So let's touch on this really simply. When you think about networks, what may not be obvious is that every network has been designed around a dominant workload. If you go all the way back in history, the first networks were all about what? Voice. We had a tremendous leap forward when we made voice mobile. Then we progressed because as we were building mobile voice networks, the internet came and it was all about data. Ultimately data became rich media data video. Video became the dominant workload and the networks evolved again. And these rich media workload enabled us also then to innovate in new services. Media let us enable cloud software as a subscription or SaaS networks and ultimately innovation came the network came but there's one thing that's different now we have a new workload. Thanks, we have a new workload and that workload is AI.
AI is already the next workload and much like we've seen in the past the network will need to evolve to support the dominant workload. And there's a lot of discussion about it. Just AI is already a scale workload. There's 1.3 trillion annual AI sessions. These are discrete AI engagements. Over a hundred trillion tokens a day, 77 exabytes of traffic per month. And what's fascinating about that is it's mobile native. Over half the traffic is on mobile. And when you think about where we are in the AI super cycle, we're very much at the beginning because the majority of AI traffic, these annual sessions we're talking about, the tokens are human enabled. They're human to machine engagements. They're chat engagements, chat bots, the early stages of Agentic, and just the incubation of the potential for physical AI. What's coming is machine-to-machine engagement. And when machine-to-machine engagement gets traction and scales, we'll see a tremendous acceleration in the number of tokens, the number of sessions, and ultimately the amount of traffic. And we're on the cusp of that. And that's why we need to plan and innovate around our networks much like we did around compute networks.
And for those of you that know this industry well, you know that compute and networking always live in this trade-off and balance. In fact, compute, memory, and networking are often usually trading off. When you think about the compute story, it started with the server. The server actually was the unit of compute. And as we scaled through the internet super cycle, what we realized is we could virtualize that device. And ultimately we then scaled that device into rack solutions. This was the early days of the internet, the cloud because we realized that scaling up into a unit of compute at the rack level was essential for performance, for productivity, and ultimately for user experience. Then we moved to the data center and this is around the time when the ChatGPT moment happened. We saw this early in cloud, we saw it in high performance computing, but ultimately what transformed AI was data center scale computing. So when the ChatGPT moment happened, we were already at data center scale, but we saw a tremendous increase in the investment of data center scale. Shortly thereafter, we had another innovation which was AI factories. And in AI factories, what we realized was the unit of compute in AI was bigger than just one data center. We needed to have multiple data centers interconnected. And what's important about this journey is that at every level from server all the way to AI factory, we've had innovation and connectivity. Innovation in terms of the scale up network, how we access the memory within the server or the rack, the scale out network, how we build data centers out of racks, and now with AI factories, the scale across network. What that simply means is if you just looked at the data center, what you would see is this story of the network changing and the innovation changing to support the workload. And it's already happening in AI. And that by the way that network that innovation that's happening today and we see this today in our optical systems and our IP switching and IP routing platforms that innovation that's happening today that's being driven by AI is reshaping the architecture of the network in data center and it's driving tremendous increases in investment in transport networks and again we're just touching the surface but what's important to note is the network is changing.
And then the another question you have to ask is, okay, is it just scale? Is it just fundamentally about scale? And that's probably the question if you were looking at me, you'd say, well, you're just going to build more of these things. But really, when you get down into what's happening, it's not just scale. AI brings a fundamentally different traffic pattern. And that fundamentally driven traffic pattern is easy to see when you inspect a mobile network. Because if you think about the traffic that we had in voice, in data, in video, that traffic, that network was pretty predictable, was quite linear. And so what we would solve for was larger pipes. That's how we move from voice to data. When data became dominant, we move voice onto data, voice over IP technologies, voice over LTE. We saw that trend occur. And when data went to rich media to video, we built even bigger pipes and we started to solve the technology. We started to solve for the technological problems that that drove. This is where some of the innovation when we move from 4G to 5G, for example, in radio networks occurred. But it didn't just occur there. It occurred in our transport networks, in backhaul, and of course in the data center. Ultimately, with AI, there's a big step. We're not just talking about more capacity. And the way to think about this is very simple. If you're talking about the most intense data traffic we have today before AI, we're talking about a rich media real-time video call, right? There can't be buffering. There has to be high bandwidth 4K video. It requires tremendous accuracy and low latency uplink and downlink communications. But in AI, something changes. And what changes is the traffic pattern is no longer linear. It's now extremely variable. One moment I might have an autonomous device projecting an image and asking for information to be processed. Another moment it might be a small token of text. It might receive text instructions back for that rich media, that rich video image that it's transmitting. It might receive video back. Those things happen dynamically. They happen interactively. And the reality in the AI world is for machine-to-machine communication to happen at scale, the network will have to be different because it has to support this variability in traffic. And it's easy to talk about, well, it's uplink constrained or it's downlink constrained. It's not about that. It's about it's uplink and downlink dynamic. And ultimately, the device doesn't care. The device wants the principles of the solution of the connectivity to occur and that's the shift in where we're headed.
And if you think about some of the things we've been doing to solve this problem on the right to date, we've been innovating in static technologies. A lot of the innovation to date has been around network slicing, priority services for example. These are very important capabilities we want to deliver in a traditional mobile traffic environment. They're valuable. They provide support for first responders and in many applications. We have many customers doing that today. They provide richer or premium services for enterprises to deploy. So we see the potential and the value of what that means. But in an AI world, that's insufficient because if the traffic is bursty and I allocate a slice, I could have all kinds of variation in the consumption of resources around that. And I can't afford to just leave a huge slice open in anticipation of the peak demand. And obviously, if I underscope it, I won't serve or deliver the SLA that's needed. And this is why we need a different network. Ultimately, the networks of the past were set on a different set of principles. They were built on SLAs. They were built on principles like five 9s. We've talked about five 9s moving to six 9s. Why? Because the reliability matters. But where we're headed is not just around five 9s or six 9s connectivity. It's no longer about peak capacity and traffic planning with Erlangs. It's about understanding the devices that are coming and delivering deterministic connectivity because what will matter to a drone, to a robot, to an autonomous vehicle is that it transmits the information it needs and it gets the intelligence back to take the action it needs at the particular moment of time. And that means a massive shift. It's a shift from an SLA or policy-driven architecture to one of deterministic connectivity. This is a structural shift in what's coming for access networks.
And a simple way to think about this is not just in looking at the top because it's easy to say we're going to layer intelligence on top. Of course, the reality is we're doing this today. We're starting around this. It makes sense. We're going to add intelligence. We're going to make slicing better by sensing what's happening in the network and automate more of slicing. But layering intelligence on top but keeping the underlying architecture the same will not transform the network. What it will do is get you part of the way down the path and that's why we're seeing the activity today. But it's not enough. We actually need innovation at every layer of the architecture. We need networks that are AI native that integrate compute, connectivity and control and constantly are realigning them based on the information and the services required at any given endpoint in any given node. This is a transitional shift for us in terms of architecture. It also means that just building one layer on top will not provide the optimizations, the performance or the experience that AI native workloads will demand. It requires an architectural shift. And this is the journey that we're on at Nokia.
You know, in the past, it's pretty easy to think about this network picture. This picture depicts a simple representation of a network. And typically all of these capabilities, all of these networks were in silos. RAN was a unique investment. Transport was separate. The data center, the core of the network was separate. The cloud was separated. The data center was separate. We'd have separate access to homes. What we see in this model represents decades of investment, decades of innovation. But we also see something that's very clear to us as we study AI workloads and AI native networks that this silo model can't continue. Because when you think about inferencing, we have a little graphic of a robot up here. When that robot needs a token, needs information, needs intelligence, it's going to request that of its access network. And what it's going to need is service. And the network needs to optimize how it delivers the answer, the intelligence back to that robot. Compute maybe is sitting in a factory but that may not meet the latency needs of the solution. Compute needs to be distributed. Therefore all of a sudden our central office may be an execution engine for compute. It's possible that even in some cases in dense areas, compute could sit in the RAN and provide low latency access because if you imagine a first responder who's got augmented reality going into a situation trying to address an emergency and they need information in real time, we can't wait for the information or the insight or the intelligence to get back to that user at the edge. It needs to happen in near real time. And this is the shift in the architecture. We can no longer just design an architecture in silos. We can no longer assume that the architecture can be hierarchical. It has to be flattened. It has to be dynamic and it has to operate in a continuous cross-domain evolution so that it operates as a unified seamless network that integrates compute, control, and connectivity.
Historically in these systems, we've also built them with a blackbox principle. What's on my right, if you think about what's on my right, this is pretty much how most networks operate. We have APIs, we keep everything inside that network function closed. It provides interoperability, there's policies, we plan traffic and capacity. This is the world we live in today in networks. We believe that AI native networks require transparency and what that means is we have to have programmability built in. This is the opportunity and the need not to just say this static performance that I'm used to is sufficient but the network actually needs to evolve. It needs to determine what actions it's going to take based on the traffic that's happening in the network. And this is fundamental when we think about tokens, when we think about the real-time intelligence that we're distributing and delivering because it's not just about performance. It's also about security. This needs to be auditable so that we know what's happening is secure and we need to have trust because what happens if those machines are getting untrusted tokens? What happens if an autonomous vehicle is getting a token that's untrusted or unauthorized? It fundamentally doesn't work. You can't run your business on it. You can't reliably get in a car unless you know that the intelligence in that car is to be trusted. And this is why glassbox programmability we believe is going to be essential. So we have visibility, we have the ability if there's a compromised model or there's a compromised performance element that we can address it by modifying it. And we believe most operators of AI native networks will end up here. We already see some of this innovation today in data center and we believe it will come to access networks as well.
You know finally I want to touch on another element of this because in this ecosystem there's a lot of talk about closed architectures, closed innovation. We don't fundamentally believe that at Nokia. What we believe is that no single company can lead the transformation to AI networks. It needs rich and broad industry partnership. That partnership extends from silicon providers to system players like ourselves to cloud providers to application platforms to expert services like security and ultimately to operators and device manufacturers. And what we're focused on is engaging this broad ecosystem in innovation, building open interfaces, enabling that glassbox model I talked about, and providing interoperability and extensibility. Ultimately, when we talk about collaboration and co-innovation, this is what we're talking about because this capability is what's needed. The industry will not move at a fast enough pace if it's just one company. The industry cannot innovate quickly enough if we depend on everybody's vertical stacks. We need openness. We need interoperability and we need the leverage of ecosystems and partnerships. And we're already doing this today. That's the other key point. We talk a lot about these innovations. We're already starting. In fact, last week in partnership with AWS and with our customers du and Orange, we announced industry first agentic AI network slicing. This is I touched on this earlier. This is the beginning of making slicing intelligent and dynamic. We're embedding AI into the network. It's sensing the traffic patterns. It's seeing the performance challenges and opportunities and it's providing real-time adaptation based on intent, real-time network orchestration. This is the kind of element that's a step in the journey to AI native networks. This is fundamentally bringing AI native capability into existing networks. But it's not just there. We announced also with Telefonica a collaboration to expand their edge nodes across Spain. This is an AI ready distributed infrastructure primarily focused on enabling inferencing solutions. If you think about the world we lived in in data and in video we had a lot of content distribution networks. These were essential for delivering video streaming, social media, a lot of the applications in AI. We need distributed compute. We need real-time compute. And we need a network that can orchestrate and dynamically manage this. This is sometime talked about from a vision perspective of the AI grid. This is the beginning of that journey in Spain in partnership with Telefonica. Distributing infrastructure, small model training, real-time inferencing, ultimately the compute optimized to where the demand is, not the users adjusting to where the compute is. What's also important about this is sovereignty because what this means is that when there's critical information, critical data, it can be processed in a sovereign manner on Telefonica's network here in Spain. Very important when you think about certain workloads that may have sensitivity. Sensitivity because of privacy requirements, sensitivity because of national security requirements. And ultimately this is going to be a critical element as we get into more and more AI workloads and AI becomes more pervasive across our networks and into physical devices.
Finally, let me talk a little bit about AI RAN. We've been working hard since we made the announcement with Nvidia in October on our partnership and I want to make a few points clear because a lot of things get covered and discussed. Let me just land the facts. This is not fundamentally about putting a GPU in and leaving excess capacity for intelligence. This is about building a dynamic programmable radio architecture. An architecture that can deliver the bandwidth, the capacity, the performance, and the services ultimately to deliver intelligence. It's about optimizing token delivery, optimizing performance, managing ROI, and managing energy. And at its foundation, it recognizes that at a minimum, we recognize at a minimum, it has to perform consistently with what we see today in existing 5G networks. So that's the innovation principle. We've been hard at work. What you'll notice up here is a few new logos. We're very excited about the fact that we're on track for our proofs of concept and we have additional customers signed up to join us in that proof of concept journey. They recognize that the future networks they need to deliver need to be AI native and to handle these workloads. They need to be innovating and that's why they're continuing to come to us for partnership. So we'll hear a little bit more about this I'm sure with our panel discussion today but we're really thrilled to have NTT DOCOMO, Vodafone, Elisa, and of course BT with us on this journey.
In closing, let me touch on one thing, which is the traffic profiles changing. This isn't something new. We've seen this through every journey in evolution. We see this already with AI in the data center. It's starting to affect our transport networks. And that innovation will persist all the way down to the fixed and ultimately the mobile edge. The workload shift is what drives this transformation because ultimately when the workload changes the network has to change. We were very successful in building networks that connected people, networks that connected information. And now working closely with our customers, with our ecosystem of partners, we're hard at work innovating on networks that connect intelligence. We're excited to be leading this journey and I believe as you'll hear for the rest of the day today, we're making a lot of progress very quickly because the pace of innovation necessary to deliver the networks that are needed cannot slow down. So with that, rather than me continuing to talk, let me get a few folks up here that are working on this closely with us. So what I'd like to do is bring our panelists up. So first of all let me welcome from T-Mobile in the US Dr. John Saw.
From Nvidia our partner. Hey John, good to see you.
From Nvidia senior vice president of telecom Ronnie Vashishta. Ronnie, good to see you. We're very honored to have with us the president, director and CEO from Indosat Ooredoo Hutchison, Vikram Sinha. Vikram.
Thank you. Great to have you. And finally, the COO and EVP of technology and operations at Elisa, Sami Komulainen. I almost got it right. Good to see you. And then because I put all these distinguished technical panelists up there, I felt like it would be important to have my CTO Palavi.
All right. Well, thank you all for joining us today. Maybe, Palavi, actually I'm going to start with you. So, you heard me talk a lot about this shift to AI native networks. Maybe you can give your perspective on what's happening and what it really means.
Yeah, I was listening into you talking about it Justin. You just spoke about how the shape of the traffic is changing. You know there was a very beautiful visual about how the existing traffic is and how we've evolved from voice to data to now AI native traffic. So if you look into AI native traffic, what we are talking about is it's about burstiness, it's about the whole uplink downlink conversation, it's about multimodal traffic that we are talking about, latency, jitter, all of those aspects. Now at the same point of time if you look into it traditionally our networks have been designed for those busy hour model but now we are in a world of agentic AI where agents don't sleep. So you fundamentally have to go about and look at networks in a very different way. And then we are increasingly now at the cusp of physical AI. And when it comes to physical AI you will have a lot of distributed inferencing happening and in fact I'm sitting next to John Saw who actually articulates physical AI really well and John you have this whole concept of kinetic tokens. Why don't I have you talk about it?
Thank you Palavi and Justin thanks for having me here. You know we all know that AI is a major force now in the global economy and McKinsey estimates that the impact of AI could be worth as much as three to four trillion dollars. And but as big as generative AI is today I mean it is still largely confined to impacting the information economy where the intelligence is really confined to the digital space where it's actually doing a lot of disruption. And AI is starting to move now away from just a digital space into the physical world and interacting with it. We call it physical AI. And with physical AI, it's going to impact the real economy because it's going to be running in systems that power the real economy. We're talking about thousands and hundreds and thousands of warehouses, logistics, millions of factories, billions of cars, of robots, delivery drones, and AI devices. And so the nature of AI is shifting. And Justin in his keynote talked a lot about tokens, right? And the nature of tokens is also changing from just information tokens where you process this in the matter of seconds to now tokens in physical AI that move. That's why we call it kinetic tokens. They take action and they have to be processed in milliseconds where space-time coherency is a critical currency. Whenever tokens move or are kinetic, I think mobile operators have the license to play in physical AI. And physical AI, many believe is going to be even bigger than generative AI. Some are thinking that it could be worth tens of trillions of dollars. And we view advanced networks like 5G Advanced and 6G as the connective tissue not just to provide connections but be the nervous system for physical AI. It's going to be the intelligent fabric that connects these AI devices with the data centers and with the network edge. It's going to deliver low latency that these physical AI devices will need, security, it will need space-time coherency like I said before and when with 6G networks when you have built-in sensing capabilities, it's going to give physical AI the ability to sense.
Yeah, absolutely. And I know Ronnie, you and I talk a lot about how both of us see that AI is going to become like a first-class workload. What according to you need to change underneath in order to make that happen?
Okay. Yeah. So firstly, first class workload I think for the networks and for telecommunications and as we've seen actually in AI and compute fabrics we constantly see this evolution and change. New models are coming up all the time. The efficiency of the model, the capability of the model and now you put that in the context of telecom networks and how static and I think Justin described that very well, the black box. How do you create a compute fabric that can run both AI and telecommunications and meet all the requirements that John talked about? Well, of course, it needs to be software defined and it needs to be more open and then you can enable AI to run on that same fabric and to be provisioned accordingly depending on not just the SLA but the actual application itself. We start to see that fundamentally you've got to redesign and I think Justin summarized it very well in his talk. You got to redesign the compute fabric of the network. And we see that happening with AI RAN today. That's one of the goals of AI RAN. Remember AI RAN Alliance was formed in Mobile World Congress of 2024, right? We're now in 2026 and it's got I don't know if the number's public, but over 130 members of AI RAN Alliance. And this is now a realization that the underlying fabric can both enable AI applications to run but also how AI can improve the radio access network itself. So the radio access network connectivity becomes an AI workload essentially. This is kind of key as you start to think about the compute fabric that needs to support that.
Yeah. Maybe Vikram, I'd be curious because you've been a trailblazer I would say in Indonesia. And obviously you touch so many elements of this. I mean you're thinking about this and from a sovereign perspective, from a capability perspective, national scaling. You're in one of the fastest growing economies in terms of per capita GDP, digital natives. I'd love your perspective on do you feel like building an AI native network is practical today or is this just a POC from your standpoint?
No, thanks Justin for asking me this question. You know as I was walking in for this event something played on my mind. I think I remember last year same kind of setup we announced our partnership with Nokia and Nvidia. Ronnie was here. People were doubting, you know, you'll not believe. The moment I walked out, a lot of people said, 'Is it another Open RAN or something? You just want to fake it?' I was quiet. Last week I was in Delhi for AI summit. I can tell you I saw your slide. Now everyone is leaning in. People want to call it agentic RAN. You call it whatever you call it. First let me tell you we started early and now we are all in. Number one. Number two I think again your leadership team Aji who has been working on this Ronnie came all the way to Surabaya. You know when you talk about serving people and doing it at scale there is no better market than Indonesia. So we started with AI RAN now we are talking about AI grid. The role of telcos from connectivity is moving to connecting connectivity with intelligence and now we are very confident. Think of it like this in a village in Indonesia if you want to have a tutor it has to be so close to the village itself and we have 55,000 sites which will become a grid and that is how in this partnership we are delivering real use cases which can solve real-time problems and scale up. The other piece I think you touched upon is data sovereignty. Every country is getting serious and they should get serious. We are also building local capability. You know our team tested something with Nokia yesterday where the Indonesian talent in Indonesia in our setup where we are trying to work on the use case they are delivering this. So I go back to what you said. We are on the day zero of AI RAN getting into grid and serving people solving problems at scale. Actually we are getting into a very exciting time Justin.
Thank you. Yeah. No I think that's great to hear Vikram and we're obviously really excited to continue to push the envelope with you. Another company that pushes the envelope is actually in our hometown in our home country together, which is Elisa. And though my favorite network of yours is in Estonia, I have to admit, Sami, that's still my favorite. We do a lot of great things together and a lot of innovation. But actually one of the interesting things to me is we actually learn a lot from you because you've been known for really pioneering the principle of zero-touch operations and automation and I know now you're taking that to the next level but as you think about scaling and embedding AI and agentic AI into your architecture how do you think about that practically? What does it mean for governance? What does it mean for performance? How do you think about it?
Well, first of all, thanks for the good cooperation especially in Estonia as Finland as well. But if you think about, if I take a short background what we are doing in Elisa. So we have automated our networks over a decade already and we have an intent-based operating model in place as we speak and that is a good foundation for autonomy development. And a couple of learnings from there I would say that autonomous networks doesn't mean that you have to keep giving up control it's actually way around. You can drive a network more precisely, you can manage it more precisely. And another maybe point is that this AI traffic, it's crucial that you have autonomous networks. You can't handle the AI traffic without that. And today we have actually multiple specialized AI agents already in our network. But the decision making is governed. So there have to be digital masters in place that you can handle it. So we drive those decisions through the digital master which is a type of the digital twin where actually those agents all the time optimizing our network on the digital side and do the optimization of intended outcome and when it's ready we just provision it to the live network. So we control it on the digital side which is exactly the same what we have in live and then we just provision those decisions to the live network. That way we handle the safety and trust and it's predictable when it comes to scaling. Sure we need different kind of orchestration and operating models because the future is intent-based. So operator just kind of describes what kind of network we need and actually AI decides and drives to that direction how we do that and end of the day most probably AI helps also to the description as well. So but on the other hand it's quite an easy formula, it's technology. So if you have strong architecture in place, strong governance in place and you're driving intent-based operations, autonomous networks are already happening.
Yeah. And I think what's incredible is the work you're doing on this is you forging ahead. I mean, we're truly co-creating as we are with everybody, which has been one of the unique things about each of these customers is you push us in this partnership on co-innovation and collaboration. And to Ronnie's point in the AI RAN Alliance, it's really an alliance, right? It's not just a vendor supplier relationship, but all of the partnership and the co-development. We continue to learn a lot in this area as well because as you said and I touched on it's not just about layering on top. It's really embedding the capability all the way in and then having that continuous learning, continuous experience particularly because as John touched on the tokens are going to be kinetic and dynamic and continuing to learn.
You know Justin, as you were talking about it one of the things that we are also talking about is that the network becomes the platform. I think there's a lot of debate about can you have the same infrastructure host AI and networking workload so I was just curious to see what our panelists thought about it. Ronnie do you want to start with that?
Sure. You know often we get caught in cliches and cliches become obtuse to the real meaning. To create a platform actually has real meaning. Up until I say relatively recently, I remember actually in Mobile World Congress of 23 I was asked a question around what will be impactful and I just happened to mention AI at that time. It wasn't the first thing I mentioned, luckily it wasn't the last thing I mentioned. And then when you come to now you think well because ChatGPT moment was really November of 22. So most of the telecom networks weren't even thinking about AI but you have to have a compute platform that can support the dynamic pace of AI and it's constantly changing and it's evolving very rapidly but at the same time be a performant radio access network. We can't make the tradeoffs in radio access networks because it's critical infrastructure to enable that to happen. There's hard work that needs to be done and that hard work is getting done. Accelerated compute of course brought us a long way towards making that happen. The ability once you hit that point of being able to run both AI and radio access network together that opens up a whole host of opportunities and we're just starting to touch onto some of those now. Kinetic tokens and physical AI support, AI for RAN and how that will improve things but then of course the ability to intelligently connect intelligence with all the devices that are going to be out in the world and how we support bursty traffic. That compute platform is really now I think here today and we're just opening up the opportunities to really say okay now what do we do. You have to be software-defined. You can't embed that intelligence into hardcoded things. You know when we go from 5G was not born in the era of AI, 6G is being born in the era of AI. It's a very different. We have to use AI and we have to use a compute fabric that can support AI and RAN.
Yeah. Anything that you want to add John, Vikram?
Absolutely agree. I like the notion that the network becomes a platform. It has to be a platform that from now on processes both bits and tokens simultaneously. I think mobile operators cannot keep spending billions of dollars building single-purpose networks just so that we can sell SIM cards and phone plans. If we want to move up the food chain, I think you have to open up the network to become a platform and expose capabilities so that other innovators in all the industries can come and create new services and new products like even real-time AI services. A good example is what we announced three weeks ago where we actually opened up the IMS and invited the top AI companies in the world to plug in the agentic model so that we can roll out real-time AI services natively on our voice network so that we can do real-time services like live translation. I think that's just a small step but absolutely I think the network has become a platform and it has to generate both bits and tokens simultaneously. That is the requirement we gave Nokia a few months ago.
No, I'll just build on that. I think let's look back 2006-07. Telcos lost all the opportunity when this whole app boom was happening. We were inward looking. We can't blame anyone else. I think now we have got again that once in a lifetime opportunity with what is happening on AI to not only play that connectivity role, connectivity plus compute in a sovereign manner and think platform and solve real problems. So whenever I talk to Justin, the whole focus on innovating. I think it's an exciting time Palavi and you will see some great work also. This will bring back the respect for telco. You know we spend so much of money and with the focus on sovereign we are very well placed to unlock this full opportunity.
Anything that you would like to add Sami?
Maybe as I would like to emphasize my colleagues. So innovation is the key. So for sure it's a platform sooner or later and for sure it's native with AI. But of course it's about the innovations what we can really do with it and what kind of value we could build for the customers.
Yeah. I think it's fascinating because we've all been talking about this concept of a platform and what's implicit and each of you do this, as operators is, I see it certainly as a partner to all of you is that you already don't accept a static environment, right? And the reality and John you said it I think particularly well the legacy, the existing approach in our industry and certainly us as a traditional supplier in it has been to build hardware-based solutions that last 5, 10, 15 years. The reality is the rest of the technology industry, the cloud, the associated services, the underlying technology moves at a two to three year cycle at the longest. And that's really the shift that we're seeing. And you're already deploying, I think about deploying software in your network. So I think about what you're describing in terms of building AI native capabilities into your orchestration, into the core of the network stack and the control plane across the board you're already pushing that cycle. It's clear, for me that in closing, we need a platform and ultimately a set of capabilities that allow you to not just have a network that operates, but a network that becomes execution infrastructure. And ultimately that's the next step. That's why we're so hard at work at transforming this portfolio because we realize Ronnie to your point for 6G to be compelling right it has to be AI native which means the network has to become a dynamic execution platform. And maybe just I'll just ask each of you to close with one thing you would expect concretely in the spirit of execution that we would see a year from now when we're back together at Mobile World Congress that would demonstrate that AI native networks are real that we're on this path. Maybe just one quick bite from each of you. So Sami, I'll start with you.
If I start most probably you will see in your HQ backyard network which is at least partly or part of it is configured by AI which means that it's self-healing. There is no human errors which again means that there is less incident or even zero incident. It's built quality so kind of a value for the customer and it's for us much more efficient to operate.
So that great network that I already have in my backyard in Finland is just going to get better.
It's even better. I will come to show it to you after one year if I'm here.
Good. Vikram, look for us.
You know Wolf is also here. Last time he was here sitting with T-Mobile I think. As telco the whole focus which we learned from T-Mobile on customer love with AI it has to translate into P&L. So we want to use it for growth and solving real customer problems which can scale up whether it is for farmer, whether it is for health, whether it is for tutor and do it in a very sovereign manner. Justin I can tell you it is no more a POC. It is now real execution at action.
So real applications, real vertical solutions that deliver value and impact.
Okay. So I'm going to go back to that term platform. So there will be a platform. That platform will be generating tokens, AI applications. It will also be generating bits as John called it which is the radio access network. And the debate about can a platform do both things at the same time will be over. We've already demonstrated it. But I think the thing that we'll be doing next year around this time is seeing that actually in real life networks. And I think that proof of that platform being able to do both of those functions at performance creates that whole new host of opportunities. So that's what I think we'll see next year.
Fantastic. So a platform that's actually generating tokens like.
So a year today, a year from now today, the Seattle Seahawks will win the Super Bowl again. No, seriously, I would be really upset a year from now if Palavi doesn't deliver an AI RAN radio that I can test in the field.
Well, that's definitely going to happen. That one I can assure you of.
All right. I think we'll get some early results from our field trials with Nokia. A glimpse into the performance capabilities, the ability to run simultaneous telco workloads and AI workloads simultaneously on that platform and also get a sense of what the total cost of ownership is.
Great. Palavi, since this is the one chance you get to sign up for your own commitment that I'll hold you accountable for. So don't miss it.
No, I think my answer to this is a combination of what I'm hearing everyone talk. What we are going to see for sure is that the network and the compute and the workloads they're all going to become very symbiotic. I think we've traditionally just treated the network as a silo but we are coming to a point where the network compute are becoming very symbiotic. And I say this I know that it's not an incremental shift like it's a big architectural shift that we are talking about but I think looking at all of us together coming together to make this happen I feel very confident that we are going to go about and make that happen.
That's fantastic because I think I kicked this off by presenting a vision, a little bit of our strategy. But we all know and you actually say this mostly, you said this far before the Gen AI revolution, which was a vision or strategy without execution is hallucination in its purest form. And so maybe what we should do is say a very warm thank you to all of our panelists. So thank you very much.
And then Palavi, why don't I turn it over to you to talk a little bit about exactly what we're doing to execute this and some of the announcements that we've made this week.
Yes. And absolutely, so that John feels very confident about where we are.
Good. Over to you. Thank you.
So we all heard the panel talk about how the network and the compute are going to become very symbiotic. And Justin took an example of robots. Let's just play out this example. We are talking about a swarm of robots which are carrying a critical payload. Now, in a world where there's no human intervention, a 10 millisecond jitter is not just a pixelated movie experience, a bad pixelated movie experience. A 10 millisecond jitter is now a critical safety incident. I just want us to now start thinking about, this is where networks are headed to. And in a world where these robots are constantly on a move, where the environment is changing around them and we are constantly chasing photon to motion consistency, there is no way that we can keep on going back to the cloud to make all the decisions. This is where we have to get into a world where inferencing starts to get distributed or distributed inferencing as we talk about. And now let's go back to the example of the robot and look at how things happen. Now the robot would do some local inferencing. In this case it does the vision inferencing but it has a stabilization model that needs to run somewhere. It cannot run on the robot because the model is too big but the model needs a 1 millisecond latency. So it runs on a compute which is close by which now so happens to be the same compute that is now running your RAN workload. Then all of these robots need to be coordinated very closely and that's where a model is now running in the central office and then you keep on going back to the cloud to go about and do all your AI fine-tuning. Now in just a very simple example you see that we have exercised the entire network. The robot speaks to the wireless network. The core does the network slicing. The core speaks to the central office using access routers. The central office gets connected to these massive AI factories using the core routers, the edge routers. And now let's zoom in into the AI factories where you have many of these AI data centers. You have racks of GPUs which are interconnected to each other using classic Ethernet-based scale-up fabrics which are the switches. And then these geographically distributed AI data centers that come across to make the whole AI factory get connected using scale-across using optical connectivity. Now this is the world that we are in. And if you look into the robot example, it's not just about any one of the domain connected to each other. What it really means is you need to have a very central and tightly coordinated execution of these domains because when we are talking about the robots, it's not about the throughput of the system. What we are talking about is end-to-end deterministic connectivity which means end-to-end latency, end-to-end jitter. Now with this let's go about and dive deeper like Justin said strategy without execution is hallucination. Let's get deeper into each of these domains and
Let's talk about how we Nokia are truly building each of these as AI native networks. Now let's start with RAN because you know this is the first touch point and this is where we have our AI RAN portfolio where we have brought in Nvidia's Arc Pro platform to run on our existing AirScale deployments and onto new CO servers deployment. Now for us as I said we see that the network and the compute are going to be symbiotic elements and which is why it's a very conscious architectural choice to bring in the silicon that powers the world's leading AI data centers. The silicon that people use to go about and write all AI applications to now power telco workloads. But as we do this, we do it within the constraint of the telco which is you know within the telco power, thermal and cost constraints. Now by doing this you know there are multiple benefits that we get as part of it. But you know let me talk to you about few of these. First we blur the boundary between the telco and the compute world. You know this is the whole seamless AI grid that we heard the panel talk about which means that now you can do distributed inferencing when it comes to the world of physical AI. The second is we believe that for us there is a new mode of innovation that now opens up because by using the same silicon we are now bringing in innovation where developers can come in and go about and contribute new AI models and that innovation cannot happen when you are in a world of vendor lock-in which is why we want to get away with custom silicon and rather bring in the same silicon that is used in a ubiquitous fashion. Now the third thing which is the most important thing and you know this was something that you know I heard the panel talk about a lot. With this we are bringing in a complete hardware to software decoupling. And what it now enables us to do is you know not wait for the next G to happen not wait for the next generation to happen not wait for the next custom silicon to happen but rather go about and innovate at the speed of software.
Now with this as we moved over to the Nvidia GPUs it gives us a lot of accelerated compute and our philosophy is not just to use the accelerated compute to go about and offload the L1 processing. Of course you know we are going to run the L1 processing on the GPUs but now this gives us a very beautiful opportunity to bring in new AI models. Traditionally, RAN uses you know has functions like beam forming you know it has functions like channel estimation. It has functions like scheduling which have been done in a very traditional fashions. You know these algorithms John and I like to talk about it. These algorithms were written even before probably we were born. But now with AI we have the opportunity of going about and bringing in AI models and making these networks AI native. And this is where you know our teams at Bell Labs research our engineering teams as they are working on these models they see that as you keep on bringing up new models and more of these models you see compounded gains in spectral efficiency. By the way, these models are not written once and then forgotten because you know the environments are changing. You know, it's a constantly dynamic environment which is why the ability to not be vendor locked in to a custom silicon now gives us the opportunity of constantly innovating and bringing in these models at the speed of software. Now in terms of the road map, we announced the partnership with Nvidia in October of last year. That's where we launched our AI RAN portfolio. And since then, you know, we've been constantly going about and making changes. In fact, in a recent demonstration with T-Mobile, we ran commercial grade RAN workload on an Nvidia platform along with an AI application that was doing live video captioning. All of this using live spectrum, a C-band radio, using a 5G smartphone. And what we saw was that the GPU handled real-time video captioning along with running the RAN workload without any of the deterministic performance being compromised. And that to us is a good proof point of showing the world that we are headed in. Now 2026 is where we will do our first commercial trial and 2027 is where we will have our first commercial release out. Like Justin talked about you know there's a lot of work cut out for us but this is stuff that actually excites us a lot and by the way when 6G happens it's just a software upgrade that's the beauty of software programmability that we bring in. We are also announcing Docksuri which is our RF8 generation in our pursuit of constant improvement this is going to be 30% power efficient 20% lighter and we spoke you know the panel spoke about physical AI. This is actually built for the uplink heavy demands of physical AI and of course it is O-RAN compatible.
Now let's move on to the core because the core is the central part of the wireless network. This is the part that manages connectivity. It manages authentication. It manages data routing. It manages mobility between the mobile devices and the external network. And you know we Nokia we have a very solid foundation in the core because the core that we've built is cloud native. And now we're going to use the same foundation to go about and make it AI native. By the way AI native for us is not just adding a layer of AI on top. When I say AI native it means that embedding AI into every single feature of the core. We are bringing in agentic workflows for policy control, for traffic steering, for admission control, for network slicing. All of these where each of these functions are being rewritten, re-engineered by integrating in AI models and then bringing in closed loop automation. And you know fundamentally when we go about and embed these intelligence into each of these functions rather than adding the layer on top what we do is we bring in operational simplicity and that goes about and reduces the total cost of ownership while making sure that our core is 5G advanced ready and also ready for the 6G evolution. In fact, I would really encourage all of you to get to our booths where we have a lot of such agentic AI in motion.
Now, let's move over to the fixed networks. Now, this is where we have launched the world's first 50 gig PON for those enterprise power users. To me, this is not just a faster pipe or a fatter pipe as I like to call it. It is a context-aware pipe that has post-quantum security and mission-critical reliability. What you can do with this is you can create access network slicing which can actually guarantee you sub-millisecond latency for all of those critical enterprise AI use cases. And one of the good examples or great example of this is how we have brought in this technology and we've applied it together with you know by combining the world of fixed access and mobile networks. This is where Nokia's all-in-one small cell which is a fully contained 5G node 5G radio. It also brings in a DU and a CU combined in it. And then along with that we bring in Nokia's optics. Now you take that box, you plug it in into an existing PON network for backhaul connectivity and by the convergence of the PON and the radio, this helps our operators to go about and monetize the existing fiber deployments. In fact, one of the very interesting use cases that I've also started to now see is that our operators and hyperscalers who've been looking into modernizing, you know, the out-of-band management network that connects them to the data centers. That network has traditionally been on copper and now they want to go about and redo that copper investment and to move it into modern fiber and now they're doing it using PON because with this they get better manageability. Of course, you know, they get like power and space efficiency, but what it enables them to do is now bring back all the rich telemetry that originates in these data center devices back so that then they can go about and make smarter controllers on top of it.
Now let's move to IP routing. This is something which is very close to heart because you know I was born doing routing like I say. I was talking about all the distributed inferencing where you know the AI inferencing can happen across multiple data centers you know which is like whether it is the far data center whether it is central office or it is the centralized AI factories. IP routing is a critical glue that actually goes about and interconnects all of these data centers to make about scale across happen. And when we look into you know some of the use cases the use cases that we are talking about we're talking about storage access replication across data centers we are talking about agent-to-agent communication we are talking about transferring trillion parameter model weights or KV cache across geographically distributed decode prefill clusters. When you start looking at all of these use cases, what you really need is low latency. What you really need is lossless. What you really need is highly reliable interconnect. And this is where interconnects have evolved in requiring high-speed coherent 800 gig transport. And it's not just about higher bandwidth. You know when you look into these DCI interconnects DCI interconnects require extremely low nodal latency. They require quality of service so that you can prioritize the AI applications. It needs max for advanced security and then it needs intelligent traffic steering and you know protocols like segment routing for deterministic paths and this is where powered by SR Linux Nokia's energy-efficient 7250 IR portfolio delivers half a petabit of capacity for extreme scale high-density data center interconnect. We also have our 7750 series which again runs SR Linux and it is powered by our in-house silicon which we call as FP5 and this provides security at the edge. It provides deterministic performance for all the data center gateway applications.
Now let's move over to DC switching. This is where we have zoomed in into the world of data centers. And like I said, you know, today we are building up these data centers as these massive racks of GPUs that need to be interconnected. And this is where our 7220 ISR portfolio. It addresses different fabrics at scale. You know, this is where we bring in the world adopted Ethernet-based scale-out fabric which enables all of these collective communications to happen. You know, we talk about all reduce, we talk about all to all, all of these collective communications that actually enable massive AI training workloads to happen. But beyond connectivity, what is also needed is cognitive routing. And this is where we bring in the smartness to handle load balancing, to handle congestion control in real time so that you know we can keep up these AI clusters and make them performant. Now on the software side we bring in SR Linux. SR Linux is renowned to bring in the automation and it has like deeply ingrained telemetry that enables these data center controllers to function. It also brings in Rocky and UEC support and we also understand that our hyperscalers and our Neocloud providers they also some of them want the support for SONiC ecosystem which is where along with SR Linux we also bring in full support of SONiC.
Now the beauty of all of this if you look into it like you know fabrics are fabrics but the most important thing that we build on top of it is something that we call as event-driven automation or EDA. What it does is it brings in intent-based provisioning and it uses AI ops to go about and design and operate the network in a very autonomous fashion. And something that I'm very excited about because I keep on talking about how you know network and compute are symbiotic. EDA in our next release is going to interface very directly with Kubernetes and workload schedulers via APIs.
Now let's zoom in into optical. This is the most important part of our AI offering. You know, this is where we have low power, high reliability, secure supply chain and multi-vendor interop. And you know, in this world of optical, this is a market which is rapidly evolving and this is where all the innovation that we do in material sciences becomes absolutely essential. You know, this is where we are bringing in full end-to-end vertical integration. You know from DSPs to our indium phosphide PICs all of them coming out from our fabs being packaged together with the system the software and the automation that comes as part of it now for data center interconnect and scale across we have coherent pluggable solutions that offer 1.2 terabyte capacity per interface over the longest distance and then you know our future evolutions will deliver even higher capacity in smaller and more power efficient form factors. Now as you see for our intra-DC connectivity for our intra-DC scale out connectivity we have our IST optical chipsets that provides a highly integrated solution that combines multiple optical functions into a monolithic chip resulting in low density low latency and power efficient connectivity.
Now you know I spoke about all of these domains. Now let's zoom back to the robot example. Now for the world of robot right it does not care if the RAN dashboard is green but the transport is saying that you know there's going to be a 1 millisecond delay because if that happens the mission of the robot has actually failed which is why it is very important to now bring in close coordination and management of all of these domains together. And this is the work that we are on. You know we have our domain controllers. You would have heard you know names like Manta Ray, Son, Aliplano, Wavesuit, Edi just spoke about. What we are doing is we will have these domain controllers now move over to be cloud native and on a singular platform so that they use common services they have a singular data foundation and they become the local brain power of the domain and on top of it we are building a fabric that then enables to go about and do the uber decision making you know the tight coordination that I was talking about. And you know one would ask like you know what give me an example of this fabric one of the use case of this fabric is as soon as this fabric sees that you know there's a congestion that threatens you know somewhere about 1 millisecond of a delay it can quickly go about it's not going to be in a world where it's going to go about and open a support ticket and something is going to happen it autonomously goes about and shifts the workload from the cloud to the edge to go about and protect the mission of the robot.
Now we do this with the glass box governance framework. Justin just spoke about it. What it means is you bring in the autonomy to go about and make these decisions happen but with clear rules with clear boundaries and with clear auditability. We are getting it into a world where the networks we want to build is where humans set the intent and the system goes about and executes it safely but within the guard rails that were set up for it. You know in some sense you tell the network the outcome that you want and it can come back and tell you that yes feasible but with trade-offs or not feasible right now. This is the shift that we are making from networks that are just up to networks that are transparent, networks that are predictable, networks that are acting with clear intent in a very safe fashion.
Now I'll just go about and close and I'll talk about you know how we see that it's no longer each of the networks operating in silos. We are getting into a world where there is a symphony of all of them coming together to make truly AI native networks and connectivity happen now and to bring all of this you need intelligence and intelligence does not happen by chance. There are two key ingredients, data domain intelligence. And this is where Nokia stands apart because with 30 years of experience, we understand the network like I like to call it all the way from the waveform to TCP IP packets. This is where we are the only trusted vendor that owns the network from the radio all the way up to the cloud. The network for us is no longer a static pipe of connectivity. The network for us is a distributed nervous system for connecting intelligence. Thank you.
Thank you so much for such a thought provoking and exciting presentation. I'm sure that what we have set here today has already provoked a lot of questions across the room. So with that, let's get into the Q&A section. Pavi, please take a seat. You're not off the hook yet. And also, could I welcome back to the stage Justin, please? And in addition to Justin, could I also have on the stage David Herd, the president of our network infrastructure, Ragav Sakal, our chief customer operations officer, and Marco Ven, our chief financial officer. My name is Maria Visman. I'm a head of our corporate communications. And with that, let's we have a room full of questions and let's kick off.
Before we take questions, if you put your hand up and we have a lovely Nokia staff here handing you the microphone. Before you say your question, can you please say your name and also the company that you're presenting very clearly. Thank you. And with that, let's get started. Okay, we have first question over there. Over to you, Kieran.
