Guy Daniels, TelecomTV (00:14):
Right. Let me introduce our next guest and get our next two guests onto stage. So Timo and Selenga are coming up next and one here, one here. So please come up and join us, big hand for Timo and Senga please. Thank you very much Timo. And just to introduce Timo Jokiaho is Telco Field CTO for SUSE. Timo, would you like to present.

Timo Jokiaho, SUSE (00:36):
Alright, thank you very much. So happy to be here. And I start with saying something I have wanted to say for quite a while in this event. So we as a community or industry, teleco industry, over the last few tens of years, we have been building something amazing on my opinion, which is a global interconnected, interoperable critical infrastructure called Telco networks giving connectivity to all of us. I think we should be proud. I'm proud and personally I've been honored to be part of this journey since March 1st, 1977, which is 48 and half a years. So that makes Dean maybe a rookie, I don't know.

(01:30):
But anyway, I'm proud and I have no intention to stop anytime soon. We'll see what happens. So I'm going to talk, first of all, I will take a little bit departure on the topic of each applications, but there will be each applications. So I'll start with a obvious, which is that we as SUSE say we are not building AI models or AI applications or any applications. In a matter of fact, what we are building is underlying platforms for any use case in Telco and outside of Telco for our customers and partners to use. That's based on open source principles, hard open source principles. That's what we do. And the other thing I'd like to say is the last bullet here, which I will touch a little bit later, is that we believe that in AI space, in telco specifically, a consistent high quality performance platform is a key to collect data consistently.

(02:40):
To format data, to process data and to have agent AI orchestration in a horizontal platform, which is a key. So I will touch that a little bit later in context of AI. So this is not going to be a product pitch, but I need to have a little bit SUSE context here. The diagram here on the right hand side, it's not a architectural diagram of our product, it's just a selection of components, high level components out of which we built products, not just one product but different products for different industries and different use cases. So we try to be flexible. We provide choice to our customers to pick and choose, not completely freely but in the limits of what we say. And AI is one of the big themes of course. So we have a SUSE AI platform connected with SUSE telco cloud, which we will talk about or I will talk about today.

(03:47):
And the SUSE AI library has AI tools, it has agent workflows, it has inference engines, it has vector databases and all those things and associated capabilities for our partners and customers to use. So then the telco slide. So I have very high level diagrams on these platforms here, which are really high level on different hardware bases, CPUX 86, Arm, AMD, you name it. And then applying AI to these use cases on the right hand side. And I'd like to touch a couple of those in core network, I mean telco network mobile network specifically. And one is UPF user plane function, which is a critical component on the user plane on telco network, which typically, or many of the instances there are several UPFs in one network to be used for local breakout and even allowing telco edge use cases now within UPF and how to potentially apply AI in the UPF in collaboration of the platform, which SUSE does and the UPF application vendor to optimize the performance because UPF is a performance critical component, delivering peaks and packets back and forth, but also optimizing or minimizing the power consumption, manipulating the sea states and P states and frequency scaling, you name it.

(05:41):
But platform cannot do that on its own decisions because the traffic has to be taken into account and only the UPF workload knows what is the traffic situation. So the UPF workload needs to give to the platform that now you can reduce the usage of course, but now at this moment I need all the power you can have. So that's a collaborative effort between the UPF workload and the platform. Very important. The other of course is the RAN side and if we look at the O-RAN alliance architecture, which was discussed yesterday in couple of presentations, there are two different rigs run intelligent controllers. There's a non-real time RIC embedded into service management and orchestration, which has a perfect place for certain AI applications, but also SMO and non realtime. RIC needs a platform which we provide. And then there's a realtime near realtime RIC, which also needs a platform which is called O Cloud by or Alliance Nomenclature.

(06:54):
That's another place for AI functionality, but that also needs a platform like SUSE Telco cloud, but it's called O Cloud. So we do that and then O-RAN Alliance has kind of a next generation group which is defining real-time RIC, I'm not sure if that's going to ever happen, but they are working on that and that if that happens, it's likely will be embedded into the distributed unit in a way or another. Again, all those things need a platform with AI capabilities. Now, just a few words on how we tie into GPUs, Nvidia GPUs specifically on the left hand side, we just on a bare metal platform, it's just use and provide, use GPU and provide access to GPU for the workloads. That's clear. The middle one is in virtualized environment, like having multi-instance GPU concept and provide access to GPUs for the virtual other applications running on virtual machines.

(08:07):
And the right hand side is rather complex. It's a concept of virtual GPU, its needs proprietary drivers from nvidia, which is something, but it kind of similar as S-R-I-O-V used to be or still is on the network interface guard size, no more on that. Then just a little bit on non telco use cases and AI in those with SUSE telco cloud. I'm not going to touch those use cases more than just showing this picture other than saying that we are working with quite a few partners right now and customers as well to tackle some of these use cases also with small language models where we likely do not need GPUs for example, which is a big cost savings in terms of, I mean the cost of the hardware and cost of the power or the consumption of the power. So important things to think about.

(09:19):
Now I will touch observability. So I take observability out of this diagram and talk a little bit on observability because in an AI space, observability is highly important and what we can do based on our acquisition of Stack State a few years back, we can calculate or we can observe the usage of tokens, which can be translated to the usage of, I mean to the pricing of the thing, how much you have to pay. And it also observability can also observe how many Gen AI requests there has been. So this is AI focused observability, but observability is highly important also outside of the AI in telco networks because for the last 10 to 15 years we have been building this multi-vendor networks and it's sometimes hard to figure out what's happening in the networks, but observability can go together with the telemetric and go into the network and collect all the information possible and then visualize and show that information to the operator.

(10:39):
And this is what's happening in the network. So I have not much time, but this is one thing I discussed about this yesterday with Philippe Ensarguet, he's not here unfortunately anymore, but I took this one statement in the middle of the slide from Sylva project public white paper that they Sylva plans to go to AI driven networks, which is public statement, but as Philippe said yesterday, Orange is already using agent AI within the silver commercial deployments in quite a few countries commercially supported by SUSE. So this became more relevant slide than I even thought. And the big thing is that we have been part of this exercise since the day one or since day minus 10 or something like that really from the beginning. And this is a ingredient of consistent high quality performance, secure telco cloud and still what drives that forward and very nicely it drives. And what we do is that since we are one of the main contributors in the Sylva, our SUSE telco cloud is as close to Sylva stack as ever possible. It's not a hundred percent, but it's as close as ever possible. So that's my presentation. So thanks for listening to a simple Finnish engineer.

Guy Daniels, TelecomTV (12:16):
Timo, great to have you as always. Round of applause for Timo. Please sit down, Timo and see if there's a quick Q and A for you. Thanks very much. Absolutely fascinating. Let's just see. Selenga let's just see if it's a quick question for Timo from, I'm sure it might be a quick, yeah, Tony, did you have one at the back there? Oh no, it's the front. He's running. He's running, he's running. Should leave it here.

Dean Bubley, Disruptive Analysis (12:38):
Yeah, thank you. No, thanks for that Timo. Question on the RAN side of things. I wasn't there yesterday, but the general sense is that open RAN is changing and so you've got some interest in specific domains. For example, military use of O-RAN May well have real time RIC, but also how far down into the round you go, do you get involved in some of the air interface AI developments that people are talking about the sort of beam management and channel estimation and I don't know, creating new waveforms on the fly and stuff like that.

Timo Jokiaho, SUSE (13:24):
So the public information I can say is that we do have a complete deployment or implementation with Parallel Wireless, which is an interesting company as such and it's all good and done and they're going for commercial, but all these air interface, even open fronthaul, all those things, they are beyond what we do. So I personally am highly interested in following all those things based on my background. But we as Souse, we don't do that. The most detailed thing we do and we have to do is to implement timing and synchronization, precision time protocol and all that stuff between the DU and RU, but channel estimation beyond what we do.

Guy Daniels, TelecomTV (14:16):
Great, great. Thanks very much Timo. We're going to change presenters now. So then again I'm going to ask you if you will kindly take to the lectern there. And Selenga is EA regional partner manager telecom systems business at Dell Technologies. Selenga,

Selega Akiner, Dell Technologies (14:33):
That was a mouthful. Thank you.

Guy Daniels, TelecomTV (14:35):
It was

Selega Akiner, Dell Technologies (14:36):
I once again stand between man and break, so I'll try to make it fun guys. You may be wondering what is an infrastructure vendor here doing talking about applications. So I'd like to explain how Dell is relevant this application space and I'll try not to sell anything before lunch. On the application side, it is more about the compute and we exist everywhere from the cloud edge to the PCs and workstations at the end that we deploy AI workloads in. And we also work in orchestrating the hardware. But the more fun part, which is my job is because we only work in the hardware space and the platform, we get to work with amazing partners, small and big from network equipment providers up to smaller startups and cloud providers. And we enable them also in our open telecom ecosystem labs specific for telco applications because the small startups may not have the investment to go into the hardware and then we make them available and then we can combine partners together as well so that we can do joint work on use cases.

(15:45):
And then we also work on improving your developer experience on-prem if you have hugging face and Llama models deployed on-prem in your industrial AI cloud. So I found some analyst data that has not been shared over the nine sessions, so I'm actually happy with this. And this is actually very concerning or exciting for the operators because the projected AI traffic growth will surpass non-AI traffic and not so far away in the future. It's like almost on point and it'll do it in the next two to three years. Now we already know that the video dominates the cellular traffic, but the upstream video is catching up and we are combining with the AI generated voice and images and uploads and then to top it off, we are going to have new media, new formats, new devices coming up that's going to strain your network. So how do we help you manage this? Now if you remember three things today from today, other than the Dell relevance, I would like this to be this. So we want to help you run any AI workload anywhere and we want to optimize network performance. We want to accelerate your business outcomes.

(17:05):
So I talked about improving your application developer's experience. So we have partnered with hugging phase. This is not news because this has been announced two years back, but we are the first infrastructure vendor to do this. This is mainly enabling you to have secure access through a portal. This gives you pre deployed scripts, this helps you with the accelerated services, pick the right language model to work on your use cases about our Llama, about our meta partnership. So we have announced this with Llama for a while back in Dell techworld in May. We've also announced that Llama 4 is coming and that's happening end of this year. What is important here is yes, you have the on-prem sovereignty, keep your data, but also the two more important things when it is on-prem. First of all, it is the cost, right? So over the four year TCO using Llama on-prem saves you 2.5 times of that of a cloud provider on the public.

(18:15):
Or if you were to just use the API based services or direct from OpenAI, that would help you reduce the cost for 4.5 times. And also the other important thing is if you keep it on-prem, you get to use the version that you want, you do the customizations and you can keep the lifecycle of it. Now we talked about running any AI workload anywhere. I know that the session is about edge. Why is it about Edge? Because this was shared yesterday in the data. 75% of the data is already produced on the edge. And to cut a long story short, I would say inferencing is where you make the money and training is actually draining. So that's why you want to make your cloud be the training and analytics engine. You want your edge to be your cash register, which applications require edges because the latency, low latency resilience, we talked about continuing to be able to do the inferencing and buffering on site without having to connect to your main data center.

(19:33):
And we talked about right sizing. So you may not need GPUs for each use case. So that is where we help. On the edge side with working with partners. This will bring you also complications. You're going to be deploying thousands of sites on the edge. We're talking sites, factories. Well maybe co-located with your mobile sites because if we're thinking OrRAN is also an edge use case and Dell can help you also on orchestration and zero touch provisioning of this. Now I want to start with, because I'm mainly addressing CSPs, I would like to work with what you have on offer right now and how can we combine it with edge offerings. So most of the CSPs have already a 5G connectivity, private 5G offering. This could be the campus networks, this could be hybrid mobility. I'm not going to expect you to read through the use cases.

(20:32):
I'm going to talk about what is up there. So this may remind you of a baseband, which it is because this is originally manufactured for open RAN use cases. Now what we have done now that open RAN is not taking up as quick as it we thought it would. We're repurposing the ruggedized architecture because this is actually really useful because it is modular sleds. Now you may have seen this little server outside. This is our XRIK. What is interesting for you here is on the left-hand side you can host the connectivity, you can host the CU/DU 5G core applications. We do it with Nokia and deck. And on the right hand side you can host your edge. So you can fit in three L four GPUs if needs be. You can do streaming cases and we can right size depending on your video quality requirements or the general use case.

(21:32):
Yesterday we talked a lot about how we partner up with Kinetica about autonomous networks, how do we work with self-healing RAN, et cetera. Kinetica is a great partner of ours that I've had the privilege to work with over the last two years. They work very large scale analytics platform and they basically can stream any streaming data that can be stamped with a time and value. So you already hold geospatial data, you have a treasure trove of it. Now what can we do? If you overlay this with the weather, with the traffic, with, I don't know, the trading. So this can create a lot of use cases for you with the data that you already have. They're real customer use cases. I'm very conscious that we're streaming. We can talk about them in person, but I'm not putting the names here, but these are in application.

(22:33):
So we have a logistics use case that is used for the real-time routing and monitoring of trucks. This is a postal service. This is used in a bank. So using the high speed analysis of the trade market and then you can make better investment decisions. I know this is going to sound a little diabolical, but it is a good insurance use case as well because this is used in the US to correlate weather data when they're prone to tornadoes and they basically trace, oh, is this person's house really impacted by the tornado? And then this impacts the insurance claims or how fast you can access your money. Of course, similar to the logistics, a lot of these use cases can help multiple applications. Imagine aviation drones, defense is already in use. And then you can do again, route optimization and target optimization. I'm going to talk about a few of the use cases that Dell is also involved in.

(23:39):
One of them is the AI enabled traffic monitoring system. You may have seen a demo of this if you were in the Mobile World Congress. This is our safe pedestrian use case, but also the one I put here is from a real APJ operator that is using the crowd movements under control. And when I mean traffic monitoring, this is actual traffic monitoring, not telco traffic. We have a retail use case I'll try to reveal but not reveal the name of the fast food chain as tactfully as the question was asked before. This is not a connectivity case, but we're there with Google. This is low latency analytics for equipment monitoring operations, the mobile app and the self-help kiosks in the fast food chain. Now this is for the use cases of reducing disruptions and improving customer experience. Now we also have mission critical enterprise AI applications.

(24:39):
That is actually with a 5G connectivity case on-prem and hybrid because we track the things that move that is also operator collaboration from APJ and then this is including computer vision drones and autonomous vehicles. Now I know I'm almost out of time, I'll be quick. So this is something I did last night. So I used to work in trying to position mobile connectivity for private enterprises before. So every time we're talking about return on investment, what we would think is, okay, make sure that AGVs are involved because then you're going to get the ROI. So I try to think of how can we present CSPs with use cases that they can actually reuse the same stack and same written different verticals. So I made it, I will have shared it, but this is mainly I want you to think about repeatable edge kits that you can deploy and then you can repeat over verticals.

(25:47):
Now I know we talked about your existing offerings, how to make them better. If you already are investing in an industrial AI cloud, keep it but create an AI edge catalog on it and then create it as edge services. So we are working with a lot of customers that try to do GPU as a service. It's not taking up, it's better when you model it after a service and you actually charge per camera, per stream, per bed, whatever your use case is and run the hot path on the edge and then only signal back the summaries. If you already are using the public cloud, keep it but use it as an overflow, not a dependency because of what happened last week. And you can enforce the data residency and latency by policy so that it only does burst to the cloud when you need the capacity, right? And if you do have a multi-cloud, we're happy to orchestrate it. Both NVIDIA inference microservices works across containers, different clouds, same as we work on the orchestration of both from the cloud to the edge. So that should not be a problem. That's it. Thank you.

Guy Daniels, TelecomTV (27:05):
Selega, comea and take a seat. Thank you. Great. Absolutely fascinating. Thank you very much for that. And we do have time for a question from our audience. Anyone? Anyone on this side of the room yet? Yes. Who? Oh you are. There you go. Thanks Tony. Yeah,

Audience Member (27:23):
Thanks for the presentation. So if we look to AI and Agentic AI applications, so we have seen that most of them are running on device and small language model are running device or on-premise for the enterprise. Now and if we look to the hyperscaler, which are getting closer and closer to the end user, where do you see the business opportunity and good ROI for operator?

Selega Akiner, Dell Technologies (28:04):
I would say the biggest, your strongest point is still your secure connectivity and your mission critical applications, which a lot of enterprises need and not the public cloud providers can provide. And what is more, it increases stickiness. Once you're there with your connectivity, you can join that with edge use cases and then it's very hard for that enterprise customer to churn from you as well. So you basically get two birds in one stone. That is where I would say enterprise play is always going to be your strongest play.

Guy Daniels, TelecomTV (28:36):
Great. Thanks for the question. That's great. Thank you very much. Do we have another question at the front here? We do think we can slide one in

Francis Haysom, Appledore Research (28:44):
Selenga, I love your presentation. You mentioned a really key point that you can reduce the cost of the computing infrastructure. I think it was 75%. My question is we're obviously a telco audience and we're really keen on the edge. Have you had any thoughts as to what the compelling argument for what would term is telco edge versus an enterprise or solution? Just buying it themselves and implementing the solution themselves, which has been an ongoing question in my mind. Have you any views on that? What makes Telco Edge better than the enterprise doing it themselves anyway?

Selega Akiner, Dell Technologies (29:30):
I think other than the connectivity and the reliability piece, it all depends on the use case and how scalable that use case is. Usually if it is a multi-country operator or if there is an industrial spectrum available for the connectivity, it helps to go with the CSP and it's all about the application. Again, I talked about the ROI of the AGVs, et cetera. When we are doing the use case, the moment you have it having pay for itself for one side, then you can scale pretty quickly. And the edge is important because the customer pays for the deployment you do for them unless you're also using your existing network.

Guy Daniels, TelecomTV (30:18):
Great, thanks very much. If it's a quickie Dean. Yep.

Dean Bubley, Disruptive Analysis (30:23):
Just as a follow on from that about the private networks, A lot of locations will have existing IT infrastructure and on-prem servers and they put the private 5G controller next to it. I was at a port on Tuesday that did this. So are you likely to sort of add private 5G to existing AI compute or AI compute to private 5G?

Selega Akiner, Dell Technologies (30:51):
Currently how it's happening is AI is added onto the private 5G or AI? Is there standalone? I have not seen it. Applications onboarded. It's usually when you deploy on site, it is for new applications or there has been one use case where we have IT and OT onboarded. It's as also a retail use case in the mail. We can talk about that, but that's usually how it's happened so far.

Guy Daniels, TelecomTV (31:19):
Great, thanks very much. Yes. Microphone's coming to you. We must get another question in. We have to, it's a great opportunity.

Ron Insler, RAD DATA Communication (31:29):
I'm Ron from RAD DATA Communication. On the one hand, as you said, we have the AI running in the enterprise to save a lot of access to the cloud and also latency and things like that. But on the other hand, we see that the bandwidth from the enterprise, at least I'm talking about fixed networks, is actually tenfold growing from one gig, 10 gig, 100 gig. So how do we explain this discrepancy? Because on the one hand we need more bandwidth to go to the AI, to the SaaS or whatever, but then we put edge AI to reduce the bandwidth.

(32:10):
They live together

Selega Akiner, Dell Technologies (32:12):
That should counteract, right? They don't live together in the same way as you do, but it is based off the investment, right? So where you want to place the investment, it's always a trade off, like you're going to trade off the bandwidth versus the application performance. So I don't have an answer to that, it's just a case by case issue.

Guy Daniels, TelecomTV (32:36):
Great. Well Selenga, thank you very much for fielding those questions. We are going to take a break. We've had five presentations this morning. It's been quite a full session. So refreshments are available in the lobby area outside this room along with our exhibitors and their wonderful pods. So go and talk to the pod people and we'll see you back here in about 25 minutes. But for now, please a big round of applause for Timo and Selenga.