00:00:15
Speaker 1: Welcome to tech Stuff. I'm Os Voloscian. We're living through a pretty extraordinary moment, a technological turning point that is or inspiring for some and infuriating for many. And it seems like no opinion is complete without an assertion about what AI is doing to the economy. Is it a boom or a bust? Is there a crash in the odds. Some people think it's the future and the only economic driver that matters, and others say it's all a big bubble sending us hurting towards collapse. But how do the economics of AI actually work? And what does the enormous amount of money being poured into this technology mean for our day to day. Our guest today is Azimazar. He's an entrepreneur, author and founder of Exponential View, a research center and substack where he just released a report called the State of the AI Economy. As he welcome back to tech Stuff.
00:01:05
Speaker 2: It's great to be back, Ols, Thank you.
00:01:07
Speaker 1: One of the things that occurred to you as preparing for this conversation today was obviously thinking about situational awareness. You know what started as a manifesto and became a hedge fund and adventure fund and then had it struggled a couple of weeks ago. But you know, there's still very much a going concern. I mean, have you thought about the exponential Review Fund or do you like to sit mainly on the kind of observation analysis side of the table.
00:01:30
Speaker 2: Well, let's talk about situational awagas. I mean credit to Leopold Aschenbrener who wrote that essay a few years ago. And I was a little bit skeptical about the numbers. I think they were too fast. Yeah, he's generally been proven to be a very very good predictor of what would happen. And what happened in July was that his fund imploded. I think it lost thirty five billion dollars of value in a month. It's still alive.
00:01:52
Speaker 1: He kept his xanthropic stake.
00:01:53
Speaker 2: Yeah. Well, early investors have still done really well. They've made like twenty times their money, which speaks a lot to what he had done. I mean, he had basically run the problem of some leverage is okay, too much leverage is like over salting your food, and it's going to be resulting in a really bad outcome. I think for what we do, I mean, I do think that it's very helpful to invest when you are looking at markets, because once you have skin in the game, you are much much keener about the things that you say, So you can't be slap dash and be in the peanut gallery and just sort of shoot your mouth off, like where are the steaks? Where are the stakes? In all of this. But I think it's really important in this particular market to get in amongst the weeds in as many ways as you can, because this is not like looking at a market that we've seen before, so you really have to equally. I talk to bosses of big and large companies all over the US and Europe to hear what they are saying about about AI, because there's just no rule book for it right now.
00:02:56
Speaker 1: I think more than fifty percent who you surveys so that they believe their job will de end on successful integration. But let's talk about this report, the state of the AI economy. What's it about? What we were reporting on? What was the problem you're trying to solve?
00:03:10
Speaker 2: Well, the main problem is is it real? Is there any money in it? How much is being spent? And it's a very very difficult thing to measure, as you often find with early markets here were used to looking at mature markets. So we went out to solve that problem. How much are customers, consumers and businesses spending on AI around the world and can we get a real economic value added number out of that? Now, explain why that's important. If you're spending twenty dollars a month with open Ai, they will be spending a certain amount of that with Microsoft, who runs the servers that serve up chat GPT, so that twenty dollars may also end up with say ten game to Microsoft. A naive count will call that a count that as thirty dollars. We will count that as twenty and apportion it. Of course we have the total number. So that's a job we did. And what we want to understand was how much is really been spent by businesses at the top of the funnel, which is then distributed to all the companies that are, you know, as a little stack serving each other to get you that response to your chat GPT query.
00:04:23
Speaker 1: And how do you do that? How do you deggregate or deduplicators you put it where the money goes?
00:04:28
Speaker 2: Well, you know this is where it gets interesting. You have to use a lot of different strategies to piece the picture together. So a lot of this is happening in the private company, so they don't have to disclose, so we will go back and reconstruct their accounts from public information and other information that has been leaked to the press. We scan over one thousand AI startups. We look at when they disclose their revenue. A lot of this ends up flowing into what we call the hyperscalers. Companies like Google and Amazon and Microsoft. They make some public disclosures and then ultimately all of this runs on chips, and so we know how many chips have been sold, and so through all of that you do a process of triangulation and you say, well, these things generally all have to agree. And that's the big part of the work. And it's not one thing, it's many, many, many different things that come together, these different skill sets, different data to give us that number that we feel reasonably confident with.
00:05:33
Speaker 1: And for the you know, for the average person who's interested in these topics, but maybe isn't such an insider that they're a paying subscriber to Exponential View, Like what what did you find that matters most? What does what everyone need to know from this report.
00:05:47
Speaker 2: Well, the first thing is you can download the reports for free on Intelligence on exponential view dot dot co. You may not get some of the updates, but it's sixty five pages long, so there's more than enough to send you to sleep for several days and maybe keep you awake if you're so inclined. I think the key thing here is that there is real spending happening. We've exceeded the one hundred billion dollar per year revenue number across the industry. It's actually as of the year to July twenty twenty six, it's one hundred and twenty six billion dollars. And that that is a number we've got to about three times faster than with mobile advertising, or the cloud or the internet itself. So it really is the fastest revenue ramp in a in any technology wave, and probably of any sector in history. And that one hundred and twenty six billion dollars is still growing very quickly, roughly three times per year, like two hundred percent additional per annum, and we were really surprised about that. We had expected that number to be slowing down from last year.
00:06:59
Speaker 1: It had so that was my next question, where does this money come from? And how much is it from the AI firms themselves, how much is it from VC backed companies who may be spending a head of profits, say, how much of this is like true demand side spending versus like supply side stimulated spending.
00:07:16
Speaker 2: Yeah, it's super important to separate those out. So the AI companies will spend on compute to train their models, and we capture that spending because those are real dollars that go into the pockets of Microsoft and Google and Amazon and so on. And they will also, in the case of open ai, be spending for those that free tier of chat GPT that most people are on that they don't yes yet monetize fully through ad revenues, So we reckon that. That's around about twenty percent at the top end of the revenue number, but it's a proportion that is dropping significantly. So if we've looked at this year and a half ago, that would have been closer to sixty percent. Since twenty percent of one hundred and twenty six billion dollars call it twenty twenty five billion dollars. And then how much of that is other venture backed startups? That is, you know these startups who raise their capital from these fancy Silicon Valley firms. That number of from the venture backed startups is you know, roughly in the order of four billion a year. Now, the question is should you discount that? Because venture back startups often do very well, and I think it's perfectly reasonable. For example, if you've got a venture backed company in San Francisco, that the blue bottle coffee shop opposite their office should be allowed to count as revenue the coffees that the company buys for the team all hands on a Friday, Right, I mean, so American capitalism and the rules of accounting have worked in a particular way, and we should be careful about where we draw that line. There is of course a question about like the sustainability of any of those companies, but we know in general for Formerllion people in the US who are employed by companies were backed by venture capitalists. Right, it's not insignificant. So then you get to the kind of cluster of relationships that have emerged from say Nvideo. So in Video is a company that makes the chips that run about eighty percent of the workloads. And there's an interesting thing here, of course, so Nvideo has gone out to ensure that the people who buy their chips and who serve up the compute have got the ability to buy those chips and serve the customer demand. So that is you know, supply side support of an emerging sector. It's really not uncommon. I mean, there are very very many sectors, and we often go back to the start of the car industry where Ford in general motors extended financing to distributors and you know, literally car salesmen and also to consumers to buy cars. When you get into a situation where a core supplier in new sector has got the strongest balance sheet and the strongest credit rating and the sector is growing very very quickly.
00:10:07
Speaker 1: It's really it makes sense they become the bank.
00:10:09
Speaker 2: It makes sense they do that. Now. Now the point is that that can also go bad in the same way. But I don't think that you can say, like upfront, but this is always a bad thing to do. This is not like you know, pilfering from the till. What it does do, though, is it does introduce a new kind of class of risk in the system that doesn't exist if customers are always paying for everything that they buy.
00:10:32
Speaker 1: So how much is this approaching two hundred billion dollars, and your run rate of AI revenue comes from like blue chip corporations, how much of it comes from super users like you know, perhaps you or medium users like me? Like, well, what's there? What's the split outside of the AI industry and the venture industry itself?
00:10:51
Speaker 2: If we break that two hundred down, which is the annualization of the run rate, you know, you're looking at eighty percent of that being enterprise rather than consumer. You're up at one sixty of that one sixty maybe twenty five to thirty at the top end is AI companies spending on themselves and venture backed companies, which leaves you in arranged, let's be conservative, one hundred and twenty billion dollars from enterprises on a run rate, so that was effectively like ten billion in the last month. And then through that it's pretty it's pretty diverse. I mean anthropic and open AI who the ones that we know that we all recognize have got hundreds of thousands of commercial customers and they've got dozens, if not one hundreds, spending more than a million dollars a year. And these are companies that we might think of as being sort of advanced tech companies like you know Uber and coinbase, but it's also you know, the big banks and the like. And then within that, of course, it leans towards technology forward sectors and technology forward companies. So I'll give you one example that may surprise people. If you look at companies in America that have been bought by private equity, so not venture capital that which is all the sort of rocket fuel, and not cash producing public or family owned businesses, but once owned by private equity, their levels of investment in AI seem to be much lower than the average. And the rationale would be that if you've been bought by private equity, you need to know every dollar that takes you off the operating plan is going to give you two dollars back. And if you don't know that, the owners won't allow you to do that experimentation. And I think that speaks a little bit too that where in general bosses are about how comfortable they are about the return they get from AI.
00:12:40
Speaker 1: Right now, Yeah, you had a headline for one of your newsletters recently, the AI adopters success and failure look identical at first. So I had three questions, obviously for us, on being why, the second being how much patience does the market have for that at a time where money is becoming more expensive? And the third is when does that change?
00:13:03
Speaker 2: Yes, I mean that second question, by the way, is a brilliant question. I'm going to give you credit. This is why you are the host of this podcast.
00:13:12
Speaker 1: And takes you see.
00:13:13
Speaker 2: We're really early in where companies in particular are spending with their AI and that could be one of two things. It could be that they're going to continue to have increasing success and that number is going to rise. Or it could be that they're still experimenting heavily and everyone's experimenting and these are experimental revenues that might not continue at this growth rate once companies wise up. And those two scenarios I think are still live. I think the balance of evidence is that it's the former, which is that this number is going to continue to rise. But why do they look the same? They look the same because when you get started with a project, whether you're going to be successful or unsuccessful, you have to make the same kind of upfront investment. And if you are successful, you will spend money on that project and then you will turn the corner and you'll start to make money from it. But you won't have got to break even. And a sensible management team will say we've done this once, now let's do it three times. So just as the first one starts to pay back, you've now got three projects that are in the red. And then they'll say let's do it ten times. And so the depth of that red, which you call a JCOB, will will continue. Now, an indisciplined company that is running into all sorts of failures might not expand as quickly, but they might let unsuccessful projects run for longer and longer. And that means for a couple of years, two two and a half years, a successful firm that has really disciplined in its scaling, that runs a portfolio will look quite similar to affirm, at least from the accounting basis, to a firm that's being a bit indisciplined and isn't having success.
00:14:57
Speaker 1: And so the second question about the patients for this, and would you be concerned that the patients for these two year you know, in the red investments at a time when money is becoming more expensive me may dry up and that the revenues from the aim may start to go down as a result, which could create the cascading crash that people worry about in terms of all of the circular deals and stuff you know.
00:15:20
Speaker 2: It might do, But who would struggle with it? I think the really interesting question here is if you're a particularly if you're a public company, Wall Street only really cares about your earnings, your earnings guidance and the fact that guidance is going up and that you're meeting that guidance. If you're starting to do a deep investment program, that's going to have to come out of earnings, and investors are going to have to believe that in that story to avoid punishing your stock, which in turn makes all that capital more expensive. And as we know, interest rates are also trending upwards. Now there is a group that is going to make money out of all of this, which are the people who own infrastructure because they they have made their investments. I mean, Google has used all its free cash flow up right to buy data centers, and those data centers are predicated on the demand from corporate America, which will start to come in the next two or three years. So those companies will start to make really significant profits at that time. And I think there is a tension there that you've identified, which is what's the patient's actually going to be. There is a there's a line in the TV series Silicon Valley from one of the characters I think is called Russ Hannigan, and he's sort of saying, you've got to be pre revenue because in pre revenue, you're sold on a dream. If you have revenues, you're sold on reality. And I think we're getting to that stage where there will be a disjuncture where the hyperscalers are big tech companies will be showing, as they are eighty two percent revenue growth in Google Cloud in the last quarter, increasing revenues on these high margin products against which still be measured, but Main Street will still be on the jam tomorrow model where they're investing to show those future returns. And I think there is a little bit of a tension there that I haven't quite resolved in my head.
00:17:11
Speaker 1: Right now you're writing the report, the open question is where the cheapening artificial intelligence can create enough volume and margin to service the build out. One of the most interesting things which really caught my eye in the report was that the kind of proportion of spend in data centers is rapidly migrating from spend on concrete to spend on chips. And we had Jasmin Sun on tech Stuff a couple of weeks ago talking about this kind of deep doubt in the Ross Belt where these data centers are popping up that this AI revolution is sustainable and the concern that these may be left as kind of artifacts of a corporate boom and bus cycle. But what do you think and what's your answer to that question? That does cheapening artificial intelligence create the required volume to continue financing this incredibly expensive infrastructure buildout.
00:17:59
Speaker 2: I mean, in short, it does. Yeah of you is that mainstream companies are really early in their adoption. Where early in the adoption cycle in two ways, one is how many companies are using it and within it company, how mature are companies in their use And the shift is really is going to be palpable. I mean, these chips are currently running in data centers at extremely high utilization. To give you an idea, about eight or nine years ago, Google launched their first and second generation chips for AI they called TPUs, and since then Google's AI compute capacity has increased by at least a factor of one hundreds every one hundred times more compute, and yet they're still at full utilization for those very very first chips, which are you know, everyone knows in the computer nine year old computer is really old. And that speaks to the volume of demand that they are that they're dealing with right now. And there's something that Andy Jasse, who is a CEO over Amazon, said about six months ago about all of this, because a lot of people are saying that these guys have lost their heads, they've gone they're starting to dream things that aren't credible, aren't grounded. And what he said was, look, as Amazon, as Amazon Web Services, We've got a lot of data about how we plan for capacity and how we plan for the future, and we've been pretty good at it over the last twenty years. And I think you can see that Amazon has been pretty good at this. So maybe you should look at the investments and the commitments. So we're making from the lens of looking at people and paraphrasing him, by the way, but from the lens of a company that has been really, really good at this for two decades, and say, what's more likely that we lost our marbles or that we're still really good at this? And I think independent of the transformer model that underpaid largelanguage models that underpins Claude and chat GPT, we were already seeing a sharp move towards the digitization of business in general, what they used to call big data, what we used to call big data, what we called spreadsheets, what we called digital supply chains. I mean, these things were already happening, and that trend was continuing, and maybe AI has brought it forward a year or two. So I don't have a question about whether the economy has got use to pay for this over the coming years, and we'll use all of those chips. But I think the interesting thing that Jasmin went off and did was that she identified and evidenced through her anthropology a lot of intuitions that I think a lot of us had about resistance to the data center, which was that it's really about power politics, distribution, and disconnection. It's not necessarily a single monoline attack on a all the data. Said to me, the data set to ends up being a physical instantiation, something you can point to that says many things are not working. And here's what the concrete example in front of us that is being planned by these people over in Washington State or in California that encapsulates all of that.
00:21:41
Speaker 1: Now a quick break switching topics to one of our favorite sponsors, Vital Proteins. As a fan of science and technology, I like knowing what I'm putting into my body and why We've talked on tech stuff about AI enhants, blueberries, superhuman athletes, and health trackers. And while I might not be joining the enhanced game anytime soon, I still want to make sure I'm doing what's best for my body. That's why I love Vital Proteins. I always know exactly what I'm getting in their products, like protein and collagen, and what they're leaving out like sugar and artificial sweetness. I especially love the convenience of their new Collagen and Protein Shake. With Vital Proteins Collagen and Protein Shake, I can grab and go without too much stress, and they have both vanilla and chocolate, so I even get a bit of variety. Not only that, with thirty grams of protein including ten grams of collagen peptides and every shake, I'm confident that I'm supporting my hair, skin, nails, bones and joints for years to come. But don't take my word for it. Try the new Vital Proteins collagen and protein shake for yourself. Thank you again to our sponsor Vital Proteins. Why have chipstocks been beaten up in the last few weeks?
00:22:57
Speaker 2: H yeah, wow, they really have been. So you know, one of the things that's gone on has been that there's been a tremendous run on chip stops. They've done really really well over the last two or three years. And the Korean market is really important for the semiconductor industry because it of a company called Eskehinix, and because of lots of the supplies to the ship industry being there, and the Korean market is very consumer led and consumers are allowed to borrow very very heavily, and consumers got very exuberant. And in fact, when the slightest wabble showed up in early July and the market dropped, at one point, the Cosby was down thirty three percent, which would have been its biggest drop ever.
00:23:40
Speaker 1: The Cosby is a South Korean strug exchange or.
00:23:43
Speaker 2: Yeah sorry, the South Arian soups is what in twenty five Korean households got what's known as a margin call. So the margin call is the equivalent of Tony soprano knocking on your door saying I want my money.
00:23:54
Speaker 1: Now got a situational awareness as well.
00:23:57
Speaker 2: Right, yeah, So that that was a sign of a lot of overheating. And now today as we record, the market is down again after a really bad day yesterday. Or two things right, One is there's a sense of unease about this particular trade because when you have a moment where there are a lot of high prices and cost of memory has gone up very very quickly, high prices cure themselves because people look for alternatives and there's more innovation. And in a market that involves a commodity like memory, if high prices do get cook cured by an alternative, you're left with an enormous oversupply. And the story of the semiconductor industry, particularly the memory industry, has been patterns of their own boom and bus going back forty years. And there are enough investors who remember that to feel a little bit skeptical and want to take them at a distance and take some money off the table.
00:24:54
Speaker 1: What are they scared might happen? There might be a new round of chip innovation that makes all of these chips. The evolutionizes how memory works in AI data centers, or what's what's the fear looming over this?
00:25:05
Speaker 2: Well, I think I think it's that fear, and it's it's other things. It's do prices get so high that they somehow reduce the demand of the thing that is keeping prices high? Is it that prices are so high and that the economic incentive for software optimizations to be more efficient with memory use rises, and so that reduces people's forward forward forecasts. And this is a little bit like a game of musical chairs in a way, right. You don't want to be the last person when the music stops, because then you sit on sit on the floor. And I think that you know, investors are smart. They look further out and they say, oh, if I think that we're going to reach this point where the memory bottleneck's clear, that other people will realize that too, and so I should take some risk off the table. Looking at it a different way, we in our models find it very difficult when we forecast to see how memory demand really ever gets met by memory supply in the next two to three years. It's really just quite quite difficult to see the manufacturers improving that much. Add see the algorithms get that much more efficient, So you know, I'm a bit more never bet against the market, but if I was going to talk about memory demanding of itself, I think it's feels and it's going to be pretty resilient.
00:26:29
Speaker 1: Let's talk about some other sort of market news. You mentioned that the Google's I think cloud computing demand or revenues are up eighty six percent in the last quarter. But at the same time, Google shook up its leadership in a huge way in the AI space right with Demis has a transition from CEO to chairman of deep Mind, and Jeff Dean, in another very senior twenty year veteran of Google Brain stepping down and the market sort of being a little bit spooked. Evidently what's going on.
00:27:02
Speaker 2: Well, Google has a lot of a classic challenge between making a lot of money renting its infrastructure through Google Cloud today and prices are going up, people are paying more or investing in the kind of exceptional, extraordinary research that Demis has done. And also Jeff Deine, who is really the one of the two reasons why all of us have been using Google Search for the last twenty years. And you've got to decide, right, is it jam tomorrow or is it sort of boiled rice today that someone's willing to pay handsomely for with a significant markup. And I think that they reasonably made a decision to say there is enough money to be made out of this that we will allocate the resources this way. And then what does that mean? Does that mean?
00:27:50
Speaker 1: Does that mean they've basically decided they want to invest more in the creating capacity for the operating business than and leading as research is that?
00:27:58
Speaker 2: That's my take. Look company in the world at pushing off disruptive innovation, and they had all of these different ways of doing it, from twenty percent time to acquiring things like YouTube and double click and working with Android. I mean they are They're not textbook. They wrote the new textbook. So for me, it's a really interesting signal that company with that deep culture would effectively signal through these allowing these departures that there's just a lot of money right now to be made off the infrastructure. And from having that focus now they've kept an option because with in Jeff Dean's company, they've got a commercial relationship and an equity relationship, so they have the ability to bring it back in or benefit from it in some way. But this is a sign not of diminishing commercial demand for Google's chips. It's a sign of increasing commercial demand for Google's chips.
00:28:52
Speaker 1: So if you're right and we're still very early in the cycle, what does this continuing wave of AI investment, how does it change how the world is for most people?
00:29:03
Speaker 2: Samulton, I think a few years ago, said back when GPT three was a thing, he said, well, I think it's going to feel quite boring to most people because you'll you'll see something, you'll be impressed, and then within two weeks you would be bored. And you know, by the third time it's just what you only even care, but you'll want it to get better. And I think for a lot of us that's how we will feel. I do think that we'll start to notice things that we have stopped doing. I mean, for example, I think I mentioned this, Lars I was on your show, like I'm a headphone addict and now I get chat GPT two, research my next year's purchases for me, and there's no amazon ing or Googling at all. So that's a behavior that has started to change. I also think that for those of us who have more discretion in how we spend our workdays. We will have much more flexibility because these things are actually really good at looking at messy information and helping us think through them. I'll give you a beautiful example. Oles My mum wasn't well earlier this week, and you know, she's in her late eighties, So I went over to to spend the night at her house and was chatting to her, and I got her to talk about the early years of her marriage. And as she did that, she broke out of English into Punjabi and was speaking very fast quickly between English, Punjabi and Urdu. And I recorded this with her permission, and I threw it into CHATCHIPC and I said, listen, do you speak of doing Punjabi? And of course it replied yes, of course I do. Take this turn it into a kind of coherent history, and it was able to things that I couldn't follow and understand because it had enough contextual cues from its knowledge. And I think something like that is really is really kind of a beautiful experience, certainly for me, with my mom at the age that she's at and my Punjabi being very very poor. So I think that those types of things will start to show up more and more as choices for people. But let's be really realistic. You know, at the end of the day, for all these things that this creativity can allow a lot of us spend a lot of our time just scrolling the instagramm reels feed. So let's also recognize that the fact that the technology can do all of this doesn't necessarily take you to great outcomes. To answer your question of how is it going to feel? Well, there are certain things I feel comfortable about. There are other things I think we need to to discover. I'm not sure we're all going to be running around with Nobel Laureates on our phone and being wiser than Socrates as a result. I wouldn't sit back and assume that you know, you'll sit there chat chept. You'll get better and better in your life miraculously will improve alongside.
00:31:42
Speaker 1: It asine, thank you, my pleasure, the tech stuff I must well. This episode was produced by Eliza Dennis and Tyler Hill. It was executive produced by me and Julian Nutter for Kaleidoscope and Katria Novel for iHeart Podcasts. Jack Insley mixed this episode and Kyle Murdoch wrote our theme song. A special thank you to all our listeners. Please rate, review, and reach out to us at tech Stuff Podcast at gmail dot com. We love hearing from you.
Speaker 1: Welcome to tech Stuff. I'm Os Voloscian. We're living through a pretty extraordinary moment, a technological turning point that is or inspiring for some and infuriating for many. And it seems like no opinion is complete without an assertion about what AI is doing to the economy. Is it a boom or a bust? Is there a crash in the odds. Some people think it's the future and the only economic driver that matters, and others say it's all a big bubble sending us hurting towards collapse. But how do the economics of AI actually work? And what does the enormous amount of money being poured into this technology mean for our day to day. Our guest today is Azimazar. He's an entrepreneur, author and founder of Exponential View, a research center and substack where he just released a report called the State of the AI Economy. As he welcome back to tech Stuff.
00:01:05
Speaker 2: It's great to be back, Ols, Thank you.
00:01:07
Speaker 1: One of the things that occurred to you as preparing for this conversation today was obviously thinking about situational awareness. You know what started as a manifesto and became a hedge fund and adventure fund and then had it struggled a couple of weeks ago. But you know, there's still very much a going concern. I mean, have you thought about the exponential Review Fund or do you like to sit mainly on the kind of observation analysis side of the table.
00:01:30
Speaker 2: Well, let's talk about situational awagas. I mean credit to Leopold Aschenbrener who wrote that essay a few years ago. And I was a little bit skeptical about the numbers. I think they were too fast. Yeah, he's generally been proven to be a very very good predictor of what would happen. And what happened in July was that his fund imploded. I think it lost thirty five billion dollars of value in a month. It's still alive.
00:01:52
Speaker 1: He kept his xanthropic stake.
00:01:53
Speaker 2: Yeah. Well, early investors have still done really well. They've made like twenty times their money, which speaks a lot to what he had done. I mean, he had basically run the problem of some leverage is okay, too much leverage is like over salting your food, and it's going to be resulting in a really bad outcome. I think for what we do, I mean, I do think that it's very helpful to invest when you are looking at markets, because once you have skin in the game, you are much much keener about the things that you say, So you can't be slap dash and be in the peanut gallery and just sort of shoot your mouth off, like where are the steaks? Where are the stakes? In all of this. But I think it's really important in this particular market to get in amongst the weeds in as many ways as you can, because this is not like looking at a market that we've seen before, so you really have to equally. I talk to bosses of big and large companies all over the US and Europe to hear what they are saying about about AI, because there's just no rule book for it right now.
00:02:56
Speaker 1: I think more than fifty percent who you surveys so that they believe their job will de end on successful integration. But let's talk about this report, the state of the AI economy. What's it about? What we were reporting on? What was the problem you're trying to solve?
00:03:10
Speaker 2: Well, the main problem is is it real? Is there any money in it? How much is being spent? And it's a very very difficult thing to measure, as you often find with early markets here were used to looking at mature markets. So we went out to solve that problem. How much are customers, consumers and businesses spending on AI around the world and can we get a real economic value added number out of that? Now, explain why that's important. If you're spending twenty dollars a month with open Ai, they will be spending a certain amount of that with Microsoft, who runs the servers that serve up chat GPT, so that twenty dollars may also end up with say ten game to Microsoft. A naive count will call that a count that as thirty dollars. We will count that as twenty and apportion it. Of course we have the total number. So that's a job we did. And what we want to understand was how much is really been spent by businesses at the top of the funnel, which is then distributed to all the companies that are, you know, as a little stack serving each other to get you that response to your chat GPT query.
00:04:23
Speaker 1: And how do you do that? How do you deggregate or deduplicators you put it where the money goes?
00:04:28
Speaker 2: Well, you know this is where it gets interesting. You have to use a lot of different strategies to piece the picture together. So a lot of this is happening in the private company, so they don't have to disclose, so we will go back and reconstruct their accounts from public information and other information that has been leaked to the press. We scan over one thousand AI startups. We look at when they disclose their revenue. A lot of this ends up flowing into what we call the hyperscalers. Companies like Google and Amazon and Microsoft. They make some public disclosures and then ultimately all of this runs on chips, and so we know how many chips have been sold, and so through all of that you do a process of triangulation and you say, well, these things generally all have to agree. And that's the big part of the work. And it's not one thing, it's many, many, many different things that come together, these different skill sets, different data to give us that number that we feel reasonably confident with.
00:05:33
Speaker 1: And for the you know, for the average person who's interested in these topics, but maybe isn't such an insider that they're a paying subscriber to Exponential View, Like what what did you find that matters most? What does what everyone need to know from this report.
00:05:47
Speaker 2: Well, the first thing is you can download the reports for free on Intelligence on exponential view dot dot co. You may not get some of the updates, but it's sixty five pages long, so there's more than enough to send you to sleep for several days and maybe keep you awake if you're so inclined. I think the key thing here is that there is real spending happening. We've exceeded the one hundred billion dollar per year revenue number across the industry. It's actually as of the year to July twenty twenty six, it's one hundred and twenty six billion dollars. And that that is a number we've got to about three times faster than with mobile advertising, or the cloud or the internet itself. So it really is the fastest revenue ramp in a in any technology wave, and probably of any sector in history. And that one hundred and twenty six billion dollars is still growing very quickly, roughly three times per year, like two hundred percent additional per annum, and we were really surprised about that. We had expected that number to be slowing down from last year.
00:06:59
Speaker 1: It had so that was my next question, where does this money come from? And how much is it from the AI firms themselves, how much is it from VC backed companies who may be spending a head of profits, say, how much of this is like true demand side spending versus like supply side stimulated spending.
00:07:16
Speaker 2: Yeah, it's super important to separate those out. So the AI companies will spend on compute to train their models, and we capture that spending because those are real dollars that go into the pockets of Microsoft and Google and Amazon and so on. And they will also, in the case of open ai, be spending for those that free tier of chat GPT that most people are on that they don't yes yet monetize fully through ad revenues, So we reckon that. That's around about twenty percent at the top end of the revenue number, but it's a proportion that is dropping significantly. So if we've looked at this year and a half ago, that would have been closer to sixty percent. Since twenty percent of one hundred and twenty six billion dollars call it twenty twenty five billion dollars. And then how much of that is other venture backed startups? That is, you know these startups who raise their capital from these fancy Silicon Valley firms. That number of from the venture backed startups is you know, roughly in the order of four billion a year. Now, the question is should you discount that? Because venture back startups often do very well, and I think it's perfectly reasonable. For example, if you've got a venture backed company in San Francisco, that the blue bottle coffee shop opposite their office should be allowed to count as revenue the coffees that the company buys for the team all hands on a Friday, Right, I mean, so American capitalism and the rules of accounting have worked in a particular way, and we should be careful about where we draw that line. There is of course a question about like the sustainability of any of those companies, but we know in general for Formerllion people in the US who are employed by companies were backed by venture capitalists. Right, it's not insignificant. So then you get to the kind of cluster of relationships that have emerged from say Nvideo. So in Video is a company that makes the chips that run about eighty percent of the workloads. And there's an interesting thing here, of course, so Nvideo has gone out to ensure that the people who buy their chips and who serve up the compute have got the ability to buy those chips and serve the customer demand. So that is you know, supply side support of an emerging sector. It's really not uncommon. I mean, there are very very many sectors, and we often go back to the start of the car industry where Ford in general motors extended financing to distributors and you know, literally car salesmen and also to consumers to buy cars. When you get into a situation where a core supplier in new sector has got the strongest balance sheet and the strongest credit rating and the sector is growing very very quickly.
00:10:07
Speaker 1: It's really it makes sense they become the bank.
00:10:09
Speaker 2: It makes sense they do that. Now. Now the point is that that can also go bad in the same way. But I don't think that you can say, like upfront, but this is always a bad thing to do. This is not like you know, pilfering from the till. What it does do, though, is it does introduce a new kind of class of risk in the system that doesn't exist if customers are always paying for everything that they buy.
00:10:32
Speaker 1: So how much is this approaching two hundred billion dollars, and your run rate of AI revenue comes from like blue chip corporations, how much of it comes from super users like you know, perhaps you or medium users like me? Like, well, what's there? What's the split outside of the AI industry and the venture industry itself?
00:10:51
Speaker 2: If we break that two hundred down, which is the annualization of the run rate, you know, you're looking at eighty percent of that being enterprise rather than consumer. You're up at one sixty of that one sixty maybe twenty five to thirty at the top end is AI companies spending on themselves and venture backed companies, which leaves you in arranged, let's be conservative, one hundred and twenty billion dollars from enterprises on a run rate, so that was effectively like ten billion in the last month. And then through that it's pretty it's pretty diverse. I mean anthropic and open AI who the ones that we know that we all recognize have got hundreds of thousands of commercial customers and they've got dozens, if not one hundreds, spending more than a million dollars a year. And these are companies that we might think of as being sort of advanced tech companies like you know Uber and coinbase, but it's also you know, the big banks and the like. And then within that, of course, it leans towards technology forward sectors and technology forward companies. So I'll give you one example that may surprise people. If you look at companies in America that have been bought by private equity, so not venture capital that which is all the sort of rocket fuel, and not cash producing public or family owned businesses, but once owned by private equity, their levels of investment in AI seem to be much lower than the average. And the rationale would be that if you've been bought by private equity, you need to know every dollar that takes you off the operating plan is going to give you two dollars back. And if you don't know that, the owners won't allow you to do that experimentation. And I think that speaks a little bit too that where in general bosses are about how comfortable they are about the return they get from AI.
00:12:40
Speaker 1: Right now, Yeah, you had a headline for one of your newsletters recently, the AI adopters success and failure look identical at first. So I had three questions, obviously for us, on being why, the second being how much patience does the market have for that at a time where money is becoming more expensive? And the third is when does that change?
00:13:03
Speaker 2: Yes, I mean that second question, by the way, is a brilliant question. I'm going to give you credit. This is why you are the host of this podcast.
00:13:12
Speaker 1: And takes you see.
00:13:13
Speaker 2: We're really early in where companies in particular are spending with their AI and that could be one of two things. It could be that they're going to continue to have increasing success and that number is going to rise. Or it could be that they're still experimenting heavily and everyone's experimenting and these are experimental revenues that might not continue at this growth rate once companies wise up. And those two scenarios I think are still live. I think the balance of evidence is that it's the former, which is that this number is going to continue to rise. But why do they look the same? They look the same because when you get started with a project, whether you're going to be successful or unsuccessful, you have to make the same kind of upfront investment. And if you are successful, you will spend money on that project and then you will turn the corner and you'll start to make money from it. But you won't have got to break even. And a sensible management team will say we've done this once, now let's do it three times. So just as the first one starts to pay back, you've now got three projects that are in the red. And then they'll say let's do it ten times. And so the depth of that red, which you call a JCOB, will will continue. Now, an indisciplined company that is running into all sorts of failures might not expand as quickly, but they might let unsuccessful projects run for longer and longer. And that means for a couple of years, two two and a half years, a successful firm that has really disciplined in its scaling, that runs a portfolio will look quite similar to affirm, at least from the accounting basis, to a firm that's being a bit indisciplined and isn't having success.
00:14:57
Speaker 1: And so the second question about the patients for this, and would you be concerned that the patients for these two year you know, in the red investments at a time when money is becoming more expensive me may dry up and that the revenues from the aim may start to go down as a result, which could create the cascading crash that people worry about in terms of all of the circular deals and stuff you know.
00:15:20
Speaker 2: It might do, But who would struggle with it? I think the really interesting question here is if you're a particularly if you're a public company, Wall Street only really cares about your earnings, your earnings guidance and the fact that guidance is going up and that you're meeting that guidance. If you're starting to do a deep investment program, that's going to have to come out of earnings, and investors are going to have to believe that in that story to avoid punishing your stock, which in turn makes all that capital more expensive. And as we know, interest rates are also trending upwards. Now there is a group that is going to make money out of all of this, which are the people who own infrastructure because they they have made their investments. I mean, Google has used all its free cash flow up right to buy data centers, and those data centers are predicated on the demand from corporate America, which will start to come in the next two or three years. So those companies will start to make really significant profits at that time. And I think there is a tension there that you've identified, which is what's the patient's actually going to be. There is a there's a line in the TV series Silicon Valley from one of the characters I think is called Russ Hannigan, and he's sort of saying, you've got to be pre revenue because in pre revenue, you're sold on a dream. If you have revenues, you're sold on reality. And I think we're getting to that stage where there will be a disjuncture where the hyperscalers are big tech companies will be showing, as they are eighty two percent revenue growth in Google Cloud in the last quarter, increasing revenues on these high margin products against which still be measured, but Main Street will still be on the jam tomorrow model where they're investing to show those future returns. And I think there is a little bit of a tension there that I haven't quite resolved in my head.
00:17:11
Speaker 1: Right now you're writing the report, the open question is where the cheapening artificial intelligence can create enough volume and margin to service the build out. One of the most interesting things which really caught my eye in the report was that the kind of proportion of spend in data centers is rapidly migrating from spend on concrete to spend on chips. And we had Jasmin Sun on tech Stuff a couple of weeks ago talking about this kind of deep doubt in the Ross Belt where these data centers are popping up that this AI revolution is sustainable and the concern that these may be left as kind of artifacts of a corporate boom and bus cycle. But what do you think and what's your answer to that question? That does cheapening artificial intelligence create the required volume to continue financing this incredibly expensive infrastructure buildout.
00:17:59
Speaker 2: I mean, in short, it does. Yeah of you is that mainstream companies are really early in their adoption. Where early in the adoption cycle in two ways, one is how many companies are using it and within it company, how mature are companies in their use And the shift is really is going to be palpable. I mean, these chips are currently running in data centers at extremely high utilization. To give you an idea, about eight or nine years ago, Google launched their first and second generation chips for AI they called TPUs, and since then Google's AI compute capacity has increased by at least a factor of one hundreds every one hundred times more compute, and yet they're still at full utilization for those very very first chips, which are you know, everyone knows in the computer nine year old computer is really old. And that speaks to the volume of demand that they are that they're dealing with right now. And there's something that Andy Jasse, who is a CEO over Amazon, said about six months ago about all of this, because a lot of people are saying that these guys have lost their heads, they've gone they're starting to dream things that aren't credible, aren't grounded. And what he said was, look, as Amazon, as Amazon Web Services, We've got a lot of data about how we plan for capacity and how we plan for the future, and we've been pretty good at it over the last twenty years. And I think you can see that Amazon has been pretty good at this. So maybe you should look at the investments and the commitments. So we're making from the lens of looking at people and paraphrasing him, by the way, but from the lens of a company that has been really, really good at this for two decades, and say, what's more likely that we lost our marbles or that we're still really good at this? And I think independent of the transformer model that underpaid largelanguage models that underpins Claude and chat GPT, we were already seeing a sharp move towards the digitization of business in general, what they used to call big data, what we used to call big data, what we called spreadsheets, what we called digital supply chains. I mean, these things were already happening, and that trend was continuing, and maybe AI has brought it forward a year or two. So I don't have a question about whether the economy has got use to pay for this over the coming years, and we'll use all of those chips. But I think the interesting thing that Jasmin went off and did was that she identified and evidenced through her anthropology a lot of intuitions that I think a lot of us had about resistance to the data center, which was that it's really about power politics, distribution, and disconnection. It's not necessarily a single monoline attack on a all the data. Said to me, the data set to ends up being a physical instantiation, something you can point to that says many things are not working. And here's what the concrete example in front of us that is being planned by these people over in Washington State or in California that encapsulates all of that.
00:21:41
Speaker 1: Now a quick break switching topics to one of our favorite sponsors, Vital Proteins. As a fan of science and technology, I like knowing what I'm putting into my body and why We've talked on tech stuff about AI enhants, blueberries, superhuman athletes, and health trackers. And while I might not be joining the enhanced game anytime soon, I still want to make sure I'm doing what's best for my body. That's why I love Vital Proteins. I always know exactly what I'm getting in their products, like protein and collagen, and what they're leaving out like sugar and artificial sweetness. I especially love the convenience of their new Collagen and Protein Shake. With Vital Proteins Collagen and Protein Shake, I can grab and go without too much stress, and they have both vanilla and chocolate, so I even get a bit of variety. Not only that, with thirty grams of protein including ten grams of collagen peptides and every shake, I'm confident that I'm supporting my hair, skin, nails, bones and joints for years to come. But don't take my word for it. Try the new Vital Proteins collagen and protein shake for yourself. Thank you again to our sponsor Vital Proteins. Why have chipstocks been beaten up in the last few weeks?
00:22:57
Speaker 2: H yeah, wow, they really have been. So you know, one of the things that's gone on has been that there's been a tremendous run on chip stops. They've done really really well over the last two or three years. And the Korean market is really important for the semiconductor industry because it of a company called Eskehinix, and because of lots of the supplies to the ship industry being there, and the Korean market is very consumer led and consumers are allowed to borrow very very heavily, and consumers got very exuberant. And in fact, when the slightest wabble showed up in early July and the market dropped, at one point, the Cosby was down thirty three percent, which would have been its biggest drop ever.
00:23:40
Speaker 1: The Cosby is a South Korean strug exchange or.
00:23:43
Speaker 2: Yeah sorry, the South Arian soups is what in twenty five Korean households got what's known as a margin call. So the margin call is the equivalent of Tony soprano knocking on your door saying I want my money.
00:23:54
Speaker 1: Now got a situational awareness as well.
00:23:57
Speaker 2: Right, yeah, So that that was a sign of a lot of overheating. And now today as we record, the market is down again after a really bad day yesterday. Or two things right, One is there's a sense of unease about this particular trade because when you have a moment where there are a lot of high prices and cost of memory has gone up very very quickly, high prices cure themselves because people look for alternatives and there's more innovation. And in a market that involves a commodity like memory, if high prices do get cook cured by an alternative, you're left with an enormous oversupply. And the story of the semiconductor industry, particularly the memory industry, has been patterns of their own boom and bus going back forty years. And there are enough investors who remember that to feel a little bit skeptical and want to take them at a distance and take some money off the table.
00:24:54
Speaker 1: What are they scared might happen? There might be a new round of chip innovation that makes all of these chips. The evolutionizes how memory works in AI data centers, or what's what's the fear looming over this?
00:25:05
Speaker 2: Well, I think I think it's that fear, and it's it's other things. It's do prices get so high that they somehow reduce the demand of the thing that is keeping prices high? Is it that prices are so high and that the economic incentive for software optimizations to be more efficient with memory use rises, and so that reduces people's forward forward forecasts. And this is a little bit like a game of musical chairs in a way, right. You don't want to be the last person when the music stops, because then you sit on sit on the floor. And I think that you know, investors are smart. They look further out and they say, oh, if I think that we're going to reach this point where the memory bottleneck's clear, that other people will realize that too, and so I should take some risk off the table. Looking at it a different way, we in our models find it very difficult when we forecast to see how memory demand really ever gets met by memory supply in the next two to three years. It's really just quite quite difficult to see the manufacturers improving that much. Add see the algorithms get that much more efficient, So you know, I'm a bit more never bet against the market, but if I was going to talk about memory demanding of itself, I think it's feels and it's going to be pretty resilient.
00:26:29
Speaker 1: Let's talk about some other sort of market news. You mentioned that the Google's I think cloud computing demand or revenues are up eighty six percent in the last quarter. But at the same time, Google shook up its leadership in a huge way in the AI space right with Demis has a transition from CEO to chairman of deep Mind, and Jeff Dean, in another very senior twenty year veteran of Google Brain stepping down and the market sort of being a little bit spooked. Evidently what's going on.
00:27:02
Speaker 2: Well, Google has a lot of a classic challenge between making a lot of money renting its infrastructure through Google Cloud today and prices are going up, people are paying more or investing in the kind of exceptional, extraordinary research that Demis has done. And also Jeff Deine, who is really the one of the two reasons why all of us have been using Google Search for the last twenty years. And you've got to decide, right, is it jam tomorrow or is it sort of boiled rice today that someone's willing to pay handsomely for with a significant markup. And I think that they reasonably made a decision to say there is enough money to be made out of this that we will allocate the resources this way. And then what does that mean? Does that mean?
00:27:50
Speaker 1: Does that mean they've basically decided they want to invest more in the creating capacity for the operating business than and leading as research is that?
00:27:58
Speaker 2: That's my take. Look company in the world at pushing off disruptive innovation, and they had all of these different ways of doing it, from twenty percent time to acquiring things like YouTube and double click and working with Android. I mean they are They're not textbook. They wrote the new textbook. So for me, it's a really interesting signal that company with that deep culture would effectively signal through these allowing these departures that there's just a lot of money right now to be made off the infrastructure. And from having that focus now they've kept an option because with in Jeff Dean's company, they've got a commercial relationship and an equity relationship, so they have the ability to bring it back in or benefit from it in some way. But this is a sign not of diminishing commercial demand for Google's chips. It's a sign of increasing commercial demand for Google's chips.
00:28:52
Speaker 1: So if you're right and we're still very early in the cycle, what does this continuing wave of AI investment, how does it change how the world is for most people?
00:29:03
Speaker 2: Samulton, I think a few years ago, said back when GPT three was a thing, he said, well, I think it's going to feel quite boring to most people because you'll you'll see something, you'll be impressed, and then within two weeks you would be bored. And you know, by the third time it's just what you only even care, but you'll want it to get better. And I think for a lot of us that's how we will feel. I do think that we'll start to notice things that we have stopped doing. I mean, for example, I think I mentioned this, Lars I was on your show, like I'm a headphone addict and now I get chat GPT two, research my next year's purchases for me, and there's no amazon ing or Googling at all. So that's a behavior that has started to change. I also think that for those of us who have more discretion in how we spend our workdays. We will have much more flexibility because these things are actually really good at looking at messy information and helping us think through them. I'll give you a beautiful example. Oles My mum wasn't well earlier this week, and you know, she's in her late eighties, So I went over to to spend the night at her house and was chatting to her, and I got her to talk about the early years of her marriage. And as she did that, she broke out of English into Punjabi and was speaking very fast quickly between English, Punjabi and Urdu. And I recorded this with her permission, and I threw it into CHATCHIPC and I said, listen, do you speak of doing Punjabi? And of course it replied yes, of course I do. Take this turn it into a kind of coherent history, and it was able to things that I couldn't follow and understand because it had enough contextual cues from its knowledge. And I think something like that is really is really kind of a beautiful experience, certainly for me, with my mom at the age that she's at and my Punjabi being very very poor. So I think that those types of things will start to show up more and more as choices for people. But let's be really realistic. You know, at the end of the day, for all these things that this creativity can allow a lot of us spend a lot of our time just scrolling the instagramm reels feed. So let's also recognize that the fact that the technology can do all of this doesn't necessarily take you to great outcomes. To answer your question of how is it going to feel? Well, there are certain things I feel comfortable about. There are other things I think we need to to discover. I'm not sure we're all going to be running around with Nobel Laureates on our phone and being wiser than Socrates as a result. I wouldn't sit back and assume that you know, you'll sit there chat chept. You'll get better and better in your life miraculously will improve alongside.
00:31:42
Speaker 1: It asine, thank you, my pleasure, the tech stuff I must well. This episode was produced by Eliza Dennis and Tyler Hill. It was executive produced by me and Julian Nutter for Kaleidoscope and Katria Novel for iHeart Podcasts. Jack Insley mixed this episode and Kyle Murdoch wrote our theme song. A special thank you to all our listeners. Please rate, review, and reach out to us at tech Stuff Podcast at gmail dot com. We love hearing from you.
