332: Money Makers Investment Trusts Podcast - Ben Rogoff (09 May 2026)

Well, thank you, and thank you for having me back on, and I hope I was sufficiently bullish when we last spoke. You know, again, I can't give investment advice, but hopefully we put up a good rear guard action in what was a pretty difficult moment for markets and for AI specifically, like you said. Before we get to the agentic moment and why we think that's important or critical, really, like you say, in a very critical waypoint, it's worth maybe just quickly recapping that the last three, three and a half years since the advent of generative AI, which now people really just shorthand to AI, as I think they must, the story to date has really been a technology that has had great promise that we felt as investors was likely to represent the next general purpose technology.

But that first three and a half years really it was a story, you know, we had to tell a story. We had to find historical parallels to I suppose give confidence to our shareholders, ourselves sometimes during those tricky moments that this was indeed going to pan out the way we hoped. And of course the snapshots at any given time in those last three years, you know, the AI story has been punctuated by bear cases and, you know, the technology starts off as being pretty inferior, quite expensive, prone to error was very much the sort of investment narrative of three years ago and we had to do our best to remind people that early applications for new general purpose technologies are often lightweight and not frankly very interesting, much in the same way that email was or I suppose airmail was in the case of aviation.

You know, and I think the debate shifted towards capex because again the technology wasn't necessarily there and revenues were also not there and so you're faced with hundreds of billions of dollars being spent building infrastructure for a new technology without apparent revenues. Again, we've been pretty, I would say, stoic and been strong in the face of that bearish narrative because we felt again that the tech was getting better and that the revenues would come.

And again, I think you can understand why investors have defaulted to sort of bubble type narratives, because I think in the end the tech sector has delivered some phenomenal things during my time at Polar and before that still, but also it has a tendency to over promise and under deliver. And so just coming through a pandemic period where the promise of a different type of hybrid world sort of fell away really, didn't it? Very fast once we had vaccines and the world returned to sort of the way it was before. So totally understand the narrative, but some very high profile investors likened it to dot com on steroids and big holes because there are no revenues and I think that was the debate we were having.

And you know, a year ago DeepSeek was all about commoditization. People are spending hundreds of billions building out the infrastructure and models that are at the frontier of AI development and here comes an open source, open weight Chinese model to just remind you how silly you all are because we're about to commoditize AI. And of course it did nothing of the sort. What DeepSeek did was remind us that we're in a race and that the runners and riders, as anyone that does the Grand National once a year will know, can change. And they have changed, haven't they? OpenAI was in the clear lead at the frontier and today Anthropic's Claude seems to be a nose ahead and Gemini from Google is a very competitive frontier model and then we have a field that's rapidly trying to give chase to those leaders, open source, Chinese, some European, obviously Meta and Microsoft in there as well.

So, it's a complicated race that's thankfully still being run, but what DeepSeek didn't do was commoditize the value of AI. And in fact what's actually happened, and I think what's underpinned our confidence, and I'll get to your question, in this first phase of AI, you know, the early promise years rather than the revenue years, has been the pace of AI model improvement. And I think we've talked about this before, humans are, myself I suppose included, you know, we're not brilliant at dealing with exponentials. There are very few things in the world that are exponentials. We tend to think in linear terms, but Moore's law, which everyone knows was a doubling of the number of transistors on a certain amount of wafer and that sort of 18 months for the same dollar amount transformed the world in terms of silicon, and here we are looking at AI models that are improving at maybe four to five times that pace.

And that's what gave us underlying confidence that the bearish narratives that naturally dominated the narrative, didn't they? The comments around AI not being great enough was ultimately going to be resolved because models were improving and we listened very keenly to what people at the frontier were saying and, you know, Sam Altman or others talking about no wall, no wall in terms of the returns to effort, the returns to capital, the returns to data and compute were still delivering, albeit maybe at a pace. But what that diminished but nonetheless continued improvement in AI models has done is get us to this point.

And this point, and I think we can take a pause if you like, but this point is this idea that AI models have now become sufficiently smart that they're able to act autonomously without humans in the loop. And not in everything, of course, but in certain critical fields, and particularly in coding, which is really the sort of test bed, the very early market where we're seeing the most agentic autonomous forms of AI. I think what that's done is really bring in a new phase for AI. The first phase, let's say late '22 to late '25, was the early promise.

And, you know, it's not an easy period, like I said, for people because bearish narratives dominate, the technology is in fear and expensive, and capex has been rising, which has been wonderful for the trust, but against the backdrop of limited revenues.

And then that all changed. And why stocks have been very strong this year, despite a war in the Middle East, and oil prices, and what have you, is that the revenues that some of those leading frontier labs, particularly Anthropic, have gone vertical. And now we have Anthropic at $30 billion of annualized revenue run rate, which, together with OpenAI, has put two companies at a similar revenue profile of SAP.

Again, is that enough to sufficiently justify $700 billion of capex? We can talk about that, but what we don't need to talk about anymore is whether or not AI is going to deliver returns, is it going to deliver the ROI, what is the revenue... Like, here are the revenues. And if you follow those two companies alone, they ought to be, by 2029, based on their forecast, a similar size to Microsoft. So, we're very clearly in a second phase. Maybe I should pause at that point, but I think that that's what's got the enthusiasm again around AI stocks is that actually the technology has now been meaningfully de-risked. We've reached a point where you can generate $50 billion of revenues, and I think that we're in that very rapid ramp now to much, much larger scale industry.

I think that what's actually happened is, and so again, there's some technical numbers that we should quickly touch on, but I think that what's happened, you know, I like to use historical parallels to help people because myself, I'm not a programmer or a tech investor. I think it is helpful to try and find parallels that maybe illuminate where we're at because this is very deep science and there are perhaps only a few thousand people on earth that are really at the very leading edge of this industry. So, the way I like to think about it is historical parallels and the one that I favored in the past has been skyscrapers as a sort of way for a number of things, but obviously an application for steel that people hadn't anticipated and how the value of land is transformed.

What's very interesting about steel and skyscrapers is what allowed humans to ultimately conquer vertical height. And that may sound ridiculous, but it obviously changes the value of land profoundly. And actually very interesting that in 1910, there were skyscrapers that had been developed 13 years earlier that were ripped down because you could now build a skyscraper that was twice as tall and the economics of doing that was compelling.

I think that we're looking at now an AI that's conquering cognitive depth in a way that I'm not sure that people have really grasped. And the best way to think about that is if you look at the time horizon that AI can complete software tasks and be right 50% of the time.

This is just as a measure, as a benchmark. If you went back to GPT 3.5, aka ChatGPT, that tool, that model was able to do something like 35 seconds of equivalent human work. You know, it was a phenomenal thing that you could ask a question and it would give you an answer immediately, but often wrong. But the duration of human task that it could achieve with 50% success within any 35 seconds. If you look at some of the latest models from Anthropic particularly, I mean we're at 14 or 15 hours.

You know, the duration of these tasks means that the work that they can be trusted with is profoundly different value. The type of things that we're using AI for internally at Polar has gone from being a quick question around where is something located or what is the forward PE or what might you know, just a simple point-to-point question is now finding five interesting chemical companies that might be pulled into the AI trade in X, Y, or Z scenario. And that's the work that a junior analyst would have taken a week or a month to do can now be done in seconds or minutes using AI and of course that has incredible value associated with it in a world where we have as a team looking after more than $20 billion of client money. So, I think that's a real example of what's happened here.

And again, this happened towards the end of 2025, the beginning of 2026, and it's gone vertical. So, the amount of time that it took for those task horizons to double was 1.7 months. We're now at 3 months. And part of that acceleration, again this is completely at odds with the idea a year ago or even beyond that we were going to reach upper limits of what AI models could do.

They were probabilistic and the benefits as I say to scale was slowing and yet it would appear to be the opposite where models are becoming better and the iterations, the time between them, are shortening. And I think there's some recursive stuff here where the models are so good that they're able to now write code without humans in the loop. They're able to not just help coders, they're actually writing complete lines of code. Something like 4% of all commits to GitHub, which is a big software repository, are now being done without humans in the loop.

So, again I don't know how deep you want to go into this, but this is not just about coding. Coding is the proving ground for AI capability because code either runs or it doesn't run. It has incredibly brutal short feedback loops in a way that writing an essay does not or doing drug discovery does not. But in the world of coding, code either runs or it doesn't. And some of the comments from people like Anthropic, where 70% of the code is now being written and committed by their own models, speaks to something very meaningful.

It's obviously created a very difficult backdrop for companies on the other side of this trade. We can come back to disruption in a minute if we may, but what this has done to software stocks has been pretty awful. But let's come back to that. The other point, if I may just share, the reason that we're so excited about the coding breakthrough isn't in and of itself around code.

You know, we had this really interesting period where we went from models that had either been trained or not trained with the answer and therefore were prone to error, to models that were able to understand a prompt, break it down into smaller component parts, and then answer the question in a much better way, reasoning models. That was a very big moment for the industry, let's say 18 months ago.

The agentic moment isn't just that I'm able to ask AI to do work without me or for me and complete work. The reason it's able to do that is because the models have reached sufficient scale or sophistication or whatever word you want to apply, that they're now able to call tools, deploy tools in a way that humans are. And if you thought about The Ascent of Man and Bronowski or something, you know, this is the moment where humans realized that with tools they're much more capable than they were without tools or using their hands.

So, now AI models are able to call tools, and one of the things that they're able to do now is call Python. So, Python is a programming language. It's not probabilistic, it's deterministic. And so, for example, it's one of my favorite parallels. I've used ChatGPT in the past to, you know, guesstimate my wife's tax return. And it's a pretty simple return, so you can use ChatGPT and the challenge with using a probabilistic large language model is that it might give you a different answer in the afternoon.

It's not in the business of math, it's in the business of synthesis and probabilistic anticipating the next word. And so, the work that models have been doing was all of that synthesis and remembering tax codes, but also having to do the math was all being done in what we call token space.

Now, models are able to do the synthesis. They've read everything from HMRC, so they are as good on the knowledge front as any human accountant, but they're not brilliant at math. So, what they can do now instead is call up Python, run the math in a sort of side pocket. This is quite technical, but I'll get to the punchline. They're able to offload the cognition part of the workflow, but still be able to deliver the synthesis that no human can match.

And so, by being able to call tools like Python, you've turned something that is inherently probabilistic into something increasingly deterministic. And now the tax return that you do in the morning and in the afternoon is the same number. And I think that's one of the parts of this agentic story that maybe others have missed. It's not just about the ability to offload work. The reason that the models can do human work is because they're able to call the same tools that we as humans use. A calculator, a web browser, a mouse, an internet connection, and because of that they're able to now do complete economically useful work in a way that just was not available a year ago. And that's why revenues at Anthropic have gone in within 4 years to $30 billion, which I don't believe any company on earth has ever done before.

I mean, there's always a healthy conspiracy theory out there if you want to look for one, and there's people out there saying it's the best marketing campaign on earth, you know. We've created something that's so good we can't sell it to you. It would make Remington look like a terrible advertising campaign in hindsight. But I'm inclined to give the company the benefit of the doubt. I think their CEO has spoken well and loudly about AI risk, and I think that we should give them absolutely the benefit of the doubt that what they're telling you is true.

And I think it speaks to, like I said, a very rapid pace of AI improvement. You know, human beings are, I mean, this is not my theory, but I'm repeating cocktail party philosophy here for you. But human beings are notoriously poor at understanding incremental intelligence. When you have ChatGPT, and then you have, I don't know, a version 4, version 4.5, like it's very hard for humans to be able to determine the improvements that are being made. And I think from a distance, it does create this really interesting sort of mismatch between what people think the capabilities of AI are and what they are actually capable of.

And I think that it's very possible that Mythos has identified a whole bunch of backdoors or AI coding holes that need to be resolved in some critical applications because, like we've said, AI can do synthesis much, much better. If you're looking at codebases, of course AI can find stuff that maybe humans have missed or that human-generated software that in the end is sort of algorithmic may have missed.

And I think it does speak to the need to sort of coordinate. And it does also speak to one of the upside risks that we've been pretty consistent about, which is that there is always a chance that AI becomes too powerful to be solely in the hands of companies, enterprises. I don't think we're there yet. And if I really believed that that was a risk that we should be worried about, I would have a very different portfolio today. I just think you're getting a sort of sneak preview about just how powerful AI has become.

And a debate that we would have had 2 years ago, 3 years ago would have been how fast are we going to get to AGI, you know, artificial general intelligence or super intelligence. These were the debates people were having. You may have noticed, no one really cares anymore. They were the debates that mattered when the technology itself wasn't economically useful. But now we've entered a world where it's already economically useful.

You can imagine worlds, and again, we try to interpolate because right now, what's so fascinating about the first 3 and a half years of investing in this AI story or this AI cycle has been that at any point in time, the investment narrative would have maybe prevented you from being invested. And of course, what the present does is it provides us with guardrails, important guardrails to understand where the debates are, but they also restrain you. They don't ever really allow you to think outside the box about what happens when this technology is five times better in a year's time or, you know, 25 times better in 2 years time.

And I think that Mythos is an indicator of scaling, and it's also an indicator of what non-human scaling looks like in this. You can call it the cognition era, call it what you like, but we're going to apply intelligence to things that we never did before because it was either impossible to or economically unviable to. And we're going to have a volume explosion. And one of the manifestations of that is that there will be more bad actors looking to exploit more digital holes or vulnerabilities, and we're going to need defenses against that.

In and of itself, I think it just speaks to a very exciting moment. Sorry, I also lost my train of thought there, which was to say you start thinking about what the opportunity sets are for this technology. Again, we're generating code at the most incredible rate, undermining a lot of the value of existing code. But I mean, how big could this market be? And you start thinking, well, the world spends $60 trillion on OPEX. It spends $40 trillion on global labor. And we're spending meaningfully less than $1 trillion on capex. So, that gives you a sense of the potential.

What happens when we start to apply AI to some of the things that we use, you know, again, labor and other OPEX items? Think about time savings. If we can arbitrage time, I remember Jeff Bezos wrote this incredible letter one year where he explained that the core proposition of Amazon was not to save people money because of lower prices, it was to save people time. You didn't need to go to shops. And actually certainly for me, that's been exactly right. That I'm much less price sensitive on items. It's much more about having them delivered the next day and not wanting to spend half of my or a quarter of my weekend buying things.

If you applied a similar time saving, I think 80% time saving is how he described it, and Amazon Prime was trying to capture 10% of the value. If you think that we have a $23 trillion wage bill in the Western world, 50% addressable by AI, you're talking about a trillion dollar market. So, what's so fascinating about Mythos, but also just agentic and non-human scaling of AI and the application of cognition in a non-human way is you start opening the door to some hundreds of billions, possibly trillions of revenues. And all of a sudden the debate around sustainability of capex and what are these companies doing seems to sort of melt away.

Well, they're both extremely exciting businesses at the forefront of probably the most important technology theme of our lifetimes. So, if they file to go public, we would go into both companies excited and with a bias to invest. Our view is that to the victor the spoils, that today 96% of AI revenues go to the frontier labs with the best models. So, DeepSeek may well have a competitive model to do X, Y, and Z, but if it can't do the agentic stuff, if it can't do the economically viable work, then it doesn't really matter. It's sort of moot.

And so, our view has been long that this is not commoditizing, that actually you need to be at the frontier with the best model, and that's why you can get to 30 billion ARR because why would I want to use an AI model that made mistakes? You wouldn't employ a person that was error-prone. Why would I do that work with a model that was cheaper but not as good? I mean, that doesn't make sense. You wouldn't build a building with inferior steel, why would you do that? So, both companies are very much at the frontier, and that's why they dominate revenues.

I would say that at the moment Anthropic's positioning, the model, looks better positioned perhaps than OpenAI. What we've seen is that agentic AI, the early application seems to be enterprises, enterprises. This is why the story is happening. OpenAI's brilliant model, I mean, ChatGPT and, you know, it's opened the market, there's 950 million users or whatever it is, consumers, you know, retail people using this product, and it's captured more than half of the market today. But now it has a bit more competition for obviously Google with Gemini, which is also a fantastic model. And at the edges Anthropic.

But the real story is that agentic AI and the productivity that you can deliver with it, you know, companies value that, which is why this is an enterprise takeoff moment. I should have said this earlier. So, it also looks like the revenue trajectory of Anthropic is steeper at the moment than OpenAI because I say their model might be slightly better right now. And the cash burn profile of Anthropic also looks better near term at least than OpenAI.

But let's not write OpenAI off. They built this industry and I think we're waiting for their next what we call pre-trained model. And when they deliver that, there's absolutely no reason to believe that they wouldn't be back in at the front of the race. And I think that some of the bearish chatter around the business will just melt away. But it's been a while since they delivered a new what we call pre-trained model. They've delivered some phenomenal models with post-training and they're certainly competitive.

We use at Polar both the enterprise ChatGPT product and we use Claude from Anthropic and have an eye for it as well. So, I think we're positively inclined towards both. I think they both are very unusual assets. What we've seen over this year particularly is fear spreading in the broader market that the value of these large language models is paramount. That these are not just things that you ask questions of, but ultimately they're sort of orchestration engines. They're the brain in a new technology stack.

And in that world, you know, what does it mean to be a software company? What's the value of standalone software in the world where you have a sort of superhuman brain orchestrating everything else that you're doing as an organization? And I think we still feel very strongly that the biggest risk to just about every other part of the IT stack is that the value goes to the large language model, to the frontier, and everything else is commoditized. And if that's true, then you would expect us to be investors in both companies at IPO.

In terms of timing, I think that expectations are that both companies come public this year. But we're watching and hope.

That's a great question. I think that history has demonstrated that the null hypothesis as an investor should be a winner takes most. Certainly that's been true at all the large profit pools that we see today, whether it be Netflix in video streaming or more importantly people at Apple in smartphone or Google in search, Amazon in retail. I think that that should be one's null hypothesis. But let's see.

At the moment actually it's quite interesting that I referenced Anthropic being in front in the race, but they seem to have, I think, the very best agentic model right now in Claude and maybe in coding in Claude Code, but Gemini from Google seems to be probably as good in multimodality, being able to pull together images and text and also audio and make sense of that. I think Google has a fantastic position there and I think OpenAI still has probably the best in class model for doing a bit of everything.

So how this ultimately works out, sorry to use a Tolkien type framework, but do you have one ring that rules them all? I think probably yes, which is why many companies are spending more than a hundred billion dollars building out their AI capabilities. I think it's really fascinating because when you look at what's happening in my world, there's a vertical integration that seems to be happening in the way that we haven't really seen for a very long time.

And we may have talked about this before, but it's one of the reasons why in my humble opinion we were pretty quick to sell software stocks a while ago. You know, most investors remember only a horizontal compute world where Intel makes the chips, Microsoft makes the operating system and a whole bunch of companies sit on the top of that and create sort of specific vertical software or horizontal software, but nonetheless not vertical in the sense that IBM was doing the mainframe world or Digital Equipment was doing mini computers.

And actually we've sort of felt that this cycle was much more likely to be a more vertical compute where you buy your chip, your GPU from Nvidia, you might buy your networking from Nvidia, various parts of the rack. The core unit of compute is the rack rather than something else and it felt much more mainframe like. And again, I'm not that old that I remember investing in mainframes, but I've certainly looked at the history of the mainframe cycle and I do think that there are more characteristics today that look like that than they have done in any sort of post-Wintel PC smartphone internet world.

So, with that in mind, I think everybody understands that you can't really afford to lose this race because just thinking about some of the comments that Reed Hastings, one of the co-founders of Netflix, made recently, like the challenge to Netflix on a I suppose 5 to 10 year view is that they're going to have to compete in a world of increasingly AI generated content with companies that own their own large language model, like for example YouTube by Google.

So, doesn't that potentially, and again, this is my words, not his, but doesn't that potentially put Netflix at risk? Companies that don't own their own large language model at the frontier are potentially passengers or spectators to what happens next. And the best thing that could happen for those companies is that ultimately AI commoditizes. You don't have to own your own steel and steel is a commodity.

But in a world where steel is not a commodity and where access to that steel means that you either do or don't get the job to build the skyscraper, actually can result in some very profound changes to profit pools. We weren't the first to realize this. I think it does explain why Microsoft, why Amazon, companies that don't have their own model are rapidly making deals with companies that do, like OpenAI or Anthropic, but also why chip makers like Nvidia are making investments in frontier labs like Anthropic and OpenAI. And why those companies are doing deals to try and get access to chips.

The industry is verticalizing, but no one really knows who the winner is. It creates this sort of web of alliances that if you're a bear, you would say it's a house of cards, but if you're a bull like we are, you would say this is a perfectly rational way for an industry to verticalize when no one, including the very smartest people on earth, know who's going to win the race.

So, the cost of losing the race, and again, this is through the AI lens that we think about all the time. And the AI lens as a reminder is something that we added to the investment sort of process May 2023, where really the lens is there as a metaphor to make sure that anything and everything we do is considered through a world where AI is a sort of dominant general purpose technology around which everything is reimagined, which was a bit nuts 3 years ago, but doesn't seem quite so far-fetched now.

And that's what again enabled us to be in the capex beneficiary stocks. We can talk about some of those if you're interested. And also steer clear of companies that are built on code. They're software companies, digital platforms, you know, the winners of the smartphone cloud world that dominate profit pools. Most of those companies...

don't have a large language model. So, in a sense, their fortunes depend very much on stuff that's happening outside of their control. And I think that that creates terminal value risk that most analysis doesn't really capture. I don't know, a booking.com. How does that stack up in the world where consumers are not the incremental buyer of a hotel room, but agents are?

Maybe lucky on the tariff front. Genuinely, I'm so happy that we made the call. We had talked about it for sort of two, two and a half years that large language models had the potential to start sucking some of the value out of the software world. What's happened of course is that things have moved much faster than even bulls like ourselves have imagined. I don't think anybody was imagining that Anthropic would be at 30 billion ARR because it was at 9 billion 5 minutes ago. So, that sort of inflection point has happened and you obviously have seen certain things happen with Anthropic and plugins and the progress that these models are making. So, we've had that right, I guess, and I'm really happy that we did, but I don't think it can stop there, can it?

I think that's still very true. Look, on the Google front, I mean, I think what's key is that a few things. The first thing is I think Mag 7, we own all of the Mag 7, you know, wonderful non-fungible assets that dominate profit pools and each has a bull case around AI that's still alive and well. The unfortunate thing for all of them is that there's also now a bear case. A more capable AI poses risks as well as opportunities to each and every one of those companies. It manifests in free cash flow because you know, look at the capex burden now that companies like Amazon and Google and obviously Microsoft are having to bear to remain in the AI race. Now, again, look at the phenomenal growth that each of those companies have had in their various AI businesses and you know, again, I'd much rather be in stocks that understand the opportunity and the risk posed by AI than companies that don't. But they are becoming much less asset-light businesses and much asset-heavier businesses and I think that's sort of taken its toll on their collective attractiveness as conduits for AI.

So we own them all. Right now there's what feels like a shortage of compute. So we've been rebuilding companies like Amazon where actually we've been reducing previously because they don't have a model of their own. And in a world where compute is increasingly driven by machines, where machines make the decision about where a workload, a unit of compute will be stood up, I think that that's quite threatening to Amazon when it has to compete with a company like Google that does have a model. So we had reduced Amazon meaningfully. Now we're in a world where it feels like there's not enough compute and all of a sudden actually Amazon's cloud business could really be a beneficiary of this sort of short squeeze that's occurring. They also have chips, you know, they have their own training chips and I think Anthropic's model has been also built around training them, so that's given the business a bit of a boost. Anyway, let's not get too bogged down there. I guess what I'm saying is that overall we still feel like it's a less good conduit than it was when being in the AI race was cheaper and therefore only really represented upside optionality for those businesses.

But around Google specifically, that's probably the most interesting one because obviously most of the businesses built around advertising and that's advertising for humans. So should behaviors change, should machines become the incremental consumer of everything, then you would think that advertising-based businesses were at significant risk. You know, where a product sits on a website is really not something of interest to an AI agent that's acting on our behalf. And that's why we are very, I don't want to say bearish because it's like we've just chosen to sidestep businesses that aggregate other businesses, you know, sitting as a digital layer between a physical asset owner like a hotel company and bringing that supply to an aggregated consumer has been a phenomenal thing in a world where humans have been faced with frankly unlimited choice and where time is valuable. I mean, let's be honest, time is valuable. Humans operate on an efficient frontier where the unit that matters is time. So, an aggregator of information services is extremely valuable to a human because we value time in a way that by the way a machine does not. And this is why I'm very anxious about businesses that aggregate and help humans, attention scarce humans, make decisions.