Guy Daniels, TelecomTV (00:09):
We are continuing our foundation readiness theme for this session, and we've got a series of presentations for you, which will take us up until the lunch break. So three presentations. Let's start first of all with Anita. Anita, if you like, a clicker is on the lectern there. So Anita Dolo, who's the CEO of NGMN.

Anita Döhler, NGMN Alliance (00:34):
Okay. So the speech today is about agentic AI for autonomous networks. What I will explain here, everything is contributed by our membership. We are an operator-driven industry alliance. We have vendors and academic institutes also in NGMN. I'm very, very happy that 6G was mentioned already in the morning by Diane and others. So our focus topic is 6G. That's why if you look at the logo, you will see that the circle has a 6G gradient just as a kind of side effect here. And we work also on sustainable AI on green future networks, other sustainability topics, smart antennas, and of course, everything related to what is needed to move ahead to autonomous networks. We are cooperating with other organizations, so of course we are not duplicating what other organizations are producing. When it comes to AI, to agentic AI, we took a holistic approach. So first of all, AI doesn't need 6G, but moving forward, we need to take care that the 6G or the next generation network is prepared for agentic AI and based on native AI capabilities.

(02:13):
This is why we published beginning of the year a document called AI Search and its impact on 6G. Later, we looked into what does it need to embrace agentic AI for operating models? And just recently we published the topic, what I will explain now, agentic AI for autonomous networks. So the team which was tackling this topic was consisting out of operators led by Vodafone and China Mobile, but also other operators contributing actively and five vendors. So this is a good example where actually collaboration can work. And the objective was to analyze whether there are any gaps and what is needed basically in a first stance to provide input to 3GBP's meeting, which happened back in August. It is clear that audentic AI is the path to the TM forum level four of autonomous networks. And the team selected three, what they called high value use cases.

(03:35):
So fraud management, run optimization and service assurance. And they were analyzing what actually is needed looking at each specific use case to make augentic AI happen. There were a few findings, of course, related to architecture topics, to the hierarchy, to the role of digital twins for validation purposes, but also the, let's say, fact that a lot of traditional tools already used for network automation and those tools, they should not be thrown away. So for instance, machine learning algorithms which are in place running well. So the agents, they should ideally make reuse of the existing running tools. And the finding was that actually it needs a telecom grade harness. It's not sufficient to have intelligence models or agents. So it needs six capability pillars, which you see on the slide to make it happen that we can generate trust in the system.

(05:00):
Further, the team analyzed all the different features and requirements. And I will not speak about those features because you can download the document. It's really a worthwhile reading. It was a lot of takeaways, but the main finding was that actually 10 organizations or even more are working at the moment on the topics which came out of the analysis. Naturally TM Forum covers a lot. I know that TM Forum colleagues are also here, but also other organizations are working on specific topics. And the conclusion was that if we as an industry do not manage to collaborate and to avoid fragmentation, we will end up with fragmentation.

(05:59):
In detail, they came up with 12 areas where we have as an industry currently a risk of fragmentation. And naturally the finding is we need a more systematic approach to collaboration between organizations. It needs a telecom grade zero trust agent ecosystem and other capabilities where this coordination of standardization is needed. And just also to mention, you saw the list of organizations. It is not only about the so-called SDOs, so standard development organizations. There are also other industry alliances and also open source communities. So what the team of NGMN members is suggesting is that we focus on those six pillars in the context of systematic collaboration. So what do we do in NGMN? Of course, we continue our work. Just today we published another input related to 6G migration options, but we also started just now in the framework of our 6G work, a project called Network for Agentic AI, where we will continue with the entire membership to analyze use cases and what does it need for the architecture requirements and for other capabilities in the network to prepare our networks for agentic AI.

(07:49):
With this, I am at the end of the presentation.

Guy Daniels, TelecomTV (07:54):
Thanks so much. Come and join us. Round of applause. Thank you. We only have one clicker this year, so we only have one clicker this year, so we have to share. Thanks very much, Nita. Sit yourself down. Fascinating. So look, I'd love to know where this work go. Where do you see this work going? Because you've had your next steps on the slide there. Where do you want to see the outcome? How do you want this work to develop?

Anita Döhler, NGMN Alliance (08:22):
Guy, that's a very important question and I kind of expected a question even though we have not spoken about it because to coordinate work between organizations, it's not easy. It's not easy because every organization is membership driven as we are. So what we are doing is decided by our membership and many of your companies here in the room are part of NGMN as well, but also you are part of other organizations. And I think there is always an opportunity to collaborate. We can also start joint projects, but it also needs that the members of different organizations coordinate this internally. We are all aware that quite often different people, different teams are serving different organizations when it comes to all the different industry alliances and SEOs, but it really needs also a coordinated approach within the companies, being at an operator or a vendor. And within NGMN, of course, we will continue to support this coordination.

(09:29):
And I can only invite everyone who is not yet in NGMN to join us because it is extremely important that we get it right.

Guy Daniels, TelecomTV (09:37):
Absolutely right. Thanks very much, Anita. Right, we're going to move on to our next presentation now. And David is returning again. So good to see you again with us. David Warner from Blue Planet is going to present next.

David Warnock, Blue Planet (09:52):
Okay. So I'm going to give you 10 minutes on what is probably the least glamorous problem in AI native networking, and that's the data beneath it. Just give you a few seconds to take that in. What's wrong with this picture? So I came across this on a social media platform maybe two years ago now, and the poster is asking ChatGPT to draw a map of the UK and giving it some parameters to deal with. But what it came out with is something quite surprising really. If you can see anything wrong with this, just shout out.

(10:33):
I mean, I think the first thing is that the south of England is actually in the north of France. Now, historically, that may have been true. In the 13th and 14th century, Britain and France or England and France were at war with each other and parts of France were in the UK, but it's not the answer that we're looking for today. Also, you might notice that there are five Bristols, one of which is in Northern France and it's called Bristol, which interestingly enough is how people from Bristol pronounce Bristol. Now this is image generation, right? It's not AI, but imagine this is a map of your network.

(11:15):
So the failure mode is identical. The model wasn't lying. It was reasoning in confidently over data that had no reliable structure underneath it. And this is exactly what happens when an agent reasons over unreconciled data. So the hard part of AI native operations is not the AI, it's the data underneath it. And Gartner has proposed that 60% of AI projects will be abandoned in 2026 unsupported by AI ready data, not because the models failed, but because the data was never ready for them. So what does ready actually mean in this context? Well, to me, it means four things. So first of all, we need to have trusted data so that we can make reliable decisions. Does the record in the network and the inventory management system match what's actually in the network? Because if our inventory management system thinks that a card is in slot three, but it was moved 18 months ago, then every downstream decision made from there is going to be wrong.

(12:20):
The data needs to be connected. So can you correlate across multiple domains?

(12:26):
When you get, for instance, a fiber break, it generates hundreds of alarms sometimes and those alarms will be across multiple domains. How do you correlate that into a single root cause alarm so that you're raising a single ticket rather than many tickets? We need to have operational context. So relationships are important, not just records. What service does this port carry? In the case of our fiber break, there could be hundreds of services running over that particular piece of fiber. So which of the services are impacted, but also there could be SLAs associated with those services, and so it could be costing us money. And so understanding that operational context is very important to us. Lifecycle awareness is also very important. A service can be in a pre-production state. It can be live. It could also be undergoing change management, and that's important for the network operations to know that.

(13:24):
So understanding the life cycle, the part of the life cycle of that service in is also very important to operations. So clean data is not a data quality project. It is these four properties held in continuity.

(13:41):
And I'm going to talk about ontology for a minute here. Everything I just described assumes that the machine knows what the data means. That is what an ontology does, not a schema. A schema tells you the shape of the field, and ontology tells you what that thing is, how it behaves, and what actions are possible against it. So just to break that down for a minute, what the data means, well, it's a common vocabulary across multiple domains and across multiple vendors, such that if we just say what a circuit or a port or an interface is, that means the same thing across all vendors and all domains. Secondly, how the network behaves, the real-time relationships and the dependencies in the network, and also what actions are possible so that an agent can make a sensible decision and can reason about the options, not just report a state.

(14:37):
So this is the difference between an AI that describes your network and an AI that you would let touch your network.

(14:45):
And this is a journey, right? And I would say that probably every organization that's represented in this room is in a different stage of that journey, but there are important milestones along the way of the journey that I think need to be considered. So the first thing you need to do is you need to be able to see the live network. You need to then understand that network across multiple domains and to give a unified view of the network such that you have a topology. You need to be able to trust the state of the data, so it needs to be live, it needs to be federated. And from there, you can actually start to think about the context of the data and you can work within the life cycle of that context. That then allows you to share intelligence, but to move to full autonomy, you need to be able to do closed loop automation.

(15:38):
Now, we have a saying in Blue Planet, you can't AI what you can't see, which is true. And I've got my own personal saying, and that is, how can we do closed loop if you don't know what the root cause of the problem is?

(15:51):
So legacy operations improve when a human looks, but autonomous operations improve when the data is shared, and that is the whole difference. So one slide on how we do this at Blue Planet, and then I'll get back to the problem. We have an inventory management system, an orchestrator, and an assurance system, and now we've just released a change in configuration management system. Underlying this is a common data layer. It's all installed in a native AI platform, Kubernetes based. On top of that is a common data layer, and on top of that is an AI layer so that agents can actually share the data across those applications and make context sensitive decisions on how to remediate various problems.

(16:44):
Back to the problem now. Now, I've been working in OSS for more than 35 years now, I think. I think the gray hair is probably a testament to that. And I used to work in a NOC back in the mid '90s, and we had multiple assurance systems for different domains. So we had a mobile assurance system, we had an IP assurance system. We also had lots of EMSs for the different transport domains. And when we got something like a fiber break, we would get alarms on all of those systems and we had to go between them and figure out what the root cause of the problem was. We'd have to do it manually because we didn't understand the context of the whole network. Our inventory management system was a whole bunch of spreadsheets. Our orchestration system was people sitting at keyboards doing command line configuration, and that's developed now.

(17:36):
So I think what we need to do from there is to actually converge those systems so that you've got a single inventory management system, a single orchestrator, and a single assurance system across multiple domains. That needs to be supported by a common data layer and underneath that a context fabric so that we have a full end-to-end topology of the network. But to move to fully autonomous systems, we need to think about capabilities such as digital twin. We need to think about context graphs. We need to understand what the context of the business is of anything that we do to the network, and that allows AI to make context and business level decisions. And we're moving away from the time that I spent in the NOC manually doing correlations and things like that to actually a supervisory mode for operations. Interestingly enough, our customers don't want to go to level five at the moment.

(18:36):
They want to have a human in the loop, and I think there's legislation, there's various legislation in different countries and different regions that doesn't allow to not have a human in the loop.

(18:48):
So automating a silo makes the silo faster at being wrong, but business aware operations means that the decision knows the customer, the service, and the commitment beyond the port. So I'm not trying to sell you anything here, but I would ask you to take these questions back to your AI program, and I'll just concentrate on a few of them. First of all, can your AI explain where its data came from? Is your data connected or just collected? And do we clean the data once? Do a data cleansing program or do we do that in every silo forever? Because if you cannot answer these questions, your AI program is going to produce a very confident map of a country that doesn't exist. Thank you.

Guy Daniels, TelecomTV (19:39):
Thank you. Thank you, David. Thanks very much, David. Come and join us for a second there. It seems to me, from what you're saying, it's coming down to what we've said a few times this morning in this particular part of the event, that it's down to business outcomes again or business aware operations -

David Warnock, Blue Planet (19:59):
Absolutely.

Guy Daniels, TelecomTV (20:00):
Is the key, and we're going back from that. Is this something that we've really overlooked, and yet with the advent of AI, we just can't afford to overlook?

David Warnock, Blue Planet (20:12):
I think we can't afford to overlook it because if I take the example of the first map that I showed, getting the data wrong, having the wrong underlying data means that AI will make the wrong decisions. I did this myself recently, actually. I was traveling across London, I live in London, and I wanted to get to one of the more distance stations in London, and I asked Gemini if it was in the contactless zone, and Gemini said, yes, it is. So I said, "Well, what zone is it in? How much does it cost to get there?" And then the next answer it gave me was, "It's not in the contactless zone." And then I said, "Well, you've just told me something to the opposite of the first thing you told me, and it said, good catch." So AI will get the answer confidently wrong if it has bad data underneath it.

Guy Daniels, TelecomTV (21:02):
Yeah. The last thing I want to hear on an AI conversation is, "Yes, you're right about that. I was wrong. Thanks for pointing that out." And I won't talk about your map either at the beginning, because as a Yorkshireman, I was aghast at that map. I just couldn't believe it. So we'll skip on that. Right, thanks very much. We've got some Q&A from the audience later, but I'd like to get onto our third presentation now, which is a joint presentation. I'm not sure if we're going to do it together or sequentially. How are you going to start this? Warren and Prashant? How

Prashant Agarwal, Intel (21:29):
Is that Warren? I think we - Right.

Guy Daniels, TelecomTV (21:33):
I

Prashant Agarwal, Intel (21:33):
Think we have a separate one after. You want

Guy Daniels, TelecomTV (21:35):
To do it separate? Okay, so who's first? You want

Prashant Agarwal, Intel (21:37):
To go?

Guy Daniels, TelecomTV (21:38):
Warren?

Prashant Agarwal, Intel (21:38):
Your name is there already.

Guy Daniels, TelecomTV (21:41):
All right. So Warren's returning to the stage again. We saw him earlier from Wind River, and he's going to start things off. All right.

Warren Bayek, Wind River (21:52):
Let's

Guy Daniels, TelecomTV (21:52):
See. If his AI is working.

Warren Bayek, Wind River (21:57):
Thanks for having me again. I'll be brief because I know lunch is coming. So I just wanted to talk through a little bit what it takes to get from where we are to where we want to be. And I mentioned it this morning, the move to AI native being an architectural change, but that doesn't mean we have to, as I mentioned, to blow up what we have now and start from zero. We're not talking about a full reset. We're talking about using what we have today and enhancing it in a way that makes this AI native infrastructure architecture possible, because we have all the pieces in place right now. We've had enough resets. We have virtualization, we have cloud native infrastructure, we have orchestration, we have telemetry, we have data, clean or not. It's all available. So much of what I like what you went through, Dave, with the progression from column one to column three to almost to five.

(22:54):
I agree, humans are always going to be in the loop if telcos are involved. But what I'm saying is the underlying infrastructure is basically there. So we're talking about enabling that infrastructure to run these different AI workloads and different AI intent engines that you talked about without necessarily understanding exactly where they're running. And that sounds a little strange, but where I'm going is back to the intent based where the supervisor, if you will, and the little pieces under that, they don't need to know that this site has a GPU and this site has a specific accelerator and that site only has CPUs and this workload needs the cloud. And so all of those things are kind of built in already. So I guess in the end, what I'm saying is cloud native, which we have now, that's not the destination, especially for 6G or 5G plus, which will be able to run more AI applications.

(23:47):
It's just a methodology to get to the next stage, which is the AI infrastructure. So this is showing with this one infrastructure to be able to run the workloads we're talking about, all of these AI ready workloads. For years, we've used this infrastructure to run network workloads. They've had to do that. The cloud's very efficient at that. We're very good at understanding how to orchestrate network workloads and how to do workflows within that environment. What we're working toward is using this common cloud native infrastructure to enable the same sort of architecture in the AI space. So not to run AI workloads, but to put AI workloads everywhere in the network, to make the network an intelligent network of which AI is inherently part of the network. It's not applications that run on the network. Now, what that means is the objective here isn't to put GPUs everywhere.

(24:48):
I don't want to ruin my NVIDIA partnership agreements, but the idea here is to make sure that we can run intelligent compute where it makes sense strategically, operationally, as well as economically. And a lot of that work, frankly, will be run on CPUs. A lot of AI inferencing can and does today run on CPUs, and that's a lot of the work we've done and why we're here with Intel. Wind River and Intel work very closely together in this domain where we integrate their leading edge CPUs, which have a lot of AI capabilities. Now they can't run the hundreds of billions of perimeter LLMs, of course, because they're edge workload systems, but you're creating an infrastructure that enables the things I've talked about today. You're creating infrastructure where the network itself can run intelligent applications wherever they belong, whether it's here, the Virtual RAN, the 5G core edge apps can all run the applications and orchestrate the correct applications at the correct layer in the network.

(25:54):
So basically you have one infrastructure model running different workloads. And the final step in this is where we're going now, this bridge from where we are today and where we need to go from automation to network autonomy. And this is something actually AI came up with, and it's very common with how WindRiver looks at the world. We have the telemetry today, the contextual reasoning, I think a lot of people have talked about that. This is something that needs to sit above the entire infrastructure, and then you decide what to do and you act on that. So you're actually performing something within the network, whether it's an application that manages a device, whether it's the network itself, healing itself, finding the MTTR to improve your understanding of what's going wrong, to hopefully be able to head off problems before they happen. In the mundane world of maintenance windows, it's one of the things telcos have a lot of problems with is our maintenance window issues.

(26:55):
So to be able to figure out how to make sure maintenance windows go within the timeframe without a hitch, these are all things that fit into this self-healing and fully autonomous network. We have some of those in place today, but as we build out more AI infrastructure and AI native infrastructure, those models will go, even those parameters and those ways of doing business will be even more efficient. So the telcos will have what they're after, truly AI native networks. So the idea is to move from where we are today to a point where AI becomes part of the network. So the network itself is inherently AI enabled and it uses AI to run itself as much as it's capable to run. And we actually have some demos of this working with some of the Intel parts and processes that Prashant will probably talk about. So thank you.

(27:49):
Told you I'd be quick. Told you I'd be fast.

Guy Daniels, TelecomTV (27:52):
We'll hand over to Prashant and you will continue the story, yeah?

Prashant Agarwal, Intel (27:56):
I will try.

Guy Daniels, TelecomTV (27:57):
Good.

Prashant Agarwal, Intel (28:00):
Good afternoon, everyone. It seems I have some sort of a special talent to pick the last session before the lunch at these sort of events. So don't worry, I will keep it short so we can all have a timely lunch. So everybody is rightly excited about what AI can do, but before even talking about what AI can do, we also need to think about what sort of foundation infrastructure is required because infrastructure, compute and connectivity, they are not anymore the side player in AI story. In fact, they are the story because without the right foundation, AI is just a very expensive PowerPoint slide.

(28:51):
So before we start talking about the AI native networks, it's worth stepping back and looking into the scale of that transformation underway. So if you look at this graph, sir, within the last five years, AI has moved from a technology discussion to a business priority. So five years back around 2020, AI represented very small number of overall workloads in the data centers. But now at 2025, you can see it's almost 23% of all the workloads are AI. There is still 77% traditional workloads also. But it's not only what has happened till now. If you think about the future, trajectory is even more important.

(29:50):
By 2035, almost 50% of all the workloads are supposed to be AI workloads. 50% will still the foundation workloads, but the speed with the AI is evolving and it's deploying is much faster than the traditional workloads. And to be honest, it will continue because we will have more agenting AI, automation networks, digital twins, and this will all drive the AI growth. So the real question today is not whether AI is relevant. The question is, do we have the infrastructure ready which can be used by AI to deliver the true value? And to answer that question, we need to look beyond the headlines and try to understand what exactly is driving this AI demand.

(30:55):
Okay, so that brings to my next slide. So the first slide was important to see how the AI workload is growing, but this one points even the more interesting things here. So when we talk about AI workloads, people think about training large language models on a massive GPU cluster. So training gets all the headlines, but yet it is only the part of the overall story. In fact, inference is where the AI delivers real value. And see, training builds a model, but inference put that models to work. Every time when a customer or application or AI agent is using the model, so inference is taking place.

(31:55):
And as AI grows, so does the demand for inference. In fact, if you see in this graph, currently we see, okay, it's not a very big amount inference, but if you see the scale by 2035, 37% of the AI workloads are supposed to be the inference workloads compared to 13% training. And of course there are 50% will be still traditional workloads. So in fact, inference workload by 2035 will be almost three times than the training workloads. And that is very clear distinction for at least for telcos, telecom operators, because most of the AI use cases in telco works are inference.

(32:55):
The use cases, they require low latency, scalability, throughput, and energy efficiency. And as my friend told earlier, they don't necessarily always need a GPU. So we need to really think about that the goal here is not just to deploy GPU, but we need to think about what sort of hardware is required to run what workloads. Most of these inference workloads can be run easily on the modern CPUs with the purpose build accelerator also and the software optimization. And as Warren mentioned, we work very closely to bring those software optimization working with the partners also. So now if we think about this, now if we agree that inference is the workload which is going to be required predominantly in future, now we need to think about the next question. How can we run this inference economically? Because we need to agree that AI infrastructure is not exclusively about GPUs.

(34:12):
The success will come that how can we make the inference economics better rather than the number of GPUs deployed? So that moves me to my next slide. So if we agree that inference workload is a predominantly, now we come, how and where we can run this inference workload economically and better way. So you see traditionally the workloads in telecom, they were very centralized. You run the workloads in big data centers and that's where most of the training takes place. But again, as I mentioned, telecom environment is different. Now most of the telecom use cases like optimization and network slicing, those sort of use cases, they require decisions to be made in milliseconds. So what we are seeing, the inference is moving to a place where the data is being generated and the decisions are being made and that's moving towards the edge.

(35:29):
So now here we are getting a distributed compute and now we are moving the architecture where we have a core data center, the regional data center, and then we have open RAN edge and enterprise edge so that our core architecture is getting distributed. But again, the key point is to have that distributed architecture doesn't mean then we duplicate the same hardware everywhere because each use case have a different requirement in terms of latency, throughput, energy efficiency. And that's the key part. We now need to put GPU everywhere. GPUs are very good for training for the large language models and those. But as we see earlier, most of the use cases will be inferencing and those use cases can very much done by using CPUs and purposeless accelerators. So what's the future? What's the future of AI infrastructure? It's about heterogeneous compute where you have a mixture of GPUs, CPUs, accelerators, and its success will not come just by the number of GPUs you deployed.

(36:49):
It will come by matching the right workload on the right compute in the right place. And that's what we need to think about. So when we move to production AI, the question is not about how many AI models can we train?

(37:11):
The real question is how can we inference at scale? And that's the key point. The success will not come by how much technology we are deploying, but by what value we are creating by deploying those technologies. Again, thank you very much. I will just leave it here.

Guy Daniels, TelecomTV (37:37):
Great. Thank you very much. I think we've got about five minutes until our lunch break. If Tony's on the side.

Tony Poulos, TelecomTV (37:46):
I'm rhyming. I'm

Guy Daniels, TelecomTV (37:47):
Here. Excellent. Thank you very much, Tony. Tony with the mic. Is there a hand in the air? I think there might be -

Tony Poulos, TelecomTV (37:52):
Hang on Francis. We're going to give somebody else a chance. Any other hands coming up? Oh, thank you.

Warren Bayek, Wind River (37:58):
It's always Robert.

Amy Cameron, STL Partners (38:03):
Hi, thanks everyone. Amy Cameron from STL Partners here. I really take your message, Prashant, about the inference at scale being driven by a latency question. What I'm trying to wrap my head around is when you need to make a decision like how you're going to provision a 5G slice and you want that to happen really quickly, coming back to David's point, you've got to make a decision that is calling data from lots of different places and that is going to drive latency into that decision making process because you have to get that information from across different places. So I'm curious if you guys have a view on how you think that's going to evolve, whether it is possible to really drive down latency, if you have to call information from agents from lots of different places, or maybe you don't have to call that information from lots of different places.

(38:54):
So I'm curious to know your position on that, whether this is a question that comes up for you.

Guy Daniels, TelecomTV (39:01):
Yeah, Prashant.

Prashant Agarwal, Intel (39:01):
Yeah. Okay. So thanks a question. It's a really good question. But as I just mentioned about the inference domain workload and what our theory is that inference would be run on the edge, and that's how we try to move this latency. So for this use cases decisions, and as we mentioned about data, data is the key, but in terms of the making those decisions, which will require a low latency, that's the key point is that we are trying to run inference on the edge and not in the central databases. And that's how we think that it should be using the latencies.

David Warnock, Blue Planet (39:38):
But just to add to that, I think also if you're making business level decisions, you need to have all that information in one place. And so you need to discover the data across multiple domains, federate it, consolidate it. And if you keep that information in one place, it means it gets rid of that latency problem that you were talking about because the AI system is talking to, for instance, an inventory management system, which is within your own infrastructure. So I think when you want to make business level decisions, I think you need to pull that data from the domains into a central place. Thanks,

Guy Daniels, TelecomTV (40:16):
David. Any more comments on this or we might look for another question?

Tony Poulos, TelecomTV (40:20):
I have another question up here guy as well. I'm over on your left this time. Got

Guy Daniels, TelecomTV (40:25):
You. Right thank you. Jump up and down.

Simeon Campos, Humanist (40:28):
Hello. I'm Simeon from Humanist. We built low latency digital twins that are called world models. And my question is to you, David. So you guys with Blue Planet are building the inventory layer where you have the live visibility of all the network. It seems like an ideal spot to be going from there to predictive sort of algorithms for AI, for making sure AI doesn't break the networks, et cetera. How are you guys thinking about this?

David Warnock, Blue Planet (41:00):
Yeah, I think that's a very good point. I think moving towards predictive data so that you can be proactive rather than reactive is very important, but it's all about the data. It's identifying the systems that you can pull that data from. And when you're doing things like predictive analysis, it's more of an assurance thing as opposed to an inventory thing. You can do things like capacity management in inventory, but the bottom line is being able to identify where you can get the data from, cleansing that data so that it's represented in the same format across multiple different technology domains, different vendors, and then being able to use those predictive algorithms on clean data. I think that's the challenge.

Guy Daniels, TelecomTV (41:46):
Great. David, thanks so much indeed. Thanks for the question. That's great. We're about out of time, unfortunately. So please, a round of applause for our guests and we'll see you back here in an hour.