CoinFund CEO: The Decentralized AI Bull Thesis in 2026 (Massive TAM)

Transcript

Ladies and gentlemen, welcome back to AI Supercycle. This is episode number 37 powered by our friends at Near. Visit their near.ai products. [screaming] Learn more about private inference and more. I'm here with Jake Brookman. Jake, welcome to the tower. >> Thank you. Thanks for [cheering] having me. >> It's good to have you. Jake is a founder and CEO of Coin Fund, one of the earliest and longest standing funds in crypto and digital assets. Now more predominantly also in AI, other sectors. I'm sure you're dabbling in robotics perhaps. Um, computer scientist, grew up in St. Petersburg, Russia. Interesting background, Jake. Good, good to meet you here. Um, and let's have a conversation about investing in AI. NASDAQ's down 3% today. IPOs are coming out. Mythos

released tomorrow. There's a lot going on. Um, >> how are you doing that? >> Doing well. Well, I just want to clarify. So when we think about investing in AI at Coinfund, which is a historically now for 11 years has been a crypto firm, we're literally we're really thinking about the intersection of AI and decentralization. So I just want to be clear that the investments that we're making are very clearly in that space and we don't really do pure AI plays but that intersection I think has some of the most interesting highest riskreward bets in AI and it's also advancing the state-of-the-art of AI itself.

>> Yeah. Yeah. I mean we've we kind of had these iterations of the AI times blockchain right there was like this early decentralized inference decentralized training kind of like the Akos era and then we moved into like this Salana memecoin attached to a Twitter account AI agent era and we kind of enter like this agent era now we're now we're in like this Venice is this really cool private uncensable chat GBT era companies like Ant Seed and other kind inference-based uh token production companies are are are getting a lot of traction and now like entering into like the agents using uh blockchain era I feel like of the space. I know you're

very very bullish on the decentralized nature of bringing these kind of AI suite to the market. Can you just give us your highle thesis for this convergence and kind of as it pertains to this evolution? >> Yes. So, first I want to say yes, we have been following this space for a while. I would say like our first investment in the space was Sam Alman's what was known as World Coin back then and world now. That was December of 2020. In January 2021, I think the deal closed. >> Um, and then in March of 2022, we made our first investment >> in really decentralized training with our investment in Jensen AI. So, when we launched our decentralized AI thesis, it was September. It was published in September of 2022. This is before chat

GPT came out >> and um it is the same today that it was in September of 2022 and that is we think that decentralized networks can start to disrupt the supply chain of how AI products are made and the core problem of being able to do that is to train AI models in a decentralized network. Now what you know what would you guys tell me if I said listen you take these processes that today are in these giant state-of-the-art data centers that some companies are building nuclear power plants to power that contain industrial strength GPUs whose who have hundreds of gigabytes of memory

terabytes per second of throughput and like this is what it takes to create a giant frontier model and what if I told you that I'm going to do the same thing by going on regular 80 megabyte per second consumer internet using people's gaming devices and even their MacBooks. >> You would say, Jake, that sounds crazy and it's probably impossible. Yeah. >> And four years ago, that is exactly what the AI experts in San Francisco and all my AI friends from Google told me as I was making these investments. Um however four years on not only did we show that this is possible we have a number of companies that are now essentially in a

race where they're creating bigger and bigger uh size models in this way. I think the the latest just from last week is 100 billion parameters from a company called Macrocosmos and Bit Tensor. And this is really like heading for certainly commercial viability but maybe even like frontier grade size >> at the end of the day trained with these decentralized methods on hardware that is commodity hardware that wouldn't be able to be used in a data center if we wanted it to be.

>> Yeah. Um, so just completely incremental compute in a market where more than ever before we need to counterbalance the centralization of power game that these big companies are playing >> and in a in a market where no one can get access >> to compute um and can't do so cheaply. >> So that's the thesis that's been playing out over the last few years. And and a part of that thesis I I I think that is going to end up working in your favor is likely this also this kind of desire to go into local models and and running local instances of an LLM at home or at your workplace, right? How does this how does this trend towards these localized

usage of LLM? How does that play into your thesis of kind of distributed compute, distributed intelligence? So earlier this year we saw a lot of people dabbling in openclaw. These days people dabble in Hermes agent right at the same time almost by accident there's a hardware trend that is putting more and more powerful GPUs on people's desks. Notably this is being done by Apple with their M5 uh suite of uh chips. And literally right now you can go and buy a MacBook Pro in the store and that will have a powerful enough chip for you to run a really significantly sized model like very fast. There's other companies other than Apple who are kind of playing

in this hardware niche where you know a proumer can put you know essentially like a you know data center grade GPU on their desk and among proumers and among enthusiasts we're seeing that people want to um inference locally they want local type privacy to their data and also they want to run agents that are that are kind of sovereign to them and go around making them money and and and doing all these things. So, I'm very bullish. Like the other kind of side of my decentralized AI thesis is we're also seeing the rise of these like agentic swarms. These swarms are starting to optimize problems. They're solving like previously unsolved math problems. They

are um optimizing Google's like quantum circuits to break ECDSA. >> And very soon they're also going to be doing inference. They're going to be doing training. um and they're going to be running your personal agent. Now, you know, the the kind of the better um the better models we can kind of provide for those agents, the more of a shot we have of collecting a bunch of people's data and then putting them in putting that in a model that they all collectively own and not necessarily like sending it to Sam Alman or Dario Amade at Anthropic.

>> Yeah. Yeah. And and so these swarms you talk about this evolution we've we've heard you know this term talked about and and we've understood like these you know agentic collaborative abilities you know you know whether it's a team for a company or a signing task or or like a a whole unit of a of a company or a team. What has been the agentics like like how does the swarm uh concept tie into the decentralization thesis? Like h how how are those two at all uh you know like correlated?

>> Well, I I I think the agent in a sense is some of the most decentralizing technology, you know, that you can have. And so if you can imagine a world like everyone in traditional VC and in web 3 VC and in cryptovc and in AI they all kind of agree it's like super consensus that there is going to be a lot of agents. There's going to be billions and zillions of agents and those agents are going to be acting on your behalf. They're going to be doing commerce. Now if you imagine a world where like every person has a personal agent for example or a significant number of people have a personal agent that are solving problems on your behalf that are doing shopping that are performing commercial activities that are investing on your behalf doing finance on your behalf

finding investment opportunities etc. Then this requires a high degree of essentially decentralized tooling. It requires this, you know, agents to be able to pay each other. So decentralized rails for payments, for settlement, you know, for authorization, for compliance. It requires um coordination among agents. So, we've [snorts] seen this amazing example the other day from IGEN Labs where Google uh put out this paper about quantum circuits and they said, "Look, we could we're now so much closer to solving with a quantum computer when we when we get there, you know, breaking ECDSA encryption." And so, we're Bitcoin, you're on notice that, you

know, you're going to get hacked very soon as soon as this happens. And so, they advanced the state-of-the-art, but what they said was, "We're not going to tell you what the scientific innovation actually is. we're going to censor it. We've been talking to the government. And about a week later, a swarm of agents came together and they worked on this problem. They kind of concentrated the agents to think about this problem. And what they then did was they discovered the same kind of optimization that Google did in about a week. And then in about a week later, they were 40% more advanced than Google in their paper. And now this thing is open source, right? So like you could really look at that and imagine, wow, that is

an incredible like democracy technology that is like opening up these things, making them available. But it also needs incentivization. It needs like the problems that it needs to be solving, need to um you know, make the agents money. And once that happens, I think there's going to be a ton of incentive for people with personal agents to put those agents out there in the world. Now, my point about um you know what why does this necessitate web 3 technology is to concentrate a swarm of agents on a problem, >> you really need to give them sort of a shared context, a shared memory.

>> And there's some startups starting to come out into the market that are doing that, right? that are enabling swarms of agents to collaborate, you know, essentially using decentralization technology. So, if you ask me, when there's going to be a ton of agents, they are going to motivate more decentralized tooling than has ever been motivated before. >> Yeah. You know, I I I think in that context, there's been this kind of like market awareness expansion as of late about the uh bullish investment thesis or the understanding of the value of inference specifically when it comes to the stack. Like a lot of people were very bowled up on okay, the compute

demand is just going to keep going higher, build data centers, the models are going to keep getting better. Like just like you know, that's where a lot of the value is. But now there's been like a growth of companies doing a lot of work just on inference and making inference as as either private anonymous or just better. Um why do you think this has happened as of late? like what has been the like the light bulb moment for the market understanding that inference is something that is very very valuable in the AI stack and and there's also now you know there's companies like ANC there's a you know couple others that are doing like open marketplaces for inference as well on chain how do you like why do you think this happened why

did it take so long and why is it valuable yeah >> yeah so so if you take a look um at kind of the broader AI market it is unquestionable Almost every dollar of revenue that's been generated this market has been generated by inference on desirable foundation model. So it's not being generated by like hey we'll give you training services you know or something else right consulting. Most almost all the revenue that's been generated has been generated by people hitting anthropics API open AI's API. How much money is that? That's about 6065 billion dollars of ARR right now.

It's about 30 billion from uh Enthropic, about 25 to 30 billion from OpenAI. There's a long tail there. Some analysts are saying just Enthropic can have 100 billion of AR by the end of this year. And the five-year outlook on that is something like 500 billion to 1 trillion of inference revenue, you know, by 2031 essentially, right? there's been $600 billion dollars invested by the big companies in AI infrastructure to make that growth, you know, kind of happen.

So, I think one of the reasons that people are interested in inference is because they sense that like this is the core modernization of AI. This is an incredibly big and growing market. >> And when I have seen now, let me let me take a web 3 perspective. So historically when I've seen inference networks that hasn't really made a ton of sense to me because the inference game to the consumer or to the customer is about making inference fast and cheap. It's about like fast cheap compute economies of scale you know and when you try to provide inference on decentralized network you're actually paying more for that because you got to pay for the decentralization. And so for a long time it didn't make sense. Like

how can you make something that's more expensive like actually cheaper, >> right? >> And now there's been some interesting solutions to that which is hey by the way we have this blockchain that when you're doing the inference those same matrix multiplications that are happening to get that output are being used to mine cryptocurrency. And this is a concept known as merge mining historically crypto. And so there's some view that like, hey, if this cryptocurrency can subsidize the inference, we actually can make the inference cheaper. So inference looks a little decentralized inference, I should say, looks a little bit better in 2026.

>> Yeah. Yeah. And and so the market like the market came to realize that that was the big revenue generator for this AI stack. U obviously it's very it's very concentrated in these in these names. Is part of your thesis just that the the the just if a couple of these decentralized inference providers or decentralized training providers just get like a little bit of of the current existing market share that is just a massive 100x from here for these companies like how simple is the math on like just chipping away at bits of the broader market here.

>> I'm so glad you asked. So this is why there's a subtlety here. So in decentralized training companies like one of the questions is what are you training? So a lot of the decentralized training processes that have happened to date essentially produce open weight models where everybody knows the weights and the problem with that um state of affairs is that openw weight models are very hard to monetize. This is why you're seeing in Chinese labs today that you know Chinese labs and meta have historically been the companies that have produced the most openweight models. Yeah, >> you're starting to see a trend where yes, we have a bunch of openweight models, but they're starting to close source kind of like the training setups and they're starting to close source like the really really good models

because these models cost tens and hundreds of millions of dollars to train, but then once they're out, nobody's like paying Meta for that. Nobody's paying the Chinese labs for that. So, they do need a business model around that. [snorts] Now, in the decentralized training world, you've had these processes that have produced open models that are like hard to monetize. One of our one of the companies in our portfolio called Plurales has an answer to the question of like how do you monetize these things and in their algorithm what happens is the model gets broken up among different uh you know GPU providers in the network and so there's no individual or company or government sound familiar?

>> Yeah. >> That owns the entire weight set of the model and so the model ends up living kind of in the network. the network like if you want to use it, you have to pay the network to do an inference. And so the answer to your question is with this type of monetizable decentralized AI, if we can win like 10% of training, we also win the inference associated with that training >> and that looks to be you know something like a 50 to$100 billion market over 5 years. Yeah, >> there's only one company in the world today that has enough technology uh to to implement a model like I just

said. >> And when it comes to the weight being o like open weights, right? Like why is it important that why is it important that there's not one entity that controls the weights behind a model? Like could you just unpack what that means a bit more? >> Yes, of course. Because if if there's if if someone has the full weight set then they can like do with the neural network whatever they want. They can optimize it. They can quantize it. They can run it you know on their MacBook >> and just inference it for free. If you don't have a s single entity that has the entire weight set, it means that um the collective of people who do can charge a fee to inference the model

because the network will require them to put up their weights for the model to be inferenced, right? And so the difference between an open model and an anthropic model is like here the the weights are totally open. Here the weights are totally closed. What I'm talking about is a model whose weights are closed which gives it a business model, >> right? >> But it's publicly owned. We can be we can be a a token holder and we can receive the revenue from each inference. >> From the inference. Yeah.

>> And and Okay. So if we get the training, if we chip away at the decentralized training part of the market, that then guarantees that we also get the revenue from the inference. >> Yes. Because if the model is trained in the decentralized network, then the customer has to pay the network to get the output. So like Akos, render, bit tensor all these kind of existing decentralized training networks their MO is to grow their adoption of their training because then the revenue from the inference that will be you know can come back to them in some in some fashion like for example this macrocosm company that that you mentioned doing 100 billion parameters sets for some of

their models. Their goal was to get as much adoption of this of this of their training network to get the inference revenue. >> Yeah. I think I think like the [snorts] the general MMO of decentralized training networks is we want to scale these processes to at the very least commercial viability. You know, I mean that can happen like as low as I would say like 50 billion to 70 billion parameters. But if you want like a really smart model, you know, you can go to a couple hundred billion. And like kind of the the the ultimate goal is to get to the quoteunquote trill, right? Um and so yes, the I think the long-term game here is for these networks to

source enough compute from from the world that they can actually start to be truly commercially competitive with things that Anthropic is doing, that OpenAI is doing. And we see that today like if you look at you know the the average commercial application of AI like you're some kind of you know consumer application you're a language tutor which has an AI backend like chances are most of your inference is running on open models. That's because they're customizable. They're 90% cheaper than the frontier models and they're like 99% as good on the benchmarks. And so the goal is for uh these decentralized models to get to that point but be a little bit more

expensive because they're actually providing like a better you know a better model overall a lower cost model etc. >> Uh kind of a bit of a tangent here. Have you seen the the goaround for this new AI kind of blockchain company called Pearl? >> Yes. >> Okay. Could you explain what Pearl is to us? >> Yeah. So, Pearl is sort of what I alluded to previously, which is like a clever like uh useful proof of work, proof of useful work, excuse me, blockchain which enables merge mining of, you know, AI computations. And so, it's it allows the cryptocurrency mining aspect and the fact that cryptocurrency is valuable to subsidize the AI computation. Yeah.

>> And actually make uh the like for example decentralized inference cheaper for the customer. >> Okay. Got it. >> Yeah. So, you know, we we were speaking about that earlier that, you know, that thing's getting, you know, sent around. I think there's like an OTC for it right now. It's it's pretty illquid, but, you know, people have been asking me about that quite a bit. >> So, so like the the um this this thread of like can decentralized things be cheaper is actually like a really important and nuanced one and I wanted to so this is like a great story like with decentralized training. So like initially as I said people like decentralized training is impossible then you had a bunch of companies kind of come to market and say well it's possible because here it is we did it we used like a Google algorithm to do it

and then we created new ones and we're now published this is now a a field which is being published in you know the biggest machine learning conferences in the world like Nurups and then people said okay you know what fine it's possible but you know what it's way too expensive because it's on these like dinky devices. And actually the reality of that is the opposite. >> Really? >> Yeah. If you're a customer who wants to train a model, like a big one, right? And by the way, we just had a news item that Kirkland and Ellis big law firm is devoting $500 million to train their own proprietary model. Not to use Harvey or some other solution in the market, but to train their own model on their own

proprietary data. So, you know, what we actually see is that a customer like that either has to build their own data center or go rent one. But the difference between a data center and a swarm of consumer commodity GPUs is that about half the cost of the data center is facilities maintenance, cooling. Y >> and capex and things like that. And when you go to the swarm, you don't have facilities, you don't have cooling. You basically like forego a bunch of costs. And what is the difference there? How is that possible? Well, it's just because we have innovated on the training algorithm itself. It now works in this context. And if you can if it can work

in this context, then the costs are lower. So what we're actually seeing is that the cost of swarm training at least for the size of model that's possible right now is actually cheaper than if you were to do it in a data center. >> Yeah. >> So first it was impossible, then it was like too expensive. >> Yeah. Then the next uh objection is like it won't scale, you know. So that's where we are. >> It won't scale to like the actual highest of quality of models like right now we're at like the 70 there's 100 billion or 100 billion parameter one in bit sensor but we get to 250 billion 500 billion where which is where these frontier models are going and this and this it'll be too expensive. It'll be

not it you you'll need more compute. Is that the thesis? Well, so another objection is like, okay, when you try to scale it, it'll just get too slow. But like all the practical data and all the practical results from these trading runs suggests that actually um it tracks the loss curves of you know normal training process like sometimes it's a little bit slower but there's like a lot of optimization to do. We have you know this is I'm talking about pre-training right now but there's also many other kinds of trainings that you could do. There's a company called Gradient Network that put out a paper that says when you're doing post training like reinforcement learning on these models, you actually get a lot of parallelization. And when you make this work on consumer devices and you can

paralyze a lot, you can actually 10 times lower the cost of this and 13 times the speed, you know, in a decentralization context really. >> So there's a lot of research that like San Francisco like AI people are just not doing because it's not like their focus, right? that actually makes this stuff in a decentralized uh context very very very competitive. And so if you look at the curve, if you plot the curve of like the number of parameters of models that have been trained in decentralized networks, it's kind of starting to go exponential. It was like 10, >> 40, 72 from Templar just a few months ago, big headline. Then 100 from macrocosmos. Then there's some early

experience that are between 100 and 200 that have been posted on Twitter, but it's not quite a full model yet, right? So you can imagine that we are going to go up that curve. What we're missing is qualified tech talent, AI talent coming into the space and, you know, doing more work. We're missing capital. We're missing more companies doing this. This is like one of the most exciting opportunities for web 3 because it has a hundred billion dollar market. >> Yeah. And so macrocosmos is the leader right now in this space.

>> I would say macrocosmos currently has the the headline number 100 billion parameters trained on >> uh you know a certain strategy. Our company Plurales is the only company that has what's known as model parallel training which is the one that breaks up the model into pieces. Yeah. >> And they're currently in an 8 billion training run which will start to ramp up in parameters over time. That's a much harder problem to solve. Um, and then there's other people, uh, you know, Jensen AI has done reinforcement learning. Gradient network has done reinforcement learning. Um, and news research has done a little bit in like pre-training, but not sure if that's like a key focus area for them right now.

>> Yeah. Yeah. Uh, have you used the the, you know, I'm a tangent here. Have have you used the Hermes agent much? >> Yes, I have. I'm I'm an open claw user and a Hermes user. Yes. >> Okay. And and is is is Hermes as good as they say? >> So Hermes I would say so I have an engineering background. I study math and computer science. Um and you know it's just like a much better designed sort of product. >> Um and I think it's really really interesting to own a harness in this day and age.

>> And what does that mean? I've I've heard harness like what does that mean? >> Yeah. It it it's so you have like the foundation models underneath and the harness kind of like puts them into a problem set, right? Like are you an assistant that's helping me, >> you know, calendar and do my daily tasks? Are you a researcher? Are you a coder? Right. And so Claw Code, that's a harness aimed at coding. >> Correct. >> Open Claw is a harness aimed at being an assistant. There's other harnesses that are aimed at like, you know, doing math problems and solving complicated like research problems. Um, I would say Hermes is kind of, you know, more like general assistant. Yeah.

>> Um, but it's just a the way it's been architected, I think, is a lot more robust and clean. Um, much cleaner than than OpenClaw in my in my humble opinion. Don't get mad at me, Open Claw guys. I know you >> put in a lot of work, but but still I think I think her Hermes is just better designed uh product. And I think why harnesses are important is because they collect user data. One of the most important types of user data in in agentic AI, right, is how your agent traverses the world and the mistakes that it makes and the things that it does to correct m its mistakes. Right?

This is data that you cannot buy. This is data that you can only generate and it is data that is like nearly infinite because you know agents can traverse the world in all these in all these different ways. And so today I would I would think that like you know people like anthropic every time somebody uses an agent on top of the enthropic API enthropic gets this it's called the trace data from the agents. Once you have a big bank of this data, then you can do reinforcement learning on the model to make your agent basically make less mistakes to become a better.

>> So reinforcement learning or training happens after experimentation with a given model. So that's like what you would call post training? >> Yes sir. >> Okay. And is that just is that as simple as like me responding to the outputs and like trying to guide it in a better direction or is that done at like a deeper like lower level mistake? >> No, no. It's it's it's almost I mean like to some extent the user could be interacting with the agent and correcting him. But most of the time the agent is just kind of like out there and it's like oh I'm going to download you know this library and try that. Oops, that didn't work and here's the reason why I should really download should should have downloaded that one for this problem. Right?

>> It does it on its own. It does it on its own, but it's like it takes so much compute >> to generate this. Remember like a couple of years ago, there was this big question. It's like, >> well, we just already put in the internet's worth of data into these LLMs. We're we're out of data. We're not going to be able to progressive intelligence. Now, it's becoming very abundantly clear that the data bank is actually like quite infinite and it's based on these like agentic traversals of the world. So the data so the lack of data is not a problem. Yeah, that that doesn't make sense. There there's infinite content out there.

>> Correct. Right. >> And there will just continue to be more infinite content. >> But now to bring it back to you know web 3 and harnesses. So today anthropic is like the premier company collecting all the traces. Right. >> Right. >> Um when you talk about open models in and of themselves they don't collect traces like they're just open weight sets. Right. And so who really collects traces? It's the interfaces to those models. It's the harnesses of those models. >> I see. >> And so in today's world, the open harnesses like Hermes, they create a very like segmented uh market that is not coordinating on collecting everybody's traces and competing with Anthropic.

>> Right. Right. >> But anthropic is just running away with it. Right. Right. And so one of the reasons I'm super bullish on decentralized training and harnesses is that you can now create a system where like let's say you're running a personal agent like Hermes agent or OpenClaw, you can point it to a special model that you co-own, >> okay? that you own the tokens of, that you own the inference revenue of, and then you feel really good about donating your traces to that model because, you know, you're benefiting yourself and not some corporation, right?

>> And so this is a way that like if enough people kind of, you know, gather around one of these decentralized RL processes and start putting it in their agent and that becomes really big. this is actually a counterbalance to like you know Claude owning everything >> right and and that seems to be one of the things that is like keeps people circling back on the decentralized AI thesis is like there's a lot of like hoops that are being jumped through and like product disadvantages compared to doing it like in the centralized way and this is this goes for everything. This goes for like why is it easier to like press your brokerage account to do a trade than like go do it on a DEX and like have to wait for the block to

finalize. You know, it's the same kind of concept across different parts of the stack. But what you're saying is like we're now reaching the point across whether it's the trace data and the ability of coordination to happen there or on the on the training and the inference side to actually create better cheaper inference. We're we're like we're reaching the point where the ideological blockers are no longer embedded into like the techn like the like they're no longer as deep of technology blockers as they once were. >> Yes. And I would say look I've been you know in the web 3 space for like 11 years. I've seen a lot of networks historically trying to offer compute, trying to offer, you know, storage on decentralized networks and a lot of those experiments have not kind of

panned out and have become like global huge projects, right? Part of the reason is there's not a ton of demand. Like even though we have these great value propositions, like it's cheaper, it's more private, it's self- sovereign. Um, a lot of times like middle market companies are just kind of like, "Yeah, but we have Amazon AWS and it's like cheap enough of course, >> right?" And so there's not like a ton, but I think like >> one thing that separates decentralized AI networks from those kinds of web 3 resource networks we've seen in the past is that there's just an insane demand >> for compute and it's just doesn't stop >> and for intelligence in general, right?

Like it's there's that's just not going to stop, right? this filecoin or are weave like decentralized storage like you know is it it was interesting but it was always that demand question right like who's really going to want to use this but it's very clear that all of us are using this technology in our day-to-day life and I think one of the other promises that kind of web 3 has made is like this idea of like monetizing your own data like I remember everything from like the early IPFS like YouTube in 2017 there was like this like IPFS thing that was gonna you know take over YouTube and then there was get paid for your data. Um, you know, things like metering, inference, these types of things. Do you think that these decentralized AI like is part of the

promise as part of the thesis kind of returning to some of that monetization of user data of your self-owned data like self-owned AI, userowned AI? near's uh CEO Ilia always says userowned AI like things like metering inference or like you know let's say if you have uh extra usage on on your plan being able to share that like are these like sharable monetizable data usage inference is this also part of thesis? Yeah, I I think so. I mean, I think like um the the other aspect is like when people are going to be able to put relatively inexpensive compute on their desk, they're going to

go to sleep at night and their agent is going to autonomously go and provide that compute to different networks and create a return for them, right? Like I I really think that in the same way that we've seen in crypto, we've seen like airdrops and and yield farming and this kind of thing, >> you know, I think we'll see that um GPU compute is also a resource that you can then use to generate income, generate a return for yourself. So, >> but what about specifically on the data front? like on the data like being willing to uh uh sell my data like I'm generating all this data anyway, right?

like like using Venice for the first time was kind of like a mind [ __ ] because I was like whoa like I know none of this is being used to train these frontier models and wow like for you know maybe it doesn't have the best context window for me but at least for once I feel like I'm not just like a uh you know a pawn on the board being monetized for you know me me you know what they say if the product's free then your data is the product right you are the product I is the data front something that you think we'll be able to tackle on maybe centralized AI thesis. A >> absolutely absolutely I mean I think that like again if you lower the reliance on kind of centralized corporations

then you're sort of leaving your data either locally or inside of models that you co-own and feel feel good about. Also like what kind of data are we talking about? You know some there's obviously like a bunch of first what they call first party data. It's like information about me, what I like, you know, what I buy on Amazon, you know, what, you know, what's do I want an aisle seat or window seat on the plane? Like these kinds of preferences that a lot of um e-commerce companies want to know about you to give you like advertising and better products. There's another kind of data that you can potentially see coming online, which is that, you know, I'm asleep in my bed.

Yeah. My agent is doing research on my behalf. He's coming up with new knowledge. He's trying to solve a math problem. He's doing research on the GPU space. He's doing research on what tokens are good in in crypto. Like whatever it is, right? That might be an area that someone else didn't think about, >> but you did and your agent did. And now you have this kind of alpha or like proprietary data that is valuable because you've spent cycles, you know, researching it. >> You're going to be able to put that on some network and sell it to somebody, right? Yeah. So again, I I think there's a huge set of financial opportunities and privacy opportunities um in data, the more like we use agents and the more those agents are

decentralized. >> Yeah. I guess Jake, just to kind of close us out here as we come to a wrap, can you just give us like a size of the TAM again? I know you've you've given us some numbers on the training side, on the infant side. Let's say we get this right. Pluralysis, macrocosmos, aos, bit sensor, near venice, ant seed, all these kind of really early stage but some more developed than others companies working in the space. How big are we talking of a market here? What is the what is the addressable market that we're really after?

>> I again I think that's a on a five-year time scale based on my research. I think that's a 500 billion to1 trillion dollar you know revenue market um based on inference and I think that decentralized AI uh to the extent that it creates desirable scaled commercially viable um you foundation models can capture like I mean even you know 1% of a trillion dollar market like it's still you know still a lot of money right um so I think that's kind on the, you know, on the training side and on the on the inference side. And then we didn't even start talking about agentic commerce,

which we could spend probably a whole other episode on and you have multiple times, >> right? But basically, you know, just to give you some numbers there. Um even today when agentic commerce is definitely like not widespread and and there's like very few agents in the world the number of ma ma of AI assistants is something like 2.4 4 billion people, MAU, monthly active users. >> Um, and the number of like personal agents like OpenClaw, you know, um, uh, Hermes agent, this is more like an enthusiast crime. It's very small. It's it's my estimate is like 20 million or something like that,

>> right? >> But the 2.4 billion number is poised to grow is poised to double in in five years. And this personal agent number 20 million is poised to grow like over 100% kagger over the next you know five years. So I think there's like this huge huge opportunity there and how the value gets captured in that market in that agentic commerce market I think is very up in the air. Is it Visa? Is it payment rails? Is it the discovery layer? Is it the people who decide how your agents decisions are made? Is it like SEO? You know, there's a lot of issues there, but everyone agrees. It's very consensus that that's going to be a huge market as well.

>> Ladies and gentlemen, Jake, the CEO of Coin Fund in the tokenization tower. Sometimes tokenization means real world assets, sometimes it means actual AI tokens. Jake, thanks for your time, man. Cheers to have you. >> Thank you. Thanks for having me.