Loading summary
Transcript32 lines
- [00:00]
A
Foreign. Hello everyone and welcome to the latest edition of the View Beyond, a series of podcasts in which we delve a little deeper into the topics on investors minds. Despite the ongoing conflict in the Middle east, we're currently witnessing record highs in U.S. equity markets. Joining me today to delve into what's behind these historic moves are Carsten Menker, Julian Spear's head of Next Generation Research, and Manuel Villas, one of the co managers of Julius Baer's Next Generation Equity fund. We're going to dive deep into the fascinating world of artificial intelligence and its ripple effects across financial markets. Just before we start, I would ask that you please listen right to the end of the podcast to make sure that you hear the important legal information. So welcome to both of you. Carsten and Manuel, thank you for joining me today.
- [00:53]
B
Good morning, Bernadette.
- [00:54]
C
Hello to both of you.
- [00:56]
A
Why don't we start with the big picture? We've seen a surge in AI development recently, particularly with companies like Anthropic gaining traction. How is this impacting the broader tech landscape? And what are we observing in terms of financial performance?
- [01:11]
B
Well, it's quite remarkable actually. Just skimming through the news, we read about the releases of the newest models such as Anthropic Slot Mythos. We also read about how quickly some of these AI companies are growing. Anthropic again has reached an annual run rate of $30 billion. That's up 30 times in around a year. And mirroring that, the demand for the data center infrastructure is booming. It's clearly exceeding the available supply. As a result, the data center construction boom is set to continue and we're seeing insatiable appetite for for everything, compute power, memory storage, even now. Also optical transceivers which are becoming more crucial for connecting these high performance data center clusters. So for me, in a nutshell, everyone needs more brain power for AI. Manuel, how do you see it? What has been standing out for you?
- [02:09]
C
I agree with you. What we see is that basically 2026 has been the year where agentic models have gone mainstream and agentic demand has triggered an order of magnitude of token demand above what we had a year ago. Openclaw, which was originally called Maltbot, then Cloudbot, might as well have been one of the most important developments of the year. We see this broad based uplift across the entire value chain, all the way from the substrates into the hyperscalers. And we see a reshuffling as well of the Frontier Lab ecosystem. And you said it before, anthropic revenue acceleration has been key to this whole puzzle. And as you said, let's not forget this was a company with annual recurring revenues of around 14 billion less than six, seven months ago. We see semiconductors, we see memory, we see networking, all benefiting from AI deployment at scale. And we try to position the fund as well and the strategy to capture some of these imbalances. There is a clear shift from training and into inference and increasingly different types of inference, changing hardware needs. And the important thing here is that this agentic demand is driving inference demand and inference demand is inference has several steps across which you need to decode the input tokens. And this has been one of the key reasons why we have seen such a shortage in some of these areas of the memory hierarchy.
- [03:56]
A
Okay, just listening to the two of you talking there, there's already so many kind of subcomponents of AI that one can focus on. But what segments have been the focus for you so far in 2026?
- [04:06]
B
Carsten, let me zoom out a bit because I think from a very high level, looking down, we have this debate about hardware versus software, right? And everything Manuel said, everything which I said before centers around hardware. And this is certainly what has been driving the AI theme thus far this year. And with hardware, there is always the question if things have gone too far, share prices, earnings, expectations, valuations. But thus far I would say the answer is no. As capital spending continues to increase, which is ultimately driving the demand for hardware and for software, I think there is quite opposing views. Clearly it's under pressure from the rapid development of AI, from the development of the agents which could take away software licenses, especially seed based ones. But will AI completely destroy software? No. Will it destroy some software companies? Yes. So for us, for the time being, we have a clear preference for hardware over software, because for software we simply can't see clearly at the moment.
- [05:22]
A
Okay, and Manuel, when it comes to your portfolio, what have you been positioning for?
- [05:27]
C
When we look at the portfolio, we have to also understand that if we look at it on a traditional way, we have around 47% positioned towards information technology. But the important thing about AI is that it is a theme. It is a theme that spans into topics like energy, into sectors like materials, like industrials. And it is a theme very closely related to another theme that we have, which is called strategic autonomy. So overall, we have a pretty heavy positioning around this AI value chain that spans from IT to industrials, to materials. So everything from the copper clad laminates to the boards to the substrates, to the fabs, which are basically the places where you fabricate these semiconductors, the wafer fabrication equipment to the compute units, the chip designers, the memory providers, advanced packaging and even power supply units. So we are really positioned for this AI infrastructure super cycle where we have focused on the picks and shovels which, which are semis memory and the data center ecosystem. But the most important thing here is that we must distinguish between cyclical and tactical place to foundational and structural place. And with that we also want to distinguish between long duration constraints and imbalance intensities. So basically we see long duration and high imbalance intensity in the memory stack where we see the most demand for high bandwidth memory, that is basically stacked DRAM that is going into the accelerators. With that the prices of DRAM and other types of memory have also increased quite a lot. Also because the memory market is a very oligopolistic competition where you have three key players being eligible for the high bandwidth memory fabrication and they have concentrated their capex and investment efforts in DRAM and, and high bandwidth memory instead of NOD and memory which is seen as NAND and it's basically flash or HDD which is this long term storage memory. Basically when they concentrate in DRAM and HBM they stop concentrating on nand which has opened the room for new players. Here we also see a long duration constraint and a high imbalance intensity across the power infrastructure going, going from the grid to a transformers and so on. On the medium duration where we see, let's say semi structural constraints with a low imbalance intensity at the moment we see testing and specialty wafers. And where we see a high imbalance intensity we see substrates. And on the short duration of these constraints with high imbalance intensity we see transceivers and we see power semis and this is more likely to be a cyclical play. So that said, we really focus on those companies that are benefiting from the rising complexity of the AI workloads, including as well everything that is happening with agentic AI where all of a sudden CPUs came back into the mix. So CPUs are central processing units and for agentic purposes, for agentic workloads they are in charge of the orchestration. And what we are seeing right now is, is that the ratios from CPU to GPU within a data center have completely changed in favor of CPUs. So we stay selective where monetization visibility is really the strongest.
- [09:01]
A
Turning to data centers, which you mentioned earlier, Carsten increased demand is inevitably leading to challenges there. Are we witnessing any bottlenecks or constraints within this booming sector?
- [09:14]
B
Well, I Think we've heard quite a few constraints already from Manuel before. So it's a simple mismatch between demand and supply which is driving these constraints across the whole value chain. Right. And we have to understand that this is not all about IT equipment, which is of course core and center to data centers, but, but it's equally about non IT equipment, as Manuel also alluded to. So this scarcity is giving equipment providers along the whole value chain substantial pricing power. And arguably the biggest bottleneck remains the access to the power grid. So not power itself, but access to the power grid. There are very long queues, especially in the US to get connected to the grid, prompting the data center operators to turn to on site power generation, for example, using a renewable energy, but also gas turbines. So these gas turbines being on site, being fed with natural gas via pipeline, of course this is producing electricity. This is helping to get the data center up and running. But from our point of view, this is not very economical. So it remains to be seen how these kind of constraints and bottlenecks play out over the long term.
- [10:38]
A
Okay then, that sounds like a potentially inflationary environment. Is that something that we're seeing play out, Manuel?
- [10:44]
C
Well, what is certain is that the bill of materials has increased quite a lot for most of these hyperscalers trying to build up a data center. And everything from GPUs to CPUs has been suffering and should suffer into the coming years because of the rising memory prices. So we've seen really memory prices double, triple, quadruple in some of these different areas. Specifically on DRAM dynamic random access memory, which is used for the high bandwidth memory, which is the HBM that goes inside the GPUs and the flash memory, which has become really important for, for these inference processes and decoding, as I mentioned before, this KV cache. And as long as you increase the context window, which means that you can put more stuff into the LLM in the context, it will put some pressure on the high bandwidth memory and the flash memory. So I want to talk quickly about, about the memory tiering for these inference purposes. And what we see is that there is a natural trade off between capacity and speed. So the fastest one, which is the one that we are using in the GPUs, is a high bandwidth memory and it is used for the most frequently accessed data. Then you have the system memory, which is the host dram, which is moderately accessed data. It has a little bit more capacity but less speed. And you have the flash memory SSD where you have these cold spills the long term context and it is less frequently accessed. And in the bottom of this tier you have the highest capacity with the hard disk drives which are used for frozen and archived data, basically storage. And for data centers you want a specific kind of storage that, that is efficient in physical terms. So basically the one that occupies the least space and has the highest capacity. And this has been the 32 TB drives for HDD and you can see lag times of up to one year. So memory has become a critical part of the bill of materials of the hyperscalers over the past couple of years. But for this year we believe it's going to continue rising and that's what we're trying to capture as well. So memory storage and the networking as well is very important for this inference process.
- [13:27]
A
Okay. And aside from these sort of obvious bottlenecks and constraints that you're highlighting, there's been some backlash in the US of late when it comes to the actual building of data centers. Do we need to be worried about that?
- [13:39]
B
Yes, that's an interesting point. We've heard about some U.S. cities, counties and states issuing bans on data center construction, either permanent or temporary bans. And I think this is a very interesting development as it seems to be based on a quasi bipartisan consensus, which is quite unusual in the US I assume so. The Democrats oppose the environmental impact of the data centers, while the Republicans oppose higher power prices. Essentially, they're coming from two different directions which put them into the same position opposing these data centers. Plus, as always with such big projects, there is a lot of nimby, the not in my backyard mentality. Nobody wants a data center in their own backyard. Still, everyone wants to use AI, of course. Do we need to be worried? I don't think so. Yes, the number of postponed or canceled data projects has increased in the US but so has the number of total projects, meaning that most likely on a percentage basis, this has not really changed a lot.
- [14:53]
A
Okay, thanks for that, Carsten. Now, focusing on the hyperscalers, we can't ignore the fact that they've recently released their quarterly reports. We've what key takeaways have emerged from those announcements?
- [15:05]
B
I think it's three points which really stood out. Firstly, the hyperscalers see growing demand for data center capacity. Consequentially, they have revised their capital expenditure guidance upwards and the market now expects around 730 billion for 2026 and nearly 900 billion for 2027. Secondly, related to what we discussed before, there is less bang for the buck. So component Prices continue to increase, meaning that the expected increase in capital expenditure reflects a mix of this component. Cost inflation and genuine capacity growth. Finally, and perhaps most importantly, the the monetization of AI is materializing. So we see the hyperscalers revenue growth accelerating to around 20% year on year last quarter, which is the fastest in around five years. And this number becomes even more impressive considering that the group's revenues are around 50% higher today than what they were in 2021. So growing revenues of a very high base is a clear indication of monetization from our point of view.
- [16:25]
C
From my perspective, what really stood out from the earnings was obviously the scale of investments, but also the backlogs. So the backlogs have grown to above 2 trillion and they have grown at a very fast pace in line more or less with the capital deployment. We see strong revenue acceleration across the board in the three main players in their cloud business. But we mustn't forget that in China we have other hyperscalers and in the US in the western markets we also have some NEO clouds. And if we add up the neo clouds and the Chinese hyperscaler numbers capex into what Carson mentioned before of around 720 billion for this year, we could be flirting with numbers very close to 900 billion and numbers that increase to a trillion if we see a similar rate of change as what we've been seeing over the past few quarters. And this trillion, as Carson said, is getting less bang for its buck because bill of materials has been rising. But we also have seen the guidance for this continued buildout globally and this build out basically trying to satisfy and catch up with demand. And this spend is tied to this inference demand and not just trading. Obviously some of these hyperscalers and these cloud providers, let's say, are built differently. Some are vertically integrated quite well all the way up from their compute units, where they are chip designers, where they have a frontier model as well, and where they have a cloud business. And these are the ones that have seen the fastest revenue growth in this business over the past quarter. And all of this is just confirming that we're still early in a multi year infrastructure build cycle. And that's what the phone is trying to capture.
- [18:17]
A
Okay, then monetization is a critical word when it comes to the artificial intelligence arena. These companies are investing enormous sums of money. Can you elaborate on how these companies are actually turning AI investments into tangible revenue?
- [18:32]
B
Well, I think what's important to remember is that this is happening through multiple channels. It's not one single channel. Which of Course makes it easier to monetize, but makes it more difficult to track. And sticking with the hyperscalers, from their perspective, increased advertising quality leading to rising advertising revenues is one of the important revenues we are tracking. A second one is stronger cloud revenues, adding to this lever they can pull. So the better they are offering in the cloud in terms of AI tools, the higher also the revenue they are able to generate on that front. And for the frontier AI labs such as Anthropic and OpenAI, of course, it's the paid access to their models which is typically usage based. More usage means higher revenues, as we discussed, or as we saw at the very beginning, with Anthropic growing their revenues 30 fold. And this can be either via so called APIs or through corporate or consumer subscriptions. Plus these AI frontier labs, they have partnerships with the hyperscalers, Anthropic with Amazon for example, or OpenAI with Microsoft.
- [19:50]
A
Okay, so then, given this positive outlook, what does that mean for investors interested in this space given that cloud computing and AI is showing strong performance year to date, outpacing global equities?
- [20:01]
B
Well, very simply speaking, stick to the theme. Our own next generation cloud computing and AI index is up around more than 30% from its low in March, which compares to around 15% for global equities. While I don't want to suggest that the index keeps on rising at this speed, we still see a very favorable fundamental backdrop for the theme. The established trends, we discussed them, I think in great detail. They remain in place, driven very much by the spending of the hyperscalers and other data center operators. And most importantly, earnings growth is keeping valuations in check. As a case in point, the price to sales ratio of our cloud computing and AI index today is at around the same levels as prior to the pandemic.
- [20:53]
A
What about from your perspective, Manuel? Are there winners that you're sticking with or even new ones emerging to replace some that might now past their best, so to speak?
- [21:03]
C
You know, I think the space has changed quite a bit over the past year and a half. We still favor the core enablers, wafers, advanced packaging, memory fabs, compute units, CPUs, GPUs, the designers. And there is a strong differentiation in these sectors, as mentioned before, but there's also a strong dependency on other parts of the other industrial sectors, let's say. So facility management firms, construction firms, these have all been strong beneficiaries of what's happening here. And there has to be some dynamism within the capacity to change between the sectors as well. And so Looking ahead, what we focus on growingly is this compute diversity that I mentioned before, from GPUs to CPUs to custom accelerators, where we see most of the hyperscalers having their own chip design and they then collaborate with a front end designer for some hyperscalers and for others with only a backend designer, we see a very strong dependency on Asia. On the Asian supply chain, where we focus on three key markets, which is South Korea, very important for memory. We have Taiwan, which is very important for basically everything from the wafers to the fabrication, the wafer fabrication equipment and to the power supply units. And Japan, where we have an industry that is also very important for substrates wafer fabrication equipment. So broadly, this is how we try to target this space. Looking ahead, we see we're not going back in terms of compute use and our token consumption should only pick up into the coming years. And if you don't believe me, you can believe what some of the executives behind some of these companies have said in terms of what they expect for software engineers to consume in terms of tokens. We watch AgentIQ AI quite closely because this is going to significantly reshape demand patterns. And who says that we can see another order of magnitude in token demand in the year ahead?
- [23:17]
A
Okay, I'm glad that you mentioned Asia there. Manuel Carsten, how international is the value chain? Is this just the US versus China?
- [23:26]
B
No, not at all. And I think this is what Manuel highlighted already. The fact that AI supposedly is the US versus China, I think is a big misconception. So yes, of course the US is developing the leading AI models, but they cannot do it without a global value chain supporting them. So we have European companies as a part of this value chain, for example, supplying the semiconductor manufacturing equipment, the power management systems, and also the cooling systems. And then as Manuel highlighted, you have the Asian semiconductor companies, compute chips from Taiwan, memory chips from South Korea, plus the equipment and testing companies in Japan. So this is truly global. The Chinese value chain is more local due to political constraints, but even they cannot do everything at home.
- [24:23]
A
Okay, and I have to ask you, what would it take for you to reassess your thematic view when it comes to cloud computing and AI? Carsten?
- [24:29]
B
Well, for us in research, certainly the hyperscalers capital spending is the single most important indicator to track. Also, as long as this goes up, hardware will very likely remain in focus. Valuations, of course, are another important indicator. So should they become excessive, we would also need to revise our view on the theme overall. But as I said, this is not the case at the moment.
- [24:57]
A
Finally then, Manuel, what subsectors do you favour currently for the rest of 2026?
- [25:02]
C
I think the rest of 2026 is a very long time, especially with the meetings that we expect this week between the leaders in the US and China. And as Carsten mentioned before, there is much nuance to these supply chains and AI has become a geopolitical negotiation tool in the form of the GPUs in the form of the wafer fabrication equipment, and China's capacity has been very limited because of its lack of access to both GPUs and wafer fabrication equipment. So we have to see what happens. The space changes quite a lot as of now and as mentioned before, we strongly focus these structural constraints with high imbalance intensity in the memory stack we prefer the more structural play which is DRAM and HBM over flash and hard disk drives we prefer power infrastructure and on the semi structural side with a high imbalance intensity, we are also positioned within this substrate space in this copper clad laminate space because we see some of the choke points here as well and we see capacity constraints that could spill over into the next few months. We have a very high conviction in these areas. We see selective opportunities in some other areas as mentioned before, for example the CPUs on agentic workflows, but the overall view is that the market is evolving from compute scaling to system level optimization. The market is evolving from compute demand for training into compute demand for inference and inference with the agents is crucial. We stay focused on these areas with pricing power and structural demand visibility. And what we are firmly convinced is that token demand is not going down in the near future.
- [27:12]
A
Fascinating insights today, gentlemen. We could keep the conversation going, but we're just out of time now. So thank you so much for shedding light on this rapidly evolving landscape. That's it for this week's view. Beyond Podcast Manuel just mentioned the upcoming meeting in China. I'd like to let you know that we're recording this podcast on Wednesday 13th May, so that is ahead of that meeting. We look forward to delving into another topic of interest in next week's episode. And meanwhile you can hear more from Carsten and our other research analysts on the Moving Markets Podcast, which we publish weekdays at 9am Central European Time on your favourite podcast platforms. Goodbye for now. The information and opinions expressed in this podcast constitute marketing material and are not the result of independent financial or investment research. Please refer to www.juliusbear.com legal forward/podcasts for further other important legal information.