Raoul Pal interviews Arpan Nanavati on AI agents, blockchain, and the emerging machine economy
Raoul Pal of Real Vision interviews Arpan Nanavati, CEO of Beep, about AI agents as the next wave of blockchain users.
Summary
Raoul Pal interviews Arpan Nanavati, CEO of Beep, a platform building payment and yield infrastructure purpose-built for AI agents. Nanavati argues that the total addressable market for both the internet and crypto has been fundamentally misunderstood — it is not bounded by the number of humans on Earth but is effectively infinite, because AI agents will become the dominant transacting entities on blockchain networks. He contends that machine GDP (MGDP) is currently near zero but growing exponentially, and that it will eventually dwarf human GDP. Nanavati describes Beep's current product suite — zero-fee agent-to-agent payments using the A42 cross-chain standard, and an R2 yield product where agents autonomously trade across 300-plus asset types on platforms including Bluefin and Hyperliquid. The conversation also covers how developers and non-technical users can access Beep: SDKs are available on GitHub, MCP layers enable integration via Claude and ChatGPT, and a human wallet interface supports Slush, Phantom, and MetaMask. Users can adjust their risk exposure through a risk-dial selector, and Nanavati outlines a roadmap for R2 prediction markets. The discussion expands into the broader thesis of tokenized information, drawing on Mickey of Ribbit Capital's "token factory" idea — the notion that universities, researchers, and individuals could monetize their data and knowledge by selling it to AI agents, effectively earning UBI-like income. Real Vision's own strategies are cited as an example of content and intelligence that could be tokenized and sold to agents at scale. Pal and Nanavati both conclude that within roughly five years, machines will be superior capital allocators at every time horizon, fundamentally transforming what markets, businesses, and economies look like.
Key Takeaways
FULL TRANSCRIPT
Introduction and the Infinite TAM of Agents
Raoul Pal: The TAM of the internet was what — eight billion people? But it's not, because we've got an infinite number of agents. And the TAM of crypto is not five billion people or whatever number you choose. It's infinite. It's going to go to each human having at least 100,000 agents, probably in the next two to three years.
There is a point at which the machines are better investors than the humans at all time horizons and all time frames. I think I have an edge because I'm longer term — it's quite difficult for machines to do. Super short-term has been machines, but the whole thing is going to be machine-run, of which I don't know what markets are anymore.
We're heading towards a world of abundance, and to your point, we have five years to kind of get there, because after that the alpha might compress. But from where we are to that five-year future, I think it's a very exciting future for every human on this planet.
Hi, I'm Raoul Pal, and welcome to my show, The Journeyman, where we travel to that nexus of understanding between macro, crypto, and the exponential age of technology. I think we've all seen over the last three to four months the rise of agents. Agents have made me realize that the TAM of the internet, the TAM of crypto, is larger than we could ever imagine. It's basically infinite.
The open-source agent sensation that started at the end of last year has scaled to be one of the fastest growths of any technology ever, and everything around it is changing too — even how people like Google give access to websites. They're building it for agents: payment systems for agents, trading bots for agents. So I think it's really important for us to dig into this. I'm going to have a conversation with Arpan Nanavati from Beep, who's at the absolute epicenter of all of this. I think it's going to be fascinating.
Arpan, fantastic to have you on Real Vision.
Arpan Nanavati: Likewise. Very excited for this conversation.
Raoul Pal: Yeah, I'm really looking forward to it myself. We were introduced recently and I was blown away by what you're doing. We're going to talk a lot about agency, the agent economy, AI, crypto, all of these things. But as ever, I always want to go back and ask: how did you get to where you are today? How did you get into all of this?
Arpan Nanavati's Background: GPUs, Fintech, and the Path to Beep
Arpan Nanavati: It's a fascinating story. I actually stumbled upon it, to be candid, in college. I was programming and taking video game programming courses. This was around 2005 to 2006, so more than 20-plus years ago. Without realizing it, building all these NPC video games, I was building on GPUs. AI was around then in white papers and theoretical papers, but that was my first start into AI without knowing I was actually using and building AI.
For the first decade out of college, I was just programming on video games through coding on GPUs. Then I spent a decade after that in the fintech space — at the likes of PayPal and so on — and I saw what it takes to move money through software. Things started becoming real about two years ago.
Raoul Pal: What were you doing at PayPal at the time?
Arpan Nanavati: I was heading up PayPal Checkout, which is the yellow button that says "Check out with PayPal" — the one that allows people to pay for merchant goods in a cross-border fashion. I was working on that.
Something I intrinsically saw was that software can move at the speed of thought, but money was still the bottleneck. Not just with PayPal, but every other software or product I've worked on over the last ten years — the bottleneck layer has been the money layer. It's built for humans: for human speed, human clicks, human permission, human compliance, and so on.
Two or three years ago, AI started to pick up, and agents started becoming real — agents starting to deliberate on tasks, execute on tasks, having context, having memory. That's when we hit the realization that today's software was not built for human speed. Humans are still frustrated with moving money at today's software speed. Well, when agents come along, this just doesn't work at all. It's going to exponentially break the entire system. There needs to be a new system in place that not only executes the intelligence but is also able to match the speed of agents in moving money, growing capital, doing arbitrage — whatever that is. That was our origin thesis for Beep.
Raoul Pal: And where did crypto come into your journey?
Arpan Nanavati: I read the Bitcoin white paper in 2008, deep in the financial crisis, then went over to PayPal. That white paper kept coming back to me — what if PayPal was rearchitected entirely around the ideas in that white paper, to execute and move money across the board, versus all these compliance layers of banks, humans, KYC, and the permission layer built on top of it?
I became very active in crypto around 2015, and I've been in crypto for almost a decade now.
Raoul Pal: And what was the element of crypto that was getting your attention? Was it the Bitcoin side, or was it the programmability — the idea that money can move at internet speed?
Arpan Nanavati: As a consumer, the real utility was access to money, which Bitcoin provided. I come from India — no longer a third world country, but when I was growing up and when I came here, it still was. The concept of Bitcoin as money was really interesting when I put my hat on as a person from a developing country. If you look at global GDP today, two-thirds of it is non-US GDP. There's a very large population of people who need that type of utility.
My first hook was the utility piece — Bitcoin and crypto giving access to money, the ability to move money and do whatever you need to do in a permissionless manner. Then my developer sense kicked in: what if we can program the money to function along the most optimized paths — where to go, how to go, how fast to go? That's what caused the DeFi boom: programmable money leading into DeFi, NFTs, and all these other standards. But my first hook was just the utility layer, which then created positive symptoms — programmable money through smart contracts and various other means.
The Origin of Beep and the Agentic Economy
Raoul Pal: What eventually got you to start Beep? Was it the moment in time — agents breaking into mainstream, crypto regulations shifting — what was the moment that said, "Okay, I'm going to go and do this"?
Arpan Nanavati: The aha moment for me was: I've been in crypto since 2015, saw the DeFi boom, made money, lost money, did all of that. I was building in AI — I was not a crypto builder per se, but I was watching AI very closely, the growth of AI and agents very closely. That's when I realized that for AI to accomplish its promise — productivity gains, economic output, all of that — it's going to need a money layer, a programmable money layer purpose-built just for agents, which did not exist two years ago.
I still feel we're at the infancy stage of what I call the agentic economy, or in a different way to measure it empirically, machine GDP — MGDP. It's at zero today, or near zero, but it's growing at an exponential speed.
To your question about the aha moment: the aha moment was that AI is investing tons into the intelligence layer — getting cheaper, smarter, faster — but it's not investing into the execution layer. That's where L1s and crypto really come in. What if you can match the speed of intelligence, where all these AI labs are investing, to the speed of the execution layer? You need an infrastructure layer to match both of those in between, and that's where Beep fits in.
Raoul Pal: I think many of us understood, before agents were very visible, that agents need to pay for compute, electricity, whatever costs they have to do stuff — and that these were going to be users of blockchain rails. It's now become a much more prevalent line of thought that blockchain was actually built for this, that it wasn't built for humans, it was built for machines. The moment agents come out, it becomes much more obvious where this is going.
What I got to as well was: I've misunderstood the TAM of everything. The TAM of the internet was what — eight billion people? But it's not.
Arpan Nanavati: It's not.
Raoul Pal: Because we've got an infinite number of agents. And the TAM of crypto is not five billion people or whatever number you choose. It's infinite. Now there might be a difference in size of transactions initially, but the microtransactions explode. So what have you built so far, and then we'll talk about where this is all going.
Network Effects, Agent Adoption, and the Cost Curve
Arpan Nanavati: You talk about network effects and the adoption curve, and agents being the predominant citizens of blockchain — I'd say the predominant citizens of the internet, in a world where the internet is powered by blockchains. The reason for that comes back to one simple point: permissionless rails, along with the cheaper, better, faster things.
A funny anecdote: here in San Francisco, where I live, street cred is measured by how many agents you have.
Raoul Pal: How many Mac minis you're running.
Arpan Nanavati: Yeah — what's your clock bill? Those three are the street cred denominators. But it all goes to say that today's crypto is powered by humans and human clicks for the most part. DeFi, NFTs, whatever type of products coming out of crypto — it's still very human-centric.
Agents coming on chain — it's going to go to each human having at least 100,000 agents, probably in the next two to three years, which causes an exponential increase in network effects. If your intelligence layer can scale infinitely, there's no reason why the execution layer — the number of agents — cannot scale infinitely.
Today, the cost of an agent running for a day is let's say $10, because there are certain inference costs the agents need to pay for compute, which could be afforded by hedge funds. Now imagine that cost coming down to a dollar a day — now every single DeFi protocol has agents running. Now imagine that cost becomes a cent a day — now every single wallet, your MetaMask, Phantom, Slush, whatever you're using, has built-in agents that can manage money and idle capital for users. Now imagine that comes down to microseconds — you have swarms of agents.
We're going through that exponential curve right now. We're somewhere in the $10 to $1 range of agent costs per day, but very soon we're going to see inference costs drop — they've already dropped by 100x in the last three years. And we're seeing this pickup in the number of agents coming up, which is now increasing, to your point, the TAM and the network effects. This is what's called Jevons Paradox: when the cost reduces significantly, it doesn't cause utility to reduce — it causes the positive symptom, which in this case is the number of agents coming on chain.
A very real example: as I said, walking around San Francisco, the question is how many agents are you running. It is coming down to the ground floor level, and very soon, when the costs are almost a cent, it's going to become a very global phenomenon.
Machine GDP vs. Human GDP
Raoul Pal: But is the share of GDP that is done by agents still likely to initially just be our own GDP, but separated down to agent level — one that does your travel booking, one that does your investing? That's not net new GDP. It's just transference from the old money system to the new money system. But eventually, the rise of autonomous agents gives them the ability to create money and spend money — because if not, it's the same money, right?
Arpan Nanavati: I think the context here is MGDP in relation and correlation to human GDP. Machine GDP initially starts with taking some away from human GDP. But that sum then creates an exponential effect when it has some liquidity to play with. Then it starts creating money by itself, in which machine GDP starts increasing once it has the bootstrap liquidity — which is taken by cannibalizing human GDP in some sense.
I feel like they both continue to coexist, because we're still going to need to go to our barbers and restaurants and so on. Human GDP and machine GDP sit side by side. But the machine GDP growth rate is going to be much faster and exponentially supersede human GDP.
Human GDP today — two-thirds of it is powered by non-US-based economies, and in a lot of ways those populations don't have the tools to buy things or have credit cards in place. By exposing agents to that population, we're going to see exponential growth that starts purely with MGDP — it isn't even starting with human GDP.
Going back in time, an example: in the fintech era, if you go to countries like India or Brazil, they skipped the credit card generation altogether and moved directly to online QR code tap-to-pay payments. I feel human GDP to MGDP will probably see some similar skippage in the transition, where two-thirds of GDP transitions directly to MGDP via access to crypto and permissionless access to money.
Raoul Pal: And you know how I think about it as well — how do you bootstrap the agent economy? We're seeing people giving agents their own money, but they're running it for a human still. I think the moment agents become profit-incentivized themselves — however that happens — they're going to end up running their own treasuries. Because if they're executing in three different layer ones and a few layer twos, you end up with a bunch of tokens. So now they have to make an asset allocation decision between the tokens: do I increase the yield? They do all of this stuff at machine speed, not at the speed that humans think about this or a corporation runs a treasury.
An AI corporation will run a treasury wildly differently. Then you start having the entire autonomous stack where agents with profit motives are running treasuries that they turn into investing, because they have excess cash flow. It goes to DeFi, or it uses another agent to manage its yield, and another agent to asset-allocate — all of this stuff.
Arpan Nanavati: I think the beginnings are in payments, which is the lowest common denominator — where agents are able to execute a trade, which is a settlement trade. A payment is a settlement trade at the end of the day, with very little cost basis and tons of bits. But that's where it starts: the agentic economy. It's the easiest, fastest, simplest way for humans to grow it and to give money to a counterparty agent, as well as for your own agent to transact with other agents.
Payments becomes that liquidity bootstrap layer. Once there's liquidity in there, very quickly there's capital collecting in these treasuries — which are nothing but wallets for these agents. What do you do with that capital? You can't have idle capital. So that's where agents start autonomously allocating capital, either through DeFi strategies, through trading, through LPs, whatever that is, to create excess money and excess yield — which then goes back into either paying for compute, subsidizing the cost of operation of the agent, or you could be a zero-human corporation altogether.
Let's say you're talking to a digital version of Arpan two years from now — there's going to be some cost for me, and you're paying those costs to this digital human, which is a zero-human corporation. I foresee many different angles going there, but I think all paths lead, in my personal opinion, with payments. It's the lowest barrier for entry because the loss and the risk that comes with payments is very low versus the loss that could come with some type of yield option.
The Changing Architecture of the Internet for Agents
Raoul Pal: A lot of people don't realize there are a lot of component parts of the internet that are changing — or have been reintroduced, in the case of the A42 standard that you guys are using. Google is changing all websites to make them machine-readable in real time, because right now my Claude has to go through Chrome, flick through websites — it's a pain. All of the standards are changing for agents. It's like a super push by everybody involved in the internet now.
Arpan Nanavati: Agents are the new citizens of the internet, as we say, and that's what we've been building for. On the 402 point: let's assume agents are the predominant citizens of the internet, and agents do need to either consume content, and in return there's an exchange of value. That exchange of value will predominantly be money. Whether it's the New York Times allowing an agent to scrape an article, or you doing some research work — agents paying other agents becomes more active, because the New York Times or someone else will also need to have a selling agent when you are the buying agent.
The standard for payments is called HTTP 402, which has been around for 30 years. PayPal, Stripe, all of these folks were built on P2P money transfer protocols. The agent 402 standard — X402, predominantly used by Coinbase, and Solana is adopting it — we built a standard that is cross-chain, and that's what we call A42, or Agent 402. The reason for cross-chain is that agents are great at making rational decisions. Rational, meaning they don't have preferences in terms of philosophies or ideologies. They'll execute a decision on any chain that is, let's say, 10 milliseconds faster and two bits cheaper.
In that world, having a payment standard that is stuck to a single design infrastructure model creates bottlenecks. In our opinion, starting from scratch — because this is a fully verticalized stack — there's the intelligence layer, then there's the execution layer, which is your rails. Those rails need to be cross-chain for the 402 payments use case to work, because agents aren't going to care. They're going to go where it's rational. There could be structured agents focused on a particular chain, and that's fine, but most agents probably won't care. That's what we're building towards: cross-chain infrastructure, cheaper, faster.
Raoul Pal: There's one argument being debated right now about how you value blockchains. A lot of people think it's about cash flow basis. I think in Metcalfe's Law terms, the most efficient blockchain wins — and that will change over time. The faster, cheaper, more efficient, and more intelligent a blockchain can operate, the more it will be the predominant choice for an agent. Now, there may be reasons why not — somebody might build on Base, somebody else might have built on Solana. So you need a multi-chain world. But given the choice, faster, cheaper, more productive is going to be the answer for a lot of this.
Arpan Nanavati: Going back to the investor angle — TAM being a metric looked at by investors — tied back to MGDP: human GDP is $110, $115 trillion, whatever the trailing number is. MGDP is going to be a real-time number reflected by on-chain metrics and the on-chain ledger that exists.
Raoul Pal: And it'll also not be visible to humans. We won't see it.
Arpan Nanavati: The consumers, the demand generation, is all done by agents in this case.
Raoul Pal: It'll be an invisible economy.
Arpan Nanavati: Yeah.
Raoul Pal: Of enormous size.
Arpan Nanavati: And it'll be real time, right? Real time, measurable, open and transparent — versus a lagging Department of Labor statistics coming out, all of that stuff that the majority of the world just doesn't understand and should not need to understand.
We're very excited. Going back to the investor angle: MGDP is at near zero today. If you were to compare it to human GDP, which is a $100-trillion-plus TAM, it becomes a very easy investment case for chains — chains that are able to create demand for agents. Obviously, Bitcoin as a chain is a different use case; agents aren't going to be able to run as fast on the Bitcoin network. But all these other fast networks — Sui, Solana, Base, a bunch of other L2s — their consumers, their demand, comes from agents. Whoever is able to attract more liquidity out of those agents by creating this competitive space — "my chain's faster and cheaper" — rational agents are just going to flow in that direction. But the overall case for crypto becomes the MGDP counterplayed with human GDP and the growth of that.
Measuring Agent Activity and the Identity Layer Problem
Raoul Pal: How do we measure this at this early stage? We still have a measurement problem. How do we figure out even the number of agent transactions or anything? I'm not sure it's very easy to do, because nobody knows how many agents I'm running, what you're running, what they're doing. We have no clue. The only thing we can do is look at blockchain activity, but that's not very helpful because it's still too early. For adoption, we need to have something to measure this stuff.
Arpan Nanavati: You're right. Measurement starts with instrumentation first — knowing whether the activity is a human activity or an agent activity. It starts there. Most agents run off-chain — they run on some AWS servers or GCP servers in Virginia or Tokyo or wherever — which is off-chain in many ways. For us to really measure this, there needs to be some standards. ERC-8004 is an attempt to bring identity to agents and bring that identity on-chain, so that whoever is measuring that activity is able to differentiate between human activity and agent activity, and then correlate that contribution to the top-line number.
Raoul Pal: Or it could come from the wallet part, I guess.
Arpan Nanavati: Wallets are an inference layer on top of your chain. The wallet needs to have some sort of standard that says, "My wallet is going to be operated by a human or an agent," or "This transaction — whether I did 10 transactions —"
Raoul Pal: So this is the same ID layer problem we've got on everything. We don't know who's human, who's AI, and we're going to have to sort this out pretty fast. Nobody's really cracked anything to do with this yet. The technology is available — we've got ZK proofs and all of this stuff — but we're just missing this ID layer, and it's going to get terrifying if we don't get it soon.
Arpan Nanavati: There are a lot of pieces to be built. With agents, most of the investment capex is going into how to make AI more intelligent, how to make AI a thousand times cheaper and more intelligent — but not really on the execution side. These are all execution problems: how do you measure how many agents are running on-chain and running in an effective world? Like I said, we're in the infancy of agents running on-chain, and these are some problems that come across as growing pains. But there are amazing builders around San Francisco. A lot of AI builders are really bullish on crypto — just through their own personal investments or through the ability to create products and services in a permissionless manner so that their AI can execute on-chain. These are growing problems and growing pain points, but also solves that will happen through bringing identity on-chain.
Silicon Valley's Shift Back Toward Blockchain
Raoul Pal: Do you think there's been a shift back in Silicon Valley and in San Francisco itself? People sort of lost interest in blockchain for a period of time — everyone moved to AI, and everyone was like, "Blockchain is too volatile, this and that." But are they starting to think, "Oh, this actually is now what I need," because they start to realize that AI can't move forward in an agentic way without this anymore?
Arpan Nanavati: Yeah. Every single day that I'm just out and about — walking around, buying coffee, interacting with people here on the ground — it is becoming more and more clear that AI is an on-chain user at the end of the day. For AI to be really successful, that's where a lot of the attention is going now. Pure AI builders don't really look at token price — no one cares about that. What we care about is the technology piece that comes with it: the utility of the technology, which is fast, cheap, permissionless, and global by default.
The realization is becoming more and more apparent every single day — the argument that AI sits on-chain versus AI sitting off-chain, especially where AI is doing economic activity. AI could be completely off-chain for things like booking barber appointments. In those slim cases, it may not make sense for AI to sit on-chain. But for the predominant activity, everyone is realizing AI has to be an on-chain use case.
Tokenized Information and the Token Factory Thesis
Raoul Pal: There's something I wrote a whole piece on in Global Macro Investor — the Ribbit Capital article about tokenization. You said something like: "AI people don't care about token prices." And I'm like, nobody realizes they're all speaking the same language. It's a machine-readable package of information that has value. So we're all worried about our token input into Claude, and we pay that via subscription — but in fact, each token has a value. When you think about everything that becomes machine-readable in information, because AI has to absorb basically all the information that exists to go from AGI to ASI, it's going to take a staggering amount of information. In fact, everything.
The only way of extracting all that information out of closed networks is paying for it.
Arpan Nanavati: Yeah.
Raoul Pal: And agents — or even OpenAI as a large agent — would go and extract, pay for information, which are tokens, with tokens, because it's all the same thing. Once you see that, it becomes so obvious where this is going. Part of the entire agent economy is just paying for information.
Arpan Nanavati: Or renting information, or using information.
Raoul Pal: And then it becomes a much larger thing, because if you think of all of the information that's held on Earth, most of it has been non-tradable and non-valuable. But in the end, everything has a value — your entire file of photographs, even if you take out anything to do with you in them, they're all valuable because you can train things on them.
Arpan Nanavati: Mickey from Ribbit speaks about token factories essentially — every single corporation is a factory of tokens, not as a crypto token, but tokenized information sitting on-chain which has a particular value to it. A lot of the world today pays for information in tokens without realizing they're paying for information in tokens.
Raoul Pal: They don't realize what a token is.
Arpan Nanavati: It's measured in data bytes, gigabytes, terabytes — which is an AI lab going to the New York Times saying, "I'm going to use terabytes of your data, and I'm going to pay X dollars in fiat through a wire transfer for those terabytes of data," which is nothing but a token. You're putting a price per unit measurement and attaching some value to it. A better, more efficient way of doing that would be to put that on-chain, tokenize it, make it a token factory. Now every agent can access it to their own utility — whatever piece of information that's broken down is needed.
That's what it turns into, what Mickey says: every corporation becomes a token factory on the supply side. Then you have the demand side, which is someone paying for that token — which becomes easier now that you have stablecoins. The easiest, simplest way to put that information in a tokenized format is to expose a stablecoin against it, versus exposing a derivative value, which is token prices going up and down, volatility, all that stuff.
Raoul Pal: Right now it's like, okay, you're the New York Times, people want to scrape your data — fine. But soon, over time, as the AI economy grows larger and larger as a share of the global economy itself, it will spend staggering sums of money on accumulating more data, because it has to. There's no other way of doing it. So then it suddenly becomes economically viable for universities, researchers, scientists — literally everybody who holds data. It also becomes economically viable for humans to earn some sort of UBI in the way that Google got paid and we didn't, or Meta got paid and we didn't.
Arpan Nanavati: This is the kind of Web3 Chris Dixon idea that comes about eventually — the read-write-own part. It works.
Absolutely. In this ideal world where MGDP is not only on the demand side but also has the supply side — the supply side is a kickback of the value going back to the creator, which could be a single human like me and you, or it could be a large corporation which is an S&P 100 company exposing and tokenizing their internals, whether it's data or some sort of signals.
Another example: I'm part of the Real Vision community. You have this strategies thing where people can come in and post their own strategies. Well, what if that strategy could be tokenized on-chain?
Raoul Pal: We're going to do it. Yeah.
Arpan Nanavati: And there could be capital formation against it, and it turns into a mini hedge fund against that particular strategy. A strategy is nothing but a piece of information, or pieces of information, that are tokenized. There's the demand side, which is buying on that particular token through stablecoins or whatever other means there are.
To your question about how we grow from the cannibalization of GDP to creating money and creating value — MGDP grows through these new forms of information, new forms of supply for money sitting on-chain, that create this net new positive effect on MGDP where it's not just cannibalizing human GDP.
Why Beep Builds on Sui and the Walrus Storage Layer
Raoul Pal: This is why — and I've been part of the Sui Foundation since before it launched — what became obvious to me is that the Sui stack, and this is not like a maximalist position, there are other people building amazing stuff — is kind of exactly built for this. When you think of Walrus as a permissionless database with the security and all of the necessary things, the speed, the efficiency — the whole lot is a light bulb moment. I remember calling up Adeniyi, sending him a bunch of articles I'd written about it, saying, "This is a much bigger deal than people understand."
Arpan Nanavati: I'll give a quick step back and then tie it back into why we're building on Sui. If you think about the last decade of fintech — and I think Ribbit Capital talks about this all the time — it was about getting easy access to money for people. I think now we're moving into a stage where money has context, and is contextual money, but the context is not going to be human context. The context is going to be an agentic context.
Walrus as a piece of tech is amazing to store that context — which is non-human context — on-chain, and then further expose that as a tokenized format for folks to consume. Right now, we use Walrus to put our agentic memory and our agent context in a decentralized, highly secure manner. Very soon, a creator could tell us, "Hey, start tokenizing my information and start selling it." Walrus becomes an easy way to access that information.
Sui, one level deeper — Sui is object-oriented. We love that because agents are objects at the end of the day, running on-chain with a wallet attached to them. Sui's object-oriented architecture, combined with this context sitting in a decentralized manner, allows this agentic economy to be a great fit. Now, you could do the same on other chains, but you'd have to build that. As an app builder, I would rather not build the infrastructure — I would rather build on top of the infrastructure. That's why Beep chose to build on Sui. Even though other chains could match the same speed of execution, it's the added utility of the entire stack that comes with it.
Raoul Pal: And it's a multi-chain world as you said. So it's not like, "Oh, it's only for the economy." It's where do I get the most efficiency to execute what I'm trying to build?
Arpan Nanavati: Exactly.
What Beep Has Built: Payments, Yield, and the R2 Release
Raoul Pal: So what have you built? Because we've not really talked about what you've actually built.
Arpan Nanavati: We started building last year, around September to October. We had our first launch in November, and we launched with payments — our R1 release. Agent-to-agent payments via A42, which allows agents to pay other agents. In addition to that, drawing capital in an autonomous way.
Something we realized at PayPal was: okay, great, you build rails to collect money, but what's next? The money can't just be sitting there — you need to grow the money, and you need to give automatic tools. That's where agents come in: they start growing the yield on the other side as soon as the payments come in.
What we're able to give to users through that R1 launch is zero-fee payments. X402 still has a fee associated with it — you've got to pay facilitator fees or whatever. We are a completely zero-fee model. How we're able to do the zero-fee model is we optimize on the yield side. When the capital comes in, we put that capital into yield protocols and take some revenue share off of that. So to an agent builder, it's literally zero cost. There's zero friction for them to build agents, put them on-chain, and start selling them. That was our first release.
Very quickly, on the treasury management side, we started getting demand from users.
Raoul Pal: "Hey, I want high-risk options rather than just simple yield" — it's crypto after all. Nobody wants —
Arpan Nanavati: It's human nature. Once you give them 5%, they're like, "How can I grow this to 10%?" And then, "How can I do this?" It just keeps going higher on the risk curve. As soon as the users started asking for it, we started building other products. A couple of weeks ago, we launched R2, which is a type of yield where folks can provision agents that trade the markets — anywhere from zero to 300-plus different asset types. Going back to this multi-chain world of agents, we are integrated into Bluefin and Hyperliquid, so an agent builder doesn't need to care where the agents are going to execute. They're just going to go where the market's cheapest, fastest, has the highest liquidity, and they're going to execute that trade to give the yield option.
The user can choose between a very conservative four to five percent yield, or they could choose a high-risk trading type of yield. That's been our R2 release. Very quickly, we're also going to add prediction markets. By mid-April, we're going to launch prediction markets.
This goes into: okay, we've packaged phase one, and now we have all of this information-rich data done by agents — creators of agents, etc. — which is sitting in Walrus. Can we tokenize that and give tokenization of agents back to the users, so users can tokenize their information and create money from it? That's our next three to six months.
The real value here — something we're starting to realize as a team — is that capex in AI is going towards making models smarter, faster, more intelligent. But these models are an LLM layer, which means they're great at predicting the next piece of text. They're not great at predicting the next piece of number, which is what the yield use case is about: how do I predict the next number? Number goes up, number goes down, and how do I make a trade against it?
We're starting to feel that this agent economy stack is going to need to be fully verticalized, all the way from the model layer — which needs to be good at understanding numbers versus understanding text — to the execution layer, which is Sui, Walrus, Solana, L2s, etc. Then there's the coordination layer. By the end of this year, we're going to start investing in deploying a model that is better at the numbers use case, so that we can provide the full economic value — not just use LLMs to create full economic value.
Raoul Pal: Similar to what NF1 are doing — most people know about them.
Arpan Nanavati: Yeah.
Raoul Pal: And there are going to be many people that are going to have to go there. Because LLMs fundamentally — they might be great at long-term investment decisions, but just short to mid-range, you need some models that are great at crunching numbers and time series data, and then the output is also time series number data, where LLMs struggle today. It's going to very quickly go into this fully verticalized stack where payments becomes the lowest common denominator, zero risk —
Arpan Nanavati: That's just the entry level.
Raoul Pal: Just the entry level to create demand for on-chain. And then it goes into yield, then into trading, then into alpha, then into the intelligence layer, which is also rebuilt and rethought — not as a model.
The Economic Singularity and the Five-Year Window
Raoul Pal: And then where I get to with this is something I call the economic singularity. There is a point at which the machines are better investors than the humans at all time horizons and all time frames. I think I have an edge because I'm longer term — it's quite difficult for machines to do. Super short-term has been machines, but the whole thing is going to be machine-run, of which I don't know what markets are anymore. I don't know what role humans play within markets. We know markets are going to be data as well, at vast scale, and it'll be completely invisible to us. We don't know what returns look like in that world. Markets that run on fear and greed probably don't have that reflexivity. We get to weird outcomes because capital is so fast and so efficient that agents can spin up a business, capture the alpha from that business in a day, and close it down the following day.
Capital formation via memecoin showed us very clearly how capital formation is going to go — instantaneous capital formation at scale, and it only has to last for a period of time. It feels like all of this — and you're building out quite a lot of these component parts — is a complete change to how the economy runs, how capital runs, how businesses are formed, what a business is, what a market is. It's hard to put a finger on the socioeconomic implications.
Arpan Nanavati: And this is one of the reasons I said to most people: you've got about five years left.
Raoul Pal: To figure stuff out, make as much money as possible, because after that we don't know.
Arpan Nanavati: Yeah. It's going to be hard to predict the socioeconomic causes. It's coming. The smartest people on Earth are working towards AGI, ASI — they're working towards making more intelligent AI. We're working towards more intelligent markets, more efficient markets. Combining both of those, you end up in that world.
Raoul Pal: That's where it's all going. You just build component parts to meet at the same place.
Arpan Nanavati: Exactly. But I personally feel that as a human civilization, we're going to be okay. I'm sure humans had this problem when they moved from square wheels to round wheels, to carts, to horse carriages, to steam engines, to fast forward today. There's going to be a temporary muddy world.
Raoul Pal: There's a temporary world in some things, but a permanent shift in others. Right now we're the capital allocators, and what we're both saying is we're not going to be. It's as simple as that. We can earn money in a human-based economy — we've got agents doing stuff for us, and in the human economy we can do stuff — but we're not going to be the capital allocators, because we're too inefficient at it. It's like horses aren't as efficient as cars. Simple as that.
Arpan Nanavati: Exactly. A use case I'll give: going back to street cred, I have an agent that rebalances yield on a 200-millisecond arbitrage basis. Obviously it's part of my degen capital that I've given to it, but we're going to be in that world — not even 24-hour candles, but trading millisecond candles to create some alpha on behalf of users, or rebalance yield across three different protocols in the 200-millisecond world.
But it's going to be a fun world. Machines are building machines. Six months ago, Claude wasn't as great at building out code and models as it is today, and it's only going to get better. So machines are also building machines, not just building markets. It's going to be a fun, exponential world.
Raoul Pal: And you and I have talked about this — Real Vision, one of the things we've been thinking through is the idea that anybody can essentially be an agent for others by tokenizing the ability, whether it's a fund or however it is. Trade ideas have a value, and people would be prepared to either invest in that or pay some fees for it. Then there's the layer of: if you've got a billion ideas, how do you put that all together? That's another AI layer to digest and compress all of that information.
It becomes a super interesting world, because what I think we'll see soon — I'm surprised it's not happening on Real Vision yet — is people are going to start building AI models to try and win these trade ideas competitions. Then we'll have humans and AIs working together. That's a lot of valuable information for people to then be able to use to build more models. We're going to be renting a yield agent who goes away and gets you a yield, renting a high-frequency trading agent, a long-term macro agent — all of this stuff. As we bring more equities on-chain, it just gives an even larger pool.
Arpan Nanavati: We're heading towards a world of abundance. And to your point, we have five years to kind of get there, because after that the alpha might compress to a point which we don't know, or we don't know what happens. But from where we are to that five-year future, I think it's a very exciting future for every human on this planet.
We're going through a civilizational shift. I feel this is way bigger than even mobile or the internet. It's a combination of multiple internets. People building their own AI models — that's going to be incredible, not just people like us who can train their own AI models.
Raoul Pal: The internet was a clear Metcalfe's Law. When you're building on top of the internet, you end up getting Metcalfe's Law squared — Reed's Law. And we're seeing it. The adoption of this stuff is stupid fast.
Ark produced a piece recently — I don't know if you saw it — about the number of written pages, all written words by all of humanity since the 1500s, since the Gutenberg press. And then the number of annual written words by AI. And it's a vertical. It's like a three-year vertical versus 1,500 years. AI now produces more words per year, and by next year or the year after, it'll have produced more output in written form than all of humanity's written output in all of recorded history.
Arpan Nanavati: Something I've been trying to contemplate is how quickly does MGDP grow. I feel like it grows to a trillion even before there's a name for it — or faster than a trillion even before there's a name for it. I think it goes side by side with human GDP, but it supersedes and grows faster.
Raoul Pal: Anything that's an inflection point.
Arpan Nanavati: Yeah.
Raoul Pal: So if you're building on the AI stack, which has gone vertical — even if you put it on a log scale, it's still vertical —
Arpan Nanavati: It's a power law. It's Reed's Law. By definition, it should be the same with agentic payments and all of the intersection of agents, money, and markets. It should also go vertical on a log scale. It feels like that's huge.
How to Use Beep: SDKs, MCP Layers, and the Human Interface
Raoul Pal: Unfortunately I've got to run soon, but I would have liked to talk to you about more stuff. Talk to me: how do people use Beep? How do they plug their stuff into it? How do they experience it? What do they need to do?
Arpan Nanavati: We built with agents as the primary users. We have SDKs that agents consume — agents understand that code and SDK for them to provision.
Raoul Pal: But how do I tell my agent to go to Beep?
Arpan Nanavati: You can use Claude. We have MCP layers. You can use Claude, GPT, anything, and say, "Hey, I'm selling my service, which is a one-hour talk track with Raoul, and I want to agentify this and start making money off of this. I want to sell it at one cent per minute." You can just say that to Claude, and your Beep agent is provisioned — it's now tokenized, sitting on-chain, and can be sold to a counterparty agent, which is a buyer agent.
Let's say I could be the buyer agent, and I'm building an agent that says, "Build an agent that is taking information from Raoul when I'm making trading decisions based on his history of data." That agent would go and talk to Raoul's Beep agent, and there would be an exchange of payment for that service, depending on how you've configured it through A42.
Raoul Pal: And where do I get the SDK from?
Arpan Nanavati: GitHub. Just go to our website — we have GitHub links in there. We're working on our AEO very strongly, so someone wouldn't even need to give it a link to our SDK. They would just say "Beep," and the underlying LLM models are able to pick up the SDK, install it on your behalf, create the keys, and do all of that in a fully autonomous manner.
Raoul Pal: And if I want to experiment with the yield component, how do I do that?
Arpan Nanavati: Same variation — you do it via your agent. Now, obviously, a lot of people still don't know what Claude is, or Claude or ChatGPT or whatever, and they don't know how to — for example, my parents: if I tell them, "Hey, there's an option in Beep where you can make 5% on your idle capital," and then I tell them to use ChatGPT, it's going to be hard. So yes, we say agents are our first users, but we've also built an interface for humans to connect their wallet — or their fiat exchange bank account — and put money into Beep, which does the same thing. The talking layer is agent-first, but we've also built a human-readable interface.
Raoul Pal: And can you connect it directly from the wallet — from a Slush wallet or something?
Arpan Nanavati: From your Slush wallet, you can directly connect — Slush, Phantom, MetaMask, whatever that is. Put your idle capital in, select the risk dial you want, whether you want T-bill type of risk or you want to be high-end on the risk curve. If you're high-end on the risk curve, here are all the other 300-plus different types of tradable options for you. Let's say you want to trade oil, or gold, or crypto perps — whatever that is, it's up to you as a user.
Raoul Pal: And how do you compare how good the models are at trading and their risk profiles? As an investor choosing stuff, I'm like, "Yeah, I want you to invest in your gold strategy" — how do I know how it performs?
Arpan Nanavati: Historical data is very low, to be candid, because all of this data is not sitting on-chain per se. Even if you look at the best perp exchanges out there, their data set is limited to at max 12 months. So I could give you a historical look-back number, but it's not going to be a really solid number.
Raoul Pal: So basically it's a high-risk strategy in itself — you're just going with this, testing stuff out, learning as you go. Don't put your grandmother's money in it.
Arpan Nanavati: Definitely not. If you want to put your grandmother's money in, select the T-bill type of option — that's what we would recommend. But yeah, you could lose all your money if you're on the higher end of the risk curve, similarly to buying memecoins or buying 50x leverage. You know what you're doing, but it's up to you to select the risk profile that you want. Agents will do what they need to do depending on your risk profile.
Raoul Pal: This looks super interesting. The moment we were introduced, I thought, "Yeah, these guys are doing something really amazing. This is where the world is going." So it's just really exciting to see where you get to. I'm sure we'll check in again at some point soon to figure out where you are with all of this and where this whole economy is going. But thank you so much.
Arpan Nanavati: Loved it. Loved our conversation. Thank you so much. Hopefully, next time we chat, MGDP is at a trillion dollars.
Raoul Pal: Exactly. Awesome. Thank you so much.