How China Could Dominate U.S. AI | Dr. Michael Power on Open Source and "The Three Assassins" of Moore's Law

I'm joined by Michael Power of Cascazi Consulting. Michael is a veteran of macro strategy among many topics. Michael, it's great to see you, but you right now have written an essay that is kind of blowing my mind. It is extremely in-depth and analysis of USAI architecture and Chinese AI architecture. You find that Chinese AI architecture has significant advantages over the USAI architecture and that basically the hundreds of billions of dollars and trillions of dollars that are currently and going to be invested in USAI right now may be basically a bus. So that this has extreme geopolitical consequences for the entire world, China and the US, but also economic and market consequences as well. So there's so many strands we can grab onto, but just how do you take us into the journey? How did you first start thinking about this? Tell us about the process of you discovering this and then we'll just get into it. - Well, thank you, first of all, Jack, for having me. I think that I'm now in semi-retirement, though I've discovered that the word retire doesn't exist in retirement. And I'm one of these people that was determined not to allow my brain to atrophy and I still continue to make plenty of speeches on those more traditional subjects that you mentioned before, but nevertheless, I've made it my business to try and understand probably the great theme of the world today, not least because it has completely enraptured Wall Street. And I felt that in order to be able to have a meaning called contribution to make, I needed to understand it. So what I did is for the first six weeks is I just immersed myself in everything AI and in the first instance, that basically meant learning the language of AI because a lot of it is jargon, which is not easy to understand. And I first translated it into language which I could understand and any recently intelligent person, such as yourself, could understand as well. And when I was doing this, I started to realize coming from an objective perspective that the narrative that dominates much of Wall Street thinking was not as strong as it was made up to be. And the Chinese, although I don't think the situation is as we speak today, one of the Chinese leading deep down the structural process that they're putting in place with regards to how they're approaching AI. I think, roll forward three years, we'll outmaneuver that of the United States and that their model, which is, first and foremost, built on the idea of open source rather than close source and we can come to that, if you will. But nevertheless, their model has a lot more runway ahead of it and it's a lot cheaper runway than does the US model, which I think is actually starting to run out of breath. And if you look at the US AI ecosystem, the valuations in that I would roughly estimate at 15 trillion, if you look at all of the publicly traded securities and then the venture capital funded companies, whereas the market cap of the Chinese ecosystem, like a lot of it is private or government backs, but I mean, you know, Alibaba has a market cap of less than a half a trillion dollars, which is certainly still a very large company, but a pale is in comparison. First of all, Michael, yeah, I just want to share some of the consequences. I mean, you said that basically the grim reaper is coming for the American AI bubble. So we'll get into that in a second, but first, what is the source of the Chinese advantage over US technology? And how does that disprove or challenge the consensus within America about US AI advantage in Silicon Valley as well as Wall Street? Well, I'm going to use a full letter word on a very erudite discussion like this. The essence of the Chinese AI approach is that it's free. And I can't say that enough. And that is because they have a completely different philosophy as to what AI should be as compared to the US model. China is building a structure at the moment where AI will be a utility like electricity. And the value that is going to be derived from using the electricity is where they're going to benefit. But the electricity itself is a pretty no value product. Yes, you can make a little bit of money on the way by generating electricity. But as you well know, the real valuation from electricity comes from what it is purpose to do. Whereas the US model is essentially that it's a service that can be monetized if they can monetize to the degree that they need to, given capital expenditure they've undertaken, but they need to service. And these two different philosophies, profoundly different philosophies, mean that there are two paths developing. And to some extent, the analogy is not precise, but it's good enough. Essentially, the Chinese approach is the Android approach. Because as you probably know, although Android is technically owned by Google, Google derives no money from that. It's actually controlled by a foundation. And that foundation is a non-profit. But compare that to Apple and their ecosystem, which is extremely profitable. Android actually follows another great example. And that is of course Linux. And if you look at the top 100 supercomputers in the world today, Linux runs 100 out of 100. And so essentially, there is this philosophy of open source or open weight, which is essentially the Chinese approach to AI versus closed source. And my sense is that over time, as with Linux, as with Android, the Chinese approach, which is open source, open weight, specifically, will likely win out. Which is why I say they have a longer runway ahead of themselves than does the United States. I should say in this that the US may be like Apple is able to do, create an ecosystem that Apple is securing. And there are four countries in the world where Apple is more popular than Android, but everywhere else in the world. Android is more popular than Apple. Those four countries, by the way, are the US Canada, UK and Sweden. And what we may be beginning to see in a very analogous sense is the emergence of a bifurcated world, where because, in large parts, the Chinese offering is free, it's winning big time outside of the core capitalist world centered on Wall Street. And so, and I've often had chats to people who are essentially Apple heads or iOS heads. And they don't really actually grasp the world of Android. They don't actually understand that there is another way. And I think what's happening at the moment that there is another way in AI. And it's winning big time in the world at large, not specifically, by only looking at what's happening in the United States. And you note that Chinese AI models have a very high rate of adoption. And that a venture capitalist, I believe, from A16z, said that 80% of the startups that they fund are using Chinese models, not US models, which is certainly very surprising. So the open source, first closed source, I think. So closed source, what it literally means is that the models' weights are not disclosed, whereas open source, they are. But you're referring to the pricing model. So you go into that one, that. As an open source model, you can play with it. But with a closed source model, you can't. And I always make the distinction that close sources like ordering beef Wellington and a perfect beef Wellington arrives on your plate. Open source, you order beef Wellington. But when it arrives, you get a list of all the ingredients. And you can play with it and mix the ingredients up and completely reform it and conceivably create egg and bacon ice cream out of it. But nevertheless, you have complete freedom. Once you receive your beef Wellington in the open source world to reconfigure it to whatever way you like. OK, but when you say it's free, so Chatchy BT is massively subsidized. And there are free versions of Gemini, Chatchy BT, all the American models. And I'm sure there are China. But are you saying that China doesn't charge its users or that the charging model, the pricing model is different? I'd say less than 5% of the Chinese models have a user fee attached to them. And it's only for unbelievably specialized areas. And to be perfectly honest, generally speaking, the answer is no, they don't have a fee attached to them. OK, but then what is the business model? Keyport throughout your piece, which is excellent and what we'll link for listeners to read it. But a key part of your piece is that China has a cost advantage, but they still have costs. So how is the Chinese LLMs and the clouds and the AI in China going to be funded if they don't make any money? Well, just take Quinn, which is by far the most powerful of all of them. That is Alibaba's large language model. Quinn is applied and used everywhere else within the Alibaba community. So Tao Bao, which is their online shopping or alley logistics or alley pay, will use almost certainly when as their model. Now, in using that model, there may be a-- and often is-- a small fee that gets paid across from Tao Bao to essentially the central resource that is Tong Yi, which is the IIT sector that controls Alibaba. And that's how Alibaba then gets funded. Now, this is not the same I must say immediately for all Chinese LLMs. There are some now that are not connected to a broader commercial network. But we'll come onto that later because there is a mechanism now arising when they can be. And but the central point is that it's like a central expense, the research and development budget, as it were, of Alibaba that goes towards Quinn. Some fees do come in from other parks at the Alibaba Empire to help fund that budget. So let's say, Michael, if you're right, in five years, what does the world look like in video, all of your company? Can I just follow on something there? Just because it's really-- it's clever. Because it's only in the last week that it's happened. As I said to you before, that Google essentially owns Android, but doesn't really make much money for it. However, it has recently agreed because, actually, every Samsung phone that I know of is run on Android. It's done a deal with Samsung. That in the search engine of every Samsung phone going forward will be Gemini, in other words, Google's LLM. And Samsung is now paying a fee to Google for the rights to have that embedded in their phones. So there is a way, an example of how it's being done in that context. But this is just happening everywhere within the Chinese community. It isn't just one-off exception. But Android is able to be partly monetized through arrangements like that. And I'm just using that as a result. Thank you for explaining that, Michael. So now we have just a sense of our terms, what we're dealing with here. So if you're right, in five years, what does the world look like? The publicly traded semiconductor supply chain of USAI, the privately backed AI models, most recently open AI andthropic, et cetera. What does that world look like? I imagine that those companies would have-- you see severe challenges. And how does that world differ from the scenario envisioned by-- imagine by many of the rosy-eyed USAI optimists who probably expect a video's market cap to be above 10 trillion and expect open AI to be massively profitable and the rest? Well, first of all, we have to be careful. When we use the term world, we don't just mean US world. We mean world world. But unfortunately, when I'm listening to Bingberg and CNBC and they use that term world, it generally speaking means, like Apple World. It's just the world that is defined by the United States. There is some international dimension to it, but it's a very small part of the whole. So if we're talking about what will US world to start to answer your question, look like. I think we're going to see that-- and we're already beginning to see-- when Anderson Horowitz is seeing 80%, 90% of people presenting to it using free software from China, that there are-- and I'm boring a word here from Jamie Diamond and you'll understand it, giving your background cockroaches all over the place-- that are essentially starting to be used. We're seeing Airbnb now is essentially moved over to Quet. But essentially, there is nothing stopping those people who find this software. And I'm going to use a phrase which is not really fair, because it's actually very good. But good enough, let's just start with good enough. And they're finding if they can use it for free, and it's good enough, and just take the example of Quet. It's fluent in something like 120 languages, which if your Airbnb is rather a good plus to have. So I think what we're seeing now is that there is leakage between the two ecosystems that have essentially been described by us, US world and world. And there is leakage from US world into world. So projecting five years forward to go back to your question, I see this leakage is continuing. However, it goes beyond that, because there are technological issues that are now beginning to question, firstly, the making of the hardware, and secondly, construction of the software. And they feed off each other. But they need to be looked at separately before we put buying them. First of all, I don't think Nvidia's hold on the chip making world is going to be anything like a strong. And again, part of this is because of what Amazon is doing with its own chip making, what Google is doing. Amazon has the training, Google has the GPU. And I think what we are seeing is that, even within the United States, a low-level civil war is breaking out between the major players in big AI. So much so that the dependency on Nvidia chips is starting to be reduced. But outside of US AI in the rest of the world, there is no doubt in my mind that we are seeing all sorts of efforts to diversify the Chinese even have a word for de-ambidiarization. But diversify away from dependence on, let's just call it expensive chips. And most of those expensive chips historically have been in video chips. So in the hardware side, there are all sorts of options that are starting to open up. First of all, also, and we'll see in the open part of my presentation, the whole area of Moore's law is starting to come under pressure because, basically, we get into a point where you can't really make chips much smaller and still hope to carry to continue the compute level that comes from those chips. There are rules of physics, rules of what I call material science, chemistry, and then rules of economics. I call them the three assassins that are actually saying that there's not much road ahead for Nvidia continue to miniaturize its chips. What China is doing is essentially creating-- I don't know, they're not learning this. That's happening even in the United States, but a whole new ecosystem that is basically built around not two or three NM's, but somewhere in the 14 to 18 NM's space. And what they're doing is that they are building what I call cognitive skyscrapers, cognitive towers. So they use the chips as the base, but then they layer other sorts of chips and memories and various other things on top of it and create sort of mega-like world. And that is increasingly the way forward. So I don't think China is thinking smaller. It's thinking smarter. And this is one way where you can continue to be very relevant to the chip space without chasing down that rabbit hole the idea of making the chips ever more small. Because as I said, Moore's law is close to dying. It's on its netbed. It may not have passed its last breath, but it's not looking good. So in the hardware space, Invidia-- and we already seen certain things that they're doing here, they're doing it more in the software space, but they too are also starting quietly to do it in the hardware space too. But Invidia is recognizing that its model up until now cannot be the only way forward for it. It's got actually diversified away. Actually, potentially look at slightly larger chips. Actually look, and talk about software yet. But actually look at complementary software with those slightly larger chips in order for it to remain relevant. The problem with that for Invidia is that the margins associated with that alternative are much, much lower than they are with the model at the end of it, too. Which then starts to moving forward, potentially undermine the whole that Invidia has, the margins that it has. And then that will potentially fade through into profits, and therefore start market value wage. So that's what's happening on the hardware side. And I truly generalized a number of areas there in order to be able to answer your question. But that's essentially what happened. And as I say, the Chinese are doing it, yes. And they've got a vested interest to do it. But even Amazon is doing it. Even Google is doing it at the moment. That then wants to diversify away from these high priced chips that Invidia has. Then increasingly, they're not built for purpose. And the new world that we're moving into for chips requires a different combination. On the software side, there's breakthroughs happening everywhere. And it's not that Invidia hasn't got a very powerful software hold on its chips. It does it through something known as CUDA. Essentially, it hides in developers to the way that they're able to use Invidia's chips. What's happening at the moment is that both in the US, but especially, especially in China, people are finding way to separate them. And this they are doing both two sides to the chips, the training of them, and the inference that derived once you train them. So you endow a chip with knowledge. And then you get that chip able to answer question that you and I might pose to. But in the training stage, which is where Invidia has a specially good hold, there are breakthroughs taking place, the most important of which happened in the last 10 days, which is something that Deepsea did. It happened on the EZ, so I don't think the market is truly understood the scale and what that breakthrough means. But it's also happening in the inference side, which is where to be perfectly honest, lesser chips are used, often chips that are previously involved on the training side after two or three years, they still got a little bit of useful life left in them. So they get transferred over to the inference side for another couple of years, where they can still be used on their point, but the margins on inference chips are much lower than the margins on training chips. But the essential point is Invidia here is now phasing attacks on both fronts, hardware, and so on. So it sounds like the consequences of what you're staying are immense, because all of this money in the United States, in public markets and private markets, is being deployed on the premise that AI margins might be slightly lower than the traditional, extremely profitable US software business model that has 90% gross margins, slightly lower margins, but still a very profitable enterprise. You are saying that we're likely headed to a world where profits margins are extremely low, and it's something of a communitarian co-op model, an open source model. And the consequences of what you're saying, Michael, is that the trillions of dollars being spent by the US is basically just cash and generation, and that a lot of investors are going to lose money. Obviously, it'll take time to pan out, and I'm not saying it's going to happen tomorrow. But I did my PhD thesis ultimately on the concept of commoditization. And I am able to recognize the traits that indicate that a particular product or service might be being subjected to the forces of commoditization. And I can now see those forces gathering both on the hardware side and on the software side. Yes. And you come from South Africa, which is a dominant player in commodities. It is no surprise that the South African stock market is a tiny fraction of the size of the US stock market, which has commodities, but also has these things. So commodities, it is hard to make money from commodities. And they certainly do not command 40 or 50 times earnings multiple. And South Africa is an example of that. But there are plenty more commodities in the world than simply those that are like wheat or metals or any of those, or even there, I say, some of the energy commodities. There are plenty of other commodities that have existed that have come into being. I mean, I would say that petrol-based automobiles are on the verge of becoming commodities. And so the concept of commoditization is not exclusive, as I say, to the traditional term that is described as a commodity. Yes. And Michael, so I've said what is in your piece about how it could-- and badly for US investors because of the consequences. But can I get you to say it? Yes. Essentially, what is happening is that China has realized that-- and part of this has come about by the fact that they've been subjected to certain embargoes or controls that take you from the United States. China has been forced to look for another way. Another tower, as I sometimes like to call it, because the tower is the other way in China, in China and in China. And what's happening is that necessity being the mother of invention, because they didn't get those immediate chips, they've had to find other ways of doing it. And as I say, they thought smaller. They thought smarter, both on the software and on the hardware side. And what DeepSeek did last year-- and what I-- it's going to do this year-- we've already had a foretaste of that with their latest paper-- is essentially challenge the margins that exist on the software side of the business. Sun Su would basically advise any Chinese general. If you don't think you've got the right number of forces to be able to reap the enemy, change the battlefield. And there are plenty of almost Sun Su pieces of advice that are now playing out in the world of AI. As I say, there's going to be no gunfight at the 3NM or 2NM corral between Chinese chip makers. And let's just say in video, because the Chinese aren't going to fight there. They know that's not a gunfight. They could win. But they are now shifting the battle and shifting it dramatically. And when you ask-- there I would say it's stuck in the US ecosystem. I don't think you can imagine that there might actually be another battle for you. But there is. And the rest of the world is catching on. And it is starting to spread both software and hardware, though the Chinese don't have a lot of chips to spare to export at the moment. But it's come to 2028. The foregast is China will be producing more chips than it needs. And it already dominates the world of what's called commodity chips to go back to where we were before. You may have remembered the story with NXperia, which was a Dutch company that got essentially shut down by the Dutch authorities in the heads of the Trump administration. NXperia produced, quote unquote, commodity chips were the particularly auto sector of Europe. And this brought the auto sector to a stance. The point being here is that the low value, the margin, chips in the world, the Chinese are ready to dominate, probably, up to the sort of almost two-thirds of the actual chips of lie in the world now, has quote unquote be. And I don't want to say commoditized at the point where it actually can't create a profit. But where the margins are very, very thin, you may have to be readably efficient if you want to stay in that space. So what's happening is that we're seeing this essentially mulch up the value added ladder. And the Chinese are now moving, I think, to the next stage, which is the chip value chips, 14 NM to 18 NM, that essentially have huge uses across the world. I mean massive use of vehicles, your cell phone towers. I mean, the areas where those sorts of chips are prevalent are almost too many to mention. In the world of-- let's call it the latest Apple iPhone-- yes, you want to have one of those tiny chips. But there are many more applications to chips in the world. Now, I'll give you an example, which is not being wholly recognized yet. But chips that are being embedded in the smart factories of China so that these smart factories can actually operate almost remotely. They don't have to have people. And those chips are not necessarily driven by constraints about size. Size is not always the issue. But they need to be able to perform the function. I'll get back to my earlier comment. They're good enough. The result is that smart factories are just spreading across China as an extraordinary rate. I mean, just to understand it, last year, China ensured more robots than the rest of the world together. So we are seeing this process take place across all sorts of areas, which are not necessarily again only acknowledged in the United States. Now, I'm being a little cruel here and forgive me. But they're not being acknowledged because there are not many factories left in the United States for their chips to be embedded in. And China is now with, by 2030, 45% of the world's manufacturing production versus 10% of the United States. The Chinese are essentially moving their industrial structure over to smart chips, smart factories. And they're not these 3NM chips that you're going to find in the world called iPhones. So China has, as I say, it's a different part. It's a different road. And it's not funny to recognize partly because, particularly in the United States, where consumption is 80% of GDP, or services is 80% of GDP. Most of the AI talk that comes out of the United States is related to service, a genetic AI that is essentially going to allow you to do, to buy an air ticket on your phone. These are what get all the talk in the United States. They get talk in China. I'm not saying that they don't. But there are plenty other conversations taking places about where chips can be used embedded in drones. It's an extraordinary degree. Chips are being used to run the electricity system. It's an incredible degree across China. The solar system, the wind turbines. Chips are-- the application for chips in China tends to be a much longer list than the application of chips in the United States. Thank you, Michael. So we're recording January 8th, 2026. A little less than a year ago, in late January 2025, deepseek a Chinese AI model company that was actually-- I think started by a Chinese AI hedge fund manager of all places, of all people-- launched their deepseek model or something, R1. And that model and this fear that emerged a little less than a year ago caused a mini one or two day crash in the US semiconductor supply chain stocks. Pretty good idea if I remember was down 16 or 17% at one time. Now, on January 1st, 2026, so the first day of the year, while everyone was partying in the West, and the markets were closed, you're saying that they have released a new model or a new paper that could have similar consequences? Tell us about this. Well, I think it's essentially-- and what deepseek is doing at the moment is a sort of dance of the seven veils before it actually drops R2, which is going to do in my prediction just ahead of the Chinese New Year, but starts at the 17th of February. But essentially, there have been a number of releases. And this is the last release, I suspect, big release before that date, and the biggest by far. Because what they've done is they found a mechanism for essentially attacking the whole idea of memory in the training process of chips. So now they have found a way where previously, a particular chip produced, let's just say 100% of memory, that's saying chip now. If properly arranged within the software, you only need 7% of its power to produce that 100% of the previous they you were able to get. So essentially, they have increased the power of a small chip by 15 times in the training process. And then this essentially leads again towards the commoditization of i-value chips. Because they found a way round the whole idea that you need our chips because you need 100. And what deep seeker said, well, actually, you only need 7% to do that. Because that 7% will give you 100. And it's all about the new phrase that everyone is talking about, architecture. And architecture is happening both on the side of training chips as well as on inference. In fact, until a month ago, people didn't really talk about architecture and the side of training. Yes, it was something, but they did talk about architecture on the side of inference in the Chinese, not just the Chinese. This big company which has just been bought by a video in Singapore Manus is very good at architecture, but it's inference architecture. What deep seek did this time was come up with an unbelievably radical way of reinventing the architecture on the training side of the chips, which is largely being explored up until now. Now, why I think this paper, which is part of a whole, as I say, is 7 veils, is significant, is that it's setting things up for the release of Part 2, which, as I say, is probably going to happen at the start of the thing in Chinese New Year. We had another indication, for instance, in mid-Semba, which only the geeks really picked up on. But the Chinese deep-seek produced something called, I'll get it right, the B3.2 special, which was essentially a mathematical standalone model. And it basically went to the top of the benchmarks. Not every one of them, but nearly every one of them went straight to the top of the benchmark. So what we're seeing is that once you put all these 7 veils together, and I'm not going to bore you with all sorts of acronyms, as to what each of those veils constitutes. But you put all of those together, and then tie them up in what is going to be coming out, I suspect, in middle of February. And that is going to be monumentally significant, because deep-seek is going to, I think, in most areas go to very near, if not at the top of every benchmark, the counts. They're going to have, and you may not know the name of the game, is the number of parameters. They'll have an excess of one trillion parameters. They will have this, what's called, mixture of experts, structure. They will have MLM, which is the paper that would just be talking about dropped on the UNZ. And they'll put all of these together to create an unbelievably powerful model. That is powerful both on the training side and on the inference side. And I think intentionally, have you need even greater effect than all one when it was released a year ago? So there's two things, training and inference. Like it'll say, the old school, creating a computer that can be human beings in chess, which at UNZ existed for close to 30 years now. But training is the process of getting the computer to learn chess and develop its strategies. inference is, okay, you're playing Gary Kasperov, now you have to actually run. Western US models of training has been extremely capital intensive. And the Chinese models appear to be far less capital intensive. And that's why Nvidia crashed 17% when this deep seek news came out in late January 2025. Because, oh my god, they don't need to spend that much on Nvidia chips. I know there was some doubt about that, Michael. Is it really true that they only spent, you know, a tiny fraction of Nvidia chips? I saw a paper yesterday, said that the deep seek actually spent 1.5 billion. First, I won. And even if it did, open AI alone is spending its own equity contribution, not co-contributions from other players towards target is 19 billion. And yet, what deep seek came up with is far more radical than anything open AI has ever come up with. I mean, on a scale of 20-30 times. And it did so. Let's take it to the worst possible extent, 1.5 billion. I don't think it was anything near 1.5 billion. But I'll accept that, that, that, that podcast. A lot of money, but a tiny fraction of what the US company was spending, a tiny fraction. Tiny fraction. So the point being is that once our one drops last year, they gave the model to Nature magazine in London, which is one of the most prestigious magazines in the world. And over eight months, Nature tested that model. And in their September cover issue last year came out and said, every claim that deep seek made as to what our one could do was verified. Now, since then, as I say, they've dropped a number of other upgrades. And you've had to be really buried in the whole process to see each of the incremental upgrades. I mentioned the one that happened in December regarding the mathematical capabilities. But I think what's happening now is that there is a cumulative effect of all these upgrades that're going to be rolled into R2. And I don't know how much it's going to cost them to get cost to that expense. I really don't. I'm somewhat skeptical of the claim of 1.5 billion, because I'm not sure that the hedge fund that owns deep seek had that sort of money to spend. Unless someone was handing them some money slightly, you know, through the back pocket. But in a way, it's irrelevant. Because none of the claims as to what these models people are doing are being disputed. They're all showing up in the benchmarks. And they're all being subjected to unbelievable peer review. And anyone who's looked at that paper on the 31st December last year has come back to it. Yep, the claim is absolutely spot on. They've done it. They've found this way around this whole problem that we've all been facing for a long period of time, which is what's called catastrophic forgetfulness, when you're training a model and you get up to a certain level. And then you add more data into it. And it just gets everything that it's already learned. And they've essentially created a very stable way for that model to accumulate and to sort out information and keep it properly organized so that they continue to add what we call scale data into that model. And the result is they now have a very, very clever, stable way to grow the database of a model. And the result is, I think it's pretty radical to be perfectly honest. And there are a number of geeks out there, and I'm not a techie. But I've read the papers that I can understand. And virtually all of them are pretty much confirmed. Later on, I've got some potential pushbacks about your thesis. But I want to get into the nitty-gritty where you're not a tech geek, but you've kind of become a little bit of a tech geek. You say the three assassins of the US AI, our architecture, the US. I mentioned just calling it chips generally. Chips generally. Yeah, and potentially the bubble in AI in US private and public markets of US AI. The three assassins you say are physics, material science, and economics. Let's begin with physics. Why is that an assassin of chips? Well, you get down to a certain level where essentially the process by which the chip operates in physics becomes unstable. You have basically switches that are either on or off. But the electrons that control those switches are able to slip through because they are so small they can slip through and essentially turn that chip into not an on or off, but a maybe. And that basically starts to question the robustness of them off. So what it does is when you're getting down to these incredibly small things, the actual main look like a piece of steel or hard polyps and silicon to you, but it's actually got little gaps in it and these electrons are finding way through. They use the expression like ghost through a wall and they move through and go to the other side and then essentially turn that particular switch into a maybe which really starts to question at what point can you continue to miniaturize everything and still have the security of knowing that when I want that switch to say off, it says off, it doesn't say maybe. You're saying that chips are coming up to a limit, a radical limit, chip sizes are getting so small that basically electrons are going crazy in there and it gets so hot that you need a lot of the chips to be devoted to cabling to control the thermic output so it doesn't overheat and ruin the chip and I will note, Michael, not you, but there have been haters of Moore's law over the past 20 years who have said Moore's law is dead, Moore's law is going to this drastic two years, every two years, the amount of transistors that we could put in a semiconductor roughly double, which is held true since the 1960s, that's no longer going to happen. I want to say Moore's law refers to the number of transistors in a chip. It does not refer to compute power, actual compute power. People compute power has way more than doubled over the past 15 years because of Nvidia and because of parallel computing and the fact that all the electrons are going at the same time, the wires are going at the same time and Nvidia invented that. So, you know, Jensen Wong CEO of Nvidia has said that Moore's law is dead in the other way that like computing power has way, way, way more than doubled every two years because of that thing, but you're saying you're critiquing, saying finally like Moore's law is going to be dead and that you can't double the number of transistors every two years. It's simply getting too small. It went from 30 nanometers to 15 nanometers to 7 nanometers and now the latest Nvidia Black Well is 3 nanometers. You're saying, obviously you can't be 0 nanometers or negative nanometers. That kind of, there's not a ton more juice to be squeezed out of that lemon. We talk of diseconomies of scale that's essentially diseconomies of physics and diseconomies of material science. But at some point, and you can't divorce this entirely from the cost of being able to achieve this, at some point it becomes prohibitively expensive, very expensive, without that much uptick in what you mentioned compute to move from 3 nanometers to 2 nanometers. And there are complications that start arising in physics, in material science, and let's leave but one cannot leave on one side the whole issue of economics because ultimately that comes in and that's probably where I come from and spoils everything. So what we're doing is we're seeing the diseconomies of physics. We're seeing the diseconomies of chemistry. And yes, there are potential workarounds to use that wonderful phrase, but they are unbelievably expensive and you asked for instance, and one critical glad to ask in this whole process is ASMR in the Netherlands. When you carry on making your UV machines, so much so that you can actually start producing 2nm chips and they will say we can, but it's complicated and it's not just complicated. It's almost prohibitively complicated to do so. People have talked about changing things from silicon to something else, there are all sorts of areas and one of the most interesting areas potentially is photonics, although I should hasten to add here is that China probably leads in the whole area of photonics, which is the whole idea of embedding data in light itself, which is a completely different way of thinking about chips. I mean it's just moving it through a whole different space. And it's still by six, seven years away from having anything that's remotely practical. But in terms of research, China probably leads in photonics, it's not an undisputed plane, but it's probably a plane. And so that is essentially saying, forget silicon, we're going to move to post silicon world, but while we still live in that silicon world, we are seeing these diseconomies of physics, diseconomies of material science and diseconomies of all economics. And to some extent these three assassins are working together, not consciously obviously, but there is this sort of strange cooperation that's happening between all three that are making, as I say, and I'm not going to say that more is dead or more is dead, but he's on his deathbed and these three guys are standing around that deathbed, so rubbing their hands saying, you know, your time is up, mate. And what the Chinese say, well, listen, we're not going to have a gunfight at the three and then Keral, it's just becoming unbelievably expensive to play at that game. And we're not going to win, because we just don't have the EUV machines that come from ASMR to be able to play at that game. Let's move the battle field. Let's fight this war in another space and using other other sort of mediums. And this is why this so called SIP system in process of these cognitive towers, is that it's not coming something which everybody doing in video is in episode. So Michael, the game at which US chip makers, primarily in video, have excelled and dominated and crushed the opposition, that game is making chips smaller and smaller and more efficient. You're saying that that game is something we're trying to say we're not playing anymore, and that the advantages in computing have mostly come from making chips smaller and smaller and smaller and smaller at pillow computing, of course, you're saying that in the future, the gains are going to be made from connecting the architecture, connecting the chips themselves, something called advanced packaging, so allowing the chip to be 3D, and that whole system, so the chips can talk to each other rather than just the most efficient chip. Because the most powerful chip is no longer you're saying going to be what is the driver of effective usable compute, which is what it's all about. And you note in your piece that the former Google executive Eric Schmidt said that the constraints of AI is not chips, it is our electricity. And that's true. I mean, when you see the amount of power that's going to be required to run the likes of Stargate, and you've seen the pictures of the maps of where all the data centers are being put up over the United States, and you don't live in Northern Virginia, but if you did, you were facing some fairly severe power shortages in the next five years, because of all the data centers, many of the military related that have been and are continually to be erected in Virginia, Northern Virginia. But there are other areas, there are five or six, what I call hotspots all over the United States, where power issues are going to be very profound, but it's not just about power, though I completely agree with what Eric Schmidt had to say. I think that is the Achilles heel, the black swan, whatever phrase you want to use, the potentially threatens the US AI model. And remember that China does not base this constraint simply because they have invested absolutely massively, in particularly renewable, but not only renewable energy. I mean massive, and this is allowing them to not think of energy as a constraining factor at all. They are able to do whatever they want to do, and they don't have to think about energy. In fact, the price of energy has actually been falling for the Chinese. So their particular model, which is what we call distributed intelligence model, as distinct from concentrated intelligence model, the sonnified by the likes of Stargate, the Chinese one is far less power hungry anyway, and they said that they need that power they have. If you go to a Chinese conference, it doesn't have to be an AI conference, and speak to all the geeks that are there, the last thing they're going to tell you about is or we worried about our power supply. The last thing, go to a US equivalent conference, and almost the first thing they're going to talk to you about is power. What about the second assassin, material science? It's somewhat related, and I always think that there's a fairly thin line between the physics and the chemistry, but essentially it's about degradation of the materials, that at these incredibly small levels with all the heat that you were mentioning rightly, you're seeing the material starting to break down. They're starting to corrode, if that's the right word, it probably isn't in the context, but it's something you'd like to appreciate. How's that, how about that? Well, appreciate it, yes, but that's we're getting into the language in Michael Barry there, but yes, or let's call it depreciate. In the sense that they're no longer useful, if that's what you mean by depreciate, yes, I completely accept that. They're no longer useful, and Michael Barry will say that a high-end chip has three years of a useful life. Amazon will claim it's five. I don't know. Accountants are going to be forcefully to follow the opposite line, but essentially depreciation, corrosion happens at these very small levels, and that creates all sorts of secondary issues, like as you mentioned, heat, and the result is the chip becomes less than useful. It starts to break down. What we call yield, which is the number of transistors that are operating within the chip at full strength, starts to fall fairly dramatic. And so it's all about essentially the corrosion of the metallic properties that exist in the chip, particularly, dare I say it, so they're all ways again, of buying time. I think it's patronium, or something like that, they get coated on the chips. It's one of those periodic table metals that you and I never got down to, but it's nevertheless, it can buy a little bit of extra time. But the point being is that we really are fiddling. It's like, as I said, we're using, it's life extension drive, if you want to think about it in that context. But it ain't going to last for long. And this is where we get to that third assassin, Michael. When I just add there, I went to my essay, and I'm going to read one sentence, two, but it basically says everything that I've just said, but very, very technically. Transistors now require, and we're talking really here about materials, science issues. Transistors now require ghost proofing, such as haphanyum oxide layers. But by 2NM, even these are just a few atoms thick. One missing oxygen atom causes a short circuit, an even deposition of that haphanyum, graze gaps that invite electron tunneling. So you're seeing one I'm talking about here is where we're reaching limits of science, both physics and chemistry that are starting to make making things smaller, incredibly difficult. And this is where the third assassin economics really comes in. You reference Michael Burry, who has reemerged and has made the following critique that the companies that are spending massively on these chips, they put the capital up front, and that is not recorded as a loss at all. It's that neutral. The cost comes and is depreciated over the weighted life. So if it had a weighted life of 100 years, every year, and it's in a report, that cost would only be 1%. If it had a weighted life of two years, it would take a 50% hit in the first year and a 50% hit in the second year. The weighted average life of chips and certain data center investments is minor standing. From like 2019 to 2021, it had been three years. It was extended to five years or maybe six years for some companies. And it's my understanding that a lot of that was for old CPUs that actually was completely legit, like three years was too short. And so it was the, it had been wrong and it was the correct thing to make it longer. The critique is that for these newer GPUs, because the transformation and the innovation is so rapid, you know, five or six years is not that relevant. And the idea that in five years, these chips are still going to have, you know, as a serious value, is something of a joke. I will also point out that, you know, Michael Burry, very smart investor, but technically, you know, someone who had been saying that a lot earlier than Michael Burry is Jim Chano, so the short seller noted for his shorting enron and being earlier there on YouTube's Deplug, I did interview him in December of last year about this very issue. So we can link to that. And people definitely should should check this out. And then also this advanced packaging thing. I interviewed Sotrini about this researcher known as Sotrini about it. And basically it's everything you're saying, that the scaling laws and the improvements are going to be coming, not from the power of the chip itself, but from the interconnection and the architecture. And basically, so that, you know, so that you maximize the effective compute. That's nothing I can say to dispute. I agree. One hundred percent. I think, you know, Michael, this issue has been raised to Jensen Wong and Nvidia CEO. And he has said that with every new Nvidia chip, it does get far more efficient and that the, you know, amount of energy it takes goes, goes way down per chip. To what degree is that a, you know, a fair pushback and a justification of the Nvidia's model, or do you find the issues with that? Look, I think that he is speaking correctly where it when he comes to capability. But as you probably know, the cost of each of those chips and the next generation chip is rising as a percentage faster than the useful compute that those new chips are producing. Now, as an economist thinking about that, it basically says we're heading towards some sort of crisis point where you can't just, regardless of cost, continue to improve the chip. If the thing that you really want it for, the usable compute is not rising at a commensurate rate with the technology. And this is where the economics really does come in. The diseconomies of scale derived in part from the physics and material science that you and I have talked about is now starting to waive very, very heavily. And I think that's where Michael Barry is coming from. I haven't seen that. I'll look it up. But there's a lot of other people that have said this as well that we're moving into a world where that monolithic chip that Nvidia beat so famous for is unlikely to continue to rule the Bruce so much. And the replacement of the Nvidia chip that dominant Nvidia chip is not so much the new dominant player AMD, a competitor to Nvidia, but rather a custom ASIC chip. Or sorry, are you saying that basically these, the companies that are building the data centers and creating the models are going to be making their own chips and probably hiring a company like Broadcom or MediaTek in order to make to make their own chip rather than just buying a chip from Nvidia or AMD. I think that's absolutely right. I think what Amazon is doing with Trainiab while Google is doing with its TPUs are particularly interesting, but you can buy them in from a third party, yes, but they're actually doing it in-house. And the chips that they're building designing are not side by side as powerful as the ones that Nvidia produces. But they're built for purpose. They work for what Amazon needs them to do. They work for what Google needs them to do. So it's a bit like buying a truly magnificent Mercedes-Benz that's got off-road capability. But the reality is that you're basically going to drive around town. It's just not needed to have that off-road capability, but it adds huge amounts to the cost. And what Nvidia's come up, sorry, Amazon and Google, and I'm over simplifying here, but they come up with a chip that works for the specific needs that they have for that chip. Now, a lot of these chips are being used just on the training side, but on the inference side. And this is something which, by its behavior, Nvidia has started to recognize. And then moving over to chips that are more geared towards achieving success in inference, but also the software that's required to get the best out of those chips, which is why they fall rock with a cube. That was precisely a recognition that Nvidia is now stopping from being simply a supplier to other big chip-based tech companies to actually becoming a player. It's building its own stack from hardware through to software, and so essentially, it's starting to shoot itself, I think, in its own revenue for it, because it's starting to compete with its best customers. And that is something which can only go on for certain period of time. I mean, if I was Amazon at the moment, if I was Google at the moment, I would just say to my chip development department, "Full speed ahead, guys." We can no longer rely on Nvidia because they're actually trying to become competitive to us. And so I think there's a very interesting low-level civil war breaking actually in the United States at the moment, between the big players in the world. What do you think is going to happen to the US model providers? So not talking about Nvidia, but I'm talking about OpenAI, I'm talking about Gemini of Google, I'm talking about Anthropic, as well as the other, let's call them lesser players. Where are they going to be in three to five years in your view? On the spectrum from they have a product, it's modestly profitable, somewhat of a success, but not flights out versus this company is not going to exist anymore. Well, I please don't think I'm trying to be a stock emoji here, but the model I like most at the moment is Google's, because they have, and I think they're doing something which is, again, happening at the very early stages, but they appear to be essentially courting Apple at the moment and bringing Apple in as in the ally. The great thing about that is that it can't be seen as a competitor or from a native trust perspective, it just can be seen as an ally. So I like what Google's doing, they have a very powerful model. Gemini is a very powerful model. They have an unbelievable distribution capability. They're actually still technically own Android, and now they are quietly cozying up to the other great phone-based software company, Apple. So I like what they've got most of the piece of the jigsaw puzzle in place already. They still need to work hard on all of them, but nevertheless, they seem to be putting it all together, almost better than anyone else. Of course, Nvidia, as I said, has broken ranks with its old model and is now trying to do all of these things as well, but the orphans, and I always think of it as an orphan, I think they're going to have a tough time with it staying independent. I think, unless OpenAI has the likes of Microsoft behind it, I wouldn't be and of course, I suppose the big Japanese companies that are supporting OpenAI, but I'm not a huge fan of OpenAI. I don't think it's going to be a winner in this setup. I think they're essentially taking on more capital costs than they were able to generate sufficient revenues from. So I think OpenAI has got its work cut out for an extraordinary degree. Amazon is an interesting pair at the moment, it's doing what Google is doing, the Trainiam chip, for instance, but they don't have the consumer reach that Google has. It doesn't have a browser like Google. What is Amazon's model? They are starting to make their own chips, but they don't have a model. I think they own a little absolute no, but Amazon is, as I said, it's interesting, and I think probably I'm making a prediction here, but it may be Amazon will buy a product and then suddenly it will jumpstart its model. The point is, Amazon is essentially now offering a data service, but its data centers are increasingly being offered to third parties, and it's doing that with its own chips. But all I would say is that, you know, not nothing compared to Google, which I think is really a great job at the moment, but I think that Amazon has got an interesting one, and they would be, for me, a potential acquirer of one of the models, the orphan models, as I like to think of them, that we're talking about. I mean, for instance, another one out there is Meta. I mean, Meta has had a disastrous year in my humble opinion. I mean, the whole Lama story, which was, and using the appropriate euphemism here, put out to brass in August last year, Lama is an orphan now. It's a good orphan, Ironicate. It's an open source all, but nevertheless, it hasn't been improved, not that we know of, in any material way since August last year, and Mark Zuckerberg is going down at the beginning of the path now, and I'm not exactly sure of what that path is. But Meta would be another company that we, on the basis of current behavior, be struggling in five years now. So, Amazon does not currently, I believe, have their own model, or any model that is, you know, serious. So they are a huge cloud provider, they're, you know, the first really large cloud provider, Microsoft is, and increasingly, Google as well, you know, the cloud computing is a profitable business, and quite growing rapidly. In particular, it is growing rapidly now. But how much of that is because the customers are, is open AI, and all of these other unprofitable AI startups. So everyone says the demand for compute is so high, the demand for compute is so high. So, you know, data set, and that is demand for people becoming customers of data centers. But how much of it is, you know, real and sustainable? You ask a very profound question, which actually leads us even back to a video, and video might be immensely profitable, and it is immensely profitable at the moment. But are its customers profitable? And so, I'm sorry, sorry, an amazing question, I want to say, technically, a ton of its customers are immensely profitable, like Microsoft. Yeah, but the customers of its customers are not profitable, that's the whole, although, and to the extent that its customers are profitable, they're not generally speaking always very profitable from their artificial intelligence activity. I mean, to the extent that its customers might be doing well, they're able, meta is able to subsidize its activities in AI because of the advertising that it gets from Facebook. But if you actually look at, if you can compartmentalize it, when I ask the question again, how profitable are, in videos, customers from their AI activities? It's a much more complex question to answer. They've got associated areas which can subsidize for now, those areas. But one of the interesting things is that we've seen a lot of the companies that are being one of the, move out of the fact that they could finance the AI activities from free cash flow to now having to borrow. Again, a slight warning sign, echoes of 1990-1992,000. I don't want to make too great a parallel, I'm not Michael Barry, but nevertheless, a warning sign. The amount of debt that's creeping into the system, both on and on balance sheet at the moment, should be of concern. It's particularly of concern in related areas like, there I said, Oracle and Huawei. The point being is, it's part of the ecosystem. So what has to look to some extent of the health of the ecosystem as a whole, though one can recognize that there are parts of that ecosystem that ostensibly are very healthy at the moment. Yes, and Michael, it has been said by others, as well as I have said, the following statement that the amount of the money being spent on AI, the people who are building the data centers and buying the chips are the most profitable and largest companies that have ever existed. I stand by that claim in its technicality. I want to add the caveat that the customers of those immensely profitable companies, named Microsoft, Amazon and Google, those customers are often VC backed companies that are losing a bajillion dollars a year. That's a technical term. So the customers of NVIDIA are making money. The customers of the customers of NVIDIA are not making money. I'm having to be, I go with your qualification. Yes. And the example, you said of meta buying a ton of NVIDIA chips in order to make its own process better and to serve ads with AI ads, that is a somewhat rare scenario. I think a lot of it is cloud computing that is profitable, but the customers of that cloud computing are losing a ton. I'm happy to accept your qualification. No, I'm not going to. So you make a lot of military analogies, and you basically compare the US architecture and NVIDIA to the German tanks during World War II, which were extremely effective tanks. It's just that the German economy and industrial powerhouse was unable to make enough of them compared to the Soviet tanks and the American tanks that the tanks were maybe not as powerful, but they were able to produce them at scale and ultimately led to defeating the Germans, thankfully. Tell us about that analogy. There's a very famous infamous apocryphal story of a rather put out German tank commander who said, "One of our tankers is worth Port Sherman, the problem is the Americans always bring by." Now, that saying has been subjected to scrutiny and it doesn't hold precise water, but the concept, everyone agrees, that in the end, if you can wovenize enough material, you can overwhelm people who have one-on-one better piece of equipment than you do. And this is something which the Chinese are essentially doing now when it comes to chips. And even David Sacks in the White House admitted as much to this, that China doesn't need our chips, because what they do is they essentially amass so many chips from far away, that they can out-shoot in terms of useful compute to your earlier term, an imidia cluster, which is really what we're talking about. So that while we're super-classists, the imidia cluster, there are just so many more Huawei chips in that cluster, and the net effect is that it out-shoot the imidia cluster. And that's what has started to happen in China. Now, it's not fully operational at the moment, and this whole, will they want a story that coming out of China at the moment with regards to will they allow the H-200 or the H-100 to be appositive from imidia, is part and parcel, it's caught in the crossbar, I mean, to my measurements, but not entirely in this whole process at the moment, because the Chinese feel that close, my own estimate is that come from 2028, they will have met with parity, are being able to match, perhaps on scale, if not on quantity, anything that can be thrown up by the likes of the imidia. The amount of effort, of money, of resources that are being mobilized to the producing of just huge numbers, or chips in China, is such that while it's touching go in the comparison today, and David Sacks may be right, or it may be wrong, or maybe technically right, but not right, but nevertheless, in two years time, the absolute will be right. My own view is that what China's setting itself up to do is to tide itself over. It basically, the by-time, probably two years, and it may take a dollar of H-200, H-100 from imidia, the 2026 and 2027, but by 2028, it won't need those chips any longer. And it's not that they won't be able to produce better chips, just that they're going to be able to produce massively full. You have several addenda in your piece. The first is a story, which I love, takes me back 20 years, when I read the story, as a child, of the Indian tale about the king that said, "I will grant you any favor," and the guy said, "Give me a grain of rice on day one, on day two, double it, and then double it on day three," and basically, by the end of the month, or by the end of two months, it was a million trillions of rice. So, you know, in a possible number, I believe that is a quadrillion. I love that tale, the sort of compound interest. How did it apply to what we're talking about right now? Well, I just wanted to, essentially, one of the problems that happens often in the whole area of talking about AI is that we get grounded big numbers. So, I essentially wanted to go to one of the biggest numbers I'd ever see, which is the number of grades of rice that will be on the 64th square of chessboard, and essentially use that to explain a story. In fact, it's 1, 2, 3, 4, 5, 6, 7, 8, 19, 11, 12, 15, 16, 17, 20 numbers on that last 64th square. And the point that I was making here is that, to some extent, the numbers that are being talked about by the likes of open AI, every now and again, start to almost approach the same sort of level of absurdity. Eventually, you run out, you can't mobilize the amount of rice that's required to cover that 64th square. I came across a statistic and I'm just going to quietly bring it up. Yesterday, when talking about, we've mentioned it earlier in the context of what Eric Schmidt have to say, but of energy. And on the current level of energy needs that open AI has, they need to increase their energy capacity over the next eight years by a kid you know, 125 times. Now, these are the sorts of compounding numbers that eventually calls me to say enough, this isn't possible, you can't carry on like that. And the whole process starts to come off the rest. I'm sure that Indian prince or king eventually, when things were starting to get 24th square, I was saying, oh my god, I'm going to bankrupt this nation. You know, the number of grains of rice on the 24th square is just more than three years worth of production. And I think that that same sort of logic eventually overtakes the likes of open AI because the numbers, if they have to increase energy consumption 125 times over the next eight years, are they mobilizing or at least it's the system mobilizing the amount of energy supplied to be able to meet that demand. And I think that those, it was just a way of playing with the numbers, the power of compounding and to use it in this context and to say, I think the numbers being talked about here are just off the charts. And I'm not really. So Michael, a key point you're saying is that all of the huge, you know, AI bulls say, Michael, the revenue is going to grow exponentially and you're saying, of course, but the costs are going exponentially as well. And that is something that wasn't true of software and software, the everything, you know, the revenues grow exponentially, but the costs stay somewhat fixed over time. And yes, you have to depreciate the commissions you give to salespeople, et cetera, et cetera, et cetera, et cetera. But software, you know, as it existed in the US over the past 25 years is a ridiculously good business. I use another copy of Windows. It costs them an infinitesimally small amount of money. So what they can sell that for, which is as you and I both know quite a lot, compared to what it cost them to produce, gives them spectacular margins, making your phone. So Michael, now I want to get the point where I want to give potential counter arguments. Some of the counter arguments are going to be what if you're wrong about what you said, some of it else what I'll start out with is let's say you're right that the Chinese AI architecture is going to reach the scale of US and soon exceed it, especially when you adjust for costs and inputs. Why does that mean automatically that China is going to dominate AI? You know, I'm, you know, not a tech geek at all. I'm the furthest thing from it, but I know that many tech geeks, people who are tech geeks say that Linux is in many areas much more preferable to Mac or to Windows. And people say, yes, Android and all this open source stuff, much preferable to the iPhone and Android. And you started our conversation by noting the success of Linux and operating systems and the success of Android over Microsoft and Google, but the Microsoft and Apple are enormously profitable enterprises. And it seems like the power of American technology companies to extract monopoly rents or monopoly like rents from their software has been true despite the fact that yes, there are cheap alternatives. Whether in the case of Microsoft, it is through, you know, somewhat anti-competitive practices maybe or the case of Apple, it's in the fact that people just brand, people just associate Mac with good things and therefore they use it way more often and and, you know, maybe not in terms of the amount of people, amount of customers, but in terms of dollars and particularly profits, you know, Apple Apple is dominant there. Why isn't that going to be the case for you for AI? Look, it may well continue to be, but it's a closed ecosystem that we're talking about and Apple has got an amazing mechanism for essentially extracting rent from what is essentially a closed ecosystem. But it gets back to my earlier comment about not just the US world but the world world. The world world is seeing this differently and I just happen to live in one part of the world that's seeing it differently. And we're not just talking geopolitics. We're talking georeconomics. We're seeing what Indonesia is setting up a sovereign AI system which it's announced it will do and yes, it will get some technical support from some of the US players. But when it comes to the software, it looks as if it's basically going to go with where. And the point being here is that the software comes in inverted commas for free. Now, we know the reason why the top 100 supercomputers in the world use Linux is that Linux comes for free. That was not the case in the mid 1990s where basically it was Microsoft that was applying the software to what supercomputers were around at the time. But what has happened in the last 20 years, 25 years, is that the geeks that run the world supercomputers basically, and there's another dimension here, but it's incredibly important dimension, preferred an open source option which they then manipulate in order to build that supercomputer in the way that they wanted it to look. And this is critical that what we are seeing is that the usability of open source stroke, open weight, and open weight is not as flexible as open source, open source is unbelievably flexible, is that it will, it's going to win in the wider world. And if all you're saying is yes, but Apple will continue to be able to extract huge profits from its core markets. And by the way, I need to correct what I said earlier, there are a couple of extra markets in the Apple world, which is South Korea and Japan, I think I missed them out. But nevertheless, that six out of what does CNN say that there are 200 territories and countries. The rest of the world is moving very heavily towards using Chinese software because it's free. And you know, there is free is free. It's a very, very powerful model. The point is there are other ways in which people are going to have to, Android is the classic. There are other ways in which people are going to have to learn to monetize the language. And it is a language we're talking about here. It's like English, you know, you and I don't actually pay anybody attacks using English. But by using English, you and I are often able, I can go and make a speech. And thereby monetize my use of that English and give myself an income. And it's the same thing. It's essentially becoming a software language that's available or free. And this is where I think that the clash will ultimately arise. Now, what we may end up with, and I'm not disputing this, is a bifurcated world with these reinforced islands that Apple can occupy, reinforced islands that the USI can occupy, six or seven or eight of them. But the rest of the world, which of course, is growing economically much faster than those reinforced islands, will be opting for something else. And where does memory come into this? I know there is a giant squeeze in memory. The postdocs of these companies like Micron or Sandisk are up 300, 400, 500 percent. It really is ridiculous. You referenced earlier that deep-seek memory usage or demands are down 93 percent from 100 percent to 7 percent. How is it so much more efficient as well? The MLA architecture that they come up with, and really I'm not a geek, so I'm not getting into the specifics here. Essentially allows them to pack down the usage of the space far more effectively, far more efficiently, so much so that they don't need to crowd the memory with 100. They just crowd it with seven if crowd is the right verb. The point being is that memory suddenly becomes scalable, whereas previously it didn't. But I fully accept that in the US ecosystem at the moment, memory is a huge issue. In fact, it's also an issue at the Chinese ecosystem. I'm not going to deny that. But after MLA, not to the scale that it is in the US, which is why you're absolutely right to reference the memory suppliers. Those people who provide that capability, the empties of the world, you're absolutely right to reference. But the point is that, again, in this bifurcation world that I've talked about, they're finding a way of doing things different, and they're demanding on memory as nothing like, let's give it three years for MLA to roll out properly. It's going to be nothing like it is today compared to if the US does not go down this part of what it will be in the US. But I have to say, I think that a lot of, because MLA is essentially available on an open source, back for namely deep sea, most of the US, the Googles and the, or actually I'd say, the NEMATrons and the Gemini's and the Open AI's, they're going to be looking at what this MLA is all about, and most likely incorporating it into their own next generation software. Because it's architect, it's available, it's a way of doing things. It's not something that's essentially owned by the Chinese, not at all, anybody can use it. So I think that the memory crunch, which we are seeing at the moment, will continue to exist for a couple of years. But if I had to make a forecast, I'd say it starts eased. And so you said that companies around the world can use this Chinese open source architecture, and they're not sharing their data with China, because I imagine a lot of companies and people in particular would say, okay, I understand if I'm using all this US technology, that all of these American corporations have my data, and they're exploiting that, and I'm not happy about it. But it's better than the Chinese government having it, and you know, using that, and they're being some geopolitical fears there. Maybe I'm wrong, maybe I'm wrong about that. But are you saying that people can use these Chinese things and not share the data with China? Yes. Is the big answer? But I'll take it one stage further. If you've got it on your phone, you can be offline and still use the software. And so if you're offline, there's no way you can be sharing the data. And that's the extraordinary thing about the open source network, is that it's very distributed. I used that term earlier. They have this phrase in English called the edge, and essentially the connectivity of Chinese software to the edge. To my cell phone here is incredible. And the interesting thing about that is that one of the problems that USAI models are facing at the moment is called data exhaustion. They're running out of high quality data to scrape. Wonderful. And they're looking to create their own forms of data, artificial data, synthetic data, which they can do. But the reliability of synthetic data is question. I won't say not worth anything, but it is question. And hallucination, imagining things as software sometimes does, is said to be more likely to happen with a synthetic data than with real, it makes sense. The point about being connected to the edge is that the data that's coming in from the edge is real worlds every day data. And so what's happening is that the extent that Chinese software is updating itself, which you can do by virtue of it being open source, it's updating itself with a whole new supply of fresh air data that's coming in from the edge. Now, the availability of that data, though there are some interesting pieces of software being developed in the US ecosystem at the moment that potentially can mimic that. And so I don't say that it's going to be possible for US closed-weight models to somewhat emulate that idea. But as extent at the moment, the Chinese just do it easy. So they basically pick up on edge data to refresh their models. Pressure, I call it. That is not accessible by and large to the US closed-weight model. And so it sounds like you think a lot of US closed-source models, in particular OpenAI, are going to have a tough time. Let's put it that. You don't want to use the word screwed, but are going to have some challenges. What about Google? You said some maybe nice things about Google, and you said, I don't want to be a stock promoter. I think with this interview, no one's going to use you of promoting American securities, at least. But Google, I believe, uses something called TPUs, Tensor Processing Units, that they invented. And Google buys GPUs, graphic, prospectus, from NVIDIA, primarily for its external cloud customers. For itself, Gemini, which it owns and produces, it uses primarily TPUs, and perhaps exclusively TPUs, as well as maybe some CPUs as well. So it's made this, I use Gemini exclusively, and I'm extremely happy with it, and many people who know a lot more about me say that Gemini model is extremely powerful and good. Does Gemini avoid these pitfalls of Moore's Law and all these scaling things that you say that Western U.S. AI is quite with? Yes. I mean, it's too sharp an answer, but essentially yes, because the TPUs are not facing the same sort of level and the imperatives, you've got to get smaller and smaller and smaller. They're basically saying, no, no, no, just build me a chip that's a horse for this course. It's built for purpose, and it serves for purpose. It's not that four-wheel-drive Mercedes that only drives around on city roads. It's just maybe just a four-wheel-drive vehicle that only goes in the rural areas, or it may just be your little hatchback that runs around in the local areas. The point being is that it is not an old singing and old dancing Nvidia pool service chip. It's a much more restrictive chip in terms of capabilities, but nevertheless it serves the purpose that Google, and indeed I'd say that with the training for Amazon, but recognizing Amazon is not as full services as Google is, but nevertheless it's interesting to see. And what this is doing is it starting to violate the most sacred piece of intellectual property that Nvidia has, which is CUDA. They're essentially their moat called, that protects the Nvidia chip, and what ultimately gives it, dare I say, the value that it has. What you can do now, it's a much more limited service chip that is being produced by the likes of Google and the GPU. And nevertheless, Google is not left saying, "Overbook, we need to be able to do that as well," or to the extent that they are, they might buy those GPUs as an extra. But for the purpose that they are talking about now, that GPU serves that purpose. It's amazing how quickly things have changed. A year ago, people said, "Google is going to be dead because of AI, because it's going to replace their core business search." Now, not only is that probably not true, search is still very dominant, but it has its own architecture that in some ways is better than OpenAI, which, you know, and the Microsoft universe. So in terms of investing implications, Michael, how are you approaching this? Look, if all my universe was Wall Street, I would worry about quite a few of the players of the so-called Max7. I mean, number one would be, nobody we've talked about at the moment, that they do have some AI capabilities or potential to, of course, Tesla. But let's leave them out of the equation. I think that they are where it comes to eb's anyway, yes, to their store. All of the other big players out there, I worry about Microsoft because I think they've got themselves attached to what I think is going to end up turning out to be a dip in over there, which is very sad because I think, although I've paid too much money towards Microsoft every year to be able to speak to you on a day like this, but I think Microsoft has done well. And I think that they are showing signs that they might be trying to diversify away from OpenAI as their only area where they're going to have exposure moving forward. I think that they, as I understand it, they've got a nice deal, a sweet deal through two or something with OpenAI. And I think that they basically have decided that in that time they better come up with something as a precaution that is alternative to our exposure to OpenAI. But nevertheless, I think it's potentially short to medium term bit of a dead weight. I think Meta is lost. I don't think, I think Zuckerberg is casting around for all sorts of ideas at the moment, and I can't see them putting together a story. And I think the fact that they've just lost their best scientists who just walked out of the door, and you may have read it the next time to interview. But last weekend, and he basically had very few nice things to say about what's happening in Meta. Now he's a Turing Prize winner, so he is no, he's difficult manner to send very good work, but he's nevertheless a genius. There's no question about he's an absolute genius. And he had some pretty harsh things to say about Meta. So I think Meta is a bit of an orphan, and they better get their act together quickly. Otherwise, I see Meta in three to five years, time being a souped up search engine attached to retail options that go into Facebook and the other members of the Meta WhatsApp and the like. I really do think, I mean, that's the length of ringing is nowhere. Yes, and actually, I believe that so Meta's model is, as you say, open source. And so they're distributing it everywhere. So it's not just at meta.ai. But I believe the web traffic of Meta.ai was something like only 10 times more than my podcast. So my podcast gets one tenth of the traffic in terms of listening time, in terms of time, as meta.ai.ai.ai. And even the movie being modest about your podcast. So maybe that, you know, such a high level of traffic that you've got. That's how, you know, poor old Meta cup. My point is that it's not a lot, you know, it's not a lot. I understand where you're coming. Yeah. But no, I think I think Meta's got its work out and losing which is what Lee can said in his interview, and losing their way with Lama, which was something that was an incredible option. And then basically having, you know, been all in Lama until about over last year and then suddenly going silent altogether on it and not having anything to replace. And obviously, yeah, they haven't got a closed source hop unless they buy Anthropic. I mean, that's something that they could do, right? Yeah. Anthropic is a great LLM, but it's a bit of an orphan and you're not sure how they're essentially going to pay for themselves moving forward unless they have a sugar daddy like Meta behind. Right. And Meta does have a hugely profitable business. And it's investing in all of this stuff to AI to make its own product better rather than serve the compute out to clients who are other LLMs. That's why I think Microsoft and particularly Oracle might get a little more screwed than Meta because Meta at the end of the day, let's say you're right, Michael, sorry, let's say you're right. And you know, Meta loses tons of money. The depreciation is immense and the revenue growth is simply not there, not enough. And then the stock, you know, goes down 80% like it did again. I think the playbook is probably pretty similar to 2022 where it's a buying opportunity because it's kind of like if Coca-Cola was spending billions of dollars on trying to, you know, go to Mars or something, it's like, yeah, they'll just eventually stop losing money and they'll do it. You know, Mark Zuckerberg hasn't noticed it yet, but it's committed to a quarter of a trillion of lease obligations for its data centers. If you were, I didn't, wasn't including Oracle and Call Weave in our conversation at first instance because they're not in the max 7th. But if you were asking me, given the choice, would I buy Meta ahead of Call Weave or Oracle, I would buy Meta. If that's my menu, I would buy Meta. And you're damning it with fake comparison if that's a phrase that we come up with. And you're absolutely right. I mean, if you want to focus on where the problems in, if there is a bubble, and let's not get into that subject, but if you were on bubble watch at the moment, you should focus on Oracle. And I mean, you said you don't want to use the word bubble, but in the piece, you use the bubble several times. You say bubble bubble, toil and trouble. And you say, you know, maybe there's bubble in Oracle and Corey, which are publicly traded, the massive, the granddaddy of all the bubbles open AI, right? Which is private. Look, I did speak about it, and I spoke about the finances, particularly underlying Stugge, for instance. However, my central thesis at the moment is that the real bubble is technological. The financial bubble is a symptom of that technological bubble. And the two are interconnected. And that's not to say if one is to burst, which one is going to burst burst, but the relationship's almost unbilical. So that if one does burst, the other one will probably suffer. But my central thesis in that paper is that the bubble is technological. Where does that leave us with the Chinese investible AI universe? There are publicly traded securities. Ali Baba is the producer of Kwen, which we reference. Deepseek is a private company. There is 10 cent, which interestingly is a lot of which is owned by masters, you know, the largest South African company. And then there's tons of Chinese companies that I've never heard of. Most people have never heard of that are up over 100%. And I think a lot of those companies are suppliers to the AI infrastructure as well. And that would certainly be what I consider a high risk area. But do you have any bullishness? You see, you have a certain caution, let's not call this bearishness, but certain caution, you know, strategists are never bearish, they're only cautious. Certain caution over US securities. What about, are you bullish or optimistic about Chinese securities? I prefer the ones like Tencent and Annie Baba where there is something else involved in the company. But when Metax sawed 700% on its IPO debut, we're seeing some unbelievable first-day pops that are happening. I'm actually deep down quite concerned about particularly the new issues that are coming to market in the US, sorry, in China at the moment. And that's both on Hong Kong and on things like the Shanghai or the Chen Sen index. So I'm not just defining myself to Hong Kong, but be careful because I think there is a bubble potentially performing in some of these seconds here. And I don't wish to sound that I don't think that they're producing some great product for some of these seconds here. If it's chip-related, I'm probably a little less worried. If it's just software-related, I'm hugely worried because I think they will face the same best-to-pressures, the commoditization pressures that I fear were over in AI, which of course is moving into providing hardware as well in the form of data centers. But nevertheless, something that is standalone and just an LLM at the moment would be something that I would be very, very careful about. So Chen sent him Alibaba absolutely by-do potentially, which is a very interesting, I mean, it's the main search engine, and they're starting to develop their own AI capabilities. They're some extent doing what Google is doing. There I said, it's not a deeply bad comparison. I'm probably okay by-do at the moment, but some of the more recent issues with those levels of first-day pops, I am really quite concerned. So I'm not to, you know, whenever it takes commentator on Chinese stuff. That makes sense. What about the more semiconductor supply chain companies like ASML, LAM Research, KLA, these companies, which, you know, semiconductors have historically been a very volatile industry, you know, bankruptcies and the like, but more recently, you know, these companies appear to be much more dominant companies that in the case of ASML have a literal monopoly on certain technologies. And therefore, extremely profitable, is this, is your thesis also a threat to these companies in the same way it's a threat to NVIDIA and the oracles of the world? Well, yes, but perhaps for slightly different reasons. And essentially because- and the ASML has got two or three years of uninterrupted blue C-com C ahead of it, but the Chinese have essentially been able to come up with their own EUV at the moment. And it's probably 2028 where it will be commercially available and being used to make chips in China. The company there is called SMEEE and recently, it basically sold its non-EUV related activities out and essentially concentrating now on its EUV capabilities. But they essentially up to what has been called China's Manhattan Project in Chen Zeng and have built their own EUV, which appears by all accounts to work. Obviously, there's going to be some teething issues, I suppose, in putting it together. But as I say, they're probably going to have it up and running within two or three years. If that's case, then the ASML is no longer the only kid on the world. And that then starts to cause issues for the likes of ASML. I think with regards to other players in the hardware supply sector, you have to look at it on a case-by-case basis, see them up if they operate in, see what the potential competition might be, not just from China, but in first instance, from China, before you may comment that yes, AMD, for instance, is super secure. There are all sorts of other players that are now starting to come in. I mean, there's a very interesting company out of China, it's called Law Threads, interestingly using more in the same sense as Moore's Law, and they are nicknamed China's Envidia. Now, they are far from being a serious competitor to the Envidia at the moment. But the technology that they're working on at the moment is pretty amazing. And again, first of all, with three years, maybe four years. More threads can not take on Envidia, but it can start impacting its margins. And I think that that's the point that's how commoditization takes place. It doesn't necessarily happen on day one, two, three, or you're one, two, three. But eventually, when there is, an Envidia is pretty much synonymous with high-end chips, ASML is pretty much synonymous with EUV machines. But when there is another kid on the block, they no longer are a monopolist and they can no longer necessarily guarantee to extract monopolies. You write in the piece that the hyperscalers are US data center companies are trapped in prisoners dilemma. No one can stop spending unless, because the rivals will search ahead. So collective over-investment guarantees systemic collapse. Is it your base case that these companies keep on spending money into the ground so that CapEx in 2026 is way higher than it was in 2025. In the same way, CapEx in 2025 was way higher than it was in 2024. And that at some point, there is going to be, as you say, a systemic collapse, a giant bust where the markets and investors and the world suddenly realizes that these investments were to put it mildly excessive. I wish I could say no. What you described, and I didn't mention it in the essay, but it has also been referred to as what's called the Red Queen Dilemma from Alice Wonderland from Louis Carroll, where the Red Queen says, "I have to run twice as fast just as they are in the same position." To some extent, when I listen to some of the hyperscalers speaking in the moment about their CapEx plans, that seems to be the logic behind what they're doing, but they can't stop now because they were followed. It was fascinating to look at what happened after meta-release, this latest set of earnings. I think the stock went down 20%. And the reason why is that people were saying, "Ping on a second, this is just because they're ridiculous." Have you seen those CapEx plans? And there is a sense that at some point, and what has to think about who are going to be the survivors? Some of the dare I say, "We compares will, probably, fall by the way, so." Now, whether that impacts negatively on those that continue to operate, I can't say, it probably does, but they will be survivors. I want you to go back to what happened in 1999. Amazon saw it's correct me if I'm wrong. Price is fall, 8% in the dot com bubble burst, but it was a survivor. It lived to bite another day, and it found something called Amazon Web Services. It's a new product. I think that there will be casualties along the way. I think that those that remain as survivors will be negatively impacted, but they will live to bite another day, and it's going to be fascinating to see which are the survivors. And to some extent, your earlier set of questions could have actually be my answers, could have been essentially, so who do you think the survivors are in? And I think that's what, if you're a serious investor in the US market at the moment, you need to start asking yourself, just as a precautionary exercise if nothing else, so who are going to be the survivors? How are you assessing the odds that, let's say someone's equally invested in the oracles and the US AI of the world and the Chinese AI of the world, public and private? Are they going to lose, I think you, based on your thesis, you probably think they're going to lose money on the US investments and make money on the Chinese investments. Are they going to make more of a lose? They're going to lose less on that Chinese. If you're talking about a broke cross-section on both sides that it includes the equivalent of Chinese coreweeps and Chinese oracles, there are some potential losers in the Chinese stack that is often used in the world of computers, just as there are potential losers in the US stack. Which one is most overpriced at the moment? The US stack. But does that mean that there are no overpriced options available inside China? No. And secondary, I mean, just like I'm moving into a different space at the moment, if you ask me to invest in Chinese EUVs, there are just so many of them out there at the moment. And they're all doing spectacularly well in terms of their technological capabilities. But I'm not sure they're all survivors. So I think there's going to be a thinning out in EUV sector and I think there will be thinning out stroke consolidation that's going to happen in the AI sector too. But who are going to be the survivors? Well, I've got my own ideas. I mentioned three of them, just not the Tencent, Alibaba, and Baidu definitely will be survivors. But will they sail through this without any scars on their face? All right. Now I'm going to ask another question. Not about just the broad ecosystem, but specific companies. There's a guy named Doug. Doug owns six stocks. He owns the USAI, which is, I'll say, Nvidia Oracle and Microsoft Nvidia Oracle Microsoft. And then he owns three Chinese stocks, Baidu, Alibaba, and Tencent. Is he going to lose more on his US investments that he's going to make on the Chinese investments or use such a good question to answer? I really don't what his in price was on all of them. But if we can say today's price is the starting price is in price as it sort of indulge your question, I would worry more about the US stock than I would about the Chinese stock. Thank you. Michael, it's been an enormous privilege to hear your views on this. You've put a lot of work into it. We will be attaching your paper, which can be read in full on your website. And you also posted it on LinkedIn. Please just tell us a little bit about some of the work that you do at Cascaldi says consulting you were a strategist for investing firm for a long time. Your main focus most of the time is not AI. Other than this, what are you focusing on and what kind of work are you doing right now? I sometimes feel as though I'm a deep seek in a sense that I'm not out to make lots of money at the moment. I'm just out to try and understand what's happening. And I get invited to speak at a lot of venues, some which I do remotely, like I'm doing with you today, others which I do in person. I have to say I say no to quite a few of the invites because I simply don't want to grab my life too much. So consider yourself on it. No, I don't mean that, but you understand what I'm saying. Oh, I am. I am. But I do speak. I do write. I'm a one man band. And so my research goes where I want it to go. I'm not driven by somebody says, right, I want you to look at this subject now or I want you to look at that subject. For the moment and for the foreseeable future, I will focus on AI as it manifests itself. And I particularly feel as though I have something extra to add in understanding the Chinese ecosystem. When I speak to really well informed people of the USAI ecosystem, I'm horrified by how little they know about the competition. So I don't know nearly as much as they do about the USAI system. But I know a gazillion time, not gazillion, but I'm multiple times more about what's happening in the Chinese system. And for me, for instance, the fact that many of them really know that Chinese AI software is offered for free. I mean, it's a simple fact, but it's a powerful one. Just tells me that they probably don't know what's coming, which is why I suppose I do get invited to speak to people like yourself, because I'm trying to understand what's coming. And I don't think I'm going to get it all right. I don't think I'm going to get it even half right. But I think I'm going to get it more right part because I'm actually investigated it seriously. I'm probably going to get it more right than wrong as compared to many Western analysts at the moment. So that's probably what I have to offer at the moment. And that is that I spend my days searching through and I can leave one piece of advice if people want to try and optimize themselves, get abreast of what's happening in Chinese tech. They should subscribe to the South China Morning Post, which has the best tech writers in English in the world. Partly because many people don't really understand it, Hong Kong is now, as it was one stage, very much the big brother to Shenzhen. It's now the small brother, and yet the two are two metro stops away from each other. And the real action is happening in Shenzhen, but also in Wangzhou, which is where Alibaba is based, and Deepseeker based. But Shenzhen, for some unknown reason that people at the South China Morning Post and I'm not paid by them, don't leave this. All I'm trying to say is if someone wants to pay ketchup and try to understand what's happening in the Chinese tech space, start reading the South China Morning Post. Yes, we will leave it there. Thank you everyone for watching. Please leave a rating and review for monetary matters on Apple podcasts and Spotify and subscribe to the monetary matters YouTube channel. Thank you, Jay. [BLANK_AUDIO]