301. Why the US won't stop China in the AI race

Now everyone thinks when it comes to AI, that China is lagging behind the US, not least because China's banned from using American chips. But there is a new impressive Chinese open source model on the scene called Moonshot. At the same time you've got Chinese President Chi, who wants everyone to join his international AI organization. So what's going on? Well, as Zim, as are, is just back from visiting Moonshot and a load of other AI labs in China. He's the founder of Exponential View, which is a leading research platform in this area. He's also a tech startup investor. So he's a man who follows the AI money. He's our interview with a Zim azar. Support for this episode comes from Octopus Energy and the founder and CEO Greg Jackson is with us now. And I want to ask about oil prices. Obviously they're very high at the moment. What's your advice for a company worried about them? First of all, if you use a lot of electricity, it may be possible to get electricity tariffs where you get charged less at certain times a day. And a lot of businesses have been able to benefit by shifting their electricity consumption. They still use as much energy. They just pay less for it. Of course, things like heating space can be very expensive and finding ways to do that more efficiently by heating to maybe 18 degrees rather than 21. Can make a very big difference. Because look, the real solution is that Britain needs to escape from this dependence on the global fossil fuel price. And that means more electrification disconnects our electricity price from the gas price. And ultimately get more of our electricity from British resources. Nice one Greg. Thanks for explaining that. Right, we're going to go to the episode. It's even lovely to have you here. We get really excited talking about it. As I know you do. One of the big things, obviously, at the minute is this new open source model that's come from China, Moonshot. This is this young lad, isn't it, who's set this up? It's really interesting because this is a real challenge here to some of the big players. I can't drop it. What's your take on it? Well, he is young and he loves rock music. I was in their offices at the end of April this year. And every office is named after a band and then it has albums of that band, Pink Floyd Radiohead or whatever, in the meeting rooms. So they're a pretty remarkable team and people were surprised by this Moonshot Kimi K3 model because it was nearly as good as some of these top American models. And what we had thought was that the Chinese were perhaps four to eight months behind the top American labs and they just seemed to be squeezing that gap a little bit. But that said, what have we got here? We've got an extraordinary team that's had to work under the difficult circumstances of export controls and sanctions. They don't have access to all the compute. And what they've been able to develop is how do you do a lot without very much? And that is a skill in of itself. And how far behind say a fable are they? Robert, that's quite a difficult question because these models, they're multi-dimensional. They can be good at one thing and not so good at another. And so you end up trying to test them across a range of different skills. And some people said, well, it's better than fable and six out of the 14 things we, they tested. And I've used K3 and I've compared it to fable. And that's a subjective choice. I think the key thing that we have to take away is that it's a significant step up on previous Chinese quality models. It's much, much better than many American models and it's much closer to where the state of the art is. And one question to ask is do companies always need to buy the very, very best or can sometimes the good enough be good enough? Because price is obviously a key factor here as well because it's a fraction of the price is it? It is. It's about half the price of the top, top end American models. And that sounds like it's very, very cheap. But previously Chinese models have been about a tenth of the price. Right. So one of the things that's happened within K3 is that it's not as efficient as previous Chinese models have been. And in fact, in some ways, it's not as efficient as American models. So in order to squeeze out that performance, they've had to do something and that something is perhaps a little bit ineligant from the perspective of efficiency. It's also open weight. Which means as I understand it, that if I'm a business, I can broadly impolite it to essentially manipulate my own data for free. Which is a massive challenge to the likes of anthropic or Jumonau. How worried should these American businesses be because their revenue model is to persuade businesses to pay them for the AI? Whereas I mean, I may be wrong about this, but moonshots seem to be giving it away. Well, it's open weight. Anyone can run it without paying moonshots. They've got other parts of their business, which many people are paying for. So if you don't happen to have a million dollar GPU cluster at home to run your models, you'll often pay them $50 a month to do that. What does it mean for anthropic and open AI? Well, Americans always tell us that competition is the best thing for the market. So at one level, it's competition and that's quite good. And what we've seen is that anthropic has responded by extending access to their top model, which is called fabled. They had been planning to take it away in the middle of July and now they're offering it as part of the offering. I think the other thing that it will show up is it will show the extent to which American businesses and British businesses value provenance, brand, trust, liability, service and support. If you have a big enterprise contract with anthropic, you're getting more than the model. You're getting an account manager. You can shout out when something breaks. Whereas if you're downloading the Kimmy model from moonshot, it's on your head. So I wanted to ask about that, you went over to see moonshot and I love your description of them as being these sort of, you know, they're not quite heavy metal head bangers, but they're not far off. You look at what they're doing and what other Chinese AI companies are doing and it looks like they are effectively an arm of Chinese industrial policy. They are behind America in the development of AI. But when they make these services free, it looks almost like, you know, industrial vandalism against these American companies, which is obviously, you know, you could argue is good for the Chinese state. You know, when you met the founders of moonshot, did you get the sense that they were in it to become, you know, rich beyond their wildest dreams or because it was a sort of act of patriotism on behalf of the Chinese state? I met the moonshot founders and a dozen of AI lab founders across three cities. And are they basically capitalists like the Americans or are they basically, you know, out there basically doing the category right there? They are motivated AI researchers who want to build the best possible AI they can. But do they want to kill American businesses? No, it doesn't really come out that way. They want to build the best AI they can. They're very respectful of anthropic and of Claude. I mean, the structure within China is, if you haven't spent time there, it's quite hard to understand. But essentially each city, Hangzhou, Beijing, Shanghai, the mayors, direct and foster industry in particular ways. And so they're competing with each other more than they compete with Silicon Valley. And they're honest about being behind Silicon Valley. But the, the, the ferocity of the competition is really with, you know, your neighbor over in Shanghai or your neighbor in Beijing. For example, like Yang, the guy behind MoonChop, he was US trender, wasn't he? And he worked for some of the big American companies. So there's probably a lot that he wants to compete. You know, he's learnt from there and he's competing in that sense. Well, I think there's a lot of that, you know, repatriation that's been going on. I mean, the MoonChop founder did his PhD with Russ Saladinoff. He used to run AI at Apple. So these are sort of very, very expert researchers. But I don't really think you don't get a sense. And maybe that's hidden from us that everyone's hiding CCP flags from us as we go into the offices and bring, you know, bring them up. You know, we, we saw people who were really, really motivated, extremely technical in the way they, they spoke. And then they left us in a nightclub in Shanghai at 1am on a Friday Saturday morning to go back to the office to check in on their AI. What do they? Yeah. And then a nightclub at 1am and then they went back to the office. I was left on the dance floor. You're not serious, Azim. If you were more serious about AI, you had a gone back to it. I remember that's what. Exactly. They're pretty motivated. And you know, they, if you look at, we went into a, a, Z, that AI has it's called and they have a big dashboard. And it showed that, you know, the US was their second biggest market about the same size as in Indonesia. And you could see the international spread, the fact that in lots of countries, being able to get pretty good AI at the 10th of the price of chat GPT is a really good deal. Yeah, that's amazing. And when you again, you were talking to all these different entrepreneurs, how frustrated are they that they don't have access to Nvidia's state-of-the-art GPUs and processes or are they, in fact, managing to get older them through the back door? Well, it's normally the first or second thing they mention. And they just state it as a fact as you might speak about the weather. It's raining heavily today. And they acknowledge the constraints and-- And they get around the constraints. Well, they get around the constraints through-- there's a gray market that exists. There are data centers in Southeast Asia that you can access. It's quite Singapore is importing more GPUs than perhaps per capita, one might expect. But they're really not getting a lot. So they're all incredibly compute constrained. And just out of interest, when you are compute constrained, presumably one of the sort of benefits of that is they have to become much smarter in a sense with the coding development, the software development, to compensate for the lack of-- everything is about efficiency there. I called it the efficiency mode, and they have internal practices. So if you've got a particular improvement and it doesn't show a 20% efficiency saving, it won't go ahead. So only the very best things can move forward. And I think the US has been here before. So in the 1970s, the US auto industry was not really competing against the Japanese. The Japanese didn't have capital. They had very high energy costs, and they competed with something called the TPS, the Toyota production system, which was an ongoing basis of making manufacturing more efficient, and they're cars more efficient. And that's why in the late '70s and early '80s, when fuel efficiency measures were introduced in the US, Detroit couldn't compete with that flood of Toyota Corollas and Dats and Sunnies that came into the US. They were unable to pivot. So I think that that's one historical analogy that we can look at, because the US labs can claim they are compute constrained. They have six to eight times more compute than the Chinese labs, and so they don't need the same discipline. Do you see within anthropic-- because we've talked repeatedly about, in a sense, the way that the Chinese-- I mean, put it in crude terms-- steel from America when it comes to IP and technology. But is America now working up? Would you say an anthropic, for example, is recognizing it's now going to learn from China or their efficiency? The researchers are pretty open-minded about what research they read. And I spoke with one of the founders of Anthropic earlier this year and asked him about deep seats. If you remember deep seats, there's a first surprising Chinese model, really, really efficient. And he said, look, we knew about all these techniques already. We had them in our research queue, but we don't need to prioritize them because we don't have the same compute constraint at the time. And then, of course, they do use them, which is why anthropics margin, the gross margin they make on serving AI, has improved so much over the last couple of years. It's because they are implementing these efficiency savings. I think it would be simplistic to say the Americans don't care about efficiency. It's just about when do they care about it? Is it from listening to saying all this? Is it that the media is the ones to create in this race between America and China? And it's not as much of a race as we think it is with AI. It's very easy to blame the media. I won't today. I think the Americans think they're running a race. Right. But the Chinese don't. You don't get a sense that the Chinese labs are sitting there saying there's a race against America. They're going to be able to make a race to produce great AI because they are researchers who think in a particular way. And some of the things that I saw are slightly a little bit like the rapture in the sense of people's ambition, which is not dissimilar to what you see in some of the Silicon Valley labs who think they're going to build this machine god. I don't believe that that's what people will build. I don't think you can build it, but I do observe that they do. And I think that the Chinese labs are thinking in those terms, the American leaders think much, much more overtly about this race against China. And when you heard President Xi saying something that we actually haven't heard really from other world leaders, which is, you know, he is really worried about the potential destructive power of AI and he wants international cooperation. What did you make of that? He said a couple of things that were interesting. One is the point you made about the risk there. He also made a statement about how valuable AI could be for the global South for economic development. Does he have a policy to roll it out throughout? He has a policy to provide these open source tools, to provide training, I think 5,000 bits of training. I'm not sure what a bit of training means. Is it one engineer? Is it a hundred? Is it a firm? It's a significant, it was to 29 countries around the world, some you'd expect, Russia, Belarus and Mongolia, but others like Pakistan, Zambia and Mozambique, countries that couldn't easily afford these two. So this is like a digital version of Belt and Road? It's a digital version of Belt and Road. Maybe I think it's perhaps closer to Bandung, 1955, the Bandung announcement where Kwame and Krumah talked about trying to establish a third nation sovereignty of some description when you were facing the Soviet and the US blocks at the time. I think it felt a little bit more like that than a straight on Belt and Road because Belt and Road had with it, physical infrastructure, capital investments and all of the controls on capital that emerged from it. It was a much, much more controlling structure, whereas this was, I think a challenge that I would hope that the US and the UK would rise to also meet with their own offerings of open source and capability export rather than just leave it to one side. If the West ignores developing economies, emerging economies and simply leaves it to China to export their AI to, as I say, these growing nations, what's the implication of that? I think it could be extraordinarily challenging because you lose the soft power battle in the first instance and the soft power has been so useful for the West for the last 50 or 60 years. You also lose access to the technology stacks as a market that gets lost because today these models run predominantly on American chips from Nvidia and AMD, but China has responded to the chip constraints over the last few years to start to build their own chip capability. You may find that you'd go into some of these countries and every element of what we call the stack from the application to the model that's being run to the chips and also the power that's being provided by the Chinese solar panels is top to toe coming from China and that is a bond that is commercially and strategically hard to break. I wanted to ask on that about the power side of it because obviously that is probably the biggest operational cost for them. Is the companies doing or who will do well or are doing well? The ones who are thinking as much about the energy generation capability as they are about the tech and the chips? It's all about power right now. One of these chips need electricity to run and to give you a sense of the scale, the latest Nvidia, it's not a single chip, but it's a series of chips that you buy in a block. Needs about 100 kilowatts, so that's about running 50 ktls at full tilt. It's an extraordinary amount of electricity and so securing the power has been the thing that has been first and foremost on the AI companies. Then you, time to power is the thing that they care about the most. They're willing to pay a premium in order to get energized as they call it. You can't get a, what is 100 kilowatts? Well, 100 kilowatts is manageable in most places, but a data center five years ago might have been 50 megawatts. The data centers that are being built today at 500, 750 megawatts up to a gigawatt. That is utility scale. It's like a big aluminium factory of power. The AI companies are now out there, first and foremost, trying to figure out how to secure power for their data centers. That is the deciding factor. If we think about what that means in the UK, we could be more well positioned, but we're not really because there's this enormous queue to get connected to the grid. The area where most of our data centers currently live, which is this corridor from Shepherd's Bush and the west of London, passed Heathrow out to Slao, is the most congested electrically. You'll remember there was a power outage at Heathrow, which is the last place in the country you want to power out. There, I don't even think you could probably add a kettle to the grid without something terrible happening, let alone an AI data center. So it really is a lot about power there. So how do you get past that? Then if you know you've just mentioned there, the problems we have here, what can you do about it? What's happening in China is that they've had an electrifying electrical system. It's growing incredibly quickly and they've solved their power issues. In the US, it's a much, much harder problem because the US power system has been, the word I used in a report I did recently, more abundant for about 15 years. In other words, after a hundred years of growth, it just went flat and the company started to extract profits rather than invest in infrastructure and growth. So there's been this injection of demand, any place that can get a connection, its price is going up very, very significantly. The second is that if you can, you will just energise whichever way you will. So that might be diesel generators which are dirty, locally polluting, expensive and chuck out loads of CO2, there might be gas turbines, but now there's a 24 month backlog to buy gas turbines. It might be to invest in renewables and we've seen some enormous renewable farms and finally nuclear. And so there is now a resurgence of nuclear in the US, but it's a very, very slow process, that particular part of our industry is like wading through molasses. This episode is brought to you by Ascension. Now when you're advertising operations fall out of sync, campaigns, slow down, insights get buried and opportunities get missed, that's why Spotify and Ascension are working together to reinvent the rhythm of ad sales. And also, we're working on automation, analytics and smarter workflows to simplify campaign delivery and access better data across the business. The result, less time spent on operations, more time connecting brands with the Mormons and fandoms that matter most. To learn more, check out Ascension.com/Spotify-EasyCare. In this incredibly important area of economic development. You have recently published a very influential report where you did something unusual, which is you actually did micro research. You went and as I understand it, talked to all the important players. And your view is that there isn't a bubble. But actually the revenues that are coming down the pipeline do justify the kind of investment we're seeing. Is that right? I'm going to allow myself to nuance it. Well, of course. I mean, you know, bubbles can emerge for lots of reasons. And there could be a bubble that emerges here. But on the specific question of are revenues from customers into the AI sector currently growing rapidly and have they started to meet some of the depreciation expenses of this multi-hundred billion dollar build out. The thing that surprised us was that at the end of 2025, they crossed that mark. They just started to cover the depreciation expense. And that's quite important. I can give you some numbers. And this is cash or this is a cruel accounting. This is cash. So in the 12 months to 2026, the amount of revenue generated in the AI sector exiling China was $110 billion. And that is controlled for duplication, you know, for spending $100 with anthropic, he then spends 50 with Microsoft. That's not 150 in our book. That's 100. That's a 3.5 fold growth over the previous 12 months. And it was growing, it's still growing at that pace. So we estimated that although it was 110 billion in the previous 12 months, if you annualize June, you get to $175 billion. And that is faster than other things like, you know, more bubbles coming in early internet. Yeah, if you look back at mobile apps and the internet or the cloud hosting, it's about three times faster than those particular waves were to this $100 billion annual mark. If you look at the investment that is taking place now and committed for the next few years, are you saying that if revenues continue to grow this rate, that is justified expenditure, or are you building in some conservative assumptions that at some point that, you know, the revenue growth will slow down? Well, I mean, if it grows at this rate, you know, it just be a very odd situation, but it would be a very good investment. I don't think it will grow at this rate. I think it will slow down. You know, these things have to. The question is where and at what point? And if you look at the level, I mean, if you think about what has to be paid back, you have to pay back the depreciation expense on these, all these chips and the cooling and the power and the building. And you also have to pay the operating expenses, so the electricity and the staff. And then you have to make a profit because you've got other costs outside of your growth margin, your marketing team and so on, and finally your shareholders aren't investing in the charity, so they want their return. So if that, if the total stack of the depreciation expense is, say, $250 billion, which it might be in the. That's a lot of depreciation. It's a lot of depreciation. That's the sort of range it might be in 2028. It'll actually be slightly higher than that. If you wanted to have a fully sustained market, you'd expect to see $400 billion, $450 billion of revenue come in, that will pay the depreciation expense, it'll pay power, it'll pay salaries, it'll leave a profit. And is that a reasonable rejection for 28? Well, we're at, you know, we'll end the calendar year of 2026 between $185 and $190 billion. And it's harder to forecast 2027, but I do think that for it to get towards $300 billion, it's not unreasonable. I mean, we have more detailed forecast. The range is quite wide. It could be 250. It could be 350. And just to be. I mean, before I get on to some of the risks around all of this, because these are huge numbers. So we're talking about 300 billion by when, did you say? The calendar year 2027. And when, I mean, just so that people that I couldn't understand just this year sort of extraordinary development of this industry, when in effect were they not? Oh, 2022. So 2022, not, you know, literally five odd years. Yeah. We're looking at $300 billion. Well, American and British businesses are still by and large dabbling. I think that's one thing for us to understand. They are. You mean the people using it? Yeah. The people who are the customers who are paying open AI and anthropocal all these other businesses are mostly dabbling. The average US business spends $11 per employee per month on their AI tools, which is not a huge amount. And very few US companies have yet come out and said, "This AI thing has completely transformed our economics." You're getting the first signs, so Bank of America, Goldman Sachs, these are the easiest businesses in the world to do this in because it's finance. It's digital. You're not moving stuff around and they already spend a lot on AI and they make money very quickly. Right. That's what a bank does. So you've got the first signs of companies getting to results that they're willing to talk about. One of the things that's difficult at Stefan Robert is that CEOs are under so much pressure to prove that they can use AI. They're talking about it a lot. So we track earnings transcripts and about 30% of American CEOs talk about how they're using AI and their earnings. You have to then do some picking parts. Say how much of that is just job protection and how much of that is real. And when I talk to CEOs, I talk to in the order of 100 in Europe and the UK and the US every year, I hear a pretty good question. Common message back, which is they're committed. They're getting some early results. It's much harder than they thought and they don't intend to slow down because they're learning and they're figuring things out and their ambition is growing. Is it a case that because you know obviously we talk to business bosses a lot. It feels like there's a kind of divide between the CEOs and the businesses who are getting their staff to use AI to kind of improve their productivity and those that are looking at it from a transformational point of view. And how do you bridge that? So it becomes, it's not just about efficiency for workers, but it is about how you use it to totally transform your business. A total transformation is very difficult. With electricity, it took 25 to 30 years in the US because you have to completely change where factory works and that just takes time. What the AI companies have started to do is they have set up joint ventures of professional services of management consultants, so they're called the deployment company, for example, as the OpenAI one. It's got billions of dollars from private equity and from OpenAI itself. And the job of those businesses is to go into established firms and do that transformation that you've talked about on the basis that you need both all of the intimate knowledge that a company has and this outside thinking that the AI companies have. So that's the theory. And I'm sure it looks really good on a slide. We'd have to see if it actually works and how long will it take to work. But there's a related point, which is I, and I'm sure like you, know quite a number of people who've set up essentially AI-only firms in specific sectors. So in law, for example, I know a couple of new law firms that are essentially employing almost no lawyers and doing almost an offering but sort of standard services. So, you know, in things like liable law or all that kind of stuff. And so it seems to be the issue is if these almost one-man bans but offering really quite a comprehensive service were to take off. Then if the existing firms don't totally re-engineer, they're going to be out of business, aren't they? Well, look, something like, like, will writing is very, very low-hanging fruit and for years you've been able to buy a will for £10 on the internet. The reality is that most law is not like that and the large part of the expense is more complex employment law issues or it is transactions. And what law firms are starting to realise is, and it's not just law firms, it's also software developers, you speed up individuals but you don't speed up the team. What you do is you create congestion because we're also busy but it has to be verified and signed off by somebody. And that's particularly true if you're a law firm because your job is essentially to be the liability sponge for your client. So the process speeds up as fast as it's slowest component. Also, I've got a mate who's a whirls and pro-brett solicitor and she spends most of her time dealing with emotional people, dealing with whirls or there's that as well which is, yeah, okay, I can help her speed up do the legislative side but not the actual dealing with these emotional clients on the phone. And that part I think stays is very hard to see how that gets affected over the years. In fact, it might get improved because she may have more time to talk to them but I talked to a manager and actually funnily enough in a company that works in AI and she said she's got a thousand engineers and every one of them is about 30% more productive but as a whole the group is perhaps 10% more productive. And what's going on? Why is 1.3 times a thousand, you know, not in the app where it should be and it is that problem of of congestions and bottlenecks because we've built all of our internal systems for an assumption about how fast people work. You know, Robert, you're extremely ferocious with your essay writing and your tweeting and so on because you go direct to the audience but imagine if there was a traditional newspaper chain of sub-editors and editors to sign that off, they wouldn't be able to cope with your the speed with which you produce. And I think the thing that companies need to do to get to Steph's question of transformation is how do you how do you get around or change the way you verify and approve and make decisions. And companies are not bundles of tasks. The task is one part. Companies are actually people sitting around agreeing to take a risk or not take a risk and that process still takes time. It does take time but even the verification process can in a sense be delegated to AI. And so all of these bottlenecks, I'm not saying this is a good thing, but all of these bottlenecks can be massively dealt with, speed it up. They can be cleared. Over time they will they'll be cleared and I have my own my own experiences with my own work where I have now verification systems which are many different AI's. That's right. A little bit like a series of Muslim Gauls is and if you overlay them all you can't they become go from being trans-ecent to opaque and and that's what I've done which sort of saves a little bit of time. So just in practice can I ask how that works? So you will you know do something and then you will put it into lots of different AI systems. That's right. Yeah. And then how do you know which is the right one? So what will typically happen is that I will I will have written something up and I will put it into two different AI systems. One might be you know, Claude and the other might be a chat GPT set of agents to do some fact checking. And if they disagree on any fact in any degree or any argument then that goes to a judge that will go off and check. And the agreement level depends on how consequential this is. Sometimes it's just the number. Sometimes I have to agree both on the fact and the academic article that it's that it's come from and that can just run that just runs that's Robert's bottleneck being cleared there. And then I end up with a like a log that is the list of facts and the list of articles that these come from. What the agreement level was and I can walk through that and say just like an auditor the senior auditor at the end of an audit and say I want to dig into this a bit more. My self-to-examination. The marginal cost of the checking is more or less nil whereas the marginal cost of a human checking is whatever that person is being paid. So you know the productivity savings here are in men. But it means trust though doesn't it that's the thing that's the key here is like businesses and business leaders have got to get to the point. Our team manages whoever it is in the case of your friend who's the engineer with a team of engineers. You've got to trust that that is right having you like and that's where we're not we're not there yet. So there are these enormous imponderables. The rate of rollout in established businesses how efficient how effective in the end this rollout will be for these businesses which takes one back to this issue you know this not not an important issue about whether there's a bubble. You know you will presumably have seen the fairly detailed academic or effectively quasi-alconomic paper by the Mac for International Settlements and just remind listeners that the point about the BIS is it does have a pretty good track record of identifying bubbles and most famously of all it was a very lone voice for a while among international regulators in the run up to the 2007 eight financial crisis. And the BIS was an early and important voice saying there is a very serious bubble here and you know we're looking at the potential for a big crash that's going to harm the economy now they have to say they put out a paper where they are saying that on their model the level of capital investment at the moment is significantly greater than what you might call the equilibrium rate which means that there is in their terms a very significant risk or significant risk that we are going to get a crash. We may well get a crash. I mean if you think about the scale of of investment that is that is going into this it's going to be into the multi-trillions of dollars by the end of 2030 and that is really really hefty and if you look at the amount of capital investment that's gone on in the US as a percentage of GDP going back to really the 1920s which is when the data set is available this year will be close to the highest year ever which happened to be I think 1930 so when GDP was heavily depressed after the the Wall Street crash next year will almost certainly be be higher so these are things to look at because this is what economists call a shock it's a shock to the system and certainly the the health of this entire system is not as strong today as it was a year ago when I started to formally track the the the bubble risk the revenue is very very healthy far ahead of where we expected it to be but that's not the only factor in where that where a risk starts to emerge and you know where where would that where would that come out the the real tell and you know of course Robert you know this from the global financial crisis and the housing crisis is when financing structures are you know borrow short for a long investment so what classically happened in the US was that homeowners were taking two-year interest-free mortgages and then moving on to ARMs adjustable rate mortgages and that's when the results happened once financing starts to take that kind of characteristic at scale in in this market you would start to get quite nervous in other words people lending borrowing for five years on a 20-year asset in these sorts of situations what typically happens is you'll get a business that says we're growing at this kind of rate And we've got these contracts in place for our services. And what they tend to do is then discount those revenues back to get a present value and then borrow against that de facto security of those supposedly contracted revenue schemes. How secure are those contracts? - I mean, some of them are much more secure than others. And there's this-- - 'Cause it's obviously a point of risk there that at the point that any individual business says, "Actually, this is not working out for us." - Right, I mean-- - Can I just get a castle? - There's this number called the RPO, the revenue performance obligation that gets spattered out and these numbers are getting so large that you can't fit them on the calculator that your kids would have used for their GCSE. And there is inherent in that, there is some risk because it's a contingent on a number of things that might play out and we were $110 billion annualized over the last 12 months and we've barely touched the surface of the FTSE 500 or the Russell 2000 US. - In terms of sales, to make it sales. - There's a lot of room for that, that's great. And so I think the risks emerge elsewhere and I think they emerge from what might happen within financing structures and financing arrangements and try to work out whether there are things that are not just the occasional deal that looks bad but something that is sort of common pattern that's emerging. - 'Cause there's been quite a bit of scrutiny in the way like Nvidia has been like financing companies that are then using the chips and there's been lots of talk about, is it like Enron with the special purpose? The renderings. What are your thoughts on that? Because that, there are people who think that looks a bit dodgy. - Well, look, I think people have, there are two rough camps about AI, there are those who, you know, think it's gonna create the machine garden or those who really hate it. And depending on your camp, you will have already made your decision. - Yeah. - So, you know, I'm gonna talk about what the middle looks like. Think about a market that's expanding really quickly. And it's expanding so fast that people providing services in that market don't have the capital to serve their customers. But their underlying supply has got an incredibly strong balance sheet. Perhaps it's five year bonds have got a better credit rating than the US Treasury, which is a case of Nvidia. - Yeah. - It would make sense for the businessman to lend his downstream customers in order for them to serve their customers. And that would feel really, really normal in most other industries except in ones where people have this, you know, spurs arsenal division about which side they're on. - Yeah. - This happened in the US in the 20s with auto finance as well. So, you know, general motors didn't just provide high purchase or financing packages for consumers. They also financed the dealerships. Now, every historical analogy is only as useful as you can stretch it. So, there's lots of things that are going on today that are without precedent. But without precedent, doesn't just means that we can't go back and look at a data set, say how do these things work out in the past. At the end, what we see is that the chips that Nvidia give out to people or, you know, sell or invest in companies and then they buy those chips are used at extremely, unordinately high levels. They're mostly on six year depreciation schedules. When we talk to people in the market, they tell us, well, after six or seven years, these chips are still commanding premium rents. - Right. - And so, what that suggests is that as long as that demands is still there at the front end, those chips will get used and the risk is-- - Demand-free I services. - Demand-free I services is there because it means the chips have to get used. If that demand softens, then the chips start not to be, start to, you know, be underutilized. And then you might start to see write downs. - And you're looking at losses. - You're looking at losses and you're looking at write downs and then you could see things unwind very quickly, you know, in a slightly disordered way. - And I guess, well, we should wrap on a minute. The part of all of this that does really worry me and is analogous to 1929, but he's not, you know, in a sense an issue of hard data, but it is one of sentiment, is just the extraordinary, irrational, enthusiasm, particularly among individual retail investors about all of this. So, I was genuinely shocked and really quite worried by the trading in the Musk IPO in the days after that. Because that was just nuts. And that was just that thing that we did see in the Twainers, which is, you know, if you don't get on board, you're gonna miss out and all completely detached from any actual rigorous analysis of what cash flows are likely to be like or the strength of the balance sheet. This feels like the Twenties to me, which is basically, you know, that these pulled particularly of US savings, just being channeled into what they think is, you know, the gold mine, in a totally, you know, vibes-based way, nothing to do with serious investment analysis. And that, it's the weight of money going in from retail that really worries me. - The Koreans have been through their 1929 moment. So, the Korean market is very heavily invested in the semiconductors that power the AI chips. And there's been a 40% decline in the cost being the last couple of weeks. And 4% of Korean household had what's known as a margin call where they've been overexposed through leverage and the guys knocked on the door. And that was the shock. But, you know, the Korean market is extremely retail driven, much more so than the US market. But you have seen exactly what happens when the exuberance gets irrational and then, you know, nerves take root. If we come back to the concern that you raised, that what, I think there's something to learn from the Korean market, the regulator moved really slowly to stop leverage around ETFs. And there's some questions to why wouldn't you've done that three or four months earlier. So, there are certain breaks that can be applied. I think the other thing that's worth noting is that US banks, tier one capital, is extremely healthy right now. And certainly compared to where it was in 2007 and 2008, very, very under leveraged. There is quite a lot of leverage in the US financial system sitting with hedge funds, investing more broadly, which I think have more than the retail risk as a risk because they're very, very overexposed. They borrow from only a handful of banks and they can unwind very rapidly. So, I would agree with you, there are lots of these risks that are there. When I look at the metrics that we track, things look healthier because of revenue, they look slightly less healthy because of the way financing, especially the debt financing sits, valuations don't look to aggressive at all across the NASDAQ. There are exceptions, SpaceX was one, briefly, but across the market they don't look particularly hairy. So, the patient is, for me, I would, if I had to give it a rating, is still reasonably healthy, perhaps not as healthy as it was a year ago, but not yet at a point where I have to call 999, but I wouldn't rule out having to do that at some point. - That's a very good point to end things on, isn't it? - Yeah, let's hope the hard stack isn't next week. - Same, thank you so much. - Thank you. - We could chat to you for hours on all of this. I love the fact that you've got a million graphs in front of you as well. - Just in case. - Yeah, thank you, it was a, that's it from us on the rest is money, bye-bye. - It is, goodbye.