DeepSeek’s Ecosystem Implications, Samsung’s Galaxy S25 Debut & CSP AI Edge Readiness – The 5G Factor EP82

Good day everyone, and welcome to the 5G Factor. I'm Ron Westfall, research director here at the FU and Group, and today it's good one. I'm joined by my esteemed colleague, Olivier Blanchard, a fellow research director and practice lead for our AI devices practice here at the Futurum Group.

In fact, Olivier just recently returned from the Samsung unpacked event. And with that, I think we have some good insights on what's going on in the 5G and mobile ecosystem realms. And with that also we'll be focusing on just that, the things that have really jumped out of recent, you know, vintage, uh, that merit our attention.

And so with that, Olivier, welcome back to the 5G Factor. How have you bear been bearing up since the last time you've been on the 5G Factor? I can't even remember the last time I was on the 5G Factor.

I've been, I've been dodging your invites for a while, not because I wanted to. I've just been on the road it seems like a lot more in the last six months than I have ever. Um, so I've been, I've been doing well.

It's just, you know, it's, uh, it's hard to keep track and, and to keep pace with, uh, all the changes in the industry and the hurricanes and, and all of the, everything that's happening all at once. So, uh, yeah, it's good to have a minute to to be back and, and catch up with you. Yes, indeed.

And, you know, perfectly understandable and right. You are. It's been a very kinetic last, you know, well, year plus really overall.

And, uh, and I think our first topic that we're kicking off has played a major role. Yes. AI again, however, this time, this is something that I think is really galvanizing, you know, attention.

And hopefully by the time our recording is published, it'll still be somewhat relevant because matters are moving very quickly. And that is, uh, just on January 20th, uh, China based Deep Sea released R one. It's a reasoning model that basically is outperforming, you know, open AI's latest, uh, oh one model.

And that is, uh, verified in, you know, various third party tests. And so what is interesting here is apparently 150 researchers at a Chinese hedge fund, also known as Deep Syn, has out flanked the entire work of tens of thousands of engineers and basically almost the entire Western scientific community. And what is this, just with a handful of modified Nvidia, uh, H 800 GPUs?

Well, let's see. Let's stay tuned on that. Uh, more likely, at least, uh, from my perspective is that Deep Seq was trained on more than 50,000 H 100 GPUs.

And while that is good, it doesn't automatically mean that the AO AI ecosystem currently has reached a point that can train on materially less infrastructure. And that means we still, despite this breakthrough, we'll probably need to continue our massive commitment to figuring out, uh, ways to optimize AI training as well as AI inferencing. And that is a naturally links to at attaining the main objective of artificial general intelligence or a GI.

Now, a little more background, uh, there's a lot of threads here, but what I'm gonna focus on initially here to kick off the conversation is that according to Alexander Wang, the CEO of scale ai, he's the one that basically I think initiated the notion that, okay, what is going on here is that the Chinese have access Nvidia advanced GPUs on a wider scale than many people realize. And the reality is that the Chinese labs, that they have more H one hundreds than many people think, and he added and shared that understanding, is that deep seek has about 50,000 H one hundreds, and they can really, uh, uh, can't talk about it because it's against the export restrictions that the US has put in place. And as a result, they have many more chips than, uh, folks, uh, really, uh, I fully understand.

But that's changed, uh, obviously dramatically. And Elon Musk, uh, also seconded the motion. Now, Nvidia, on the other hand, has insisted that deep seat is excellent AI advancement and a purpose example of test time scaling.

So this is an important technique now that will, I think, get a lot more attention because of what, uh, deep seek has basically achieved. And, uh, in addition, uh, deep seeks work illustrates how new models can be created using this technique, leveraging widely available models and compute that is fully export control compliant, uh, emphasized nvidia. And so, uh, this means that ING requirements, uh, that require significant NVIDIA GPUs and high performance networking will still be needed into the foreseeable future.

And this aligns with what they see as that there are now three scaling laws, pre-training, post-training, and now test time scaling, uh, the newest one on the block. So in effect, uh, NVIDIA's adjusting that the deep Seeq used export compliant equivalents of an A MC gremlin and pacer and soup them up to the top line equivalents of Maseratis and Bugattis. I know it's a loose analogy, but it's like, wow, this is a pretty remarkable that if they only use export compliant GPUs to achieve this.

Well, I think there's a lot of counter appending out there. And, uh, as a result, there's this mystery, like what are is the GPU cluster that deepsea actually used? I don't think you're really gonna know for a little while, but this is definitely, I think, fueling all the speculation just how replicable is this?

Just how cost efficient is this really? But the outcome is like, okay, great. Now we have a AI training that could be done on a much more cost effective, much more energy efficient level then yes, and it's also open source space.

That's good news for the rest, the AI ecosystem. Uh, now officially, uh, Deepsea claims have used only about 2048 Nvidia H 800 ships, which, uh, to train the R one model. And that is alongside the, uh, 10,000 older generation, a 100 GPUs that had already attained, uh, before the US uh, imposed export controls.

Now, all this, uh, I think is going toward NVIDIA's credibility on US export policy. Now, naturally, NVIDIA has to claim that yes, that the, uh, chips that were used were export compliant. And that's understandable.

However, I think as we dig deeper, and if we take, you know, the opinion of, you know, folks like the CEO of signal ai, uh, then I think that means that there's more going on here that can be officially disclosed. Bottom line is that, you know, many of the actions that are backed by China's government, I think as we understand, you know, we've seen this with the Huawei for example, is that it constitutes what can be characterized as a giant psyop. And it's, in this case, China is attempting to create a little bit more uncertainty among the investment and tech community about AI's prospects in terms of it being, you know, primarily driven by, uh, US or Western technology knowhow.

And as you can see, it did effectively shift the spotlight from, you know, the half trillion dollar project star, uh, uh, uh, um, uh, uh, which has, uh, uh, own set of, uh, major, uh, questions. And so, uh, we have been, you know, really looking at, you know, how can we improve scaling laws and seeking efficiency with AI despite the deep seek breakthrough. And, you know, uh, again, we're been using open source, uh, rag fine tuning and forking capabilities, all these AI techniques to allow smaller models to be more performant.

And we've already seen companies such as IBM, meta minstrel and others have stepped up and really have made, uh, I think, an impact on how this can be better achieved. That is, you know, right size, large language models that are better performant. And, uh, with that, uh, Olivier, from your perspective, I know you've already, uh, provided some, I think, very valuable insight on this.

And what is your take on what's going on here? I know this is just one aspect here, but what else do you see going on here? Right.

Well, that, that, that was a lot. Um, I feel like the, the meme, you know, with a guy like this, and it's like he's got the, the board with all the red string just kind of connecting everything. Um, so I think that everything that you said is, is, is on points, uh, including the, uh, the, the, the theory or the hypothesis around Chinese ops.

Although I, I would caution that, um, if, if we equate the AI race between US and China to any other arms race or space race in the past, the, the, the, the prevailing sort of tactic to, um, harm the other side, your opponent isn't to necessarily undermine credibility, uh, of their model. It's to get them to spend a lot more resources chasing the same thing instead of trying to make them efficient. So by, by releasing deep seek, um, China, assuming that the Chinese government is involved in some way in this strategy, in the timing and in the, uh, the, the tone and tenor of this release, um, the, the, the objective for China should be to just let the US overspend on infrastructure and AI resources instead of taking this sort of more efficient scrappier approach that they took to achieve the same results with fewer resources.

So I, I, I take the, the notion that that China is trying to undermine US markets and US investments in AI with a grain of salt a little bit. Um, but having said that, I, I think, I think most people are actually missing the point with deeps six. I don't, I, I don't, I don't know that we're, we'll still be talking about deep seek in six months.

I think that the, the larger trend, which is something that we've been talking about for the better part of the last year, is that there's been an evolution. There's been a, a trend line in, uh, in AI training and AI inference that transcends deepsea, uh, and, and any other company like it that is very likely to pop up in the next, you know, six to 18 months. And it's this, the, we're we're the confluence of, uh, an increased and, and sort of systemic, um, improvements in the efficiency of the model on the training side, on the inference as well, but definitely on training, um, where, um, models that, that had to be trained in the cloud with massive resources behind them, massive inputs of power of chip and compute to, to train these models have become so much more efficient in the last year that what used to be only trainable in the cloud can now be trained on PCs and in some cases on mobile.

So we're seeing these, these models sort of shrink and get, get faster and cheaper to train, uh, just by virtue of the fact that they're, they're, they're becoming more efficient. On the other hand, we have this, this other element, which is the chips are getting more high performance as well, which is the whole value proposition between NVIDIA's Blackwell, right? You could use fewer GPUs to, to generate more outputs faster and more cheaply.

So there's, there might be a slightly more upfront cost on, on the GPU side, but you need fewer of them to achieve the same results and few less power, less, less, uh, water, all of these other resources, less footprints, uh, data center wide. So we have these two efficiencies basically working together to make AI training a lot cheaper and a lot faster. This was happening already.

It has gonna continue to happen. Uh, it's, it's one of the things that is fueling the proliferation of AI processors and AI models from the cloud to the edge, and creating this more hybrid ecosystem of basically sort of like cloud to endpoint, uh, AI training and, and inference, uh, which, which I think is the reality that we'll be existing in, in, in, uh, as, as early as later this year. Uh, and we'll touch on that again when we get back to our coverage of the, uh, Samsung event last week, because there's actually an element that that plays into this.

And so, um, the issue with, uh, with, um, stargates and the enormous investment numbers that we were talking about a week ago, and sort of like this injection of excitements in, um, in AI infrastructure specifically in the United States, seemed a little bit, um, I don't know, warts last week when the announcement was made, the, the impression that I had given what we just talked about and, and sort of validated by the deep seek announcement, uh, a few days ago. And the, the sell off, uh, of, of US tech stocks in, uh, in the last, you know, 72 hours, um, is this, um, this is already happening. And, um, it, it's the, we don't need to spend trillions of dollars building massive data centers that are going to house millions of chips and, and, and servers and racks and, and require nuclear power plants and, and some kind of, you know, weird re-engineering of our water supply and our water management systems.

We don't, we don't need to, to completely just stop everything that we're doing and build these massive projects, um, because the models and the chips are becoming more efficient and this federated distributed AI training and inference is already happening and the costs are getting lower. So I think that what we're seeing last week is, um, let's go back two weeks ago, two weeks ago at CES Jensen, um, introduced sort of like his vision for the next phase of what NVIDIA's about, and we're gonna talk about Nvidia, because NVIDIA has sort of like the, the market power and the best position in terms of, of GPU infrastructure, um, and, and IP when it comes to training, um, AI at scale. Um, his model at CES, the introduction of Blackwell, all the things that he wanted to do, what he talked about in terms of, uh, training for automotive, training for industries, doing all these, these virtual sort of, um, you know, digital doubles and digital twins of the world and, and different environments, it's still valid.

None, none of that has changed. Um, but, um, it, it's the 2025 spend on this. And the reality of, of where technology is today in 2025 with 2025 chips, when 25 models, the, the entire 2025 current layer of, of AI solutions and A IIP is absolutely not where we will be in 2028 and in 2030.

And I think that Stargates misses the point in that it, it looks like, um, the type of spend that we would put together if we assumed that these efficiency advancements, these efficiency improvements were to stop this year, right? Uh, so we're looking at bills for 2030 with budgets that reflects where we are today in 2025, not where we will be in 20 28, 20 30, 20 35. And so I think that they're, they're a little bit bloated.

I think it's, it's warped expectations. Um, I think the front end investments, uh, may be accurate. It, it may be the right number, uh, dollar wise, but I don't think it, it's, it's necessarily, I think it may be over-indexed on the data center side, is what I'm saying.

I think that, um, taking a more holistic approach to the entire ecosystem of, uh, of research of chips all up and down the, uh, the, the AI value chain from basically wearables all the way up to the data center is, is a more realistic approach. So I think that, that we're gonna end up spending a lot less on data centers and a lot more on production, on supply chain and on all of those middle layers to, to create a, a more sort of hybrid, uh, ubiquitous AI ecosystem. Um, and, and deep six is announcements and obviously the reaction to the market, which I think was an overreaction, but that's a whole other story, is just one of several proof points along the way that the increasing the rapidly increasing efficiency of, of AI training and AI infrareds workloads for that matter, uh, and the, the, the sort of like diminishing curve of costs associated with that is, uh, is going to be a disruptive force in some of the calculations that we've had about how many chips we're gonna need to be able to, to, to power this new, um, ai uh, AI economy, right?

Um, and that's not necessarily a bad thing for Nvidia and for a MD and for everybody else who's involved with this, but it, it, we should perhaps adjust these, uh, these massive numbers that we've been looking at and, and look at it more as a term in terms of, uh, diversification of chips. So if you're in Nvidia, for instance, don't just focus, or if you're looking at Nvidia for, uh, you know, in, in your investments, I'm not make giving any investment advice. Uh, so, so, um, don't take that for this, but if, if you're looking only at, at Nvidia for its data center GPUs, you're missing the point.

You need to look at Nvidia all up and down the value chain from, from PCs and devices and iot all the way up to, um, all the way up to the data center. And if you're looking at other companies like Qualcomm or Intel or a MD or, or, or media tech for instance, uh, you also need to look at those, those device layers and those intermediate layers where there's gonna be a lot of scale to deliver AI organically through this hybrid model, if that makes any sense. There's a lot Yeah.

And, uh, topic warrants that I think we're only gonna scratch the surface, but I think that, uh, our outstanding, uh, viewpoints there that you shared Olivier, uh, you know, first of all, you know, let's look at the big picture here. I agree this is not an existential threat to Nvidia by any stretch of imagination. And yes, their portfolio has diversified over the last few years.

It's no longer, you know, about, you know, only, uh, GPU, so to speak. Or at least, you know, there's that perception. It's, you know, the software and the services, it's definitely been diversified.

And yeah, I think, uh, one important takeaway from CES among many was, you know, what, uh, Nvidia is doing with Cosmos that is, you know, taking real world video and applying it to AI training and capabilities so that you have, you know, these outcomes that can be very useful. And, uh, just that real world settings, you know, how to optimize, uh, say what's going on in the warehouse as an example, and, you know, advancing robotics and so forth. And so, you know, a lot is going on here, but I, I think, uh, what is going on, uh, with a deep seek and, you know, uh, for that matter, uh, project Stargate, uh, I think, uh, that's important to note here is that it's like the throt thing of, you know, the AI hype cycle here.

And to your point, I think, you know, okay, what about the half trillion dollars that's been allocated toward, uh, project Stargate? Uh, what are the implications of, you know, a lot more efficient ai, uh, capabilities throughout the AI ecosystem? And this is a variation of Jevons paradox that is, if you make something a lot more efficient, that will actually increase more demand for it because it becomes more affordable and more broadly available.

And to your point, yes, when it comes to, uh, hybrid ai, okay, a lot of the heavy lifting, uh, AI trading that's done in its GPU clusters and data centers, you know, throughout, you know, say hyperscaler, uh, networks, uh, a lot of this, uh, can now, uh, uh, be done in a more distributed basis in your edge data data centers and so forth. But also, uh, to your point about devices, yes, you know, the ai, uh, interesting, let alone some of the training can now be done at the outer edge. And that includes, you know, devices that are powered by Qualcomm arm, uh, media tech and Apple, as well as, you know, across the, you know, the broader edge, you know, uh, including, you know, Broadcom, Marvell, Intel, a MD, you name it, that, okay, this doesn't necessarily mean you have to enlist Nvidia GPU, uh, GPUs to do this, uh, more distributed edge, uh, training and inferencing, but it's really, I guess, uh, just that, uh, welcome news for the entire a AI ecosystem as well as for the other, uh, chip players out there.

And, um, and on that note, let's, uh, now segue to the next topic, which is really AI at, on the device level. And as we saw at Samsung unpacked, which you were able to join, is that Samsung announced the, uh, galaxy S 25 Ultra. It's, uh, galaxy S 25 plus and Galaxy S 25, all aimed at setting a new standard for true AI companionship with, uh, you know, basically our, you know, context to where, uh, mobile experiences.

And I think that's a, a very poignant way to, you know, position this and market this at the onset. And what we saw is that Samsung introduced multimodal AI agents, and as a result, the Galaxy S 25 series is really a first step in Samsung's vision to change the way users interact, uh, with their phone, and also, you know, with the, the real world for that matter. And so it's using the, uh, customized Snapdragon eight elite mobile platform to attain a lot of these, uh, uh, goals.

And also the Galaxy chip set is, uh, designed to deliver, you know, greater on device processing power for Galaxy AI and, you know, improved camera range and so forth, and let alone galaxy's, uh, you know, uh, pro visual engine. So I'm gonna stop there because you were there. Uh, you have, I know in depth, uh, uh, perspective on this.

So what were some of your key takeaways from, you know, what Samsung announced with the new Galaxy S 25 Y? Right. So it was, it was a really interesting event of, of all the, uh, of all the Samsung impact events that, that I could have gone to this, this was probably, uh, one of the more significance, and I think, um, somehow a lot of, of that significance got missed by the coverage of the event.

So let, let me give you my, my take on this and why, why I think it's so important. I've, I've been operating under the, the assumption, uh, as an analyst and as a, a practitioner of, you know, where AI is going and how, uh, I've been operating under the assumption that ultimately what we want is ubiquitous AI, device agnostic ai, where you walk into a room, it doesn't matter where you are, um, the, the devices that is nearest you and most capable of delivering the agentic AI experience that you're, you're, as a user, you're expecting, uh, and doing it more efficiently is going to be the one to deliver it. So if I'm talking to an assistant and prompting it and saying, uh, I wonder what the weather is today, or, tell me, you know, what are my appointments this morning?

Uh, it doesn't matter if it's my watch, my, my headphones, my smart glasses, my pc, my smart speaker, whatever it is, that's where it's gonna go. And with, with Agen AI specifically, one of the really interesting user experience advantages or value propositions is this hyper personalization where your agents learn from you. They learn from your habits, they learn from your needs, from your patterns.

They become sort of like your best friend. They, they know what you want and in what, in what format, and what style and what speed before you know it. But definitely when you prompt it so they can anticipate your needs and, and adjust your environment, your workspace, your calendar, your, your shopping planning, all of it, uh, for you.

Um, however, what that requires is a lot of device to cloud integration. And, and I'm not gonna get into the, the specifics of, you know, how that needs to work from an architectural standpoint and an orchestration standpoint, but it's extremely complicated, and it requires some data to be on device to be cloned in the cloud and to be accessible both in the cloud and on-device, uh, so that there's no, there's no lag between, you know, the prompt and the response. So you can have natural, uh, naturally times natural language conversations with, with your agent and your ai, uh, assist it, uh, without having to wait for a response or like, Hey, I'm thinking about it, gimme a second.

Um, what Samsung did though is something very different from that, which is super, super interesting. And at first I thought, okay, this is the wrong approach. And then the more I thought about it, the more I realized, wait a minute, this is actually smart.

What Samsung did is they prioritized personalization and data security and data integrity and, and, and safety, uh, by essentially moving a lot of those processes, the training and the inference to the device itself, and blocking it from the cloud. And so essentially what they did is they have two parallel systems. They have, uh, uh, because they're Android devices, there's, there's Gemini, right, which is kind of like the Android device to cloud, uh, AI platform that remains untouched.

So anything that you do with search, it's gonna go out for Gemini, and, and you have this, this normal integration, but all of the super personal stuff that it's learning about you, um, essentially sort of like the personality graph that's, uh, that an agentic AI is, it needs to be able to build and train itself on in real time based on your needs is, is super personal. And, uh, and Samsung made the decision to keep all of that on the device secure, no contact with the cloud, nobody else is gonna train on your data, nobody's collecting your data. Google is not collect, collecting that private data.

It's on, on your device. Not only that, but it's protected by their NOx security solution and it's post quantum encryption, which means that currently, at least in, in theory, I haven't validated this, uh, but, but based on on what Samsung is telling us, um, it's, it's not decryptable, uh, with quantum computers, which is, which is really nice. So pros and cons of that.

Um, pros, obviously data security, privates, uh, all of your data remains private. And, and you have this, this, I think added sense of, of, of privacy that you nor wouldn't normally have with a lot of AI products and, and solutions. Um, and it's tied directly to Samsung.

The, the, the con is that if you lose your device, you have to start all over again with all of that, uh, the agent training. Um, but device to device, when you upgrade from an S 25 to the S 26 or whatever the next generation is, you'll be able to do that. You just can't like upload your, uh, that private data to the cloud.

Um, but all of this is done because the, uh, the chip that they're using, the SOC, the system on Chip is, uh, is capable of doing that. And in this case it happens to be Snap or Qualcomm's Snapdragon eight, uh, elites, which is the, the sort of like flagship platform that, that Qualcomm introduced, uh, at their summits back in October, um, of last year. So it's, it's the latest and greatest, but it is a custom chip made specifically for Samsung.

So it's, it's actually the Snapdragon a eight elites for Galaxy, uh, which has a few additional bells and whistles for, you know, some camera, uh, improvements and also for this, uh, this enhanced, um, uh, agent AI and device capability. But it's sort of illustrates, I think the, um, the, the, the new paradigm of this distributed, uh, agentic AI where sure, you can do, you can train a lot of models and, and do a lot of things in the cloud, and there are things that work best in the cloud, and that should be operating that way just as a cloud service. But there is also, oh, alright, I'm gonna continue it, it pause for a second.

Um, that'll be a nice place to cut. But there's also a, um, uh, a, a, a really huge leap forward in capabilities of these, these ARM-based chips that are on mobile phones, that are on PCs that increasingly are showing up in, in smart watches as well, uh, in smart glasses, uh, in, in essentially every digital product that, that we touch. Um, there's more and more capable of, of doing this.

And, you know, the, the size of the, the models that you can train and, uh, and, and do inference workloads with on a PC versus mobile versus smart glasses depends a little bit, first of all, on, on the, the system on chip itself, but also on the size and the form factor, right? You can put a lot more processing power in a PC than you can in a phone, and you can put a lot more processing power in a phone than you can on smart glasses or on a smart watch. Um, but you also have to think about how all these devices work together and how they can pool their processing resources so that instead of processing something on the watch, you're processing some of it on the watch, some of it on the phone, some of it on the pc.

Uh, and a platform like Snapdragon, which is in all of these devices, might be able to just kind of work altogether more efficiently than cross platform, uh, combinations to give a user, um, uh, an enhanced AI experience that is primarily on device or that can sort of separate, um, the needs of, you know, pushing some of the, uh, the workloads to cloud services and then keeping some of them private for a variety of reasons for speed and efficiency, for power efficiency, for cost efficiency, but also for privacy. And so it has implications for, uh, consumers, right? You and me just wanting to keep the private, uh, and cheap and not having to pay for all of this cloud inference stuff.

Um, but also on the commercial side for businesses, because now the more data they can, they can have in-house, the more secure their data might be, um, and the more processing they can do in-house, also, the lower the costs, uh, of, uh, I mean, they don't necessarily have to pay for as many, uh, instances of, of training or workloads that they're pushing out to a cloud through cloud service. They can keep a lot of that stuff in-house. Yep.

Yeah, I, I think it's good news for say, the health monitoring use case, this, and it aligns. And that's, I think, one constant theme. Uh, when we saw, you know, with the launch of Project Stargate, Larry Ellison step up and say, Hey, this AI investment has warranted because of advances that could be made in medical research or, you know, tracking medical records and so forth with that privacy respected, built in and so forth.

And likewise, you know, just being able to have a device that somebody can have confidence, it's gonna maintain my privacy, but it's also can just do that, you know, detect issues, uh, before, you know, they get outta control or, you know, uh, just across the board any health, you know, benefits, uh, out there. So that I think is, is certainly the good news. That was one of the, the, the main use cases that, um, I'm not even sure that it was, you know, I think it might have been under indexed a little bit at the announcement itself, but in the pre briefings, uh, we spent a lot of time on, um, on that particular use case, how, uh, Samsung specifically.

But I think as a concept, having all of that data and that agentic ai, um, processing and analyzing and recommendation engine on the device as opposed to in the cloud, first of all is more immediate. But also, um, I think if, if, and I think especially with medical issues, um, I think people want their data to be protected. And we've all been burned so many times with hacks and, and, you know, our, our private data, especially medical data, getting out into the wrong hands, just getting out in the public this, this takes care of that.

And so you have a, a personal assistant that is able to, through the use of different devices and sensors, whether it's a, a Galaxy watch or a ring or, or other sensors, um, e essentially guide you and give you feedback on how well you're sleeping, for instance, your eating cycles, what your caloric intake is. Uh, we were talking about things also about a more granular approach to your diet, which you can sort of enter manually, but also, um, the AI knows what you're eating, and so it can sort of extrapolate the types of nutrients that you're getting and not getting. So it's, it's creating and painting a, a, a much more, uh, again, granular and, and detailed and complex picture of your overall health.

And it is intelligent enough to understand your cycles, understand your patterns, understand cause and effect of good and bad behaviors, make recommendations, um, monitor your health. And it's, it's, it's amazing because it's all on the device and it's all secured. And, um, I love this, it, it, it moves some of the, the, I think the, the real true immediate benefits of a gentech AI out of the cloud in areas where it doesn't need to be in the cloud.

Um, so some things need to be in the cloud, some things shouldn't be, and it's, we, we now have the capability of, of parsing that, uh, according to our needs and according to our preferences. And that's gonna, that's gonna change that, that changes the equation a little bit for, uh, for how we think about AI investments. And again, it, it speaks to that diversification of, you know, focusing on, on cloud AI training and inference, but also focusing on this, these edge use cases that are becoming much more prevalent and with a, a huge potential for, uh, for adoption.

And the, the footprint. If you look at the mobile industry, if you look at the PC industry, um, even if, if those numbers, uh, essentially the install base doesn't really grow that much, there's a refresh cycle here, uh, that's gonna bring a lot of these AI chips to this, this broad install base. So the market doesn't actually have to grow to show results.

It's the refresh cycle that's gonna show those results and, and push a lot of these, uh, uh, these AI chip numbers out. Um, and that's what I'm hopeful about. Yeah, and I think, uh, that's a great segue for who else can, uh, benefit from, you know, the, uh, the potential large s of the AI ecosystem, uh, overall, uh, growth.

And this ties directly to, well, you know, 5G service providers, uh, certainly, you know, uh, the folks who provide mobile services to, uh, consumers and, uh, businesses, uh, and specifically here in the us. And what, uh, I think is interesting is that we're seeing, you know, major US operators such as at and t and Verizon looking into, okay, how can we play a more integral role and monetize, you know, AI in terms of, you know, uh, being closely linked to the services they provide. It's not, uh, exclusively mobile, but it certainly includes, uh, for example, their fiber services, business services.

But I think what's interesting here is that they have the real estate that we touched on, Olivier, how do we push more AI capabilities, training and inferencing closer, uh, to, you know, where the customer is, you know, bringing the AI to where the data is is certainly the mantra that, uh, we've have heard a great deal about. And so what's interesting is that now we see, uh, that Verizon business has, uh, revealed basically a, a bundle of products that are designed, uh, for enterprises and also cloud providers as well as hyperscalers to, again, deploy those AI workloads at scale by, you know, basically offering a single platform, uh, that is designed to do just that. And what it is, it's called Verizon AI Connect, and it's offering a blend of really the operator's fiber infrastructure along with its power space and cooling capabilities, uh, and with those resources using it, uh, to deliver it, but also it's backed by its virtualized 5G programmable network to really make the AI workloads, um, more, I, I guess you'd say, customized to the, uh, customer needs, uh, out there.

And so what it's interesting is that already, uh, Google Cloud and Meta have already onboarded meta platforms specifically, and I think that's important because, uh, these are, you know, going to be logically the early adopters of AI infrastructure, uh, platforms such as this that are offered by a, a major, uh, service provider. And, uh, not to mention Verizon and Google are also looking at ways to, you know, advance AI services for, you know, things like network maintenance as well as anomaly detection. So this is in play, uh, this is something that I think will kind of help the, uh, service providers.

Okay. Finally, uh, we have a way to use our real estate, uh, to take advantage of, you know, ai. And this is not going to be necessarily a repeat of things like, you know, mobile edge computing and so forth.

There's still that risk, you know, know the operators still not, not be able to figure out how can we be integral to this. But I think, uh, this is demonstrating that, uh, they're all to a decent start. And not to be outdone at TT recently secured, uh, $850 million from the sale and lease back of its underused, uh, co facilities, uh, from, uh, a property company, uh, capital rain as part of its copper retirement plan.

And so, uh, this deal closed in January, it involves the asset transfer of 74 properties across the US spanning more than 13 million square feet. So the bottom line here is that the operators are becoming, uh, smarter about how they can take advantage of the real estate assets that they do have today, you know, that is retiring co assets as well as other edge infrastructure and, you know, working with, you know, the, uh, the major, uh, AI players out there, that's IE the hyperscalers to, you know, basically come up with mutually beneficial ways to monetize AI in the near future and certainly longer term. And, you know, Libby, what are your thoughts on this?

Do you see the service providers really being able to step up and actually, you know, playing a meaningful role in this? Or is this something that, okay, it's another missed opportunity and the service providers will be reduced to kind of a commodity like provider of the infrastructure for the AI services that are running across, you know, the, the clouds out there, Right? Maybe a little bit of both.

Um, so yeah, no, I think, I think it's smart for them to do it. So, uh, you know, the, the situation that we're in is, uh, we're expecting, obviously if we're spending, we're looking to spend half a trillion dollars on building data center, um, infrastructure. It means that we're, we're looking for a lot more computes to come from somewhere.

Uh, a a lot of these, you know, you know, builds are, are years down the road. We, we can't really start yet. So, um, the expectation is that we're gonna need a lot more compute power very quickly.

Where can we find it? And if you have data centers already out there that are underutilized, right? Um, or that can be sort of, you know, a lot of that process and think power can be retested for, uh, you know, higher price or, or more premium services, um, it makes sense for a business to go for that, for that ROI, right, for that opportunity.

So one, if, if a lot of those data center resources are underutilized or not used at all, suddenly they can be, you know, assigned to this, if we can charge a little extra for this because AI is more valuable than, you know, being in a 5G network, um, maybe there's, there's value in that as well. I think though, that, that you're right, ultimately there's, whether it's successful or not, that that is gonna depend on them and how they package it. And, and it, there are a lot of variables there.

I think at some point, those data centers and those resources age out, they're no longer, um, the most efficient, the most cost efficient, uh, and, and maybe they just get retired or the Verizons and at and ts of the world upgrade their systems specifically for those AI workloads. And now we start seeing a, a, a transition of spend where they also become, you know, they're, they're buying the black well powered, uh, racks and, and, and they're becoming more AI focused, and they become part of this sort of like, uh, ecosystem of, uh, of, of AI workload, uh, training and inference services, and it's all interwoven. Maybe that's, that's possible, but for right now, at least for the next few years, while we wait for all of these, these massive data center builds, um, they're there with capacity.

And so absolutely they should go after that market and, and see what they can make of it. Um, worst case scenario, it doesn't work at all. Middle best case scenario, uh, they make some money, it becomes commoditized, and eventually it just kind of dies out because, uh, they're outperformed by other outfits.

And best case scenario, they'd build a whole new business model that's, uh, that's gonna be really lucrative for them. So it's worse the shot. Yes.

Yeah. And I think, uh, this is again, you know, AI and it's open-ended possibilities. Uh, there's just a myriad of variables here.

So on the one hand, it's good news for those sa uh, service providers. Certainly a test time scaling, uh, is, is introduced to the possibility of, okay, AI as we do it could definitely become more commoditized, as you pointed out. And thus, you know, that's good news for the ser uh, service providers if this become a lot less expensive to, you know, do the, uh, AI infrastructure hosting and so forth.

Uh, now they, however, is, uh, will this result in monetize outcomes for them? And so, yeah, that is again, that kind of, uh, one of the many, uh, uh, I would say the three doors that you presented, uh, possibilities. And it's hard to bet right now on the operators because the entire a i ecosystem is basically going through a lot of flux as we speak.

And so, uh, we'll, uh, we'll come back to this. We'll talk more about it. Yeah, we should, we should come back to this six months from now and see Definitely, or let alone you say after Mobile World Congress, uh, you know, a lot could changes from that.

Um, so this has been great. Uh, thank you so much Olivier, for coming on board and again, appreciate, uh, the, uh, opportunity for you to share your thoughts. Yeah, thanks for, thanks for having me on.

And I am not going to Mobile World Congress as of now, this year. I'm skipping it. I'll probably go next year.

Um, but, uh, but I'll be, I'll be looking forward to, uh, to the announcements surrounding MWC. 'cause I'm, I'm sure this, this very topic is gonna be, uh, is gonna be one of the major, uh, themes of, uh, of the trade show. Yeah, Yeah.

That we can bet on. That's something I think we kind all agree on. A I might as well call it AI World Congress for the time being.

But, uh, and, uh, well, great. Yes, uh, I know, uh, we'll certainly be sharing, uh, thoughts on Mobile World Congress and, uh, and beyond that, uh, thank you everyone for joining the 5G Factor. Again, you can bookmark us on the future and group, uh, website as well as we can be, uh, viewed on Tech strong, uh, tv.

And, uh, we certainly, uh, again, appreciate, uh, taking the time to listen to our thoughts. And with that, everybody have a great AI and test time scaling as let alone 5G day. Again, thank you all.