The AI bubble and new frontiers in tech

Welcome to Beyond the Benchmark, the EFG podcast with Moz Afzal.

Moz Afzal:
Hi, everyone. So we're going to do actually quite a controversial topic, this podcast on artificial intelligence and everything tech related to artificial intelligence. And I'm very happy to have Dominic Jacobson on our podcast today. Dominic, welcome.

Dominic Jacobson:
Thanks for having me, Moz.

Moz Afzal:
So Dominic is a senior equity analyst in our US growth equity franchise and based in the US. And Dominic, maybe give us a, given this is the first time you're on the podcast, give us a little bit of insight into yourself and how on earth did you land up EFG and in the US?

Dominic Jacobson:
Yeah, no, sure. Happy to. So I guess I'll start off by saying I'm an Aussie. You probably couldn't tell because of my accent, mainly because I grew up in Hong Kong. It was a bit of a melting pot, just around a bunch of Americans, Chinese, British and Indian. So I have a bit of a mixed up accent, but I've lived across seven different countries. And I guess my career path is non-traditional in some ways. I went to university to study chemistry, mainly because I tended to be a bit better at science and maths at school. But when I was doing my master's in Imperial College on route to doing my PhD, I started a business with my friend Parlia because we were becoming less and less enthusiastic about the idea of not making money for another five years. And long story short, it was more of a concept as opposed to a business, to be honest. I was making a lot of money. But we won a couple of competitions, including the mayor of London's Low Carbon Entrepreneur Award. And I think that experience really gave me the bug in terms of wanting to be at the interface of business and science. And I came to the conclusion that being an investment analyst focused on tech was a way that I could kind of really explore both of those interests at the same time. And I was really fortunate to get my start under Louis Garve, who's one of the best macro investing minds out there. So I started off there at Gavekal around about 12 years ago. During my time in Hong Kong, I met my now wife. And as things go, she wanted to be back in New York. So that's what happened. And I took a job at VanEck, an asset manager based out of New York, global presence. And I was a tech analyst there for about four or five years. Also had a short stint of BlackRock. And then now I'm at EFG with the US team, which I'm really enjoying my time with.

Moz Afzal:
Them. Well, great. So let's go into the topic at hand. And I think probably the most controversial question that certainly I get on the road at the moment is, is AI in a bubble?

Dominic Jacobson:
Yeah. Look, I don't think we're in a full-blown bubble. I think maybe we could be at the foothills of one, and I respect there are kind of various definitions of what constitutes a bubble. I think from our perspective, we're not there yet. So let's take a look at a couple of things. The first is that price action in the markets over the last couple of months, specifically some of the AI darling stocks. The next thing is fundamentals and what are management teams seeing. And I think maybe bringing a little bit back to our process here at EFG, part of what we do is that we're fortunate enough to get in front of management teams or portfolio companies on a regular basis. And about a month ago, we spoke to a lot of our tech holdings and we literally asked them in a variety of different ways, "Are we in a bubble?" And the answer was almost unanimous where we're not in a bubble according to these guys.

And they're saying either we can't get enough compute, we can't make enough of it, or we don't have enough power to light up these data centres. So walking away from these meetings, it really seems to be boiling down to two gating factors. The first is that there simply isn't enough compute. And the second is that there isn't enough power available to light up these data centres. So these two things like act as natural governance and prevent a full bubble from materialising from now. And I guess just one last thing probably worth touching on from taking it from a different angle, maybe a historical perspective, people like to draw parallels to the AI or the dot-com bubble and the internet bubble of the '90s and how GPUs are analogous to fibre. You look at the peak of the internet infrastructure build out, don't quote me on this exact number, but something like 98% of the fibre footed to the ground was dark. It wasn't being used yet. Literally no company we're speaking to at the moment are talking about dark GPUs. In fact, in many cases they're being used so much that they're melting.

Moz Afzal:
Yeah, that certainly is a good analogy to draw against, I guess, the internet bubble history. And there's certainly plenty of evidence of that. I think your earlier point is, I think it's a very rational one. I think the market is rational. When something gets too hyped up in how much they were going to spend over the next 10 years in the trillions of dollars, and then you look at how much their revenues are and you think, "Well, this is just not feasible in terms of how much they're going to spend." The markets get very rational and drive those stocks down. And I think that is actually a very, very good point. We don't see that behaviour permeating across all sectors. And yes, there are many points that happen and maybe on a day or on a two-day basis, but what you find within the next month or two months later that over hype has just been wiped out.

Now, I guess one area where there is certainly a lot of attention is, and again, you alluded to with GPUs and semiconductors, tell us what's going on. History tells us that the leader in the pack usually makes 90% plus of all the economic profit that's made out of any sort of hype or any product cycle. What's going on in the kind of semiconductor space? And obviously there's a bit of a competition developing between, if you like, Google's TPUs, which has obviously been a great win for them over the course of the last six months versus NVIDIA's Blackwell chips and now Ruben coming up and how the competitive forces between those two are interacting.

Dominic Jacobson:
Yeah, no, that's a great question. I think the debate around GPUs and ASIC is probably one of the ... I guess I like to think it as one of the most important battlegrounds in tech today and has kind of implications for the broader market just because of how much market cap is tied up in that kind of complex of stocks. But maybe just to set the stage very quickly for the listeners who are a bit less familiar with the space, GPUs tend to be general purpose. They have the ability to tackle a whole wide range of use cases, but tend to be quite expensive. ASICs on the other hand stands for application specific integrated circuits. These are basically custom built chips for very specific use cases. So not as flexible as a GPU, but they can handle a narrow range of workloads, but tend to be quite a lot cheaper.

So kind of bringing it back to your question, I think for a while it looked as though NVIDIA was going to be this runaway freight train when it came to advanced compute for AI data centres. It wasn't clear who was going to be number two or a legitimate competitor. But this, I think this year we kind of got an emphatic answer to that. And it turns out Google with their TPU or their own ASIC chip that they developed in- house in partnership with Broadcom is kind of the big thread in the room for them. And that became clear when Gemini three was released and they announced it was exclusively trained on TPUs. They didn't even use NVIDIA GPUs. So Google has really emerged with this 800 pound gorilla in the AI room as it were. And that's in, which mind you is in stock contrast to how the stock was perceived at the beginning of 2025.

And with their TPU and Gemini-3, they've crucially become the lowest cost producer of tokens, which is important because it's only becoming more expensive to run these models. So if Google's able to run their models on thin or negative margins, they're putting an enormous amount of economic pressure on companies that have to buy GPUs and pay that quote unquote Nvidia tax, particularly if they're relying on external capital to fund their loss making businesses so that ChatGPT kind of comes to mind there. But I guess the big question is like, how does this play out going forward? And I think there's a decent chance that we see a bit of mean reversion in the trade that took place at the end of 2025, specifically companies that were tied to the NVIDIA GPU supply chain underperformed the Google TPU supply chain. I think that trade reverts. And the reason for that is that NVIDIA has rolled out that Blackwell chip next generation of GPU has been deployed.

And we're going to begin to see the first models trained on these clusters. And if scaling laws remain intact, which I think they should, these models will be noticeably better than Gemini-3 and perhaps maybe an oversimplistic way to look at it is that these models trained on Blackwell will become the lowest cost producers of tokens.

Moz Afzal:
And I think that's, I mean, it's going to be quite fascinating. And I think maybe many of the listeners and certainly some of our investors probably haven't quite understood this sort of push and pull dynamic that is developing between sort of GPUs and ASICs and then of course ChatGPT and Google. I guess the other dimension is, and something you can alluded to was power. And then when you think about China in this regard, where China obviously has much better power dynamics in terms of solar, nuclear, and obviously a lot of investment over the next 10 years to make sure that they are very well endowed respect to cheap energy. How does that look relative, but they're behind on technology? Maybe take us through the dynamics of that sort of power to token ratio and how does that play out?

Dominic Jacobson:
Yeah, no, that's a great question. So I think it's funny, I think Dan Wang has some really interesting thoughts here. So he's an ex- colleague of mine. He wrote this really good, I think it's bestseller now called Breakneck. And it's about China's rise as a global superpower and what the future holds. And a real thought leader in the space, and he talks about the geopolitical dynamic between China and the US. It's kind of a battle. He calls it between a loyally society in the US and the engineering state in China. And in the US, our political elite tend to have backgrounds in law and our economy tends to reward things like IP design and software. Whereas China, you look at the CCP and the Politburo, they tend to have backgrounds, engineers and scientists, and they tend to prioritise manufacturing and infrastructure. And this kind of feeds down into these dynamics that we're seeing in AI.

So the US dominates the choke points in the semiconductor supply chain and things like EDA software and semiconductor equipment, which makes it kind of close to impossible for China to compete at the bleeding edge because the US limits through litigation, their access to mission critical, let's call it components or products. But when it comes to power, China seems to have a very distinct advantage, specifically when it comes to speed, because the way Dan Wang puts is, China is run by engineers. They prioritise output over procedural correctness. So for example, right now, China's currently building 30 nuclear reactors. The US is virtually building zero because of red tape and regulations. China is adding more renewable energy in one year than the US has done in its entire history. So they kind of have what each other need. It's just a matter of if the US can either get out of its own way when it comes to red tape and regulations or whether China is able to close the innovation gap. And I think it's going to be fascinating to see it play out, but I think if we're looking at the next one to two years with Blackwell and Verubin coming out, I think the performance gap between the US and China is going to widen in favour of the US, at least over the next year or two.

Moz Afzal:
No, I think that's a very, very good point. And also, I guess the key thing is specifically to NVIDIA technology and the new GPUs is that they use a lot less power, I guess, power per token than the current versions, as well as I guess the Chinese homegrown versions.

Dominic Jacobson:
Yeah, no, that's correct. So GPUs ... Yeah, so the way that I would view it is that like the Chinese tech, like the infrastructure layer relies on semiconductors that are a few generations behind that of which we have in the US. And because of that, they tend to be less efficient with the energy and the electricity that they can use. So the cost per token is much higher, but the advantage that the Chinese have is that they have an abundance of power, kind of like what I was talking about before. They have a wealth of energy through a variety of different sources, whether it's renewables, increasingly nuclear, and a variety of different others. So the cost per token is less of a barrier to them than it is to the US. So I think that that means that the US, despite having a significant tech advantage when it comes to like leading edge GPUs, et cetera, it's not going to be this runaway freight train if you want to call it that, because China can kind of subsidise it with excess cheap power, if that kind of makes sense.

Moz Afzal:
Yeah, no, exactly. I think it's exactly the point. I think where today's relative difference is, is that one has a technology ed, the other has the power edge and I think somewhere in between is ... If they could combine together, they'll be very, very powerful, let's put it that way, in terms of technology stack as well as the power stack. Now, I guess the other sort of key development in this sort of geopolitical AI war is, if you can call it that, is open source versus closed source language models. Is there any value in models if one has an open source and one has a closed source?

Dominic Jacobson:
Yeah. So I mean, I guess ultimately you're asking who wins. And I think the way that I think about it is that both close and open source models will both continue to improve over time. I think there's room for them to coexist in many ways, but ultimately closed source is going to keep their lead over time, just because the barriers to entry around these frontier labs are becoming much, much higher, specifically in the form of capital, data, talent, and distributions. It's effectively turned into a fortified player oligopoly. We're moved away from a world where you could train a model for $10 million, which sounds like a lot, but it really isn't. Now you need somewhere in the range of 20 to $100 billion cluster to train a next frontier model. So these closed source frontier labs will always be at least a generation ahead because they're the only companies that can really afford the chips and electricity to run them.

I also think that context, I guess you kind of think that of that as data in some respects and vertical integration of some of these companies create these really high switching costs. So once you're on a closed model, you're unlikely to leave. Take Gemini, for example. If you've been using it for a while, it has a set of unique data points about you that an open source model doesn't. And then on top of that, Google can feed more personalised data about you into the model from things like your Gmail, your YouTube, your maps. So the more you use it, the better it gets, the harder it is to leave. And because it knows what you're interested in and how you like things being done. So I think as we begin to move more and more into an agentic world, that level of personalization is going to be another kind of like sticking point as to why these closed source models are probably going to maintain their leads.

Moz Afzal:
Yeah, no, absolutely. I think that as we move to now, let's maybe pivot a little bit more to Agentic AI and agents. I was listening to a podcast the other day and there was two technology, I guess, technology companies talking about the fact that in their now meetings, there are more agents than there are humans in the meeting, which I've found quite amusing that we've come to a world where you can have a meeting with 10 people, four are actually humans and six are actually agents listening into the meeting and all doing their own summary versions of it, but that's kind of the world we're moving towards, I suspect. But let's talk a little bit about that kind of agentic and I guess how that now fits into the software world where clearly we're starting to see software companies bifurcate between kind of winners and losers as some take on agents and some try to hang on for dear life for their high margins that they've been so used to over the last decade.

Dominic Jacobson:
Yeah, no, you're right. I think SaaS or just software generally has been in a rock and a hard place over the last year or probably even more now. And that's definitely been reflected in the share prices in many instances, at least in public markets. And I think that the reason is that there's this perceived AI threat, specifically that AI will be able to replicate many of these SaaS models and mass produce them at low cost, kind of effectively leading to the commoditization of these types of products and services. Or as I guess Sam Altman put it last summer, we're entering the fast fashion era of SaaS. I guess I partially agree with that sentiment. I think our view is a little bit more nuanced. For example, I see infrastructure software companies doing quite well as the world transitions to AI. I think a lot of the companies that we speak to, whether it be in tech or outside of tech in our portfolio, talk about all this unique and proprietary data that they have that would be really high value inputs to LLMs, et cetera.

But the issue is that the way that they've stored it and the way they query the data just makes it difficult to implement, at least in the real world. So some of these infrastructure software companies help their customers prepare their data real estate or estates, as it were, and leverage it in a way that they can actually implement it in the real world and into LLMs. Another fear is that AI code generation is going to lead to this kind of mass extinction events of software developers. And admittedly, I might be fairly ... I have an out of consensus view maybe here, and maybe that's been driven by some of our portfolio companies that we've been interacting with, but I think demand for developers isn't going to go away. And in some unique cases is actually going up. Although AI code generation tools are writing a tonne of code, writing code is not the whole job of being a developer.

Sometimes it's like 20% of what they do. And the other 80% of the time is managing all the code that's being produced, whether it's checking it, implementing, making sure it's compliant, et cetera, et cetera. So in some instances, you're going to need more agentic software tools that help software developers manage all of this code. And I think in a number of companies, this thesis isn't really fully appreciated, or at least the scenario isn't discounted into the share prices appropriately. I think the last point where I may agree with the negative sentiment is that I think you're going to see some real disruption amongst application software companies that don't have proprietary data. And I think you're beginning to see that play out in some of the bigger names in the stock market. And so if your software is offering purely a system of record and isn't paired with some form of proprietary data, I think you're at a decent chance of being disrupted.

Moz Afzal:
Yeah. And again, I think we're certainly starting to see that play out. So let's move on to, I guess, a couple of other topics. And I guess we're moving much more into the physical AI frontier, which are basically another word for robotics and autonomous driving. Maybe give us your thoughts on that. We're getting very excited here because we're starting to see some of the early signs of autonomous driving coming here in London and we're starting to see the first set of Waymos driving around, although they do have drivers in them at the moment, but I suspect they will be coming probably sometime in 26 to London. Maybe talk us through both autonomous first maybe and where you think that's running to. And then maybe I guess where it is more, where Asia in the frontier is probably on the robotic side.

Dominic Jacobson:
Yeah, no, definitely. So yeah, I mean, we're seeing the same thing here in the US. You go to San Francisco, San Francisco or Arizona, there are Waymos everywhere. You can't go down the street without seeing them. I think when it comes to autonomous driving, you're right. I think we're kind of potentially at an inflexion point maybe in 2026, maybe in 2027, because you have this confluence of positive forces. I think thanks to advanced compute and neural networks, you're getting this exponential improvement in driving performance, which is then leading to more and more autonomous vehicles on the road. And I think people are beginning to see that the unit economics are starting to make more and more sense. And I think that's one of the big reasons. I think Waymo actually raised a bunch of money not too recently. And then finally, you have this regulatory streamlining.

Sorry, these three factors are coming together, and I think we're going to see an acceleration of deployment of autonomous driving, at least here across the US, and likely in China, but I'm a little bit less familiar with that geography. I think robotics on the other hand ... I'll start off by saying this. I think it is the biggest opportunity out there at the moment. It's like tens of trillions of dollars opportunity. It's maybe the biggest opportunity across any industry right now, but I caveat by saying that there's still some pretty meaningful technical challenges that this technology is facing. I'm probably oversimplifying a very complex topic here, but just for simplicity's sake, cars tend to operate in highly semantic environments. There are lanes, stop signs, traffic laws, and that kind of leads to a narrower range of behaviours that the AI has to predict. I think we're a little bit further away when it comes to having a robot in your home.

The home is a highly unstructured environment with many different types of objects, different whether toddlers, dogs all over the place should you grab a cup very hard or not so hard. And so there's that, the environment kind of operating. And then, let's look at the car, for example, autonomous driving, effectively two degrees of freedom there, more or less. The human hand alone has 20 degrees of freedom. So we're just dealing with much harder problems when it comes to physics and engineering. And while I am bullish, I just think that it's going to be a little bit slower than maybe some management teams out there would have you believe.

Moz Afzal:
So maybe coming to the last question, and I guess it's the open question, how do you think about 2026 as you think about your space? You touched upon many of the topics we talked about today, but where do you think you, from a technology perspective, an AI perspective, where do you think we kind of land by the time we get to the end of this year?

Dominic Jacobson:
Yeah, I think I'm going to be a little bit boring. And I think I'm going to say that we see a continuation of the same. I think the AI trade continues from here. And look, I think for a couple of reasons. The first is that there amends a sizable delta between the guidance of some of these AI bellwether stocks that they're giving and where consensus currently sits for them. So that in my eyes suggests that there's room for estimates to creep higher over the course of the year. Valuations, again, I think a lot of people point to them. I think that they're not bubble-esque multiples, at least on many of them. I get the spending and the digestion narratives, but people have been worried about those things since the AI trade started in 2023. I think ultimately they probably will materialise at some point, but I don't think 2026 is that year. And the reason being is that the committed, you look at the committed volumes for a lot of these companies, they have visibility into 2027. So I think at a high level, we continue to see more of the same this year, barring any exogenous events, which to be honest, we've already started this year.

Moz Afzal:
Well, Dominic, listen, this is very, very fascinating. It was very interesting. Certainly I learned a lot and I'm sure our listeners learned a lot as well. So thank you very much for coming on the podcast. I'm sure we will have you again very soon.

Dominic Jacobson:
Thanks, Moz.

Moz Afzal:
But thank you very much for coming on. That wraps us up today for Beyond the Benchmark podcast, and we hopefully will be again on very soon.

 

This podcast is provided for general information only and assumes a certain level of knowledge of financial markets. It is provided for informational purposes only and should not be considered as an offer, investment recommendation, or solicitation to deal in any of the investments or products mentioned herein and does not constitute investment research. The views in this podcast are those of the contributors at the time of publication and do not necessarily reflect those of EFG International or New Capital. The companies discussed in this podcast have been selected for illustrative purposes only or to demonstrate our investment management style and not as an investment recommendation or indication of their future performance. The value of investments and the income from them can go down as well as our up and investors may get back less than the amount invested. Past performance is not a guide to future returns, return projections or estimates, and provides no guarantee of future results.