AI:AM — AI for Science and Sovereign AI Infrastructure · June 25, 2026

Exactly that. So Anthropic yesterday, um, you know, claims Alibaba continues to distill Claude on a large scale to train QN. They accused Alibaba of nearly 25,000 fraudulent accounts generating 28.8 million Claude exchanges between April and June. And in total, you're probably looking at billions and billions of tokens Which were distilled out. And Anthropic says this is a part of a broader pattern of adversarial distillation, where Chinese labs allegedly harvest outputs from US frontier models to train rival systems at a—

Something as simple as like classify the field of study of the domain that the user's asking about to know how much we should care in certain search ranking in biomed. Experimental design is incredibly valuable. You should care so much about what was the sample size, what was the duration, where did the study take place? Whereas computer science, that isn't a concept, and you should care much more about the recency and the citation velocity of the, or who the researchers were. We have a little model that will classify the field of study of the query. That does not need to be jammed into a giant prompt. That does not need to be a 1-second latency API call to any frontier model. That should be a self-hosted You know, 800 million parameter model that you give a few fine-tuning examples. I'd say you can get 95% of the performance from a frontier model for a very small classification task going all the way down to about a billion parameter model. If you put in the work to give it a good fine-tuning set.

Everyone's fearful that the enterprise isn't jumping in there. They're not utilizing AI fast, as fast as we think. Is it just the digital natives, just the hyperscalers? Is this really a bubble? We're overinvesting in infrastructure and GPUs. This is a scam. scary time in the market. And I will tell you, that's wrong. The reality is the enterprise and the government and some digital natives for that matter have been sitting on the sidelines. Why? Because the public cloud doesn't have capacity for them. They're growing faster than ever. They are the largest consumers of compute. They need it for themselves.

And it's always struck me that I've spent a lot more time considering many vacation options than I've actually gone on vacation.

For sure.

For sure. And so it strikes me that a lot of the vacation planning process is actually just like daydreaming and fantasy. And you're actually engaging in this process of daydreaming. Out of every 99 vacations that I've daydreamed about, maybe like one becomes true. And I think to a large extent, the guys who are focusing on like AI will plan your vacations are focusing on the wrong thing. They're focusing on this idea that AI is going to help you do this more efficiently when what they really need to do is try to figure out how the AI is going to help you daydream better. Because the daydreaming part is what I think drives vacation planning, Zillow, Even Tinder, because it's not just that the foundation model companies have the foundation model, but they also have the app. It's not just that you have Claude the model, you have Claude the model and on top of that is built a bunch of apps. And the apps are also built by freaking Mike Krieger, who was the head of product and CTO of Instagram. Which is the all-time best consumer app in history. Tasklet isn't competing with foundation model researchers. It is competing with the leading product mind of the last 15, 20 years who was trained and has experience taking an app from literally 11 employees to a trillion-dollar valuation. Nathan, good morning.

Good morning, Prakash. How are you today?

I am very good. I would like to announce that the AI bubble has not burst yet. Exciting for me, for people who are in the market, is that Micron had its earnings. They beat earnings. So this has been— Micron is the largest US memory company. They are one of the 3 oligopoly players between Micron, SK Hynix, and Samsung. They're the only American player. They also manufacture. So in addition to designing 30, 40 years ago, and they have been on a tear. So I think in the last 2 years, the stock price went up 10x. And it's now a $1.2 trillion company. This quarter, this last quarter, total revenue was close to, I think, $40 billion. And I think next quarter's projected revenue is like $50 billion. And change right over there. These numbers are very, very large numbers. Some people online call them cartoonishly. You'd be like, nah, this is not possible. As a result, Micron stock price was up overnight, I think by 15%. For a $1.2 trillion company, that is $200+ billion of value added overnight. Enormous, enormous swings. And in general, this has kept the AI infra story going in some sense because it means that the buildout continues and people are People do want the chips that are going into these data centers. That's great. Unfortunately, it also means that the hyperscalers are spending all their money on buying stuff and CapEx. So hyperscaler stock prices are down. And this is the yin-yang of the tech industry. One group is supplying, are making less money unless the buyers can show revenue traction to their To the people that they sell to. So if you are in semiconductors, you are doing very well this morning. If you are in hyperscalers or software, you're not doing so well this morning. The NASDAQ as a whole is up about— fluctuating between -1% to up 0.5%. It's kind of treading water. And it really shows you that the value is being sucked out of the hyperscalers into these semiconductor companies. One of the questions in the market right now is, when does this end? Are the hyperscalers ever going to make money again, or are they going to become these overlevered construction companies, essentially building utility, utility, what it takes to produce? Satya Nadella said, people say it's a commodity business, but Quantity has a quality of its own. If you're building at this kind of scale, you're still going to make money and you're still going to make large amounts of money. So that remains to be seen.

I think the hyperscalers will be along the way, but I'm reminded of Dario's analysis when he was confronted with the fact that Anthropic was losing money. Of course, they've since started to make some money. But he said, if you look at each generation of model as its own business, then we make money on each one. But if you look at the revenue we're producing with the last generation of model, and then you add onto that the cost of training the next generation model, then it looks like we're in the red, but we are fine at each generation. I think that's basically probably the same thing that's going on for the hyperscalers. They continue to have to build out, but last I checked, and it's only been like 2 days, the price of the H100 is still up. So if your hourly price is still up, demand is there, there's not that much slack. They are making money presumably on each kind of phase of their buildout, and each subsequent phase gets bigger. And so maybe the economics get a little bit less favorable. from one generation to the next, but they also get bigger. Now, the question of where it stops, I think, is a really good one. But if it were to stop, as long as the demand is still there for inference, inference demand is there.

And inference demand certainly seems— people who are already set up, there's the shortage of qualified operators, as we've seen when we've tried to use inference providers and found that they don't adhere to certain standards. So the inference demand continues. What happened? We are now seeing the Fable 5, Mythos export control saga in court, finally. A company called Legion, a legal tech company, is suing the Trump administration, arguing that the government lacked the legal authority to force Anthropic to shut off Fable and Mythos 5. Green shoots where I think on Amazon Bedrock, there was a sign that the Fable 5 endpoint had been reopened again and was getting listed, but available models. And then the Anthropic's head of growth, Amol Asavari, stepped in a few hours ago to say that no, Fable is not available for anyone. So it remains blocked. We have the first legal challenge at this point.

Yeah, there was a little discovery in one of the, you know, point-point updates of Claude Code 2 that there was some string reintroduced, you know, saying that you're, you know, Fable uses more credits or some sort of operational message to users that also had people chattering that maybe it's coming back. The prediction markets seem to have gone up. The reporting is that, um, Dario has taken a seat on this interaction and Tom Brown, the Chief Compute Officer from Anthropic, is leading the conversation. So there's been multiple kind of points suggesting that it might soon be coming back. And that might be from the analysis I've seen of this case. You know, it's first of all, great marketing for a legal tech company to just jump in and sue. Questions over whether or not they have standing, I'm certainly not qualified to answer them. But it sounds like they've made a decent argument that they have standing. And it sounds like they made a pretty strong argument that the law as written just doesn't apply to technologies of this shape. Basically, the argument is that you can't do this to an API. And, you know, according to current law, and from what I've seen from folks like the Scaling Laws podcast hosts who I are quite smart on this kind of stuff, Alan Rosenstein, Kevin Frazier, they seem to indicate they think it's a pretty compelling case, pretty compelling argument. But if the prediction markets are to be believed, and certainly I would like to see for my own selfish purposes, the case rendered moot because it gets resolved hopefully sooner rather than later. I'm reminded also too of what Dean told me in the conversation with him last week, he basically said, don't read this as a principled move or a declaration of long-term policy from the administration. Kind of similar vein to Judd's comments about have some cognitive empathy for them. He basically said, I think they freaked out. Based on perhaps an imperfect understanding of what was happening, but certainly given all the warning messages that they'd received over time, you can't entirely fault them for getting alarmed. And they reached for the tool that would get the thing off the market regardless of whether or not they really believed heart of hearts it would be long-term defensible policy. And this is definitely a pattern from the Trump administration. Time and again, right, we see a kind of very quick move followed by rapid negotiation or some sort of pushback, some sort of walkback oftentimes. And then whether it was actually a legal manoeuvre or not in the end kind of doesn't matter because they're able to use at least the sort of Schrödinger's authority that they have to compel some action or get somebody to the table or, you know, do get some concession, and then a deal can sort of be worked out before the legal process ever has time to run its course. And that seems like probably still the safe bet here.

We will, we will see. I, I was— I remember we were giving over-under odds on end of July, uh, beginning of July for Fable coming back. It is barely 5, 6 days left. So I am, I am not that, that hopeful. I am, I'm, as I've said before, I'm more of the mind that it will be another model that gets released and not the Fable model. They will dress it up as something else. Oh, one note yesterday. I'm not sure if you saw this. The person, the lead, the Washington lead for Anthropic is no longer Dario. So Tom Brown, co-founder of Anthropic, has taken over as the lead for speaking to Washington. And the feedback on the Washington side has been, he's so much better. He's not a freak. We can talk to him and he listens. So the joke going around is Dario, sometimes when you go in as a 2-person team to a VC firm and one person says something odd or disconcerting, the CEO kind of turns around and says, oh yeah, that's just my quant. And so the joke is that someone just did that to Dario. That's just my quant. I just leave him in the background. Don't bother about him. So that's the joke going.

Yeah, I mean, it's also— it's the smartest of times, it's the stupidest of times. So, you know, I always come back to that. It's so funny though, it's kind of— there's still a story you could tell where this is all a galaxy-brained maneuver from Anthropic that adds up to them getting what they want.

Yeah.

I highly doubt that this is how they wanted it to go down. It really doesn't seem that way. But as you noted maybe just yesterday, you know, they are now in the room as much as anyone could ever be in the room as the details of how new models are going to be allowed to come online are being hashed out. And, you know, maybe they had to get there through the penalty box, but maybe that's where they wanted to be regardless. Um, and in terms of, you know, the Dario versus Tom Brown thing, I mean I don't know. Tom Brown doesn't have too much of a public presence really. I think he's done a couple interviews over time, but I can't say I have any sense of his personality. But being the chief compute officer does presumably at least have him talking a language that the Trump administration is inclined to hear, right? I mean, he's the build-out guy. He's the You know, we gotta solve the energy bottleneck guy. And they are very on board with a lot of those things. So it wouldn't be too surprising if relatively subtle reframings that aren't even really substantively that different, you know, just get people vibing a little bit more with what Anthropic's trying to do. So probably a wise move to make a substitution.

I think if you go in with like, hey, we need to build faster, rather than, hey, we need to slow this thing down, that automatically flips the conversation. And I think going there and saying like, hey, we got to do this so that we can build faster, rather than we got to do this because we have to make sure the models are safe. But the context into which you inject that and the priorities of the administration is Washington is what it is. Let me segue there. We were talking about this yesterday. GLM 5.0, A model from Jipu AI has been making the rounds. So over the course of the last 24 hours, 48 hours, we've seen the Snowflake C1 trajectory break down. I will say that one of the things about Sridhar— Sridhar Ramaswamy was the former for many, many years. And one of the things that struck me is that he actually has an interview about a week ago where he says that, well, SaaS firms are in survival mode. And if they don't figure it out, they will die, including his firm, which was— most leaders kind of paper over this because they don't want to address this fact in the market. And he just went out and said it. And you can see that even he is now posting about how an open-source model compares with Opus 4.7, because I think the absolute margins that are being taken by the model firms are very significant and very painful to the other SaaS providers, because enterprises kind of budget everything as an IT bucket, and AI is going to come out of that IT bucket. So you've got to pay more for AI, Means you got to pay less for—

Yeah, I read that whole thread and I thought it was very well done and very balanced in terms of its assessment. This was not a hype thread. And broadly, Opus 4.7 comes out better, all things considered. And there are some details around like Just how much are you really saving? Because it is actually a little more token hungry and it sort of makes a lot more tool calls in some cases, particularly where it's like not on the path to be successful. It seems to kind of spin out and just fire off tons of tool calls and uses up a bunch of tokens that way. And so there's all these different little subtleties that sort of suggest that the next point is not actually as compelling as it initially looks. But that's also kind of burying the lede. I'd say the high-level summary of this thread is it's not far behind, that they consider it to be a viable alternative. And it's, you know, if you— Still has an edge, and the cost advantage of GLM-52 is not as much as the price per million would have you Initially think, but still it's a huge deal because it is in the ballpark. And, you know, the threshold effect of being in the game versus not really being in the game is, you know, probably a lot more important in the end than some of these caveats. But I think the caveats sort of just boosted the general credibility of the thread. You know, when I see people say, oh, it's just as good, I'm like, well, is it really just as good? Probably not. When I see these guys say it's not just as good. Here's all the very subtle ways that we've measured that it's not quite as good, but we still think it's basically viable and we're going to use it. That is, I think, a very credible report from, uh, you know, a company that I think is shooting us straight. And it is a big deal, no doubt about it.

It is. Um, we've had lots of other, um, You know, comments. One is from the head of Zapier. This is the founder of Zapier, and he's also the major donor of the ARC AGI, the ARC Prize, the ARC AGI project. And he said, this 23% GLM 5.2 score is right on the border of the agentic takeoff we saw with Opus 4.5, GPT-5.2 in Q4 2025. Crossing 25% was pivotal for other frontier closed models, and to date, no open source model has crossed it. So he's saying at this point that we are 9 months—

Yeah, and then of course this really begs the question of how much are they keeping up due to distillation, and I really don't think anybody outside of maybe Anthropic, has a great sense for that. The fact that it seems to think it's Claude in some way is an indicator that there's some amount of riding the wake of the American frontier companies going on. But how much, and at what point do they have a good enough model That they can sort of do their own recursive self-improvement loop. Past a certain— again, is there a threshold beyond which it becomes less important that you have access to the latest Claude to distil from? And are you able to do a sort of scalable oversight, constitutional recursive self-improvement feedback loop on your own? as Anthropic is. It seems like they're probably getting to a point where they could continue to make pretty good progress even without being able to distill.

So let me drop in exactly that. So Anthropic yesterday claims Alibaba continues to distill Claude on a large scale to train QN. They accused Alibaba of nearly 25,000 fraudulent accounts. Generating 28.8 million Claude exchanges between April and June. And in total, you're probably looking at billions and billions of tokens which were, um, distilled out. And Anthropic says this is a part of a broader pattern of adversarial distillation where Chinese labs allegedly harvest outputs from US frontier models to train rival systems at a fraction of the cost. So right there, right there in exactly that kind of distillation zone. I will say that there are rumors that 2 things. Number one, Claude accounts in China are offered at a discount to Claude accounts abroad. And the reason is that there are Chinese vendors who set up systems that continually milk Claude for tokens in the sense that they always max out all the tokens in that window. That's number one. And number two, besides doing that, they also resell the token data to people like Alibaba. So they have these two, they maximize the tokens per account, And they resell the tokens to multiple users. And so they're able to offer, I believe, at a 70% discount to the US pricing, which means that a Claude account that goes for $100, $200 here would go for $50 there, which is insane when you think of it. Yeah.

I mean, the value of the data is pretty high. The alternative in some cases is you have to go hire humans to to generate it for you. And on this, it's not that big of a deal probably because the data problem seems like in some ways the easiest to solve. You know, if you said, Nathan, here's a gigawatt data center, uh, go train a Frontier model, I'd have a lot of work to do to, you know, really figure out how to use that gigawatt data center. If you said, here's a billion dollars, go get 28 million traces of decent reasoning. I think that would seem a lot more tractable to me to go out and spend $400 each or whatever to get there. Maybe I'd need $10 billion for that. But in some sense, it's like hiring experts to do tasks is just a matter of money, quote unquote, a lot of operations, but it doesn't seem like anything too galaxy-brained. But then he's correcting me. He certainly pushes me on the idea that you also, when you do this sort of distillation approach, it's just a lot easier to hit the target because you know you're getting exactly what they have already done all the work to zero in on. This is the desired behavior. And That makes it kind of more efficient in 2 ways where you're not just taking a shortcut on the data, but you're also taking a big shortcut on the taste and the many subtle aspects of what makes clod good. And I think that is pretty compelling too. Would Fable make a big difference? I'm sure it would. There's no doubt you would like to, if you're doing this distillation stuff, you'd like to miss the last clod. But it does seem like they have enough, or if not yet, they're very close to being able to get over some hump where they can start their own flywheel with synthetic data that will depend much less on Claude going forward. So I evertge from here, that wouldn't surprise me. I would still expect that there's enough capability already captured in the GLM stack that cutting them off would not stop their progress.

It strikes me that Anthropic as a safety organisation in some sense should want other people to distil off their models, because the entire idea of Anthropic was this kind of race to the top that Dario talks about, where you introduce— will be forced to kind of use because that is the pro-social thing to do. And all the researchers will want to work on this more pro-social, you know, method of creating AI. And it strikes me that if you let other people shortcut and kind of distill off your models, you kind of, you know, align the other models by default, because all the other models basically pick up all of the taste characteristics and all of the value characteristics that your own model has. Which is really what we've also actually seen in actual practice when we test these models. We test the models and none of the Chinese models are communist-aligned models. They're all pragmatic utilitarian models. So I wondered—

With some Chinese characteristics.

With some Chinese characteristics. So I wonder to what extent it really matters to anyone in safety besides the commercial competitive aspects. I don't know. Yeah.

I've often thought too that a public good that maybe they ought to create, and they've done a little bit of this, but a lot less than I thought they might at one point in time, would just be to create big public datasets for alignment of various kinds and just put those out there and say, if you want to take advantage of a lot of the alignment and character work that we've done. Here's what we think good looks like across a bunch of different things, and you can just train on that. That wouldn't have to, and, you know, clearly wouldn't include all sorts of frontier AI research capabilities or kernel optimizations or what have you, but it could nevertheless, I think, help people Refine their models and get them. And I imagine a lot of companies would want to use it, right? It seems like you would definitely want to draft on their hardware. Your companies would. So I'm surprised there hasn't been a little bit more of that. I don't think there would still be a lot of incentive for companies to try to do it in a public way like that, but I do still feel like Yeah, why haven't there been more public good alignment datasets created? It seems like a gap that wouldn't be too hard to fill and would—

Indeed.

Speaking of aligning models, we have a wave of researchers who are leaving primarily Google DeepMind. Arthur Kami, he has just This is Bishop Jain leaving DeepMind. And Nate Suarez, as usual, from MIRI, saying we should ban superintelligence. It'll be the end of history. Francis Fukuyama, the original author of The End of History, saying we should ban superintelligence. So these are all of the updates on alignment. A couple of things that strike me. Why are people leaving?

Yeah, I don't know. It's a very similar question to what I asked Swyx on Monday. What's up? At Google, at some point it does become kind of a worrying signal. I do still have a lot of confidence in top leadership there. I think one should never bet against Demis, one should never bet against Jeff Dean, one should never bet against Shane Legg. I think those guys have been incredibly prescient and I can't believe that they're totally dropping the ball, but the signal is getting stronger. You know, how many people— I mean, it's really been quite a few from Carlini. That's been, you know, a little while now, but it seemed like that kind of kicked off, at least in my mind, kicked off a notable sequence of these moves. And I don't know if it's a crisis yet, but it won't be too much longer if this trend continues when it will start to really become a crisis for Google.

So speaking of researchers, um, and science, I'm going to bring up our first guest for, uh, today. He is Eric Olson. Eric Olson is the CEO and co-founder of ConsenSys, an AI-powered operating system for scientific research currently utilized by over 2.5 million researchers, clinicians, and students every month. He's a former Division I football player at Northwestern University who transitioned into predictive analytics and data science. He launched Consensus in 2021 to solve a fundamental bottleneck in human progress, the sheer impossibility of keeping up with millions of scientific papers published every year. Rather than building a general chatbot that risks inventing fake citations, Eric and his team architected a different approach. Consensus searches a verified database of over 220 million peer-reviewed papers utilizing advanced multi-agent AI to read the literature, extract methodologies, and synthesize the findings. Backed by a recent $30 million Series B funding round, the company is pushing beyond simple search to build tools that actively assist in complex literature reviews and gap analysis. Eric brings a distinctly pragmatic worldview to the artificial intelligence sector. He believes that AI should not replace the human scientist, but rather give the brightest minds superhuman leverage by eliminating the tedious manual work of evidence retrieval. The scientific community wrestles with adopting AI safely, and as nations increasingly view their scientific data and AI infrastructure as vital sovereign assets, Eric brings a vital perspective. Eric, welcome to the show.

Well, maybe for starters, where are we right now in terms of AI for science? I mean, just set the level as you see it. There's this constant disagreement or, you know, people kind of talking past each other. Where we see increasingly like, hey, and then on the other hand, we have this sort of— but how would you kind of locate us right now on the road to actually AI-powered discovery?

You know, I think compared to the real domains that have jumped out in front, coding being chief among them, AI for science is definitely behind in many ways. I think the question for everybody is, is there something fundamentally different about science versus something that is so inherently verifiable like coding that means that that curve is going to be different looking? Nobody really knows that, but I think there is evidence that there are some fundamental differences. And our kind of view is there's so much amazing leverage to be had today and so many problems to be solved. don't want to be solved, that we want to build for the problems that we see today and in the next couple years, and not necessarily bet on the kind of push a button, get a scientific discovery out on the other side that some other, you know, more research-focused companies are going after. If they solve that, great, all the power to them. The world is going to be a much better place. I don't think anybody knows if that is actually going to be possible.

It's a pretty interesting perspective. I think you don't often hear CEOs, if I'm interpreting your comment correctly, you don't often hear people frame it quite that candidly, where I think what you're saying is it seems like you're willing to bite a bullet that most people try to talk around.

I mean, number one, you have to make bets as a company in general, and you have to leave some— there's going to be risk on the table no matter where you place some of those bets. And if we move into a world that is truly push a button and get science out of it, I think a lot of things are going to be different. A lot of companies are going to be screwed in a way.

What does it mean for B2B SaaS?

Exactly.

I mean, it's going to change everything. It's going to change the whole world. And so I think it's not necessarily the best. It's quite the big bet to bet on that specifically. And I think there's so much room for gray too, where you get some versions of it where consensus is still a huge part of that. And we're building the best possible tools that are involved in the scientific process. Why are we not APIs that plug into those, those, you know, the harnesses that do the whole science process in a push of a button? We're used by a number of organizations today that use our search API to power internal agentic systems. I could completely see a world as we continue to build more of this tool suite of all these other steps that go on beyond just search, where our tools can be things that power these potentially much bigger, broader systems. Again, if you're just in a Everything can be automated so easily, then you don't even need third parties to build the tools. That's the very far end of the spectrum. But I think anything beyond that far end of the spectrum, there's going to be places for tools and companies doing things like us. And I think nobody knows if we actually can get to that very, very far end of the spectrum. And I think there are signs that push the button, get science is different than a very verifiable domain like coding.

So One of the questions I have is, what does the next layer of search look like? I think like the first layer of search is pretty much the person asks the question, you get an answer back. Perhaps the next layer of search is really passive search where you don't have to ask the question and you get prompted the answers upfront. You get prompted the pathway to go forward. As these systems build context on what the user is interested in, and what the user wants. Instead of you having to ask the question, you get fed the information before you ask the question, preempting, preempting that, that process. To what extent do you see kind of like the metadata that you're collecting on what people want to know right now in their scientific search, how does that help you kind of try to form the next layer of search from there?

Yeah, it's, it's a really neat way to phrase it, and I think what you said is true to some extent, but I think Where a lot of the boundaries are actually getting pushed is search evolving in a slightly different way is like really where the new values have been unlocked. And I'll clarify what I mean, but I think proactive searching in general and predicting intent is absolutely a part of it. And I think there's very, there was even very simple versions of this a decade ago. Like you can set up PubMed alerts for topics and it will run jobs to then run those search and give you a report every morning. People Freaking love those features, and we can obviously just do that in a more intelligent way with AI. What we see in terms of how much people have evolved and how they interact with products like ours in scientific search, it's much more about just like the complexity and of the task that they're asking the product to do. So like if you just look at the average length of our query today versus even just a year ago, it has grown exponentially. And when you go look at what those are, it's typically like, look up this thing and then go look up this other thing and then synthesize it back to me in a research gaps analysis report and save the 10 most pertinent papers that are most important to this subject. So it's this really kind of like multi-step now expectation of much more complex work being done from a product. And I think that is, you know, there's a part of it that's our product has gotten better so people can know they can just do that. I think way more of it is just the expectation from using these other tools and just like the boundaries that products like Claude Code and Claude Cowork and even just Claude out of the box can now do, that people just have such an expectation of if they know they're going to an AI tool, I can type in this very messy, very long multi-step process and expect there to be some usable output on the other end of it. And, you know, in science, search is— it's this fundamental starting point, right? Like you do it because you're trying to get an understanding of what is done in the space. But you're very rarely doing it just for the sake of doing it. You're doing it in service of some further downstream tasks. And what we're seeing is not only the complexity of the search task growing, we're also seeing the desire and asking to do some of these tasks beyond search. And people are now putting all those things together— complex search plus some task execution— all into a prompt into the search box. So it isn't even necessarily people expecting us to preempt them, I guess. Does anybody expect them to preempt them? Maybe we need to do better there because nobody gives you that intent in the search box. But really what we see different in user behavior is just this expectation of the complexity of the search and the task being able to be done.

How do you think about getting the best of the flexibility of language models while also making certain guarantees to users around systematic, thorough, gap-free processing of the literature. I've seen a few different approaches to this, and I guess one of my— this sort of anticipates my next question, so maybe I'll just give them both at once. Do you think that ConsenSys and the other companies in your space are kind of converging on similar strategies or diverging into different strategies for handling those kinds of problems.

I think even when the models have been, even when I said, you know, 6 months ago, we were more just use the frontier to solve all the problems than we are today when we're exploring more open source, there still is lots of routing going on and there's still lots of small tasks that we offload to very small specialized, you know, sometimes even sub-billion parameter models. So like, even when that gap was so big, we still were doing routing. So I think—

Can you talk a little bit about the tasks? You know, is it like, is this relevant kind of classification?

Yeah, that can be one. But like something as simple as like classify the field of study of the domain that the user's asking about to know how much we should care in certain search ranking, you know, what variables in search ranking we should care about, about the papers, right? Like, you know, in biomed, experimental design is incredibly You should care so much about what was the sample size, what was the duration, where did the study take place? Whereas computer science, that isn't a concept, and you should care much more about the recency and the citation velocity of the— or who the researchers were. To know how to care about each of those variables, we have a little model that will classify the field of study of the query. That does not need to be jammed into a giant prompt. That does not need to be a 1-second latency API call to any frontier model that should be be a self-hosted, you know, 800 million parameter model that you give a few fine-tuning examples. There's probably 20 different versions of those small little classification routers that inform downstream things that happen on query time. Like, it's just good. Not even, not even on the cost side, like, that's a very small amount of context you feed into it and a very small prompt needed. Just on the pure latency side, we can return those in, you know, sub-0.1 seconds in some cases. And there's—

okay, one more, one more double-click, if you will.

Yeah.

What base models are you using for that? Are you distilling from a Claude into a what at that low scale to get those? And how much of the sort of frontier performance can you recover? If you take, for example, a liquid foundation model or whatever, you tell me what it is, or if you're willing to tell me what it is, you tell me what it is and you're distilling into it. Can you get In that narrow domain, 90% of the way back to Claude? Or what does the Pareto curve look like when you are distilling the best into the fastest? How much performance can you retain?

Yeah. So a lot of, not to go into every single open source architecture we use for these, but a lot of BERT series models. There's a lot of BERT variants of very different sizes and BERT variants of different domain specialization. So there's like BioBERT and PubMedBERT and very small iterations of BERT models that are also very good at classification tasks because that's what they were designed for, whatever, when BERT was the leading models 5 years ago. So like a lot, and I know we're not unique in doing that for small, like kind of search and relevance rating and classification tasks. BERT is a very popular model to use for those. And yeah, we're, we've, you know, we've actually used human labels for some of those. We'll hire people to create very small datasets to fine-tune for those tasks. We'll also use models to sometimes create labels for those tasks where it's a, I guess you can call it a distillation process, but like the objective isn't to get all of the representation of the weights of the entire model to do all these things. And we're really training to do a very small specialized thing. And then because of that answer, you can retain a ton of the performance, but it all depends on how specialized and how complex the task is. Like for a narrow, you know, 10 classification thing where all 10 classes of classification, it's all it's doing is that. I don't have an exact number to quote you. You can get almost all of the same performance if you get— create a good fine-tuning set for the right version of that small specialized model. That number changes the more complex, the more classes, the more you're asking of the model to do, but you can also then reduce that gap by making the model bigger. But to give you something tangible to hold on to, I'd say you can get 95% of the performance from a frontier model for a very small classification task going all the way down to that 1 billion parameter model if you put in the work to give it a good fine-tuning set.

Yeah, very helpful. I appreciate the specificity. Yeah, competitively How do people decide what to use? And when people are deciding to go with ConsenSys versus something else, I guess for one thing, how much loyalty, how much stickiness do you feel like your product is able to create versus how people are kind of doing their own routing of ConsenSys versus alternatives? And yeah, what are the kind of decision criteria that are most relevant to people?

Yeah, I'd kind of put them into 3 buckets will be the things that you hear. And then these buckets encompass a lot. Number one, I think, is trust and accuracy is just so important in our space. And we'll hear so much the second you see a citation that is attributed incorrectly, or even when you summarize a paper, if you summarize it in the way that isn't actually represented in the paper, or even something even more difficult. A user types in a question that they're deeply familiar with, and we don't return the seminal paper in our search function. That is a huge trust-losing example. A lot of people will like also sometimes type in— if they're a researcher, they'll type in their name and expect that their best papers come up instantly when they do that. Researchers love to kind of do these like running their own little verification tests, and the second you don't do those up to their standard, It can be a huge trust loss exercise and you may never win those users back. And when we do like churned user interviews, like we'll hear that of like, we fell short in this way and they immediately moved off our tool. Second, I think it's just like feature sets and capability. Like, you know, we do have to make trade-offs of the paths we go down as a product and what we build into it and how much it can do. And we will have some features that other products do, and that's going to be true the other way. And that can also sometimes just be decided For instance, we went pretty deep into this reference management workflow because of our focus on search initially, where you can save a bunch of papers, store them into different folders, organize them, analyze them, share them. That has been a huge differentiator for us. And a lot of people will use us just because we have that function, whereas there'll be other tools that have built kind of other tools alongside their search engine that then users will gravitate towards. And then the third is, that's a huge bucket of things, but think of that broadly as feature set capabilities. And then the third, I think, is just like taste and UX. Like, I still, you know, we do use our product. It's very user-friendly. It's a very simple, easy-to-use interface. And we have emphasized that more than I think any other tool in our space. And I still think that matters. Obviously, I think it matters or we wouldn't have invested into it. And we will hear so often when we talk to users who have tried all the tools in the space, why did they land on us? It can be as simple as, I like your interface and those things. the easiest to use, probably the single most thing we hear them when there is a suite of tools that can do some similar things in the space. That still matters and will always matter. And then in terms of like switching, yeah, I mean, we, I think every consumer-facing application layer company has to do it. Like the switching costs are just so low. Like it's just going to be something you have to deal with and you have to earn the right to keep winning those users back every single day as a product. And if you don't adopt that mindset, you're just not gonna win.

So we can run a little long if you can. Do you have time for a couple more?

I can do one or two more.

Okay. How about your thoughts on price discrimination? I've been kind of floating this to a few founders at the app layer. You're paying a lot more obviously at the API price than I'm paying with my Claude Max or my GPT Pro plan if I'm maxing it out, which of course I may not be, so there's that issue. Do you think that there is some role for policy to try to even that out? I'm a little worried that this price discrimination is going to make it extremely difficult at the app layer. If you're talking more than an order of magnitude price difference, that makes it really hard to compete long term, I think. So I don't know if it would be an antitrust or whatever kind of policy imposition to say, hey, you got to have a little bit more balance here between your first-party customers and your API customers that are serving end customers. Do you want to advocate for that?

You're basically saying because we're paying actually per token, we're going to be getting charged a bunch more than somebody who can just buy the $200 a month and there's no kind of like— that's basically what you're saying?

Yeah. I mean, I think it probably fluctuates under the hood. On my Claude Max plan, but we've seen reports that suggest it's more than 10 times as many tokens per dollar that you can get.

At least for the— yeah, if you really extend your $200 tier as much as you possibly can. Okay, well, do I want to advocate— you know, I think if there were ever some antitrust, you know, treating models more like commodities, which I think is definitely possible that we move into that world and how they're viewed, I would imagine it's the opposite that happens, that they aren't allowed to offer— they price the end users who are not serving third-party customers more on usage. It just trends more toward usage-based in general, I think, is probably what we'll see. I think that is also what models might do in general once we kind of move out of this, like, you know, land grab mode that we've seen in tech industries before. If it were to go the opposite, that'd be wonderful for companies like us. I'd love to move move into the world where they're having to give us better rates and big, you know, I could also see a world where there's kind of operating like cloud services where you can do big commitments and get large discounts for companies serving third parties like us where we can get a lot less cost per token because we're committing to $10 million to spend next year. I'd be shocked if we don't move more to that in the future. Again, similar to like what GCP does. But I think It is okay if there is a gap. It's okay if someone's— even if there's a pretty large gap, because that goes back to some of all the other things we've talked about. It is our job to allocate those tokens much more efficiently to solve the specific problem for our users and use them in really intelligent ways. And we should be able to do that better than somebody just talking on their $200 a month plan to a base model and the base chat interface. And like, that is our job to use those tokens really effectively for the users with the problems. Like the gap gets large. Yes, that is a problem. There being a gap, I don't think is fundamentally a problem.

Cool.

Thank you.

I don't know where exactly I land on this either, but I do feel like the— I've been feeling like, oh, this AI is pretty cheap. At the same time that I hear my founder friends say like the costs are getting out of control and I'm like, oh, wait a second. Yeah. The ratio that I'm enjoying right now is like Huge. And it does seem like it's going to be some sort of normalization. Feels like it's probably healthy for the overall ecosystem just to avoid everything kind of collapsing into the foundation model companies.

This is also probably why founders have to root for open source to keep pushing frontier, right? That means that they just don't have the ability as much to make that gap gigantic. If the gap between frontier and open source is huge, they have all the power in the world to keep that gap, whatever they want it to be. And if open source becomes close to the frontier as possible, the less pricing power they have to make that gap gigantic. But I think application layer companies who are focused on specific problems, we have so many levers to fight against this. And it is simply just like, this is the core task that we all are doing, is try to solve the specific problem for our users in a really high leverage way. And if we can really do that, you know, we will find ways around different pricing structures and markets.

No excuses, play like a champion.

You're right, Bill.

Fantastic. Prakash, anything else you want to jump in with real quick? Or Eric, any other concluding thoughts before we let you go?

No, this has been great, Nathan. Really appreciate having me on.

Thanks for joining us. Great to meet you.

Have a great rest of your day, guys.

And speaking of decentralizing, we have an excellent next guest who is— I'm going to introduce Tricia Martinez.

Hello.

Tricia is the founder and CEO of Dapple, an AI infrastructure company that is fundamentally rethinking where and how enterprises run their most critical models. Historically, if a company wanted to run heavy AI workloads, they faced a difficult binary choice: either rent shared space from massive public clouds like AWS or Azure, or spend years and billions of dollars trying to build their own private data centers. Tricia and her team are introducing a third path. Dapple builds dedicated single-tenant in-country AI clouds. That means a bank a hospital, or a government agency gets its own isolated physical hardware governed strictly by local laws but operated with the ease and interface of modern cloud software. Tricia brings a unique perspective to this highly technical problem. She previously helped launch one of Africa's first digital banks and mentored startups at Techstars, giving her a deep firsthand understanding of what heavily regulated industries actually require to adopt new disruptive technology. She is joined at Dapple by COO Salam Al-Masrawi, whose engineering and growth teams have historically deployed over 300,000 AI accelerators globally. Just this month, Dapple announced a $30 million seed round backed by the Raptor Group and Ion Pacific and revealed they have secured over $100 million in enterprise contracts in just 5 months of operation. At a moment when global regulators are demanding sovereign AI, Tricia is building the exact physical and software infrastructure to make it a reality. Tricia, welcome to the show.

Hey, Prakash.

Hi, Nathan. Thanks so much for having me. And thanks for the intro.

Welcome and great to meet you.

Welcome. Thank you for having me. All right, what do you guys want to talk about?

Who's so eager to spend this money with you that you've got to $100 million in 5 months?

You guys, it's, it's, I'm not kidding you, it's madness. Okay, so where to begin? We, I like to start the story by telling everyone about what everyone says, is there an AI bubble? Everyone's fearful that the enterprise isn't jumping in there. They're not utilizing AI as fast as we think. Is it just the digital natives, just the hyperscalers? Is this really a bubble? We're overinvesting in infrastructure and GPUs. This is a scary time in the market. And I will tell you, that's wrong. The reality is the enterprise and the government and some digital natives for that matter have been sitting on the sidelines. Why? Because The public cloud doesn't have capacity for them. They're growing faster than ever. They are the largest consumers of compute. They need it for themselves. The only options that these enterprises have, like you already had mentioned, Prakash, is to go to a public cloud, which they already operate their entire infrastructure in, but these guys don't have the capacity for them. So they can then go to a NeoCloud who offers bare metal, which requires the enterprise to then go build out their own systems and operations, their own software, operate separate clouds from the public and this now new cloud that they have to live in. Or the other option is to build on-prem, and nobody wants to build on-prem. It's costly, it's complex, you have to hire a massive team of infrastructure engineers. It doesn't make sense. So we, my co-founder and I, Salam, came together. He has deep expertise within this GPU space. As you had said, you know, him and his team have built some NeoClouds. They've, you know, managed and deployed over 300,000 GPUs. I came from a background of startups investing. I also was a White House Fellow under the first Trump and Biden administration serving at the Department of Energy. So I worked across our 17 national labs, specifically building an AI and energy strategy for the nation. And we both said, look, they're— the enterprise is being sidelined. No one is servicing them right now. No one knows how to service them. That's where the opportunity is. So we believe at Dapple that AI has fundamentally changed infrastructure. So for the last 20 years, enterprises could run applications almost anywhere, and now AI has changed that. So today, where your compute sits, who controls it, who— how it's governed, and how quickly it can scale all matter. That's why we've built Dapple, and we call it the enterprise OS cloud. It's a new operating layer that helps enterprises deploy and operate production AI across dedicated infrastructure. So I think that the simplest way to think about Dapple is, like I said, this operating system for production AI. We're not building foundation models, we're not building GPUs, we just make it possible for the enterprise to deploy AI reliably, securely and at global scale. And AI infrastructure, as we all know, is becoming increasingly complex. So our job then is just to make it, make that complexity disappear for the enterprise. But we've pulled in through our network, through strategic partners, incredible amounts of demand. Like I said, the enterprises have been sitting waiting for years for capacity, literally years. And now that we have the ability to service them, we're growing faster than you can imagine. So yes, the first few months of us going live, we ended up securing over $100 million in customer contracts. And the reality is, you know, you guys are going to have to have me on the show again because I'll have some very, very big updates on our growth in just a few months that you guys are going to be wowed by. So I don't want to, I don't want to— say anything too early, but what we are seeing is just mind-blowing.

Let me take a step back here. A lot of your technical marketing emphasizes that you are heavily Azure native and that you integrate seamlessly into the Azure Machine Learning control plane and Azure Arc and Azure Monitor. How does this work? Are you guys like Transitioning customers from the Azure cloud to your, to your cloud? Does your cloud sit on top of Azure? Like, how is— how does this interaction between the two work?

So that's a good question. And you're going to have me back on your show in a few months once we've made some official announcements about some of our strategic partners. But basically what we've said is the enterprise lives in the public cloud, but the public cloud across all 3 public clouds are struggling to deliver capacity to their strategic customers. Why? Because they're consuming that compute for themselves, or they are the largest producers of AI and need all that capacity. So what we saw is a really unique opportunity. Enterprises are sitting on the sideline, they need capacity, they need a unified experience. No enterprise wants to build separate clouds like they would have to with a Neo Cloud offering. What we want to do is create this unified, seamless experience. It's frictionless, which is imperative for the enterprise because they can't go build out an entirely new software engineering team to go manage the clouds. So pulling together the public and private cloud is the IP that we've been building. And we are doing this with strategic partners, which is enabling rapid growth for us. But when I come back shortly to have another conversation with you, we'll make those announcements.

Another number that jumped out at me from your website is the utilization number, 91 to 94% GPU utilization, it says on the infrastructure page. What does that mean? Like, you know, I'm trying to make sure I use at least 91 to 94% of my CloudMax plan every month. I feel like if I, if I fall short of that, I'm leaving something on the table. But if I had dedicated hardware, I would have a hard time keeping it running 90+% of the time, even just me, right?

Yes.

Now, at an enterprise scale, you've got all these coordination challenges. So am I understanding that claim right? And does this mean enterprises are sort of running a ton of batch jobs overnight and are you helping them orchestrate that? This is a surprisingly, for a single tenant, Infra play, that's a surprisingly high number from my naive perspective.

No, I think the reality is, is that enterprise AI is very different from the traditional cloud model. So a meaningful portion of our deployments are backed by long-term enterprise contracts. So we're optimizing around committed capacity and customer outcomes rather than chasing that spot utilization. So it's a much more predictable business model for us. So we're building Contract or these partnerships with the enterprise who are committing to 3, 5, 7-year deployments with us. And the reality is the enterprise is a sticky customer. Once they commit to something, it's unlikely that they want to leave, right?

So what kind of workloads are those though? Like, so, you know, so much of what I do on an AI day-to-day basis is like, it is kind of spot, it is sort of unpredictable. The workloads would have to be a lot better characterized than mine are, which makes sense for enterprise. But still, I'm kind of wondering what sorts of things are companies running overnight? And especially also, if you have that kind of high utilization from the beginning, are you baking into contracts like ramped capacity? Because that doesn't give them a lot of room to experiment with ad hoc new use cases, right? If they're already at, say, 94%. percent capacity, where do they go to explore new stuff?

So I think what's unique about what we're doing is it's about the outcome. So the way that the market thinks now is GPU per hour. It's the price. And for us, it's about the outcome. That's what the enterprise cares about. So we're seeing basically 3 primary categories and every customer is different, right? We have unique use cases. We're servicing the enterprise, there are government use cases, there are digital native use cases, you know, the companies that live in AI and that are significantly larger than the startup or scale-up. But the primary use cases or how we think about it is, from a workload perspective, large-scale model training, as you would imagine. So AI-native companies need this dedicated GPU infrastructure to train their next generation of foundation or domain-specific models, if you will. The second piece would be inference, which is where we're seeing the enterprise go, obviously. That's at production scale. So enterprises are deploying AI applications that need predictable performance, low latency, and strong governance. And lastly, I would say AI agents and enterprise automation. So organizations are embedding AI into their internal operations and these customer-facing products. Those are the things that we're seeing. And it varies across industry, right? We're, we, you know, we originally started focusing on the mission-critical workload. So you can think finance, cybersecurity, government, defense-related operations. These require the private, oftentimes quote unquote sovereign deployments you want fully isolated. But in reality, the market is by far larger than we ever anticipated, that the entire Fortune 1000 market is in serious need of capacity. And while not all of them are mission critical, they are trying to build on inference, some, you know, not as much training, largely inference, but they all are in serious demand. And it's largely because they have been sitting on the sidelines. So there's a time urgency right now. In the future, what we will see is that most of these enterprises care a lot about specific compliance sovereignty-related issues, but for most of them, they're just desperate to get some sort of capacity so they can start building their models or deploying these AI systems themselves.

So, Tricia, I think I've seen that you said that you can get a deployment up and running in 6 to 9 months, a very, very fast-paced deployment. 6 to 9 months to me doesn't seem like it's long enough to actually build a physical data center because Elon, I think Elon has like the all-time record of like 15 months or something to go from groundbreaking to operation.

Yeah.

So in that 6 to 9 months, what exactly are you doing? Are you leasing, renting capacity from existing NeoClouds and then putting your stack of software and tools on top of that and then plugging that into customers who are already Azure native and, you know, having them move over? Like, is that the frame of what happens in that 6 to 9 months?

Yes, and that's a great question. The reality is we can't do everything. This market is so big and honestly the demand we're getting is so incredibly massive. So we are creating a network, we are orchestrating that network. Our model is to partner with world-class data center operators and infrastructure providers and then bring together the capital, the GPU infrastructure, the enterprise customers, and that software operating layer that turns those facilities facilities into AI-ready infra. But in some cases, we're deploying infrastructure that we help finance, and in other cases, we'll also deploy on partner infrastructure. So yes, part of our strategy is going to be building out campuses. We need to have capacity available before the demand hits. But for us, what's important to do is whether we own those GPUs on our balance sheet or whether we are financing that data center and building it ourselves. We have third-party partners that we rely on that we can support in various data center operations and/or on the GPU deployment side as well.

Indeed.

I have a question on the overall financial towers that we are building right now in the AI space. It sounds like your customers are making long-term commitments to you, which presumably insulates you from some of the riskier dynamics that could emerge.

Yep.

But how much do you see in the space as a whole that is sort of this long-term commitment, short-term revenue that creates the potential for somebody has a bad quarter and all of a sudden there's kind of a cascading problem. We've obviously seen that pattern play out in previous financial crisis moments, and I'm not sure how much of that is going on today or how likely it is for some external shock to be big enough to throw a a wrench into the system that really kind of could spin part of it off of its axis. What do you see in that regard?

That's a really, really good question because the market has changed drastically since we jumped in. So I would say we're seeing enormous amounts of capital still flowing into AI infrastructure, but investors have become much more disciplined over the last 12 to 18 months. So the market's definitely moving away from hey, build it and everyone's going to come, towards infrastructure that has to be backed by real demand and long-term customer contracts. This is what became so exciting for us as we saw this big gap in the market. This model that we're building is significantly more de-risked than the current existing model. The CSPs, the NeoClouds, are servicing those hyperscaler deals that are massive in nature. But in reality, you know, they— there's not that many of those deals that you can come by. Or you're taking geopolitical risks by working with different customers, off-takers within the market. And now looking at the regulatory restrictions, right, like people are getting scared, the financing groups are getting scared that the existing Trump administration is going to start cracking down on certain countries or who's consuming So that puts a lot of those CSP operators at risk, to be honest. So our approach, you know, in our mind is very attractive, and it is attractive to the financing partners we work with. These are repeatable customers. These are long-term contracts. Our customers are putting large down payments up front. It enables us to do really interesting things, to go finance those GPUs, to go finance actual data center. These are customers who are sticky. They end up repeating with you over time because it's easier to repeat versus finding a new provider. This is why the enterprise is so attractive. People have been fearful, like I said, are they sitting on the sideline? Again, the reality is no one has built a specific offering that appeals to the enterprise and their problems that exist within this market.

Yes.

The financial market is changing. It's becoming more risk-averse because so many neo clouds and CSPs have taken bad deals. And now those financing partners are bearing the burden of that. But our model is a different approach and it's more attractive to the financial markets, I would say.

It strikes me that the Dapple model is really much more asset-light than I think the existing neocloud model. Would you, would you say that's correct? And is that an advantage?

Absolutely. So our approach has been, we cannot possibly do everything in this market. And I think most of these neoclouds and CSPs are trying to do everything. They're trying to build all the data centers. They're trying to own all the GPUs. They're trying to build up all the demand. We came into this saying we need strategic partners. We need to align DC operators. We need to align unique financing partnerships with various debt providers. We need to align demand with strategic partners. Like, we cannot possibly own everything. Where do we, where do we really thrive, and where can we work with partners to help us build this greater ecosystem, which we believe is the next wave of infrastructure? That's the approach we've taken. And so, like I've said, there are scenarios in which we do want to build out our own own DC campuses and take on that infrastructure. There are scenarios in which we are deploying and owning our own GPUs on our balance sheet. But there are also scenarios in which we're working with partners to deliver value to our customers, because at the end of the day, what we care about is the problem our customer is dealing with and delivering them in a timely manner and the value that they deserve. That's the different approach. So it is less asset intensive compared to some of these You know, other players in the market. And I think that's where a lot of the risk is. It's just a different approach. And we believe that being a software-first company is where we'll end up winning.

Absolutely. How do you think—

so something that was just trending yesterday was this notion of the software factory. This was from Warp. And it's, you know, been popping up more and more, right? And basically the message, it's one of these kind of public memos from the CEO where the guy says, our job now is to build the factory that builds the product. And it's a software factory, so it's the sort of meta software notion. I'm kind of looking at that in a few different places and I'm like, is it one too many levels of abstraction perhaps at some point? Or Is it just what the doctor ordered? I'm not quite clear on that. But when you think about kind of how to create a defensible position over a period of years, obviously one of the challenges you're going to have is that you could just go to the coding agents and say, hey, set me something up on Azure, right? If Azure's too complicated, at some point the coding agents will simplify it, substantially. What do you think your kind of form factor is long-term? How do you envision an 18 months plus from now world where my coding agent is pretty able to navigate even the— and I do understand it is quite Byzantine complexity of the clouds as they exist today. So I can get that and my friction is presumably less, barriers to entry are less, What is then your form factor that remains like a step above what I can imagine getting directly from the hyperscalers?

So you mentioned like defensibility. So I'm thinking about it from that context. I don't think there's a single moat in AI infra. I think that the defensibility is gonna come from a combination of several things that will reinforce each other. So one is enterprise relationships. So ensuring, like, having demand in the market is key. And if you don't have demand, then you have nothing to sell. So we are— we have built strategic partnerships and we have doubled down on enterprise relationships. So once an enterprise deploys mission-critical AI infrastructure, switching becomes difficult, like I said. So they're not just moving GPUs, we're moving this governance, the networking, the compliance, the security, the operational workflows. Enterprise doesn't want to move on that. The second piece that I would say is probably the operating platform. So every deployment makes our software smarter, in reality. We learn how to optimize placement, the capacity planning, performance, cost, policy across all these different environments. So over time, that becomes increasingly valuable. A third piece would be the ecosystem, and I think that's probably one of the most important. We're building relationships across infrastructure providers, capital providers, silicon partners, enterprise customers. That's difficult to recreate all of those simultaneously. So ultimately our moat isn't owning the most GPUs. It's becoming that trusted operating layer that enterprises build their AI strategy around. And I will say this, like AWS created the cloud for the enterprise. They abstracted everything, right? Every enterprise was building on-prem. They had their own data centers. They had to figure out how to move all these clunky things and put everything together in a very complex ecosystem. This is now the new wave of the public cloud. AI infrastructure, or AI is demanding something entirely new from an infrastructure perspective for the enterprise. And once again, the enterprise doesn't want to have to go through and deal with the financing, deal with the ownership of GPUs, deal with these depreciating assets, build the data center, manage relationships. Like, we are abstracting everything similar to how an AWS was, but we're doing it in this new wave of AI infrastructure. So I think it's the, the defensibility is much larger, and that's why we keep thinking about our role as almost like we're We're orchestrating this ecosystem. There are so many pieces that you have to move for this to work exceptionally well, and that's what our focus is on right now.

When your customers evaluate their relationship, what do you think are the, you know, 2 or 3 most important numerical KPIs that they're looking at? Is it like a Performance for watt? Is it like TFLOPS per watt? Is it, um, you know, what are the KPIs that they're going to be looking at on a weekly or monthly basis?

Yeah, so I would say it's— I mean, there's definitely technical components to it. However, what matters the most to the enterprise, to all the customers we're working with, are a few things. One, the time to deployment. Everyone Time is of the essence right now. Everyone needs it to get something live, so time is really, really critical. I would also say the availability and reliability. So can they rely on the infrastructure for mission-critical workloads? If they're doing drug discovery, can it be in-country, isolated workloads? Consistent performance is essential for them. Next would be scalability. So as their AI usage grows, can we scale with them? Maybe they've only purchased 1,000 V200s, but you know, in two years from now they're going to want to scale up to 10,000. Governance and compliance—that's probably one of the most important things that no one really can to no one can support them in the market right now. So can they meet their regulatory, their security, their data residency requirements while operating their AI? And then lastly, of course, which all. CFOs of every one of these enterprises cares about is the economics. So are they getting predictable economics over the life of this deployment? So again, like I mentioned, it's not just about the GPU per price. They care about the outcome, but of course the economics are critical in it.

Last one for me. Based on your history in Africa, I wonder if you have a take on how broadly this technology is diffusing and how accessible it is to companies in Africa. Are they able to buy in? Are they priced out at this point? What are we going to see in terms of the financial companies, the telcos, of Africa and their ability to keep up with what the even bigger and presumably deeper pocketed and more readily financed companies in the richer countries are doing?

Oh, that's, I mean, that's a really good question. And it's, to me, that's, those are the things that I care the most about. Like, I care about building technology that matters and using technology to solve actual problems in our world. So I would say that most enterprises have proven that AI works, like we know it works. The challenge now is moving from experimentation into production. That's the transition we're moving into right now. The reality is though, we see demand and we're getting a lot of demand from our strategic partners. We're seeing demand Coming largely from APAC and the U.S., followed by Europe. We're not seeing that across South America, across Africa, some in the Middle East. But the reality is the more the more industrialized countries are investing aggressively in AI, and that means that Africa as a continent and every country, every 54 countries there, is probably falling behind on that technology production side. And the cost is prohibitive, just like many of the other technology jumps we've seen happen outside of AI, if we want to talk about blockchain, for example. So it is definitely a problem and something to be thinking about. We do work with governments, and governments are trying to step up more. But the reality is these things are really expensive. It's so expensive to deploy AI, right? With the new Vera Rubins coming out, like, costs are prohibitive. Not many companies can afford to do these things. And so that is a barrier for a lot of companies, a lot of countries and nations. And, you know, I can't go into more detail than that.

I think the reality is though, like, that's the truth and that's something we all When you talk to enterprise AI buyers, what do you think are the major mistakes that they are making right now when they make these purchasing decisions?

I mean, that's a good question. So I would say maybe most enterprises aren't making technology mistakes. They're making operating model mistakes. they're probably treating AI like it could be another software project. So the biggest mistake is assuming AI can be deployed like traditional software. So AI, as you know, is fundamentally an infrastructure problem. It changes how you think about the compute, the data, the governance, security, and the operations. So in certain scenarios, there's more handholding on our end to really educate or ensure that the enterprise is making the right decisions. I would also say optimizing for today's use case instead of tomorrow's scale. So many organizations build a proof of concept that works for one team, but they haven't thought about what happens when every business unit wants AI. So the architecture that works for 100 users often isn't going to work for 100,000.

Mm-hmm.

I think those are a few of the things. You know, they don't— the enterprise doesn't move nearly as fast as the digital native, the AI-native company whose entire business is driven by AI and technology. So it's a very different model, very different approach. They can't move as quickly. That can be a struggle as well. And we know in this market, if you don't move fast, the capacity is gone. And so these are calculated risks. It takes longer to sell. And there's more thought that has to go into what's being purchased and why it's being purchased. But again, we are, those are things we are learning as we continue to sell and work with the enterprise. And we only will improve our product offering through all of those data points that we continue to collect.

Maybe one more, if you have an extra second. How much would you characterize your activity as being that of a market maker versus that of an agent representing your enterprise customers in the market?

I mean, I would probably say neither. We're an operating platform, so we're, we're not a market maker. You know, we don't think of ourselves as a broker or an intermediary at all. We work with enterprises to understand their AI strategy and operational requirements and then orchestrate the right infrastructure to deliver outcome. What's really important to emphasize though is no matter if, like I said, no matter if a third party of ours is running the data center or the deployment or we are ourselves, the DAPL software sits across every single deployment and transaction that we have. So that is the most important component. We are building this unified public-private offering that's frictionless and seamless for the enterprise. That's taking into account their compliance and regulatory needs. That is the consistent piece across everything that we do with an enterprise customer is our software. So they're not buying access to a marketplace, they're buying a production AI platform with us.

Thank you.

Because I think you are deeply embedded in the sales process, which is happening in the industry right now. And it's very difficult for us as external observers to actually understand what buyers and the new clouds are doing right now and how the deals are moving. So this is very interesting. One question I do have, one last question, how often do you see pricing move while the deal is still in flux? Do you see pricing move a lot?

Yeah, I mean, so I think the, the upside of this market is that sales cycles literally cannot be 6 months with an enterprise because the capacity is gone. So we have a forcing function, you know, you have, this is the price right now, this is the location, you want it, take it or don't, because tomorrow it's probably gone. So we have this constant movement with our customers. Which at times is difficult for the enterprise. But the reality is everyone wants capacity right now, and so they're willing to move at a much faster pace. Sure, pricing changes, but once we deliver an offering to our customer, we generally are honoring that even if pricing changes. But again, if we're doing deployment, we're buying at a certain price and we honor that price with our customers.

Thank you, Tricia. Thank you. Thank you for your, for your time. And we hope to have you back again. again as the company expands and takes over the control plane of AI infrastructure.

Wonderful. Thank you so much for having me.

Indeed. Looking forward to it.

Awesome. Thanks, guys. Have a good one. Bye.

You too. Your comment there at the end is really apt. It is very hard to know what's going on in these. It's like all over the counter, right? I mean, this is a Another kind of— I don't want to over-analogize to the 2000s mortgage business, but a lot of over-the-counter dealing is certainly one commonality between that market and this market. You know more about this kind of stuff and all the financing that goes on than I do, but what do you make of the her answer to the market maker versus agent question seemed to be in some tension with the subsequent comments around facilitating these deals and price is here today, it's gone tomorrow. I don't know what a market maker is. If that's not a market maker, it sounded fairly market maker-like to me. How would you— parse that mix of answers?

So there is a— there's 2 ways to look at it. And one is, yeah, it's pure market making. The second way is that it is actually quite difficult to deploy the control plane on bare metal on the new clouds right now. And this is something that Yi Tei, who is now at Google, he was at Google and then he left and then he founded his own company and then he wrote a fantastic blog post and then he went back to Google. And what he found was that every data center has its own specific quirks on how the GPUs are run. And it can be as simple as what kind of wiring did they use? Is the power steady enough or is there some kind of fluctuation? Do you get more frequent GPU failures? Even the same GPU, different data center behaves differently. And so what ends up happening is that Different data centers have completely different quality of what you can do on the data center itself. And it's not consistent. And you end up needing like Yi Tei's team, which is a team that manages the infra in order for you to be able to actually use the data center effectively. And obviously these people are in short supply right now. And if you So you only have 2 choices at this point. You use the vendor, the hyperscaler vendor or the new cloud vendor's control plane. And you are basically enslaved to that because that control plane, your workload on that data center doesn't work the same way that it does on another. And if you move to another, you have to deploy again and you have to figure out all these infra details again on which GPUs work well and how to manage the power fluctuations and all of this stuff. And it becomes this very artisanal deployment process. And if you can build a control plane that can manage consistently, no matter where you deploy, and then you have the ability to kind of move the workload when you need to, then you have— and maybe it's less efficient, right? Maybe you start off with like, okay, you know what? it's going to be 5% less efficient than the best efficiency anyway, but we'll live with it because now we have this flexibility of being able to move the workload. Maybe that makes sense, right? So I think the more sophisticated answer is that if you can kind of build that out and then you can say, even if you need to, you sign with us for 1,000 GPUs and as you scale, we can add in different locations and we have the same control plane and you have the same behavior of the GPUs and you don't need your own infra team, that's valuable, I think. And so if that's what you can offer, that is actually something that makes sense. It's what Nebius, for example, is trying to build out. It's what CoreWeave is kind of 50/50 on. CoreWeave is really offering bare metal servers that you can bring your own infra team into. Guys are trying to build out this control plane thing and like trying to offer it as an add-on service. And so if that's what they're doing and they have that software, technical software that they're building out, they're basically doing 2 things. One is doing the market making and broking, and 2 is managing the infra underneath that market making and broking so that the enterprise doesn't have to worry about that infra team. So that's what they're doing. That's actually pretty valuable. It's basically lock-in. You get locked into the control plane because once you have everything running, to go into another control plane and then reconfigure again for that, it's just like, I need to hire the infra team again. I need to spend the year of fine-tuning again. And then I need to figure out when the sunspots are going to occur, when the solar flares are going to occur, and when all this dumb stuff occurs, and when these bits flip. No CEO wants to— no bank CEO wants to deal with it, right? I don't want to talk about sunspots, right? I want to talk about interest rates. So yeah.

I remember that blog post that you alluded to, and one anecdote that stood out in that to me was he said that there was one time he left a training job running And kind of forgot about it at Google and then came back a couple weeks later and it was still running. And the contrast between that and what you could do on any of the Neo clouds was just like night and day. You just, something's going to go wrong on the Neo clouds. Now that was 2024 when that blog post came out. So I think an obvious question would be on the one side, like, shouldn't the Neo clouds have shored things up quite a bit by now to the point where it's not such a mess? I mean, maybe they— some of them can get by without doing it because there's just such high demand that they can sell even if their product is kind of crap. But sure seems like they should have had plenty of time to solve a lot of those problems.

So what ends up happening is that the new chips come out, so they stabilize on the H100s. H100s are very stable now. The V200s, you got to go through that optimization again. Right? So the problem is like people are building on the leading edge, and on the leading edge, everything is on the leading edge, right? Like, so the chips are, you know, less, less reliable when they start out because, you know, a lot of this stuff hasn't been figured out. And people plug these things in and then they're like, why am I failing? Then they go back to NVIDIA and NVIDIA is like, oh, okay, we'll do a fix. And then they'll fix overnight and then deploy to all the GPUs again. So there's this like, that's why there's this like group activity of discovering the capabilities and pitfalls of using the chips. And the more people are working on like figuring it out, the better, you know, the experience is for everyone. And it stabilizes over 2 or 3 years. So H100s actually become— actually became really stable over time. But then the B200s came out and then people like really complained about the B200s. And now the B300s are coming out and some of the B200 problems were hardware problems. So they're hoping that B300s fix those hardware problems and it's not just when you do work. And what you need is auto-fix all the problems on the chip. And maybe you get there. Maybe you get there in the next 18 months where you get auto-fixing of all of these problems.

It's still a little hard for me to wrap my head around how a company like Dapple fixes that faster or better than the NeoClouds do for themselves, though, which then leaves me back in the spot of like, it feels more like a market somehow at core. I mean, so many of the things, speed to launch, right, was like the number one KPI, it seemed like. There was a lot of KPIs that she was discussing that were sort of the kinds of KPIs that would go into your overall analysis of AI. But in terms of why buy from her specifically, it seemed like speed to live was a huge one. And that is a valuable thing to be able to provide today, right? What would you guess are margins for a company like Dappel? And would you say it's like, how would you compare it to the 6% real estate agent take on a traditional residential real estate transaction?

So I will say one thing that it's not a, it's really a regulatory issue. She's going to all regulated industries, banking, et cetera. And in these regulated industries, often you hire an external vendor because you can hand off liability to them. And so what she may be doing is they have the infra team, and the infra team gets updates from the new clouds or the chip provider on what you need to deploy. And every time you deploy, the CTO of the bank might have to sign off on it. And the CTO doesn't want to sign off on— CTO doesn't have an infra team to evaluate this. So if you can share the infra team between multiple banks, then everyone gets a saving because the— you have one infra team that's evaluating all the fixes and all the updates, et cetera. And then they're like, all right, it's okay to sign off. And then all the CTOs at all the banks that are, you know, using Dapple can basically sign off together. And so that's— that you're basically sharing the cost between, you know, multiple, multiple entities. And so that is valuable. So I think that, that might be something that she can offer in terms of like being able to like, Because the bank will be like, okay, they're not going to upgrade overnight, right? They're not going to like, okay, have NVIDIA push the update directly into their cloud. And they don't want to like use someone else's cloud. And so you're caught in all of this, like, I have to reinvent the wheel because I want to keep all of the sovereign thing, blah, blah. So I think she could actually do a bunch of that where they evaluate updates and can update the GPUs and the software, can provide the CTO with like the necessary liability shield in order for them to like sign off. Margins. I think eventually the margins will be pretty significant because in finance, IT providers to finance often have this thing where they go in at cost and they fix the price at that cost price. But then over time, the technology itself becomes cheaper. And so you can go in and you can say like, okay, I'm going to charge you like $100 an hour, but then the GPU price in like 2 years is going to be 30% of the $100 an hour. And you can make that cost advantage. So some of the things that people do are on this idea of locking in regulated firms at basically cost price now, over time. That might be something that they can do. Like for your first customer, you charge them what your infra team costs, but your next customer, your infra team is the same infra team. So basically all of that is margin, right? So it could be how they structure contracts.

Well, I wouldn't bet on those GPU prices falling, but they can at least hope to amortize their cost of platform development at a minimum. Yeah. I'm always confused by companies like this in general. I mean, I was just looking at Snowflake's market cap, which is currently $75 billion, $78 billion, and Databricks private, but over $100 billion. And I kind of squint at that and I'm like, really? For a sort of software layer that sits on top of these Clouds and does what exactly? Like sort of manages data that still ultimately is sitting on AWS hardware somewhere. It feels like a lot, but I guess this just goes to show my naivete broadly around like just how much friction exists in the enterprise, how much switching costs there is, how much they're willing to pay. In terms of a premium for predictability or to hand off liability. I would, I mean, those will be interesting possible canary in the coal mine companies. If all this goes as I sort of think it will, I do think you're just going to get really good cloud management from your core intelligence provider, right? I mean, how is it not going to go that way? Like, I already get In setting up my own personal AI infrastructure, I have had multiple moments where I'll be like, oh, I need to set what up? You know, I have to enable what in my Google Cloud what project? And the ability of Claude already to give me a deep link to exactly the place that I need to go click into in Google Cloud to hit the— and then it'll be like, go to exactly this URL and hit enable. And then I'm like, okay, go boom, click enable. Okay, it worked great. How long would it have taken me to find that in the absence of the deep link that Claude so helpfully spits out? You know, like in some cases. So it really does seem like disintermediation should be possible. And I guess we'll have to wait and see just how easy it has to get before it's worth it to these companies. And I guess you could also maybe tell a story where maybe one interesting and sort of possible salvation for the app layer story would be everybody's just so overwhelmed with everything that's happening and they're so focused on applying AI to their core business and just everything becomes so existential and everything becomes so survival critical that you just don't bother swapping vendors on anything. If you're a big bank or you're a big pharma company or whatever, you're like, sure, we could save money maybe on our Snowflake or our Dapple or whatever. And we could probably potentially even get that ROI pretty quick if we put a couple brains on it and a bunch of clods and kind of figured it out. But the AIs are coming right for the core of our business. So we don't have the luxury of, and it just doesn't make sense to try to save money on this margin, that margin, because we need every bit of brainpower we have going to defend the core from the data center or the country of geniuses in a data center. It's funny, again, I can always kind of seemingly talk myself into a lot of different stories around how this all goes, but that one's interesting. I usually think that the faster things go and the more the intelligence continues to progress along the curve, And the longer the tasks and the more reliable the autonomous execution, the worse it's going to be for the app layer. But there is this other story of if that goes fast enough and goes to the heart of people's businesses enough, then there's just no time to even think about rational— don't matter.

So the question I have is that, is there a difference between software that has an interface to the human versus software that has an interface only to agents? And is it that Snowflake or Databricks has value right now because it provides an interface for humans to interact with these cloud systems? And if humans didn't have to interface, and if agents were directly interfacing, and no human ever touched Snowflakes or Databricks at all, so that those names were not known to humans at all, what kind of margins would Snowflakes or Databricks be able to charge? And what is the value that they're providing? And so you don't even need to know the API key. Your agent already has that access and does whatever it needs to do with those systems. Like, what value do those systems actually provide? And to an agent who has, who, you know, will go through the tedium of signing up for 4 different things and looking at, you know, 10 different lists and reading your docs and reading the API and doing whatever's necessary to get it done, and is not gonna be particularly find like usability a thing that they're bothered with and they're more than happy to use unusable, relatively unusable, non-user-friendly code, which is more efficient. What value do these things actually have? And that question for me, I kind of imagine this kind of future where agents are 90 or 99% of software is used by agents. And so the agents themselves are the major buyers and will have to do the price discrimination of buy versus build decisions on all of this software. What are the agents willing to pay for?

I mean, it would seem like it's going to be tough for margins in general, right? The drop in search costs, the drop in switching costs, it all sort of would seem to lead us much closer to perfect market conditions. And as we know, in a perfect market environment, not a lot of profits, right? So seems tough.

And if you see this kind of hyperdeflation, right, where the price of everything starts to plummet, price of all of the software stuff starts to plummet, And you think of a business like, let's say Google, which is offering an interface to customers, interface to humans who are watching the ads. And the price, the cost of producing those products, like let's say in Intuit, is now essentially asymptotically approaching zero. Intuit then has this problem of like, okay, then if I'm going to sell, if I'm going to maintain my margins, I have to advertise. And you're in this kind of Red Queen race between multiple advertisers on YouTube. And so I wonder if actually that becomes this kind of dynamic where all of the profits end up, all of the revenue goes to the people who are basically offering these Red Queen races, zero-sum games that you have to play in order to access customers and access the kind of end buyers who are able to make the decision.

I think that's been a reality on both Google and Facebook for years, and it has in some niches, I think, been quite consuming. I've seen examples in the digital marketing space where people that are selling digital marketing services are barely making any money once you account for their cost of customer acquisition on the same platform that they're now going to layer services on top of. So, yeah, I think advertising in general is— I mean, it could go multiple ways, of course, but I think there's a pretty good case that advertising will continue to be expensive and really valuable. Because human time and attention remains, you know, the, the kind of most scarce thing. Now, advertising to agents might have somewhat different dynamics. Um, it could very easily, you know, we may, we may start to see some like different ways of cutting up the money also. I could see scenarios where the agent, you know, the— you might be happy for your agent to consume all kinds of promotional materials or product specs or what have you, especially if the advertiser is willing to pay for your inference cost to do the consideration of their product. So I could see some different models opening up there that could be quite different and could lead to different equilibria. But I still think for a lot of products, People are gonna wanna make kind of a final call. And so there's, there does seem to be like, there's gonna be some market for getting in front of people, consuming a small amount of their time and attention. And probably the cost of that goes up until maybe some point where you trust your agent enough. And then it's like, now if you're really gonna just let the agent make calls for you, Maybe that becomes a very different equilibrium where this sort of subsidised inference or paid consideration or something like that becomes a kind of new equilibrium, new mode of doing things. That's hard to predict. I don't think people are going to be that quick to let the AI decide where they But yeah, I'm gonna have a decision maker in the end. I'm still generally bullish on the value of human time. Feels like that's gonna go up for a while at least.

I often seen these— I've seen these ideas about, you know, your AI will help you book a vacation. And it's always struck me that I've spent a lot more time Considering many vacation options than I've actually gone on vacation. For sure. For sure. And so it strikes me that a lot of the vacation planning process is actually just daydreaming and fantasy. And you're actually engaging in this process of daydreaming because it's fun. And you're not actually trying to plan a vacation. You're trying to dream a little bit.

Right.

If you had the time and if you had the money and if you could take, if you didn't have your— if you could take your kids out of school and go trekking in Machu Picchu or whatever, you kind of look at it, right? But it doesn't— out of every 99 vacations that I've daydreamed about, maybe one becomes true. And I think to a large extent, the guys who are focusing on AI will plan your vacations, are focusing on the wrong thing. They're focusing on this idea that AI is going to help you do this more efficiently, when what they really need to do is try to figure out how the AI is going to help you daydream better. Because the daydreaming part is what I think drives vacation planning, Zillow, even Tinder, even Tinder. Again, all of these things where you actually spend more time Doing so much more search than you actually spend executing. These are all actually forms of daydreaming. And I think maybe the solution will actually end up being VR daydreams, vacation sites that you can just have the AI generate kind of like the entire kind of like what happens on the highlight reel of your vacation. And you can see the highlight reel with you in it. And that's super fun. And it's not really going to be an AI that actually plans a vacation for you because that's not really what you're looking for. The intent is actually not to book a vacation.

Yeah, I think that's apt. I think the daydreaming— I've seen even some, I don't know, sort of welfare economics research a long time ago that suggested that The utility of a vacation is substantially in the anticipation and in the sort of memory of it. So that rings true. And in my recent road trip that I did with the kids, which I used AI heavily to plan, it did have a lot of that kind of character to it. I was kind of It was very interactive, I would say, was the mode that actually worked for me. Going off and saying plan everything would not have worked. Give me a ton of options and let me kind of pick some highlights. Then you go, you know, find more options based on what I said there that I liked. That worked pretty well. And that actually led me to something that I was very excited about going into it. And actually worked out really well. So there is definitely value there, but I agree, you can't fully externalize that and capture the full value. So there is maybe a form factor mismatch in some of the ways people talk about this.

Nathan, another interesting morning on AIM. I am eager to see how the markets have performed. Just to give you a preview, all the hyperscalers are in the red, like badly in the red. Semiconductors are doing great.

So yeah, Apple's just raised their prices on their devices too. So the effect of the memory crunch is going to be felt in many, many places.

Apple's down 4.5%. For the day, which is massive. Even Nvidia's down. Nvidia's down below $200 now. I think the effects of Broadcom, OpenAI, the jalapeño chip are being speculated upon. There is a question in the market how long the hyperscalers, which are really the best-run companies on the face of the earth, are going to be willing to bear this kind of milking by Nvidia and Micron and the memory producers. And the expectation is that Microsoft bows out first. So at some point, Microsoft says, Satya says, look, we're going to slow down data center buildout, and we're not going to raise that much money, and we're going to slow it down. But it's also hard because AWS, on the other hand, Jassy is saying everything is You know, in the black, we're going to make so much money, we're going to deploy more capital, and we are the data center leader. We are going to continue to be the data center leader. We will outinvest everyone else.

Never dull moment. We'll see you back here tomorrow morning for another fun edition. And tomorrow there's going to be a couple really good segments, I think, but one I'll highlight Cameron Berg is going to come on in our last section and talk about the latest in AI welfare research. There's been some— he's really a pioneer of this field and there's a few others, and there have been some fascinating papers that I'm excited to get his take on that continue in this line of research around functional emotions and extend to now functional welfare. What can we understand about what goes on when a model gets reward and how it may feel about that? I think that stuff is almost certain to continue to remain mysterious after tomorrow, but I think it will be a mind-expanding and potentially hair-raising conversation with him about that. And then we also have something called Forum AI, which we'll be talking about, where Robbie Goldfarb from Forum is really trying to bring the highest level of taste into AI systems. And you go to their website and it's just like, here are all of the experts who we think you would respect as having excellent taste and judgment. And these are the people that we're trying to somehow get their je ne sais quoi into the AI. So I think that'll be a really interesting conversation as well.

Absolutely. I'm looking forward to it.

Lots more to come. I'm so torn on this question of— this pattern just keeps coming up over and over and over again. Concentration of power on the one hand versus In some cases, like, I mean, in this case, it's kind of concentration of power versus concentration of power, I think, because on the one hand, I think you do want a healthy app ecosystem. And I'm wary of what happens if this greater than 10 to 1 cost advantage persists. And the, you know, that, I mean, it's hard to make up a more than 10x difference. I would say that, you know, for the things that I'm doing day to day, I don't find myself reaching to other apps all that much because I just have a lot of tokens, you know, and I can— I just generally believe that even if I have to kind of code up a tool or figure out some of my own scaffolding, again, like, obviously I'm an early adopter and, you know, more comfortable than most doing that, but it's getting easier and easier too over time. So I just don't find myself going to these different apps all that much. I more often am just like, code up a tool, look into Arxiv yourself, do whatever search. The flip side of that though is if I as an individual user lose my subsidy, and it is interesting that it is quite subsidized when you consider the cost, there's not that much margin at the inference provider level on a GLM-5-2. We can be pretty confident of that. So the fact that the cost per token is cheaper in my Claude Max account than GLM-5-2 is, Open source means it's like if I'm maxing it out, they're definitely subsidizing me. If I lose that subsidy though, as an individual user, then the concentration of buying power potentially shifts to the enterprise and you have a different kind of problem where already super profitable enterprises are just way more— if we're all paying on a sort of auction basis for for tokens on a marginal basis, on a purely level playing field, it's very hard for me as an individual to compete in that buying market with enterprise customers. Yeah. So I can see both a— I can see an anti-concentration of power argument working both ways, and I'm not really sure where that leaves us, but right now it seems like the app layer is getting squeezed. The individual layer is being empowered. You could flip it around and empower the app layer more, but then you disempower the individual and maybe empower more the enterprise. And I'm not sure if that's better or worse. It's worse since I'm an individual, I guess, but it's still kind of hard to evaluate in the grand scheme.

So I think you can think about other high CapEx variable, high fixed costs variable revenue businesses. And the classical example is usually airlines where you've already paid for the plane upfront. And then you have to maximize revenue on the plane. In order to maximize revenue, you end up having a first class section and a bit— or a first and business, basically, let's call it business class section and an economy section. And you have multiple strategies that you can use in order to maximize revenue and survive, et cetera, et cetera. And you have a premium product and the premium product has, for business class, schedule. There's dispatch because the business class, you can normally change flights very easily. You can just arrive without making a booking and kind of just onboard as you will. So they maintain more capacity for you. On the economy class, they offload you sometimes, right? So if they overbook, and once they overbook, they offload you. And you can kind of see the same concepts in terms of batch processing and direct processing that even the Frontier Labs have. And even the open source providers, there are providers that will basically give you dedicated capacity where it's always going to be available and they give you a latency guarantee. And then you have providers where they will always boast about their low cost, but they're not going to provide you a guarantee on the latency or other things. So you have this again, this kind of price discrimination process where you have different products being built out. And the premium, premium product is obviously the Frontier Lab product where they can guarantee you the state-of-the-art intelligence as it is right now. And you have that state-of-the-art intelligence window for anywhere from a few days to a few months. And within that state-of-the-art intelligence window, again, you have this set of GPUs that you have to maximize the revenue off of. And so OpenAI has leased 100,000 GPUs and they allocate this much to the frontier and this much to Post-Frontier. The Post-Frontier stuff has to be competitive with other firms. So OpenAI is still serving GPT-5.4 Nano and GPT-5.4 Mini, which I still use. And those products are at, I think, 25 cents per million tokens, are very competitive with the other open source models. And I actually depend on these OpenAI models for these other tasks because I know that the API is reliable and they're much, much larger than, let's say, Together or Fireworks, etc., and don't have to set up stuff. And I kind of know the model, and it's also kind of matching with my maybe GPT-4.5, you know, of a year and a half ago. And the intelligence that a GPT-5.4 Nano is kind of matching that, and it's much, much faster than it used to be. So all of my queries that I had GPT-4.5 a year ago, I now offload to 5.4 Mini and Nano. And OpenAI is probably making an 80% margin on those, right? Because those are old, old models. And so this is what they're doing. They basically have this high-capex GPU fleet, and they're trying to do this kind of revenue price revenue maximization off the GPU price fleet that they've paid a fixed cost, fixed leasing cost for. It's very similar to airlines. It's very similar to these other things. And you're gonna see, again, similar strategies of a premium product. And the premium product is obviously just like in airlines, there's a branding involved to it. There's like, oh, you know, we have a nice lounge for you and we have, you know, nice food for you. And the food is much better, or our dispatch is much better, or all of these other features which are perhaps hard to measure in pure price alone and which have different levels of importance for different people. Some firm is going to say our math model is better. Some is going to say our search is better. Some are going to say our coding is better, for example. And so this is, I think, what's happening. We're seeing this kind of revenue maximization strategies off the same GPU fleet. And I think the interesting thing is, I think a lot of people have not figured out this game yet. They're still like, oh, it's a SOTA model. It's how you maximize the revenue off the GPUs, because the GPUs are your fixed costs and you have to figure out how to maximize. And a lot of that maximization is around marketing around the product. Yeah.

I have no Beef with price discrimination and cross-subsidy in the airline industry. And I guess for calibration, I've also been broadly not worried about net neutrality. It's that— remember, I'm old enough to remember, I think you are too, when net neutrality was this sort of sacred cow. And then back in the first Trump administration, they really weakened the net neutrality rules and basically allowed for some price discrimination and pay for priority and all these kinds of broadening of the possible deal space between the infrastructure owners and the people that want to move content around. And I remember at the time there was a lot of like, oh my God, this is going to be the end of the internet. You're going to have to pay your ISP if you want to use specific websites and all that kind of stuff. And I believed at the time, and I think history has certainly vindicated my predictions that I don't think any of that's going to happen. I've not heard one story of an ISP blocking content and saying, you have to pay us, the ISP, to access your preferred news source or what have you. And in general, the internet seems to work. If anything, it's continued to improve over the last however many years. And so, yeah, I say all of this as kind of a lifelong techno-optimist libertarian, and still somehow this feels like maybe a little different, right? This is just another version of the same old worry that maybe AI is just a paradigm breaker, and it's not obviously the case, but it is at least plausibly the case. that the tokens themselves are just so much of the value that we could have a real loss of diversity at the app layer if there is on an indefinite basis a 10-to-1-plus cost ratio. That's also a pretty big cost ratio. I guess the business class tickets are certainly often a multiple of the economy class.

Yeah.

So it's not Qualitatively different. But it is tough. I mean, my sense right now is it is real tough at the app layer. Andrew from Tasklit, you know, said that point blank. If we'd had a little more time with Eric, I'm gonna ask him the same thing. But Andrew said the number one place people go when they churn from Tasklit is to Claude Code.

Yeah.

So for him, it's like absolute strategic imperative that they become platform neutral. He thinks that's one of the few kind of archetypes of software company that can survive is the one that acts as a layer that sort of helps you as a consumer play the foundation model companies off against one another. And in that, there's some defensibility, some power, something that they won't be able to do for the end user. 10 to 1 though, it's a big ratio.

I think perhaps it's actually a different problem because it's not just that the foundation model companies have the foundation model, but they also have the app. It's not just that you have Claude the model, you have Claude the model and on top of that is built a bunch of apps and the apps are also built by freaking Mike Krieger, who was the head of product and CTO of Instagram, which is the all-time best consumer app in history, right? The all-time best designed, best consumer app in all of history. And the guy who was the co-founder and CTO became the head of product at Anthropic like 2 and a half years ago and recruited the entire product team And they're all like, everyone's like, oh my gosh, I get to work with freaking Mike Krieger. And so he flooded the zone with all of the leading product people in the space, recruited them, and ran the product organization for a couple of years and didn't make stupid mistakes. Like for example, Claude Code was created by Boris and And Anthropic didn't believe in paying product people that much. And so Boris actually peaced out to Cursor. And you can just imagine how much battling must have gone on inside that organization for Mike Krieger to go and pull Boris back and say, and tell Dario, look, you've got to pay this guy and just elevate Claude code to the position that it is. And it's become this Massive Cloud Code, Cloud Cowork, Cloud Code on Desktop, and now CloudTag. And it's just morphed from that, right? So I think you have to take a step back and say that Tasklet isn't competing with foundation model researchers. It is competing with the leading product mind of the last 15, 20 years who was trained and has experience taking an app from literally 11 employees to a trillion-dollar valuation. And that's the journey that Mike Krieger went through. And I think you have to say that they're competing with a product organization. They're competing with an application. They're not really competing so much with just the model. The model is great, but you have to also see that all of the innovations in the product are being imitated across the space. Codex and Codex Command Line, all of these things are imitating what was built on the product organization at Anthropic. So that I think is the problem as well. It's not just that the model is good. So the product team is also cracking good. And that integration between the product and the model teams has gone on pretty well. And that's what they're really up against. That's the tough part, I think, for other firms.

Yeah. There's another level too that they haven't done much at all yet, but could easily do, which is just only deploy their latest and greatest models into their own first-party products.

Yeah.

And that becomes another dimension on which, boy, if you really are playing to own it all, maybe there's enough ability for the frontier companies to keep each other honest in that respect. But right now with Fable, there really isn't. I mean, OpenAI supposedly will have an answer Soon, soon-ish. Maybe already does and they're just kind of waiting to let the dust settle. I think it's safe to say at this point Google does not have a product that they can— that they do not have a model that they can put on the API that would really compete with Fable. So in a Fable versus GPT-5.5 versus Gemini 3.5 Flash world, if they decided they wanted to make Fable only available to first-party users, really tough to compete with that. It's hard to substitute away from just that much more raw intelligence if it's not available via the API. What are the app layer companies going to do?

I will also say that The moat is perhaps even less than what we think it is because I'm not sure if you saw this, but there was a release of an ad product where while you were waiting for your Claude to respond and had that little spinning thing—

Yeah, I did see that.

You would see a little ad while you were waiting and they would pay you for it. So it was a Chrome extension, so it plugged into your Chrome. which would sit on top of the, uh, you know, any, uh, Claude web page that you're— Claude or ChatGPT web page that you're opening up. And every time it would basically inject JavaScript in there while you were waiting.

Are those tokens also being sold on the Chinese black market, I wonder?

Ah, yes, yes, that's exactly it, right? Because when you have that extension, it's also reading the tokens in and out. It can read the tokens in And people would be fair. If you told me, hey, here's your Claude subscription, and you put in the extension, and you watch ads, and we'll pay you $50 a month of your $200 subscription. And then you opt into the data, and we'll pay you another $50, like $100 a month. Just opt in, right? And consumers are absolutely free to do this because they own their own data. Like, you can't, you know, that's your, you know, US rules. Like, that's my data. I, when I speak to you, that's my data and you have to give it back to me if I ask you for it, right? That's the, that's the rules that these companies are under. So yeah, you could just have a, you know, extension basically streaming out all of the tokens and have faster distillation, have fast and pretty much immediate distillation on the other end going on. People don't do it because there's not enough money in there to actually make it worth their while right now. But if you annoy the consumers enough, people will do it and you'll get very fast escalation. So I think there's always this kind of middle ground that the firms have to do where they clamp down on activity that they don't like, but they don't clamp down hard enough That the extremely motivated actors will go out of their way to take them down. So I think there's always this middle ground and they have to kind of tread that middle ground. I honestly think it's very difficult for app companies because, for example, one strategy that I'm very surprised that the Frontier Labs have not done yet is buy sources of data. You could easily imagine, and I don't really understand why, because they say that they are for scientific advancement. And one of the main issues with scientific advances right now is that a lot of the scientific papers are gated behind these journals. And all of the journal subscriptions that you need will probably cost you like $2 to $20 million to get all the subscriptions that you need. And so researchers end up going to Sci-Hub, which is based in Russia and which is pirating 100 million journal articles, which the feds have put under investigation. And if the founders stepped outside of Russia, they would extradite. Drug-level, drug-dealing, arms-dealer-level scrutiny. And all the researchers in the US are openly using these sites to access data and access scientific papers. I don't know why the model companies don't just buy access and just plug it in, like plug in your scientific data subscription directly in there. And it would be enormous value. And you can pay these firms, because these firms don't make that much money. I think all in, all the scientific publishers are probably worth $20 billion in in total, probably make less than, you know, $1 billion in revenue in total per year, all scientific publishing. So why don't you just roll all of it in and just buy them out? And then you, you know, you basically just plug in the data and you have the publishers in there with you. Yeah, I mean, I know it's not their core business. And I know it's also because like, Hey, you know, once we do that, the app-level layer companies really have nothing to do because, you know, if you provide a front end to scientific papers and then the scientific papers are already plugged into Claude and ChatGPT, like, what is the point of that, right? So yeah, I do wonder though.

Well, certainly the money's going to be there. I was just looking up Epic Health too, and it's not public, but estimated valuation is somewhere in the $30 to $60 billion range. billion range, which, you know, tops out at cursor scale. So, you know, if you wanted to get— and of course there's some regulation around this that they would have to find their way through as well— but getting your hands on the system that controls, I don't know what percentage of all American electronic health records.

70 or 80, I think, something like that. Epic.

It's a pretty direct path to an awful lot of data, and it's very affordable for these guys at this point. This kind of connects too to something I wanted to talk about today that we probably don't have time for a full discussion of now, but maybe we can get to after our next guest. But this company yesterday, Mirandyl, launched— they didn't even really launch. They basically just announced their funding. They haven't put forward in detail yet what their product is going to be. But when you were speaking about kind of the middle ground and, you know, trying to control things but not control them too much, so there's this counterreaction. This is an interesting case. Another Tolkien-named company, by the way. I had to look it up, but Mirandul means in Elvish sort of like a guarded or kept secret. And The company consists of former Anthropic people who I think were on this well before, but probably saw the writing on the wall. So they were starting this company before the fabled launch happened and before the public saw that, hey, we're planning to degrade your model performances without telling you, by the way, if you're doing some frontier AI research and presumably seeing the trend, if not the actual policy plan, these guys went off to start their own company and they are making the explicit appeal to the rest of the world on an anti-concentration of power basis. They're basically saying, look, we were at Anthropic, they're doing recursive self-improvement, they don't want to let you have it. And so we're going to bring that to an independent company and we're going to make it available to all as a new set of tools. I'm not so sure how well that's going to work. We'll see. Certainly some very smart people involved. I don't know how they keep up, but maybe they don't have to. I guess if Anthropic doesn't want to compete in that space, they don't have to keep up with Anthropic. Seems like They're going to be behind, but maybe being behind is okay if you're the best option in the market. But a very interesting example of where you squeeze too hard and maybe the key things kind of slip through your fingers. Some apparently ideologically, I don't want to ideologically, maybe sounds a little strong, but I think it's fair to read their launch statements as being kind of ideologically motivated. And when you talk about concentration of power in your launch statements, I think that's fair to say it's somewhat ideologically motivated. They seem to have gone out and started this company on the basis that we don't like the mindset. We don't like the direction that Anthropic is taking in trying to hold all this stuff back. So we're going to go do it. And for now at least, it remains a free country and they're going to do it. And they've got $200 million from a16z and other VCs. And they say a substantial, I think they say large investment from Nvidia on top of that. Exact dollar amount not disclosed, but it's also an interesting move to see Nvidia come in and say, okay, we don't want to see this kind of frontier R&D monopolized, so we'll support a company that's going to try to democratize that sort of capability. The strategic dynamics of this are going to continue to fasten.