The Accidental Supercomputer: How SF Compute Started | Worldbuilders

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 amount, 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. They'll get compounding advantages over time. So our goal with our cluster that we had bought was, okay, so we've signed up for a year.

We really only need one month.

You said we only need one month because we can get a first thing out the door in one month.

Mm-hmm.

And our plan was we have about $500,000 raised. 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 thousands of thousands of dollars.

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—

Yeah.

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 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 supercomputer. 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. Um, we, because we didn't own the cluster, um, like you didn't control it, we didn't build it, we didn't have any supervising over it, um, 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.

Right. Um, but like not enough to hire 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 just did this and has the thing. 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 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 GPU cloud to, wait a second, there's actually a structural, like a massive structural opportunity here?

Once we had the cluster and we had agreed to do the company, then the next part was, how do we 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 going to make great software. And because we make great software, customers will reward us with higher margins on our GPUs.

Hmm.

And because we have higher margins, that means we don't have as much risk of not selling the next month. If we made a bunch of money this month because we were rewarded with high margins. Um, then, um, even if we missed a month, it would not be that big of a deal because we'd made it up before. This is the model of every CPU cloud, and it is what all the GPU clouds at the time were attempting.

Hmm.

The difference between us and all the other GPU clouds is we had no money. Problem with that is because you have no money, um, you have to do the, like, right startup thing, which is you go and you 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 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.

Yep.

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? 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. And while we were doing that, we built up like a manual order book of all the clusters that we were operating and selling and buying and selling. This plan eventually turned into like an automated order book for it. 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 like people doing derivative things. There's like people doing aggregators, all sorts of stuff. The thing that I think is different about what we were doing is we, we explicitly built the market to solve the problem that we initially had, um, which is we were, 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 had like a $100 million contract. Um, and if the revenue doesn't come in to support that, uh, the company doesn't exist anymore.