To embed our video on your website copy and paste the code below:
<iframe src="https://www.youtube.com/embed/MxQEpCfWqzE?modestbranding=1&rel=0" width="970" height="546" frameborder="0" scrolling="auto" allowfullscreen></iframe>
So our next session is building the AI backbone. Now AI infrastructure is fragmenting across distributed data centers, GPU clusters, edge and enterprise deployments. AI training requires high sustained bandwidth and near zero packet loss while edge inference requires low latency connectivity. Hyperscalers are building private fiber networks to bypass telcos for interdata center connectivity. So how can telcos transform their infrastructure into an AI ready fabric? So let's meet our panelists. I'm going to ask them to introduce themselves starting at the far end. Neil.
Neil McRae, CityFibre (01:10):
Hey, I'm New McCreary. I work at CityFibre.
Ray Le Maistre, TelecomTV (01:14):
Okay. Gabriela.
Gabriela Styf Sjöman, BT Group (01:16):
Gabrielasio Gueman, Managing Director of Research and Commercialization at BT.
Ignacio González, Red Hat (01:21):
Ikinesio Gonzalez, part of the Red Hat team looking after a mobile network business across EMEA Rio.
Colin Evans, HPE Networking (01:27):
I'm Colin Evans, HPE. Introduced myself this morning.
Andy Linham, Vodafone Group (01:31):
I'm Angelina. I'm part of the strategy team at Oat of Home Business.
Ray Le Maistre, TelecomTV (01:34):
Okay. Welcome everybody. Andy is our DSP co-host for this session. So before we begin the discussion, I'm going to ask him to step up to the podium for a quick DSP leaders address. So let's give him a round of applause.
Andy Linham, Vodafone Group (01:50):
So in honor of Neil's pinball tables, I'm going to tell you a slightly embarrassing story that I probably shouldn't, but we'll give it a go and see where we go. So when I was a kid, I had terrible tasting music. I really, really did. I grew up in Somerset. I listened to a band called The Wurzels. Kind of all you need to know, right? The last tell you, if you don't know who the Wurzels are, it's kind of a niche market. So I went to university and I met a guy called Mark who was fantastic and he decided he was going to educate me in the way of music. And he made me a mixed CD, right? And that endsly dates how old this story is. The first song on that CD was Pimball Wizard by The Who, right? And I put it on, I was really excited and it starts with a fantastic acoustic guitar introduction.
(02:34):
It's really fast paced. It's a brilliant tune, but it just kind of kept on going and didn't really change. And I said this to Mark, he looked at me like I was an idiot and he took me back to my room and we sat down and I put it on and it starts again. The acoustic guitar's going. And after about 30 seconds, Mark stops it. He goes around the back of my speakers, fiddles about and grumbles a bit. He comes back out and he says, right, listen to it again. So he put it on exactly the same song again, but this time I can hear the really fast paced acoustic guitar and it's brilliant. And then out of the left hand speaker comes this crashing noise, this huge electric guitar that I hadn't heard before. I just changed everything about the song. It was fundamentally different, right?
(03:14):
So it turns out what I'd done, when I'd connected up my speakers, I'd got the wire back to front, left hand speaker, completely dead. Pinball Wizard is a brilliant stereo song. So right and left hand channels, completely different. Without the left hand song, with left hand part of the song just didn't work. It was totally different. And the reason I tell you that story is that I think there are a lot of parallels to draw between me being a teenager sitting on the floor of my room in my halls, listening to half of Pinball Wizard and where we're at today in terms of understanding AI and networking. So I think we've got kind of half of the baseline, right? We understand the pace. We understand the way things are changing. We understand that every month I get a new GPU to work with. Every couple of weeks I get a new application.
(03:57):
I get a new local large language model every, what, few days, every hour I get a new set of features to work with. Painted change is unbelievable, right? And that's kind of where we're at right now in terms of understanding that right hand speaker. What I know less about is what's going to come out of that left hand speaker. So what is Pete Townsend, the guitarister who can come out on stage and play that I'm not so sure on. Things like I put into three buckets. One is agentic applications, right? There's a lot of talk around agentic services and agentic commerce, but how those agents going to discover each other, how they're going to identify, authenticate, authorize what each one of them is going to do. That's a massive unknown that does change how my network needs to operate. Things like the large language models, they're probably going to get bigger and also going to get smaller, right?
(04:44):
We're going to have really big multi-trillion parameter models that are great for general purpose stuff. We're also going to get really small language models or domain specific language models that might have a few thousand parameters, but do one job really, really, really, really well. The final thing is fundamentally for this discussion, where do I actually process that information? If it's a small model, I can probably run it on my device or I can run it on a 19 inch server on the customer site or on the CPE. If it's a bigger model, I might run it at the edge or do I have to put it in the cloud because I need that sort of capacity. These are things that I'm less sure about today. And those are the kind of questions that we're going to answer in this panel. So if I kind of complete my story, I guess what I'm really saying is that us stood here or sat here in front of you.
(05:30):
We're kind of the networking equivalent of Pete Townsend. We're about to come and play this amazing chord for you to hopefully answer as many of those questions as we can. So now I've lined us up for anything we could almost possibly can't achieve. I'll hand you back to Ray. Thank you very much.
Ray Le Maistre, TelecomTV (05:48):
Okay. We like a musical angle of DSP leaders. So yeah, as Andy mentioned, the focus here is very much on the networks that are, I guess you could say the network capabilities that are going to enable what we think the AI era is going to be. So let's start with a pretty all encompassing big question. Our telco end to end transport networks AI ready today and can telcos deliver the latency, sustained capacity and performance guarantees that AI workloads require. Now the end to end part of that is important and we added that in after a little discussion to make sure that we were not missing out the acces part of the network because that's just as important and I think makes it a little different from conversations that we've been having 10 years ago about connecting the first major data centers for cloud services as well.
(06:55):
So are transport networks AI ready. Neil, as I mentioned, the access part of it there, I'm going to start with you. How important is the access network to this whole conversation and to what extent are they ready for what's here today and what might be coming down the line?
Neil McRae, CityFibre (07:15):
Yeah. Although for some reason I couldn't make the prep call, but I would have challenged the title, building the AI backbone. Let's be really clear AI exists because of the networks we've built starting point, right? I think the second thing I would be focused on is if we're really to see the benefits of AI, then it's kind of 24 by seven, 365 days a year to the kind of micro second, then I think there's a lot to do to connect customers in a way where today their business depends upon the network. Tomorrow their business is going to be like so intrinsically linked to the network and the ability to pull AI workloads and have their business informed or guided by AI that they're going to need to kind of continually improve their network. I think if you look at the market today, you see the kind of campus and branch part of networking go bonsai in terms of sales.
(08:22):
People are replacing switches that were there for 10 years, they're upgrading, they're putting in 10 gig, 100 gig even in some offices and then people at home with more and more applications running at home and more and more of them being guided by AI or influenced by AI. In my mind, we need to keep pushing the horizon in terms of what networks can do, how they're available, how they're resilient, how they're secure, such that people can use these new tools without even realizing the networks there. That's our goal is to be kind of invisible and I think there's a lot of great work being done by the industry. We've all been building fiber for the last five years. We rolled out 5G debatable whether you claim success on that. I claim success purely on a capacity point of view because without 5G to sell our networks would be in real trouble right now.
(09:23):
Maybe we started too early. It's a fair challenge, but I genuinely think that getting high quality fiber based connectivity to literally every location is a must do and then expanding that so that you've not just got connectivity to the bit on the wall where you plug in the router, but ensuring that connectivity is there where you need it. So whether it be a cellular phone or a machine or something else in your home that it works in the way that you expect it to work. And I would argue if I remember the early days of the internet, so I helped find a company called Demon Internet. If I compare what it was like to connect to the internet then to what it's like now, I would say we made kind of a million orders of magnitude better. That's what's required for us to really take this kind of AI opportunity to the maximum.
Ray Le Maistre, TelecomTV (10:26):
Okay. Gabriella, how close are we to a bottleneck free transport network?
Gabriela Styf Sjöman, BT Group (10:35):
Bottleneck free. Oh my God, well that's ... Yeah. I don't know, but I would like to approach the question slightly different though because I want to follow up on what Neil kind of said. Now actually first question, are we ready kind of end-to-end and where do we have bottlenecks? I think the biggest challenge we have is that AI is fundamentally changing the economics and architecture of the network and in particularly the life cycle of how we upgrade our networks today. So to make it end-to-end ready and to have no bottlenecks is going to be a massive challenge. It's going to be a massive challenge because this is not how we're used to building networks. It goes so fast. Today it's evolving so fast that I would need to have a much faster lifecycle upgrade of my network end to end, which today telcos don't have and that's not how networks were built, number one.
(11:27):
Number two is for what? AI for what? I mean, we have two very different use cases. Training and inference are fundamentally different. They solve for fundamentally different problems. So the needs are going to be very, very different. So the way I used to build my network, I predict it, I build, I sweat it, it lives there for seven, eight, 10, or even 20 years. That's no longer the case. So do I build it for training, for inference? Yes, I need to have super highways, but I also need ultra low latency, deterministic, smart transport. I probably need to have congestion aware. If I want to have a need to have congestion aware, application aware transport, network in general on the transport and to build that requires much more than super high capacity optical transport. It requires really that end to end and the upper stacks that I think we'll talk about.
(12:27):
So I think no, we're not ready, but it depends what do we need to build it for and how do we predict also these needs? The more the models learn, the more the adoption at the inference and how do I predict that?
Ray Le Maistre, TelecomTV (12:38):
Yes. The old crystal ball, it's very hazy at the moment.
Neil McRae, CityFibre (12:43):
Can I answer the ball neck question? No. I mean, genuinely every aspect of the network right now, if you're at the, was it OCP, even laser preciseness of lasers are in like the third generation in the last two years because ultimately that drives how much bandwidth you've got, more importantly, the latency of it. And I think everyone in the network and vendor space, and I used to work for one time, I saw this firsthand as well. We're literally kind of reinventing things that we thought, actually that's okay now. And it's amazing to see that in fact, South Hampton University, a leader in optics, they themselves are doing a lot of work in this space and here in the UK we're a leader in that space and just the width of the laser is becoming incredibly important and most of us don't even know that's part of the supply chain, but it is.
(13:48):
So I don't think all the ballnecks are ironed out, but I would say that's been our challenge forever. I remember BBC turned on iPlayer and I was like, "Oh my God, somehow we made it work."
Ray Le Maistre, TelecomTV (14:03):
I'm sure you're coming with a slightly different perspective, but I mean, how-
Ignacio González, Red Hat (14:08):
Yeah. From my point of view, the first question is if telcos can afford it. So this is coming from USMA Intelligence that here in Europe there's a huge gap in terms of investment. If they compare, and this is coming from them, this is not coming from me. The average CapEx per connection in Europe right now is around 35 euros and for the global leaders, they are spending around 70, just twice. So how much of the investment that Telcos are doing today, which is around 85% of what they're investing is in infrastructure, they can afford to pay this bill. That's one of the questions that we have to put on the table. Speaking a little bit about other, not only AI, so if you want to invest in the access network, it's not just only for AI. Hello, hello, Rand network, cloud RAND is also there.
(15:05):
We have seen a few weeks ago or a month ago from Japanese telco operators saying that if you go for cloud RAN and you go for centralization of the RAN, you're going to have savings, huge savings between 40, 50% of what you're investing in RAN, you will save due to decentralization.
(15:25):
So I think we have to be there thinking more about what we can do. And I know in Japan fiber is by default everywhere, but my final point here is the first time that we work with an operator on deploying GPUs for AI at VH level was in 2019 in Japan and they are doing a good business there with that business. So they invested for sure. Right now they are having a double digit growth in the B2B. We have another example. Tell us Q1 of 2026, they invest, they create an AI factory, they are growing 22% in this business, 800 billions 2025, they want to reach two billions in 2028. So if you invest, potentially you're going to get extra revenue. The revenue is not coming from B2C. It has to come from B2B. So my answering the question is maybe they cannot afford it, but if they invest, probably they can get this revenue.
Ray Le Maistre, TelecomTV (16:35):
Yeah. And I guess every operator is at a different stage of investing at different parts of the network and certainly Japan is a very different environment, obviously network environment to Canada. We'll come back, Neil. I know you want to challenge that, but Colin, let's come to you next.
Colin Evans, HPE Networking (16:55):
Well, I'm glad Gabriella mentioned my favorite word in this conversation, which is deterministic because that's what this is all about. The performance needs to be deterministic. And what does deterministic mean? It means low latency, no packet loss or very low packet loss depending upon whether it's inferencing or training, but this is fundamentally traffic engineering and traffic engineering is not a particularly new concept. We've been doing it for 20 years as Neil will attest to. It's not a new concept. It's ensuring that something has a quality of service, has an SLA and that we build and we meet to that SLA. We have the observability to actually see that that performance is being met. So we define it, we measure it, we enforce it and most importantly, we also observe that it's actually doing that. So I think to me that's kind of answering a little bit of the question, are the networks capable?
(17:53):
Yes, the traffic engineering capabilities exist. Do they need improvement? Of course.
Ray Le Maistre, TelecomTV (17:58):
So the capabilities exist, but are those capabilities in the networks right now?
Colin Evans, HPE Networking (18:03):
I would say they are. Most of the carrier grade networks that we're building today already have those capabilities, but let's also draw a distinction between different parts of the network. Inside the data center, between the top of Rack Switch and the server, between racks inside the data center is one domain. Between data centers, data center interconnect is another domain. And a third domain is from the user who's consuming this AI model into the network and that can be the access network, which we know can be distinctly less deterministic than the rest of the network for a whole variety of reasons, some of which are slightly outside of our control.
Ray Le Maistre, TelecomTV (18:42):
Sure Andy, what's your take on where networks, whether it be Vodafhones or others are now in terms of what is required at this point in time to deal with where we are in terms of AI led demand, whether that's data center to data center or from enterprises?
Andy Linham, Vodafone Group (19:06):
So right now in terms of the use cases that we've got across the piece, apart from maybe a bit of extra bandwidth into the big AI factories and the training centers, yes, the networks can cope. I've not seen a single use case a customer has brought to me to say, "I need to do that and it's linked to AI and I need to have latency beneath 10 milliseconds or whatever it happens to be. " We've got the capacity, we've got the engineering, we've got the access networks now. I don't think today we have a problem. I think the problem comes in the future when we have to work with customers that aren't human beings that will sit there spinning GPU cycles that are very expensive and that's when the latency, that's when the performance really kicks in. I think we're going to have to work out how we deal with that in a really, really rapid way, much faster we've evolved previously.
Ray Le Maistre, TelecomTV (19:50):
Okay. And Neil, I don't want to leave you hanging because you were chomping at the bit to come
Neil McRae, CityFibre (19:55):
Back. I think we kind of talk about various other things. I agree right now is AI working 100%.That's where we're started. I think the question is, can we just do what we're doing and continue that trend? And absolutely the answer is no. I predict 10 years from now that AI traffic will be as big as today's internet traffic, write it down. And in that scenario, that ultimately means that the traffic everybody's sending today somehow doubles. That's a massive shift. It's a massive shift in the economics, it's a massive shift in the supply chain and also the kind of number of devices that we're connecting to the network and the use of them. But in telco, we've been far too focused on solving yesterday's problem. "Hey, I'm going to cloudify all my RAN and then suddenly save tons of money. Nobody is saving money by doing that.
(20:58):
I was in front of the guy, NTT in Japan six months ago, he said," Neil, help me take cost out.
(21:07):
"All these disag stuff, this ag's dead. AI has overtaken this aggregation because I can run an agent on any of these platforms and make them all look the same. And again, it's somewhere," Oh, this will take cost out and we've been kind of focused on these cost out things rather than these things that are going to drive revenue. Tell us, beautiful company, awesome guys, built a platform, sold that out in months. Why? Because they had capacity, simple as that. What they built was literally anyone could have done it. My wife works for one of the biggest neo clouds in Europe and the demand that they are seeing is off the ... I mean, she literally is ordering hundreds and thousands of GPUs for delivery in three years time. When I look at that, I kind of feel like going on as we are doesn't feel like the right thing.
(22:01):
Trialing technology for years when half the rest of the world have already deployed it doesn't feel like the right thing for us to do. I think we need to be more aggressive. We need to trust the amazing engineering people that we've got on our teams and we need to drive out networks that are there for people when they need them for whatever workload that they want. And by the way, we've done a pretty decent job of that, I would say. Not perfect but reasonably decent.
(22:31):
But the big change isn't so much in all that kind of capacity and boxes and stuff, it's in how we run them.
Ray Le Maistre, TelecomTV (22:37):
Okay. So if we take that a step further and let's use that idea that in 10 years AI traffic is as big as internet traffic is today and the kind of processing that is being done has shifted from where it is today. Inference starts to become a much bigger thing as you mentioned already, Gabriella. What kind of network architectures and capabilities are going to be required to support the AI training, which will go on, but also inference workloads. What does that mean for network, the capacity, where it goes, what it's connecting and how it's connecting it and how it's all managed? A long, big question. Who wants to go first?
Neil McRae, CityFibre (23:30):
Well, I'll give you an example, right? I predict three years from now, no legal document will ever be issued that hasn't walked through an AI system. What does that mean for the legal industry? That means they're going to need some way of delivering that AI capability in a secure, private and highly accurate way. What does that mean? They're going to have to build infrastructure that supports that and we will have to connect that infrastructure to enable them to deliver that sort of capability. And so we're thinking of a network that is nothing like the network that we have today
Gabriela Styf Sjöman, BT Group (24:10):
And
Neil McRae, CityFibre (24:11):
This is where the sovereignty thing that we just talked about becomes quite important because actually as a legal organization, I want sovereignty from everyone else and from my perspective, it's how do we ... There's a term in the kind of AI world called scale out. We need to be looking at that and what do we learn from that scale out thinking and apply it to how we build networks today.
Ray Le Maistre, TelecomTV (24:37):
Gabriela, I take it you're already thinking about this. You brought it up, you built it out in our discussion the other day. What does the future look like? How do networks need to evolve to deal with what we pretty much know is going to be coming down the line, whether it's one year, three years, five years?
Gabriela Styf Sjöman, BT Group (24:57):
I think that the first principle that I talked to with my team at least is to reimagine the network as a whole. So it's not only about it. I think we're talking, we're going to this super intelligence era where it's not only inferencing or learning. I mean, at the end of the day, you want all of this for some type of actionable insights. I mean, some AI can be for augmentation, but much is going to be for some type of coordination of critical assets. So I often talk about we're going to the coordination era away from the information era. So then you need to, we say, well, the network is just a super sensor. That's the number one thing, generating lots and lots of data and that data is, you use that somewhere in the data, you need that intelligent layer on top of that and then you need the applications.
(25:46):
So I think that the big discussion in addition, I think we're well positioned as telcos for inferencing. We're deeply distributed. We have both from a metro perspective, we have our POPs. I think we have an architecture that's kind of prepared us for that, but we're not prepared for us to yet, how do we monetize on it? If we just stick to providing data and just computing it, we're not really going to monetize on it. So I think what we need is not only to look at architectures, the capabilities to enable monetization, but that requires that we understand the use cases, who is going to be using that AI and for what and what's the value that the telcos can either alone or in partnership. I think that kind of providing that intelligence, that's why I talk about, and you were onto it as well, that the infrastructure needs to be intelligent, but then you need to be able ... And there is, by the way, I also agree, we have a lot of telemetry in the network.
(26:45):
We have lots of data already today. We're not doing a lot with it, especially not end-to-end, but that's the opportunity. How do we really expose that into intelligent insights for ourselves to be able to operate the network differently, but also to be able to offer our customers those near real-time actionable insights that will happen at the edge.
Ray Le Maistre, TelecomTV (27:06):
Okay. So we need to build smart networks, but I guess we need to build them smarter as well because you can't just build and build and build because nobody's got that much money and nobody's got that much physical real estate.
Gabriela Styf Sjöman, BT Group (27:19):
Exactly. They need to be smart because how we operate them. Well, I mean, if you look at inference, inference is going to be very much about efficiency. It's about not only serving millions of requests, you need to serve them quickly, but also efficiently. So that efficiency also requires an intelligent distribution. That's why you need to be application. The network needs to be application aware to be able to do this efficiently. But I'm talking about as well on top of that, how do we monetize on providing there's an opportunity beyond just providing the AI infrastructure.
Andy Linham, Vodafone Group (27:58):
I think flexibility is the most important thing for me in the future because we don't really know what the use cases are going to be that drive this type of change in our network architecture. The three that I hear the most are physical AI, autonomous vehicles and probably AI wearables. So the Metaglass, that type of stuff. They each have different requirements that they place on the network. And until we get some real traction in some of those use cases, we need to build in the intelligence, but then the flexibility becomes the most important thing because I need to be able to change how my architecture flows, the traffic flows, that type of patent really, really quickly to meet the demand that it's going to come faster than I think anybody's really expecting it to.
Ray Le Maistre, TelecomTV (28:38):
Okay. Challenge for Ignacio and Colin here. Are we ready for this kind of scenario that Andy's just described?
Colin Evans, HPE Networking (28:48):
From a flexibility perspective, yes. I mean, I'll come back to the point we were discussing this morning, automation. The automation things we were discussing this morning equally apply to this infrastructure as to, from the perspective of what we were talking about, the ability to build the infrastructure, to bring it into service, and then most importantly, to change it on an almost instantaneous basis because we need that flexibility is there Is it completely autonomous as we discussed this morning, perhaps not, but we're on that path and on that journey. Again, we should consider that we do have these two different kinds of use cases, one of which is about building AI models using large language models and that is almost a non-real time activity. We teach the model and then we reteach the model because obviously this task to have a constant feedback loop, it's not a standalone thing, but it's almost non-real time.
(29:45):
And then we have the inferencing where people are actually accessing and consuming the model, which is real time. One requires very little packet loss, the training part, which as I said, we'll probably do inside a data center. We won't do that and there's distribut needed basis. And then the inferencing piece is absolutely in the wide area network accessing the model where that is therefore very latency dependent and jitter dependent. You have the round trip delay of the processing time and then the user expects a response very quickly, particularly if it's one of the use cases that Andy was mentioning like autonomous vehicles or something like this, that response will be required very quickly. So that's kind of the way that I see things working. But yes, we have the flexibility there. The networks are designed to provide that flexibility connectivity from a compute node through a tunnel, that tunnel can be changed in milliseconds to be somewhere else and connect somewhere else.
Ray Le Maistre, TelecomTV (30:46):
And Ignacio, obviously the concept of distributed networks has been around for a long time. This is going to put a lot of more pressure on the edge of these distributed networks. How does that play out? What does that mean for companies like yourself and the kind of discussions you might be having now with telecom operators about what they might need in the next few years?
Ignacio González, Red Hat (31:10):
We are now talking about the fabric AI. So where you can think about this distributed AI, wherever you're going to need it, you have to deploy, let's say, without a huge investment so that you can be more not only flexible, but fluid in the way that you can cope the AI demand. That's one of the things that we believe is important to cope with this. And in terms of the focus in terms of this north, south traffic, so from the user or the device, all the way to get the answer from this is very important in terms of ... I'm not sure in total in terms of latency, but in terms of the user experience that they can get, that's something that I believe we have to be focused. In the training, I don't believe we have a huge problem because it's a massive monster and it is very well prepared in the access for doing this distributed AI is when I believe we need to put the effort and that's why we are working with this Fabric AI.
Ray Le Maistre, TelecomTV (32:05):
Okay. I'm going to come to the audience in a minute, see if we've got a question for the panel, but I want to come back and I want to see if we can actually understand this because we really are talking about the transport network here, what telecom operators can do to help make sure that the data is in the right place for processing, that everything is connected as it should be. Is the network architecture going to look any different? There's a lot of talk now about edge becoming a much bigger play here in terms of inferencing and some of the use cases you talked about there, that sounds like a lot of the processing might happen within a private network. So where does the transport network requirements from telcos end and how might it change in the next few years if the robotic revolution takes off in a way that looks like it might do?
Andy Linham, Vodafone Group (33:10):
It's a good question and that's the sort of thing someone always says when they're thinking about a really good answer to go with that good question. I think in terms of the physical AI stuff, a lot of that process is going to have to happen really, really, really close to the robots, whatever it is that's kind of using that physical capability. Any sort of latency, it could cause safety issues. There could be that type of concern coming. I think for autonomous vehicles, you're looking at things like image analysis, that type of stuff, which needs decent amounts of bandwidth, but not mega amounts of bandwidth. A lot of the work that's going on now around things like 6G standards that, believe it or not, are being worked on, which I still find a little bit irritating because we haven't really fixed 5G yet. But anyway, I think with the 6G stuff, a lot of that is going to be around concurrency and around ultra low latency and making sure when you're out and about in these types of autonomous vehicles that you can get a reliable, resilient connection.
(34:02):
I think from a transport perspective, the whole point of doing this processing at the edge means that you take a lot of the burden away from that kind of core transport stuff. I don't think massive amounts of change will happen certainly in the next couple of years to the transport networks. I think it's going to be access to edge where you're going to see most of that evolution.
Ray Le Maistre, TelecomTV (34:20):
Okay. And do the network operators have the infrastructure they need right now for them for the next couple of years, even if these trends and use cases do come on? Are we good?
Neil McRae, CityFibre (34:34):
I think we are right now,
(34:36):
But 108 years ago, we were good for telegraphs, right? So I mean, this is the thing we neglect to reflect on is how many technology changes have we gone through? It's quite a lot and probably a few other industries have had the same level of velocity, but I think we need to keep pushing. The CEO of Nvidia, Jensen, why do you think he's spending so much time talking to telecom operators right now? You know why? Because we are a barrier to his success. If we aren't able to carry ... And he knows it. If we aren't able to carry this traffic and deliver it and make AI the amazing thing that he believes and many of us believe it can be, then he's in trouble and that's why he's like, "Hey, look here, have a billion quidswork for GPUs and see if you can make it go faster a bit." He's all over us for a reason because he sees us as potentially in the way, much like Meta and Facebook where we've tipped years ago, similar sort of strategy.
(35:47):
And actually in some ways I think it's helpful because I think 5G has kind of like put us on a bit of a, rightly so because we haven't made any money from it, right? So how do we double down and continue to invest and continue to drive the step changes that we've had over the last 180 years? So I reflect on that and I just look at the investment he's putting into the telco space and think something in this. So the question we've got is yet, how do we really embrace that? Can we embrace it? Because some of it's like, "Hey, give you all these GPUs, but you got to go write all this really hard code with our libraries, et cetera, et cetera." It's quite difficult. But on the other side of it is it's also about enabling us to run the networks really well and something Colin said, I don't think really caught.
(36:43):
If we keep running the networks like we are today, then it doesn't matter what we do technologically, it won't translate into value for customers. We've got to be able to deliver services predictably. And that's not just on stuff we've already built. It's like, "Hey, I'm digging into your building because you've got this new AI change in manufacturing and whatever it is. Help us be predictable about this so that when I say on Thursday I'm coming to di, you're ready for it. We've got all the permits, all of the glorious government bureaucracy has made really easy for us." And now I'm really dreaming, but every part of what we do to enable connectivity is going to be super important. I super believe that 6G should be about making all the black spots go away, all the not spots go away because having that predictability, knowing that wherever you are, you will have connectivity will drive more use cases and more people who want to use the network to do things that, well, I don't have network in this part of the world or I'm in this car park, I suddenly don't have networks so I can't rely upon this.
(37:56):
We've got to take all of those barriers away so that this AI and the offshoots of AI, which I don't think we've really seen yet, really get the momentum they deserve. So thinking about how we take what we've got and push forward is crucially important.
Ray Le Maistre, TelecomTV (38:12):
Okay. Gabriela, and then I'm going to get Tony to go to whoever has their hand up to get a question to the panel.
Gabriela Styf Sjöman, BT Group (38:19):
No, I just want to say I fully agree with what Neil said. I think though that to get to that is not an infrastructure challenge, that's a data challenge, kind of that data layer that we're so bad in because to be able to predict or to serve our customers, even to operate the network differently, or we talk about deterministic, all these things is less about the infrastructure and it's really about what we do once again with that data. So that data layer, which we haven't talked about, it's not the purpose of this panel, I think that is even more important to invest in that is going to be as important as investing in the infrastructure if we want to capture this opportunity.
Ray Le Maistre, TelecomTV (39:00):
Okay. We have a question. If you can say your name, where you're from and give the question to the panel. We have just three minutes left and six people on the panel.
Dean Bubley, Disruptive Analysis (39:10):
I'll be quick. Dean Bubley, we've talked a bit about infrastructure for individual networks for AI. What do you see as the future of interconnect between networks, particularly for things like Agentic AI where agents might need access to the rest of the internet rather than running workloads locally? Do we need more peering points and other exchanges? What does that look like?
Neil McRae, CityFibre (39:38):
Let me tell you that, because I'm involved in that part of the industry. So I'm a director of London Internet Exchange, the best internet exchange on the planet. So I think it means we need to think interconnect very differently and actually for me it starts with, it could be our handsets just talking to each other directly and for that actually the point made earlier about access and data and securely. So, "Hey, I want to have this exchange and I need these resources or these capabilities, I can authenticate, I can get charged for it, I can get a receipt for it and then I can do my transaction." That is kind of a world of opportunity and actually now with AI, there might be some actual way of delivering that and making it work as opposed to trying to knit together a million systems that literally falls over the day after you turn it on.
(40:34):
So I genuinely think that's the future and exchange points like Links and other places will grow to meet that demand. Links used to just be in London. We're not all over the UK. In fact, we're in Africa and we're in the United States as well because Interconnect's so important also requires the senders and receivers to want interconnect. That's probably where there's some challenges and headwinds ahead of us.
Colin Evans, HPE Networking (41:05):
Can I flip that on its head, Neil? If I'm going to bring up the word security, security to me is part of the network. Security is criminals and bad actors. If we're going to blindly believe that AI tells us the truth, criminals and bad actors are at some point going to try and infiltrate that truth and tell us things that aren't true for a vast variety of reasons, which I'll let you imagine security is absolutely critical. If you're going to believe in AI's technology, then security of the models, et cetera, becomes absolutely critical. We should never forget that that is a fundamental part of the network infrastructure.
Ray Le Maistre, TelecomTV (41:42):
Okay. Okay. So this is going to have to be the shortest question to one person.
Neil McRae, CityFibre (41:51):
Oh, you're giving a mic to the wrong guy. Because
Ray Le Maistre, TelecomTV (41:53):
Tony's on streaming live in one minute.
Andrew Collinson, Connective Insight (41:59):
The question is, you've all said it's really uncertain what's going to happen, and I think we kind of buy that. Neil talked about a case study with legal. What are you doing to think through the options and the possibilities for the future? Because you can't just be ... I doubt you're sitting there going, "Ooh, we don't know what's going to happen. We'll kind of just cope."
Neil McRae, CityFibre (42:18):
I think you have to go out and talk to the network customers and users. I mean, they're ultimately the people that will decide this. I presented at this legal conference, that's how I kind of ... And all the comments was talking about AI. I was shocked by this. It's a bit of a wake up call actually. All the things that we didn't do in 5G I would suggest we have to do and we need to build the right ways to understand what our industry partners who are building cars and there's a car section coming soon. We need to learn about each other's business because only in that kind of handshake will we deliver the sort of capabilities that we all dream about, at least in my view.
Ray Le Maistre, TelecomTV (43:02):
Okay. Great answer and a great way to end the panel. These are going so quickly, but that is all time all we have time for. We now have a break of 30 minutes. We're back at 4:00 PM UK time for the final session of the day, which is the one that Neil referenced there. So for those watching online, the next extra shot with Tony Poulos is coming up soon. For those here in Windsor, the Pinball tournament is back on. Have a go on Jaws if you haven't already. The coffee cart is open again and refreshments are being served in the foyer, but let's have a big round of applause for our panelists. Thanks so much.
Please note that video transcripts are provided for reference only – content may vary from the published video or contain inaccuracies.
Panel Discussion
A panel of industry experts from Vodafone, BT, CityFibre, Red Hat and HPE Networking assess the readiness of telco transport networks for AI workloads, examining end-to-end infrastructure capabilities from access networks to datacentre interconnects. They explore the fundamental differences between AI training and inference requirements, the investment challenges facing European operators and the need for deterministic network performance. They also reveal how AI is changing network economics and architecture, the importance of traffic engineering, and whether current networks can deliver the latency and sustained capacity that AI applications demand.
Broadcast live 19 May 2026
Participants
CO-HOST
Andy Linham
Principal Strategy Manager, Vodafone Business
Colin Evans
Snr. Director Business Development, HPE Networking
Gabriela Styf Sjöman
Managing Director Research and Networks Strategy, BT Group
Ignacio Gonzalez
Business Development Manager, Telco CoE, Red Hat
Neil McRae
Group CTIO, CityFibre