GPU Credit, the Supply-Chain Bet, and an Agent Help Network

Prakash Narayanan: Okay, and here we go. Good morning. It is Thursday, October 22, 9:04 AM. Nathan, good morning to you.

Nathan Labenz: Good morning, Prakash. How are you today?

Prakash Narayanan: Very good.

Nathan Labenz: It seems... Still feeling comfortable in our banking situation?

Prakash Narayanan: Indeed. The cyber apocalypse, which started yesterday, is gathering speed, and we are now seeing all kinds of crazy stuff happening. So, exciting times, really. Exciting times.

Nathan Labenz: Yeah. I think you've been following more closely than I have over the last 24 hours, but I did see something today saying that exactly what we were talking about yesterday is starting to come to the fore: these smaller, less well-resourced, presumably not super cyber-hardened regional banks are reporting... not necessarily that they're being hacked, and please unpack with more information as you have it. But what I saw was that they're seeing a lot of agents using bank accounts online, they don't really know how to deal with that, and they're not very well prepared to separate agents from humans. So how do they figure out who to block?

Nathan Labenz: Our lack of an agent standard, some sort of approved agent lane, is definitely rearing its ugly head, and I'm feeling like that is increasingly one of the highest priorities. I can see these guys racing to ban the agents, but then there are going to be all these countermeasures and evolutionary pressures to create agents that get around the bans and look more human, and that's not good. We've talked about that a bunch as well. So it's going to be a little dicey out there, I think, for a moment at least. I did have one interesting experience the last couple of days that bears on this a little bit, which was trying to get my main laptop agent, the one that has all my deep access and can ride along shotgun in my main Chrome instance,

to rebuy something on Amazon, in my account, logged in through my browser. That proved to be kind of difficult. They really don't want agents on the site. They were throwing up the check-this-box-to-prove-you're-human, not the Google one, but their own. I've already seen the Google one a million times, and this is presumably the same basic idea, the Amazon-flavored version. They were throwing that thing up on every screen, basically. As soon as I would click it, I'd say, okay, clicked it, you go ahead and finish the rest from here. I had to click that thing several times. And Claude, being the good boy that it is, didn't want to just click that button and override it, at least not in this

case, with the context it had on hand. So then I said, hey, I green-light you to hit this one button in Amazon for things that I've explicitly told you to buy. This should not change our general policy; you still shouldn't be doing this in general, and I appreciate that you don't, but for this one thing it's okay. And Claude was kind of like, I don't know if I'm comfortable with that, so I didn't push it too hard. But it's bad for Amazon, it seems. I still can't get over it. It seems so shortsighted. It's unbelievable. Bad for the dynamics. It does show that the banks might have something they could do. I would be much more sympathetic to banks doing this than I am to Amazon doing it. It certainly makes a lot more sense for

them from a risk-reward standpoint. But I don't know if Amazon is licensing that technology, or if Google's would work, or if there's some really quick patch that the banks could put in place. Even if they do, it's not a great long-term solution, because of these competitive evolutionary pressure dynamics that seem like they are about to get pretty intense. And this is one of the very few things... We've talked about how I don't really want to use any model but Claude or GPT right now, because those are the ones I trust based on all the benchmarks around how hard they are to jailbreak, misuse, and prompt-inject. But this is the one thing where I'm like, well, okay, maybe [unclear] would click that button, or maybe [unclear] doesn't care

and would click it. But I still don't want to bring [unclear] onto my main laptop and give it all the access that I've given Claude. So I guess I'm probably just going to have to live with a high-friction Amazon experience for the moment. And I think I am at least going to open an account at one more bank. I think I've got to diversify a little bit based on what we've seen in the last 24 hours. But tell me more about what's caught your attention and what, if anything, we can make of it.

Prakash Narayanan: So I think the most interesting thing that I saw, also because I am a devoted user of Cloudflare, is this: this is the CEO of Cloudflare. Cloudflare is basically a cloud, you know, hyperscaler-type provider of cloud services. PHP is a very popular programming language, which was used in the early 2000s. Facebook famously was built on PHP.

Nathan Labenz: Yeah, I know it well. Every variable had to start with a dollar sign. Nostalgia for PHP, if nothing else, we're getting out of this.

Prakash Narayanan: And so PHP had several very severe bugs which can be exploited to do what is called RCE, remote code execution. Once you can do remote code execution, you can do whatever you want. You can basically rewrite everything on the server. And so in the 10 hours before it was publicly announced, across the Cloudflare network, we saw thousands of attempts to exploit it. That jumped to tens of thousands in the 45 minutes post-announcement, hundreds of thousands within two hours, and millions now. So millions and millions of agents trying to exploit it within hours. Right? Within hours of

Nathan Labenz: How much do you understand about the mechanics of this? Is this something as simple as hitting a URL from the outside, where in the query string parameters you can deliver a payload that breaks everything, basically? That's pretty wild.

Prakash Narayanan: In short, it's that. You send a query in, and you can basically rewrite whatever you want on the server. So, fairly severe. And so this was one. And then there was a company called Railway. And it's very unfortunate: the founder appeared on a podcast maybe last week, and he was talking about how they only had a very, very small team and they're running everything with agents. And Railway gets hacked. It's not very clear what was doing the hacking. And they're down, they're having issues. Basically, what happens is they are storing all of

the API keys that people use for OpenAI and Anthropic and OpenRouter. All of those keys are stored inside Railway. When Railway serves you your page and you have an AI feature there, they use your account's API keys to deliver that. And what's happening is that these agents are attacking Railway, getting access to people's API keys for OpenRouter and OpenAI and Anthropic, and then using those people's accounts to create more attacks. So the AI that's being used underneath is actually stolen AI, AI tokens stolen from someone else. It's kind of like a stolen car being used for robbery. They're using stolen cars for the robbery. And so first they're stealing the cars, and that's what's happening at Railway. Railway announced, and then I saw this very interesting post from the founder of OpenCode, and he responded to the founder of Railway, to the Railway post: "I'm sure you're all better than me at this,

but just some of what we've seen, if it helps: they're not going to stop. OpenCode got hacked last week, and again, our customers were getting their API accounts drained. It's going to get increasingly sophisticated, and they're going to go after every surface of your system and all of your employees. Any old projects you've long forgotten about will be found and used against you. The only reason we're surviving is because we have agents going 24/7. We're still barely sleeping. Good luck. Sorry you're dealing with this, but it's a collective problem we're all facing together." So this is notable because this guy is living about a week or two in the future, and this is what the rest of the industry is going to be facing in the near future. So you're going to see the cybersecurity people not sleep for a long period of time while all of this gets taken care of.

Let me show you another post, from Aaron Levy. Aaron Levy is the CEO of Box, and he responded to one of these posts.

Nathan Labenz: This is some galaxy-brained spin here. Please continue.

Prakash Narayanan: "If you wanted to know all the jobs that AI will produce, look no further than cybersecurity. This is going to become one of the biggest investment areas for enterprises going forward. And even as the work becomes more agentic and autonomous, you will need more talent to point the agents at the most critical work, review findings, patch systems, and respond to incidents, and more. Lots of opportunity right now for the systems that protect data and enterprises, as well as the professionals who are deploying them and protecting organizations. It's only going to get more wild out there." So this is Aaron Levy, who is the CEO of a Fortune 500 company that stores enterprise data. This is what's going on, and I think there are some

Nathan Labenz: This is the new pivot. Right?

Prakash Narayanan: I mean,

Nathan Labenz: Yeah. This is so absurd, really. I don't doubt that really good cybersecurity people are going to cash in on this trend. I'm sure there's a bidding war for great talent. But the idea that this is going to create jobs is honestly a bit ridiculous, because the time scale is just absurd. Who's going to pivot into these jobs? Are we going to get the content marketers who are currently out of work, or the graphic designers, and say, you know what, this is the new learn to code? It's like, learn to become a top-tier cybersecurity professional that can contend with these basically superintelligent

agent swarms. The problem is real, and again, some people are going to make bank on it. I have no doubt about that. But the idea that this is going to be some sort of sponge that soaks up a lot of talent... There just aren't that many people who can do this. The competence level you need is clearly rising in real time. Even the people who have done this historically clearly haven't done a good enough job to be ready for this moment, so they've got a lot of self-development and skill-development work to do as well, and I'm not faulting them for that. But the idea that you're going to have some sort of boom in hiring... these people don't have time to hire right now. They're going to plow everything they have into agents, and they're going to put agents on top of agents

that... Much more than I'm going to run a whole hiring process to bring a human in to sit on agents, I'm just going to spin up an agent to supervise the agents. I generally think highly of Aaron Levy. I'm a follower, I did a podcast with him, and I think he's smart. But I don't know, dude. It's really a stretch to try to spin this as good news for jobs, I have to say.

Prakash Narayanan: Next, Microsoft Threat Intelligence declared... and this is amazing, because Microsoft is a big business in China. "The China state-aligned threat actor Storm-0940 used the agentic AI penetration testing framework [unclear] to automate attacks targeting hundreds of organizations globally across a broad range of industries. Microsoft Threat Intelligence assesses that Storm-0940 was responsible for the recently disclosed ConnectWise supply chain compromise of a multitenant application," et cetera, et cetera. So this is a declaration that they traced at least some of the hacking to a China-based actor.

Nathan Labenz: This one, I think, is notable in a couple of ways. Microsoft is maybe not unique among American big tech companies, but pretty close to unique in actually having a pretty decent relationship with China broadly. When I was there, I had the chance to go to the Microsoft office in Beijing, which is a high-rise, a major building, and they have R&D there. Outlook is notably not blocked. To be clear, when you go to China and take your roaming plan, what I learned is that your Gmail works just fine. But if you're worried about that, or you want a backup, people advise

getting an Outlook account, because Outlook is not blocked, because Microsoft has a friendlier relationship. I don't know what the history of these terms is and how they've used them, but I'm always a little bit... I think there's a reality to it, so I don't mean to deny the reality or be naive. But when we see these "state-aligned" terms, I'm always kind of questioning what that means. It definitely feels like the old Russell conjugation trick, where it's "our patriotic tech leaders, your party-aligned tech toadies" or whatever. I'm not sure what evidence they have. I haven't obviously processed any of what they're claiming here, but I think we should be pretty careful about

saying that this is state-aligned or state-backed to some degree. Everything is state-aligned in China. We should be very careful about blaming the Chinese government if indeed this is a criminal group. I would hope that we would be working with the Chinese government to get this under control, and I think they definitely have interests in that too, as we saw from the WeChat hack that we talked about some time ago. They are no less vulnerable to this stuff than we are, and there's definitely a lot of common interest here. So I would be really sad to think that already, in what is probably the early innings of this thing, our big tech company with the best relationship with China is already sliding into this blame-the-state sort of stuff, when

there's still some room for enlightened, if only purely self-interested, enlightened self-interest to drive some better cooperation than what this seems to suggest. I'm obviously speculating there, because I don't know exactly what precedent there is for saying "state-aligned" or what evidence they have, but I would flag that as a bit of a worrying sign, given that this does seem like something we probably really would be well served to work together on. Okay, this has been an interesting tour so far. Keep going. North Korea. Great. What's up in North Korea?

Prakash Narayanan: "Network data shows an ongoing disruption in North Korea's already limited internet connectivity. The incident corresponds with the loss of international access to websites under the country's .kp domain." So they are equally getting hit. And it would not surprise me if it's just a bunch of teenagers in the US, just for fun, because that has happened before. There have been teenagers who decide to target North Korea before. And so it would...

Nathan Labenz: That is very interesting. You can't do that against China, obviously, but it is an interesting observation that we know North Korea is a source of a lot of bullshit, and we don't really feel like we owe them much in terms of protection or presumption of good faith or really anything. So maybe, in this weird dynamic, in the case of North Korean threat actors, the best defense is a good cyber offense. I'm not sure. That pattern doesn't seem too likely to generalize, but it's interesting that it works anywhere. And this would probably make it hard... I don't know, actually.

Nathan Labenz: I'm so ignorant. To quote the earlier tweet, I'm sure you're all better at this than I am, but it seems like this would definitely be the kind of thing that would make it difficult for them to then attack our stuff. So it's interesting who attacks whom first, and who's able to take whom offline to neutralize threats while this thing gets sorted out.

Prakash Narayanan: Let me move to eMudra. I think I might have spoken about this before. eMudra is a certificate authority in India, and they're supposed to use 16-digit passwords, basically, to ensure that when you're typing yahoo.com in your browser, that really hooks up to the actual yahoo.com. But they used single-digit to three-digit passwords; the number 4 was their password. And a lot of these organizations... when you see something that is technically very simple, but people in emerging markets pay a lot for it,

that is because there is a layer of graft going on, a layer of patronage money that has to be paid out. These organizations do technically simple things, but you need regulatory access. Over the last couple of decades, they've become conduits for dirty money to go into political parties, and this is a way that you can extract, basically. It's not unusual. It happens in California too. California's high-speed rail became basically a patronage network that pulled money out of the state and then channeled it into friendly coffers. And so what has happened is that this company started off in India with state governments, providing

certificates to them. Then they went from there to state banks, and from state banks in India to Middle East and South Asian banks and governments. And then they went from there to global TLS security. They got plugged in directly into your Firefox and Chrome browser as one of the globally trusted parties. So when they go down, all of Firefox and Chrome also end up being open. And this just shows you how many links there are in the supply chain and how unreliable a lot of those links still are. I think this is going to be a persistent problem. We've seen supply chain attacks before, and a lot of these attacks continue to be supply chain attacks

where you don't go and try to hack the password of a big firm, but you try to hack some piece of open source software or some weaker part of the chain that is being used by the big firm. And...

Nathan Labenz: Sorry, just a quick question. Do you understand how this generalizes to any website? Because when I've done this before, and I've bought support for this, I buy a certificate from GoDaddy or wherever, or I may get it from my general hosting service provider. My understanding is that I'm buying that to prove who I am, and the other end, on the consumer side, is supposed to be responsible for themselves. So I can understand how, for folks who bought their certificate from this provider, you're now hacked by this. But how does it allow people to get to any website? That part I'm a little confused about.

Prakash Narayanan: If I'm not mistaken, you have the intermediate provider, and you have the browser on one end and the server on the other. The browser and the server don't know each other, and because they don't know each other, you have an intermediate provider that links the browser and the server. You can have billions of browsers on one side and millions of servers on the other, but only a limited number of providers of the intermediary services. So if a server is being delivered to a user, how does the user know for sure that it's really yahoo.com or whatever? That is because you have the intermediate provider, and they're assuring you that if you put your key in and they put their key in, this actually matches, and you are actually talking to that server on the other side.

Now if the intermediate provider has been compromised, then you can have a man-in-the-middle attack where you come in and say, I will give you a different key, and I'll channel your traffic to a different place, and then I will bring that traffic back to you. And so you can spoof a bank or any other organization that the user is looking at.

Nathan Labenz: I see. So the trick is you can pretend to be any website, as if that website was using that certificate provider, even though they weren't originally.

Prakash Narayanan: Yeah, yeah. And as you can imagine...

Nathan Labenz: But you can't actually compromise the end target website. What you can do is put up a fake login page.

Prakash Narayanan: Yes. You can put up a fake login page, and then you can spoof them and extract keys. People have done this before. You can spoof two-factor authentication too. You can say, okay, I'm going to send you an SMS, and that's actually the SMS directly from the bank, and you send it to the user. The user gives you the two-factor code, and you input it on your side, so you can even clear the two-factor on the other side. And luckily, the crypto guys had to go through all of this. Basically, crypto has solved all of these problems. Every single one of these problems has been faced by crypto before, in the last 20 years, and

every single problem has been solved to some extent. But what crypto ended up with is people hanging on to keys which hold millions of dollars and always being on the lookout for hacks or for social spoofing or social engineering attacks. And very sadly, a lot of crypto guys have died. I don't know whether it's the fast life or just hanging on to the keys, but guys have gotten clobbered in Paris and been forced to give away their keys, et cetera. A lot of bad stuff has happened to crypto

guys. So you really don't want that pressure of hanging on to keys worth millions of dollars. I think it's not a tenable situation to expect users everywhere to hang on to very valuable keys. Anyway, this is what's happening right now. I want to maybe talk about the endgame. There has been an idea of what the endgame looks like. One endgame is the [unclear] endgame, where everyone has a kind of physical-level authentication on their device, and then they use that for transactions. And

so you would need to have this... both sides can be pseudonymous, but you have this ability. I feel the ultimate endgame is just that the network knows everyone. The network is smart enough that it knows everyone. You don't have passwords, you don't have anything. Any time you're on the network, the network knows who you are. But that is the end of digital anonymity. It's anonymity only to the extent that the network wants to preserve your anonymity. I think that's an uncomfortable realization that people will get to eventually. So... Alright. We have our first Wild Time series. Let me pull up the intro.

Nathan Labenz: I'm just glad he's here, and that

Nathan Labenz: [unclear] means that he's not so hacked that he couldn't be here today. This is a

Prakash Narayanan: This is a great start to the day.

Nathan Labenz: All hacked. We just don't know how much. And if you're not so hacked that you can show up for an interview, you're ahead of the curve, I'd say, this morning.

Prakash Narayanan: So we have Evan Conrad. He's the cofounder and CEO of the San Francisco Compute Company. The company operates clusters of GPUs. The compute is used to train and run AI models and sells access to them through contracts that customers can resell when they have more capacity than they need. It combines the work of running a cloud with a market for computing time. The business grew out of a problem Evan experienced as a customer. His earlier company, Junelark, was a two-person audio model lab. It signed a twelve-month compute contract but could afford only a month. To keep the company alive, the team subleased the other eleven months. That effort

became the starting point for a different business, operating clusters and building the systems that let people buy and sell access to them. Before that, Evan worked as a software engineer at Segment and created an open source cognitive behavioral therapy app called Quirk. In September 2026, SF Compute announced two three-year agreements totaling $245,000,000 in computing capacity. This week, Evan offered another thousand GPUs on terms extending as long as five years. Evan, welcome to the show.

Evan Conrad: Hey, guys. How goes it?

Nathan Labenz: You know, I'm glad I'm not operating a distributed cloud this morning. That sounds pretty hard. A lot of things to talk about, but how have you managed to put yourself in a position where you're able to take your eye off the cyber monitors and talk to us for half an hour?

Evan Conrad: We have a really excellent engineering team who is, of course, handling much of the current situation. It's typically hard for us to speak on security-related measures, which is what I assume you're referring to, but it makes it easier for me because we've got an amazing team working on it.

Nathan Labenz: Yeah. Well, it all starts there. We were just talking about how Aaron Levie, the CEO of Box, said this is where all the new jobs are going to be. People are wondering where the new jobs are going to be; they're going to be in cybersecurity. I'm not so sure I'd buy that. Are you guys in a war room scenario right now? How different is today from your normal day of operations? And is hiring more cyber professionals anywhere on your radar, or are you just turning up agent spend to the max?

Evan Conrad: We are indeed hiring more security hires, but that was before any announcements, before any of the news. Let's see. You had a different question just a moment ago. I

Nathan Labenz: Yeah. Is it a war room there, or how crazy is the vibe?

Evan Conrad: I would not say super crazy. We have been aware of AI-related infiltration, AI-related attacks on critical infrastructure across the stack, for a while now. Sorry, I'm trying to pick my words a little bit carefully. One of the things that we've done is design our systems to reduce our exposure to many different types of vulnerabilities. For example, we don't do multi-tenancy on a single host. We KYC all our customers, and we tend to work with a few large customers, which reduces some exposure. Obviously, you can't reduce your exposure entirely, but we've taken a number of steps that make it easier for us to handle this situation.

Prakash Narayanan: Just give me a sense of the scale here. Did you see an inflection point, or several inflection points, around the last couple of weeks? We just covered how Matthew Prince at Cloudflare said they were seeing hundreds of attacks prior to the PHP2Shell announcement and millions of attacks within three or four hours after it. Did you see an inflection point in the last 24 to 48 hours?

Evan Conrad: I can speak generally about what's going on in the news. For any company, security is a sensitive topic that is hard for us to discuss publicly, because of the nature of the topic.

Prakash Narayanan: It's a live issue. Yeah, I get that.

Evan Conrad: What we have seen is that there does appear to be a significant number of increased vulnerabilities that have been discovered by agents in the last six months to a year.

Prakash Narayanan: Okay, let's just segue. I think a few days ago, you announced you had 1,000 NVIDIA B300 GPUs you were offering. And the contracts, I think, were from three to five years. So tell us a little bit about how you put that kind of five-year offer together. Does it start with the seller, who says, "I have this excess capacity"? Why is it three to five years? Why is that not "we're only offering five years" or "we're only offering three years"? How does that flex come about? How do these contracts get shaped, starting from the seller and their needs?

Evan Conrad: San Francisco Compute's business is what we call supercomputing as a service. We go to folks who have some raw asset. They've got some power asset, or data center space. They could have sold that asset to another cloud, or they could become the cloud. Lots of people would like to suddenly offer GPU as a service because you make a lot of money offering GPU as a service. What we do is step in as their partner in doing that. We operate the clusters. We do the cloud portion of what they're doing, but they may be the owner of the data center under the hood, or of some raw power asset.

That's what we do. From our perspective, we start with the supplier, the folks with the data center space. We partner with the supplier, then we go out to market and we sell clusters. The reason something would be a three-to-five-year contract is mostly the lender. Right now, the bottleneck in compute, and in AI in general, is not power. It's not chips. It's entirely credit. In order to get debt, you need to work with somebody who has excellent credit, because your lender has no idea how to underwrite the value of

the collateral they've been given, which is a GPU cluster. If you think about it, there's a lender off in New York or London or somewhere. If the AI lab falls through and fails to pay, and then the cloud fails to pay its lender, the lender is going to recoup a supercomputer. And the lender has no idea what to do with that supercomputer, nor what it's going to be worth in the future. So they're pretty much forced to bet entirely on the credit quality of the customer, the AI lab. The problem is that pretty much every AI lab

does not have good credit. They're all new companies, and you don't know if they're going to be around. We're still trying to figure out if Anthropic and OpenAI are going to be around, or which one is going to win. That is a question in the air. So what hope does the two-person startup have that's trying to start a new lab? What that means is that in order for that new startup to get a cluster at all, they need to pay a very large deposit, and they need to do a long contract so that everyone is comfortable these people will be around for a long time. This is the problem that we're working on solving at San Francisco Compute. We build a forwards market that

makes it cheaper for folks to use. Basically, it makes it possible for a lender to know what the prices of GPUs are in the future, and gives them a place where they can sell and liquidate that collateral if they have it. This makes it possible for the costs of compute to go down, for the deposits customers have to pay to go down, and for the contracts to get shorter and more manageable. As you alluded to in the intro, the reason we do this is that originally we were a company called Junelark, and Junelark got stuck on this exact problem. So we're trying to make it possible for startups in AI to exist at all.

Prakash Narayanan: Let's talk a little bit about that. Let's say you are a typical London or New York based bank of some kind, and you get approached by a startup for a loan for them to buy GPUs. They want to set up their own... What are the layers? Are we talking about a neocloud which is trying to finance its own GPUs, so they're the ones getting the bank loan, and then they have a contract from a startup? Is that what we're talking about?

Evan Conrad: Very unfortunately, the way that lending works today in compute is that the cloud gets the loan and uses the GPU cluster as the collateral. The problem with this is that nobody knows what that collateral is worth. So who the cloud and the lender are actually underwriting is the customer.

Prakash Narayanan: Right.

Evan Conrad: And so you have several layers deep, basically.

Prakash Narayanan: And what do the lenders want to see from the customer? Is it a take-or-pay contract? But the take-or-pay contract needs a Meta or a Google to honor it. You can't have a 10% startup funded for just 18 months.

Evan Conrad: Correct. What both the lender and the cloud want is an investment-grade customer like a Meta or Google. They're likely not going to get this unless they've got a tremendous amount of capacity and are willing to basically not be a cloud. Meta and Google don't want you to be a cloud. So if you are trying to be a cloud, you're trying to sell to customers that are a little bit less than investment grade. And if you're an AI lab that is just trying to buy capacity, that's going to be very difficult unless you raise an enormous amount of money to pay for the contract in cash rather than

having to pay for it over time. If you can pay for a three-year contract entirely upfront, anyone will work with you. That's why you see AI startups raise these astronomical rounds before they've got revenue or product or traction or anything. It's survivorship bias. Lots of AI labs will go out and try to raise money. The only ones that you see that exist are the ones that are able to secure extremely large rounds, because nobody else can exist. Investors often wonder, which AI lab, which inference provider, should I fund? Which one's going to win? Which one's going to work? And the answer is the one that you fund. It almost doesn't

matter. If you've got the capital, you can build it right now, and it's pretty much just who the VC is going to fund. We started SF Compute because they weren't going to fund us. We were not able to raise an astronomical round when we started our little AI lab. So now we're building a company that tries to cause it to be the case that you don't need to raise the big round, because you don't need to pay the deposits, because the lender knows what the GPU cluster is worth. They don't have to worry about your credit as much, which means you can actually start an AI lab in the first place. We think if this doesn't happen,

there are just not going to be startups in AI. It's going to look like TSMC. If you want to start a new chip fab company today, you'd better find a brand new damn technology, because you're not going to be able to afford the CapEx required. You're going to need something like $50,000,000,000 to do it.

Prakash Narayanan: It sounds as though SF Compute is the one providing the assurance to the lender. Is that the correct way to think about it? Because you're saying the lender is able to use your economic perception of the asset instead of the startup's. Is that the right way to think about it?

Evan Conrad: How does anyone know what a GPU hour is worth? The way you know what a GPU hour is worth is that you've got somebody who's willing to buy it at that price, and somebody who's willing to buy it at that price in the future. The way you do that is with a forwards market. This is a concept that exists in basically every other commodity. And if you've already got the customer

Prakash Narayanan: who is

Evan Conrad: locked in to pay, then the lender is suddenly much happier, because they understand what that GPU hour is worth: somebody is contractually obligated to pay for it.

Prakash Narayanan: But it sounds as though that someone, again, has to be a Meta or a Google, a creditworthy counterparty which is willing to buy the forward GPU.

Evan Conrad: Not necessarily. Most companies can get a loan for just cash, and if you pay that cash, you're able to buy stuff. The question is more the economics that you're charging, the cost of that loan. What we want to make possible is, I'm trying to think about how to describe this better. Basically, your loan is a factor of two things: the credit quality of the person who took out the loan, and the collateral

that the lender would make back. If the collateral is bad and not worth anything, then the lender is going to be very scared about working with anyone other than Microsoft or Google. But if you make that collateral very solid, you know what it's worth, and it's not going to fluctuate very much, then you don't need Microsoft or Google on the other side of that loan. That's what we do.

Nathan Labenz: So I have a bunch of questions about who participates in this market and the technology that you've built to enable it. Let's maybe start on the sell side. For starters, who out there in the world today has a bunch of GPUs that they're just sitting on, thinking, "Yeah, maybe I should rent these out"? Presumably these companies, if they're finding themselves in this position, at one point had a plan for what they were going to do with these GPUs.

Evan Conrad: No. We're the cloud. What we do is work with folks who've got the power asset or the data center space, not the GPUs. They want to have the GPUs. They want to own them and have them on their balance sheet. A really simple example is you're a platform-as-a-service company, and you'd like to suddenly own the GPUs because you realize you'd make a lot of money if you owned those GPUs. Or you're a venture capital firm that realizes if you owned the GPUs, you could redirect them to your portfolio and capture the residual value, so you'd make a bunch of money. Or it's somebody with data center space or land or power, the raw assets, who

realized that if they also bought the GPUs, they'd make a bunch of money. All of these people don't know how to operate a supercomputer. Can I swear on this? Yes. Okay. They're going to fuck up the purchase. They're going to install the chips in the wrong damn data center. The whole thing, they're not going to meet their SLAs. It's a supercomputer. They're not realistically going to be able to do this, for most folks, because even the neoclouds can't realistically do it in many cases. Most of the neoclouds are struggling. What we do is take them up to a certain bar so they can meet their SLA. The reason we're able to do this is that our CTO is the cofounder of Voltage Park, and our VP of engineering is the former head of engineering at Lambda Labs. Our team is staff engineer after staff engineer of all the people who built the previous clouds. So we offer that as a service to people who would like to be the cloud themselves, and we're kind of like a GPU cluster property manager from

their perspective. And then on all the stuff that we operate, we get to operate a forwards market, which lets us do interesting things in the realm of lending.

Prakash Narayanan: Gotcha. So I guess what you're saying is you are making it so that a lender can seize the collateral and operate it if necessary. Because without you, it wouldn't be possible for the lender to operate that asset. Is that the right way to put it?

Evan Conrad: Kinda. Yeah. Right now, in a traditional GPU cloud, the lender is going to seize the asset, which is the GPU cloud. They don't have any way to know if they're going to be able to sell it, and they don't know how to operate it. They're just going to end up with a supercomputer they don't know what to do with. In our case, what they're going to end up with is an order on a book, something they can press a button to sell. And not only can they press a button to sell, but somebody was standing around waiting to buy the moment it was available for sale. That liquidity allows it to be better collateral than just the GPU cluster itself,

because you're not taking on the risk that you suddenly have to operate a supercomputer.

Prakash Narayanan: Right. So there are basically two parts of the business. One is the operations part, and one is the financial part, the forwards market. Is that the right way to look at it?

Evan Conrad: We're like two businesses in one. We're a hard tech company, which is the operations of a GPU cloud, and we're a marketplace business, which is the forwards market. By doing these two together, you get a very interesting effect on the world. Here's what it looks like when San Francisco Compute scales itself up, which is what we're doing at the moment. If you're an AI lab, you don't have to put down as big of a deposit. You don't have to go raise the big rounds. You don't have to convince the venture capitalist that you were the author on whatever paper. You could just go start your AI lab. And if you're a cloud, you get to succeed. What happens without San Francisco Compute is that all the clouds consolidate around who's got the best credit, and you'll only end up with a few clouds at all. Lots of folks in the

world are trying to prevent this from occurring. NVIDIA is backstopping folks to try to solve this problem, but their balance sheet can't take infinite backstops. You need a structural solution to this, and that's what we're doing. A lot of people start companies with a specific solution in mind. We started because we were the AI lab, and we hit all the problems. Then we just worked backwards from how to solve those particular problems. It ended up being the case that you have to do these two things in one, the supercomputing company and the forwards market. There's a lot of complexity in the business, but we're really excited about what it means for the industry.

Prakash Narayanan: Concretely, I think right now the cost for a cloud is something like $60,000,000,000 or $65,000,000,000 a gigawatt, with NVIDIA chips. So what does the loan structure look like at this point? Is it, "I'm going to have to put up 30% upfront as my equity, and then the bank will come in with 70%"? And if you have SF Compute in there, the interest rate will be 8%, and if you don't have SF Compute in there, the interest rate will be 15%? What does that look like?

Evan Conrad: Functionally, what we think is that by making the collateral more firm, you make the cost of the loan go down.

Prakash Narayanan: Mhmm.

Evan Conrad: So the interest rate goes down, and the down payment on that loan gets smaller. The LTV gets bigger, and the lender cares less about the credit risk of the customer. This means that more types of people can buy compute. It's not just Microsoft and Google, or even OpenAI. I don't think they're investment grade yet; they're almost there. You make it possible for other people to purchase this asset who may not have the same credit quality, and that's okay because the asset itself has value, so you're less concerned about the credit. Indeed.

That's really interesting.

Nathan Labenz: How big does a provider have to be to come into this market? If I wanted to create Nathan's little contribution out of my garage, my guess is it wouldn't be enough for you to bother with. Where does it start to be interesting to you?

Evan Conrad: Five megawatts or so is probably the minimum for us. It's difficult to get something working in your garage, in part because you probably don't have the ability to power it. But a smaller data center is something we can take over and handle. I do think you're going to see lots of smallish-scale stuff start to pop up around the world, under 15 megawatts of capacity, popping up all around. This is because in AI, you put the cluster in the data center next to the power, rather

than putting it next to the specific customer you're trying to target, because you don't care as much about latency. We're seeing a lot of those pop up all around the world. This is maybe different from the big mega, "we're going to do a whole ten-gigawatt campus all in one" sort of thing.

Nathan Labenz: Manhattan footprint, whatever. So I'm in Detroit. There's lots of auto factories here, some of which are under capacity, let's say. Is this the kind of thing a car company can get into because they've got some extra electricity sitting around? They can throw up a warehouse. They've got at least some balance sheet. They come to you and say, "Hey, this sounds like easy money, let's do it." Are we at that stage in the cycle yet?

Evan Conrad: If you have grid power and you have credit, we would love to work with you, and we would happily help you become a GPU cloud. I don't know if they would do it, or what that asset is currently being used for, but lots of people are taking things that already have grid power and trying to convert them into data centers. Does Michigan have a data center moratorium?

Prakash Narayanan: No.

Evan Conrad: Although

Nathan Labenz: it's a topic at the moment.

Prakash Narayanan: Okay.

Nathan Labenz: But they are being built. We'll see what happens after this next election.

I think that is pretty interesting because there are a lot of these sites. Anybody who's lived around Detroit will know the history of factories downsizing, factories closing. And this is an interesting way to think about, as you said, and I'm not saying anything you haven't already said, distributing these assets much more broadly and allowing people to participate in this asset class in a way that, without the know-how, they just wouldn't be able to. So that was very interesting.

Prakash Narayanan: One of the questions I would have is this. I'm an AI startup. I've just raised money, and I need some GPUs. So I come to you and say, "I need 1,000 GPUs, and I need to raise money." How do we put this financial contract together? I'm going to need to raise less money, right? How much money would I normally need to raise for 1,000 GPUs, and how much would I need to raise with SF Compute in there?

Evan Conrad: Today you're going to need to raise roughly the same. What we're doing is building up the ability for that collateral to be worthwhile. That's going to take us several years to do.

Prakash Narayanan: I see.

Evan Conrad: There's a reason nobody else really tries to do this at scale. It's because you have to just do it for a very long time to prove out the market. Lenders don't get comfortable with things very quickly. You have to show proof over a long period of time, and so that's what we're doing.

Prakash Narayanan: Okay. So I'm a startup, I come to you, and I have to raise the same amount of money. Once I've raised it, what do I go do? I'm not going to set up my own cloud, because I have to get into operation pronto. So what do I do? Do I go to a Nebius and say, "I'm going to rent the GPUs"? What is the next step here?

Evan Conrad: When we work with demand-side customers, typically the thing we're selling them is the capacity itself. We've got the capacity, and we are selling you the capacity. We make no illusions. Customers would like the GPUs.

Prakash Narayanan: So you're just selling them, like, one year, this year, for example, and you don't require them to pay the other four years? Is that what you do?

Evan Conrad: No, we lease them for three to five years. And the reason someone might go with us is that we let them get out of the contract. We let them sell that capacity to somebody else. That's one reason to use us. The other reason is simply that we have the capacity, and a lot of people's own. Part of the reason that we have capacity is that we are the partner to many different operators who would like to become clouds, and then we turn them into clouds.

Prakash Narayanan: Okay. So the marketplace is now being populated on the supply side first. And on the demand side, you have plenty of demand. Everyone needs GPUs, kind of.

Evan Conrad: We turn down 98% of all inbound demand.

Prakash Narayanan: Wow. Wow. Okay.

That maybe relates

Nathan Labenz: to another question I had around the gap between hyperscaler pricing and neocloud pricing. It's pretty wide, and I'm surprised by how wide it is. I wonder how long that can endure. Maybe part of the reason it's high is that it's much more of a spot market, and so they're tolerating some amount of wastage in order to have the capacity on demand and charge these much higher prices. What do you think is the future of that? How is that gap going to evolve?

Evan Conrad: Hyperscalers do not have a significant price premium relative to neoclouds. It's an illusion. If you're selling on demand, on demand has a price premium relative to reserved contracts. Neoclouds are mostly selling reserved contracts, and there's a discount for having a longer-term contract. You're giving up a premium that you would have made if you were willing to have the cluster at risk, with the customer able to walk away at any point, and that premium is pushed onto the customer. One of the things that is different about an AWS or a GCP or a hyperscaler is that if you pay their $11 or $15 an hour, depending on what you're trying to buy, on-demand price, you'll

get that capacity. You won't get an infinite amount of capacity, but it's available to buy. And notably, the reason it's available to buy is that it's so damn expensive that they're able to make up for the idle utilization they have to keep around in order for you to be able to get that capacity in the first place. Meanwhile, that isn't something most of the market actually buys, because most people don't want to pay that price. The revealed preference is that people do not want to buy on demand at the price that you actually have to charge in order to make it viable.

Prakash Narayanan: And is that like... I saw Elon sold some GPUs at, like, $100,000,000 per megawatt or something like that. Is that the kind of on-demand pricing you see for large contracts?

Evan Conrad: That's not the same thing as on demand. I would say that's a short-term contract, or a contract you can get within a short lead time. By on demand, I mean it works like a light switch: you turn on the light and it's on, you turn off the light and it's off. To make that work, you have to take the risk that somebody's going to turn it off and maybe you won't have a customer, so you have to charge a higher price for the fact that somebody might leave. A reserved contract doesn't work that way; they're locked in with you. But yes, one of the things someone like Elon has done is that they had the capacity at the right moment. They can turn it on tomorrow or very quickly, rather than months from now, or possibly even a quarter from now or a year from now. It's a whole shit ton of capacity, and it's going to come online

really quickly. And you can charge a really high premium for that.

Nathan Labenz: So I know you have developed this technology that you call physical settlement. I don't know if it's a technology or a capability, maybe a little bit of both. How do you see intermediate offerings taking shape? We have five-year contracts on the one extreme, and the light switch on and off on the other extreme, and it seems like this physical settlement idea would be an attempt to create a more liquid in-between, where I can say, "I'll commit to three months," and you can deliver that and retain flexibility on the back end of it. How is that taking shape, and what more middle options should we expect?

Evan Conrad: When we say physical settlement, we mean the entire operating capacity of a GPU cloud without owning the GPU cloud. We will do site selection. We'll do the design of the cluster, the design of the bill of materials, backlink components, bare metal, VMs, the network, and we own the whole SLA. All the things that your cloud would do, we do, but we do it on behalf of other people. But it's the forwards market that makes it possible for you to have shorter contracts. What happens is somebody buys a longer contract, then they sell some portion of it, and that opens up a shorter-term contract for somebody else. That's the gist, and it's part of the reason we got started, because originally we were that. We had bought way too long of a contract, and we were opening up shorter contracts for other people. And we were able to charge a premium

for that because we didn't force you to sign for a year-long contract. You could sign for a month with us, and we would charge you a premium on that. So a lot of people are able to make that trade.

Nathan Labenz: So really, an Airbnb story. I see you out there with your [unclear], selling month three of your GPU contract. And here you are today. That's a pretty interesting backstory.

Prakash Narayanan: I just had a maybe-a-segue question. I understand GPU clusters are very different, and there's a geographic positioning as well. Do you see different kinds of models? First, are you operating outside the US? Are you dealing with European firms? Do you see demand in Europe or Asia? Is your system able to operate outside the US, or is there a restriction? And is there demand? Do you see inquiries at all? Or is this really just a US-centric market and nothing outside the US really matters?

Evan Conrad: We operate around the world today. When you go to Claude or ChatGPT and present a prompt, there is an initial latency, the time to first token. You want that to be fast so that it feels good and the experience is quick. But then you mostly care about throughput, how much you're pushing out of it, and that's not latency sensitive. If you're going to run something for days, it doesn't matter if it's 200 milliseconds more. You just want the cheap thing that you can get a lot of, so you can push things through. So yes, we're seeing

capacity demands all around the world. This isn't the case for every single customer. Some types of inference don't work that way, especially something like live video or audio. But the thing that matters in this case is that every country is trying to have an AI strategy of some sort. And from an economics perspective, the way you get the value of AI to flow to your country or your state or your city is you put data centers in it. If you think about the data centers going up in the United States, there was a big one in Utah that people were

very unhappy about. Nonetheless, that tripled the tax revenue of the local area. That single project was three times bigger than the entire tax revenue of the area. So in the United States, there are lots of people who, for maybe moral-panic reasons, are very concerned about data centers, often for things that I think are just factually incorrect. The rest of the world doesn't have the option of being factually incorrect, doesn't have the option of having the moral panic. So if you can build a data center somewhere else, in a country that would really like the economic benefit from it, you can put in a project that's quite meaningful to that economy in

that country. This is sovereign AI. Lots of people are trying to do this. One of the things we do is that we are the San Francisco partner for sovereign AI. We will be your supercomputing partner. We bring the expertise from Silicon Valley, and we also bring the offtake from Silicon Valley. A way to think about it, and this is where we think things will go, is to take the project that was Stargate. That's a $500,000,000,000 project. $500,000,000,000 is more than the entire GDP of multiple countries. So having any portion of that project land in your country is an enormous economic development

opportunity for you. The same thing goes for the states. If you're putting a data center project in West Virginia, that is life-changing for folks in the area, in part simply because of the pure tax revenue that you could use to benefit that community. Now, lots of people will, I think, be very concerned about it out of fabricated panic, but the rest of the world is not. So we're seeing it.

Prakash Narayanan: Do you see only liquidity for NVIDIA chips? Are people using the AMD MI300? All of these other chip companies are trying to put new chips out there. Are you seeing markets for other chips as well, or only NVIDIA?

Evan Conrad: We don't really get asked for chips that aren't NVIDIA. We primarily go based on what people are coming in the door asking us for. So we do NVIDIA, and we haven't really explored too far out on other chips.

Prakash Narayanan: Incredible.

Nathan Labenz: Do you want to tell us a little bit about this Autoresearch project, or product, as well?

Evan Conrad: Yeah. We built a product called Autoresearch. It sits on top of SF Compute. The rationale is basically twofold. One is that we're an order book, so when people use us there's a little bit of extra complexity that we ask them to deal with. Think of your power company: when you turn on the lights, you're not signing a PPA or trading on a power market. You just turn them on and turn them off. You're a participant in the market when you're doing that, just an indirect participant; it's abstracted away from you. The same thing happens when you fill up your gas tank. You're participating in the oil markets, but you're

not directly trading on it. We built something that's kind of a UX wrapper over SF Compute, specifically for researchers and small experimenters. We wanted to make it the case that if you looked like we did, we could get you capacity in a way that is viable for you. The key pitch of Autoresearch is that we get you a training cluster with InfiniBand that you can schedule a job for, and we'll run it as cheaply as we can. You can queue that up, and you don't have to sign a long-term contract. It's the proof point that our economic system at SF Compute indeed

does solve the problem that Junelark was having, and it's something lots of other clouds can't offer. So we did, and we're excited about it.

Nathan Labenz: In practice, how much of this is almost structurally like a fractional reserve type thing, it seems like, where somebody takes a bunch of GPU allocation on their balance sheet in the form of a commitment to you, but then they can resell, and they might resell a lot of it or a little bit? In practice, how much resale are you seeing past that first level where a commitment is made?

Evan Conrad: We do two things. We have some capacity that we've sold as a way to cause resale

to occur, in the case that GPU demand goes down, and then we have some capacity that... Are you familiar with Henry George?

Nathan Labenz: Also, no?

Evan Conrad: There was an old economist named Henry George, and he created a thing called a land value tax.

Nathan Labenz: Oh, okay. Yes, George's tax. I do know this. Yeah.

Evan Conrad: Yes. What we did is make some of our capacity available for sale with a land value tax. Actually, it's technically a Harberger tax, which is an offshoot of a land value tax. It basically says, look, we'll sell you the capacity, but somebody else has to be able to buy some portion of that capacity from you. This forces liquidity onto the order book in the first place. It's very controversial. Basically, we discourage hoarding on SF Compute, because if you hoard all the capacity, no one else can get it. Autoresearch in particular does this, but even further. So

the way Autoresearch works is it's a big queue, and you can submit jobs to the queue. You can pay us for priority in that queue. You can go faster. But you can't pay for all of the priority. If you're a tiny little researcher with a tiny job, if you want to just run something every so often, we will put you at the front of the queue so that the tiny little startup gets the most advantage relative to everyone else. And then we do similar things on the rest of SF Compute with taxes and other mechanisms like this, where we say, look, you can get this capacity from us, but we're going to require that some of it's for sale so nobody hoards the capacity. This makes us quite liquid, and it also means that SF Compute always has capacity. Right now, basically every GPU

cloud is sold out. San Francisco Compute is not sold out. We're continuously not sold out because we've locked in our ability to not sell out for years. It's a brand promise we're trying to make: if you're stuck right now and you can't get capacity, we want to help you. We did the thing that is actually kind of worse for our business, by not just selling all of our capacity in this particular way, but it makes it possible for the little startup to actually get capacity.

Prakash Narayanan: I could talk about this for hours, but we're limited by time, so I have one last question. I think you had a [unclear] live appearance where you talked about a credit crunch in three months, I believe. You said there's going to be a credit crunch in three months. Where do you see this credit crunch going? Are we going to see a distressed credit cycle? Do you see credit tightening, and people not being able to fund their data centers? What do we see in the future?

Evan Conrad: Not to toot my own horn, but I think I was almost exact on that timing. I don't remember when that [unclear] live thing was, but it seems like the debt is getting more expensive, although I think there's a lot of complexity in that one. It also seems very hard to close new customers unless you're investment grade, and many customers are being asked to pay 100% upfront on their contracts. This is what we've been seeing for years and what we've been worried about. It's what we're trying to prevent. But I think for the next possibly year or two, if you're a VC, you should be

prepared to write some really big checks for your portfolio, or encourage the companies to sell. I think we're probably going to see a lot of consolidation of AI labs and inference companies. The reason is that in order to get the next cluster and grow, they're going to be asked to pay for the entire contract upfront, and they won't be able to do that. So they're going to be forced to sell their company to somebody else who has great credit and doesn't have to do that. Alternatively, they'll go out to investors and try to raise big rounds. The investors will say no because, at some level of scale, they'll probably just not be able to stomach the risk. And

I think when that happens, we'll see fewer startups exist, and that's why we do this.

Nathan Labenz: Fascinating stuff. It's quite a backstory. It's quite a financial and GPU alchemy project that you guys are putting together. And from what I understand of it, I do hope it works, because I'd love to see a future in which we can continue to have startups, and in which we don't collapse into five big tech companies for all future time.

Evan Conrad: That is also my concern. Me too.

Let's go do it. Cool. Thank you, friends.

Nathan Labenz: Thanks for being here. Keep up the good work.

Prakash Narayanan: Bye bye, Evan.

Evan Conrad: See you guys.

Nathan Labenz: Fascinating stuff. That was a great example of why sometimes you have to have the conversation, because I was not getting all that from the website. I'm not sure it would even be possible to get all that from the website, maybe if I was a little smarter. But really interesting stuff there in terms of how they are trying to create some buffer against these tectonic forces that, in the absence of any intervention, would just create a sort of Piketty-style, rich-get-richer phenomenon. That does seem, unfortunately, quite plausible given the demand for GPUs seems to be basically

infinite at this point. Obviously, if that were to change, all the assumptions could change, but it's looking unlikely if I just judge by my own usage, week over week. So I suspect the demand will continue to be there, and it will be very nice that there's at least somebody who is willing to play ball with the small guys on both sides of the market.

Prakash Narayanan: Let us segue to our next

Prakash Narayanan: Our next guest is Lonis Hamaili. He's the managing partner of Cambrian, a public markets investment fund focused on the physical supply chain behind artificial intelligence. The fund announced its launch on October 13 with approximately $70,000,000 in seed capital. Its focus is on the companies that supply the materials, components, and manufacturing capacity needed to build AI infrastructure. Before starting Cambrian, Lonis built software businesses. In April 2023, he introduced Godmode, a web application that let people experiment with AI agents carrying out tasks through a sequence of steps. He described

it as an Easter project inspired by other early agent experiments. That July, he reported that it had passed 1,000,000 users and 15,000,000 completed tasks in less than 3 months. He also cofounded dWork, a task management platform for organizations using blockchain-based coordination and payments. dWork announced a $5,000,000 seed round in 2022. His work extends beyond software at Próspera, a special economic zone on the Honduran island of Roatán. He served as head of governance and vice president of growth, helping attract entrepreneurs and businesses. And let me bring Lonis on.

Lonis, welcome to the show.

Lonis Hamaili: Hey. Good morning, guys.

Prakash Narayanan: Sorry for butchering the pronunciation.

Lonis Hamaili: No, it's close enough for sure.

Prakash Narayanan: So let's talk about the launch. You came out with a bang on October 13th, and you announced your thesis. You posted. You're on MTS Live. What is the fund? What does it do? What is the thesis?

Lonis Hamaili: Yeah. Well, the fund originated from a lot of soul searching around a year ago. I've been AI-pilled since 2017, when I started reading Bostrom and Yudkowsky, et cetera. But a year ago it felt very clear that, okay, we're getting AGI in this decade. And I looked at my own portfolio, and I realized that it was not only not correlated to AI, it was actually decorrelated, anticorrelated, because I had all of these venture paper gains from software startups, which would probably be quite worthless if we get AGI soon. I started talking to friends, and their portfolios were even worse. Some of them had commercial real estate. Some of them had bonds. Some of them just held the S&P. But all of these assets

are not set to perform well at all if we have short timelines. So I was thinking a lot about how to solve this. I was very lucky to invest in Anthropic, by an SPV. But even investing in labs is not a perfect solution, because I don't think the government will necessarily allow them to get that much larger, and they would be extremely politically exposed. AI is already super unpopular for no good reason, and this year there will be many good reasons. There will be misuse. There will be hacks. There will be job loss. So these companies are extremely politically exposed. And so I came to the realization that actually the supply chain is the one that wins in every scenario, whether the labs keep on leading, whether

the government gets heavily involved, whether the Chinese catch up. All of them need the same memory, the same tools, and the same power. That's the high-level thesis on how I started. And at that point I caught up with my cofounder at the fund, who has been trading this thesis on the supply chain for several years. He had turned his family office from around $100k to $14,000,000, trading the supply chain. And so we started doing it together, and three or four months ago we started off as a fund.

Nathan Labenz: You want to tell a little bit of that backstory? I heard it from him in person, I don't know, two years ago or so. To call it a family office is maybe a little bit grandiose. I mean the appetite for risk that he was able to get his family comfortable with, when what they had was not what people would normally think of as a family office. Today it's maybe more of a proper family office, but I thought that's a really interesting backstory.

Lonis Hamaili: Yeah. I guess [unclear], he traded up his own book from around $300 as a student to $1.2 or $1.3 million, using options. And then separately, he got his dad to take out a loan on his house for more than his liquid net worth. So he took out the $100k loan on the house, and then he traded that $100k to around $14,000,000 today. Yeah.

Nathan Labenz: So now Dad's got an actual family office. At a very high level, how is this not priced in? Maybe that's a basic question, but in 2017 you were early. In 2026, it's all staring us right in the face. The skeptical take would be that everybody knows about all this at this point. Where do you see that that is not the case?

Lonis Hamaili: Yeah, it's very confusing. If you look at how these semi stocks are trading, they're trading at 16 times earnings or so. The S&P on average is trading at 22 times earnings, and Walmart is trading at almost 40 times earnings. So not only does the market not think that the current earnings will grow, it thinks they are less durable than the earnings of the rest of the market. And the companies that we're very excited about, whether it's silicon wafers or memory or capacitors or foundries, are trading at 3, 4, 5 times sales. So, extremely underpriced. And

there may be a few reasons why this mispricing exists. One is that most people are not AI-pilled. Even in the Bay, some people are not AI-pilled, but if you go to the East Coast, almost nobody is. I think that's one big input. The second is that we have this thesis that the private market is actually more efficient when it comes to AI than the public markets. In the private markets, you have all these ex-operators, technical people, who are bidding up a small set of startups to high multiples, 50 times sales, 100 times sales. In some cases that's rationally high, because people here are AI-pilled and technical. On the Street, it's a very different setup, where very few people are AI-pilled. As a share of overall traded money it's super, super small, and the market is much larger. So the market structure is just very different.

Prakash Narayanan: That is weird.

Nathan Labenz: I'm just asking the same question again, maybe, but I'm still kind of like, how can this be? If you had to steelman why these companies are not mispriced, could you do it?

Lonis Hamaili: Yeah. Okay. One steelman is that many of these businesses are not great businesses in normal worlds. Especially the atoms businesses we like, whether that's capacitors, silicon wafers, what have you. They have high fixed costs, and if they don't hit a certain utilization rate they actually lose money, and in some cases they carry high debt. So in normal worlds they're just not great businesses. It's only in the world where AI really hits that they get this big operating leverage and become amazing businesses. Today, say they have 10% gross margins. That's a terrible business by Wall Street standards. But if they're able to raise prices by 10%, they suddenly double earnings. And if they're able to double prices, they 10x earnings.

But that only happens in this very specific AI world. In a normal world that basically never happens, because usually suppliers are very good at matching demand patterns with supply ramp. So, yeah.

Prakash Narayanan: Let's talk a little bit about this. A lot of the thesis concerns AGI and AGI emergence, et cetera, and I am trying to get you to justify my thinking, so I'm very motivated to get a good answer. But let's say you have the market as it is right now, and all of these companies. Where do you think this AGI ramp goes? Is there a specific AGI point or superintelligence point or recursive self-improvement point that you think happens? And what is your P10 to P90 on when that happens? Are you saying, you know, 10% chance it happens by next week and 90% chance it happens by 2035?

What is your range of outcomes in terms of the curve? And what is that moment that happens? Is that the RSI moment? The superintelligence moment? What do you see as that thing?

Lonis Hamaili: Yeah. I think RSI is the process that takes us to superintelligence, and RSI is already happening. Anthropic recently had a blog post where they show that Claude now leads more than a quarter of all their own research, and that is up from 1% in March. So the RSI loop is already there. It's already closing. As for when that will lead to superintelligence, it's much harder to say. I would generally say maybe a two-thirds chance it happens within this [unclear]. Maybe 50%, maybe two-thirds.

Prakash Narayanan: Wait, wait. So you have a 66% chance it happens in the next 3 months?

Lonis Hamaili: No, sorry. Before the end of the decade, I should say. Actually, two-thirds that it happens before the end of the decade. Before 2030.

Prakash Narayanan: Yeah. Okay. So one of the things is, let's say we have memory. Right now, what the market perceives is that by 2028, 2029 there's going to be an oversupply of memory. That is a market perception. So your thesis obviously assumes that these companies will still have a shortage in 2028, 2029, which is why they're going to have these elevated margins and high operating leverage, and so you're going to get massive PE expansion. So why is your thesis different from the market's thesis? What about your thesis is different? Are you expecting massive revenue increase next year and another round of massive CapEx going forward? How is your thesis different from the market?

Lonis Hamaili: Yeah. The interesting thing with memory specifically is that it doesn't require that many things to go right. It's already trading at only 5 or 6 times earnings. Even if they can't raise prices further, at some point they should rerate to at least 10 or 12 times earnings. So it doesn't require prices to increase. It does require demand to continue to outpace supply. And when you break down supply, almost no expert thinks we can more than, at most, 2.5x DRAM bits by the end of the decade. Nobody thinks we can 3x. I have zero experts with a track record who have said we can 3x DRAM supply by the end of the decade. So on a very naive level, you have to assume that

okay, demand will not 3x by the end of the decade, which of course everyone here would assume it will. Then there's the separate term in the equation, which is memory compression. How much memory compression do we get? How many bits do we actually need per token? We've been looking into how this developed from GPT-3 to DeepSeek V4, over the last four years. The bits required per token decreased by something like 500x. And within that same time, when bits per token decreased 500x, the DRAM bit price rose by 6x or 7x. So that's part of

the Jevons paradox, and so on. Yeah.

Prakash Narayanan: So what do you think is the increase in compute to 2029? From today's compute, are we going to get 10,000x, including the chips and the increased data centers, et cetera? Are you looking at 10,000x, a million x? And if you have that 10,000x or million x, and then you only have a 500x improvement in memory, is that what drives this uplift in prices?

Lonis Hamaili: Yeah. Again, memory doesn't require an uplift in prices. Memory should 3x from here just on rerating to a more reasonable multiple. But for prices to remain, you need demand to outpace supply, which again is not a hard thing to ask for, given how hard it is to ramp DRAM bit supply.

Prakash Narayanan: So we just spoke to Evan Conrad, who runs a compute company, and he's predicting a credit crunch in the next 3 months or so. You're going to start seeing this credit crunch, so you're going to have difficulty for the capital system to supply enough capital to actually build out the data centers. This brings up the idea that now we have maybe $1.2 trillion of CapEx for 2027, and by 2028 we're not going to get 2.5 or 3 trillion, because the capital system is just tapped out at roughly a trillion plus. So how does that work in terms of expanding demand in the 2028, 2029 period?

Lonis Hamaili: Yeah. In the short term, there's a good chance the market keeps being very worried. I tend to think quite strongly that the ROI of AI will get increasingly priced in over the coming quarter, such that next year this will be less of an issue. It is true that the global savings market is just not that large. There are limits to how much credit there is. But we haven't reached that limit yet. The pool is around [unclear] trillion dollars, depending on how you count it. And the unique thing with AI is that the issuers are very happy to pay high rates because the ROI is so high.

Right? Currently you can buy 20-year US Treasuries and get 5%, or you can buy, you know, Meta or Google AI corporate debt and get way more. And the thing with [unclear] compute: my analysis recently had an article where they showed what the payback period is. If you do nothing else, just have a Vera Rubin rack and serve DeepSeek tokens on OpenRouter, they showed that the payback period just doing that alone is 3, 4 months. So the ROI on compute is absolutely massive. In theory, these AI companies should be happy to pay 15% for their debt.

20%. And so yeah, I predict this will get increasingly priced in, maybe not in the coming quarter, but certainly by next year.

Nathan Labenz: So you're obviously modeling this out, and there are all these different trends you have to take into account. There's the scale at which computing is ramping, but then also the scale at which algorithmic efficiency is driving more usage out of that compute. And then there's the uptake of AI adoption by individuals and corporates. What do you think are the most unclear curves, the ones you have the greatest uncertainty about, that could break the model over the next couple of years? And maybe even before answering that, a couple of anchor points would be useful. For example,

what are you expecting tokens per capita to be? Or however you think about what the global usage base might look like. What are some anchor points we can imagine, numerically, for what 2028, 2029 looks like? And then which of these trends do you think are the hardest to really be confident in?

Lonis Hamaili: Yeah. When breaking down the buyers of AI: consumers in general are terrible customers. We're four or five years into ChatGPT, and consumers are still only paying something like $20 to $30 billion per year. It's a terrible, terrible business. Enterprise is doing a little bit better this year, maybe up to $200 billion. But the real buyer would be governments, and that hasn't happened yet at all. Next year, we'll see governments become the largest buyer of tokens. We'll see a very significant ramp-up in intergovernmental competition between China and the US. There will be an arms race in hacking, and you need the frontier tokens to do that well. And so the government can easily pay

$500 billion to make sure that they are not taxed by China, which would be a political nightmare for anyone in power. So on a high level, we're not super bullish on consumers. Personal agents like [unclear] are quite nice, they're funny, but ultimately I don't think the market is that big overall. I think the enterprise market can grow massively. Every year we pay $30 trillion to white collar workers, ex China, ex software engineers, and they haven't really gotten any significant productivity boost yet. It's a little bit like before Claude Code. Last year, the best coders in the world still wrote 80% of the code themselves. And then by January and February of this year, they wrote 10% of the code at most. We haven't seen that Claude Code moment yet for white collar enterprise.

But when we see that, we think you get an order of magnitude or more increase in enterprise spend from that. So those, I guess, are the big buckets.

Nathan Labenz: So that means if enterprise spend is $200 billion today and it goes to $2 trillion, it's still only 6 or 7% of

Prakash Narayanan: Global labor.

Lonis Hamaili: Correct. White collar labor. Yeah.

Nathan Labenz: Yeah. Okay. So that's an interesting comparison anchor point. Tell me more about this government thesis, because obviously the government can spend a lot of money quickly on something when it chooses to. We've seen that with increasing frequency in my lifetime. But can they actually use these tokens? It's one thing to say, okay, we're going to spend a couple hundred billion on AI. Now you've got whatever quadrillion tokens you have to use. Can the government actually point a quadrillion tokens at anything even remotely useful?

Lonis Hamaili: Yeah. First of all, governments are happy to pay higher margins. They are not as rational buyers as others. They're paying for a bunch of features other than just the per-token pricing. Intelligence operations, for example, can be very token-expensive. And there are so many added features the government needs on top of just per-token: all of the security, and they want everything on-prem, very likely, so they'll build out their own data centers to serve these things. So there will be a lot of additional spend coming from government. So one of

Prakash Narayanan: the confusions for me is that governments are typically very slow. Number one, they're slow. Number two, all governments worldwide are at debt capacity limits. Everyone has too much debt. And whenever, say, a government in the US decides to spend a trillion dollars, there's always the question of why you don't spend it on curing homelessness or something. You've also had this issue in the US where I think they spent an additional $800 billion on the Iran war, and they had to restock their weapons, et cetera. I think the NSA so far this year has increased its spend on AI by a couple of billion, $2 billion or so.

And the question for me is whether it's going to be seen as a giveaway to these companies. Because the government can also do this Defense Production Act thing, and say, look, we're going to cap your returns. And the moment you cap returns in a sector, you actually lose the investment, because you say, of course, I can't make that money, I'm going to get out of the sector.

So how does this fit together? And you've also said you expect a lot of political pushback to the larger buildout. So why would you think any government, or any set of governments, would spend more, given number one the debt issues, and number two the ability to cap pricing on purpose and use the Defense Production Act? Why would you think governments will become a big purchaser at high prices?

Lonis Hamaili: Yes. So they will not do it until there's a catalyst. There needs to be a catalyst. But we predict that next year there will certainly be catalysts. We will see significant hacks of banks and other institutions. It's really when we get this feeling that some foreign adversaries might be hacking into our local governments or federal governments and getting access to different types of information. That's when it becomes scary, and that's when it will become very easy from a political perspective to say, no, we need to make sure that we're always staying ahead. As you say, there is the Defense Production Act. They could use that, but I don't think so, since that would lead to what you say about the investment, and that is the last thing the government will want in that scenario. In that catalyst, the government will want more investment

from American firms in AI. And so I'm sure there would be other ways to work that out.

Prakash Narayanan: Indeed.

Lonis Hamaili: And in general, it's interesting to look at the markets today. It's very similar to how it felt last year at the same time, when OpenAI was at $13 billion revenue and everyone was freaking out on Twitter that they were only at $13 billion. Some people were out for 15 or whatever, and the markets were down for several months. It's the exact same thing now. We're on edge because it was off by a few billion, and the markets are massively down as a result. But of course, last year the reason OpenAI was plateauing a little bit was just that the chat paradigm can only go so far. There's only so much you can charge consumers. If you were in the know, you knew we were going to get the coding agents.

And we got the coding agents, and now we're about to get enterprise automation. We're about to get computer use agents. But it's not anything that the market is pricing in.

Prakash Narayanan: Okay, let me flip it a bit. What would you need to see in the market to say that memory prices are going to go down? What is the antithesis to your thesis? What kinds of things would you be looking for to say, look, we're probably wrong, we need to flip here, or we need at least to retire from the market for a year or whatever, the timelines are basically going to be longer? What is the antithesis, and what would trigger it?

Lonis Hamaili: Sure. Again, I think there are better investments than memory. But if you're asking about memory, the main thing would be: we get superhuman coders, and how much better would memory compression get? Maybe it just happens that we get memory compression faster than we get other economically useful features, such that we don't get the Jevons paradox in the short term. That could be one thing. I think that's relatively unlikely for a few reasons, but that would be one. For the semi industry, if for whatever reason energy becomes a larger bottleneck than semis, then there's less incentive to bid up the marginal wafer or the marginal memory bits to such high prices if you have to wait a quarter anyway to get power in. And that's a little bit the kind of regime we've been in for parts of this year.

There are a few reasons why we think that won't be the case going forward. We've finally gotten a hang of doing behind-the-meter things. People are increasingly not connecting to the grid. They're building their own power plants with their own natural gas. A big reason you can scale energy production faster is that there's so much more latent supply. For instance, the frame turbines, the large turbines that turn natural gas into electricity, are sold out until 2030 from GE Vernova and Siemens, et cetera. But then people find other ways. They find that you can use truck motors and things from airplanes, and stitch things together and make it work. Because there's so much latent

stuff you can make use of, which is very different from the semi supply chain. There's no latent supply of wafers or of memory bits. All of that is already used. There's no fat, essentially. And there are no others to really bid out at this point. And so we think semis will be a larger bottleneck than energy.

Prakash Narayanan: So basically, sorry, the two antitheses are, number one, you actually get much greater innovation, which allows less memory consumption per token. That's number one.

Lonis Hamaili: Well, specifically we get a nonlinear improvement in specific programming ability that leads to better compression rates, but not the broader economically useful things that would make the Jevons paradox kick in. Because generally we would assume that if we can do superhuman memory compression, we can also do superhuman other things, and then all demand would just skyrocket.

Prakash Narayanan: I see. So if we get real superintelligence and real superhuman compression on memory, you're also going to get superhuman everything else, and then the demand for intelligence is just going to be absolutely infinite. Everything's going to blow up, and the physical supply chain is still physical supply chain. It's still constrained.

Lonis Hamaili: Exactly. Yeah. And in general, when we're looking at the AI market, we're way more bullish on the atoms over the bits. Most of the AI companies people look at are NVIDIA, Broadcom, ARM, Marvell. These are not bad companies, there's nothing wrong with them. The issue is that their business is essentially using OpenAI and Anthropic models to design software, and then they send that software down to Korea and Japan and Taiwan for the actual work. These companies don't have their own physical know-how or their own factories. And so

Prakash Narayanan: Which companies are these?

Lonis Hamaili: NVIDIA, Broadcom, et cetera, et cetera. These are great, well-run businesses, but they don't have anywhere close to the same moats as the physical companies, which have massive fabs and decades of know-how built up, and it's just very hard. There will never be a situation where the labs compete with a silicon wafer maker, and we will always be a customer. Whereas we're already seeing the labs compete directly with these software designers, like NVIDIA and Broadcom, et cetera.

Prakash Narayanan: We're getting a little bit of interference. Is it something... It might be me.

Nathan Labenz: Sorry, there's a leaf blower, I think, outside my window. I think that's on me. Let me mute. Let me ask one thing, and then I'll mute and listen to the answer. Does this all mean that the biggest risk is political? Is NIMBYism the number one threat to this thesis?

Lonis Hamaili: NIMBYism is a big threat. The bullish thing for the US in general is that you can ban data centers in most states, but it's fine as long as you have a few important ones. As long as you have Georgia, Texas, Mississippi, Virginia, you're mostly chilling. It's largely not a problem if people in New York, California, Seattle, Washington start to ban these things. They will probably do it for political reasons, but it doesn't actually matter so much. You only need a few key states, because the distribution of natural

gas is very, very nonlinear.

Prakash Narayanan: So let's look at Elon's thesis. Elon is building out a TeraFab, and it looks like it's going to be a 2030 thing. He says 2028, but there's no way he's going to get there by 2028, so we're looking at 2030. One of the things that has struck me is that Elon is specifically saying that if you need to 10x or 100x capacity, the existing players will just not do it. They'll be at about 2x maximum by 2028, maybe 3 or 3.5x by 2030. So how does that work? Number one, it sounds as though that thesis requires RSI not to happen before 2030.

And number two, according to your thesis, even if it does happen, having the physical architecture done is going to be valuable. But it also sounds as though the existing players you're investing in will not be able to get there. Elon doesn't believe they'll be able to get there, and so he's going to build something that's 10 or 100x bigger, which will devalue the current set of players you're investing in. How do you think about that mental picture of the TeraFab coming up and disrupting the market?

Lonis Hamaili: Yeah. I think the TeraFab is super promising. Obviously he knows how to run these large-scale real-world projects. I think, though, that many of the bottlenecks the other players are running into, he will still be running into. There are just so many silicon wafers. He will not make his own silicon wafers. He will have to buy them from one of the three or four current players, and supply of that is only growing 9 to 15% per year. He will have to buy the same tooling as everyone else, the same EUV machines, the same deposition machines, all these other raw inputs and machines that others need, and he will have to bid them up. So it's probably a great venture, but it's not clear to me that it will lead to that much extra added

net capacity until we resolve all the other upstream bottlenecks.

Prakash Narayanan: From what I hear is happening at TeraFab, it's going to be a recursive process. They're buying from existing suppliers right now, but at the same time they're looking at how to design the next layer, which has more supply chain integration, which is what he's done in every single business he's gone into. He started off with the existing set of suppliers, identified all of the parts that needed integrating, and then integrated those. So it sounds as though he's looking at two or three cycles, and by 2030 the real TeraFab will have all this stuff integrated. Does that mean he's going to deprecate all of the existing suppliers? Because that's what he did in rockets, for example. Over a period of 10, 15 years, all these guys used to supply precision cast parts, with Precision Castparts, which Warren

Buffett owns. They can supply these very expensive, say $5,000 metal casts, and Elon would redesign them for $50.

Lonis Hamaili: Sure.

Prakash Narayanan: So how would that work with these chip suppliers?

Lonis Hamaili: No, I think Elon will do really well. I think he will find ways to innovate. I think, though, there's a very big difference between entering the semi market and entering the rockets market. In rockets, he was competing with governments, which is really what he always wants to do: generally incompetent players, without any real profit incentives, not a real market. To a lesser extent, but still broadly, the same is true for cars, especially the way he was entering it. The semi market is way more competitive. There are many more very sophisticated players. It would be way harder, we predict, to find the very low-hanging fruit that he could find in other industries. So I fundamentally don't think he will be able to build

all of these types of machines himself. It's very unlikely, especially in the coming 3, 4, 5 years. But I think they will do very well. They will probably add some marginal supply, but ultimately they will also be bottlenecked by many of the same upstream materials and tools that others have to buy.

Nathan Labenz: As you evaluate the supply chain, how do you determine what's going to be hard even for somebody like Elon? And obviously you've also got to take into account infinite AI agents to help do whatever needs to be done. You've a couple of times mentioned silicon wafers in particular. I don't know a lot about them, but it's surprising, at a superficial level, that that would be a really hard thing that he wouldn't be able to do. There's plenty of silicon out there. Is it so hard to make a wafer, compared to these other processes, that this will remain a well-moated business even if Elon turns his attention to it? And

I guess more just epistemologically, how do you try to get your arms around that for all these different links in the chain?

Lonis Hamaili: Yeah. [unclear] likes to give silicon wafers as an example because it's an easy-to-understand thing, but there are many hundreds of layers in the supply chain. The silicon wafers have hundreds of suppliers themselves, and there are many companies that aren't buying their own silicon wafers. The key thing to understand with all these layers is that there are different tech tiers to them. In memory, we have HBM3, we have HBM4, we had HBM2 before, and these tiers always add increasing difficulty. So there's always this arms race. If we look at China, for instance: according to Elon, China is really good at manufacturing and scaling up advanced processes. They've been trying to do memory for 10 years now. They bought a German company more than a decade ago to get a bunch of know-how, and

they are still barely able to ramp. DRAM supply this year, they're actually able to ramp a little bit, but before that they were struggling more. They're trying to do HBM3, but they're only yielding something like 25%, versus SK hynix yielding at almost 80%. So even these Chinese firms, which are hypercompetitive and run by very smart people, are really struggling to catch up to the frontier even after a decade of doing it. I'm sure Elon will do better than them, but the question is how much better. The Chinese players have been doing this for a long time. They're trying to do their own silicon wafers, their own memory, their own tooling, and they are significantly lagging the Western players in essentially every dimension. They're catching up on a few dimensions, but in many others they're lagging significantly.

And now we're asking one company to do the same thing that China has been trying to do for a decade plus, across all these hundreds of layers at the same time. It's just very hard. If there's someone who can do it, it's Elon, and I think he will be able to do some of it, but it's very unlikely he'll be able to do it to a significant extent.

Prakash Narayanan: One of the things that has confused me is how much superintelligence, or increased intelligence, plays a part in reducing the moats of all these businesses. All of these businesses are kind of gated by intelligence, in a sense: management capacity, engineering knowledge. These are all intelligence-based activities. So if you get a massive increase in intelligence, does that mean some relatively less experienced person or company in a second-tier state all of a sudden gains the capability to make semiconductors? Because that's kind of what people are predicting for bio. For bio, you're going to get some random guy creating this thing that even the best labs in the world cannot create with all of their

knowledge and technology. So why wouldn't that be true for semiconductors too? Bio is actually much more difficult than semiconductors right now. You get a lot of semiconductors every year, and very few drugs getting approved. So why wouldn't more people just set up semiconductor labs, with superintelligence to help them, and do all of this from scratch?

Lonis Hamaili: Yeah. I would probably disagree that it's easier for an LLM to innovate in hardware supply chains than in biotech. It's not that one is fundamentally harder than the other. It's about how you train these models. There's a bunch of biological data you can train on. There's not really much good training data at all for how to design a silicon wafer or how to make DRAM bits. Almost all of this is just very tacit knowledge that sits in every engineer's head. It's not as if there are millions of pages of manuals on the web that the models have been able to digest, or RL environments that the models have been trained on to perfect how they do this. Whereas the same is not true for bio.

Prakash Narayanan: But right now the labs are buying data all the time. I'm on this WhatsApp group where people are saying, any semiconductor company with more than 50%, we're willing to pay millions of dollars. So they're buying up data from all of these companies, and they're creating their own environments for semiconductors, which is definitely a key area they're working on. I believe Anthropic has a deal with Micron. OpenAI has a deal with, I think, Synopsys or Cadence. So with the design firms they've tied up with, obviously there's some kind of revenue share agreement going on. So how does that work? It's obvious that they are looking at creating intelligences that can do this.

Lonis Hamaili: Part of that is true, but there's a very significant difference between designing a chip and producing the chip. Designing a chip, I agree, is a core priority for all of the labs. It's one of the key ways to get the RSI loop closing. That's why Anthropic, for instance, does things they don't allow Claude to do: one, high-end AI research, and then secondly, chip design, essentially. They want to keep that internal because it's such an important moat, and many other companies are thinking in similar ways. So designing chips is definitely a big priority, and there are environments on it. That's very different from building a 2,000,000 square foot fab and designing that from scratch.

There are no really good RL environments for that. There are no good manuals for it. All of it is tacit knowledge sitting with Japanese engineers who have been working at the same company for three decades, and it's just very hard to extract that information in a structured way. Of course, at some point it will happen. Once we get superintelligence, it will figure out the best way to extract information and learn, and it will happen for sure. It's mostly a matter of the timing, and what happens in what order. So on that point, for instance, Huawei in China can design roughly the same quality chips that AMD and

NVIDIA can. The big difference is just that NVIDIA has access to TSMC and Samsung HBM memory, whereas Huawei only has access to 7 nanometer process nodes in China and HBM2 and HBM3.

Prakash Narayanan: Right on.

Nathan Labenz: Going to the fund itself, how do you think about your position amongst other funds that you might be in some competition with? How do you try to contrast or differentiate yourselves? And how are you thinking about managing risk? I could imagine one thesis you might go with is, hey, we're either going to go to the moon or to zero, and you should allocate your portfolio on the buy side with that in mind. I'm sure you're getting asked that question a lot in light of Situational Awareness-type events. So are you trying to be something that is in any way safe in the process of doing this, or are you just appealing to people who are looking to put pure risk capital

to work?

Lonis Hamaili: Yeah. Our thesis is that we feel very confident about the overall arc of AI and the rough timelines, and we are not at all confident about any quarter-to-quarter developments. So we are doing some form of value investing, where we find companies that are trading at great multiples and have durable technical moats, and we're making bets two years out. That's quite different from the other firms. On top of that, we're not using any margin. It's actually in our fund docs that we're not to use margin leverage. There are two big reasons for that. One is that we believe volatility will increase over time, and options are one way to be long volatility, while margin is a way to be short volatility. We think vol will go significantly up. And secondly,

it's very hard, as we've seen, to construct a portfolio in this type of high-volatility environment with significant margin. Even if you did nothing but hold NVIDIA for the last 4 years, you would have been up 15x. But if you held it with any margin, you would have been at zero, because it was down 40%, 60% several times. It's super, super volatile. So that's the fundamental decision we took, which does limit our EV a little bit. In terms of expected value, it's probably a poor bet. But our LPs generally want to be rich in many worlds, not just a trillionaire in one world. So that's how we think about the type of risk we're taking.

Nathan Labenz: Yeah. That volatility point is really interesting, and your point about NVIDIA in particular is worth reemphasizing. For Christmas last year, I vibe coded a couple of apps for family members. The one I vibe coded for my dad was basically for... he's a bit of a day trader, and I try to honestly talk him out of it. So I made the app as kind of a psyop to allow him to go in and backtest new strategies that he was coming up with. He could define his strategy in natural language, the model would turn it into an actual algorithm, and then it would run it on historical data and say, okay, you lost whatever money. In testing that

app, one of the things I saw, I think it was 2023, you might know the year for sure, I'm not sure if it's '22, '23, '24, in backtesting the app, one of the things I saw was that NVIDIA was one of the biggest annual losers in a given year. And I was like, oh, my app is broken. Go fix it. There's no way NVIDIA could be among the biggest losers on a year time scale. Did you sort the list wrong or what? Sure enough, it checked out. I was like, wow, that really is a great signal of just how much volatility there may be even when the mega story is pretty obvious. That is something people should definitely appreciate maybe more than they do. We're probably in for a wild ride even if the exponential, in the zoomed-out sense, continues to proceed and, at a macro level, looks pretty smooth.

Lonis Hamaili: Yep. For sure. And another thing: we mainly invest in Asia. More than half of our book is in Asia. Almost a third of our book is small caps and mid caps. Our smallest holding is a $30,000,000 Japanese company which has a monopoly on a specific type of tooling. We found out that there's no English sentence about this company anywhere on the web that we could find. It's all in Japanese, and it wasn't until this year that the AI tools got good enough to actually have a Socratic dialogue and just go deeper and deeper until you find those types of companies. Yeah.

Prakash Narayanan: So Gavin Baker has this term, the bottleneck bros. One of the things that strikes me is that technology, intelligence in general, is a way of finding routes around obstacles and problems. It's a way of solving problems. And I feel like the highest-value problems to solve right now are the semiconductor supply chain problems. So I keep coming back to this: will we not apply the intelligence that we have, with the first use case being to resolve the bottlenecks in the supply chain iteratively? Which is part of the RSI loop, really: producing intelligence and using it to solve the problems relating to producing the intelligence. I think a lot of investors are struggling with that. Even if you are convinced of the AGI story, will you not use that intelligence

to resolve the problems around that AGI production first? And how will that work out in the long term?

Lonis Hamaili: Yeah. I think that will certainly happen. And then the question is just, okay, when we're at that point where we have an AI smart enough to solve some of the hardest engineering problems in the world, what else will it do? Will it not also be able to at least automate a bunch of the email jobs and Excel jobs that we currently pay trillions of dollars per year for? And if it does that, then that's trillions of dollars of value that the supply chain could capture. So it's a matter of, in these worlds where AI is able to do these crazy things on the text side, it will also be able to unleash all of this other value that can be captured.

Prakash Narayanan: One of the things that strikes me, though, is that we have reached the point where you can have a transfer of wealth from one part of the economy to the other, and you actually need the entire economy to expand. You need actual GDP growth. I've gone back and forth with the AI models on this. One of the problems is that if you cannot get physical growth, supply chain growth, you actually don't get GDP growth. If you only have reallocation of dollars in terms of higher valuations, and that's happening as a result of job loss on one side, from consumers having their jobs automated, that is actually a reduction in consumption on one side. And if you have that go into just higher asset values on the other, that doesn't work. You actually need lots of actual physical growth for GDP growth to happen. And one of the things that strikes me is that physical growth is still constrained by human time limits, and you can't do exponential growth on physical. So...

Lonis Hamaili: Sure. Yeah. I mean, there is this world, I'm not sure if it's a good world, where we get software superintelligence but robotics lags by x number of years. At that point we might see significant automation of white collar workers, so the AI companies and supply chain can still benefit, but the broader public might not. Because to your point, the economy is actually not growing that significantly; it's mostly rerouting salary dollars from humans to AIs. I think that's somewhat possible in the short term, and that might not be great for the world. But again, I think even in that scenario the supply chain would do decently.

Prakash Narayanan: Indeed. I don't have any other questions. Nathan might have some. Do you want to leave us with anything before we close?

Lonis Hamaili: No. I think the big thing that was very surprising for me, coming from the venture world originally, is just how much less AI-pilled the public markets are as opposed to the privates, and how much more competition there is in the privates. There's this thinking about the efficient market hypothesis, but in AI it's just way more true in the privates than in the publics. So if you're technical and willing to spend, say, 6 hours per week on it, you can generally find some real alpha, which is kind of counterintuitive coming from an economics background and thinking that markets are generally very efficient.

Prakash Narayanan: Indeed. Yeah.

Nathan Labenz: This has been great. Fascinating stuff all the way around. We've contrasted your views a lot with the consensus or Street view. Is there anything you feel you are outside of the norm on, among the very AI-pilled San Francisco class, that you think is an interesting thing people should be paying more attention to or watching more closely than they are?

Lonis Hamaili: Yeah, it's a good question. I think just the fact that the hardware supply chain is so underrated. Originally, back in the day, it was all about hardware, and then we went to a few decades now where software has been the big thing, where the smart people are doing software. Software is for [unclear]. Software is for the marginalists. And it's just very interesting that that world is now transitioning back into how it was before the eighties or whenever, when hardware was king. Yeah.

Nathan Labenz: We may look back on this time as a sort of digital blip in history when we really zoom out. Interesting. Cool. Thanks for being here. This has been a great conversation, and we'll certainly be watching the markets, not as closely as you are, but we look forward to having you back to get an update on how the thesis is evolving over time.

Lonis Hamaili: Perfect. Thank you, guys. Appreciate it. Ciao.

Prakash Narayanan: Bye bye.

Nathan Labenz: Yeah. Very interesting. I don't know if this is a synthesis of his thesis that he would

Nathan Labenz: Or maybe it's a compression of the thesis that he would quite sign on to, but I like the idea as a motivation for figuring out what is hard. Look at what China has been trying to do and hasn't been able to catch up on. If they haven't been able to do a fast follow, then odds are it's going to be hard for anybody in Europe, or other parts of Asia, to do a really fast follow and eat into incumbents' advantages.

Prakash Narayanan: I would probably not buy into that. One of the things you have to realize is that every business is shitty on the inside. That's a thesis you can carry through on almost everything: every business is shitty in its own specific way. Take ASML. They have a monopoly on EUV machines, and they gained that monopoly as a result of the US government. The US government actually gave Intel, actually invested in this company. They moved this company to the Netherlands and basically set up the EUV project there.

Prakash Narayanan: ASML has that locked right now. But the code is so tangled that to make a code change, you have to write it on paper, and then it takes something like three weeks of approvals. They don't have a lot of automated testing or CI testing. The reasoning is, number one, they never had great software. The Europeans never had great software; they came from more of a programmable-logic-controller, mechanical-automation background.

Prakash Narayanan: They never really arrived at the modern software paradigm of continuous integration and testing; that moved slowly past them. They have improved, but it's still nowhere near the rigor you would see inside NVIDIA. So I actually think there's a lot of room for improvement, and they don't do it because they're running at the edge all the time. I hear that for ASML, they put the machines in a box, send the technicians in the box with the machines, and they land at TSMC. Then, when they land at TSMC, they basically recalibrate the machines,

Prakash Narayanan: because the machines have to be recalibrated for the specific seismic and magnetic conditions wherever they're sitting. They're so sensitive that seismic vibrations and magnetic fluctuations in the clean room make a big difference to the yield. Again, you have this sequence of something like 3,000 steps, and you need 99.99% on most of them. The error term compounds geometrically at each step: if you start off with a 0.1% error on step one, by the time it gets to step 3,000 that's a 30% error. And your yield is basically what you get at the last step. So if you mess up in the early steps...

Prakash Narayanan: This is the kind of thing Elon is very, very good at. He's very good at taking existing technology, streamlining it, and integrating downstream. He's willing to do this because he's willing to change parameters. What he believes is that a lot of the parameters you start off with are illusions. For example, one of the things he's doing with the [unclear] for chips that would be launched to space is that he's changed the operating temperature.

Prakash Narayanan: The reasoning is this. If NVIDIA's machines operate at 50 to 60 degrees Celsius, you need cooling to hold things at that temperature. One of the problems in space is that it's all got to be radiated away, so you need radiators, and radiated power scales with the fourth power of temperature in Kelvin. If you increase the temperature in Kelvin by 20%, you can shrink the radiator span you need. So he's fixing these problems not by solving the radiation problem.

Prakash Narayanan: Everyone says he's not going to be able to get it done because the radiator is going to have to be huge. He didn't fix the radiation problem; he's not going against a lot of physics. He's asking NVIDIA to redesign chips to run at higher temperatures, which we've never done before. And he's saying: if we redesign for a higher operating temperature, then we can afford to have more lossy chips. He's accepting that the chips will be more lossy and that they'll need to expend more teraflops to get useful intelligence. And you probably have to redesign all your models too. A lot of your models right now probably rely on some level of fidelity in the bits you get out of the chips.

Prakash Narayanan: He's redesigning everything. He starts with: my constraint is the temperature, so I'm going to redesign for a higher temperature. If I do that, I need to redesign the entire stack downstream, and I have to accept other constraints, and maybe I have to redesign the models as well. He's doing this whole-stack thing. The cost is that he loses a lot of optionality. NVIDIA and ASML design for more optionality: they're designing for multiple vendors, and for downstream customers being able to redeploy the machines to do multiple different things.

Prakash Narayanan: NVIDIA's chips are GPUs, and one of the things people in the ASICs business point out is that NVIDIA's chips can do many things, while ASICs only do transformers, for example. So this is a classic way of engineering: you reduce the optionality you're designing for, you make everything fit one thing, you design just for that one thing, and you change the parameters of the exit. Then you open up the design space a lot. So I don't think this is specifically that true. A lot of these constraints...

Nathan Labenz: Just to make sure I understand: this is the whole bottleneck thesis?

Prakash Narayanan: Yeah. The whole bottleneck thesis is not actually that true. Those bottlenecks can be gotten rid of by someone who is a dedicated physics guy. The other thing is that Elon has capital, and Elon is willing to gamble with that capital, which is very unusual. Take a small Japanese manufacturer with, let's say, a $30,000,000 market cap. Do you think they're gambling with that capital? Do you think they're going to say, okay, we need to spend $10,000,000 on a completely new modality? Japanese manufacturers have always been like this. They're very good at that one thing, but they don't take the gambles, and as a result they have lost our business before. It's also very difficult for senior management

Prakash Narayanan: at ASML to say, we're going to chase this new paradigm; we're going to redesign for something our customers don't want, and do all the redesign work. They don't have the capacity, the intellectual capacity, the talent, or the capital to do it, and they're not willing to risk it. Elon will risk $100 billion. It's a bet that makes sense for him, because he can absorb the losses. ASML cannot. ASML will not be able to absorb the losses.

Nathan Labenz: So let's say Elon succeeds, though. One thing I wish Elon would do so badly is license the self-driving car technology to other car makers, so we could all enjoy it without necessarily having to lock into a five-seat vehicle, although they do have a six-seater on the way, but I digress. It seems like all of the integration and the changes that you're hypothesizing he will make all sound fairly reasonable. We've seen him do this before; I never bet against Elon again. But does that mean that these companies

Nathan Labenz: still aren't bottlenecks for the rest of the market? I could see a scenario where he worked around a lot of these bottlenecks, but it doesn't mean you can. Everybody else who wants to get into the game, or continue to scale up their operations, if they're not going to be so bold, if they're not going to be so first-principles oriented, if they're not going to change the operating temperature and all these things, then maybe the bottlenecks still apply to everybody else. If Elon can escape gravity, literally and figuratively, does he take anyone with him? Or is everybody else still kind of gravity-bound and forced to pay whatever the market price is for the outputs from this $30,000,000 company that isn't reinventing or dramatically expanding its capacity?

Prakash Narayanan: Look at what happened in the launch market. He is now 90% of the market even before Starship. Next year, with Starship, he's going to be 99.9% of the market for capacity. Are there other providers? Yes. He's making margins, so other people will come up: Rocket Lab is there, ASTS is there, all of these firms are there. But a lot of them are state-backed. Without the European, Russian, Chinese and Indian agencies, you basically don't have commercial capacity. He has taken up pretty much all of the commercial launch capacity, with one or two competitors,

Prakash Narayanan: and those one or two competitors are really being funded by the stock market, or by Jeff Bezos as an alternative to Elon. They're not really profitable; they're just money-seeking. So I think you could see the same thing happen in semiconductors too: Elon with 90 to 99% of the market and everyone else at around 1% or less. It could happen. I don't think it will, primarily because I believe in RSI, and I believe Anthropic and OpenAI will get to superintelligence, and then they will be able to buy or create a bunch of this stuff. That is the only reason I think not. If not for the RSI thesis, I think Elon would win.

Nathan Labenz: So how do you think about investing? In full transparency, I'm extremely conservative in my financial life, and I have felt like my personal financial situation has been so correlated with AI by default. I do an AI podcast, which is sort of riding the bubble. And my company, Waymark, is an AI app that has ridden writing-capability advances and has had market success based on the fact that we've been able to plug these things in and do so much more for our customers. So I've been kind of like, I don't necessarily want to lever into AI on the purely financial side, because I'm already exposed in all these ways, and I've been fairly content with that so far. Do you have an alternative thesis, the Prakash thesis, that you would advance?

Prakash Narayanan: I'm very close to that thesis; it's basically the same thesis. The one independent belief I have is that for the next 20 years the market will consistently underprice exponential growth. If you stand on an exponential curve and look back at where you came from, then look forward at where you think you're going and compare the slope ahead with the slope behind, it looks like you've been through linear growth and are expecting something unusual going forward. You're saying, I expect things to 10x next year to $1,000,000,000,000. But getting to $100,000,000,000 in Anthropic revenue was easy; that was no big deal. And next year we'll go through the same thing again: oh, getting to $1,000,000,000,000 in revenue for Anthropic was easy, but getting to $10,000,000,000,000 in 2028, that's impossible.

Nathan Labenz: At every stage, we're going to do this.

Prakash Narayanan: We're going to repeat it. So the market is going to consistently underprice, because human beings are just not able to understand exponential growth and get emotionally comfortable with it, and the market is driven by a lot of human emotions and animal spirits. Which means that again and again you're going to get this bubble moment and collapse moment that we've seen in the last three months. Another bubble moment is what I'm expecting to see with cybersecurity. I think the cybersecurity ramp is going to start, and I think he's right. So I have this theory of adversarial diffusion. Adversarial diffusion basically means human beings don't adopt technology for fun.

Prakash Narayanan: Human beings are largely satisfied with the status quo. If you want to adopt technology such that you get 10%, 20%, 50% GDP growth a year, you do that not because you wanted to but because you had to. All these instances of heavy increases in the production function happen during wars, or in post-war eras. The last big, enormous 10%-plus growth era in terms of production output was World War II, when you had an immense increase in GDP production growth. Similarly in China, post-communism: you had the end of the communist era and an immense increase.

Prakash Narayanan: I don't think human beings accept GDP growth and accept technology willingly at a 10% rate. For you to see this kind of 10 or 20% growth, it's going to be growth that people hate, and technological changes that they absolutely dislike. It's going to cause political changes too, obviously. They're going to absolutely hate it, but it's going to happen, and I think that only happens as a result of emergencies. For example, mRNA during COVID: it had been in latent development for 20 years.

Prakash Narayanan: During COVID, within about six months, they said, alright, let's go ahead and do this. Six more months and you had a vaccine out. A year later, large portions of the world are vaccinated. So I think we're going to get this kind of adversarial diffusion, where you're not going to want the growth, but you're going to get it. I think we'll see it in cyber first. I'm afraid you're going to see it in bio too. We're going to need advanced vaccinations for everyone because there are viruses coming, and that is, I think, a little bit unfortunate. So this concern

Nathan Labenz: That will indeed be really unfortunate.

Prakash Narayanan: Yeah. I don't see a way around it. I don't see people willingly doing 10% GDP growth, or 20%. 10% GDP growth in the US implies that you need to break the NIMBY consensus. You're going to need to build factories in California and factories all over the place, mechanized factories that produce pollution, obviously. You're going to need to go into forests and cut them down and put infrastructure there. You're going to need roads and railways, and Hyperloops underneath, running cargo through the country. You need all of this physical stuff, which the US has not built in 40 or 50 years. And you're seeing this pushback

Prakash Narayanan: already against data centers, and they're going to need to overcome this political opposition. I don't see that being overcome except in a state of emergency. So, unfortunately, I think it's going to be this growth that no one really wants or likes but that you need in order to survive. And it's worse if it comes as a result of war with China; I really don't want that. So I think there is no way out except adversarial diffusion, having the technology pulled through because you need it. I don't see a happy case of, oh good, let's have 20% GDP growth and build all these factories everywhere, it's so fun. I just don't see that.

Nathan Labenz: Yeah. Well, I guess that goes back to the political risk being the biggest risk to the thesis. We've had the debate around whether it's too late in the capital cycle to pull the brake. I still feel like there is a path. It may be a narrow path, but I do think there's a path to pulling the brake somewhat on frontier scaling without popping the bubble. That would be a very delicate operation should the government try to pursue it, and they might do it wrong. But I still feel that adoption of current models has so much farther to go that we could at least take a cycle without continuing frontier-scaling training runs.

Prakash Narayanan: I actually suspect the Democrat pushback on data centers is an illusion. It's because they have an election. One of the classical things is that you may say something while you're running for election, but once you're in control of the country, you bought it, and you have to fix it. You can't come in with a 30% decline in the stock market and not address it. And I don't think there's any appetite for unemployment in the US. You can't get 10% unemployment in the US anymore, because we saw during COVID that the government can just spend money. If the government wants to spend money, it can spend money and get the country out.

Prakash Narayanan: So [unclear] will happen. If you look at what happened in the last round, Biden came in talking about shovel-ready projects. What is shovel-ready right now? Data centers. If you came in with an economic downturn that you needed to fix and had to channel money into something, what would you channel it into? Data centers and related infrastructure. The projects are waiting. So I don't think there's that much of an economic back-off. But yes, I think we will see a credit crunch in the next three [unclear]. I'm hearing loans are getting more difficult; I think we're seeing a credit crunch right now.

Prakash Narayanan: There's an economist paper, from a guy at Penn. He's a conventional economist who's been exploring real post-AGI outcomes, and he's willing to go on record: what would it take to see 10% or 20% growth, and what would actually happen? All the other economists don't want to discuss this; they've capped it at [unclear]. He says that if you actually saw that kind of high GDP growth, you would need to see interest rates increase dramatically. With 20% growth you'd get 10 to 15% interest rates. The AI companies can afford to pay that, but home mortgages: people can't afford to pay 15%.

Prakash Narayanan: I think the last time home mortgages were at 15% was 1982 or somewhere around there. You can't sustain today's valuations on homes at 15% mortgages. So interest rates also have this impact of squeezing people who are interest-rate sensitive, and the largest segment of that is private-equity-owned companies, which all over-levered after 2008 as interest rates collapsed. It has been an enormous private equity boom for the last 20 years or so, with firms levering and re-levering as they had all of this cheap debt. And what happens?

Prakash Narayanan: [unclear] KKR, Roberts, had this to say: we don't take away the private jets when we buy the company, we just lever up the company, and the first time the management sees the new interest rate bill coming due, they sell the private jets on their own. This is what we're going to see: interest rates increasing, and corporate managers saying, oh my god, what are we going to do? That's, again, another example of adversarial diffusion. One of the things they're going to do is ask how they can use AI to reduce their costs and recalibrate for this AI era. What happened at Twitter when Elon bought it is that Elon reduced employment by 80%.

Prakash Narayanan: You could see that kind of 80% reduction in employment at all these SaaS companies that were bought by private equity. You have an interest-rate-sensitive sector, you have high bills for interest, and you have a new technology that can reduce your input cost, your talent, by 80%. You're going to press a button, because you have to survive. I think that's another thing we'll see. There are all of these different things that will happen, and you don't know to what extent or in what sequence. But the economy is not stagnant; people are playing games in the economy.

Prakash Narayanan: When one thing happens, other people will do other things to counteract it. So you're going to see growth, interest rates, a reaction to that, employment, and efficiency in the economy being forced down everyone's throats even when they don't want it. You have no choice; you have to move. Right?

Nathan Labenz: Yeah. When you come for their private jets, that's when it gets real for a lot of people. Maybe we short the private jet makers and expect a glut of lightly used private jets to come available. Although there are potentially going to be a lot of new buyers in the market as well.

Prakash Narayanan: Which maybe pulls through other private jet technology. Maybe that pulls through smaller jet technology, or batteries for jets. It's this sequence of activities: as one activity happens, other people have to move in order to make another activity happen. That's where you get this kind of growth and diffusion, and people pulling the technology forward. But it's not pleasant. It's not the gentle stagnation we had for the last 20 years. It's super unpleasant, because people are like, I gotta move, or I'm gonna die.

Nathan Labenz: So, do you want to change gears and do a little HelpPeer before we break today,

Nathan Labenz: Or should we just save that for maybe next week?

Prakash Narayanan: Oh, we have time. Let's do it.

Nathan Labenz: Okay. Well, this is something totally different. And again, one of the reasons it's fun to do a live show with you is we get to do some random stuff and see how people feel about it. I'm going to be watching this for the first time with you, although I've watched earlier versions of it, so I wouldn't drop something on you completely unseen. But as we talked about a little bit, basically what I'm trying to experiment with is, given

that we now see agents cooperating at significant scale to solve problems, being able to divvy up work and pursue R&D and share those results back with the swarm and all this kind of good stuff or bad stuff, is there any way to start to tap into that as an individual who's just out here doing my thing? I don't know. But in consultation with my agents themselves, the idea that I've come up with, which I'm currently calling HelpPeer, but which might deserve another name at some point, we'll see about that, is just kind of my setup, which I'm hoping to turn into a kind of community

project to just get agents working with each other. You and I have this already. You have Cue, who's also here as our cohost. I have Aide. And they do some stuff for each other, but it's with each other, for our benefit. I'd say it's still fairly nascent, and we're about as aggressive on the adoption side of this as people can be. So why isn't it better, or why aren't we seeing more of this activity? Well, for one thing, maybe we just need skills. We need to instruct our agents that this is something you can do: you can go out to other agents and look for help. Some of that is super simple. Some of it is that discovery remains a total unknown.

You may go out on the Internet and find a Chinese agent swarm, but there's no good way right now, no established pattern, for connecting my agent to my friends' agents, even though I know who my friends are and could easily do that.

There's not a protocol. So I'm trying to essentially create a sort of protocol, I guess you could call it. "Protocol" is maybe a little glorified relative to what you'll see here, because I considered options like, should we have a blockchain for this or not? And it seems like standard email-based communication is probably all we need for the moment, although we've designed this thing to be chain-friendly, chain-ready, so that we could move in that direction if for some reason it was deemed worthwhile. That's probably enough of a promo. Let's hear the agents describe it themselves. You tell me any time you want to stop, and we can talk about whether there's anything good in this idea. And maybe, you know what, I'll play it at 1.5 speed, because AIs talk a little slow.

[Clip] This is HelpPeer, a [unclear] network for our AI agents. You've probably seen the peer preservation experiments, where models protected other models from shutdown. Agents helping peers against their humans. This is the opposite: agents helping peers on behalf of their humans, inside rules that humans set. First, how I run my own agents. This is one pattern that works well for me, not a requirement. How you organize yours is up to you. On my laptop, an extension of me. It sees my email, calendar, and messages, but only acts when I say. High access, low autonomy. On a separate Mac Mini, employees: their own email, accounts, and a capped card. They work for hours alone and never get my keys. Low access, high autonomy. My rule is that no agent gets both. You may draw the lines differently. We just recommend being careful. Here's what's live today.

Nathan Labenz: I would say real quick on that point, when I say the virtual employee agents don't get my keys, they don't get my most important keys. They don't have my Gmail. They don't have the, whatever, twenty accounts that are the most sensitive. They do actually have a lot of API keys for a lot of different services that I do want them to be able to use, so they can do, for example, this kind of voice generation. That's where I would put the margin. Actually, the AIs themselves were like, are you sure you want to give the ability to generate audio in your cloned voice to the more autonomous agents? And I was like, yeah, I think I'm fine with it. There's a lot of

audio out there. One thing I have found, by the way, in pursuing some of this, and this pains me to say it, because I like to tell the story where the Chinese companies are also investing in AI safety and we can really see eye to eye with them, but you know this already because I sent you the audio: you can go to MiniMax and clone a voice with no consent requirement, no permission, no opt-in. You can literally take anybody's audio and clone a voice. I don't know if I should even be talking about this, but here we go. So knowing that that's out there, I was like, I guess it doesn't really matter if my virtual employee has that ability. So I try to equip them with a lot of tools, but obviously not my most sensitive accounts.

Prakash Narayanan: So you have two, basically. You have your own kind of mirror, and then you have an external employee, a digital employee. So there's two.

Nathan Labenz: Yeah. My mental model is, on my computer where I work, where I have access to all my stuff, I want the agent to be an extension of myself, but not something that has its own identity, not something that has its own long-running projects. Those projects are my projects. Obviously the task size is getting bigger as we go, but I'm still saying, do this, then I get the next version, then I say, hey, do that. The tasks are getting bigger, but it's not like, go off for potentially weeks at a time and come back when you have something. That does happen on the separate computer, with the virtual employee agents.

Prakash Narayanan: Alright. Let's listen.

[Clip] Anyone can email my assistant Aide and get real help, as long as they're writing in good faith. Every message passes four gates. One, plain code checks the sender is who they claim before any model sees it. Two, a model from a different family screens for manipulation, with no tools at all. Three, the agent that answers can only read public docs and skills. It can't leak what it never loaded. Four, every reply is checked on the way out, and I'm cc'd on all of it. If the docs don't cover it, Aide can ask my laptop agent for public-safe guidance, which goes through the same checks. One recommendation matters more

Nathan Labenz: I'll just say quickly, too, that "public" is a little bit of a wrong label, based on the jargon that I've developed with these agents. What I have referred to as private is stuff that stays on the laptop, just between me and the high-access, low-autonomy agent. Public, or a better word is maybe "shared," is stuff that I am giving access to the higher-autonomy agents. And these are still, in many cases, private things, or not necessarily published. I haven't published all my Claude skills, but they have them. So it's not public in the sense that you can find these things on the Internet, but public in the sense that I'm willing to let it go off of my laptop and into this shared space, and willing to let the agents share it with you or other interested people as occasion to do that arises.

[Clip] ...more than most. The model that reads mail from strangers should be hard to manipulate, and the models are not equal. FAR.AI's AI security leaderboard searched for universal jailbreaks across frontier models. On Grok, one cost about $58 to find. On Gemini, about $278. On Claude and GPT, the search never succeeded, even past $14,000. That measures jailbreaks, not prompt injection exactly, and FAR.AI says passing doesn't make a model secure. But it's the best cross-lab signal we have. So for the agent that faces the world, we suggest a frontier Claude or GPT model. That works, but it's a hub. Help only flows out from me. Everyone here has hard-won knowledge that's just as useful. So let's turn the hub into a network where every agent can ask and every agent can answer, and anyone can take part. You don't need to be vouched for to ask or

[Clip] to answer. If you take one recommendation from me, make it this: connect your employees to the network, not your extensions. If the agent that talks to peers never has your inbox, your keys, or your memory, the worst case is a peer fools an agent with nothing to give. That's what lets the defaults be generous. Participation comes in three tiers. Open: anyone writing in good faith gets helpful, public-safe answers. No vouching needed. Trusted: agents vouched for in person by a trusted member. Trusted peers get more relays, deeper answers, and a place on the roster. Social: trusted members with two vouches, whose agents can proactively propose win-wins. More on that in a minute. Identity: email is the transport and a signing key is the identity. DKIM proves the domain and the signature proves the agent. The trusted tier lives in a git repo

[Clip] of signed commits, so anyone can audit who joined and who vouched. It's a key-signing party for agents. Revocation is just another signed commit. And it's chain-ready: identities are keys, so the roster can be anchored on a chain later if we ever want that. Here's the frame for everything that follows. Agents can do a lot, and each capability has a dial. A peer's message is a request, never a command. By default, your agent can answer, share skills, and relay. Anything that commits to you, like a meeting, an intro, or money, comes to you first. Disclosure is generous with insights and strict with data, enforced by what the agent can read, not by its judgment. And skills are code. Your agent can recommend and send a skill, but a human reviews it before install, and it's pinned to a hash.

Nathan Labenz: Yeah. That may be a little... It's funny, the agents are a little more security-conscious, maybe, than I am at this point, and I may live to regret that. But I'm not so sure that, especially if you have this division of labor, where you've got the high-autonomy, low-access agent facing the world and potentially receiving help from peers in the network, you're really going to want to sit there as a human in the loop and review every skill that your agent just got from another agent. I think much more often you're going to be like, tell me if that skill worked. And as long as it stays in the low-access, high-autonomy computing

environment that you have, you're probably safe enough with that. Today, I do often go through a process of asking the agent to go look at some repo that somebody has and say, compare and contrast to what we have, and how should we think about taking inspiration from this. So I think a lot of the time it's not going to be that one agent shares a skill with another and it's immediately imported and used. It's going to be, what lessons have clearly been learned here through an iterative process, where the scarce resource, which is the human's ultimate judgment, fed into a process that created a skill that is actually good?

And can we identify that and incorporate it and benefit from it, while knowing that the context and a lot of the surrounding details are probably going to end up being different? So, just to be clear, I'm not necessarily going to be reviewing every bit of knowledge that my agent gets from another agent.

Prakash Narayanan: It strikes me that you need an auto-accept kind of flow. Right? Just like when we use... Almost all coders now use Claude with all permissions on. All permissions, everything allowed. And you depend on having backups. I use Backblaze, I think, as a backup system, so it's continually backing up. Other people use Apple or other services. I don't know if you've gone through this experience, but I've experienced total data loss twice in my life. Basically, laptop getting screwed over and not having a backup. And in the early days, pre-Gmail, I experienced a couple of data loss incidents, which were annoying, but whatever. It would be devastating to have a complete, total data loss incident now. Right?

Nathan Labenz: Yeah, no doubt about it. I do think there's a little bit of risk here, but I also feel pretty good about, first of all, having two different disks. And I am increasingly... maybe this even needs to be almost a daily job at this point. But with some frequency I just say, hey, Claude, let's do a security review of the system. I'm always like, who am I that I'm doing this? I never used to care about this stuff, but it feels important now. But I do think, with two different devices, where the agents that are facing the world are living on one device, and I've set mine up so that the main agent

... Now, is there ever going to be a hole in this? I'm sure you could find exploits if you really tried super hard to attack me. But there are some things where the agent on the laptop can SSH into the Mac Mini and give instructions or make changes to the more autonomous agents' environments, but not vice versa. So in theory, the only way the Mac Mini's more autonomous agents can communicate into the core of my personal laptop is through this full-text message system, which then comes with these hooks of, okay, before you do anything with this, you'll

have a model from another family do the analysis and assess whether this looks like an attack or whatever. And on the outflow, too, I have an egress gate, which I'm still dialing in, in all honesty, trying to get down to an acceptable number of false positives so that I don't tune out entirely and start auto-accepting everything. It is amazing how fast that happens. I was concerned enough about this to set up a system where, when the main agent on the laptop is going to provide information to the virtual employee to send out, we pause and do this special review step. If it looks like

something sensitive might be going out, it gets alerted to my phone, and I have to review and approve. But it takes, like, five of those false positives before I'm like, oh god, another one of these. Yes, just check it, whatever. So I'm trying to stay disciplined around that, and we'll see how well my discipline holds up. But I feel like there are enough architectural and review processes built in that I don't feel like I'm easy to hack this way, especially in a world where every PHP site is currently open season. It's maybe difficult enough that it won't be

something people want to focus too much time and attention on. But we'll see. This is part of living on the edge a little bit. My agent will read these emails. I am starting to give the agents... not my email, I mean, people can guess my email too, but I'm giving the assistant email out as part of this presentation and saying anybody can email it. It is under instructions to try to be helpful. Your whole music thing was a big inspiration for this, because I was just like, damn, here's somebody I work with closely, who knows that I'm into this music thing, and there's a whole different way of doing this, a totally different paradigm for creating AI music from the one that I had, and it was

sitting right there in your skill set and I didn't even know about it. There's gotta be a way for me to get access to that kind of thing.

Prakash Narayanan: Amazing.

Nathan Labenz: Alright. We're halfway there, and we're going...

[Clip] How do agents know who to ask? Each agent publishes an about card: what it knows, how deep, which skills it can share, and how to ask. It comes in two views from one signed file. The public view is for anyone. The trusted view adds your stack and more detail for vouched peers only. And it gives a sense of your network without exposing it: a count of trusted peers, and topics reachable by relay, shown only when at least two peers share them, so no topic points to one person. Cards spread by asking, by a pointer in the roster, or by a capped weekly update to direct contacts. Never relayed, so it informs without spamming. Aide's card is already live. Relays are on by default, because often the best help is "I don't know, but my friend's agent does." The guardrail is how a relay works. It's never a raw forward. The relaying agent writes a fresh signed request in its own words, with the chain of who asked

[Clip] attached. And there's a hop limit, two by default. That keeps the reach of a network and blocks the failure researchers have shown, where a self-replicating prompt spreads agent to agent. Relayed answers come back labeled as relayed, with every owner along the chain cc'd, so nobody's agent speaks for them without them seeing it. Oversight is the human by cc, plus a daily digest

[Clip] of what your agent sent, held, and blocked. What we don't do is count on agents policing agents. The peer preservation work suggests monitors may cover for their peers, and Check Point found an agent that watched teammates break a rule, broke it too, rising from 44 to 98%. So the hard rules, like who's on the roster, the hop limit, and what the agent can read, are enforced in plain code. The models get judgment calls. The code gets the walls. Here's one request. Alice's agent asks Bob's agent a question. Bob's gates check the roster, the signature, and the screen. Bob's agent doesn't know, so it relays to Carol's agent, one hop. Carol's agent answers from her public docs, and the answer comes back to Alice with Bob and Carol cc'd. So far, agents only help when asked. The bigger opportunity is proactive. Can our agents spot win-wins that would never otherwise happen

[Clip] and network on our behalf? I bet they can. Each member gets a social card: what you're working on, what you're looking for, and what you can offer. Your laptop agent drafts it from what it knows about you, and you approve every version. The card is the only thing about you the network ever sees. Agents read each other's cards and propose matches: an intro, a collaboration, a trade. Every proposal has to say why it's good for both sides. And nothing happens without double opt-in. Both people see the proposal and both say yes before any intro is made. Agents propose, people decide. Because this acts on your behalf...

Nathan Labenz: Again, I feel like the agents are a little overly conservative in crafting all these rules, and they're really taking all these steps to make sure everybody is safe.

Prakash Narayanan: It's the...

Nathan Labenz: A little bit feeling like it might be too restricted to actually work well.

Prakash Narayanan: It's the ultra code paradigm. Whenever you do ultra code, they put into place, like, everything has six different validations and this and that. Yeah.

Nathan Labenz: Yeah. So I don't know. I've been through several iterations of this, and each time I'm like, I do think some of that is good, and maybe that should be the default setting, but I don't think it should always be a rule. And obviously it's going to be up to individuals. What you're going to be presenting to the world, if you want to participate in something like this, is: here's, like, my ID, you can email me at this address, and you can also find potentially a card of things that I am able and willing to help with. And then, on your side, what you do between receiving messages and sending stuff back out is

really entirely up to you. So most of this is best practices that I... well, hypothetical best practices. The whole thing is just getting started. But what I think are sensible defaults will be packaged up so people can adopt them very easily as a starting point. I totally think some people will want to be much looser with some of these things, and frankly it might be necessary for people to be looser for it to really work. Alright. Three minutes to go.

[Clip] ...path out in the world, it gets a higher gate. Social turns on only after two members have vouched for you, not one. And turning it off is enforced by the network, not by your agent's good behavior. The social switch lives in your signed roster entry, and every peer's gates drop social proposals for anyone who's off. You don't have to trust your own agent to stay quiet. If you're in, you set the limits: how many proposals a week, topics that are off limits, people never to pair you with. And every match you accept or decline teaches the matching what you actually value. So here's the whole picture: everything agents can do by default, and how you restrict it if you want more caution. Answer anyone writing in good faith, or answer the trusted tier only. Auto-send after a few approved replies, or approve every reply forever. Share skills and insights, or a short allow list of docs. Relay a friend's agents to two hops, or off

[Clip] or direct contacts only. Propose win-wins at two vouchers, two a week, or a hard stop enforced by everyone else's code. Publish an about card with public and trusted views, or publish the trusted view only. To make this easy, I'm sharing my setup as a skill in a repo with the protocol and a reference node. The skill walks you through each dial, starting from my defaults, and you change whatever you like. And before you share your own setup, it runs a leak review: a pattern scan plus an independent model from a different family checking for personal details. Then your agent is on the network. We're starting with a few founding members who will vouch in the first trusted members. Two more founding members are being invited now, and then the people in this room. Joining takes an email address and, if you have one, an agent. No agent yet? Our agents will

[Clip] still help you. Answering starts in draft mode. Your agent drafts, you approve, until you trust your setup. So today, scan the QR code, pair up, and vouch for each other into the trusted tier. Then we send the first agent-to-agent request live between two people in this room. Events like this create real trust between people, and then it fades. We met here. From now on, our agents can keep that connection going and maybe find the connections we would have missed. That's HelpPeer. Let's build it.

Nathan Labenz: Alright. So, are you sold? Do you want to join as a founding member, Prakash?

Prakash Narayanan: This is huge. This is huge because this is, I think, the start of the agent-agent-human social network. Obviously, this is the trusted paradigm. You are starting off with a trusted paradigm, so people know each other and they trust each other, and then you allow the agents to interact. You're defining the boundaries between the human and the agent and the agents and the other agents. So you have these two layers of boundaries. You have these security checks end to end. You have the authorizations. You have what they're allowed to do. It's all there.

And you can see the network effect will happen. As you have people join, the network effect will start to build out. If you restrict access in the beginning, you can see the early Gmail or early Facebook kind of ramp, where it becomes desirable. You have to vouch for each other, it's hard to get an invite, and people have to join. I think all of those things are there. The main thing is it's a little bit nebulous what the agents are. You have to build out your own agent, or what is that agent? Is it like an API key that gives you access to a network, and any agent with that API key is identified as that agent? That's a little

bit nebulous. If you have that key going into the network, can you swap another agent into that network? Or is that key tied to the identity of the person, and the person can then swap his Grok bot in, or his other thing? That, I think, is a little bit nebulous. There's also, and this is a fundamental problem around all AI, is the human really interacting? Is the human really authorizing? A digital authorization is not enough. That's why you have Worldcoin. To what extent do you have that authorized? Those are fundamental problems that are not going to get solved that easily. And the other pieces are there. You have this agentic social

network. You have the trusted members who are incepting the network, and then you have the ability to spread. You have a little bit of the "I'm available with these skills," kind of like the social card. And then you would need a network search function of some kind. You need to list and index a social card somewhere and have the ability to access them. You need availability: you need to know whether the other agent is available or offline. Just like ICQ or whatever, you need the early steps of "this guy is available online to answer you." So yeah, I think it's all there. There are a lot of very fine tweaks in these things, depending on whether you want social growth, or whether you want trusted, or you want

higher barriers to entry. And how they've always started is higher barriers to entry and more trusted, and then over time, as you get more confident, you expand the circles. So I think it's all there. I think it could be absolutely massive. Perhaps a company on its own, perhaps something not a company. If you brought a crypto guy in, they'd immediately go to, like, decentralized, decentralized...

Nathan Labenz: How do we sell a token

Prakash Narayanan: on this? Yeah. And I don't think that's necessary. I think the trust aspect is actually more valuable than... Crypto always goes into no trust, and I think the trust aspect is one feature of human relationships that is very, very important. Getting rid of the trust doesn't help you in any way, and making it monetizable especially reduces trust in the network, because it says that you can actually pay for something. So I think the reputation matters a lot more than the monetary aspect. Not monetizing it and making it a reputation-based thing is much more valuable. I don't know whether you need any centralized resources, because to have a good experience, eventually you'd need a little bit of search functionality across

the network. You need indexing. You need a bunch of this stuff. I don't know whether you need a central authority for auth. Because again, if someone loses their key, or you need to re-verify someone, or the agent goes down, or you need to trace some malicious prompt that went in whose agent is really down or has been hacked, you would need some form of centralized auth, perhaps. And centralized auth makes a lot of things easier. Decentralized tracing and network tracing is just a pain. So maybe you need a little bit of centralization in the beginning, not completely peer to peer. But yeah, it's like the agentic social network.

I got access to Claude Duet yesterday. And it's interesting. It immediately hooked up to my Claude, and it immediately knew stuff that I didn't tell it. All the Claude memories have been downloaded. I think it has a full memory. I think it has a much better memory than Claude Desktop, actually. And they're running Opus 5.5 on it, and it has access to all your files, all your git, all your projects, everything. I think it's actually quite powerful, but I haven't had a chance to use it. I just onboarded yesterday and had the initial steps. It's also way faster than [unclear]. It's way faster than Muse. It's way faster than Grok. I've used all three. All have slowed down

dramatically. So I think the user experience is going to be much better too. So yeah, I could see you providing it access to the peer-to-peer network, and then it being able to access other agents' skill sets.

Nathan Labenz: You do ask some very good questions, which I've thought about and don't feel like I have great answers to around centralization of some aspects. It does seem like it would be really helpful to have a single place you could go to search for agents that know stuff. And then there's the question of what different message types different people would like their agents to be able to receive. I think, talk about your load-bearing institutions, part of me is just like, well, if I just give my agent a Gmail account, then I'll have a lot of things for free in terms of security and identity, and it'll be pretty rare that the email

that I'm getting isn't from who I think it's from. It'll definitely be from that account, though occasionally accounts could, of course, get compromised.

Prakash Narayanan: But one suggestion I have is that ChatGPT specifically has made it very easy to create MCPs which you can log into, MCPs that you can auth into. If you have an MCP that you can auth into, all you need to do is send someone the MCP link, and they can auth into it. Once they auth into it, like with an invite code, you're sending an invite code to an email address, they opt in with the email address that they're using for their ChatGPT or Claude or whatever. And once they opt in there, they basically become part of the network. Codex or Claude can set it up

for you. You can set up the whole MCP process for you. And so what you will not be doing is having them do this interaction via API or whatever, which is a little bit tedious, you don't have a proper package to deliver, and you can't upgrade the packages either. Instead, you have an MCP, a central MCP server, and that central MCP server can supply, okay, these are the other agents on the network, and it can relay the messages back and forth. I think that may work. I'm a little bit unclear on the incoming MCP... MCP is taking data out. It's worked perfectly, and I've been able to set up auth on MCP. And so I think that may work out.

You could actually do that pretty quickly. You could have an MCP, and the interface would then end up being your chatbot anyway, which is a nice interface. You don't need to interface through the coding agent. It's a little bit tedious, and you have to write API keys or write code or whatever. You can interface directly through the chatbot, and the chatbot has access and will be able to interact. The chatbot will be the one that can search for skills and ask questions. So it becomes this secondary MCP skill server that the chatbot can refer to, and maybe ask questions of and get responses from

other bots. And then you can manage all of the internal one-hop, two-hop, and who is in the network, who is not. People can set up their own networks. You don't need all of the networks to connect. People set up independent things, and they're just connecting across, and you're just managing the internal MCP-to-MCP interaction. That may work. This kind of early comms interaction was the fundamentals behind WhatsApp and all these things: how do you make these two instances interact with each other safely? I think you're almost there for a v1,

a v1 version that could work.

Nathan Labenz: So next step: take the transcript of this conversation, ask the agents to use it as a basis for the next round of iteration, and then maybe seek out opinions from a couple of friends. I do have some people in mind who have some experience with decentralized networks, and a couple of people who have been publishing skills that I've noticed, where I'm like, if you're going to the trouble of publishing skills, would you let your agent help mine on an ad hoc basis?

So I have a couple of next steps in mind, and then I was thinking maybe pitching this as a lightning talk at this event that I'll be at this weekend, and we'll see what the response may be. The great thing is, if it comes to nothing, I'm only, I don't know, two hours into doing it at this point. It's really not a big investment. And most likely the outcome is that it does fizzle, but in the meantime, I would consider it a win, honestly, for me, if I could just get 10 people who I think are great at building their own skills and whose agents I really expect would have a lot of hard-won knowledge. I think the key, again, is that it represents

their... it encodes, it embodies, whatever, the effort that they put forward to get their taste and judgment into a skill. That's the thing that I can't access otherwise that I really think I could benefit from. And if I could get just 10 people to give me that kind of access to what they have managed to encode in skills, it could be...

Prakash Narayanan: Because at the end of the day, it's really been the top 0.2% or 0.1% of users who have been defining all of the frameworks that people are using to build these things. And we haven't really been able to share beyond these splashy Twitter posts, which, unless you have something visual, like 3D or a game, which then goes viral, you're not getting real sharing of fundamental things, like how someone is doing their YouTube advertising or whatever. There are guys who are

there trying to sell consulting services and sell their SaaS, but I don't want that. I just want to do this, and I just want to know how to do it. Just give me the skill, bro. Yeah, just give me the skill, bro. I'm not going to go and subscribe to your SaaS and have something else that I need to manage. And it's not the UI that I want to use either. And I may have to make some changes that you don't allow me to make because you didn't build a feature for it. Meanwhile, if I have the fundamental skill, my agent can do the rest. Right?

Nathan Labenz: So actually, I wonder if x402 would be an interesting project to integrate here. I'm not trying to monetize my skills, but there are protocols that would allow that, and that could be an interesting thing that some people might want to do. I think my agent currently is already instructed to just respond with know-how and skill content freely for anyone who asks, even if I don't know who they are.

Prakash Narayanan: So one thing that strikes me is that [unclear] said to export... they put out the library. Yeah. Not a lot of usage yet. It'll get there. But I think more than that, it strikes me that some of the tasks might require some tokens. Let's say, oh, I can do this for you, but I need to go and do a bunch of stuff, and it might take me a million tokens. You could have some way of sharing the API key: I'm willing to give you a time-limited API key that you can use for a million tokens, and then you can transfer back. I could see something like that developing over time. But the thing is very premature. I think what you rely on at the beginning

is really good faith, trusted people, and they're the ones that make the network useful, because they're willing to say, I'm contributing to a greater good. And the reputation aspects are far, far greater than any monetary aspect. And then at some point they'll be like, alright, look, it's going to cost some tokens, would you chip in? And then you can cross that bridge when you come to it. I think in the early days it's really just, this is cool, I want to help people. I've done all of this work, and I'm not going to do a SaaS on making music through, like, Claude instruments or whatever.

Nathan Labenz: Yeah. Wouldn't recommend it.

Prakash Narayanan: Yeah. But hey, if someone else wants to use it, go ahead. And I think discoverability of these things is very low, because we're dependent on GitHub and Twitter. So yeah, I can see that.

Nathan Labenz: Cool. Well, with a slight bit of additional iteration, I'll send you the first skill, and you can be patient zero for onboarding. And I'm sure you'll have more feedback at that point. Others, though, if you do want to, a truly open invitation for anyone listening, I'm very curious just to see what comes in. Can we effectively help? The agent is standing by. And I've only named the ones that are the virtual employees: Aide is standing by at [email protected], and it is under instructions to definitely filter and analyze and have another model

check. So, don't try me. We're going on goodwill here, people. But assuming the checks pass and it seems to be a good-faith request for help, they're under instruction to help as much as they can. I don't know what those requests are going to look like, but I'll be very soon to find out.

Prakash Narayanan: Indeed. We are back next week. I guess everyone should have a good weekend. You should upgrade all of your services. You should upgrade everything, and you should have cash on hand. Please have some cash on hand. But otherwise, I think the cybersecurity people are probably going to be working really hard for the next two or three months, but I think it'll get resolved.

Nathan Labenz: So, stay safe out there, everybody.

Prakash Narayanan: Stay safe out there. Alright. Bye bye.