The Largest Infrastructure Project in History — Evan Conrad (SF Compute) | Worldbuilders

The current scale of the GPU buildout is the largest infrastructure project in the history of the world, larger than the cost of wars. It is multiple times the Apollo project. We are the company that takes no bullshit. Our entire culture is we never had the opportunity to be fake. You will never see us say we're gonna democratize compute or powering the AI revolution or something. It's like, no, we sell GPUs.

Right.

We rent them to you. The reason why you use us, price, reliability. If there is a bubble, here is where the bubble is. In order to afford a large GPU contract, you go out and you raise the big round. So the bubble isn't in the GPU cloud? I don't think so. Instead, I think the bubble is in whoever owns the equity of those companies that just raised a whole big round. Because if they don't succeed, they can't get that capital back. And there's also nothing else you can do with that.

I'm Sumeet Singh, and I'm guest hosting a new series for the Village Global Podcast called Worldbuilders. I run Worldbuild, a thesis-driven investment firm that backs creative technologists. We spend our time exploring rabbit holes because some of these rabbit holes are actually tunnels that open into the future. In this show, I'll be taking you down those tunnels. We'll have conversations with the people shaping what comes next. break down the frameworks for where value accrues when the rules are changing, and figure out what the builders actually see that the rest of us don't. So come take the leap with Worldbuilders. My guest today is Evan Conrad, the founder and CEO of the San Francisco Compute Company. SF Compute provides GPU infrastructure to AI companies. They make it easy to get access to the compute you need to train and run models without the long-term commitments and complexity that usually comes with it. What I love about Evan's story is that SF Compute started as something completely different. An audio model company. They bought a big GPU cluster, realized they couldn't get out of the lease, and ended up building an infrastructure business almost by accident. Now they're at the center of what Evan calls the largest infrastructure buildout in human history. We talk about the real economics of GPU compute, where Evan thinks the actual AI bubble is and isn't, and what it looks like to build a company that sells the picks and shovels in a gold rush. Enjoy.

Hey Evan, how are you?

I'm good. I'm good. How are you doing?

I'm good.

I'm good.

Excited to Talk more about the world that you're building here with the San Francisco Compute Company.

Oh man, you said the name.

There you go.

There you go.

You know, you have an amazing origin story with the San Francisco Compute Company, or SF Compute for short. You know, you started as this AI lab that was trying to train music models. You bought a year-long GPU contract, used it for a month, and then you had to sublease the rest or go bankrupt. Can you walk me through, and the audience through, like the first 12 months? Like what does don't die mode actually feel like when you owe $500,000 every 30 days?

It was not very fun. San Francisco Compute started as an audio model company. We were going to make something kind of similar to Suno or Udio, but prior to their launch or existence. And the first thing you do is you buy a big GPU cluster. In order for us to buy a big GPU cluster, both at the time and now, the scale that we needed, you had to reserve for a long period of time. So think about it in the same way that you can't get a lease on an apartment typically month to month. Unless you're in some weird situation, like you're renting from a friend.

Right.

No landlord would want that. Yes. The landlord wants a year-long lease, sometimes multiple years, especially if you're getting a bigger or nicer place. Um, and we were in the same situation. We went to all the GPU clouds, all the GPU clouds said, sorry, um, several-year lease. And like, of course they did, because if you're a GPU cloud, you have to go get financing for your cluster. Um, and your loan is for, you know, 3 years minimum. Um, that, or that's the payback time on that loan. And so if you sold month to month and your customers dropped off, you're bankrupt.

So— You're left on the hook.

You're left on the hook. So you really want to lock in a price for a long time. And obviously you maybe don't want to work with the 2-person startup in a San Francisco Victorian. Nonetheless, we found a GPU cloud that would, but as long as we signed up for a year. And we did—

That was their bare minimum, essentially? That's the bare minimum.

That's the bare minimum. Most people want 3 years. That's the— actually, most people want 5 years, but Most people are going to settle for 3 years.

So these would be the CoreWeaves of the world, the Nebulas of the world, et cetera?

All the GPU clouds. Any GPU cloud that exists, their financial structure is going to make them want as long of a contract as possible, and they're often willing to go down on prices for that. That's sort of the advantage of you working with them on that. But obviously they want long-term contracts, and they should. The problem for us is that we really only needed 1 month. Our goal, and the goal of basically every startup, especially ones that are training models, is to Use as much compute as possible to get your model out the door as quickly as possible, so that way you can get a product out the door and then raise your next round, so that way you can then get another big model and buy more compute and continue on the cycle.

It's a very bitter lesson, pilled.

Yes.

If you instead took a shorter, like, a smaller amount of compute and had it go longer, that's very suboptimal for you because somebody else who did the opposite will beat you to market. you know, get compounding advantages over time. So, um, our goal with our cluster that we had bought was, okay, so we've signed up for a year. Um, we've raised—

We really only need one month, you said.

We only need one month, um, because, um, we can get a first thing out the door in one month. Um, and our plan was we have about $500,000 raised. Every month it's about $500,000 to pay for the cluster. Um, I think it's actually a little less than that, like $400,000 or something like that. And so that meant that You know, month 1, good to go, totally fine, excellent. Month 2, that's a problem. And so there's 2 solutions that we had at that time. One solution was immediately raise another round, like take your seed capital, immediately raise another round, big round. This is what all the startups do. They go out and they raise the monster round.

Right.

At massive valuation, pre-revenue, all this other stuff. The problem with that is that if you then deploy and spend that monster round, and you don't make a product, you can neither sell the company, because the preference stack like messes you up, nor can you raise another round. Because no one wants to do the— like, you're already so high value, it makes it very difficult to raise another round.

Right.

So you're sort of in like, shut down the company or recap, or like zombie company forever. None of those are good options.

Right. So, or, I mean, you get, try to get acquihired by someone that actually does have endless compute capacity, endless compute budgets, et cetera.

Which is what happens to some people if you're like super lucky. Um, most of those people of course had much better backgrounds than we were. Um, we were like 2 kids in a, in a Victorian. Um, and so assumed that wasn't going to happen for us.

Right. Um, so, but it is true that, you know, there is the saying constraints breed creativity and I guess it actually happened.

Yes.

Um, so what we did is we, um, thought, okay, we've got a year's worth of compute. We can use the first month. Um, we'll just sublease the other 11 months. Um, we'll find somebody else who wants to take it off our hand. And those people are probably in the same situation where they also want one month.

Another Victorian next door.

Yes. Um, and they literally were in Victorians next door. Um, like that's, that was the gist. And so we called it the San Francisco Compute Group. That was the idea. Um, and we would just, you know, have this sort of come and go, um, hacker house style So the problem of that is that you are then operating a supercomputer. And it was initially just me and my co-founder at the time. And you don't have a lot of other time on your hands to do the model thing when you are also functionally running a cloud for lots of other people.

Right.

The clusters break all the time. We, because we didn't own the cluster, like, we didn't control it, we didn't build it, we didn't have any supervising over it. had very little, um, ability to fix it for other people. We could identify the problems. We could go to the person that we had bought the cluster from and like try and get them to fix it. But, um, you know, if somebody, if it broke, we were blamed, obviously.

Right.

Um, like as you would, as you should and would. Um, and so a good portion of the first year was try to sell the next month of cluster. Cluster has broken. We don't have the ability to fix it very well. Um, because we don't have control over most of the hardware. Um, and Um, meanwhile, try to build some sort of like minimal experience around it, um, on almost no venture capital because we had spent the first few hundred thousand dollars on, um, the first month. Um, and we would make, so we would charge more money per hour, um, than, you know, we were getting. So we were, we were making some money on it. Um, but like not enough to hire like a massive team.

You had essentially become an accidental Broker of GPUs.

We were an accidental cloud initially. Like initially we were just a cloud. So, um, and not one that was trying to be a cloud, definitely one that was just trying to like train a model and get out of this situation. Um, it was only after we kind of like agreed that, okay, we are in this situation where we are now on the cloud. And I distinctly remember this moment because, um, I had fallen onto a couch, um, being like, okay, what if we just did this, um, and as the thing, um, like we had to basically agree that, okay, we're not going to get back to training models for a moment. We have to figure out what to do with this GPU cluster. So, um, this is just the company now. We have to make this good for the people who are using it. We'll try that.

What was the moment for you where you went from that sort of accidental, you know, GPU cloud to, wait a second, there's actually a structural, like a massive structural opportunity here?

Yeah, so this—

That led you to sort of doubling, tripling down instead of going back to, okay, we sold, you know, we took care of our risk and now we could go back to training models.

Once we had the cluster and we had like agreed to do the company, then the next part was how do we like make this into a thing that works? And so our initial thought was we would do the same thing everyone else does, which is we're gonna make great software. And because we make great software, customers will reward us with higher margins on our GPUs. And because we have higher margins, that means we don't have as much risk of not selling the next month. You know, if we made a bunch of money this month because we were rewarded with high margins, then even if we missed a month, it would not be that big of a deal because we'd made it up before.

This is very similar thinking to how sort of traditional clouds charge, make money.

Exactly.

Extremely profitable, high multiples in the public markets.

Yes, this is the model of every CPU cloud, and it is what all the GPU clouds at the time were attempting. The difference between us and all the other GPU clouds is we had no money. Um, and when we went out to try and fundraise, obviously everyone looked at us and said no, um, because we were underwater like crazy, um, had basically no reason to do this, um, and were objectively a terrible investment, um, at the time. Um, uh, thank you, Atlanta Coyle, who is a, um, and you yourself and, um, other folks who eventually, um, and lots of other people, um, along the way. The problem with that is because you have no money, um, you have to do the like right startup thing, which is you go and talk to your customer and you try to pre-sell them. Um, and if you pre-sell them, maybe you can say, um, okay, if we build this product for you on this spec, um, will you pay X dollars per GPU hour? And everyone said, ah, that's a great product. Um, I, I'm super glad for this thing you're pitching me. Um, except for the fact that I would never pay you an additional markup on your GPUs. Like, obviously I can't do that because I am in the position where I raised the monster round. Um, and I, you know, have $40 million to spend. Um, that $40 million, uh, if you charge, you know, an additional markup on it, well, now I have to spend $60 million. And do you understand how there's only $40 million in the bank account? Um, like there's—

They can't pay for software.

They can't, they can't. Like, obviously they cannot. There's no, there's no product that you could possibly do that would justify you increasing your margin by any reasonable amount because the money's not in the bank account. There's no more money. That's it. That was it. Now we're, now we have 2 failed, 3 failed things. Um, The audio model is not working out. Um, the, uh, initial, like, sort of whatever the hell we were doing with trying to sublease our cluster was not working out. And also, um, no product that we seemed to ever pitch could ever do a higher margin activity. So then we were stuck with, what do we do, um, to make some sort of thing that works?

I view them as glass ceilings to break through, by the way. So you did break through.

Yeah, I guess so. And so the only thing we could think to do at the time was technically we have a website that says you can buy GPUs. Um, and technically people come in and ask us for GPUs, more GPUs than we actually have. Um, and we also know where to get these GPUs. Um, and so why don't we become brokers for a moment? Um, and that will help us, um, you know, figure out what to like, get enough money to kind of continue on and figure out what to do. Um, and so then we would do that. Um, and while we were doing that, um, we built up like a, a manual order book of all the clusters that we were operating and selling and buying, buying and selling.

So no longer necessarily building a software cloud, but actually building something more akin to a financial Yeah, yeah.

So this is— this was spreadsheets. Um, it was 2 people on the phone. Um, it was like Eric and Ethan, um, who would sit on different ends of a room and basically like one person would be on the phone with one person, another person be on the phone.

Slug GPUs.

Yes.

Um, and, um, this plan eventually turned into like an automated order book for it. The way that we had designed the order book, because like I— so as far as I understand, we like, we are the first compute market in existence, as far as I know. There are lots of other people who are trying to do various compute markets now.

Almost like financial derivative products, though, it seems.

Yeah, there's people doing derivative things, there's people doing aggregators, all sorts of stuff. The thing that I think is different about what we were doing is we explicitly built the market to solve the problem that we initially had, which is we had bought a big slug of GPUs, like a big block. Um, and then we had to sell capacity back over a time dimension.

Right.

Um, meaning that you're like, um, you're not saying, oh, I have some GPUs now and it's like on demand, and then I can sell those GPUs and release them. It's more saying you're selling blocks of time, um, over some period of, um, like years. And that was the thing people actually cared about. Um, because if you're— if you would raise the big round, um, you would raise the big round to buy a long-term block of compute. And then, um, if your product didn't work, you were in the same position that we were, um, just on a bigger scale. And you need to be able to get out of that contract because if you don't, your business blows up.

You're back on the hook.

Yeah. Because, um, at the time that we were doing this, um, businesses blew up. Like there is no maker of like, uh, I don't know if I want to say the name, but there are, there are, there were big AI labs that, um, sort of imploded, um, because they had bought too many GPUs because they They had like a $100 million contract, and if the revenue doesn't come in to support that, the company doesn't exist anymore.

Why did the suppliers on that side take that credit risk essentially then?

I think they were assuming that it would work like CPUs. In CPUs, people take credit risk all the time, just at a much smaller scale typically relative to their balance sheet. I think everyone who took credit risk at that scale basically lost money or blew up or made less money than they should have. Coming soon, um, there's not really going to be on-demand at any serious scale for the next chips. So like in A100s and H100s, um, yeah, people built, um, speculative clusters, um, and then, um, were trying to sell them on demand or were taking on a lot of credit risk. And so it meant that you could kind of get this, um, in Blackwell, you just can't, um, like the customer has to sign a long-term agreement, um, before the cluster is built.

Why is that? What's the sort of Key difference, you think?

Because everyone got wiped out.

Um, we just learned from the last cycle, basically. Like, there was a cycle that happened, and, you know, insiders know, outsiders don't necessarily see that.

Correct.

Yeah, like, we, we saw it because we were the brokers and we had an order book, and you could buy GPU— we could buy H100s for 40 cents an hour on us because the, like, people had built—

The people that own the chips actually had to get rid of them as well.

Yes. Um, because, um, they had built speculative clusters, um, and needed to get out of their speculative position. Um, and, um, capacity on that time was idle. And so prices went down. Um, nobody in their right mind would do that today. Um, people will build Blackwell clusters, the new, the new chip, um, in small scales and then offer that on demand at really high prices. But nobody's going to give you, um, a really big Blackwell deployment, um, unless you sign up for a, um, contract upfront. Um, small scale, totally. Um, that, that's like reasonable and it's a good decision to do. Um, but big scale, like You have to get offtake. Um, that's like the keyword for anyone outside of the industry.

So what—

can you explain, like, what do you mean by an offtake engine, right? You recently raised a Series A to build this offtake engine. Yeah, like, what does that mean? What, what are the different parts of it? You know, what are the different components? Like, what does it actually mean to build an offtake engine for the world?

Offtake is like the most important word in all of AI. Um, offtake is the contract that you get presumably before you build the cluster, that you take to your financier, the person who's providing the money of some sort, either internal capital or external capital. And you say, I'm financing against the credit risk of this offtaker. It's a long-term contract. That's all it means. We're building a company right now to, like, the way that we've designed our order book is specifically to generate offtake, to make it so it's really easy for someone to get offtake. That's what we do.

How do you make it easy to get off-take?

Everyone who is in the position of being an off-taker, someone who's willing to sign up for a long-term contract, was in the same position that we were at the beginning of SF Compute.

And these would be people on the demand side of your equation?

Yeah. So anyone who's an AI lab that is trying to buy capacity right now, when they are positioned with 2 options, one option is I can sign a 3-year contract, or I can sign a 3-year contract that I have the ability to get out of.

To sell back.

To sell back at some price. It is so much better to get the one that you can sell back than not. However, None of the GPU clouds want to do this because you will obviously make less money if you do that.

What is going through the head of the demand side customer here, the AI lab? Like they're thinking, hey, if the demand for my product changes, I'm still on the hook for this contract that I've signed with the cloud, right?

I think if you were a smart AI lab, you were looking at all the AI labs that blew up. Um, like a lot of them did and we don't talk about them anymore and they wiped out, um, everyone involved.

Including the suppliers?

Um, so the suppliers sometimes, yeah, sometimes, uh, the suppliers got hit pretty bad from it. Um, sometimes actually the people that got wiped out was the venture capitalists. If you were smart, if you were smart cloud, what you did is you got payments upfront. Um, and so what happened when, um, what happened when the cluster was idle and you weren't bringing in any money, um, is you just wiped out that initial capital that you paid down.

Mm-hmm.

Um, and the person who's paying for that is a venture capitalist. Um, so, um, both I, I hope, I hope the venture capitalists, um, are being smart, um, and encouraging people not to sign long-term contracts. Like, um, to be clear, the GPU cloud should always ensure people sign long-term contracts, but the demand side—

They're financially incentivized.

They're financially incentivized. Like, everyone, if they were operating smart, would be trying to do the opposite. They're like negotiating the, the midpoint between this.

Yeah, I often describe SF Compute's role then actually as like helping solve for the Rubik's Cube of problems that the supply side is facing and the Rubik's Cube of problems that the demand side is facing.

Yes.

That is the offtake engine.

Yeah.

And so what does that end up ultimately shaping up to be, this offtake engine? You know, the different sort of components, different paths that you could take here.

Yeah. So are you asking like, what are we doing today?

Yeah. What are you doing today?

Yeah. Okay.

It's quite— it's been a journey since the early days, right?

It has.

As you walk through.

Okay. So the first thing is that we're building our own clusters right now.

Okay.

But the way that we're doing that is we're working with outside financial parties who would like exposure to compute. We build clusters for them.

We then list them with us.

And they're viewing— these are new participants to the market? Like, are they viewing this as a—

Not always. Sometimes it's people who've already been doing this for a bit.

Okay.

Um, but a lot of the world's money would like to turn into GPUs. We help them do that. We look like a property manager. Um, so if you're used to investing in real estate, um, you might work with a property manager or a developer, um, who builds a building for you, who then, um, operates some sort of like rental agreement on it, um, a hotel, like does something with the building that generates a revenue stream that comes back to you. Um, so And that's what we're doing. We're taking— we're basically cutting out the GPU cloud in the middle. And the reason why we ended up doing that is because we couldn't get access to the rest of the components of the cluster. So our problem in the beginning was we had this cluster and it would break all the time and we didn't have the ability to fix it.

Right. It wasn't yours.

It wasn't ours. And we also ended up having to build all our software stack that is built on top of the hardware, uh, the, the worst possible way. Um, so in, in a big GPU cluster, there is, um, like the node you are selling, the bare metal node you're selling, but there's also everything else. There's the switches and the DPU and the BMC and the UFM. Um, it's like all these other components around it. If you don't have access to those other components around it, you end up needing to go down a technical path that is much, much, much harder to do. And you kind of can't use any off-the-shelf components, and some things you can't do at all. That's very painful. So to solve that problem, we ended up saying we're going to design the bomb, we're going to build the cluster, and then we're going to own the whole thing.

You're going to build the cluster and You build a cloud essentially as well?

Yeah, so we build the cluster. So, well, build here meaning we're working with other vendors to do the physical construction and so on. But we're designing the BOM, the bill of materials. Because we have built and own the whole cluster, we also operate everything like a cloud does. That's a much better technical position to be in than the one that we were in before. But what we do is we work with other financial parties who would like to own clusters and like to get the revenue stream from them because they're taking a position on The industry of compute, the industry of AI.

Participating in this model economy is what I call it.

Yes.

As you said, they want exposure to this asset class, to this economy. It has essentially become its own economy. It's the model economy, as I call it.

This is an objectively worse way to build a cloud in some ways, um, because the— you just like, we as a company end up making less money than other GPU clouds. Um, and that is kind of on purpose because what we're doing is we're splitting that money with the the capital party that's putting down the capital.

Right. You don't take your own sort of capital risk. You don't take your own credit risk.

Yeah. So the, the trade-off here is we have less risk as a company. We're very happy to do that because of the— our early days as being massively risky. We really don't like that. Um, so we get to be a stable entity. Um, the benefit of that is we're like selling that risk, um, to someone else who's gonna make a lot of money from it. That's a pretty good trade-off from, I think, both parties' perspective, um, because it gets So it gives the financial party exposure to AI. It gives us stability so we can build a product that's great over time.

And these financial parties are not necessarily AI people or cloud people, right?

They're— Sometimes they totally are. Oh, okay. Interesting. Sometimes they totally are. They're funds.

Yeah, interesting.

Yeah.

That want exposure to this.

Yeah. So it's a different way to get exposure. It's a way to diversify what you already have. It's a way to not depend on a specific cloud provider. Like, you could invest in a particular cloud provider, but if that cloud provider's contracts go bust, you are not in a great position. So it's a way to diversify your portfolio, and then it's also a way to just get more direct access instead of having this intermediary between.

If SF Compute didn't exist, how do you think the story ends for both sort of NeoClouds, these financiers, even startups?

One of the things I think that SF Compute does is it reduces prices for compute broadly. The way almost all the clouds have set up their capital structure And the way their company works is we're going to make great software, we're going to charge a high markup for it, and then we're going to get really great margins from it. The problem of this is this is divorced from what customers want. It's what the public markets would like them to be like. This is why they will— why they will present this. But it is not what any buyer of GPUs wants. What every buyer of GPUs wants is they want reliability, They want security, and then they want better price, and then they want shorter contracts. Um, and as long as you are reliable and stable, um, and you are secure, um, you are mainly focused on price.

They view it as a commodity, but the reality is like, it's not as simple as a commodity, I guess.

Yeah. So, um, the caveat of this is there are, there are lots of little details, um, in GPU clouds that matter tremendously.

Reliability, security.

Quality, price, contract duration are the key things. On the micro scale, it's little tiny details about the way that your system is implemented.

So now we know if SF Compute doesn't exist, how the story ends. But if SF Compute does exist and we do exist, what do the other clouds do? What should everyone else in the market be doing?

If you're a new cloud, you have 2 paths that you could go down. And many clouds today are doing both. Um, one path is you go straight to the top of the market and you deploy large clusters without a lot of services on them for reasonable prices to OpenAI and Anthropic.

Mm-hmm.

Um, you get offtake from the biggest buyers. Um, and those biggest buyers are cost sensitive and they have internal teams that can do most of the managed services that you offer. Um, And you should totally try and sell your managed services to them, but you should be okay if they don't buy them because you're mainly trying to scale volume for your clusters. The other thing you should do is you should make managed services.

The other path.

The other path. So this is things like offering a great Slurm product or great storage products, or, you know, you can push really high up and offer something that looks closer to like Modal. Modal is an excellent managed service. But I think these things are hard to do in the same company because you end up focusing your, um, your resources on both at once and they kind of don't necessarily overlap with each other. Um, an amazing example of this is Modal's product is amazing and really good. They are an excellent managed service. They do not build clusters.

They're focused exactly on that one thing.

Yes.

And as they scale and they raise VC and they make profit, Um, their money goes into making their thing even better.

Mm-hmm.

And their managed service is leagues ahead of any GPU cloud, any normal cloud.

Mm-hmm.

It's way better than anything anyone has ever produced because they have focused directly on that. Um, meanwhile, Modal cannot sell like a big GPU cloud, uh, or GPU cluster to OpenAI.

Mm-hmm.

Like, I, I don't think that's going to happen because, um, you, the amount of expertise and experience you need. in order to site and land a cluster that is signed off by OpenAI is a lot. And I would suspect that you will see pretty much every cloud split into doing one of the two. Probably a lot of them will keep going on one side or the other, but the winners will be the ones who specialize in one area. That's my guess.

Where do you think amongst that sort of two paths does SF Compute ultimately end up sitting in then?

Our goal is to sell to people who look like managed services. Um, part of what we're doing is trying to make the best cloud provider that would sell to managed services. It is a place in the market that's almost underappreciated because if you're a big cloud, the big clouds have competing incentives right now. One is the public markets want them to have managed services because they want them to be a high-margin business. But then OpenAI comes along and says, here is a massive contract Um, we're gonna make AGI. Um, would you like this money? And obviously, um, you should do the OpenAI thing.

Right.

Um, and so they try to do both. Um, and we think the thing we're trying to do is split those, um, and make the greatest thing we can for people who want to do managed services. Um, and the things those folks want are the same things that we cared about in the beginning, which is Short-term contracts, um, or contracts you can get out of. It's price. Um, it's specifically price. And SF Compute is not trying to be a thing that captures a massive margin on top of it. We're explicitly trying to make the prices go down.

Right. And the price is so important because of that sort of difference between CPUs and GPUs that we spoke about earlier.

Yes.

Like, can you double-click on that?

Um, so in CPUs, um, almost by definition, like, your customers are like SaaS companies. It's Gusto and Rippling or something. And Gusto and Rippling does not need more CPUs to make more money. Gusto and Rippling can make great software Um, and then they can, um, sell it like crazy and their CPU bill doesn't really go up, not, not significantly. And they end up with, you know, top-of-line margins because— They end up with SaaS margins. They end up with, you know, 70-plus, 70% margins.

You can build it once and sell it infinite times. Correct. That was the beauty of SaaS.

That was the idea. Um, that doesn't exist in AI. It's just like never going to exist because of the bitter lesson.

Here at Worldbuild, uh, our model economy thesis, which I mentioned, really was based on the bitter lesson, right? That the general models with scale, with more compute, with more data will beat the clever architecture, or even for me, the workflows and lightweight applications built around those models and with those models. Can you walk through why GPU buyers, people that even believe the bitter lesson, those GPU buyers, why are they different from CPU buyers? What is the Can you explain the economic reality?

In CPUs, the customers of the cloud are SaaS companies, and SaaS companies almost by definition have really fat margins.

Build something once, you could sell it infinite times.

Exactly. And then it gets better over time and so on. The problem of GPUs is that is not how AI works. In AI, you have the bitter lesson. And the way to conceptually think about this is if you're Doing some sort of generation, like you're generating a video or something, your customer's paying you for the video, and then you turn around and you use compute for that video, and you do that every time. And so as your volume scales, your margins stay roughly the same, and they don't look like SaaS margins. They look really thin. You as the customer need to be price sensitive. It is a requirement of your business. And that means that you push down the prices for your GPU cloud too. And so then your GPU cloud also has lower margins than the CPU clouds do. And you care less about fancy managed services. Um, and you care more, um, about, um, like price, as long as reliability and security and, um, so on are there.

Is the why for you here ultimately then, like the why around SF Compute to lower the cost of compute?

Like, yeah, totally.

Price is everything.

Totally.

San Francisco Compute is completely distinct in GPU clouds in that we are trying to give you the best possible price. Um, almost every GPU cloud in existence, and every cloud, um, is positioning themselves to the public markets as we are the premium experience, we offer these great managed services, um, and therefore you should, um, reward us with a high markup for our, our GPUs. Um, they tell the public markets that. Um, meanwhile, uh, customers don't agree. Um, customers instead say, if I'm buying a 3-year contract with you, um, I better get a discount. Um, and the GPU clouds behind the scenes do that, um, when they can.

Mm-hmm.

Um, we are explicitly trying to make the, the cheapest possible GPU cloud. Um, that doesn't mean we're sacrificing on reliability. That doesn't mean we're sacrificing on the product. Um, but if you wanted to design a business model that produced the best possible prices, that's, you would end up with SF Compute basically.

Lower prices, the more compute can proliferate.

The world gets better when you make prices lower. The world gets better when things are more affordable. It's almost like a rule of reality. And at the moment, I think a lot of what SF Compute has been doing for a while is implicitly betting on people like lower prices. And so that's what we're going to do.

And I guess, you know, through the various twists and turns, you know, now it's let's build supercomputers essentially to lower the cost of compute.

To be clear about like why our model makes cheaper prices is that if you're a fund and you got a 20% return on your fund over the year, that's amazing. That's really good.

Yeah.

If you got— if you're a startup and you make 20% margins, it's terrible. You don't want 20% margins. That's really bad. And so the expectation differences between like what a fund is doing and what a company is doing are different. And so the more your company is actually a fund, the more you should be a fund, and the less you should try and be a big sort of company with high margins. The problem is all the clouds have set themselves up and positioned themselves to investors as companies. Meanwhile, their assets under their balance sheet are growing and growing and growing, and actually they look like funds. Um, and they're not, and there's actually better people who are out there in the world who have cheaper costs of capital, um, even than the hyperscalers. Like there's, there's sovereign wealth funds and like, uh, massive players right now who are entering the space. They have cheaper costs of capital than most GP funds.

Right. That is the economic reality and you are playing for the economic reality.

Yes. We are the company that takes no bullshit. Um, we do what we say on the tin. Like our entire culture is we never had the opportunity to be fake. Like, we never had the opportunity to pretend to be anything other than the reality of the world. So in the branding of SF Compute, um, you will never see us say we're going to democratize compute or powering the AI revolution or something. It's like, no, uh, we sell GPUs, right? Um, we rent them to you. The reason why you use us, um, price, reliability, right? Like super simple shit.

It is, it is straight up. It is, it is for an economic reality.

Yes.

Versus some grand vision. And ultimately, the buyers of, you know, the folks that you work with, they definitely should have those missions.

Correct.

But I think it's just you make that— you make it possible. You know, part of my model economy thesis is that there will be bouts of volatility. Uh, we're in a GPU glut one day to all of a sudden we're in a GPU shortage. Like, you know, there is this constant chatter of a bubble. Are we in a bubble? Yes, this is a bubble. But where does SF Compute sit within that sort of Bubble spectrum.

If there is a bubble, here is where the bubble is. It is that in order to afford a large GPU contract, you go out and you raise the big round, and you then spend a bunch of that money upfront to pay for your GPU cluster because you're buying it for a long period of time, um, often longer than most companies fully have a plan for.

Like past their projections.

Yes.

You do that right at the beginning of the company. And you pay a— and you're not paying overtime and you don't have any way out. From the GPU cloud's perspective, they're relatively safe. So the bubble isn't in the GPU cloud? I don't think so, unless I'm wrong about this and lots of clouds are just like way overleveraged on accounts payable risk. I don't think that's actually the case. I think a lot of people are being smarter than they were before. Instead, I think the bubble is in Whoever owns the equity of those companies that just raised a whole big round, because if they don't succeed, then they can't get that capital back. And there's also nothing else you can do with that.

So how do you sort of frame SF Compute as like, you said this thing before, as like the anti-bubble company? Like, what does that mean?

If you think the bubble is in companies that are buying GPU clusters and it's their ownership right over those GPUs, that's where the, the sort of coyote that's walking off the cliff is. Um, then what you need to do is you need to give those people liquidity. You need to get those people an out, um, that they don't have to, um, blow up their company just because they bought too big of a contract.

Just because maybe their demand changed or—

Yes.

It was different than their projections.

And if a bunch of companies sell all at once, Totally. The compute price will collapse and so on, but it's not going to be zero. There's something you can recoup. Right now it's zero. Right now there's nothing you can recoup if you have signed a long-term contract. You're just dead. So that's the only way out as far as we see. So SF Compute's goal to some extent is to reduce the risk of an AI bubble, but where the risk actually is, not where everyone thinks it is. You have to be right. You have to know where it is. It's a very complicated subject. And lots of people on the internet say all sorts of things, and most of them are wrong.

Yeah. So as this anti-bubble sort of company, you know, I remember when we first met, the idea was that SF Compute would be this exchange for GPUs. Like, is that still the right way to position yourself to be the anti-bubble company, or, uh, is there, is there something else?

So I think that was the way that we were framing it at the time because it seemed like the easiest way to get the point across. Um, I've currently stumbled upon just calling it like Marriott, which I think is the better analogy. Um, if you don't know Marriott, Marriott doesn't own a lot of their hotels. Instead, they, um, the hotels are like owned by somebody else, but Marriott manages them. Um, and then, um, you know, they have their own genre of market, um, for their, um, their hotels that they're managing. Um, that's kind of what we are doing. Um, the clusters, financially not owned by us, but run and operated by us. Um, and to do that, what we had to do was we made a finance-grade ledger and an order book, um, that had all the weird intricacies of compute built into it. Um, and we built a KYC program and an AML program and the whole thing that you would otherwise do to make an exchange. But typically when we say the word exchange, most people kind of get the wrong idea. Um, or they go so far into the like financialization of compute thing.

They go derivatives.

Yeah. And that is totally a thing we may do in the future. There's lots of people who are trying to build that right now. We think there's lots more infrastructure you have to do before you even get close to that.

You have to settle the compute, as I call it.

Yeah. So for example, lots of people are trying to make indexes. The indexes, as far as we can tell, are built on people's wishful prices. It's like, what do you want your prices to be rather than—

Back to the economic reality.

Yes. And I would be very terrified of betting against an index that doesn't Settled to a real price or settled to the price that someone wants the thing to be.

By being Marriott, you are actually, you are at the forefront of that. Like you have information.

Yes, we have real deal pricing information. Um, and we have, um, yeah, we have the actual settlement price, um, that other folks don't have.

Right. So to finish the conversation, you know, the current landscape, let's say in AI, uh, It feels kind of like a cyberpunk dystopia, right? Everyone going crazy about Maltz book recently. Let's say there's anxiety about AGI, it's about scarcity, it's, it's quite frantic. But when I think about like the world that SF Compute is building, like the San Francisco compute company world, it's blue skies, green grass, landscapes.

Like, why is it important for you that the future of supercomputing One thing you will never see SF Compute do is position ourselves as, like, democratizing AI or revolutionizing compute or something. We like to be very does what it says on the tin.

Right. Anti-hype by nature, almost.

Yeah. I think you get, like, clouded by hype. Your brain doesn't work well, and people's brains collectively doesn't work very well. And if you actually think that the current moment is really important, and we do, you actually shouldn't be very hypey because it makes it hard for you to do anything real. The current scale of the GPU buildout is multiple times the cost of the Apollo project. For a while, people were saying there was gonna be a Manhattan Project for AI. We're like 30 times the cost of Manhattan Project by now. You know, it's the cost of a war. It's more than the GDP of various countries. It's a really big thing. It is an important thing. should think seriously about. And when you get consumed by the hype, you end up kind of being in this weird mental state that doesn't let you make good decisions about what you're doing. And if you actually think it is important, you should slow down, calm down, and think over a longer period of time. SF Compute right now is trying very hard to think, how are we going to do things over the next 10 years?

Mm-hmm.

Not how are we going to do things in the next 6 months? Um, and that is in part because in the beginning of the company, we had to think in the next month, month by month. Um, and that's so bad for your brain. And it's also so bad for like the things you were doing for other people, like your product and your, um, your employees and everything.

Well, thank you for joining us and, uh, hope the world learns something new about this model economy and where SF Compute sits in that. Yeah.

Thank you for having me.

Hey, this is Ben Kaznoka, co-founder of Village Global. Thanks so much for tuning in to the Village Global Podcast, where we go deep on all of the biggest topics in tech. If you enjoyed this conversation, please subscribe to our YouTube channel. You can check us out on Spotify, Apple, wherever you get your podcasts. We'd love to see you for the next one.