The Pomp Podcast - Why No Company Will Win the AI War: The "Rebel Alliance" Thesis | Nick Grossman

Episode Date: July 20, 2026

Nick Grossman is a General Partner at Union Square Ventures. In this conversation, we break down his "Rebel Alliance" thesis for AI — why he believes the industry is too big for one or two companies to dominate. We also discuss USV's internal multi-agent platform, model routing economics, AI's growing role in venture capital and financial markets, data privacy risks, and what happens to jobs and society as agents get smarter.=====================BitcoinIRA: Buy, sell, and swap 80+ cryptocurrencies in your retirement account. Take 3 minutes to open your account & get connected to a team of IRA specialists that will guide you through every step of the process. Go to https://bitcoinira.com/pomp/ to earn up to $2,000 in rewards.=====================Looking for a better place to trade? BloFin gives traders access to deep liquidity, advanced futures products for crypto AND TradFi assets, fast execution, and a clean, intuitive interface—all in one platform. To celebrate their partnership with us, they're giving away $100,000 in Deposit & Trade Rewards. Deposit, trade, and earn rewards based on your activity during the campaign. Check them out at ( https://partner.blofin.com/d/Pomp ).=====================Simple Mining makes Bitcoin mining simple and accessible for everyone. We offer a premium white glove hosting service, helping you maximize the profitability of Bitcoin mining. For more information on Simple Mining or to get started mining Bitcoin, visit https://www.simplemining.io/pomp=====================Arch Public is an agentic trading platform that automates investment strategies across Stocks, Commodities, ETFs and Crypto. Whether you’re rotating into AI & Gold, allocating to the S&P 500, or accumulating Bitcoin, Arch Public executes your plan 24/7 without ever taking custody of your assets or funds. Sign up today at https://www.archpublic.com, and start your FREE automated trading strategy! =====================0:00 - Intro0:42 - The "Rebel Alliance" thesis: why one company won't dominate AI7:15 - General-purpose vs. specialized models in USV's portfolio8:40 - Compute costs & the rise of model routing11:14 - Inside USV's internal multi-agent system16:25 - AI's role in venture capital & autonomous investing22:50 - How AI reshapes market signals & information edges27:06 - Apps building models & the fight over your data35:31 - Five-year outlook for the AI stack37:57 - US vs. China: the "philosophy" of AI models40:30 - Personalized AI & where the Rebel Alliance thesis plays43:53 - The dark side: jobs, data centers & society47:44 - AI as a personal superpower & neural implants

This episode is brought to you by Accenture. When your advertising operations fall out of sync, everything else follows. Spotify and Accenture are working together to reinvent the rhythm of ad sales. Using automation, analytics, and smarter workflows to simplify campaign delivery and access better data across the business. The result? Less time spent on operations, more time connecting brands with the moments and fandoms that matter most. Learn more at Accenture.com slash Spotify. The 2026 Chevrolet Equinox, awarded the most dependable compact SUV in the U.S. by J.D. Power, is designed for your everyday. And with available all-wheel drive, you can handle your to-do list

with total confidence. Start your build at Chevrolet.ca. Details at JDPower.com. We're going to absolutely be living in a multi-agent world, thousands, millions of agents, and that there will be agents that kind of look like people or teammates that you talk to, and then there will also be agents that are just embedded in all of our systems. And what's going on, guys? Today, we've got a great conversation with Nick Grossman. He's a general partner at Union Square Ventures. And in this conversation, we go deep down the AI rabbit hole. We talk about a thesis that he has called the Rebel Alliance. It's the idea that there's going to be tons of companies and that the AI trend is way, way bigger than you're thinking about. On top of that, we get into some of the nuanced details about what's happening in terms of building this technology, where value is going to accrue. What are the different use cases? Where is the technology going to be valuable? And what are some of the negative tradeoffs? All of that and much more in this conversation with Nick Grossman. All right, Nick, you guys have a great piece that you put out that talks about the Rebel Alliance when it comes to how the AI industry is going to evolve. What exactly is the Rebel Alliance? So the Rebel Alliance is our view that AI is too big an opportunity to be dominated by one or two or three companies and that it's actually forming into a massive ecosystem.

And we're seeing that happen in real time as agents and agentic approaches start to take hold. And that's really two things. It's like agents that interact with consumers as AI doctor agents, AI financial advisor agents, these sort of personified agents. They're going to be approaching every sector and every use case. And then also under the hood, it's agents as infrastructure. What we're seeing now is the agentic stack is forming like the compute stack formed, you know, in the prior eras of the web and mobile, where, you know, we're not just talking about what are the products that people interact with, but how are every product going to get built. And that's agents under the hood, agents as infrastructure, fleets of agents that are going to get orchestrated in millions of ways. And what that says to us is that agents and the agentic AI approach is taking over, and it's going to look more like the fabric of computing and apps than, I think, the thing that a lot of us are maybe afraid of, which is everything collapsing into one or two apps, like a ChatGPT or a Cloud.

Now, if you were talking to, like, Peter Thiel, he'd be like, competitions for losers, you know, I want to find a monopoly. He's done a great job of that throughout his career. You're really arguing that, like, that is not going to be the end state here? Well, there are real network effects in AI, right, in model training and, you know, quality. And so, you know, so far we've seen compounding advantage going to the big labs. And the big question has been how far will that take us and how much of advantage will that accrue, especially as they go farther up into the application layer and get access into more data from more places. So that's the worry, that it's going to just compound and that there's going to be an unassailable gravity into the big AI companies. But I think what we're seeing is a couple of things.

We're seeing real competition among the models. So it's, you know, it's the frontier models from the big labs, but also the leading open source models are really good and they keep getting better and better in lockstep. And developers are experiencing a lot of choice and are pitting models against one another. And then we're also seeing that like the world is big and wide. And if you think about the advantage is being able to capture data and use that data to improve systems, that data is going to show up in lots and lots of places. And a lot of that data value is going to accrue not just to the model layer, but also the application layer and other layers. So it feels like a huge, huge, huge opportunity, you know, for a lot of folks to win. Now, let's talk about a couple of different components. And I was telling you before we started that we've been building Silvio, which is this AICFO. And so in a weird way, I have half my brain as an investor, half my brain is like operating this business, building the technology and seeing a lot of these nuances that I frankly would not see if I was just sitting in the investor seat. One of the things is you're constantly worried about the big labs who have tens of billions, hundreds of billions of dollars, trillion dollar valuations, thousands of employees saying, that's a nice little business you have there.

I'm going to take that for myself. Now, we have both had the blessing and the curse where some of the labs have launched competing products and some of them have chosen to go after enterprise or something different. How do most of the companies in your portfolio think about, is it good when there's competition because that validates a space? Is it bad because they have a lot of assets? are they distracted? And so maybe they like go into a space, but they're not very good at doing it because they're not as focused. Like, what are you seeing? Yeah. I mean, it's, it's definitely

both. And I think any entrepreneur building this space is excited about where they can go and also worried about the shadow of the big labs. You know, we refer to it as the kill zone, the AI kill zone. And, you know, where does that land and how big is it and how real is it? I think we don't really know the answer. I think folks who are closest to the core skill set of the models today are probably the most worried. And that's things, you know, language, media, law, code, you know, things that are really natively digital, where the core experience of the big products from labs is really good. But I think as you get, so I think there's a lot of worry there. But even within those areas, you know, there are really good companies that are building defensibility and

specialization and connectivity into the real world, if you look at, like, AI lawyers or AI doctors or even AI financial advisors, like, there's a lot to build. Or even in AI music, you know, it's about developing an experience that is special and unique and really resonates. And, you know, our bet is that while the big labs are amazing at what they're doing, the idea of being the best at every use case is too much, too much to chew. And do you feel like that is, you know, kind of, it feels like there's like general purpose models and approach. Then there's this like specialized workflows. And, you know, again, I'll just go right to the example I understand, which is when I look at opening, I launched a competing product.

The number one takeaway I had was there was some core functionality that was missing if you were looking at it through the investor lens. So the AI CFOs, you know, trying to serve these investors. And so it's missing some of this core functionality. If you look at it through an engineering lens, it's exactly what you would build because it's like, hey, Plaid and SnapTrading, these things have the connectivity. And so it was like, oh, wait a second here. I've read about in legal or medical or, you know, but I don't understand those sectors to the depth I understand the investing stuff. And so to me, it was almost like, are these engineering problems or are they actually like almost like customer pain problems that need to be solved? And where does that expertise lie?

Yeah, I mean, I think you're taking the bet that, you know, as someone with experience in that field and knowledge and, you know, a feel for the customer need, you can pull the product in the right direction. And I think a lot of entrepreneurs are doing that, not just in categories that are kind of close to the big labs, but in other categories, too, like factory automation and robotics and all these areas. Like, there's so much expertise to be had in the world. There's so much connectivity to the real world to be kind of implemented. And I think there's a lot of knowledge at the edges. And on top of that, there's just like a lot of different form factors that tools and applications need to take. And I think people are going to pull them in many, many, many ways. When you look at your portfolio, how much of it is you guys betting today on like general purpose approach versus specialized workflows?

If you had to put percentages. That's interesting. It's definitely a combination of both. You know, we have investments up and down the stack from, you know, open approaches to model training to, you know, AI applications focused on, you know, medical and music and other things. And so I think in robotics and lots of others. And so I think one of the things that makes us bullish about the idea of the Rebel Alliance is that as we watch our teams build applications across all these spaces to build a really good application, you're using a combination of general models and specialized models. So if you're, you know, building a robot that needs to, like, navigate around and do patrols around the outside of a factory or whatever, you know, there's general reasoning. There's also spatial reasoning.

There's local mobility. You know, there's, like, a whole lot of functions that you need to have. Some of those can run on the cloud. Some of them need to run locally. Like, as you start going out into all the places of the world, you start having – you start seeing a real hybrid approach to model usage functionality needs. To me, it says the world is going to be multimodal, and most use cases are going to be multimodal, a combination of general purpose and specialized. Yeah. What I see on the model side is we're one of the fastest growing, I think, on a percentage basis, customers of Clodge, right?

just month over month we've exploded the wake-up call actually you'll find this interesting uh i was talking to one of the executives at a public company and he asked me he said how much are you spending on computing i told him and his eyeballs like thought i said you're spending more than us this is like a really big public company and i was like oh that's a problem right like how is that possible and so we basically went and almost created like an internal task force like let's get the token cost down. So my takeaways were one, the token cost was going to become a topic for us. I started talking to other CEOs and they were like, hey, this is too expensive. But the second thing was I still wanted the intelligence and productivity. And it's not like other CEOs did

too. It was just like, I want a lower unit of intelligence per dollar, if you will. And so I thought that was going to have to be, that means we have to go find a degradated performance in order to get the lower cost. Now it seems like the consensus is like, well, this whole idea of model routing and the ability to predict the complexity of a query and which model is best at answering it and is it specialized or not? You're seeing that proliferate across the companies that you guys are invested in? Definitely. And it's both cost and quality. I just saw a report that came across last week about compute performance using a bunch of different harnesses on top of a bunch of different models.

And the best performing coding harness was the Pi open source harness on top of a combination of models that was beating out Clod in Clod code or Opus or in Clod code. And so, and we're seeing a similar story across

pretty much everything in our portfolio, where when you use, when you can intelligently use a mix of models, you can not only optimize your cost, you can improve your performance by using the best model

for the best task. and orchestrating models that way. So I also came at it thinking it's just about cost and everybody just can't be just like with their pedal on the metal for the most expensive cloud model, but it's also about optimizing for quality. And I think as everybody moves

from kind of prototype phase to deployment phase, right? The whole world has been prototyping with cloud for the last, you know, in OpenAI models, GPT models for the last nine months. a lot of stuff's moving into production now. And I think we're going to see much more sophisticated orchestration of agents and models, both for cost and performance. Explain that a little bit more in terms of the orchestration. Like my experience is people think they're talking to one model, but actually the underlying kind of technology is a bunch of

models doing different things. Yeah, I'll give you an example from something we're building at USV. So we have an internal platform that tracks deals and companies and people and ideas and all this stuff, kind of like a CRM, like our platform. Like a Palantir for VC. Exactly. And we built it ourselves over the last nine months, which is crazy. And we just launched something yesterday where every deal, company, idea, potentially person in our platform has an agent attached to it. And they're all orchestrated through software. So it's not just a single agent you talk to.

It's like thousands of agents that are under the hood in our system. And that's just one way of designing it. But what are like the, so like you as a user of this product, you'll come in and say that you've been tracking a company, you want an update on it or something? Yeah, so we have, so we have a, this example makes sense. We have a single agent that's like our deal analyst agent. Okay. Lives in email, lives in our messaging, whatever.

You can ask that agent whatever questions you want. It feels like a team member, right? And the types of questions you would ask this agent is like, hey, what's the latest on this deal? Yeah, or like research this company for me, tell me more about the founders, who are the competitors? You know, what are some risks? What should I ask them in the next follow up in the next meeting I have with them or whatever. So that's like our main analyst agent that kind of knows about everything. Under the hood, we have a sub agent that is only focused on that deal. And

so we all go to bed at night, and that agent wakes up and goes out and listens to Twitter and listen and does web searching and sees the notes that have come in from our system during the day. It's like reads them, kind of dreams about them, thinks about them, goes and does a bunch of work. And And then that work gets, like, pushed back down into the main, like, shared context memory data layer so all the other agents can, you know, benefit from it. So, you know, I don't know that this is the ultimate architecture for agents, but what it says to me is that we're going to absolutely be living in a multi-agent world, thousands, millions of agents, and that there will be agents that kind of look like people or teammates that you talk to. And then there will also be agents that are just embedded in all of our systems. And, you know, we're building a version of that. I'm sure you're building a version of that in Sylvia.

How much of, let's say, this internal product that you guys built is you're generating some action, task, question, whatever, versus it has become, you know, it's somewhat sentient. I'll use sentient, not scare everybody, but like it's almost thinking about here's the thing that he's not asking that he should be asking. Yeah. Well, I think there's two parts to that. One is you want to prompt them to always be questioning and thinking and, you know, like reasoning and being proactive. And then the second piece is, like, what are the triggers that they're working off of, right? So our fleet of worker agents trigger off a bunch of things. Some of them have, like, a daily, you know, wake up, think about this company, you know, do some work on sort of a timer, like a cron job.

Others of them are responding to things that are happening, whether a meeting gets pushed into our notes database or an email comes across our investment team list or some other trigger that happens. And so there's sort of triggers that are firing agentic work within our system all day long. And, you know, what we I'm sure where this is going is because as the agents get smarter and more capable, they can do more long running tasks. They can make more decisions on their own. You're seeing that today in coding agents where you can give a coding agent like build me this whole app and they can just go do it like fables. Really amazing at that. So as they can do more long running tasks, they can also be more proactive, you know, across each step. So I think we're moving from a kind of a chat based call response paradigm to more of a proact to like a trigger based paradigm where you have cloud agents that can trigger off of things that are happening in your system to more of a kind of proactive, you know, model.

And we're already seeing that, you know, move that way. Today's episode is brought to you by Bitcoin IRA. Are you a crypto investor with a retirement account but don't have any crypto in that retirement account? Then listen up. This message is for you. Bitcoin IRA is revolutionizing the way Americans save for retirement. They're helping smart investors diversify their savings with access to over 80 cryptocurrencies with world-class customer service, military-grade encryption, and a vertically integrated licensed

trust company. It's no wonder more than 200,000 Americans trust Bitcoin IRA to secure their financial future. Getting started is quick and easy. It takes just three minutes to open an account. Once you're set up, their team of IRA specialists will reach out to guide you through every step of the process. Whether you're transferring an IRA from a legacy bank, rolling over an old 401k, or starting fresh with new contribution, the Bitcoin IRA team is here to help you access real crypto in your retirement account.

And here's the best part. As a Pomp podcast listener, you can earn up to $1,000 in rewards. All you have to do is add funds to your account. search for bitcoin ira in the app store or visit bitcoin ira.com slash pop and you can join 200,000 americans on their journey to upgrade their retirement that's bitcoin ira.com slash pop to upgrade your retirement today now if we go and let's just use the venture capital example um right now uh i would say that this sounds like it's a system that you guys built it's human you know generated or kind of human initiated then there's a bunch of automation and ai and stuff that's happening the information is coming back but like you as a human you make the decision right you're you're you're the you're the final press yes no you know caesar type uh type thing

there have been many people we 10 years ago i was like why didn't the software just do it at the earliest stages you know there's not a lot to look at is you're kind of just like betting whatever. Can investing get there? I think certain kinds of investing can get there. Okay. Explain. Well, like we do venture investing where so much of what we're doing is, you know, really episodic and you have to spend time with humans in real life to get to know them and know their company. And it's a very human process, you know, today. And also, you know, we're not trying to have our agents make investment decisions or kind of execute on those investment decisions. Although there are people who are trying to build agentic VC,

like fully agentic VC. We have a portfolio company that's working on that too. So I think for certain types of deals, it could happen that way. And then outside of venture capital, I mean, already, I think the first, you know, fleet of investing agents were like trading crypto, you know, and things like that, where, you know, it's really more programmatic, the market's accessible. You know, this is a kind of a separate conversation. But I think one of the things that's exciting about where the world is going is agents, you know, having wallets and making payments and being more autonomous financial actors. And the places where that's going to be the easiest are the markets that are the most open and accessible, either via API or because they're, you know,

on crypto rails. And so definitely see autonomous investing as a thing that is already happening, going to keep happening. And then will it happen to VC? I think certainly parts of it and maybe eventually all of it, but I think it'll be a little, probably a little slower. Some of the, I mean, as much as I think we all love and respect YC, like there is somewhat of a joke of like, you know, it's a 10 minute meeting or whatever it is. And so there's a bunch of work that gets done that, again, it's not fully automated, but you can easily see maybe the human component is being compressed. And so the question is like the last mile, right? How much of that is necessary? And I do wonder, it's almost impossible to know,

but like how much of that meeting is less about making this decision versus it is establishing the relationship to open the opportunity for the next round or something. That's both. And so I think when you think about venture capital deals, the style of deal you tend to focus on most is the lead investor who's like really building the relationship and crafting the deal and, you know, taking a big role within the company. But in every venture round, there's also lots of follow on dollars. And so one area where I could totally see more automated capital coming to venture deals is like filling in edges of venture rounds where there's, you know, 200,000 or a million or whatever it is left. And maybe there's some signal that a VC, kind of like a pool of capital agent, could follow parameters and, you know, act pretty autonomously and pretty quickly. There's a firm that did something like this. I forget the name of them.

I know what you're talking about. Okay. And I can't remember the name either. We have a portfolio company. But they basically just followed. They were like, hey, there's, you know, here's the tier one VCs. Right. If they, like, lead an up round and they're the lead and it's done in some time, whatever they're like.

It's programmatic. Correct. That looks like programmatic. Humans still said yes, but it was pretty much done. And you could imagine more and more and more of that getting automated. And could you go into other parts of venture maybe probably? Now, when it comes to agentic investing, one of the things I've been somewhat surprised by is digging a little into the details of like how some of these systems work. There is like automated execution and then there is like full on agentic decision making and automated execution.

I don't think anyone has a problem yet or is concerned about like automated execution because there's still some element of like oversight or parameter set or, you know, if then type statements, whatever. The AI models, I have seen very dispersed outcomes. I've seen some people be like, tell me a stock to buy. It's going to double by the end of the year. Right. And like, you know, like right goal, pretty crude, you know, just like general prompting. I haven't seen too many success stories from that. I've also seen we see on Sylvia.

Um, we actually have one engineer who, uh, every couple of days he'll like post, I made X dollars and he'll like explain what he did. And it's some combination of like stock screen plus, you know, uh, maybe he's got some unique view on an industry or something, but really what he's using is again, it's kind of his idea that he's then using to find the idea, the full on, like, here's, you know, a hundred grand, like knock yourself out. I feel like we're all excited about it. I just don't know if anyone's like fully there.

Yeah. And I can't say that I've heard too many actual stories of that actually. I know one person who is building this, but what he has been doing is less like high frequency trading type stuff, because obviously those guys are excellent at what they do. And this is almost more like index creating. OK. And so it's like less frequent trading and it's more, I think, just like you can kind of look at, hey, what's going to happen for the next, you know, 90 days type stuff, right? That to me feels like, okay, and in his case specifically, the performance is like amazing. But like the average person just like prompting the model, I haven't heard too many.

Yeah, I've heard some, you know, I know a team that is, you know, using agents to trade frontier markets. So, corners of prediction markets and other kind of obscure places where maybe there's some early dislocations that, you know, maybe it's largely retail and maybe an agent can see patterns, you know, before the big institutions get there and arb everything out. I don't know. So, I do think there are probably existing corners of real markets where you could probably just set an agent free and have it – give it a wallet, load it up. Those guys are making money. They're making some money. Yeah. It's like it's quote-unquote working on a –

Working on a small scale. Early. Got it. Okay. The other aspect of this that I find fascinating is if you know that the agents start to become more involved in financial markets, well, like every company in the world, every founder in the world starts saying, like, how are the decisions made? What are the things that matter? And we know this because humans do it. if they know that a bunch of, you know, VCs are all looking to invest in physical AI, all of a sudden every company's got physical AI in their deck, right? So, like, there's always this, like, cat-and-mouse game of, like, the signal

gets arbed away. That feels like where AI is really good at, like, dealing with the complexity, identifying some of the stuff, et cetera, right? And arbing that away faster. Is that what you mean? Yeah, yeah. It's just like, okay, where's the trend shifting? Like, I don't know if you know who Jordy Visser is, but I do this

episode with him every Saturday. And one of the things he told me, the best idea he's ever told me, he said to me, a lot of the AI leaders are more public in terms of interviews and stuff than any business leaders ever have been. Like Jensen gives like five interviews a week, right? People want to know what he's thinking. If you take those and you take the transcripts and you put them into the systems and you say, hey, based on what Jensen is talking about, what are the companies that are going to benefit over the next 12 months or something? Like scary accurate. Yeah, I bet. Right. And so, you know, he tells me this is like a cool idea. Gavin Baker on stage at his own conference. So what I do is he's talking about Tranium being like a big H2 trend. Okay.

Okay. Take it. I go ask the model. So if Tranium is the thing, listen to this interview, whatever, what do you think? A full risk on portfolio, 30% allocation, Marvell. 60 days later, Marvell's up like 100%, right? Now again. Did you trade that? No, dummy me. I heard Jordy had told me about Marvell, then this, right? And so I'm like, oh, of course, like, you know, I can't now FOMO, right? Right.

But what I started to realize is like, that is almost exactly what great investors do, is they listen to a bunch of complex data points and information, whatever, and then come up with, here is how I'm going to express this view. That seems like a perfect AI use case. For sure.

And then the question is, you know, how much of that capability is just generally available to everybody. And then if it's available to everybody, then everybody has it. And then the market, you know, goes away. And then the question is, what's the difference between having

like general intelligence and like specific intelligence? How do you think it plays out? Like, we know that obviously the model labs for some period of time, they have access to the best model that isn't yet publicly released. Forget for a second, the government,

even like approving who gets it, who doesn't, right? Just like in general, they got to test it before they release. We also have heard reports of maybe some of the less popular large models like basically trading in the market with these models um does that seem like a path for these guys is like to like hold back the best models for themselves well and to do what with them is that yeah like i and like trade the market with them that you could just turn uh the labs ssi ssi i think is the one that is supposedly rumored to

have done this and i'm kind of torn right like on one hand if you can navigate the market like that is the ultimate sign of intelligence for sure so okay that would mean the model is really valuable but like if everyone's just trading against each other with models like maybe there's not as much alpha so like maybe you should go sell to cyber security companies or whatever right uh yeah and it's you know that's i i guess a a point you're making is like i think we're just at the beginning of understanding how superintelligence goes to market in what package, right? Because like right now the labs are selling the models as APIs or as consumer applications, but you can also internalize them, right? So like you could, in theory, develop a frontier model

and not sell it and just turn it into a trading fund or something else, you know, where like it's a little bit like Google, you know, channeling back in the day, channeling all of its data into making its own product better. And we haven't seen any of the big model companies go this way, but it's not inconceivable that there could ultimately be a better business model for developing super intelligence than the ones that we're familiar with today.

I also think there's another way to come at this. The model companies, their business right now is making the model, right? If you look at, in the crypto world, like Robinhood, the Robinhood chain, now is all the rage, everyone's talking about that. If you look at Revolut, Revolut went And they actually trained a foundation model using much of the data they have.

Like, it does feel like everyone is always talking about vertical integration as, like, model up, you know, to the application layer. But, like, Revolut getting into the model game was not something I think a lot of people were thinking about before they announced that. That's, like, the opposite direction. Going down from the application layer. Correct. Like, they just have unique data, right? And if you have that big of a footprint with that much user, you know, coverage and that much data, I think it's potentially possible to do that depending on what type of model. I don't know how broad their model is versus how narrow, but presumably they have enough data to train.

And I think any really scaled application company is going to have enough data to train models to do certain things. And then there's a question back to your point at the beginning about the pace of generalized models versus specialized models and how those continue to shake out. But I think we'll continue to see application layer companies go down and all to the model layer and also the sort of big labs like go up to the application layer and we'll see where it all lands. When you're talking to founders, one of the big things I hear people talking about is I don't want to use XYZ model because I'm like giving away my data. I'm giving them whatever the zero data retention policies like do they work, not work. There's a lot of questions right now about are you essentially handing over this really valuable resource in exchange for the model? What is the conversation you're having with the founders, and do you guys have a specific point of view? Yeah, I mean, this blew up in the news over the weekend with Alex Karp and Palantir and everything else.

He has a tendency to be able to get attention. He does. He's really good at that. You know, I think certainly for starters, everybody is using the enterprise plans with the best retention data policies and assuming those are the best they can get. And I think a lot of founders are parallelizing across different models and not necessarily putting all of, you know, their data into any, you know, all of their experience or their data streams and traces into one place. And then, of course, there's a whole stack of data that never reaches the actual model itself, right, which application layer companies can uniquely see and learn from. So I think at the stage that we invest at, which is the seed and the A, like very early stage, companies are – most founders are mostly focused on capability, performance, user experience. Can they deliver something that works?

And then I think as they scale, they tend to worry more about these structural problems of, you know, making sure that they're sort of protected as best possible, which is why you hear more about it from the biggest companies and less about it from the early stage startups who are just trying to get something that works going. Blowfin. They built a trading platform focused on fast execution, deep liquidity, advanced futures products, and a really, really clean user experience. They don't overcomplicate things. You can trade hundreds of spot and perpetual pairs. They integrate directly with TradingView, and they've built a platform for both active traders and people who are just getting started. Security is a major focus, obviously. Assets are custody with fireblocks, they provide proof of reserves, and they've built a strong reputation for platform stability during those high volatility environments. And to celebrate the partnership with us, Blowfin is

now giving away $100,000. That's right. Blowfin is going to give away $100,000 in deposit and trading rewards. So if you're looking for a better place to trade crypto, go check them out today using our link in the description. Again, that's Blowfin, a trading platform built for serious crypto traders. Go sign up, pay attention because we're giving away $100,000 in deposit and trading rewards. Today's episode is brought to you by Simple Mining. Bitcoin mining has a reputation for being complicated, risky, and hard to evaluate as a real investment. If you're considering mining in 2026. What actually matters isn't headline profitability. It's uptime, repairs, and whether the operation is run like a real business. That's why I've been using Simple Mining.

They're based in Cedar Falls, Iowa, and they run a white glove hosting operation where you own your miners. You choose your own pool, and you have Bitcoin sent directly to your wallet. They were featured on the Inc. 5000 list as the fastest growing company in Iowa with over 40,000 machines under management what stands out to me is execution they have the number one rated asic repair center and for the first 12 months repairs are included if mining margins get tight you can pause with no penalties and if you want to resize or upgrade your fleet there's a marketplace to resell equipment instead of being stuck to help people think it through whether mining actually makes sense right now they put together a short resource called the 2026 bitcoin mining blueprint

it walks through the five mistakes investors make when allocating to mining and they also explain how to avoid them before deploying capital if it sounds interesting to you you can get it for free at simplemining.io slash pomp that's simplemining.io slash pomp go check it out today and see if you should get into the mining game take a step further like one of the things we've always been tempted to is like oh we'd be an amazing data labeling company right like you know all these people you could have them opt in you could pay them some percentage whatever we're probably not going to go down that path but you could see all these companies they're like wait a second I'm generating tons of data, and especially if I get my users to almost like a YouTube rev share.

Hey, if I go monetize this data, you get a piece, whatever. What is the conversation with some of these businesses that are like, they're building AI products, but I know the model labs are calling companies saying, hey, we'll pay you a ton of money. Is that a business? I think it depends. Okay. You know, you saw the first wave of this happening with all the UGC content, like the Web2 era UGC content companies seeing selling data to the labs as a potential adjacent business model. And I think that can work. This is like RedEd and Stack Overflow and folks like that.

And then today, there were the first round of big data generation companies like Scale AI and others who uniquely built a business model around selling data to the labs. And I think that has worked pretty well for a category of companies. Within that category, there's a question of how much it becomes commoditized and flattened as a business model versus really defensible. We have a portfolio company that does DNA synthesis and, you know, uses like a really patented chemical process to generate data for folks building biomodels. And, you know, that's an example to me that feels more defensible and like specialized versus just sort of doing work out in the world. And then I think we see all the time companies in the middle, like new startups that are getting out and building something. And, you know, the question is, should I take my product direct to consumer, should I internalize and verticalize, or should I, you know, sell data or something else as a service to either one of the big labs or some other incumbent? And I think we're generally, just as USV, less excited about that category because it feels like you should be looking to capture the most value you can.

And depending on how defensible your kind of your spot is, selling data to the labs, you know, probably may not be the best option. So I think it really kind of depends on where it sits. I saw somebody online say something. I think it was a guy, Justin Welsh, said if you only get paid a flat salary, so you have no equity, no upside, no anything, somebody has figured out how to monetize your skills better than you have. Yeah. In a way, the data, if you're selling it to the model lab, have they figured out a better way to monetize it than you? Like there's a parallel there. Totally. And then you could also argue, you know, one layer above that, you know, those labs are selling inference to, you know, developers who are building applications. Maybe those developers are, you know, capturing the value on top of the inference. So maybe, maybe, maybe, you know, and it depends. Yeah. Yeah. How long do you think we need until we have some of these answers? Because I think you're very, I think, open eyed about like, as you guys are working through this, just in talking to you for a little bit, there's a lot of problems. It could go a couple of different ways. I'm assuming as a smart venture capitalist, you're going to make bets in, you know, multiple potential outcomes. Do you guys have a sense like, hey, five years from now, we'll likely have answers to some of these questions, 10 years, one year? I mean, I think it's going to be a little while. I think five years is a decent time horizon.

Things are changing so, so, so fast. But what would we want to see in five years, five years from now? I think we will see a number of massive public companies in the consumer AI or the business AI space, whether that's health, legal, finance, that aren't the big labs, right? So or or industrials or whatever at the application layer, basically. And so many of those companies are being built right now. A lot of them have really robust medium sized valuations. I think we're going to see a lot of those like mature. Then I think there's a enormous category of what you might think of as middleware companies,

like companies in the middle of the agentic stack, whether that's memory or harnesses or orchestration or payments or whatever. A lot of those are kind of just getting going as agent orchestration is, you know, becoming more of a thing. They're all going to become more important. How many of those become like really big, important companies? We're going to see that happen over the next five years. And then even a layer below that, what does the landscape of actual models look like? Is it open-ended and anthropic plus some open-source Chinese models and that's it?

Or can we develop open models here in the U.S. in other ways? We have an investment in a project that's trying to do decentralized model training. in an open way. There are a bunch of projects that are trying to do some versions of that. So I think what I'm saying is in five years, we'll have a clearer picture of the model layer, a clearer picture of the middleware layer, and a clearer picture of the application layer. And the Rebel Alliance thesis posits that there'll be big, important companies at all three layers that aren't the big labs. And this is not to say the big labs are not amazing companies He's building incredible products.

Like, what they're doing is magic and just insane and inspiring. It's an and, not an or. Now, you talked a little bit about, like, open source versus closed source. I think there's also this, like, America versus China component to it. Sure. It feels like, to me, and maybe I'm missing some, but there's, like, a performance component. So, whenever, like, an open source model comes out and people are like, you know, which one's better? There's a cost component.

And then there is like a data security component of like, I don't want to give my data or I can self-host or whatever. The fourth one, though, which I have not heard a lot of people talk about is what I'll just kind of call like the philosophy of the model. Yeah, I knew you were going to go there. The personality. Yeah. And just like, and maybe if I just make it really concrete, and as somebody who's building something in finance in particular, you could almost think of like the West is built on capitalism. Sure. And the East is built more on like a socialist, communist style kind of thought process.

Again, huge generalization, but let's just use those to illustrate the point. If you have an American model, whether it's open or closed, it kind of comes with certain ideals, certain perspective, certain risk-taking is part of the culture. The East model may not have that. It may be more of a socialist type approach. And so even if you have all the security, the performance, the cost, if the answers, the weights are different. For sure. So how do you think about that type of stuff? I think it's yet another vector of competition and choice.

Yeah. And this is like in the last week we started to think about this, right? This is not something I've been thinking about for a long time. Yeah. So it feels like it's like a wide open thing that people just haven't really come to a conclusion on. Yeah. And it's going to take us a while to understand these nuances. You know, we have another portfolio company called Cradle that does these very nuanced model challenges to try and tease out the personalities of the models.

Oh, interesting. They just published a report a couple weeks ago about which models lie to you the most, whether it's Grok or Fable or whatever. It turns out Fable, in their first test, lied the most. Anyway, so it's just, I think these kinds of questions are nuanced and are not covered by your standard, like, software engineering benchmarks, where I think so far it's just like, how well can the agents code? That's more or less been the most important question. As we have agents doing more, you know, different kinds of things up and down the stack, you know, these questions of its personality or philosophy are going to come up and we're going to have to figure out ways to test them and understand those contours. And then, you know, depending on what you're building, you may want a model that's more eastern leaning or more western leaning or more woke or less woke or whatever, you know, and I think we're just starting to figure that out. When you start thinking about different use cases for different things, like, in a weird way, I was having this debate with a friend of mine.

Think of, like, therapy. You know, my, like, general framework, I think, is, like, if you want a child to learn, you give them one-on-one tutoring, right? If you want somebody to be healthy, like, personal, private, you know, healthcare, whatever. We think finance, you get, you know, personalized experience through Sylvia. But therapy is, on one hand, you can make a strong argument. Like, actually, people just need tough love and, like, you know, shut the hell up and go out there, which I think, again, you know, doesn't – it's not a one-size-fits-all solution, but there's plenty of people who would make that argument. There's also a huge population of people who would be like, hey, we need to be more empathetic and all this stuff. Well, like, kind of the model you choose may have different, you know, perspectives.

Well, the model and the application on top of it, right, and model and the prompt and the system. But you could help somebody, or in that case, you could hurt somebody. I mean, listen, this is why competition and experimentation are good things because as we – and also it speaks to the concept of the Rebel Alliance, right? Like, I don't want there to be a single AI therapist called ChatGPT. Like, that's not a good world for us to land in. 100%. We should want to live in a world where there are, you know, lots of them with different philosophies and different approaches and different underlying, you know, frameworks and models. And, you know, people can, you know, try them all out and move to what feels right and see the results and have the results speak for themselves.

And so I think, you know, that's like a core piece of our point of view on all this. And it's about the models themselves and it's also about, you know, how we build on top of them. Because of the Rebel Alliance perspective, do you guys – are you less interested in the large models? Are you less interested in like public – I know you guys are venture capital. But, like, how do you think about if the Rebel Alliance is the right thesis, which I agree with you, does it box you out from certain areas and maybe other people are investing? No, not – I don't think so. I mean that, you know, the whole point of it is that this is so big. It's just unimaginably big.

That's really the core idea. And it's going to – the answers are going to come from everywhere at all layers. And the big labs are hugely important because they're innovating in the models, they have massive consumer reach, and also they're going to be the first to land in the public markets and they're going to get valued on fundamentals before anybody else. And so they matter a ton. And when OpenAI and Anthropo go public, you know, we'll start to get real analysis of, you know, their business models and their margins and their ambitions, you know, and that's going to ripple back, you know, through the whole venture ecosystem, which it always does. As far as where it lets us play, I think it lets us play kind of everywhere. And, you know, I mentioned the model layer, the middleware layer, and the application layer. We're looking at all three.

We're also investing at the energy layer. um, because energy is an input, you know, across the board for, for all AI. And we've been investing in energy for the last, you know, six, six or seven years now as, as part of all this. And so, no, I think there's no shortage of areas for us to invest. And, you know, it's it, the, the big labs and the big models make up a really important part of what I think of as the infrastructure of the next internet. Um, and then we're going to build everything, you know, next to them and around them now the last part that i always uh kind of think is like okay let's keep kind of going down this path is um there's an entire group of people social uh belief that this stuff is

negative for humans so we see data centers local communities we see the like job displacement you know concerns um i can go on and on about all the different arguments yes are they right are they wrong? Is there nuance? How do you think through maybe some of the like non-technical impact if this is as big as you think it is? Well, there's no question that it's as big as it bigger than we all think it's going to be. And so that makes it hard to reason about. And I think the other thing that makes it hard to reason about is the rate of change is extremely fast. And so humans are not good at reasoning about exponential change, uh, myself included. Many humans may not be good at reasoning at all. I'll put myself in that

bucket. Right. Um, and so I do think it's going to be really big and I think it's going to be really big in lots of unexpected ways. And, um, you know, I am as somebody like, like yourself, who's close to using the technology today, you can feel, um, the magic in it. And so to take the optimistic side of things, are we going to get exponentially better at solving hard problems, whether that's, you know, making investment decisions or, you know, finding discovering new drugs, you know, discovering new forms of energy, doing, you know, things in ways that unlock efficiency and make things cheaper and more accessible? Like, I think all of that's going to happen.

And those are all things that we want as humans. And at the same time, it's going to be massively disruptive. It's going to be massively disruptive in terms of jobs. Already, you know, the sort of data center energy thing is both a tradeoff. I think, you know, data centers are the biggest, you know, new consumers of energy. They're putting a lot of strain on the grid. Their, you know, communities are revolting around them. There are also real economic questions around that, which is, you know, what's the right way for, you know, not just on data centers, but in AI broadly. You know, if you believe that a lot of historical jobs are going to go away, you know, what's the right way for humanity to benefit from this? you know, in some sort of ownership model or some other, some other way. These are questions that we don't really have good answers to right now. And I am an optimist because I just am. And I,

and my job is to, you know, imagine the good possibilities and invest behind them. But I am absolutely also, you know, fully real around the unknowns and the scary things that can happen. And also the, you know, I think with any new technology, whether that's like the steam engine or the automobile or the internet or AI, you know, there are these waves of implications, right? If you just look at something like the automobile, you know, you think of it as a way to get from point A to point B, but you don't think about it as a way that land is going to get developed and, you know, business are going to develop and energy and the environment and the internet, same thing. You know, there was sort of the optimistic first phase of the Internet. And now we have, you know, all the implications of, you know, social media on mental health and politics and all this instability that can kind of come from it. So that that's the pattern that keeps happening. I think that's a pattern that's likely to happen here, too, possibly even faster. And so I am honestly like equal parts excited and scared. when um when i think about some of the job stuff a personal experience i had is uh went to the doctor with my wife uh she's pregnant time doctor told us some information we basically double

checked it with the ai turned out the doctor was wrong went back the doctor was like oh my god i made a mistake i'm so sorry whatever right so like there was this like very personal experience that i had and it made me think that one somebody should create double check your doc.com just like you know dollar per per check or something and like you'll have a great business um but second was i am not of the belief that like doctors are going to be replaced by ai sure maybe there's like you know reading x-rays or something right but i do think that there's something about like not augmenting just the doctor but like augmenting the patient who's talking to the doctor sure and i think people kind of already do this right they're like i'm

gonna go buy a car let me go and talk to the ai about like what are the things i should care about or whatever and that feels very underexplored as to like there's like a displacement argument but there's also like a consumer protection argument yeah well it's like personal superpowers right so everybody gets super intelligence whether whether that's entering the doctor's office or entering the car dealership or you know dealing with a legal you know situation you know we use a lot of ai to support our legal decisions but we still work with really talented you guys use lawyers Outside counsel, you know, and I think we're going to keep using both. But I, you know, because of at least today's AI that I have in hand, I can go to our attorneys a little more prepared and a little faster. That's just the same way that you can go to your doctor, you know, a little more prepared. So that's like a today thing, I think.

And I do agree that that's an optimistic, empowering, you know, view of the world. And maybe in a couple of years, you know, we have eyeglasses or an earpiece or, you know, neuro implant, you know, that gives us that augmentation in real time. And I don't even know what to make of that. This is one of my favorite questions to ask friends right now. How many people have to have successfully gotten the neuro implant and safely, it's understood it was safe, would it take for you to do it? Yeah. 10 100 a million 10 million less than a million less than a million okay probably yeah so like more than 100 more than 100 less than a million but like it's kind of interesting to think through

you're like uh the reason i started thinking about this um a long time ago was just like you know that okay here's the first one how many yeah i saw a piece of news recently that neural link supposedly is on 26 of them that was a bigger number than i thought that's i i wouldn't have been able to guess that number that is a real number right so like okay we're at 26 again i'm not gonna go do it but then you start thinking like well if i was like a chess master and you know like give me some advantage like again you know then they're gonna outlaw this is the new steroids right it's gonna be the enhanced games yeah exactly the mental games i'd feel better and i i don't know you know this is not something i've really gone deep on but i'd feel better with

the try before you can buy, try before you buy approach, you know, if possible. But it's got to be coming, right? And I think, you know, I'm not really a sci-fi guy, but you kind of have to go that far now. And that can take you both to really exciting, you know, places and really dark places. And everybody's going to both of them, you know. You need both. You have to get both. In a weird way, you don't get the benefit without the scary potential downside.

and I think a lot about similar to how you have to be an optimist in the private market it's like impossible to make money as a pessimist you also as anyone involved in technology

you have to understand there's these trade-offs and like do your best to kind of bend it positive if you can yeah all right

where can we send people to find you or read more about the Rebel Alliance yeah sure I'm nickgrossman.xyz and you can follow usv at usv.com

all right and then that's where they can read Rebel Alliance there's blog.usv.com amazing all right thanks for doing this

all right thanks Anthony