great if we could get to that point on networks. But what does it all look like in practice? What technologies are making this possible today? Yeah. You know, I mean, SD-WAN kind of, you know, I mean, let's just look at the buzzword, right, of, you know, the self-healing. I've been hearing about self-healing for at least 10 years, right? Yeah.

[00:15:29] So, and I've even implemented a lot of this early on with the DMVPN days, you know, and then the early SD-WAN. So, you know, you've got kind of this telemetry-driven, you know, routing decisions that are being made. That's great in the SD-WAN platform, but that's got to extend now into the data center at data center. You know, and if you look at data center at data center, then you've got cloud to cloud. So, this is where there's network, you bring in that network as a service kind of backbone between the two.

[00:15:59] And it also has to have that same type of intelligence for self-healing and predictive, you know, route changes, et cetera, within the different fabrics that you've got there. You know, and if you look just recently, you know, you can see that the big ISPs are, see that that is the fundamental change, right? It's because systems are going to be talking to each other much more frequently. You're going to get a lot more cloud to cloud. You're going to get a lot more, you know, site to site.

[00:16:26] East-West, you know, East-West traffic patterns instead of just North-South, which is what we've been designing for for the last 10 years. We just saw a recent acquisition, you know, from Lumen to Alcura, for instance, the Lumen-Alcura. It's really a signal to the direction the industry is looking. Where's that single control plane that can manage traffic flows, not just at the SD-WAN, so the North-South layer, but how do we actually look at East-West as well?

[00:16:56] And for people listening in organizations that are running hybrid or multi-cloud environments, how do you design connectivity that is both resilient and flexible, but also without adding unnecessary complexity or cost? And I appreciate that question is almost a podcast episode on its own, but anything you can share around that or kind of advice that you would offer for people in that situation listening?

[00:17:18] Yeah, you know, I think it's, you have that design principle that, you know, it's integration and management competence across many underlays. So you shouldn't be choosing between resiliency and simplicity here. Yeah. I think, you know, your concrete building blocks that I'm seeing and what we're putting into ours are those kind of network-as-a-service fabrics. I mentioned them before. You know, multi-carry underlays orchestrated.

[00:17:45] That's where you're orchestrating the SD-WAN and the SASE or the on-ramp to those network-as-a-service fabrics. And then you have your consistent security policies. So you have to start thinking of not just infrastructure as code, but policy as code across these, you know, so that you're designing not connectivity, but you're also, when the network changes, it's easy to make sure that your policies are adopting, you know, the changes well.

[00:18:14] So that's really kind of it. It's really, to me, it's those three building blocks there. The network-as-a-service multi-carry underlays, bringing in that centralized single control plane. And that right now, you know, that's kind of more of a goal than a product because there isn't really something that's taking your SD-WAN SASE and your multi-cloud routing overlay, and then your network-as-a-service fabric and looking at it holistically.

[00:18:43] I mean, I think MSPs can come in and actually perform that for you as kind of, hey, we know this vendor. We know the network-as-a-service layer. We know the cloud routing. We can build an architecture and we can observe it for you and we can better manage that. But it's still more of an architectural goal than a product today. So that's kind of how we build today, right?

[00:19:05] We build them that on-ramp SD-WAN, SASE, and then we build a network-as-a-service that's connecting multi-cloud, all of their cloud providers, resiliently into this resilient fabric. So it's typically two multiple network-as-a-service providers. So you get that resilience. So you get that cloud-to-cloud and that data center on-ramp or that data center to data center. And then you've got the cloud workflows and the east-west in the cloud.

[00:19:32] So if it's Azure, it's the cloud provider and you've got multi-regions of Azure, it's best to keep that inside the cloud, the CSP's fabric. It's going to be lower cost and it's going to be lower latency than to take it out. But when you go cloud-to-cloud, so like Azure to AWS, for instance, you want to take that through your network-as-a-service layer.

[00:19:51] But being able to put all that together and observe it all and then understand where things are potentially at capacity strengths or where you need to augment and dynamically improve, that's where an MSP brings a lot of value. And that's what Global Gig does. And I think without looking into a crystal ball, I think the one thing that is incredibly easy to predict as we look ahead is the amount of bandwidth that we all use, whether we're in the workplace or at home, is going to scale dramatically.

[00:20:21] And you mentioned that great example at the beginning of you using agents and how many websites those agents access. But what do you see a future-ready global connectivity strategy looking like for enterprises that want to scale AI safely? And what would you say are the maybe first technical steps teams should be taking right now to begin moving in that direction? Yeah. Again, I think it's that programmable network, right? That one control plane. That's the future. Everybody's got to get there.

[00:20:49] There's going to have to be products that actually are evolved there to get us there sooner. The underlay diversity thing we talked about, you know, verified physically, though. Not just by, you know, hey, he said, she said, you know, that this is two diverse carriers. It's actually getting the KMZ files and looking at the paths and then looking at the peering relationships and making a decision there.

[00:21:17] Inference, you know, where's the organization placing it? I think jurisdiction, latency, data governance, all those things are going to play. So it's not just going to be, you know, driven where compute is cheap, right? It's going to be all these things are going to play. So you have to have a network that's adapted to that. And then again, the end-to-end visibility. So the observability piece of it all.

[00:21:41] And then leveraging AI, you know, to assist humans in the beginning to run it and then ultimately automating a lot of those functions. You mentioned technical steps as well. So I think the place to start there would be, you know, with an audit. You know, I mean, I think it's kind of, are the tools that you've got in place?

[00:22:10] So that's the SDGLAN fabric. You know, do you have multi-cloud infrastructure? Are they capable of getting you to the next steps? You know, tech your audit, your circuits as well. You know, how many of them are broadband circuits? How many of them need to be, you know, that diverse circuit that you're buying need to be upgraded to a symmetrical type service? You know, whether that's a, you know, FTTX or DIA Lite or even a full DIA based on what sites you believe are going to be there.

[00:22:38] So I think you need to, you know, get that audit done on the underlay as well. And then really kind of just working across your teams. So that's, you know, infrastructure, whoever owns your AI strategy and inference, understanding where they're going so that you can make sure that you're ahead of that on the network side.

[00:23:00] I think a lot of these teams still work in silos and where a lot of them may be taking shortcuts because there's a lot of risk kind of that I believe organizations are taking right now because their CEOs are wanting to be first. They're wanting to be fast. They're telling people to embrace AI. There is a lot of that inference is going out to the big players right now.

[00:23:26] But as, you know, people are starting to redline these MSAs with their customers that forbid them from taking, you know, PII and customer identifiable information to these public LLMs, these public models, you know, people are going to start bringing that stuff inside. So that's going to become part of their infrastructure. And I think the network team and, you know, security team need to get engaged now, you know, to understand what that looks like so they can start building.

[00:23:57] That's what we do. We bring those teams together. We have these conversations. We let them know what the industry is doing, what our customers, you know, our other customers are doing. And then we come up with a plan and help them get there. Well, I've had so many conversations around AI this year and everything from agents to the importance of data and the models, etc. But I love how you've been able to shine a light on the importance of the networks that we all take for granted today.

[00:24:26] And for anybody listening that would like to continue that conversation with you, maybe you've set off a few light bulb moments. Maybe somebody's listening thinking, at last, somebody's talking about this. Where would you like me to point them so they can find out more information about Global Gig and maybe connect with you too? Yeah, for sure. I'm not big on social media outside of the industry standard, LinkedIn, you know, so you can do Global Gig, Jamie Pugh, and you'll probably find me. I should be the first result there.

[00:24:56] And alternately, our website is a great source of information too. So that's www.globalgig.com. G-L-O-B-A-L-G-I-G.com. Awesome. Well, I'll have links to everything that you mentioned there. I, for one, love how you're helping transform networks into a strategic advantage via managed network and security solutions. I'd urge anyone listening that's enjoyed today's conversation, please go check them out, connect with you.

[00:25:24] And I'd love to hear back from you on anything that you thought. But more than anything, thank you for starting this conversation today. Really enjoyed talking with you. Yeah, thanks again for having me, Neil. Appreciate it. One of the things I loved about this conversation with Jamie today was how it shifted the AI discussion away from the obvious talking points and into the infrastructure that many people only notice when it fails. Yeah, we often talk about AI in terms of models, data, agents, applications.

[00:25:52] But I think Jamie reminded us all today that all these things depend on connectivity. Connectivity that can handle new traffic patterns, new latency demands, and new levels of business reliance. And his point about AI changing traffic from people to applications into systems to systems must have set off a few light bulb moments.

[00:26:14] Because agents querying tools, models connecting to data stores, inference engines reaching across clouds, and employees relying on AI assistance. All these things place more pressure on networks. Networks that were often designed for a very, very different era. And the lesson for me is that resilience cannot be treated as a backup plan. They need to understand their underlays, validate carrier diversity, review cloud on ramps, assess broadband usage,

[00:26:45] and bring network security infrastructure and AI teams all into the same conversation before AI workloads move deeper into production. So I'll add links to Global Gig and Jamie's LinkedIn profile in the show notes. But I'd love to hear what you think. Have you or your business spent too much time talking about AI models while underestimating the network needed to support them? TechTalksNetwork.com.

[00:27:12] That's where you can find me, talk to me, work with me, browse through 4,000 different interviews. A quick thank you to NordLayer for supporting the podcast and helping me make these daily conversations possible. And if you are listening and you're responsible for security or IT, you will know the reality. The reality of most of your risk now sits inside SaaS apps and browser activity.

[00:27:37] That gap is exactly what NordLayer is addressing with its new business browser. So instead of bolting security on from the outside, it builds it directly into the browser itself. This means you can control access, monitor activity, enforce policies, and reduce shadow IT all from one single place. And most importantly, it does it without adding deployment headaches or complex onboarding.

[00:28:05] You get things like browser-based data loss prevention, SaaS access control, and zero trust browsing, but delivered in a way that your team can actually use. So if you've been trying to simplify your stack while improving visibility, please check it out at nordlayer.com slash browser. Love to hear your thoughts on this one, but that's it for today. I'll be back again real soon with another guest. But thank you for listening as always, and I'll speak with you then. Bye for now.

[00:28:35] Bye for now. Bye. Bye.