Transcription
7 Clarman has been managing an investment fund for 44 years and in almost half a century, he has only experienced 5 losing years. So perhaps that doesn't impress you much because we've had few losing years, but you should know that we're talking about the 2008 crisis, the bursting of the dot-com bubble in the 2000s which led to a lost decade in financial markets. So naturally, when an investor of this caliber tells us that the current market has the characteristics of a bubble, well, we listen. But be careful, what he's telling us is a bit more subtle than "AI is a bubble." He's not saying that the technology is useless, nor is he saying that the market is overvalued. Ultimately, he's asking himself pretty much the same question as I am, the one I ask in almost every video I make about AI. That is, who will capture the value? Will it be the foundries that manufacture the chips? Will it be the chip designers? The hyperscalers, so Amazon, Microsoft, Google, Meta, so the cloud people. Will it be labs like OpenAI or Anthropic that create the models, the software, or ultimately the customers who will reduce their costs or improve their productivity? Well, if we're honest, we're not yet sure who the big winners of AI will be. And this poses a problem for Clarman because he says that when you pay 40 times earnings for a stock, sometimes much more, you need to have very strong conviction about a very distant future, and he simply doesn't see how one can have that conviction. He says the market has become binary, almost a caricature, with on one side the AI winners that everyone wants to own, on the other the AI losers that everyone wants to flee, and in the middle, ultimately, companies with no direct link to AI that no one looks at anymore. But, and this is important, Clarman is not, as is often the case, someone who sleeps on a mattress of cash repeating that everything will collapse, etc. He does, in fact, hold a few stocks related to AI, let's say the obvious winners like Amazon and Google, which he calls "cash flow machines." He has even bought land with access to electricity that could potentially house data centers. But ultimately, if we look at his exposure, he is moderately exposed. He says he has about 10% of his portfolio exposed. So he is a little exposed but cautious, and above all, he spends most of his time looking where no one else is looking, i.e., at companies with no link to AI. So, to summarize his thesis, yes, AI is and will be very useful, but investors are paying very dearly, too dearly, for distant prospects. So there are probably valuation aberrations, and he will look elsewhere for opportunities, leaving them behind, because all the cash, all the investors' money is rushing towards all areas of AI.
So, I'm going to go a little further because I find his discourse rather basic. It's important to understand that the market has completely changed its way of viewing AI. The market has stopped buying only Nvidia. Before, that's what everyone did. In short, we buy Nvidia, and investors gradually move up the value chain. So they abandon Nvidia, they also abandon Nvidia's customers, and they start valuing all the physical constraints necessary for AI. So they move up the value chain like that. And to understand this movement, we need to go back to Moore's Law. For decades, computing has progressed thanks to a rather simple idea: we were able to put more and more transistors on the same chip. And Moore's Law stated that we could double the number of transistors on a chip every 2 years. So, chips became more powerful and generally cheaper. So, all of that was summarized by saying, well, that's Moore's Law. Doubling the number of transistors every 2 years. Today, this logic continues, even if you'll see everywhere that Moore's Law is over, that it's dead. So the logic continues, but it's ultimately much slower and especially much more expensive to have much more complex chips. And especially this Moore's Law now relies less on the miniaturization of transistors. We continue to gain power because that's the philosophy of Moore's Law. But manufacturers also have to stack chips. They have to use advanced packaging. They have to bring memory closer to processors. Heat also needs to be managed better. Power consumption needs to be reduced. So we are achieving an increase in chip power, but we see that it's no longer just basically the number of transistors. So what's interesting is to understand that progress is no longer just happening within the chip; it's happening in everything that surrounds it. And ultimately, computing has become so powerful that the challenge now is to feed the massive data, to connect them, to cool them, and to provide them with enough electricity. And that's what the stock market has understood in recent years. And that's where the stock market started looking for the next winners, the next ones after Nvidia, of course, because in 2023, the story was quite simple. The story the market was telling was Nvidia first, the most in-demand GPUs in the world. Hyperscalers, so cloud managers, buy these GPUs from Nvidia. So investors buy Nvidia on the stock market. So, we are typically in the years 2023, 2024, 2025. And then, by understanding where computing power comes from, the market understood something: a very powerful GPU without ultra-fast memory is not very useful. Memory is also needed, HBM, very high bandwidth memory, which is placed right next to the processors. And suddenly, once the market understood this, memory manufacturers became the heroes of the stock market. So now we are in 2025-2026 with incredible performance like Micron, which gained +800% in 1 year, SK Hynix +1000%, Samsung +500%. And by the way, on this memory explosion, and particularly what happened at Samsung, we published a complete analysis at the beginning of the year in the Boursecho newsletter. It explains HBM well and why memory has become so vital for data centers. So, it's a free newsletter where I learned a lot. So, I'm putting the link in the description. And the surge in memory is not necessarily irrational because there is a real shortage. And Morgan Stanley estimates that this shortage will last until the end of 2026. And in addition, we've seen a new phenomenon: hyperscalers are starting to sign long-term supply contracts. This is very rare in the sector. Before, there were no issues with memory, shortages or otherwise. It was something we bought cheaply and in large quantities. So, we see that there has been a paradigm shift, and hyperscalers are prioritizing supply security over price. So yes, there is a real bottleneck in memory. But be careful, because memory has historically been a cyclical sector. So, profits regularly explode. Stocks can appear cheap in terms of valuation. If we look at the price-to-earnings ratio, the P/E for example, sometimes they are cheap and sometimes they are not at all, because the market knows well that if we look at profits at the peak of the cycle, they won't necessarily last. And indeed, if you look at the profits of SK Hynix, for example, one of the stars, as I told you, of last year, you will see that they historically have a very unstable growth. We see profits that are real roller coasters. So one of the questions today is whether memory, and I want to extend this to semiconductors as well, whether it remains cyclical or not. And for many investors, semiconductors, memory, etc., are no longer cyclical since the arrival of AI. So, does AI permanently change the market for memory, semiconductors, etc.? Yes, that's certain. But that doesn't mean the laws of supply and demand are eliminated. The idea here is that if everyone builds too much capacity because they see huge profitability, whether in semiconductors or memory, or if hyperscalers slow down their demand, prices can turn around much faster than the market imagines. Which doesn't stop me personally from really liking the semiconductor sector, but it's something to keep in mind. Rather than saying it's no longer cyclical, I would say that the cycle is extremely long, and that, temporarily, for a few years, it's true that the cycle has disappeared. And after memory, the stock market chased another link in the chain: storage. And we've seen companies like Western Digital gain over 1000% per year, Seagate +700% in 1 year. Now, let's be very clear, hard drives are not at the heart of AI. They are not used at all for AI calculations. They are mainly used to store gigantic amounts of data, videos, archives, copies of models, for example. And obviously, the more AI we build, the more storage we need. So they are benefiting from the AI wave, but ultimately without being the most obvious winners of AI. But despite that, they are part of all these sectors that are growing along with AI.
Next, the market discovered photonics. Now, the word might sound scary, but the idea is relatively simple: the more data centers grow, the more data needs to circulate between them. It's like a huge giant brain spanning thousands of square meters, and at very high speed, copper quickly reaches its limits. As soon as you exceed 2 to 3 meters of copper, the signal degrades significantly. So the solution is to circulate data using light. And that's why companies like Lumentum have gained 900% in 1 year, Coherent +500%, Corning +300%. The stock market gains are completely astonishing, but there's also something real behind it. Very real. We have Lumentum, for example, which published 90% revenue growth in one quarter, a record, and went from a net loss to a net profit. And even better, Nvidia is not just working with these companies; it's financing them. Over 6 billion dollars have been invested in photonics since March 2026, including 2 billion in Coherent, 2 billion in Lumentum, and 500 million in Corning. So we see that Nvidia is investing in its own suppliers, but I'll talk about that a bit later.
Let's return to Clarman's question, and mine as well: who will capture the value? And if I quickly review the value chain, at the very beginning, we have the equipment manufacturers who produce the machines necessary to manufacture chips, including the largest European company, ASML, which I think you know. So, these companies sell machines to the foundries, who do the work, TSMC, Samsung, etc., they are the ones who actually manufacture the chips. Then, we have the chip designers, so Nvidia, AMD, Broadcom, who tell TSMC what to manufacture. We also have the memory manufacturers, whose names I've already mentioned, and then all the companies that supply servers, networks, optics, so cables and cooling, electricity, everything needed for data centers. And after that, we have those who manage these data centers, so the hyperscalers, the cloud giants, Microsoft, Amazon, Google, Meta, who are the ones financing all this infrastructure and who will rent out this computing power. On the side, I would put labs like OpenAI, Anthropic, who create the models. And finally, at the very end of the chain, we have the end-users, so individuals, but especially companies that use AI in their daily lives. So any company, I mean any individual, even if it's not the case yet, but any company that does IT development, research, wants to target its advertising better, etc. But I think today, almost everyone, all companies use AI and pay for it. So these are the end-users. And what's interesting is that in all businesses in the world, except for AI for now, as we'll see, the money comes from the final consumer. So if you take any value chain, automotive for example, it's the car buyer who pays for all the components in the chain, the car seller, the subcontractor who made the steering wheel, the engine, the bodywork, etc., all the way to the foundry that produced the steel, to the oil that was transformed into plastic in the interior. So all of this is paid for by the final price of the car, and today money is flowing abundantly into AI, there are no major financing problems for now, but it doesn't come from end-consumers. I almost want to say that this money has been advanced by the hyperscalers, and this year, according to Goldman Sachs, hyperscalers will spend about 770 billion dollars in investments. And you know what that corresponds to? It corresponds roughly, at least that's what we estimate for 2026, to 100% of their operating cash flow. In other words, hyperscalers are spending all they earn on AI, and by 2026, we already forecast 1200 billion in investments. So we clearly see the question of who will pay, because the pockets of hyperscalers are not infinitely expandable.
If I project myself a little into the future, I see four possibilities. The first possibility, the healthiest one, is real demand. Here, I'm going back a bit to my car example. So, we have end-users, companies and individuals, who are willing to pay a lot for AI, and these revenues make the expenses profitable. That's the normal system. Second possibility, debt. Well, as long as demand is not sufficient, the giants will have to cover the shortfall by borrowing, because we see that they are reaching the end of their cash flow. They don't earn enough money for everything they want to invest. So this works, but only as long as the markets lend. Third possibility, no one pays enough, so the end-consumers, and ultimately the bill ends up absorbed in margins, particularly the margins of hyperscalers. And fourth possibility, the money circulates in a closed loop. Nvidia finances its suppliers. Hyperscalers finance labs whose revenues return to their own cloud. Nvidia also finances labs because they know well that the cloud will buy their GPUs, etc. And ultimately, we have this financing of the entire ecosystem. And this phenomenon has a name: circularity. So, it's as if the money gets stuck in the system. And this works as long as each link holds. So the money circulates, everyone shows growth, etc. But the real risk is that not enough money comes from outside, because we see that growth will quickly be limited if hyperscalers stop constantly buying more. And the other risk is if one link falters. And when we think of a faltering link in AI, we especially think of the labs, because they have the most fragile business models. In short, the current business model is to burn cash without a proven economic model. And if they start to have a problem, they will spend less on the cloud. If cloud providers earn less money, they will slow down their investments, particularly at Nvidia, but not only there, but throughout the entire semiconductor supply chain. So everything will be affected, memory, optics, electricity, everything will take a hit, because ultimately the same loop that made everything go up together can make everything go down together. And the fundamental point is that the market doesn't need demand to disappear or even decrease, because we see that the demand for AI and computing is enormous. It's enough for demand to be a little less spectacular than expected for something to get stuck.
Now, be careful, I don't believe at all in the idea of a generalized bubble in AI. I reason more in terms of risk layers. I would say there are three layers of risk. The first layer is everything related to the big platforms, so especially the hyperscalers, of course, so with Microsoft, Amazon, Meta, etc. Why? Because these companies have such massive cash flows, they are so diversified, they have such solid balance sheets that they can obviously absorb mistakes, and particularly investments that yield less than expected. The risk here is obviously not the survival of these companies, and it's ultimately a question of valuation. But be careful, even for them, we can still have some minor concerns in the coming years. There's Goldman Sachs, for example, which expects the profitability of the seven largest tech companies to decline by an average of 7 points next year. Why? Because the depreciation of all the data centers, you know, the hundreds of billions they invest each year, will start to be deducted from profits, as any good accountant must do. And so, we can have results that remain excellent, but less good than expected, because the clock will start ticking to say, well, now, we need to bring in more money. In the second layer, I would put the specialized suppliers because they are obviously much less diversified. We've already talked about memory, so Micron, SK Hynix, Samsung, chips, of course, Nvidia, AMD, Broadcom, further upstream, TSMC, ASML, all the machine manufacturers, optics, which I've already mentioned, storage, energy with companies like Vistra, networks with Schneider Electric, etc., electricity producers, of course, like Vistra and Constellation. Now, there is real demand today in this layer, but there is always this risk of cycles and dependence on a few large clients, and a risk of possible over-investment. In this layer, we must understand that the risk levels are very different. We cannot, of course, compare TSMC, the huge Taiwanese foundry, or Nvidia, to, for example, a small player in photonics. So, I am making a clear distinction between TSMC, Nvidia, ASML, etc., and smaller players, so to speak. But I wanted to separate them from the third layer, which I would call "companies with promises." First, all the labs, OpenAI, Anthropic, etc., because these companies need a lot of cash to stay at the forefront, and if they falter even once for a quarter or two, they can really have serious problems. And in this group of much riskier companies, I would also include Necloud, which rents computing power. I'm thinking of Corwil, for example, which is valued very highly on growth promises, which has 67% of its revenue from one client, 77% from two clients. And then, ultimately, all the companies linked to AI that sell a narrative rather than figures, whereas the first two layers really had revenues, figures, growth, etc. And it is obviously on this third layer that we have a real risk of a bubble. And I tell myself, fortunately, fortunately that Anthropic and OpenAI have not been listed so far, because we would undoubtedly have a bubble. But let's see what the IPOs will bring.
So, concretely, what should we monitor to know if this situation is weakening? First, I think you're used to it. What we continue to monitor quarter after quarter are the actual spending of hyperscalers and potentially their discourse, but especially their spending, because ultimately the entire chain, as I've explained today, as long as we don't have a relay from end-users, the entire chain depends on these expenses. Then, we have memory and optics orders, because these are leading indicators. When orders slow down, the market can already understand that the party might be ending or at least calming down. Third signal, customer concentration. This is really something very important. You absolutely must look if a company depends, for example, on two hyperscalers. Potentially, at the moment, it can have completely astonishing results, but it can become vulnerable very quickly. It's enough for one of the hyperscalers to decide to change suppliers or to develop its own chip, for example. Now, I'm not talking about Nvidia, which is currently indispensable, but for smaller players, they really need to be careful. And fourth signal, you've guessed it if you've followed me a bit so far: it's the return on investment for customers, for end-customers. Are the companies buying AI really making more money, or are they ultimately spending on computing, on tokens just hoping to earn more? And the last point that should not be overlooked is long-term interest rates, because the higher interest rates remain or the higher they are, the more distant profits lose value. So this can potentially lower the valuation of all these growth and even hyper-growth companies.
And to finish, let's go back to our friend Clarman from the beginning, because even though he sees signs of a bubble, he's not sitting on his cash waiting for the collapse. He's even doing the opposite of someone who would talk about a bubble, meaning he's still investing in a few companies. I've already mentioned them: Amazon, Google, and a bit in data centers. And what worries him, remember, is the binary market. The fact that everyone is rushing to the AI winners and fleeing the potential losers, while he decides to dedicate himself to searching for opportunities outside of AI. So he's trying to see if, since all the money, and this is important to understand, all the money is being sucked up by semiconductors and everything around them, it leaves companies that are often undervalued and underpriced. So, even though I agree with him on part of his point of view, meaning it's very difficult to distinguish winners and losers, especially in the software sector, I wouldn't want to be in his shoes and be under-exposed to AI and semiconductors as he is. Perhaps he will be right. Perhaps he will end up like Warren Buffett when the dot-com bubble of the 2000s burst. That is, everyone said, "Ah, well done, Warren, he resisted temptation, he stayed with solid companies, etc." But it's true that when I look at Set Clarman's performance over the last few years based on his reporting of US stocks, we see that he is significantly underperforming the market. He made 12% over the last 5 years compared to 89% for the S&P 500. So I think, yes, this extreme of fearing a bubble too much is another risk. And on my side, I'm trying to find the right balance. So, not playing a single scenario, not being like some people I know, almost 100% AI or at least all the companies that orbit around it. Even though, when we look at recent performance, that's what we should have done. And I don't want to be under-exposed either, because I truly feel that AI will bring a lot of value to the economy, even if I don't know exactly who will capture it. As long as we own roughly the entire chain of machines, a bit of foundries, a bit of design, and hyperscalers, etc., I less need to know where the money comes from. It will be generally captured, the money will be generally captured somewhere, and since I also have more traditional companies that will be users of AI, ultimately I am on the entire chain.
So, as you can see, AI, and particularly the semiconductor value chain, is complex. So, if you want to go further, I will provide you with two other links to Boursecho newsletters, in addition to the one on Samsung memory. I will provide one on French AI winners, and another newsletter on TSMC, which is one of the most difficult links to replace in the AI chain, given that almost all the world's chips are made at TSMC and also at Samsung. But obviously, I have to ask you the question: is it a real bubble, or is it only a bubble on a few stocks, and which stocks do you think are still in a bubble? Perhaps some of you will think that no stock is in a bubble, but I'm interested in having a little poll like this. Put it in the comments. Thank you again. I know I talk a lot about AI, but it's important to understand all these concepts when investing in the stock market today, including when investing in ETFs, because you should know that in ETFs today, you have a large portion, I don't have the exact figure, but I think it's around 25% in semiconductors. You have 8% of Nvidia in your portfolio if you are invested in the S&P 500. So it's something to understand. You need to understand what you are invested in, even if you are invested in ETFs. There you go.