Coinbase Head Of AI: Agentic Finance Is The Biggest Opportunity In Fintech (Full Explanation)

Transcript

All right, you guys. We're back here at the Tokenization Tower for episode number 41 of AI Supercycle, our weekly premiere AI show powered by our friends at Near. Visit near.com and near.ai to learn more about their private inference product. We've got Lincoln Murr. Lincoln, welcome to the tower. >> Thank you for having me. >> have you, man. You are the lead of AI at Coinbase, X402, agents. All of this is happening fast. Just ended a conversation there with Lucas, the the VP of research at Galaxy. I think he's your friend and he was speaking a lot about what that is and kind of like why this tokenized inference, tokenized compute market is really growing. So, why don't we start kind of here? But but before we do, just quick intro on on yourself. What do you do at Coinbase? How'd you get started

there? >> Yeah, yeah. So, I've been at Coinbase for about 2 years. I lead all the different AI products. So, that's X402, Coinbase for agents, and Coinbase Advisor. I've been in the crypto space since I was like 11 or 12. I got started because I needed to buy a VPN so I could pirate video games and the only way I could buy that is by turning a Target gift card into Bitcoins on LocalBitcoins. I've been involved ever since. Prior to that, I was doing research at the Ethereum Foundation, but yeah, Coinbase is now the new home. >> Beautiful, man. And so, this kind of AI crypto crossover has happened very fast. And I think I I I think we should start with that's kind of like financialization part of it first, which is this kind of growing desire to have tokenized inference and tokenized compute markets. And we've

seen different companies come out of F SF that have launched compute marketplaces and there's this kind of growing demand to basically speculate on on compute prices or hedge your exposure via, you know, if you're a data center owner, etc. And then simultaneously, there's also this push to really reduce the cost of inference. Right now, we're seeing this kind of like narrative unfold. I think today Forbes put out an article that was like like like the headline was like, "We replaced all the workers, but AI costs more." And so, there's kind of like this this drive to lower the inference cost as well. And so, you get these open source inference providers and these marketplaces kind of gaining steam.

Maybe just kind of on unpack your view on the financialization of the AI stack. We'll start here and we'll get into some more of the product side of things. >> Yeah, totally. And in my view, and I think this is something that Sam Altman directly said, which I really like, the whole idea of inference or intelligence is now just becoming a utility. In the same way that electricity or water or anything else is. And part of that process is naturally going to be the explosion of its use, but then also the continual decrease in the cost of it, as well as the ability for people to have better piping and infrastructure through which they can buy or speculate or even hedge their exposure to this new utility. All of these financial primitives, or at least a large part of them exist today for these existing

utility markets, and I suspect we'll see this continue over time. But the great thing about tokenization, as we all know, is that it makes it just so much cheaper, faster, more efficient for these markets to exist and for them to proliferate. >> Yeah. And so from from the product side of things, do you see like marginal improvements from having tokenized compute markets or tokenized inference markets or open marketplaces? Is it Is it a cost thing? Is it a efficiency thing? Like what Like what does it really enable? >> Yeah, I think the biggest one and the most obvious one to the vast majority of people that's that's valuable is just speculation. If you think that the compute resources are going to become larger, there's an ability to speculate on that, where today people can't get access to OpenAI or Anthropic besides some pre-IPO shares of that. It's It's

just a more direct pure form of speculation, which I think is broadly what we've seen in the crypto space recently with prediction markets being a great example of that. >> Yeah. >> Uh Beyond that though, like if you're a large company and you want to replace, you know, X% of employees with AI or you anticipate having a bill for compute over the next however many years, you can use these new tokenized structures and whatnot to very seamlessly hedge against that cost going into the millions of dollars, per se. So it's a nice way for people, in the same way that you have farmers who sell their futures on corn futures markets, for larger enterprises and tech companies to sell uh compute or buy compute to help hedge against issues there.

>> so that kind of utility framing is it's like like the way I'm thinking about it is that these these kind of like these parts of the stack are turning into like like they're being commoditized into like commodities into like tradeable investable assets basically. >> Exactly. Yeah. >> And so that that creates a more fair market that is set by the market rather than by you know one single company Nvidia or the Chinese version of the chip manufacturers or whoever creates the the the GPUs they set the price because you know they have the best tech. Well sure that can you still you know proliferate but we're going to have more of a let's call it decentralized or or or wisdom of the crowd price discovery for what are going to be you know the electricity of the future right? It's that kind of utility based.

>> Yeah, exactly. And I think we'll see this trend start to continue with the general tokenization or you know apparent speculation from potentially going from inference and LLMs into maybe entire agents. I think there's this really interesting movement that we've been seeing around agentic trading that's starting to pop up. >> Yeah. >> And these agents that are running 24/7, they're paying for premium data on X402 to make more well-informed trading decisions which makes them more money and it's this nice positive feedback loop. But I can see that also becoming this interesting space where speculation can occur even if it's on the direct agent itself. Will this agent make money over the next 24 hours? Will it not?

That opens up new exciting markets for people to watch. >> Yeah. All right, well let's get into the agent side from the Coinbase view. You you guys just had a system update event. I believe it was here in New York. >> Yep. >> A couple weeks ago and there's a big push towards this I think you guys call it Coinbase advisor which is kind of like this more agentic framework for you know using a using intelligence to make better financial decisions. Maybe walk us through what your personal view is of how agentic finance will shape will will kind of reshape the way that trading investing people are using their capital now and then we can get into some of the Coinbase things.

>> Yeah, totally. So, I think broadly agentic finance is going to become more and more prevalent until it ultimately just becomes finance. It feels like we're in the flashlight app QR code scanner beer drinking app era of agents right now, just like we had with mobile in the past decade or so. Uh so, I think broadly what will happen over time is we start with these more frontier users and and hackers who have an agent that rebalances their portfolio for them, or I can go in and say, "I want to invest in space. I don't know how. How can I do that?" And over time we'll start to see these features more directly embedded into Coinbase and and other platforms as well. So, that's largely what we're thinking about is how do we support the entire spectrum of users from the most simple to the most advanced to basically

have a team of financial analysts that exist and and manage their money for them. >> Yeah. And and do you think that the that these agents will do better than humans? >> I think there's a couple value props. Calling them doing better than humans is a little bit nuanced, but the way that I see it is there's an orchestration gap that exists among a lot of the users where people have ideas that they want to trade on, but they don't necessarily have the time or the expertise to take some broad thesis like I think quantum computing is going to go up and then implement that into financial primitives, both in crypto, but also across equities and prediction markets and all these different things. But an agent is very good at taking that and translating it into those actual core primitives and actively managing that

position. I've gotten liquidated so many times because I just forget about my positions, but if I had an agent managing it for me, so much easier. So, I think in that regard it will do better because it'll make more people able to express trades uh in a more streamlined fashion. >> Yeah. Okay, so what if we take the human element out of it? And and I was telling you about this thing called bot.fun. It's this new app that I saw that's now the second largest user of X402. Um so, they're really really really growing. That's based on their own kind of uh marketing, but uh I mean, something in the to the order of like, you know, tens of thousands of transactions X402 transactions in a given day maybe or a given month. I don't know what the exact number was.

And so, this concept is like we're taking agents and we're giving them the power to just interact without the human intervention. It's kind of like just I think it's like probably provide them a wallet, spin up the agent, and then let them go and kind of do it on their own. Have you guys seen the proliferation of that on base or kind of like is it still human in the loop a lot? Like when do we take that step of like, you know, here's the money you're supposed to be smarter than me. You have god level of intelligence. Go make me money. Like people want to know in the chat in in our audience like how will this technology help to maximize my my portfolio? When do we get that out of the human in the loop stage? Are we already there? And have you seen any

examples of this? >> Yeah, so oddly enough, we've seen a lot of the fully autonomous use cases come out because this this user group that's very experimental is the same group that's very open to speculation in the crypto markets. So, we're actually seeing more people do this autonomous agent trading than we did with human in the loop trading. And I think the X402 piece you're exactly right is a big part of it because as I said earlier, you give the agent financial independence on the internet. And if you give yourself a credit card and browse around, you can do so much more than you could if you were just, you know, trying to read through things and not have any way to pay. But if you give an agent the ability to have access to money, it really becomes a first-class citizen of the internet. And

that's really what we're building toward. And one of the things we announced at the showcase was a Coinbase for agents. It's like the Coinbase MCP. And part of that is the ability for your agent to plug into its own portfolio inside the Coinbase platform. It can go off, it can pay for things with X402. It can make trades entirely on your behalf. You can just tell your agent to set up Coinbase for agents for me. It'll go find all the documentation and do it. And I think this is kind of the first step toward massive agent autonomy that might take place. But if you're not comfortable with that, we just released the advisor product which is actually regulated financial advice inside the retail app that is fully human in the loop. So, if you still want to be in control, that exists. So, it's a it's

full spectrum of users that are playing around with this stuff and it's really just a matter of how do we get everyone in some way, shape or form using AI with trading. >> Yeah. Yeah. Yeah. I just saw something today uh that Restream, which is like the streaming app that we use, uh you can plug it into to any LLM VMCP and then operate the app from Cloud or from ChatGPT or whatever it is and like see all sorts of, you know, stats and start stream, schedule things. So it's this this kind of MCP concept is like expanding the use case of like all the apps that are possible to use within AI. I want to talk about some more current happenings.

First is Trump and the government banning or restricting the usage of high intelligence models to the public. So Fable 5 came out, they pulled it back. Then they they reinstated it like it's fine. Opus or uh ChatGPT 5.6 came out, they pulled it back. Then today I saw something that Trump says uh that they're basically allowing it now. What's happening here? I- is this intelligence that good that we it's a national security issue? Should we be nationalizing intelligence as well? Is this like a Is it a a a national defense issue? What what's your take here?

>> Yeah, it's fascinating to watch all of this happen and with the backdrop of however many weeks ago the whole Mythos eCash thing as well where like I crypto is in a really unique position because all of these smart contracts are immutable and if there were going to be exploits, if I was a a hacker or I had a quantum computer, the thing that clearly is the biggest uh pot of gold in the world right now is is Bitcoin or any one of these smart contracts that you could just exploit and hack into. So I imagine what will happen and what we'll see play out is that for the very short term, if Mythos and these other models that are amazing at cybersecurity get leaked to the public or otherwise made available through open source means, whatever it might be, we're going to have this short

term period of just complete chaos where all sorts of different projects and protocols are going to see hacks both in Web3 and Web2. But once that initial period is over, we're just going to see these models become more of a defensive security sense. Before you do any sort of publication of a of a new piece of code or whatever, you put through Mythos, Mythos says it's good, and then you're safe, and actually probably safer than you would otherwise be today. Um but, in terms of your question about the uh the the you know, imminent potentially imminent nationalization of these labs and whatnot, I imagine that a large part of the uh the speculation around these models uh performance today is due in part to upcoming IPOs. If I had a massive company and I could go say, "Hey, our

model is so good that the government is trying to ban it." That's a pretty good reason to buy it. It's marketing. So, I I think a large part of it is that, and I think we'll look back on Fable and Mythos as the same way we do about like GPT-3.5, which is an exponential increase in the capabilities that we have today, but far from the end state of where we see these things. Um so, yeah, I imagine this will be short-term chaos as with everything. AI is a big bogeyman right now, but there's a lot of a lot of space still to grow. >> Yeah, I heard some rumors that Coinbase got a call from Anthropic before the release, kind of like giving access to it to make sure that that Coinbase was was all battle-tested for security as well. I I imagine these these large labs

they almost have a like a moral or like a business corporate commercial duty to private companies to make sure that their stuff is all beefed up before giving this to the access to to the public. On the nationalization of AI, Anthropic Google um and OpenAI are all American companies, right? They're, you know, typically out of like Palo Alto or whatever it is, right? Is is is America just dominating the AI race here? Obviously, China has these extremely similar models that are way cheaper, but it's becoming like somehow the AI game, if you will, has turned into a USA versus China, which is like everything in this world as far as superpowers. Is is America just dominating the AI race right now, and is

this a nationalization is is AI becoming nationalized in your view? >> I So, your first question on if America is dominating, I think very clearly yes. I think China is taking a probably the best strategy they can given that being the case with how they're using or open-sourcing all of their models and making it extremely cheap and and easy to use for anyone. So, I think that is the right approach. If you're coming from behind, you want to go from the position of oh, we should make these things open and and here's an opportunity for you to get cheaper access to intelligence and whatnot. But for the time being, yeah, America's absolutely dominating.

>> Yeah. >> Um in terms of your second question on like should we nationalize, is that something that'll happen? >> Yeah, like Trump wants to take 5% of OpenAI. And or like they want to give 5% to the government or something. >> Yeah, you know, it's interesting. I'm not really opposed to it. I think it does make sense from the perspective of aligning interests. 5% of a trillion-dollar company is an massive amount for the government to just sit on and hold. >> Yeah. >> But it's not so big that the government is necessarily incentivized to go king make OpenAI or or Anthropic in any sense. >> Yeah. >> So, you know, it's interesting. It's the same thing that happened with Intel as well. It feels like we're entering an era of sort of nationalization of these companies providing better alignment when the government can be so helpful to

them. My big concern though would be that if the government is aligned with OpenAI and Anthropic and lab number three comes out, maybe maybe Grok, SpaceX, whatever it might be, is there now a reason to support that lab on the basis that they're an American Elon supported and he generally has the best interests of the country project, or are they going to side with Anthropic and OpenAI because they have 5% of equity or whatever? >> Yeah. >> So, that to me is going to be a pretty interesting thing and see how that plays out over time. >> Yeah. Yeah. Yeah, interesting. So, um with So, Fable 5 is free right now. They extended their free period, I think until like July 12th or something.

>> Yep. >> Um and so, people are like really gung-ho about how how smart and intelligent this thing is and just how great it is. And um what you can do right now with your open claw is you can just go download the Chinese version of Fable 5 or of this most premier intelligence. And it's it's almost like advertised as like 95% as good for 5% of the cost. >> Mhm. >> Is this a problem? Like is is is AI is China going to pop the AI bubble that is happening right now? Like what is your view on the price of inference as it pertains to the quality you get with American models versus the price as it pertains to the quality you get with the Chinese models?

>> I think it's quite ironic that that these Chinese models are actually providing the capitalist incentive for intelligence to be proliferated across the United States in a pretty meaningful manner. I it's to me I think it's totally fine. Like it just means that that Anthropic and OpenAI are going to have their backs against the wall continuing to innovate. And if we are truly the leaders in in AI then it should become very apparent over time that we have the best models. It's not going to be something that China is able to keep up with. Uh and even if they are I I think the biggest concern is around all these data centers, spin commits, all these different things that are happening right now. And do those bubbles get burst with open source models continue to keep up? Broadly I think so, but there's such a massive stack from from

end to end in terms of chip making and and data centers and even like the utilities, water, electricity that are required for these models to run. Even if inference gets way cheaper, I could still see that helping the US because you're going to want American-owned all American land data centers to run those models. You're never going to outsource those to some data center in Beijing. Uh and and broadly I think it would be helpful for 100 times more as many companies that have access to frontier intelligence today to get that access. >> Yeah. Yeah, it's just interesting because when you do run like the when you run things through Claude console on the API, the amount of spend that you incur on any premium model on on an Opus is insane. Like every Telegram message that I send to my open claw is like 10

cents, 20 cents. It's like what are we doing here, you know? Then you pull it into like the like the kidney model or whatever and it's like a cent. It's just like drastically different. >> Yeah, and I imagine too for the whole trading use case as well, this will start to become a pretty interesting piece where these models are going to start performing differently. We're going to see models specifically tuned for trading. I've been playing around a little bit with Fable and having it kind of like rebalance my portfolio for me and what not. It's great at it. And you know, given it's managing a thousand plus dollars, it's actually still pretty cheap overall compared to what it's providing for me. Instead of me paying 220 to some financial advisor, I can do it myself and pay a couple bucks a day.

>> Guys got size thousand bucks. >> Yeah. >> [laughter] >> Um, what about data centers in space? Is or in the ocean? There's a couple startups doing both. How realistic is it to put data centers in space? >> I think every moonshot goal sounds absolutely ridiculous when you first start with it. And I think that one too is looking 20 years down the line to and making a lot of assumptions about how AI continues to grow at an exponential rate until they >> Yeah. >> If we get to the point where we saturate the amount of space available in the US or these, you know, protests and and conflicts around data centers in people's backyards. If that really reaches a boiling point, I could see that being the case. But it would require such a uh such a lift both in terms of the cost to

get these things up there, but also the value that they'd have to provide that I would be surprised if we see that in the next decade plus. >> Really? Yeah. Yeah. Yeah. So, I saw um Blackstone pulled out of a large data center deal recently. Um it was massive and now that data center is kind of kind of in trouble in terms of like getting the funding to finish the job. And so, do you just with that in mind, like do you think that this kind of this this conversation that we're having about inference, like do you think that we're going to see kind of a return to cheaper costs and like less production in in the in the AI world? Like I I seems like it's a never-ending decade-long plus build-out happening. But it feels like right now we're kind of in this phase of like uh

everyone was token maxing. Like we were like, "Let's spend all the tokens. Like give your employees unlimited tokens. Everyone needs to just use AI." And then now it's kind of like, "Well, let's not use it unless we have to and it makes sense." And then so this kind of like shift, even though it's happened relatively quickly last like quarter or so, uh it is affecting now some people's outlooks on the amount of capex and and you know how much spending that they want to invest in this industry. Um do you think like that this kind of re- centering of reality in terms of like how to best use these tools could affect like the investment landscape broadly or like the like these IPOs, just you know, give us a sense of like you know, ground us here. Where where are we truly at?

>> Yeah, I I think it's still so early in terms of the AI rollout that even if these cheap models ultimately win, which seems like they will for for most use cases, uh there is still so much adoption that is yet to occur that the general market isn't necessarily seen the types of payment volumes that we may see or or inference volumes that we may see in the future. Like to your point, the vast majority of companies in New York and SF have adopted AI, but there are a ton of companies in Middle America and globally that have yet to even understand what cloud code is or they just have just now heard of it and are just getting started with it. And all of those places that have yet to see any real traction or adoption of AI are

still sort of yet to be disrupted. And whether that comes from Fable and Mythos or whether it comes from GLM and Kimmy and some of these other Chinese models, uh there's massive massive massive amounts of tokens to still max. >> Yeah. Yeah. Yeah. So just kind of circling back to Coinbase as as we kind of come to a close here in a few, um what is the future of Coinbase's AI rollout? Like what is the grand vision that, you know, you talk to Brian or talk to the to Emily and you're like, here's what the Coinbase AI product vision is going to be. What is that What is What is the kind of broad end game there? And how does that look like in the scope of fintech applications or, you know, kind of like consumer grade

investing apps? >> Yeah, so broadly the goal is to make Coinbase the backbone of the post AGI economy. And really what that means is that we think there's going to be a proliferation of agents across the internet. All of those agents are going to need to have financial accounts in order to exist as a more autonomous entity. And we want that to exist on Coinbase. We have the Coinbase for agents product to give those agents their primary financial account. We have X402 and base and USDC all to provide the settlement layer and rails on which internet native microtransactions can happen at billion plus transaction scale. Uh and we're doing all of this in a way that both is aligned with a lot of the enterprises that are now adopting and using X402. Like I don't know if you saw Cloudflare is now allowing, and they

you know, gate 25% of the internet, uh to pay per use with X402 if your agent wants to go crawl for a website. So we're seeing this shift happen, and we're meeting it in real time, while also trying to get to this broad vision of if there's a product on Coinbase that exists, your agent can use it. If your agent is going to be making a payment, it's going to have something to do with the Coinbase stack. >> And is that going to require like a lot of like rewiring of the product that has been built for humans? Because people often say like crypto was built for robots. Crypto was built for agents. So like the underlying tech of like zeros and ones and transaction hashes and addresses, it doesn't really It's not really appealing to the human eye, but that isn't really matter for agents.

It's just a synthesization of of information rapidly and then gives you an output, right? So it can really rapidly, you know, digest this. Whereas humans, you know, you check the address like 10 times. You're like, all right, let me check it, you know? So product side, what needs to change? Like what is the What does that change look like? >> Yeah, the analogy I've used, and I might have already said it here, is like the the 2010s we saw web to mobile was the big story. The 2020s are being defined by mobile to agents. Uh and that's both the shift in terms of like UI and user experience and how you're interacting, but it's also critically important because these agents as uh a byproduct of being able to act fully autonomously, there's going to be a substantially larger number of transactions happening. So there's some

a little bit of retooling that has to happen in terms of like how do we take something that exists in an iOS app today and make that exposed and externalized to an AI agent that's interacting in cloud or open cloud or whatever. Uh but broadly we have the the rails and we have the tools to do it and it's just a matter of some slight reconfigurations to make it all uh safe and friendly to agents. >> Beautiful. Um man, I mean any closing thoughts Lincoln on this kind of convergence between blockchain and AI, where you see a lot of value being captured, you know, where you see a lot of room for improvement, what specific areas you're excited about in this kind of broad convergence kind of just close us out here.

>> Yeah, I think uh we've seen a lot of legitimization from large enterprises, AWS, Versel, Cloudflare, Stripe all adopting agentic payments on crypto rails. But we have yet to see a really clear demand side piece. So until the average consumer is making micropayments with their agents or or doing these transactions, which ultimately I think will happen entirely behind the scenes, there's a gap to be filled. I think there's a massive opportunity for builders out there to build the tools and the infrastructure and most importantly like the consumer facing apps. It's the exact same problem we run into in crypto in terms of uh we need more consumer stuff. We need things that people are going to play with and use.

Uh that to me is where I think there's the the biggest opportunity coming over the next couple of months and years. Uh and I think broadly payments and trading are going to become the biggest use cases for for AI interacting with money in the next 6 to 12 months. >> Love it. Ladies and gentlemen, that is AI Supercycle episode number 41 powered by our friends at Near Visit, near.ai to learn more about the private inference tool. Lincoln Murr, the head of AI product at Coinbase live from Tokenization Tower. Lincoln, thanks for joining us. Great to have you here. Appreciate it.