The Five Nine: Are enterprise networks ready for AI?

Is your enterprise network truly ready for AI — or are GPUs getting all the attention? In this episode we break down why AI readiness depends on more than compute power, and why networking, private connectivity, latency, security, automation and network as a service are becoming critical to enterprise AI success.

Guests Brandon Ross of DE-CIX and Bill Long of Zayo explain how agentic AI, real-time inference, AI-powered customer care, analytics and large-scale data movement are changing what enterprises need from their networks. 

Catch the video at top, listen to the audio edition and read our transcript below, or watch this and future episodes on YouTube. 

To learn more about the topics in this episode, check out: 

This podcast is written and hosted by Diana Goovaerts. It is edited by Diana Goovaerts and Matt Rickman. Liz Coyne is our executive producer. Special thanks to guests Brandon Ross and Bill Long.

Diana Goovaerts: Everybody wants to be AI ready, but if your answer starts and ends with GPUs, you may be missing a key element that determines whether AI actually works for employees, customers, and machines in the real world: the network. That's because enterprise AI is about to stress networks in ways many companies haven't fully planned for. 

Brandon Ross, DE-CIX: For many years we looked at the network as just basic plumbing, right? We didn't pay a lot of attention to it in the enterprise side of the world. But the more AI we deploy, especially agentic AI, makes the network really more of a strategic asset. 

Bill Long, Zayo: You gotta think about what is the architecture for AI. The traffic volumes are much bigger, the security requirements are much higher, and the performance needs to be much bigger, that's kind of what folks are solving for. 

Diana Goovaerts: I'm Diana Goovaerts, and this is The Five Nine.

AI readiness is becoming a networking problem, and the winners may be the companies that can move data predictably, privately and fast enough to keep up with AI. For months, enterprises have been talking about whether they have enough compute for AI, enough GPUs, enough storage, enough access to models.

But two recent conversations with networking executives made one thing clear: enterprises may be underestimating the network layer. Brandon Ross of DE-CIX and Zayo's Bill Long both argue that as AI moves from experimentation to production, and especially into agentic AI, real-time inference, and large-scale analytics, AI-ready has to mean more than having compute capacity.

It has to mean having the right connectivity, visibility, automation and control.

Let's start with the basics. When companies say they're AI ready, they often point to compute, GPUs, servers, storage, model access. That makes sense. Without compute, AI doesn't happen. But Brandon Ross says that that picture is a little bit incomplete 

Brandon Ross: What enterprises tend to measure is things like GPU and compute power, maybe some storage. But that really doesn't tell the whole story. It's really more about the network, and it's not just about having a quality amount of bandwidth, but it's about moving data in a predictable and a secure and an efficient way. 

Diana Goovaerts: Ross's point is that AI performance depends on moving data in a way that is predictable, secure and efficient. If the network can't reliably connect users, applications, clouds, AI providers and data sources, those shiny GPUs that you bought may not translate into business value.

And Bill Long framed it a little bit differently. He says that enterprise AI is following the same arc as cloud adoption, but the stakes are higher.

Bill Long: I think of it as kind of cloud on steroids, where in the old world where you had all your IT stack in the basement of your headquarters, that was-- it was a pretty simple world, and you could put a perimeter around that, keep things safe.

But then really with cloud, it became a distributed hybrid multi-cloud kind of architecture where you had some infrastructure up in cloud, you still had some in your basement, you had some in third-party data centers, and then providing the networking where you could connect all of those things together and start to manage it, that was pretty complicated.

That was enterprises adopting cloud, and AI just put that on steroids. 

Diana Goovaerts: In other words, the problems enterprises already had with hybrid cloud, things like distributed environments, security boundaries, data movement and application performance, are all amplified by AI. One of the biggest shifts both executives flagged is the move from relatively forgiving AI interactions, like browser-based chatbots, to AI systems that act in real time and trigger workflows or coordinate with other systems

Brandon Ross: When you start moving to agentic AI latency and reliability becomes so much more important. Really what happens is that the network stops being just transport. It starts really become part of the application, so all of those metrics that we use to measure networks become even more important when we talk about AI, or especially when we're talking about agentic AI. Any loss of data on the network, while again, in a traditional sense just means retransmitting it, maybe taking a little bit of latency. Those kinds of network impairments really exponentiate on each other and make the network much worse, which then make the application response much worse. 

Diana Goovaerts: That line, "The network becomes part of the application," is the key. In a chatbot, a small delay may be annoying, but in a robotics system, a manufacturing line, or an autonomous workflow, a delay can directly affect the customer experience, trust, and business outcomes. Long agrees that performance matters, but he did offer a contrarian take.

Bill Long: I think latency matters a lot less than people think. You know, it takes roughly 100 milliseconds to blink your eyes. So, when people are like, "Oh, you know, you need to have, you know, one to five milliseconds of latency for things to work," eh, I don't really buy it.

I think that there's a lot you could do within a 20 to 30-millisecond sort of window and things will work just fine, with the possible exception of, like, robotics.

Diana Goovaerts: His view is that for many inference use cases, compute in major metro areas may be close enough. The far edge may be reserved for high latency sensitive environments like factory floors or robotics, where the economics of small GPU pods just become hard to justify. So which AI workloads reveal network problems fastest?

Well, Ross points to anything that is real-time, so think voice AI, call centers, robotics, industrial automation and autonomous systems

Brandon Ross: If you call into a call center and you're not getting natural voice responses because there's a long delay between when you finish saying something and when the other side responds, it's much less natural. And as human beings, we start to lose confidence that what we're dealing with is actually working right.

And so those kind of applications really become very important. And of course, you can imagine things like robotics, industrial applications and autonomous systems are gonna be even more impacted in those cases. So, it really all comes down to the network has to perform in predictable and reliable ways.

Diana Goovaerts: Long's answer, though, is more focused on analytics and enterprise data movement. In his view, one of the hardest problems is connecting large pools of proprietary enterprise data to a growing ecosystem of specialized AI platforms 

Bill Long: You're gonna want to take your data to where the compute is or make sure your different sources of compute can operate on your data.

And I think the enterprise network is not very well set up for that. You need really big pipes. Those pipes need to be private, and those pipes need to connect to a large ecosystem. 

Diana Goovaerts: And that ecosystem point also matters. Enterprises may no longer just be connecting with two or three cloud providers. They may need access to 10, 15 or 25 AI providers, each with different services, models and connectivity requirements. Both Ross and Long also took aim at the limits of relying on commodity internet or best effort connectivity for business critical AI traffic 

Brandon Ross: On the open internet, you have no control over how your traffic is routed or routing changes. And when I say no control over how it's routed, that could mean both topologically and geographically. And so, there's geographic risks just as much as there are topological risks.

So if your traffic is getting rerouted from a path that worked well to a path that doesn't, that's something that's gonna affect your performance. 

Diana Goovaerts: For Ross, the risk is predictability and visibility. If AI traffic is rerouted in ways that the enterprise can't see or can't control, performance can suffer, and so can data sovereignty commitments

This is where private connectivity starts to look less like optimization and more like AI infrastructure. It helps enterprises reduce exposure to the public internet, control traffic paths and support higher throughput data movement 

Bill Long: Where you run into problems both from a security standpoint, bandwidth, latency, cost, especially if you're data center to data center connectivity or data center connecting to clouds or to AI providers, you need the scalability and security and performance of larger pipes.

And if you do that over best effort internet, you're gonna be paying way too much. The performance is gonna be bad. The security risk you're taking on is not optimal. 

Diana Goovaerts: The next question is management. So even if enterprises know they need more private connectivity, how do they actually operate that environment when the number of clouds, AI vendors, data centers and policies keeps multiplying? 

Bill Long: Now that you have not just two, three, four clouds, but you know, 10, 15 25 AI providers associated with that, more infrastructure across third party colo providers, with an estate that's that complex, you cannot manually manage that anymore. 

Diana Goovaerts: Long sees network-as-a-service as a usability and automation layer, a way to configure, provision, monitor and change network services in real time rather than through slower manual processes. And Ross also puts automation at the top of his list, too. If AI demand is unpredictable, if new workloads and providers can appear quickly, enterprises need a network approach that can flex with the business rather than lag behind it.

Brandon Ross: The AI marketplace is developing extremely rapidly, as we know. We don't really know what our loads are gonna be like. We don't really know what our users are gonna demand of us in advance. And so being able to automate the provisioning of additional interconnection, I think is critically important. 

Diana Goovaerts: So what is the biggest blind spot? Well, for Ross, it's pretty straightforward. Enterprises are still treating the network as an afterthought.

Brandon Ross: Enterprise has really focused so much on GPUs and servers and storage and the network, you know, so far has been kind of an afterthought. You know, for many years we looked at the network as just basic plumbing, right? We didn't pay a lot of attention to it in the enterprise side of the world. But the more AI we deploy, especially agentic AI makes the network really more of a strategic asset.

Diana Goovaerts: For Long, the mistake is planning only for what enterprises can see today and not the unknown shape of the AI market of tomorrow.

Bill Long: The biggest error that I see people making is not preparing for the unknown. I think, you know, the only rule that we have right now, the level of of change that's coming, now the only dependable thing is change. So I think folks are not planning enough for how different they're gonna need to plan for.

So that, that manifests in, in terms of how do I select to put my infrastructure in locations where I can connect to a lot of different ecosystem providers, flex my bandwidth up and down, add security where I need to add security. So they're not planning enough for the option value of what the future world might be like, um, instead of just what they have line of sight on today.

Diana Goovaerts: Put those two together and the message is clear: enterprise networks need to become more strategic, more programmable and more adaptable as AI moves deeper into production.

That is all the time we have for today. Make sure you like and subscribe wherever you are listening or watching, and stay tuned to fierce-network.com during the week to get the latest news in between episodes.

And then we will see you next time.