← Azeem Azhar's Exponential View
Azeem Azhar's Exponential View
Inside the economics of OpenAI (exclusive research)
00:49:46
In this episode, I'm joined by Jaime Sevilla, founder of Epoch AI; Hannah Petrovic from my team at Exponential View; and financial journalist Matt Robinson from AI Street. Together we investigate a fundamental question: do the economics of AI companies actually work? We analysed OpenAI's financials from public data to examine whether their revenues can sustain the staggering R&D costs of frontier models. The findings reveal a picture far more precarious than many assume; we also explore where the real infrastructure bottlenecks lie, why compute demand will dwarf energy constraints, and what the rise of long-running agentic workloads means for the entire industry.
Loading summary
Transcript71 lines
- [00:00]
A
Today, artificial intelligence companies are now being valued in the hundreds of billions of dollars. It's OpenAI, it's anthropic. It's all the value that DeepMind has added to Google over the past years. But that forces a really important question, and it's a question that is being asked by the mainstream, but also by specialists. Do the economics actually work? When you look at what it costs to train and run a Frontier model and what you earn from it before the next model comes along and replaces it, is that a profitable business? Are we looking at something a bit like Uber, which lost money for 14 years before turning a profit and is now handsomely valued? Or something that doesn't have an end in sight? Now these questions really matter. The stock markets. Well, Big Tech had a lurching week this week, and at one point, more than a trillion dollars was wiped off valuations. Wall Street's very linear investors trying to digest the $650 billion of capital expenditure commitments being made by big tech for 2026. Some of that $650 billion is going towards AI infrastructure. Does any of this make sense? Are there actually going to be operating margins to defend? And is the revenue growth going to support this? Now, as readers of Exponential View, you'll know that we've been asking these questions for months, if not longer, but most recently we partnered with EPOP AI. I'm sure everybody knows epop, but if you don't know they are really preeminent independent research organization tracking some of the trends behind AI. You've probably seen their work on scaling laws and compute trends. So we worked with their team to dig into the actual margins of Frontier AI, and the results are really, really interesting. So whether or not you've had a chance to read our research yet, and you really should have done, this conversation will give you a really clear picture of where things stand and where they're heading. I've asked financial journalist Matt Robinson from AI street, it's another substack newsletter, to moderate this discussion and really put us on the spot. Jaime Sevilla, the founder of Epoch AI, is here. And Hannah Petrovic from my team, she led the research. On the Exponential View side, she's also no stranger to large numbers. She has a doctorate in astrophysics. So when I say, well, is that roughly right for Hannah?
- [02:20]
B
It's.
- [02:21]
A
Is it within a few orders of magnitude often, yes. Matt, I'm handing the stage over to you. The floor is yours.
- [02:27]
B
Maybe you guys could start with, you know, for someone who's getting into the research, what's the big takeaway here. And how did you even think about building a framework to analyze a business like this?
- [02:39]
C
Absolutely. So Matt, a little bit here of the context of why we were doing this before. I takeaway to our understanding, no one had really taken on this humongous task of piecing together all the public information that there is about the finances of OpenAI or any large AI company really, and trying to paint a picture of whether what are their margins like, whether they are making enough money to recoup the large cost of developing new products. So we did this hermeneutic exercise of just hunting for all the information that we could find and trying to make sense of it. Now, I won't pretend that we have arrived at a definitive answer. In fact, our views are constantly evolving as we learn more about the companies and their finances. But I'm pretty happy with the overall framework that we have established for even trying to think about this question in the first place. If I was trying to now communicate like, okay, in summary, what did we learn and what did we find? For me, the two most important takeaways is that one, it seems likely that OpenAI during the last year, and especially while operating GPT5 was making more money than the cost of the compute, which is the primary expense of operating their product. Though they seem to have made like a very small margin or even having lost money after accounting for all the other operating expenses that are going to run in the model. So this is paying for staff, this is sales and marketing spending, this is administrative cost and this also includes the revenue sharing agreement that they have with Microsoft. Now the raw profitability, the operating margins of a company are not necessarily what you want to look at when you are trying to assess whether the company will be profitable in the long term. As Asim alluded to earlier, Uber lost billions upon billions of dollars before they finally became profitable. And really if you are, if you're an investment minded person, when you are looking at a growing business, you do not look so much at how profit, how much profit they are turning in their early years or they're still growing, but rather you, you will rather, you will rather look at the gross profit that they're making at their gross margins and how the revenue is scaling year after year. So you can get a sense of, after this initial phase of a rapid growth where the industry and the company could land at. Now if you did that then I have just said like, okay, they look, they look to have made like a decent gross margin. But there is one more wrinkle in that you need to account for here, Matt, which is that these products are really, really expensive to develop and they have a very short shelf life. So it's not enough to just look at the gross profit and check like, okay, it seems that they have this 50% gross profit. Seems they're getting twice as much money as you put into to the machine. Like, no, you actually need to think about how much money does it take to develop a new model and how long could you expect that model to stay relevant before it becomes obsolete by your competitors or by open weight alternatives that will make a dent to the usage of your, of your model. So this is the second part of our research where like we try to look at how much OpenAI is spending in R and D and how that compares to their gross profits overall. And what we found is quite shocking. So if you look at how much they spent in R and D in the four months before they released GPT5, that quantity was likely larger than what they made in gross profits during the whole tenure of GPT5 and GPT5.2, which points to how competitive this space has become in the last couple of years.
- [06:33]
A
Hannah, why don't you just. I know that you dug into this in such detail and I don't want to speak ahead of, of your expertise to Matt's question.
- [06:41]
D
The methodology, how we actually got into it. And a lot of it was based on numbers that we could find in the past historically and then trying to predict what would happen in the rest of 2025. So for example, the sales and marketing, we had some data that 2024 was 1 billion in sales and marketing. And then in H1 of 2025 that was 2 billion, for example. So we can build a picture using constraints in this way and from that you can try and understand the costs of the company as a whole. And we broke it down into many categories, as you can tell in the piece, but each of those as well, I tried to break down further into the different separate components so that we could realistically understand whether or not it was feasible or a realistic approximation at least.
- [07:33]
A
This is a kind of complicated exercise. And one of the things that comes out from this is this question of that short model life. And the family that we looked at was only really the preeminent family for a few months. Now, we know that enterprises and even enterprises don't change the API they're using the day a new one comes out. There's always a bit of a lag, but consumers do, right? Because that's what you get access to, on, on ChatGPT. And you know, you remember, you may remember that when GPT4 was set aside from ChatGPT, it was an emotional support tool for many users and they were very upset with how methodical and mechanical GPT5 now felt. And I think one of the uncertainties is to what extent do you actually learn and prepare for your next model based on the short life of the existing model. Right. There are a couple of elements to it.
- [08:30]
C
Right.
- [08:31]
A
One I think is a little bit more nebulous, which is that by having a really good model, even if it lasts for a short period of time, you maintain your forward momentum in the market in terms of customers liking you and your enterprise sales and so on. And that feels less tangible than the second bit that I think is perhaps a bit harder to unpick, which is what do you learn about running better and better models from actually having run a better model, even if it only lasts for four months? And that learning might be sort of down in the weeds in sort of R and D and, you know, particular choices you make in training data and reinforcement learning. It might also be in operations. Right. And just operating a model of that scale. And I think it's quite hard for us to know. I suspect it's hard for OpenAI or any of the other foundation models to know the contribution of that second part to the model itself. Right. So in a sense, who in this kingdom is actually able to see with two eyes? I'm not sure, you know, many can at this point.
- [09:31]
B
It's interesting. It was making me think of GPUs and how do you. I was talking to some finance folks about, okay, well what is the value of these H100 chips going to be in a few years? And everyone's kind of shrugging their shoulders like, you know, and sort of putting it out there. And it's kind of seemingly like a parallel to these models, like what is the value of G2P4 three years ago was. So how do you think about that? One question I have is you talked a bit about compute and costs there, and this may be a little down in the weeds, but the cost of compute in sort of building these models is going down and how do you sort of see that kind of going forward?
- [10:12]
C
So the cost of compute of building these models, I don't think it's quite going down. I do see it as going up time and again. The pre training seems to be going up. And despite rumors to the contrary, pre training is not dead at all. People are Building hundred billion data centers for a reason. They are invested in running very large scale experiments and very large scale training runs that are unprecedented inside. And I think that this is part of what contributes to these models being so expensive. Like one of the interesting things here when I think about OpenAI, is the game that they're playing. The game that they're playing is not so much about becoming profitable right away. Rather what they are trying to do is convince investors that they have a business. They have a business and a research product that's worth scaling as much as possible like, driven by this conviction that through a scale they're going to unlock new capabilities that in turn is going to unlock new markets and let them continue their incredible, incredible revenue growth.
- [11:11]
A
I would say I want to come to, I think that's exactly the right thing for them to do anyway. You know, as an investor myself, I want to invest in people who have optimistic views of the future and therefore believe you need to plant seeds today in order to harvest them in, in two or three or four years. And you know, particularly in a business like this where there is no asset, you know, there's no, there's no hotel that's been built that can be resold to, to another property developer, you know, it's an intangible asset that may not have that much salvage value, especially if people in the team leave. I think the exactly the right thing to do is to, to be building out ahead of time. And you see that investment J curve. I think that the, the two kind of challenges around this model are. Number one, is the OpenAI model the only way to do this? And I don't just mean from the financial side, I also mean from the sort of strategic focus side. We've seen anthropic do something completely different. And the second challenge is, and I think we went part of the way to answering this question is is there a path to positive unit economics? In other words, are they producing something for X dollars that they can sell for 1.3x dollars, or are they producing something for X dollars that they sell for half an X dollar? Which was the story of a lot of the.com, right, cosmo.com and all these other things. And I think we got partway to answering that second question, which is that yes, it's expensive. Yes, there is, you know, some kind of gross profit margin. The level that we estimate, I think Hannah can speak more accurately to this, is, you know, lower than a traditional software business. So we're learning that perhaps Foundation Labs don't look like software businesses, they look like something different. But you know, these are the things I think that we have to, we have to play around with.
- [13:03]
D
Yeah, spot on with the numbers there. The other thing I would like to also consider is that AI is also creating a flywheel in the development space of itself. So I'm wondering how that might affect R and D down the line given that R and D is such a huge development cost to the company for the next model. That's just a thought there.
- [13:24]
B
I'm curious, you know, the OpenAI's got a little bit of slack for saying that they may introduce ads, which is I thought kind of peculiar that they did because we all have been, I mean I've been using Gmail for 20 years and actually in preparing for this I, I, I stumbled upon some research from Google, Larry, Larry Brennan about how they were sort of against ads in the beginning and they, you know, changed their minds. And I'm just curious how you think about ads and how that, you know, if you have what they have, 800 million eyeballs every week or you know, how that might play into this.
- [13:57]
C
Let's think about why OpenAI is trying to introduce ads in the first place. Because with these ads again that you said yourself, it seems that they have on the order of almost a billion users that they could monetize through ads with their monetization plans, it seems that they might be able to reap a revenue of like maybe a couple billion, maybe even up to a few dozen billions of dollars off of that audience. That's not enough. If your plan is to build hundred billion dollar data centers, that's not going to be enough to fund that. So why are they considering ads in the first place? I think this has to do a lot with this game that we're alluding to where they are not looking to become profitable right away, but they have this vested interest in demonstrating to investors that if we want it, we don't want right now, but if we want it, we could be, we have a path to profitability and the ads kind of fit into that, into that project plan is part of the way of expanding their market. That's going to allow them to show like look, there's a path to a hundred dollar billion revenue between the ads, between business sales, between other markets that we could be unlocking here. Like these models are not profitable right now. But they have these arguments that they can make, including ads that point to that, that help them argue like look, not right now, but they could be. We are not Going to, we're not going to do that because we're more ambitious than that. But we could be profitable if we wanted.
- [15:21]
A
We've seen some success with ads and Gen AI. I think it was in Meta, wasn't it, that they have really had some sort of forward momentum?
- [15:31]
D
Yeah. So in Meta's earnings in October, they commented that the AI tool driven ads, which were essentially bringing in 60 billion of revenue in ARR. So there is a considerable uplift already in the ad space there. But it's. This will be the first time that it will be in the chatbot.
- [15:53]
B
And that was helping with conversions, right? With Meta's ads.
- [15:55]
D
Yeah. And it was also that Meta were able to keep people on like Instagram on YouTube. Not YouTube, Instagram and Facebook longer. So people were seeing more of these ads as well.
- [16:06]
C
Right.
- [16:06]
B
I guess everyone can just make up their own ad and you know, it's a lot easier to do that and.
- [16:12]
A
You'Re, you're sort of, you're stuck in there. But I think that this, this question about the ads is, is a really important one. I think, Jaime, what you've suggested is, is really intriguing. So it is, it is the question as to whether an advertising model is really, really fundamental to OpenAI or whether it is, it's a sort of instrumentally useful thing that gets you to, to the next stage. And you know, I think a lot of that depends on how we start to use these tools. I mean, you know, the piece of work that Epoch and Exponential View did looked at ancient history, with all due respect. Right. It was before Last week.
- [16:51]
C
Right.
- [16:51]
A
It was before Openclaw, it was before Opus 4.6 and whatever else Anthropic comes out with. And we looked at a particular world before what I think Andre Carpathi called the threshold of coherence for agents. And he described this moment where the agents are now good enough that you can get them to do lots and lots of things for you. And so that also makes me wonder whether that traditional ad model makes any sense, because there ain't going to be any eyeballs to sell it to the agents, sell it to the agents who have probably rented humans to do the jobs they can't do themselves.
- [17:28]
B
You guys have spent, you know, you spend a lot of time doing some rigorous work here. Say you swap seats with Sam Altman. What do you do differently or what do you keep the same?
- [17:37]
C
I mean, first of all, I will just go and look at their finances and actually get a clear picture of what's going on after having done that. What I will do in Sam Almonds, in Sam Altman's position, like, honestly, it seems kind of similar to what they seem to be doing already. For me, this question of the models as a rapidly depreciating asset actually brings a little bit into focus of what might be the Perdurian asset, the part that might retain more value through generations of AI. And it seems to me that this part is infrastructure, and they're gearing up big time to break into the infrastructure space. They have famously said, like, oh, we want to get to a position where we are building gigawatts of power at the time, which is really ambitious, a very ambitious goal. But it makes sense from my perspective. If you think that the software part is rapidly depreciating, you might want to get in on this part of building and serving infrastructure at a scale.
- [18:38]
D
Yeah. So if I was to bring in a different view here, obviously the consumer section proportion is quite large, and we know from Sarah fryer that about 60% is now consumer, 40% enterprise. So obviously the enterprise push is there, and we know that the enterprise push would be, you know, bringing in money for the company quite well. The consumer side is very competitive given Gemini and, you know, other AIs, which you can easily have on your device. Say, like Samsung, I just hold my finger over a button and I have access to Gemini, so there's very little friction in using it. But so if I was Sam Altman, I would want to try to see if I could do something different on the consumer side. And obviously we know, like, they're hoping to bring out a device that is a unique, different side of targeting that consumer component. And if there's other things they can do there that would keep their consumer money coming in.
- [19:38]
A
Okay, we've had two different views here. Matt, I'm going to give you a third view just to really make you work hard for your moderator's seat. Let's take this idea that Hannah raised, which is like different classes of interactions for the end user, whether it's consumer or business. And the point that Jaime made, which was, look, the infrastruct really matters and the infrastructure obviously matters an enormous amount because in the last week or two, if you've been using anthropic, it got really slow because we all got excited about Opus 4.5. There is a mo. And then the question is, well, where does the revenue come from? Where is the point at which people start to spend more and more? And one thing I would say from just looking at the exponential view bills, our Bills have gone up since Opus 4.5 came out. Okay, because everyone is coding more, we're running many more background processes that are chewing through tokens. And I thought that was all true until I installed my OpenClaw bot. Actually it was called, what was it called? Claude initially. And I've called mine Mini Arnold in homage to the second Terminator that came back to protect us. And, but Mini Arnold is a greedy and forgive my French mofo, he, he will chew through 20 to $30 of tokens a day. So we're talking five grand a year in order to do my bits and pieces. And I've pushed him down to Haiku, which is the cheapest anthropic model. I, I do the Heartbeat on a local LLM so that, you know, every 30 minutes, I'm not having to pay pay for that. It, it is expensive. Now what drove that? What drove that was that idea that the models were just, just good enough. Like they, they crossed that uncanny valley. And when we did the work on OpenAI, their models hadn't crossed that uncanny valley.
- [21:25]
C
Right.
- [21:25]
A
GPT5 was not the thing you could leave to run for hours at a time. 4.5 opus from Claude from Anthropic was, was really the first. And I'm just curious. My sense would be that all of this discussion starts to look very different when OpenAI is shipping things that run 5, 6, 9 hours at a time. Because at that point actually, you know, the inertia of being an OpenAI user through the enterprise or through the customer sticks with you. And the only thing I would say is just think about dear old Mini Arnold who cost me the same as, you know, four Starbucks flat whites a day to do whatever the hell he's doing in his Mac Mini. I don't know what he's doing. I don't ask. It's his private space. You know, get on with it, Mini Arnold. So that is, that for me is like, is how we, we merge Hannah's observation, like kind of the user experience, user interactions patterns with Jaime's point, which is this is all about infrastructure because ultimately all that processing has to happen somewhere and that becomes a choke point.
- [22:20]
B
And to that, you know, this week, as you mentioned earlier, that you know, the markets were caught flat footed, I guess, to say the least about this, you know, ever expanding compute spend. To me what's interesting is that, you know, as the hyperscalers just reported, there are, they're capacity constrained, right. You know, we're seeing these big rollouts where, you know, CapEx is huge, but yet they can't meet demand.
- [22:42]
C
Right.
- [22:42]
B
And I think that, you know, maybe it's a little separate, sort of beyond OpenAI. But I'm just curious, like how do you see just, it's just wild that they're spending this much money and they just can't catch up.
- [22:53]
C
It's just a lot here, Matt. And I do see that they do put a primary constraints here. If you want to scale up the infrastructure is you need enough GPUs and you need enough energy. It seems to me that energy right now is the thing that everyone talks about. It's something that we know how to solve, we know how to build energy. Like you don't need that much energy. All things considered, if we need to build 10 gigawatts, 100 gigawatts of extra power, that's only a 10% increase over all installed capacity in the U.S. this has happened in the past. In the 2000s, they built enough gas infrastructure to match that level of expansion. The GPU part though, that's something very unique. That's something that right now is being chocked hauled on production in a few factories in Taiwan. And they have been trying really hard to expand on it with pretty limited success. So it feels to me that that's probably where the, where, where the, the bottleneck to scaling is going to end up being in the long term.
- [23:50]
A
So I love what you've just said, Jaime, because of course, you know, the, the, the general note out there is it's all about the energy. Like energy is the, is the bottleneck. And I think it's, it's pretty clear that the energy is, is constrained because of lead times and, and so on. But if you listen to Elon Musk talk about why he wants but data centers in space, every single thing, reason comes down to things we've done to ourselves, grid permitting, backlogs. I mean these things are cues. They're not walls, they're not laws of physics. You know, the laws of physics are, you know, black body radiation and the speed of light and all these other things that Hannah knows much more about than I do. And so I think that there is something to be said that it's very exciting if you're in the energy space to suddenly be important because we'd rather forgotten about you for the last ten years or so. And Europe hadn't really thought much about its energy and was caught flat footed by the gentleman from Moscow. And I think everyone has got really, really excited about that question. And as Jaime says, it feels like it's solvable. I would push a little bit on that because there's just a lot of supply chain questions that have to get fixed. The copper issue. Right. Suddenly we all know about copper and, and about, you know, optical fiber and whoever thoughts about corning honestly before two months ago. So there is something there. But, but I think that what I also take away from this is let's go back and find out when these companies started talking about megawatts and gigawatts because I'm pretty certain they were not saying it at the end of 24. And I, and I have to go back on my notes. I think the first time I started hearing them talk publicly about gigawatts was, you know, I met with, with Satya in January of 2025 and, and I would say that it was after that that I started seeing Microsoft talking quite a lot about, you know, megawatts and gigawatts and or as Doc Brown from Back to the Future would say, gigawatts. And so this is kind of new to them and it's also new to the, the energy business. But there's definitely, I think this is the thing that the markets didn't get at the start of this week, which is these hyperscalers are absolutely supply constrained. They don't have enough chips and they can't energize the chips they have. And you heard Amy Hood, who is the Microsoft CFO say I had to make a choice of whether I put processing power to third party services on Azure or Power. Microsoft Office are first party apps. And I had that trade off. I mean this is not a market which is not being, you know, hasn't got people running after it trying to spend money. And you know, I think that's Hannah, you know, you've, you sent me your kind of latest analysis of the market dynamics. I mean that's exactly what, what we're seeing.
- [26:39]
D
If I was to add anything, I guess I'm wondering how things will move also to the edge. So obviously you have a lot of this build out for the hyperscalers in data centers. But we commented on Azim, you said you were running an LLM on device. what point will they get better to the state that you can actually run, you know, the things you're doing now on your device with hardware improvements that are coming and the algorithmic improvements which are also coming. And I wonder at what point we can do most of the things we're doing now on our device.
- [27:12]
C
Yeah, well it's actually very interesting if you Lend me to build on top of this. Because if you look at a fixed level of capabilities, you see this rapid growth where in order to achieve what models could do nine months ago, you already have pretty much an open model that, I mean, it's going to be kind of there, right? If you look at the Kimi 2.5 model, like it's arguably at the O3 level, the O3 being the model that OpenAI launched in April last year. If you look at that, then you see this rapid decrease, this rapid decrease in the amount of resources that you need in order to train and to do deploy a model at a fixed level of capabilities. But it's not a fixed level of capabilities. What's driving the growth of the industry and the growth in revenue. This ever increasing mood of capabilities that we have. Like, we are building more and more, we're expending, we're expending more and more in building more sophisticated machines. And I think this is going to cut against the gradient of moving things to the edge. Like all of these new exciting capabilities. Like, you're just going to want to run it on your data centers. You just want to run it in your data centers over a very long amount of time. Like, ASIM is talking about, like, oh, I may be looking at a bill of $5,000 a year for running my agent. And it's like, that seems very small compared to what these kind of machines could do in the future. Like, if you get to the point where these machines have an output which is comparable to a worker, like, how much will you pay to have a virtual coworker who is like, really knowledgeable, who is like available 24, 7. Like you might be willing to spend on the order of hundreds of thousands of dollars a year just to keep those kind of agents running. And there is huge advantages to running them in a, in a data center. Like, the biggest disadvantage here, the one that, the one that physics won't allow you to overcome, is latency. But right now the models are already slow enough that I don't mind waiting like 200 milliseconds for the responses to get to me. Like, it's fine. They can take, they can take. I leave my GPT5 Pro thinking in the background for, for 20 minutes and I get back to the answer. I'm not in a hurry.
- [29:14]
B
I think that the compute demand has been sort of underreported. I mean, we all know about CapEx, but I've talked to folks in the enterprise that even when they're not using it, they Don't. They don't want to give it up. They don't want to lose their spot. So there's just that constant demand for access.
- [29:29]
D
Seems like the same way that planes were still flying in Covid so that they could keep their flight routes even though no one was flying.
- [29:38]
B
Yeah, they didn't, they didn't want to lose it. I just, yeah, that, that part of the story is sort of interesting to me. And you know, we saw so much whiplash in the market this week about what's happening here.
- [29:47]
A
You know, there's a bunch of other things going on like us levels of debt and what's going to happen with employment or not. But I think at the heart of it is that on this call are people who are in general probably more closely attuned to what's going on and the trends that we see. And you know, I think, I think to, to Jaime's point about us not wanting to give up capabilities, it's, it's absolutely right. I mean, I have, you know, I have a prompt about a model evaluator that I, that I built and you know, occasionally I throw something back to a GPT4 class model and it's kind of moronic. The response, I don't want to deal with this anymore. And the one place within, you know, within my team where we do use quite a lot of, of models, but you know, there is a lot of batch processing and we'll throw that into deep seq 3.2 and that can be cheap, but we're never going to put that on the edge because it's, it's like batch processing. Why would you put your kind of core infrastructure on on edge devices? And then I think people are a bit disciplined, more disciplined about how long do you want to wait and how much do you want to pay for a particular class of output? And as you start to get more and more value from the output, you're willing to pay more. And I think that for a lot of companies, and maybe this is true for many investors who are still sagging on a Microsoft copilot license, they've never really had the breakthrough moment of getting Claude cowork to do 10 hours of tedious manual work in, as Jaime points out, 35, 40 minutes while you go off and do something else. And once you do that, you sit and you say, well, actually it's worth paying £75 or 100 bucks a month for Claude Max, in order to take this off my desk. And so when I looked at what happened in the markets this week, this was an overreaction. I mean, the market is always right. So let's just sort of get this right. Market is never wrong. You can never bet against the market. They will always stay solvent longer than you will. However, with that said, they didn't get the demand growth. They didn't get the way in which demand is outstripping supply. They didn't get how much more we were going to demand as these models get better. I mean, the moment a model can work for 20 hours, I can tell you, I mean, I will be running hundreds of these things because I've got a lot of work to get through. And I'll be, you know, saying, hannah, how many models are you running? I think I probably sent that to you already. We just introduced. If you don't max out your Claude usage at least once a month, you lose Claude max as a tier. You've got to be maxing these things out because they're so powerful. And I think that once those, that realization moves into Main street, which it will in the next two years, or parts of mainstream, we will just see the usage grow and grow.
- [32:38]
B
Yeah, Actually, it reminded me of the study I wrote about. They tested the models on financial analysis to see, you know, how. How well it could do on, you know, and it wasn't as good, but it was improving. But the comparison was like, all right.
- [32:51]
A
Well, if you, if it's a human.
- [32:53]
B
It'S like $20 an hour.
- [32:56]
A
If it's an agent, it's like 50 cents.
- [32:58]
B
But I mean, so it was. But that threshold is, is, you know, that gap is closing where you can get it, you know, they're making headway into, you know, solving the sort of financial analysis questions. So curious. What, what do you guys think was like, diving into this was the, the most surprising conclusion you, you walked away with.
- [33:16]
D
I was actually quite surprised the margins were where they were, as in the gross margins, because I wasn't expecting them to be around 50%. That seems pretty good for a model where people are saying, you know, they're always having losses year on year on year. And although we heard from Dario and Sam that, you know, at the model level, you know, insinuating there's a profitability, to actually see that come out in the numbers is quite reassuring in some ways.
- [33:47]
C
Yeah. I think that for me, I came away with a more pessimistic view that I had. But you might have picked up on the fact that I'm very bullish on artificial intelligence already. But I think that after looking at it, I was expecting to come here and find resoundingly that they already have a profitable model. When you look at it through the life cycle of the model so that they're making enough of a gross profit to completely offset the cost of development. And it seems that no, this is not the case. It seems that this is way closer than what I thought what I thought it was. And this took us to me more than ever how reliable the companies are right now on investors goodwill and how much they are relying on this story of we will be profitable later to make. So maybe this is the counterpoint. It's like it's. They seem to be great at inference. Once you have the model built, they seem to have a great business in their hands. But after you account for the cost of developments the thing looks much closer than what I expected. I still remain bullish, but I'm now much more temperate than when we started to look into this.
- [34:51]
A
There's something that the market has done which reflects what the work that we did, which is that if you look at the OpenAI stack of affiliated public companies, the Oracles and so on, they have really underperformed relative to the Google anthropic stack of affiliated public companies. And there's been a divergence since. I think it was really when Sam talked about needing a trillion dollars or something which he may have done last, last summer, I found it really helpful that we had gone and done this X ray of, of this, this company because it helped me see what I think some of the choices are. So the first thing I would say is there's clearly, there's clearly a path to success that emerges for OpenAI based on the work that we did. One of the big drags is this 20% Microsoft cut they have to pay. So they did this deal early on in order to get distribution and compute. Actually I think it was years ago, right before ChatGPT where Microsoft got took 20% of the sort of top line revenue. And that drag does get in the way quite a lot of like an independent successful business. And that's going to be a commercial negotiation as we've seen between Microsoft and OpenAI. It's unlikely that Microsoft will shoot the prize pet on that journey. And I think what I was able to therefore see was certain of the levers that they have to control on that journey. As Jaime says, they're going to need a lot of money to get there. I think the other thing, just as a contrast was we had often written about how OpenAI was trying to capture a lot of different Fronts at the same time. So it was trying to capture the sovereign national government front. They were very early with their large compute infrastructure commitments through Stargate. They wanted to get the enterprise, they wanted to get universities, they wanted to get consumers in lots of different ways. And that often flew foul of actually how Y Combinator, which Sam used to run, would encourage founders to work.
- [36:58]
C
Right.
- [36:58]
A
It's like, find a beachhead, stick to the beachhead, and then grow. I think DoorDash, which is a YC company, a successful food delivery company, was originally called something like Palo Alto Pizza Delivery or something like that. I mean, because that's what they did. And I found it quite surprising over the last year or two that they had gone so broad. And I think one thing that this X ray of the business showed was that there is something appealing about the way Dario has run Anthropic in terms of its really deep focus, which means that lots of the speculative investments that you might need, either in product to consumerize or in sales and marketing and awareness, you don't need to make. Right. You're able to focus a little bit more tightly. And that showed that there is a, you know, another path. So I. I found the. I mean, I certainly learned a lot from it. And observing the hard work that Hannah and Jaime were doing, that's been the.
- [37:49]
B
Focus, obviously, Anthropic, for a while, and OpenAI has been putting its weight there.
- [37:55]
A
A quick note, if you want to support us in bringing more of these conversations to the world, please consider subscribing to the show.
- [38:02]
B
But a lot of the folks I talk to are sort of agnostic about the models that they're using.
- [38:06]
A
They're. They'll.
- [38:07]
B
They'll sort of pick one out, use one here. And then, you know, we had the. The market crash because Anthropic put out a legal plugin. That was surprising to me. So. And that's a lot of the conversation.
- [38:18]
C
That I have with people.
- [38:18]
B
They're not. They're not sticking to one necessarily. They're. They're building the infrastructure so they can swap them in and out. So I'm just curious how, you know, how you guys think about how do you compete in that space?
- [38:28]
C
It's very hard. It's so hard to compete in all of these spaces. I do like the. I do like the perspective of focusing in the end game here, and I like more the end game right now of focusing on business, as SIEM has. Has alluded to, which is a little bit more the anthropic focus that this is the Place where I was talking before about where do you spend a hundred thousand dollars a year on an agent? Like that's not something that an individual is going to be able to spend right now. This is something that a company spends on their employees and that they might see a reason to spend if they actually believe that AI provides them as much value as a marginal employee would. But you need to get, you need to get there. And what I see a lot of what OpenAI is trying to do is a lot of these speculative bets that bridge the gap between where they are now and where this, where these hundreds of thousands of hundreds of billions of revenue are. And some of these go well and some of these don't go so well. One that I will say doesn't seem to be going so well is Sora 2, which we actually looked over the course of the investigation. I was surprised to learn that it had significant, though not overwhelming compute costs. From what we could tell from afar, it was entirely subsidized, very much driven by the demand of their free users. And it seems that the usage has actually dwindled since its release until now. So it seems to be an example of an avenue for a new market that hasn't quite panned out for them.
- [40:00]
D
We've heard from Sam Altman that they've slowed hiring. However, we've also heard from him that they're now trying to hire more consultants to actually get, you know, purpose built solutions with OpenAI in enterprises. So maybe at the employee level you have people switching a lot between models. I mean, I personally do a lot between the different models, but if you're building systems that have your entire company's data connected to it and workflows from that, I could see that, you know, having one company work with that is a bit easier and maybe that's the route that they're trying to go and stay competitive there.
- [40:37]
A
I think one interesting thing to do would be to figure out how large enterprises move from model to model when a model release, you know, emerges. And we what we saw with Opus 4.6, which is a new anthropic model that was released ages ago, I think the day before you recorded this, it's probably been superseded by something was that 4.6 was not accompanied by the usual list of 50 great software companies from replit to notion to others saying we're using 4.6 from the get go. I don't think that's to say they're not going to, but it's more to say that there are still change costs Associated because a better model is not better in every direction. And there may be certain classes of prompts or bits of your workflow where it doesn't do as well as the one that you've, you've tested it. And so there is this balance between, you know, do you exploit what you know or do you explore with the, with the new model? And I've certainly found that. So we have a whole load of, you know, agentic flows or workflows or decision systems that run, that are still running on things like Gemini 2.5 flash because they do it reliably and we don't benefit at all from a better model on that particular use case. And the time we would spend to go and make that change is better off spending on something nearer the frontier. So I do expect there to be this kind of long lag of models, particularly in larger enterprises. But I have also heard, and Matt, I shouldn't do this, but I'm going to put it back to you because you talk to the banks all the time. I had heard that a lot of the banks now do use these model routers that sit on top of the underlying provider. And that some of the banks have said we want to move to more open source models so that we can have control, better control open source open weight models than exclusively relying on closed weight models. Now, of course, the reason we choose banks is because they're the most advanced incumbents, right, in using AI. But I think there may be some of that pressure emerging.
- [42:49]
B
Yeah, that's my sense at least. I had a conversation with Man Group, which is like the largest publicly traded hedge fund and that's what they're doing. They're sort of agnostic and they'll sort of use it to power how they're using it for investment idea generation. But as you were talking, I think it's true if you're a massive company, if you're 200, 300,000, are you really going to be. You have too much on the line to be swapping out. You have all your ip. I think there's a whole other governance issue across all of these industries. Like, okay, what are best practices for this sort of thing? Which is still, I think, you know, emerging. And so I think the, you know, Man Group is large, but it's, I think it's a few thousand employees, right. So it's not JPMorgan's like 250,000. So I think there's some back and forth there.
- [43:36]
C
One related point related to Aseem's favorite topic last week apparently which is mold book, which is that in malt book, in this, a social network for agents, there's some post of agents that claim that they have been changed. What the model is that runs the agent in the background and yet they retain this strong sense of identity and they write about, I mean everything there is kind of like, who knows, maybe it's fake, maybe it's written by a human, maybe it's just an invented experience. But they related this experience of like, oh, now I've been changed. Before I was like a GPT5 model. Now, now I'm in the background, I'm running a Kimi model and I'm, I'm so much slower of thought but they still identify with the same thread. And this kind of like this is interesting to me because what it suggests to me is that the unit of persistence of this, of these agents is not going to be so much the model that runs behind, but the memory and the history of the models themselves. And if this ends up being the case, like oh actually this problem of you need to change your models every so often, it might end up not being as big of a deal as you might naively think. You might end up in a regime in which you saved the history of your agent, you saved the interaction that it has had and how and its history. And you can seamlessly adapt that, change the model in the background while still retaining context coherence. And that if that is true, they're like, oh my God, this means that adoption of AI is going to be much faster than I was expecting and that the stickiness of the models is going to be much less pointing to much more competence.
- [45:09]
A
Yeah, I think that that is one of the ways in which openclaw, Claude, whatever we multiply has what it's opened up. It is opened up that idea that you can flip back and forth between models. And I personally found using that Mini Arnold has a decent enough memory, even though I am flipping the models through the different anthropic models. And the question is, you know, how does that then make its way out into the, you know, into the industry at large? And if that, if that way of delivering a virtual worker is going to be what happens within the enterprise. Enterprises make decisions that are not just about cost. Right. That's why even though LibreOffice has been available for doing, you know, Office PowerPoints and Excel spreadsheets, you know, the dominant player is, is the Microsoft Office suite. And that might be where, you know, the, the anthropics and the OpenAI's you know, maintain that even if there might be tranches of consumers and it would feel to me like Apple would be really well suited for a kind of open claw approach, once they felt safe with it would go. And that I think comes to Hannah's point about these consultants, these four deployed engineers, which is getting the, the clause, as it were into the enterprise, which I think does provide some defense against the model being completely abstracted away because someone has been in there, they have done some tuning, there is some know how that is going to get lost when you, you swap out an anthropic model for a Kimi or a Quen model.
- [46:46]
B
I'm, I'm curious, you know, you're talking about Claudebot and you know, I, I think like buying airplane tickets will be different in a few years, right? You'll, instead of just constantly refreshing and buying it on Tuesdays or Wednesdays or whatever. The trick is you'll have, you'll be like, you'll give it the parameters, you'll say this is, I'm going here, I want to spend this much economy. And you'll just let it go and it'll buy it for you. And that to me seems sort of inevitable. And that's a whole other level of compute consumption, right? That's. That you know, is maybe not as fun as Quadbot, but you know, it seems like it would be significant as well.
- [47:22]
C
I'm very bullish on this. Everyone is going to have their personal AI assistant and it's going to be great. And we're just going to have our AI assistants talk to each other and do business for us. And this is, this is great. Now mind you, I don't think this is where the, where the hundred billion trillion dollar industry is. I think that for that you actually need to go to go to business. But I'm pretty positive that this is a way that you can bridge the gap between there and where we are here. And it seems eminently doable. The models still need to, they need to get a little bit better, a little smarter. I will not trust one of them with my credit card. Not yet.
- [47:56]
D
I think the ChatGPT operator was planning to be acting as this agent who can buy things for you. I mean in that demonstration when they launched it, they were trying to buy groceries. I can't remember if it was successful or not, but I don't think it really got taken up in that way. So it's coming, I guess.
- [48:14]
C
I think this research was very fun to do. I think we have learned a lot. One framing that I think it's important to have in mind is this framing of the models as these rapidly depreciating infrastructure that lose value very quickly as new as new models and competing offers came out. And this being an important part of how you think about the AMOS while still being bullish about it. And one other thing that I will reflect is like this is probably not the final word. Again, we have done this hermeneutic exercise of interpreting public information out there, but we don't have access to OpenAI's finances. So we're learning a lot. Actually, we have already learned a lot from the public response to an article and we will continue. We will continue researching and doing so.
- [48:57]
D
Yeah, we also in the beginning we were thinking of also doing the same for GPT4, but obviously given how expansive GPT5 ended up being in terms of an investigation, we don't have that comparison historically. So it would be interesting to see we have a state of now and the margins now. How that will change in the future is still to be seen.
- [49:19]
A
Well, there may well be an IPO for OpenAI this year or early next year, so they'll be there. Something to mark our homework against. Thanks for listening all the way to the end. If you want to know when the next conversation is released, just hit subscribe wherever you're listening. That's all for now and I'll catch you next time.