My OpenAI thought experiment: $500B to $1.5T?

0:00Today, I'd like to speak to you about a question that really got me thinking. So I was on WhatsApp with my buddy and we were discussing the news that OpenAI is allowing insiders, that is early investors in the companies and employees, to sell up to$10 billion of stock at a$500 billion valuation. My friend asked me this very pointed and helpful question. At this price, would you be a buyer or seller of OpenAI? Now, investors have lots of choices when they decide where they should put their money. That's based on one level risk and return. It's based on their worldview. It's based on building a diversified portfolio that rides out the bad times and wins in the good.

0:49So when my buddy asked me this question about open AI, yes, it's about whether I reckon open AI's value will go up or down in the coming years. But it's also about the other choices one might have with this sort of notional investment in the firm. Would you be better off putting your money elsewhere? Now, given OpenAI is a technology stock, that's somewhere else. Easiest be the NASDAQ index. It's the stock index that aggregates most of the big tech players. And so in my interpretation, the real question is, do I think that OpenAI is going to outperform the NASDAQ over the coming years? Now, what I should remind you is that nothing I'm saying is financial advice, not a red cent of it.

1:41So it's a fun experiment. I've been thinking about it. So let's start with some context, right? There is a benchmark here, right? Over the last decade, the Nasdaq index has returned roughly 13 % per year. It's pretty generous. Over 15 years, it's returned about 17 % a year, nearly 23 % over the recent three years as the world has got excited about AI and this new technology. So what does 17%, that 15-year track record mean? it means that roughly every four or five years, if you'd put your money in the NASDAQ, it would double. Not just adjusted for inflation, but just in sort of nominal terms, it would double.

2:23That's pretty punchy. With that punchiness comes some volatility. Prices on the NASDAQ bounce up and down. Nothing in life comes for free. You get the return of the NASDAQ. You get a pretty excitable stock index. So if you buy an index like a NASDAQ, which you can do in many different ways, you get a bit of diversification. And every single company in that index is a public company subject to the governance and reporting of public companies. And if you bought that index, it's immediately liquid. You can take your profits and losses whenever you like. You could sell when it's a bit high, so you could go off and treat yourself to whatever you fancy.

3:03Private companies can't sell their shares. You can't sell those shares as easily. They are somewhat illiquid and they don't have the same information rights or governance controls as you're forced to do on the public markets. So all of that should add what I'd consider a risk premium. You can't be buying OpenAI when you could be buying the NASDAQ because you get the benefits of that diversification and public governance. 17 % is not enough. It's got to be much more than that. It's got to be, I don't know, in the case of this, 20, 25%. So what's going on in OpenAI today is that there's a secondary sale that's allowing insiders to sell about$10 billion of shares reported at a$500 billion valuation.

3:47Now, if I recall correctly, the last time the company raised money, it was at$300 billion. So that's a big step up. And within all of that, OpenAI is quite a complex beast. It has this relationship with Microsoft that involves licensing the technology. It involves distribution on Azure, which is Microsoft's cloud platform. And OpenAI has to share right now 20 % of its gross revenues with Microsoft through to 2030. Now, OpenAI has indicated they want to push that down to about 10 % by the end of this decade. So you have this very complicated, complicated, but a little bit unique setup that has with it a bunch of risks.

4:31What if that relationship really doesn't work out? If we think about that benchmark discussion, you put your money somewhere else, to match NASDAQ over five years, that 17 % return, coming in at$500 billion would imply you'd need a trillion dollars at the end of 2030. But that's just to match a longstanding stock index. As I argued, you need to add a risk premium to all of this. So call it 25%, so an 8 % risk premium on that. At 25%, OpenAI would need to turn into a company of a scale of$1.5 trillion by 2030. Now, that's a big, big company that is somewhere between Tesla, Broadcom, and Meta platforms, all established companies in the case of Meta and Broadcom with really substantial revenues and quite low multiples.

5:27So my gut reaction to all of this was there's no way this could be the case, right? That was what I was thinking while I was on my phone walking home, did some mental maths, and I thought, hey, there's no way they can get there. But when I sat down with Excel and a pencil and paper and went back over my own model of open AI, I started to think through whether that gut reaction was right. And so here's a bull case, just presenting a little bit of maths. Today, OpenAI's revenue run rate is approaching$12 billion per annum. And it's likely to end the year at an annualized rate that'll be closer to$20 billion.

6:06It's growing incredibly quickly. In 2023, it had revenues of about$1.6 billion. There have been a number of leaks of the investor decks that OpenAI has been shopping around to the capital markets. And a recent one showed that they were forecasting$125 billion of revenue in 2029. And actually, that was an increase on a previous leak, which had had that number of$100 billion. And by 2029, they will finally be profitable. They'll be generating more than$12 billion of free cash flow. Later leaks suggested that even at the point at which they're raising revenues of$125 billion, they're still growing pretty quickly.

6:52What does that mean in terms of profits? Well, the leaks that I have seen, which have been in the mainstream media, don't talk about how OpenAI envisages its profit levels in 2029 or 2030. But roughly EBITDA, which is a key measure of that, is roughly the free cash flow, So, you know, plus and capex, minus depreciation and amortization. And if serving costs of serving these complex models fall through the use of efficiency techniques, improved algorithms, improved routing, custom silicon, we just heard today that Broadcom is going to be building specific chips for open AI, tens of billions of dollars of EBITDA could be plausible by the end of the decade.

7:40And on top of that, for the bull case, ChatGPT has become the cultural shorthand. It is the hoover of vacuum cleaners, the Kleenex of tissues. In that sense, people say, I'm using ChatGPT, or occasionally they say, I'm using Chat. That brand is really, really important. And in the last couple of weeks, there were changes in leadership at OpenAI, not the chaotic changes that we've seen over the past year with co-founders leaving, but Fiji Sumo, who's a very, very well-regarded executive, was previously the CEO of Instacart, has come in to be the CEO of applications in OpenAI, acquiring another AI company to bring the founder there on as CTO so that Sam Altman can, rumor has it, work on the big framing of OpenAI and global AI infrastructure.

8:31So there is something that is being put together around thinking about how do we build an organization that can take advantage of our technology, our brand, our distribution, and our growth rates. To get to that$1.5 trillion number, it implies big increase scale in the business. But when you turn back that$174 billion revenue number in 2030, you take off Microsoft's cut, whether it's 10 or 20 % by that time, and you turn that into multiples, it looks a little bit expensive, certainly more expensive than Meta is today, but much cheaper than, say, a Tesla is today as well. So the bit that surprised me in all of this when I did this pencil maths, and actually I looked at some other people's work, was that there is sort of this credible path.

9:20I mean, it's not impossible. There is a path towards that number. Of course, to be doing$170 billion of revenue in 2030 means that the market has to be much, much bigger than that. So how big could that AI market be? Well, today, although it's relatively hard to measure, estimates of the generative AI market in 2025 end up as between about, let's say,$40 and$45 billion across infrastructure and apps, which is pretty remarkable considering there was none just a few years ago. It means that OpenAI probably has a rather less than 30 % share. If they held that share through to 2030, that would imply$550,$580 billion in AI spend.

10:08If they lost some of that share, which would seem reasonable as competition emerges, that would imply an even bigger market approaching about$850, maybe$900 billion. Again, I'm estimating these numbers a little bit. None of this is financial advice. This is a thought experiment. Some of the major investment banks actually project that$1 trillion level of spend within five years. Now, look, that sounds like an astonishing number. It's almost like Dr. Evil in the Austin Powers film,$1 trillion, although he was three orders of magnitude lower than that. But let's see where we are today. Roughly 9 % of American firms claim to have a scale AI project up and running and delivering results.

10:58Now, that 9 % may seem low to you, but it's very quick compared to what happened with the transition from co-located servers to the cloud. To get to about 9 % of firms was perhaps seven to nine years. It was even quicker than the rate at which enterprises moved from non-smartphones to to iPhones, which was probably a two to three year, maybe three to four year process. And that 9 % is a small portion that's going to become 95%, maybe 100 % within a few years. And on top of that, we're going to go from single use cases in companies to dozens or hundreds. And if agents start to become reliable, agents churning over more and more use of these systems.

11:48And it's not just the US, of course. Every other company around the world needs to think about how they're going to implement AI. It's not, of course, just about enterprises. It's also about consumers. Today, we're at 700 million active users on ChatGPT. That number has lots of room to grow. So too does the depth and frequency with which they use it. So too does the way in which they interact with it, because we're still many of us in a question response. We're not leaving agents running. We're not leaving our AI running in the background the whole time. So in that sense, even this idea that the market couldn't possibly be that big doesn't feel to me that it couldn't be.

12:35There are lots more people to come in, start to use this technology. But of course, there's a bear case. And we've got a question come in, what's the single biggest constraint that could derail this? Is it technical scalability, regulatory pushback, or organizational complexity? These are all great bear case answers. There are all questions that we should consider. Of course, there will be competition from other companies. Google, having launched its Code Red a couple of years ago, really feels like it's got into third gear. It's shipping more frequently. It's willing to experiment with using its distribution across their different platforms to put AI in front of people.

13:17It, of course, has really deep technical depth and infrastructure, not just within DeepMind, but across the rest of Google. Meta, Mark Zuckerberg, he's like Decimus from the Gladiator, you know, on my command, unleash hell. He is chasing after this. It is absolute flooring the pedal. Just today, he was with the president of the US and he made some enormous commitments about how much Meta was going to build, how much infrastructure running into the hundreds of billions of dollars over the next few years. And, you know, Apple in all of this has been fumbling around looking for its keys. I think it's found its keys.

13:54It may even be able to get them into the ignition. But these are three companies that in different ways understand how to monetize consumers and it's consumers who make up 75 to 80 % or more of OpenAI's revenue. And on top of that, there is this strange channel conflict with Microsoft that has to play out. So there is competition from traditional technology companies. At the same time, the technology is improving really, really rapidly. And we've written about this in Exponential View about how the cost per token has dramatically fallen by many, many orders of magnitude over the past few years.

14:35and pricing may well fall dramatically further. And that starts to kind of crush your operating leverage, right? Even though as you get more efficient, you're in a competitive space and those margins start to fall. Of course, there's that Jevons effect that sees more usage, but it's not making up for the fact that perhaps you're making less money or maybe even no money serving these systems. We can't forget about the way in which the market might take shape. If you look at the database market, for example, about half of the market is proprietary databases and half of the market is open source.

15:12And the open source challenge, the open weight challenge that we normally associate with Chinese AI companies is a really important one because at some point you may get to a stage that for many workloads, you don't need the biggest, heaviest state-of-the-art model. What you need is a lightweight model that is fast, that is cheap and does what you need in that context. You'll also start to see more and more inference happening on the edge, that is on devices that we hold in our hands, in factories, in cars. These are places where OpenAI could play, but they're not places where OpenAI currently does play, right?

15:49It doesn't really, we don't think of it as a provider of edge-based models. And then I think something that certainly will happen, and of course they can lean into this, is that specialist vertical models, particularly in enterprises will persist because they can be much smaller and faster and cheaper than big models, but do exactly what that enterprise customer needs them to do, whether it is do OCR on insurance claims forms or look at security breaches across a network. These models will be highly, highly tuned. And of course, we shouldn't forget that there's enormous amount of capex that is required and will continue to be required over the coming years, which is going to potentially be a constraint, right?

16:29If the debt providers and credit providers are not willing to back this build out, which they are so far, and I think will do for another 18 months or so, that could end up being a constraint. And the next constraint will be energy, whether the American energy system in particular will be able to cope with ongoing demands. I mean, it already seems like most demand growth for the next few years will get soaked up by AI data centers. I mean, on this front, of course, OpenAI has built relationships in the Middle East and up in Norway and elsewhere so that it can start to serve customers in those geographies.

17:06A quick note, if you want to support us in bringing more of these conversations to the world, please consider subscribing to the show. But perhaps one of the big unknowns is what might happen with alternative architectures. Large language models pushing for scale are doing kind of amazing and interesting things, but they are pretty lumbering. And there are other approaches. There are alternative architectures that people are using today. For example, Liquid AI, which is an MIT spin-out, which has models that are about 10 times as efficient as transformers. You've got Jan LeCun at Meta talking about his framework, that what is needed to move beyond these autoregressive, of exponentially decaying large language models are frameworks that have a much better, stronger sense of the world.

17:59Demis Hassabis is a Nobel laureate, runs DeepMind, built some of the greatest AI technologies of the moment. What we'll say, look, we don't know whether there aren't going to be really big breakthroughs required to get to the next level of AI capability. You know, my own view is you probably would, right? You generally have always needed scientific breakthroughs, right, to improve things, even if it's a car engine or the way in which you build a bridge. But, you know, Demis, who knows much more about this than me, will likely say, without putting words in his mouth, we don't know. We don't know how far this will go.

18:33We don't know whether it will necessarily take us there. So all of those things, I think, are frictions on OpenAI. They add to the bare case. They talk about really, really deep scientific and technical risk. They talk about some market risk. They talk about general execution risk through which OpenAI needs to sail. And of course, to the question that's just come in about the recent MIT paper, which caused cold water on the bull hypothesis, it's worth saying that the evidence for actual value is, in my view, stacked toward there being value, but there is contrary evidence emerging. That MIT paper, which made it into the Financial Times and elsewhere and amongst investors, is worth exploring.

19:22I wrote about it in an issue of the newsletter a couple of weeks ago. It said essentially that enterprises said that in 90 or 95 % of the case, generative AI projects were not working. But I had some real kind of issues and problems with the methodology of that paper in terms of the way in which participants were sampled, the way in which the data was gathered. It was not a representative sample. It didn't feel like the questioning adhered to the kind of standard that you'd expect for an academic paper. And also, I felt the timeframe was a bit awry. IT projects, any kind of project in a company, takes a few years, a couple of years at least, to start to pay back.

20:07And even with the best will of the world, because we misspecify these projects, a large portion of them, 60, 70 % of them will fail. And to ask in six months whether these things are working, well, it just feels like it's way too short of a time period. It's a bit like getting angry with the pregnant cow elephant because four months into her 15-month gestation period, she hasn't given birth. I mean, it's just too short a period. So yes, I read the paper. I thought obviously it was of interest to people in the world, that I don't think it necessarily pours cold water on it. But I think this is a really important point, which is we don't really have super robust evidence about this.

20:46We don't really know exactly what this is going to mean to companies and their willingness to continue to invest in these types of AI projects. But the evidence is slowly but surely building up. I think there's another point with open AI or with any of these AI companies, and that is the disruption premium. You know, when you're a venture investor and you're investing in an early stage company, you're not really thinking about whether it will beat the Nasdaq. You're thinking about convexity. You're thinking about, does this give me a chance for a really, really exceptional outcome? A Google-style outcome, a Figma-style outcome, a Spotify-style outcome, or indeed an OpenAI-style outcome if you're one of the earlier investors.

21:32That bet is something that if you look at technology companies at this early stage is in a way the one that you might be thinking. It's really interesting historically when you look at stock market returns, and I recommend people read a paper by Besson Binder, Arizona University Finance Prof, where he looks at stock market returns over 100 years and sort of identifies that they're very much clustered around a few dozen companies. And when I looked a bit deeper, many of those few dozen companies were connected to the infrastructure technologies of the time, like the car or electricity or computing.

22:09So there is this sense that when you are in a technology transition, that value will accrue disproportionately. And I think we're to certain sectors and we're starting to see that in the fact that the Magnificent Seven is driving the bulk of America's stock market returns. So perhaps there is this disruption premium, This sort of wager that you're not really off to 17 or 20%, you're off to the chance of something that's much, much higher. I'm not even sure that any of this depends on AGI, Artificial General Intelligence. Long-term listeners and readers will know that I have lots of problems with the way that term is conceptualized and thrown about.

22:51I don't think it's a particularly useful definition or a helpful one, but let's just play with it for the time being. whatever AGI means to you. I don't think that any of this is dependent on super capable AI systems that can do thousands of hours of work, you know, without anyone supervising them, even though that's what the evaluations seem to suggest they'll be able to do within 5, 10, 15, 20 years. I actually think that disruption premium could exist with the capabilities that the technology has today. So we had a couple of questions come in. One is, do you foresee OpenAI revenue coming mainly from corporate clients or users?

23:27It's a really great question. I mean, today, most of it is coming from end users, consumers, and there have been deals done in the UAE and in India that make it widely available or will to those populations. OpenAI's API revenue, that is the enterprise revenue or a big subset of it, is actually apparently, according to a research note from Barclays, one of the banks, in a similar range to that of Anthropic, which of course is somewhat smaller. So the way I think about this is that it's very hard to serve two masters, that is the enterprise and the consumer. They move at different clock speeds.

24:03They need to do different things. And if you look at companies that are big and successful that have done that, Google would be a great example there with, you know, Google as a consumer search engine, then moving into publisher and advertising services, then into Google Enterprise Services and the cloud. Another good example would be Amazon with going from retail through to AWS. Another example would, of course, be someone like Microsoft. The computing industry was just so different when Microsoft embedded itself in the late 1980s. So I do think it's hard to serve both masters. But that said, the team in OpenAI has history to guide them.

24:42They've got a really deep and talented bench now with people like Kevin Weil and Fiji Simo and others. And perhaps that allows them to run across both markets. But there's another thing that I think is really important to bear in mind and what is, I think, underappreciated about these LLMs today, the software that we're using. I ended up in a chat with a friend of mine who every morning uses ChatGPT to build social stories for her son who's autistic that allow him to understand what he will have to navigate that day. And she described it as being a lifesaver. Really, really remarkable story. You know, she has some agency, she's got some ability to understand what he needs and can deliver that to him every day.

25:34I also separately read about how KPMG, the accounting firm, had had dozens of its tax specialists build a 100-page prompt that can do several weeks' work of tax expert work on inter-jurisdictional tax issues. So complex that you have to be a trained tax expert to use And I want you to just hold those two use cases in mind and think about the range, the incredible range that this technology is actually delivering on today. You know, for all its shortcomings, growing hallucinations, you know, inability to be deterministic, to be a bit sycophantic, you know, sometimes lumbering and slow. It's got that range.

26:24And so when we think about, will OpenAI be able to compete in the enterprise space, even though it's got this very strong consumer proposition, I think we need to look at that and say, well, maybe sort of the past view of serving two masters being too difficult doesn't hold as much. So there's another question, which is, can we really tell the difference between any of these models in a double blind? And if you can't, aren't they already commodities, in which case they'll compete on price? I think it's a great question. On the consumer side, it doesn't matter. People are locked in, right? They are using ChatGPT in the same way that they used iPhones rather than the Android device.

27:04And so, you know, I think on the consumer side of the business, the fact that they own the word for it will continue to help them for many, many years to come. On the business side, I think it's a really fair question that for many use cases, models will end up super serving that use case. if you need to do something really simple like get a summary of a meeting you were probably already at the point where the models are just generally good enough to do that in the same way that you know vision models a few years ago approached 99 % accuracy actually i've made that number up but they exceeded human accuracy let's think about what that means for models i mean the truth is i can tell the difference between a chat gpt output and a gemini output and a claude output maybe it's because of the custom prompt I've put in chat GPT.

27:56It's given a prompt to be unduly difficult and to not answer my questions directly and to leave with open questions and critical perspectives and also to push harder than it thinks I might understand. So I'm constantly confused by the answers and I have to put them into another LLM to make sense of them. But you know, my experience there is designed to push me as a cognitive partner. So I can tell the difference. I would also say that in our experience with an exponential view where we run a bunch of AI-based workflows in the background, helping us with research and analysis, we can also not necessarily tell the difference, but different models perform differently on a cost and latency basis and a does it work for this particular use case.

28:43And so, you know, I have a particular workflow which still uses Gemini 2.5 Flash, which is the quick Google Gemini model. If I put it into 2.5 Pro, it doesn't work as well. If I put it into Claude or GPT-4, it doesn't work as well. And I think that that is a, you know, an interesting observation. The models will have many different classes of capabilities. And the question is, will you be able to serve a sufficient portfolio of them? And I do think that, you know, these open source models will be something of a threat. Let's just come back to the original question, which is, you know, would you be a buyer or a seller of OpenAI at that price?