Moonfare Insider Talks Episode 01: Azeem Azhar

Michael Sullivan: Hello, I'm Michael Sullivan, chief economist, uh, at moonfair. And you join us, uh, for the first, uh, Moonfare Insider Talk, a new series, uh, where we debate the fast evolving world, uh, with the very best minds, the most original, uh, thought leaders across the world. Uh, the setting here is quite special. It's vestry House in St Giles in the heart of London, uh, a building that sprung up in the first Industrial Revolution. And I'm delighted to be joined by an old friend, Azeem. Azar. Azim is the author of the best selling book, uh, Exponential. Uh, the man behind the blog, Exponential View. He's an analyst, a scientist, uh, and an entrepreneur. Uh, Azeem, you're very, very welcome. It's great to see you again. And what lovely surroundings, which I guess came into being in the first Industrial Revolution. Now we're in Industrial Revolution four or five.

Azeem Azhar: Absolutely. Well, thank you so much for having me and also bringing me to this wonderful building.

Michael Sullivan: And Azim, we're going to talk about artificial intelligence, uh, which is beginning to take over the public debate. It is becoming, uh, geopolitical in terms of the race between the US And China, uh, and Europe to master forms of AI. Uh, there's a big debate about how it will transform, uh, our jobs, our economies, uh, our societies, uh, and also I think really importantly, from a Moonfare point of view, it's transforming, uh, the investment world not just in terms of how we invest, but the new companies and the new sectors, uh, that are already beginning to bubble up. So we have a fascinating hour ahead of us. Uh, so, Azim, just for a bit of context and so people, uh, understand exactly what you do, uh, in February this year, the share price of Nvidia crashed by 20%, uh, because of the release of Deepseek. And what had happened? I scratched my head because I thought, Hang on a second. I put up a slide of, uh, Deepseek's model results about three weeks previously to my, uh, inferior colleagues. Uh, and that slide came from you. You'd sent it out, uh, over Christmas. And then I went back to all your blogs and I found that you had been talking about Deepsea back in, uh, 2023. So you really have, uh, your finger on the pulse of what's happening in AI. And just give us a bit of context as to how this new technology is becoming strategic across the, uh, three or four big regions. I know you're off to Japan soon, for example.

Azeem Azhar: Well, you know, my team and I do track the major research labs and we track what researchers are finding interesting And Deep SEQ did, as you say, pop up, uh, a couple of years ago. And I think what we are starting to see now is that labs in China are taking a different approach to the big labs in the US The American approach is a little bit like many things that are American. Build it big, stack it high. Very, very large models, uh, which require a lot of computing power, are targeting this idea of a very generalized AI system. What we're seeing in the Chinese labs in particular, since they haven't got the capital to spend on the chips and they can't get the chips anyway, or the most recent chip they can't get access to, are more innovative approaches. And we're seeing really, really genuine, uh, breakthroughs. And so deepseek has emerged with a model that's nearly as good as some of the best American models. But more recently, in the last week or so, Moonshot, AI released a model called Kimi, which is even, in a sense, more remarkable. If deepseek was the Sputnik moment, um, Kimmy I described as the Yuri Gagarin Vostok moment, the first man in space.

Michael Sullivan: And I'll ask you to explain that in just a second. But, um, I think one of the things that is both eye catching and alarming to a lot of people is the speed of AI, because every week we hear updates about what AI is doing in medicine, et cetera. Um, and for most people it was the launch of ChatGPT that defined the advent of AI, even though it's been happening in different forms, uh, machine learning, neural networks for a long time. And I know you like to talk and write in terms of frameworks. Can you just give us a framework for how AI has evolved and where it's going to go in the next three or four years and what kind of milestones we need to watch out for?

Azeem Azhar: Absolutely. Well, I think the first, most important breakpoint in the last 15 years was the, uh, ability for deep learning to work on Nvidia's gaming chips. Uh, that was in about 2010 and 2011. And deep learning revived a method that essentially said if we do big statistics on lots of data, we can get these things to learn how to recognize images or recognize speech. That electrified that first AI boom. Recent AI boom we had, and it took us to, uh, deep mind beating, the world's best go player, and some other breakthroughs. But the thing that has really picked things off was the development by Google of this large language model architecture, which is what underpins ChatGPT. They developed that themselves in 2018. And in so many cases, of course, the inventor does not exploit and ChatGPT then shows up and it is visible in the Data. Up until October 2022 when ChatGPT, uh, is released. The big AI labs in America are releasing a new AI model collectively about every six months, not very often after ChatGPT, we've now counted it, it's every two weeks. And so you've gone from this cadence that is rapidly accelerated. But it's not just that the cadence has accelerated. The models before ChatGPT cost about 50 to $100,000 to train. Um, the new models cost between 50 and $100 million to train. So there are many more of them and they're much more expensive. And it's absolutely visible in the data. You can look at this graph. On the left hand side it's barely populated and on the right hand side it's just a forest of X's.

Michael Sullivan: May I ask why has the cadence, uh, picked up? Is it more capital, is that they're learning from each other?

Azeem Azhar: It's a combination. What ChatGPT showed was we can now use this bit of technology. You can use large language models for things other than chatbots. But the chatbot constructed an enormous consumer pool, uh, customer pool, enterprises and end users. Once you get that customer pool, once you see it working, I think then the competitive dynamics, uh, emerge. And the trouble is that AI, if you put yourselves in the heads of one of the big tech companies, AI, uh, uh, could be existential to your business. It could be the digital camera to your Kodak. And so you have to calibrate what you choose to do and in action is a type of action and the markets will see it. And so I think we can see Apple floundering a little bit without an AI story. And it doesn't look good. And so I think that that is one of the motivators. But then the real. Let's take my cynical hat off and be optimistic with the data that we've seen demand for using these systems, uh, I don't even have the phrase for it, it's off the charts. So we call it, I mean Google saw 50 fold demand for actual AI requests in just a 12 ah, month period. And that is an enormous demand growth. No one is forcing these customers to use it, they're having to pay for it and they're choosing to.

Michael Sullivan: Yeah, and I think it's going to change, um, whole industries in terms of how they're structured and they behave. And actually a lot of the channel checks from our investment team At Moonfire pick up the use of AI across a whole range of industries. Let's jump back to Kimmy, because you sent out a note, uh, I think it was last Monday, describing this model. And when I read it, it seemed to me to be quite obvious what they were doing and quite clever. Uh, but it's a big leap forward in terms of changing the, uh, maybe the financial dynamics and the scientific approach. Do you want to, uh, describe it? Because I suspect that many people haven't heard of this.

Azeem Azhar: Yes, well, Kimi is a model that's come from Moonshot AI. Uh, it is optimized for reasoning. So the idea of reasoning is that it's less of a textual parrot, should be able to reason across, logically think through problems. It's also been optimized to be used in agentic applications. And that means that you can let the AI work on its own for a little while without having to constantly check at it. And one of the things Kimi did was that they trained a very, very large model. Uh, we count these things in parameters. It has a trillion parameters. A trillion of anything is a lot. But it only activates a small number, a few billion, about 30 billion at a time. And it has effectively a series of experts within that model. And I think any of us who run teams would appreciate that. When you put together a team for a particular deal, it may look very different to the people you bring on for a different deal because you understand the dynamics. And the mixture of experts approach is a great way of getting very, very efficient because you only use the experts you need. That has meant that Kimi is extremely cheap to run. And at this moment it's about 100 times cheaper than some of the U.S. um, uh, reasoning models. Although I should correct myself when I checked on Tuesday last week, so a week or so ago it was 100 times cheaper. It may no longer be 100 times cheaper. I mean the world's moving very quickly. But I think the thing that is most interesting is that, ah, Kimi's model is what's known as open source and open weight. And what does that mean? It means that anyone can download not just the models, but the weights that construct and define the model, adapt them for their own use, and run it on their own infrastructure. And that's really distinct to the approach that's being taken by OpenAI or Google or Anthropic. In their main models, which are uh, entirely closed models, you have to use it through their infrastructure or through a partner infrastructure.

Michael Sullivan: Yeah, and just uh, to take That a step uh, further without knowing much myself about AI in China. It's quite a revolutionary approach uh, in China which we associate with being quite a closely guarded economy and quite closely guarded in terms of R and D. So this is really evolutionary I guess.

Azeem Azhar: I think it's evolutionary. I Deepseek took the same approach. They had the most permissive style of open source licensing and there are a number of different ways to interpret it. It's hard for me to intuit intentions but let's talk about what impacts are. By offering very very high performing open source open weight models you actually undercut strategically people. With closed models you say look here's something that may be nearly as performant but it's much cheaper. You can optimize it for your own applications. Why would you go to a, ah, closed source shop like OpenAI? And I think it's a very very um, it's a very strong play particularly if you are the challenger in a market whether you're Chinese or American. It's a good thing to do to go out with open source. Why? Because it builds a developer ecosystem which builds uh, use cases and then it builds reasons to purchase and that turns into some sense of longevity. I think that that is a really smart move by these firms. Maybe it's something that one can describe as China, but I don't know if the order has come down from Beijing. But I think one of the things that is a bit concerning is that uh, these models are very, very sophisticated machines. We don't know what's gone into them and we don't know how they have been trained. We don't know how they'll respond. Kimi doesn't really publish many specifications on what it was trained on, nor does it tell us how has it been um, uh, culturally tuned for use. And that could present a security risk if you use a model in the US or in the uk. I don't mean it would exfiltrate data but I do mean that there might be certain subjects that it won't give great answers to. It's less performing CCP type questions.

Michael Sullivan: For example, I've mentioned your blog twice already so I'm a fan and recommend people look at it. Um, but let's go back to base and your book Exponential Mhm uh, which is a best seller. Do you want to just describe for uh, the viewers your idea of the exponential economy? Because I think they see it happening in different industries but maybe just put, put it together in a whole.

Azeem Azhar: Yes, I mean the foundational idea uh, is that because of learning by doing and learning rates, uh, one school is Wright's Law. But you could connect it to, uh, economic theories like endogenous growth theory, uh, is that, ah, effectively we are compounding our knowledge. And what that results in is certain classes of products which end up being inputs in the economy getting cheaper and cheaper every year. We know this with Moore's Law and silicon chips. It's also true with solar panels. They've dropped by, uh, 99.9% in cost over, uh, 30 years or so. And that would continue to go so long as that learning persists. And that has very interesting impacts when it gets rolled out into the economy. Because of course with the price that low, the breakeven point for a business to use that technology also drops. And so you'll just get a lot of demand, uh, which in turn will give you economies of scale and more learning and will bring the prices down very, very low. I think also importantly, it means that it tends to break the boundaries of existing markets. So in venture capital terms, we call that the tam, the total addressable market. And what you find in technologies that, um, have these exponential characteristics is the price comes so low, complementary uses and services get developed quite quickly. The TAM expands dramatically. That's why there are more than five computers in the world today, because computers got very cheap.

Michael Sullivan: Yeah. So we are in the middle of a productivity recession in many parts of the world. The UK is a prime example.

Azeem Azhar: We're sitting in the middle. We're in London. We're literally in the middle of the epicenter of the productivity funk. Yeah.

Michael Sullivan: And if you were to, um, advise the government, advise the state, and it's not just the British state, but because it's a problem, uh, writ large. We're at this period of the end of globalization and people are searching for new sources of growth. Where does the idea of the exponential economy fit in a growth model? And maybe you could give one or two practical examples. If the chancellor or the minister of finance in a given country was to do something, uh, how would they incorporate that into, um, policy?

Azeem Azhar: Well, the first thing is that you can see the trajectory. You can see that solar panels, despite tariffs in the US Are going to continue to get cheaper. And we understand the mechanisms. And one of the things I discovered in talking to policymakers over the past few years was that they didn't necessarily understand that dynamic and they thought that Moore's Law was just an exception, nor did they necessarily want to do the second and third order, uh, thinking the second order and third order Thinking being, well, um, we've got, we use, each of us use one AI today, but in five years we'll be using a thousand. And in 10 years we'll be using 100,000. We'll have this constellation. So I think one of the starting points is really about that degree of, um, belief. Do you believe that that's going to happen then? I think there are two little policy suggestions that I can think of. One is that when technologies that lend themselves to, uh, exponential price declines emerge, they tend to be quite expensive. And what gets them cheap is if industry can generate learning effects. And so there's a very positive intervention that governments can do, which I think appeals to the left and the right, which is to make sure that technologies that exhibit those learning effects get into the market as fast as they can, whether it's batteries in the home or heat pumps or AI systems. The reason the left likes it is because it democratizes access. The reason the right likes it is because it creates a new infrastructure on which entrepreneurs can build new businesses, employ new people and generate new tax revenues. So that, for me is the easiest policy possible. We can forget about whether we think of it as industrial policy or interfering in the market too early. Um, I think it corrects a market failure, which is I know today that heat pumps, exponential technology are as expensive as they ever will be and that in 10 years will be much cheaper. The fact that the market is not pricing that, the fact that finance packages are not reflecting that is to me a market failure. And that is an opportunity for government to intervene.

Michael Sullivan: Yeah, interesting. Um, policy usually lags, uh, technology and particularly breakthrough technologies. You, uh, know, regulators are scrambling to catch up. And we saw a moment when ChatGPT was released. You had a lot of very nervous discussions in the press about the positive and negative effects of AI. And you had a lot of the AI leaders writing open letters and newspapers saying, you know, uh, we won't open Pandora's box, et cetera. And Vinod Khozler, who's a very famous, very high achieving, uh, venture capitalist who'd been known to the, uh, Moonfare audience, um, he coined the phrase, is it a utopia or is it a dystopia?

Azeem Azhar: That's right.

Michael Sullivan: And I'll ask you for your own opinion on that in a moment. But maybe just give us, because you're in touch with all these people, what's the debate in the AI leadership community on this and where is that going

Azeem Azhar: in the US the general view is, uh, it's too early to regulate, uh, that would be the view that could be seen as being self serving. Of course, which business person wants regulation, but I don't think that's quite right because I think they are all quite thoughtful business people and most people who make cars or pharmaceuticals or even financial products are pleased for regulation. It allows them to know their boundaries for operation. But I think the issue is that uh, we can't quite work out exactly how this will play out and where the risks will lie. And I think Vinod's view would simply be that there's too much uncertainty and so intervening particularly heavy handed ways is going to be unhelpful unless you really really have the capabilities. An example of someone who wants to intervene in a heavy handed way and doesn't have the capabilities is of course the eu. And so the EU AI act, uh, has emerged and it's really problematic and it seems clear that the American companies are going to not comply and they're going to ask Trump to help them not comply. But the act in of itself has all sorts of issues. The first is that we don't have enough evidence yet about what the technological path is going to be, what the harms will actually end up being. So the first point, the second point is that we're in the post rules based order right now. And so we're not in that world of 2018 where Microsoft said, look, we're going to apply GDPR standards globally even when we don't have to have them enforced in particular countries. That's not the world we're in. And so that sense of being Brussels and shaking the world stage, that's a distant dream. But the third thing is that AI is incredibly important. And if the EU wants European values to persist, Europe has to be strategically, financially and economically a player in the game. And by squeezing innovation, by forcing regulation right now we're, I think what they do is they take themselves off the pitch and that doesn't serve their stated intent. And unfortunately I think this is a great example where this type of regulation will harm your long term prospects if and only if you believe, as I do, that AI will fundamentally transform firms, then the economy, then society at large.

Michael Sullivan: Okay, let's dig into this a bit because this is really interesting because AI is, ah, as a sort of wizard technology of the future, perhaps the dominant technology of this century. It is highly geopolitical and I've had similar conversations with people in Brussels on things like the capital markets union or savings and investment union. And I have the same, uh, observations as you do. There's Just not enough capital flowing and it's not flowing quickly enough. There's too much, uh, risk aversion. I mean, do you really think that this will hamper Europe, uh, in the sort of race with the US and China to be relevant in the 21st century? Uh, and just as an extension of that, what could they change quickly to speed AI up?

Azeem Azhar: In Europe we have lots of advantages. In the UK and in Europe we've got very deep human capital. We've got great universities, we've got great universities that are in the AI fields, but also broadly in science fields, particularly in Germany, we generate a lot of industrial data which CHATGPT is not trained on. It could be useful. And in the uk, City of London, we generate financial data and we have these hotspots. So these are all strengths. Data, of course, will travel across optical fibers to data centers wherever they are. Uh, and you also have to construct that milieu where people feel they're welcomed to work in this area, which I don't think by and large the messages have done. So consider what's happened in the Gulf. So Saudi Arabia and the UAE have both announced pretty staggering deals to provide what all AI services need right now, which is power. Uh, and in doing that, data centers are going to be built in the Gulf area.

Michael Sullivan: There's the G42.

Azeem Azhar: The G42 as well. And every citizen in the UAE now has access to ChatGPT. Plus, I think those data centers will be able to serve every consumer in Europe, except for those in Iceland and in the Shetlands. Well, by serve I mean it will be 100 millisecond round trip ping time to data centers in the UAE from Frankfurt, from Paris, from London. And so of course it end up being cheaper to surf there, it'll end up being cheaper to build there. You will construct an ecosystem around you. And so I think one of the things that Europe would need to do is they would absolutely need to have a very, very aggressive power and electricity policy, which, um, having spoken to some of their leaders in Brussels in the last two or three months, I really don't think. Yeah, they have, they haven't cottoned onto an era of electricity demand growth. After 15 years of being driven by efficiency, there is enormous demand growth from heat pumps, from electric cars and from AI. I don't think they're thinking like that and I don't think they are willing to cut through whatever tape is required to sit around the table with the kkrs and the Brookfields of this world and the nvidias or the AMD's and say, we're going to facilitate this. We want you guys to bring the $500 billion worth of capital and we will make sure you will get the power and you will get this up and running in three years. Now I know they can do it because I was in Paris three weeks ago and Notre Dame seems to have got built far faster than any data center can do. So we know that there is political capability to speed things up.

Michael Sullivan: Yeah, and at the risk of kind of being, uh, AI, uh, energy geeks. Let's just dig into that one more time because I was actually quite impressed by the UK's AI opportunity plan, which is the plan for the, uh, build out of AI and the use of AI in the uk, but also very detailed policy, uh, for energy, for data centers, new sources of energy. Looks very, very good on paper and I think you probably had an input into some of the policy making. Um, do you think that's feasible? And where might they get the money? Will it be from the private, Will it be private capital?

Azeem Azhar: Uh, well, I think it'll end up having to be largely private capital because the sums get big. But I also think that they're more well stewarded and there is a lot of private capital around. And I really do think that the Prime Minister's AI advisor or the Prime Minister's former AI advisor, Matt Clifford. Absolutely. Superb job. I think the challenge is that it's pushed from the bottom up and one needs to really be convinced that the leadership knows and understands it. I mean, I'm old enough to remember Margaret Thatcher. And when Margaret Thatcher stood up and would say, talk about the power of the small business person, you knew she believed it and you knew she meant it. When Tony Blair would talk about Cool Britannia and our cultural capital, you knew he believed it, you knew he meant it, and in both cases you knew they'd follow through. And I think that if you're investors, you also have to get a sense that the government is going to be capable of following through. We are starting to see signs of that. Bat tunnels, fish aquariums and so on are starting to, to disappear. But I also think there is still a bureaucratic lens over a lot of this articulation, which is not quite. I think what electrifies people, what electrifies people is political leadership that says, this is who we are, this is where we're going and we're going to do it.

Michael Sullivan: It's a vision problem.

Azeem Azhar: It's a vision. Vision driven. Yeah, I think we think we see a little Bit of that in China with the various five year industrial plans, in the sense that, you know, many members of the top cadre, including xi, understand this, care about this deeply and are willing to intervene as needed.

Michael Sullivan: Let's jump from geopolitics to economics. If I was to walk out into central London, um, how will AI change economies? How will it change the way we live, the way we get healthcare? Will it make us healthier, live longer? Uh, will people lose their jobs? Just give us some of the headline, uh, effects you're seeing.

Azeem Azhar: Well, in truth we don't know. Uh, it's a general purpose technology and uh, that means how it will deploy in the economy is a little bit unclear. But there are some things we do know. It's going to get much, much cheaper, uh, to use AI within enterprises and that AI will get more capable. What that means is that lots of work that involves certain classes of thinking will become much cheaper and companies will become more efficient at it. It will almost certainly generate more work as it always does. Um, and you will start to see, although I think it's too early to see it beyond a few, uh, specialist firms, productivity benefits. But it's not just productivity benefits, it will also be improvements in service culture. So if you talk about health care, for example, um, a world where you have a bunch of AI sitting in healthcare looks quite different to today. We get telemetry from smart devices, we get MRI scans turned around much faster. There's much less likelihood that anything will be missed. Your doctor will have a medical scribe with them, um, meaning that notes are very, very well taken. And complementing a doctor's own analysis will be a series of AI agents looking for exceptions or outlandish possibilities or errors that may have emerged. So I think you should see both the speed of care, the quality of diagnosis, the quality of follow up, the methods we can use to make sure people comply with their treatment regimes, all lend themselves to be improved by systems that have a little bit of AI.

Michael Sullivan: So you may, for example, see uh, a dramatic improvement in longevity in emerging markets where healthcare systems are not maybe as good as in other countries where life expectancy is much lower.

Azeem Azhar: Right.

Michael Sullivan: Um, that can really jump them forward, I guess.

Azeem Azhar: I mean three or four things. Um, the ability to be able to do frontline diagnosis on your Android device would be one. Uh, the second would be that, um, the ability to better manage cold chains and supply chains through AI based optimization. So, so there is more medicine available because a lot of it is getting lost on the way. I mean these things all Add up.

Michael Sullivan: Yeah, yeah. Um, Azim, we're going to talk about investment and the, and AI. I mean in the last three years, investment in AI is, it's just dominating, uh, you know, parts of growth. PE dominating venture capital.

Azeem Azhar: Right.

Michael Sullivan: Do you want to start by just giving us a kind of a very landscape view and what you're seeing across the investment, uh, uh, scene. And then we can dig into little.

Azeem Azhar: Let's do it this way around. You said venture capital and pe. So we'll start there and then we'll go to the big one, which is infrastructure. So you know, in venture capital, uh, there has been, I mean essentially at the seed level, right at the small end, Y Combinator, everything is AI, uh, and everything's being built with AI. And the teams are much smaller, they're getting to revenue much faster. They're growing much more quickly. Uh, and you know, you can Google stories of firms, firms getting to $50 million of revenue in their first year of operations. I mean it's remarkable. Then there are the very big rounds in the large companies like OpenAI and Anthropic and Perplexity, which are the bulk of the capital that's going out of venture. And the thing is that these companies are growing incredibly quickly. Uh, OpenAI went from, sorry, Anthropic. Pardon me, Anthropic went from a billion dollars of revenue to $4 billion of revenue in less than 12 months. At that revenue growth rate, what multiple are you going to apply? It's a company that when you peer below the hood, has started to dominate certain subsegments of the market, suggesting it's got a pretty tough beach hold that particularly in the coding tools market. Uh, so those companies are pretty frantically growing and are in those investments merited? Well, my, this is not investment advice, but my, the question I would say to people is, well, make your decision now. Would you think it's worth it or not? Write it in an envelope, open up in five years. If you're right, pat yourself on the back. And if you're wrong, you can send me a tenner. So that's one side. Then you've got these staggeringly large investment rounds in two OpenAI founders companies, Mira Murati and Ilya Sutskeva, where they don't yet have products. And that I think is showing the bet that exceptionally good talent will deliver outsized returns. I'm not sure I would have been comfortable with either of those deals. $30 billion valuation for a pre product company is pretty juicy. So those companies are genuinely growing very, very Rapidly. And just to give you a sense of um, at the end of December 2000 or the start of 2023, we were spending less, my team, less than $5 uh, a month on AI because you couldn't get paid for OpenAI. We had some API access. We're now spending well in excess of $1,500 $2,000 a month in a team of six, uh, in just two years. So that is a, going from $60 a month to 24 grand is a big load of revenue expansion. And people's uses are actually gated, um, not by how useful the tools are but the amount of time they have rather than anything else. But we need to talk about infrastructure.

Michael Sullivan: Let's talk about infrastructure. Yeah, because I guess for a long time it's been sort of a boring industry, but it delivered very solid returns. But what the sense I have is that there's a new industry structure being developed and you have the uh, AI teams, you have the models, then you have the energy sources. So maybe just paint a picture of the uh, AI infrastructure and the investment in that.

Azeem Azhar: Absolutely. So what you need is you need loads of chips. The chips get really hot, they need lots of cooling and the cooling and the chips need power and they need to be in buildings. And you generally have to build new buildings because the racks are so heavy and they pouring out so much heat that you can't.

Michael Sullivan: It's an engineering problem.

Azeem Azhar: It's an engineering problem as well. Right. And yes you're right. Infrastructure we'd normally think of as in terms of being a, ah, sort of steady state, uh, low volatility style business. We're going to see 3 to 5 trillion dollars minimum. I will review my estimates when I get back from my summer holiday. Uh, going into the chips, the networking, the cooling and the buildings, uh, before 2030 and that is there to serve this demand. Remember Google's demand for AI increased 50 fold and at the same time they've made orders of magnitudes of improvements in efficiency. Uh, so the question here is how should you look at this as an infrastructure investor? On the one hand these things are not like the water pipes in London or the sewer that's been in place bazalget sewers for 100 years. GPUs according to accountants depreciate in five years. Although some people are pushing that out to six years now if you talk to people on the frontier they'll say well within three years they're burnt out. Okay, so this is exponential time. You're going to have to juice that asset in three years. Uh, and so I think that that makes it complicated if you're an infrastructure investor. But I think the secular question is, when does this actually slow down? And a lot of us are sitting there thinking, well, I remember the shale gas boom and suddenly it just fell off a cliff. And I was left holding onto these assets that were going to go nowhere. Or you may remember the telecoms bubble, where effectively everyone lost their shirts, but then infrastructure investment growth rates exceeded revenue growth rates, whereas that's not happening in AI. And I would contend that even though by 2030 there'll be 10,000 times as much computational capacity in data centers to serve AI as there is today, it's probably not going to be enough. And that's why Mark Zuckerberg has gone off and said, I'm going to build a data center the size of Manhattan. Uh, because they're looking at the exponential curves, they're looking at, uh, how prices decline, they're looking at elasticities of usage. And they can go off. Sure. With wide M error margins and get a sense of where this may end up.

Michael Sullivan: And, um, by the way, governments don't have the money to do this. So the money will come from private capital, from PE firms who get it from pension and firms and maybe sovereign investors as well. Yeah, yeah.

Azeem Azhar: I mean, I mean, Zuckerberg doesn't have enough money on his balance sheet to do this, which is why they just did a debt facility for particular data center. What he does have, and I think this is an advantage over, um, Sundar at, um, Alphabet and Satya at Microsoft, is he basically controls a company.

Michael Sullivan: Yeah.

Azeem Azhar: So he could take the bet and you've just got to come with him and you have to go with your shares. Uh, or not. Uh, but I don't think their balance sheet is deep enough to do this on their own. And I think that creates an enormous opportunity for people who are willing to supply that liquidity and the risk to somehow get involved in putting these together. But I do think governments can play a role, and I do think even the British government. Even the British government, uh, could play a role here by essentially convening, promising stability, promising rapid decision making. And I think they should be smart and figure out how to get some of the upside. Why don't they take some of the sweet equity?

Michael Sullivan: Take some of the equity. Exactly. Um, AI can be very disruptive for certain industries. So at moonfair, we notice a couple of things. One, a lot of investors are using AI probably just to speed up the process, speed up elements of diligence. But also a lot of the industries that say PE has invested in. You mentioned healthcare, uh, finance payments is another one, uh, are being changed, uh, and actually in some cases made more dynamic by AI. Do you have examples of industries where there's been an equilibrium for quite some time but that's now being obsessed by the advent of AI?

Azeem Azhar: Well look, the challenge with these technologies is that you can often spot the winners but you can't spot the losers. And the losers like Blockbuster uh, are going to go to zero. And if you've got a portfolio company in pe, one of the things that AI is certainly going to help you with is bringing efficiencies and bring your costs down pretty quickly. Where AI is going to harm you is if AI enabled firms actually redefine what the service proposition to the customer is. It didn't matter how efficient Blockbuster uh, got, if you're still shoveling VHS cassettes out, you're going to fail. And I think the difficulty here is as a PE firm, um, your portfolio is built on probably 10 year bets. You're going to hold the company for three years but you need to know someone's willing to buy it and they need to better that someone will buy it off them later on. And I think that that is a little bit of a challenge. I think it's a bit early uh, to found uh, equilibriation, uh right now and I think the signals are very, very hard to read. Uh, so the tech industry is news abounds with layoffs. But the tech industry over hired during the period of 21 to 22. Right. And there's lots of other things going on. But then let's go off and look at the finance industry. So because of uh, ETFs and algorithmic trading and systematic trading, the vast bulk of equities and other asset classes are moved without touch of human hand. And that's happened over a 25, 30 year period when we go back to the first ETFs. Right. But uh, the finance industry employs many more people now than it did. So what happened?

Michael Sullivan: Technology usually either displaces jobs and, or creates jobs.

Azeem Azhar: Right?

Michael Sullivan: Uh, I think is the.

Azeem Azhar: Yeah, exactly, is the measure. And so I think it's hard to your question about, you know, who's at risk from a firm perspective is challenging. I would say that you know that there are areas where one would feel a little bit more vulnerable in media, in certain parts of insurance and finance, where you're in the back office, you're perhaps not dealing with risk but you're dealing with, uh, paperwork. Uh, there are also, um, issues. If you are in retail and you've been very dependent on search volumes for your traffic. And then there are industries where you might sit and think, well, these guys are probably okay if you own. I'm in love with heat pumps at the moment, but if you own a bunch of district heat pumps across northern European cities, I think you're probably all right. You're probably all right.

Michael Sullivan: Good. So your vision, I guess, is that there will be much more private capital, sovereign capital, going into, I would say, the AI ecosystem, infrastructure, energy, et cetera, um, in Europe. Do you think we need to, uh, change our capital markets? Go back to my question on savings, investment, union. Do you see more, uh, public private capital and pensions savings being released into some of these areas?

Azeem Azhar: I don't know enough about how across Europe, pensions think about, uh, where they can play and the degree to which they can move into other structures. But my sense would be that the infrastructure needs to be built, therefore it needs to be financed. And you have to go off and find your sources of finance. And if you're a finance minister in Italy or Spain or Portugal or Belgium, you need to look at what's holding capital back.

Michael Sullivan: Um, can we dive into the.

Azeem Azhar: Let's do that.

Michael Sullivan: There's some really fascinating ones. Uh, so we've had one, we've discussed Kimi K2, uh, will it outpace closed models? But there's a question here on a company called Base44, which I don't know. The question is, uh, will we see more, uh, AI bootstrap companies like Base44? And, um, will we see smaller teams within companies? And maybe you can explain what base 44 is.

Azeem Azhar: I don't know what base 44 is.

Michael Sullivan: Okay, right.

Azeem Azhar: Either. Uh, I'm behind on the question, but I think there's a general question about how small can teams be?

Michael Sullivan: An organizational structure, which I think is very interesting.

Azeem Azhar: Yeah. So we're seeing, um, AI companies. I mean, if you think about OpenAI's revenues or Anthropic's revenues, for the size of company they are, they have a very small workforce and their compensation matches. And I think you see that in emerging areas consistently. The same is true with Jane street or G Research. Lots of revenues, not many people. Uh, and I'm seeing this in the startups I invest in, that the ones that are built from the ground up in the last year or2 where AI tools were good enough, have much smaller headcounts, uh, have a smaller core team. And some contractors helping them because they're going so fast by using AI tools. The question also asked about small teams within larger companies. And I don't expect large companies to do much that's particularly interesting. There will be some stories of it for all the same reasons that we've seen with every other technology revolution. The managers of the large company up to the CEO, uh, have been hired by shareholders to do a particular thing with a particular set of priors in place. And if the shareholders want to get involved in something funky and exciting that's AI based, they can move their capital to that.

Michael Sullivan: Can I dig into that a bit? Because I've had this conversation with a couple of big companies and one idea I've thrown back at them is that they should have a corporate venture unit, uh, that if a big company is too, they admit it's just too slow and clunky to do innovation but have a team m investing in relevant, uh, technologies that can be used in the business. Is that an answer or how can big companies remain relevant?

Azeem Azhar: Well, look, corporate venture has matured now and there are lots of companies that effectively, uh, run corporate venture outside of the traditional tech industry and Microsoft and Google. But you have Airbus as an example. You've got BMW, um, you've got a whole slew of firms that do that. The question is, what purpose does it actually perform? And I think that then really depends, not just sector by sector, but firm by firm and management team by management team. Can you take advantage of, of the bets that you've made externally? Um, and they're not normally financial bets. They should be bets that have a financial return but also have a strategic implication because it's a new material or you can use in your aircraft or it gives you access to, uh, a new market. But that doesn't really swing the ship. And for many companies, the challenge that they're going to, to face is going to be about what, uh, do their customers want, what is the product offering the customers want, um, and how well can this company end up delivering it. And if you just run analogies that have happened during the, you know, the digital wave, I think newspapers are really salutary and local ones in particular. So in the early 90s, local newspapers, real customers were classified advertisers. And then the Internet came along with, um, Gumtree, Gumtree and Craigslist and so on, and those customers went somewhere else. And I think that this is going to be the challenge with a lot of, uh, AI products where the touch point, many of our touch points will end up changing because we want to have this de novo uh, product. There may be physical products, let's think about furniture or we think about cars. We think, well, how could AI replace that? I want this chair to sit on. I don't want a large language model. But this is where AI starts to challenge those businesses as well because they've built an assumption of how they acquire their customer through Google search. How will they price, probably with fixed pricing, how will they persuade the customer to buy and ultimately how will they serve that back to the customer? Well, that front end now is really vulnerable to much, much higher marketing costs because AI based tools and AI based channels will be occupied by people who know them. And so while it doesn't actually completely disarm the value of the chair, what it does is it slices away a whole bunch of your margin because your marketing costs have doubled, tripled, potentially quadrupled.

Michael Sullivan: Got it. Very interesting.

Azeem Azhar: Sorry, long answer there.

Michael Sullivan: No, no, fascinating stuff. I have quite a punchy question here. Azazim is an investor himself. How does he spot the difference between true AI breakthroughs and overhyped noise when allocating?

Azeem Azhar: Well, the way uh, that we invest is that the bulk of our capital is in a less saucy uh, set of structures. Um, and then I will invest in early stage companies generally through very, very uh, qualified referrals. Um, I've been lucky. I've been in the tech industry for 30 years. I've been looking at AI for a decade. Uh, and where I end up looking, I look for uh, the intersection between a team and the founders and how good they are, what their track record is. Things like Having worked at DeepMind or OpenAI are a good signal. Uh, and the market they're going after because things will change and I look for that intersection and there are always exceptions to what one chooses to do. But I tend to be right now further down in the infrastructure, the models and the tooling than I am right at the top at the application end of the consumer or the top end of the enterprise end. So the vast bulk of what I've done has been can you deliver a particular model or a toolkit for building customized models for enterprises that are very customized for their specific requirements. That tends to be where I operate right now. I think at some point there will be more opportunities in applications and consumer applications, but there are sectors that I just don't understand. I don't understand gaming and entertainment, I don't understand retail. So ah, I wouldn't stay away. Yeah, I'd stay away from that, yeah, very good.

Michael Sullivan: Um, someone has come back on our discussion on the EU AI act and they've said provocatively, is there anything useful in this? And actually are some of the safety provisions from the point of view citizens, uh, will they in time will they actually be seen to have been, to have foresight?

Azeem Azhar: Uh, well, you know, I think the challenge here is that the world has descended into a knife fight. We're not saying it, but that fact is what effectively it is. And we've got strong evidence that uh, E Privacy and GDPR really, really harmed smaller and mid sized companies that are the backbone of Germany's economy, for example, and they made not an iota's difference to Google and Meta. Uh, and when GDPR was rolling out, I was a product manager at a $7 billion market cap European company. We had to spend tens of millions of euros making ourselves GDPR compliant and that the harm is now manifestly visible. And on top of that we all have that cookie banner that takes 400 million hours a year in Europe to click away, that serves no purpose. Energy, um, and energy and time. And Europe is not so productive that it can afford to squander 400 million human hours a year. So I think we've got this sense that both the environment has changed and that what was conceived of last time ended up having these tough second orders, uh, costs. There may be high risk things that may emerge if the EU AI act is curtailed or delayed. Um, so we'll get to avoid some high risk edge cases while being in fundamentally less dynamic and much poorer economies. And I think that that is the hard reality that as I said when we talked about this, if Europe wants its values to stay in the game and be, be relevant, it absolutely has to stay in the game. I mean the US is dictating a lot of this. Not because they have wonderful values, not because they're super nice people.

Michael Sullivan: This is the future because.

Azeem Azhar: And they matter. And they matter. So I think the first thing to do is um, figure out how do we stay in the game just on that.

Michael Sullivan: I mean, do you think that the Trump White House has a more coherent AI policy than the Biden one, or will this be driven by the industry?

Azeem Azhar: I think coherence is um, not a word that you generally associate with the current um, administration. I think directional, I think vigorous, I think willing to pivot. These are all agile, these are all great attributes, uh, that one might want in a character and that will be the overarching direction. But there's a lot that goes on that's unclear. So the new regulations they're pulling out to prevent woke AI, which is um, a little bit like uh, King Knuth's advisors telling him he can turn back the tide. That's a very difficult thing to do. I think the way the Department of Energy is promoting the use of coal as the energy of the future is also something that's directionally a bit odd. Um, so I think that what's coming out of the US is just headroom for businesses to get on with it.

Michael Sullivan: Okay, we have a ah, couple of minutes left. One thing we haven't talked about is data.

Azeem Azhar: Yes.

Michael Sullivan: Um, because it strikes me in my old fashioned kind of going back to when I was doing regression analysis on parts of economic data, uh, uh, the use of AI is really defined by the quality of data. And maybe I'll just get you to talk about this for a bit and then talk about specialized data sets. Will these become more prized within say fintech companies, within healthcare and military, et cetera?

Azeem Azhar: Yeah, I mean data is one of the input sources into AI and higher quality data certainly helps. I think one of the reasons why Claude is so good, we've learned this through various uh, legal uh, documents is that they heavily trained and over trained on a corpus of books and books are very high quality data. You know, brilliant author like you and fantastic editors have turned this into high quality uh, insight. And so we are going to start to see a shift towards needing more and more experts to contribute data and the big labs will start to do that I think at the same time there are many sources of data that are valuable economically. Uh, genomic data, uh, multi omic data in the healthcare space, time series data and industrial um, environments and in financial environments where the data sets are not available on a Google site or a BitTorrent. And so that makes them quite interesting for those particular areas. But the last part of that data is what will happen in terms of the settlement with IP owners, particularly the publishers. And I think that this will ultimately end up with there being some kind of payment from the AI companies to the, to the publishing industry. You uh, know it's going to be interesting but I don't think it'll change anything.

Michael Sullivan: And will, for example banks or payment companies uh, begin to realize the value of the data they hold on consumers. From the point of view of AI and keeping that in house and doing analysis.

Azeem Azhar: I'm not sure how banks will necessarily uh, act. But we certainly know that where MasterCard and Visa understood this from 20 years ago. I think the other question though, is whether, given how good AI systems are at characterizing individual behavior, whether that data really adds enough alpha. Uh, now that the models are quite as good as they are, I don't know. We could find out, perhaps.

Michael Sullivan: Azim, we're coming up to the, uh, end of the year. It's been a brilliant discussion. Can I just maybe ask you to give not so much a summary, but a couple of pointers for the future, Whereas, I mean, it's happening so rapidly. But, uh, what should we look for in terms of the development evolution of AI?

Azeem Azhar: Well, we are just past the foothills, uh, of this. And because financial capital and because people's expectations move more quickly than even this technology does and it can be deployed, you will see, uh, you know, divergences between what's really valuable and what people are willing to pay. So there will be ups and downs. Uh, no question about that. But, you know, over the medium to long haul, we're going to be using much, much more of this. I do think that deployment in market takes longer. Even with a sophisticated model, or perhaps especially with sophisticated technology where it can touch on so many other things. The way in which it'll get introduced will take a little bit more time. And there is a tendency, particularly if you read my newsletter, it's a fault of ours to think that the world is moving right at that speed. We're sitting 1 millimeter behind the bleeding edge, but the rest of the economy is on the trading edge, uh, of that.

Michael Sullivan: So, Zane, this has been the first, uh, Moonfare, uh, insider talk. You set the bar incredibly high. It's been a brilliant discussion, uh, on such an important topic. So from all the team, thank you very much. And, uh, thank you to everyone, uh, who's dialed in. I hope you've enjoyed this.