Hello and welcome to another episode of the Odd Lots podcast. I'm Joe Weisenthal.
Tracy, we're here in Chicago, second year in a row that we've done a live show here, and a great city for us.
That's right. And we're going to be talking to someone who we spoke with last year, but a lot has changed in 12 months.
So this is someone who is trying to build a market for compute, right?
So we're going to be speaking with Don Wilson. He is the founder and CEO of DRW. He is really at the forefront of this effort to build futures markets for things like GPUs and so forth. So check out the conversation.
We talked to you, Don, basically exactly a year ago. Last time we were here, we talked about things like GPU futures, but in like AI time it might as well have been ten years ago because so much has changed. But we want to talk GPU, market structure, all that stuff. The last couple of days have actually been kind of interesting in the AI trade world. Everyone's watching CoreWeave. There was an Anthropic investor meeting before the IPO. A little jitters. Maybe their revenue only grew, whatever, 50% versus 51%. I don't know the exact numbers. Give us your read right now. Like, is there a little slowdown happening in AI? Are we shifting into a lower gear, or is there anything in the trajectory that's changed, any new thing?
Yeah, that's growing this fast is not going to grow at a constant rate. It's not going to grow. And, you know, that first derivative isn't going to be constant. The second derivative isn't going to be constant. So to me, this is just kind of expected volatility. I don't know, maybe people went on vacation in August so they didn't consume as many Anthropic tokens.
But you're not reading anything more to it than that.
Okay, wait, so with something like CoreWeave, a big part of what you're doing with Compute Exchange and Silicon Data is this idea that if you create a market for compute, then cloud companies can hedge their future revenue streams, and the cost of capital can come down. In a week like this, when you see the yield on a CoreWeave bond at 15%, what does that tell you? Is the suggestion that, like, well, if compute futures were up and running, all the CoreWeave problems would be solved, like here?
I mean, I do think that it's highlighting that massive need for risk management products. And so, you know, think about if you're a neocloud and, you know, you sign contracts with the most creditworthy companies in the world, then you're fine, right? But there are a lot of other consumers of GPUs that are not the most creditworthy companies in the world. And if you're a neocloud, you want to be able to sell to them. And so I think what the market's saying is if you're selling on-demand compute or you're selling to less creditworthy entities, then your cost of capital is just going to be higher. Now imagine once GPU futures exist, you can, as a neocloud, buy the GPUs, deploy them in your data center, sell a strip of GPU futures, and then just sell on-demand compute if you want, or sell, you know, six-month contracts to less creditworthy companies that can post a big percentage of the contract value upfront. And the market is going to, you know, reward companies that do that, that basically just have, on a giant basis, trade heavily. Their cost of capital is going to be much lower.
Terry mentioned, okay, the GPU futures don't exist yet, and there's some message from the White House like national security, etc. And he's like, I don't get it. Do you get it? Can you tell us something about why these products don't exist yet? Maybe what the regulatory hesitancy is about the listing. And just remind us what the state of regulation is.
Yeah. So right now there's a comment period that's open that the CFTC extended it. You know, if they don't do anything, then, you know, the futures can get listed November 9th. How conveniently, November 9th is after November 3rd. And so, you know, maybe some people in Washington will be more comfortable with the compute futures launching. And I don't know, I can't explain what the concern is. I've heard all sorts of different theories. You know, one theory is that some of the indices are not very robust.
So I'm actually very curious about this, the robustness of the indices, because a lot of GPUs are sold via bilateral offtake agreements. Right. So it's like Anthropic has some deal with SpaceX, right, to use their compute. We don't really know the price of that. Why should anyone feel confident then when they look at an index of GPU prices that, like, it actually represents something resembling the state of the market?
Yeah. Listen, I think it's buyer beware. Learn about the index that you're trading futures on. Okay. You know, Silicon Data, that team has done an amazing job of collecting massive amounts of data from, you know, and it's focused on on-demand compute from the clouds. You know, there are a lot of neoclouds. There are very few hyperscalers. So their index is focused on neoclouds. Okay. There are many. And there is a lot of transparency about on-demand compute. And, you know, they've done a very good job to create an index that I, you know, will be IOSCO compliant. That's the benchmark. Some of the other providers, I don't think I've ever heard the phrase.
I ask you, what should I ask? Sorry.
So, IOSCO is, I don't know, it's the kind of group of international regulators that I think in 2013 came up with best practices approach for indices.
We should do an IOSCO episode one day. Now that you've been working on the futures contract for a while, can you maybe tell us what you've learned about, I guess, the pricing trajectory for GPUs over time? Because this is one of the concerns, right? You can say you're bullish on the compute market overall, but it doesn't necessarily mean that you're bullish on the value of a single GPU. It can depreciate over time. Maybe it gets more efficient and people move on to new things. So...
That is the trillion-dollar question. That is why I'm glad I asked. That is, yeah. And that is why these futures need to exist. Because, you know, nobody knows what the value of a B200 will be four years from now, five years from now. And, you know, and so by introducing futures contracts that trade, you know, way out, I think initially the CME's futures contracts are going to go out three years. And, you know, as they become more liquid, they'll go out more. But, you know, that enables people to hedge. That back in November of, you know, almost a year ago, the price of H100s, the hourly on-demand price, had just been declining, declining, declining. And everybody was saying, why do you need GPU futures? These things just go down. And then, you know, right around Thanksgiving, they turned around and kind of went up and to the right. And what happened, you know, it was Anthropic's newest model came out, which was a massive step up in usefulness. And all of a sudden there was a ton of demand for inference. And, you know, more recently we've seen, you know, looking at B200, B300, and H100 prices, another surge upward in price. And by the way, at the same time that the Silicon Data token index price has plummeted.
Yeah, it's interesting. And so, yeah, I mean, these things are evolving. They can go up, they can go down, and a hedging instrument is badly needed. I take that point. But just as a hypothetical, let's say GPUs depreciate over time and there's more demand, but in general the value goes down, which is something we've seen in a lot of commodity markets. Right. In that case, does having a market where you're able to hedge your future GPU revenue streams, does that just enable people to, I guess, borrow more and overbuild versus genuinely reduce risk? You're just allowing people to borrow against the same risk, ostensibly with a hedge, but the underlying isn't actually moving in your favor. Does that make sense?
No, no, I mean, here the reason it doesn't make sense is that for a futures contract to exist, there has to be a willing buyer and a willing seller. Right. So your hypothesis would work if they could just go and sell futures contracts at some price that was made up, maybe too high, and there was kind of infinite demand for that. Okay. Well, sure. But there has to be a rational buyer on the other side. And so I think what we'll see is that those forward contracts are going to be, are in fact going to be at lower prices. Just because there's so much hedging demand to sell in the forward space, and there will need to be, you know, investors who want to generate a return by buying those futures contracts.
What is the cast out a few years? What is this sort of ecosystem of clouds and inference providers that needed to exist for this to be a rich market? Because like, if the future of AI were just, let's say, Anthropic and OpenAI, and the future of cloud were just Alphabet and Amazon, there wouldn't need to be a futures market. It would just be all bilateral agreements, offtake agreements, etc. Yeah. So how much, like, how sad of a tale of non-OpenAI and Anthropic inference do we need such that this is a useful market that cannot just be solved by bilateral agreements?
Well, I mean, I think we're already there. I mean, look at the growing use of open-source models. I mean, it's exploded, and you're seeing companies with, you know, many billion-dollar valuations that have just been created to put open-source models on GPUs and sell the tokens. Right. You know, like Fireworks and Together. And so we already have a really rich ecosystem.
Okay, interesting. Yeah. Should we talk about perpetuals?
Yeah. Can I ask? Yeah. Can I ask, GPU futures. So one of the other things that people say, the skeptics, is they say GPU or compute is not homogeneous enough for a futures market. And there's a few different things they say, like there is different software arrays and that could significantly change the value. There's also the fact that there's like, at least in the inference part of AI, there's a lot of other players who want AMD and Cerebras and Etched and all these other Trainium and whatever the Microsoft one was and TPUs, etc. And this is just way too diverse of a market to really be captured in an index. How confident are you that either the market is going to be homogeneous enough for futures, or that there will be enough sophistication about the sort of basis between the H100 index versus something similar, such that this market could serve a diverse provider base?
Yeah. I mean, first of all, the amount of money going into, you know, just H100s or just B200s is large enough to support a single futures market. Now, in addition to that, these other things are very correlated. And so, you know, people will be able to trade a basis on these things versus whatever it is. And, you know, maybe, obviously, you know, these are going to be off of kind of the baseline standard configuration, whatever. And so maybe somebody is planning on doing a training run in maybe a giant cluster. Well, that's going to bear a premium, of course. But, you know, if the demand for inference goes up massively, then guess what? That training run is also going to be more expensive. So it's just a basis between the things. That's one thing. The other thing is that eventually, and, you know, we may be getting close, you'll be able to trade futures contracts on the output, on the tokens themselves.
Oh, token futures, token futures.
And so, you know, again, there's of course a diverse number of tokens. And people will put up the same objections. But I think that over time that's going to be possible. And by the way, that will introduce a really interesting possibility, which is now think about if you go and rent GPUs, you pay the on-demand price, and then you load up an open-source model and you start selling tokens through OpenRouter or something. Yeah. Right. Well, can you do that profitably or not? Well, that profit margin that you can generate is akin to, you know, what an oil refinery does. Right? And or what a soybean crush plant does and sort of a crack spread. So in oil, we call it a crack spread. In soybeans, we call it a crush spread. So you heard it here first in the year. We call this the crunch spread.
That's great. But token futures are something you're actively working on, it sounds like.
Okay. And what sort of stage would you say?
