The King of Chicago Trading Wants to Build a GPU Market Bigger Than Oil

0:00Your best restaurant location gets 5 star reviews. How do you make every location like your best location? Your best paper mill has been operating at peak productivity. How do you make every mill like your best mill? Your best data center has optimized every drop of water. How do you make every data center like your best data center? The answer is Ecolab. Better performance, better outcomes, better impact. Ecolab. Now every location is your best location. For enterprise organizations, managing all your food needs is a tall order. But with EasyCater, you get a single workplace food vendor with the tools and resources to make it easy.

0:40Giving teams across your organization an easy way to order from a huge variety of restaurants, all on one platform. All while consolidating your corporate food spend so you can control costs, streamline billing and payment, and simplify reporting. EasyCater, your business tool for food. To learn more, visit easycater.com slash podcast. When you own your own business, you own every decision. Now own the card that rewards you for it. Chase Sapphire Reserve for Business is a painful card that elevates your travel experience and offers premium benefits that can take your business to the next level.

1:16Sapphire Reserve for Business offers 8x points on all purchases through Chase Travel, 3x points on social media and search engine advertising, airport lounge access, and more. With over$2 ,500 in annual value, it's the card that gives back all you put in. Learn more at chase.com forward slash reserve business. Chase for business. Make more of what's yours. Accounts subject to credit approval. Restrictions and limitations apply. Cards are issued by JPMorgan Chase Bank and a member FDIC.

1:47Bloomberg Audio Studios. Podcasts. Radio. News.

2:03Hey there, OddLots listeners. You are about to get a conversation with Don Wilson, founder and CEO of DRW, sometimes called the smartest man in trading. This was recorded live on stage at Chicago's Untitled Supper Club. We had a blast and we hope you'll enjoy the show. All right, Don. Well, thank you for being here. Really appreciate it. Great to be here. Truly the perfect guest to talk about what's next in trading. But just to begin with, why GPUs? Well, obviously, AI is becoming more and more useful. And as it becomes more useful, people use more of it, which means they need to use more GPUs to run inference or train new models.

2:47And I actually have this theory that within the next 10 years, the world will spend more per year on GPUs than it does on crude oil. And that would, of course, make GPUs compute the largest commodity in the world. So it seems like you would kind of need a market for that. A very modest call. Just the largest market in the world. Yeah. It's funny because, you know, I associate oil often coming out of, you know, sandy deserts. But now they're literally turning the sand via chips into the commodity itself, or like breathing life into the sand. Just to back up, I have a million questions about this.

3:27For those who don't know, why don't you give us the sort of, you know, the 30 second or the 45 second description of what you do or what DRW is? Yeah, so I started off standing in the trading pit in Chicago and the Eurodollar option pit yelling and screaming. And then I would go home and write code on my Macintosh computer and build models. And essentially, you know, I don't stand in the pit and yell and scream anymore. most of the bits are gone, but we kind of do the same thing now with computers. I heard a story that you were once on vacation with your family and you were in Italy, I think in Florence.

4:04And instead of, I don't know, eating gelato or something like that, you decided to invent a new Greek letter for derivatives trading. This is cool. Yeah. So here, I mean, you're confusing two stories. So actually, what happened was there was a new exchange that had launched an interest rate swap futures contract. It was called IDCG. And I looked at the contract and I figured out that actually they had not designed the contracts properly. And so although they were telling everybody that it was economically equivalent to a regular interest rate swap, it wasn't because it had this additional convexity bias in it, which is we could talk about convexity bias.

4:51It goes even more in the weeds than a lot of your podcasts go into. But so when I was in Florence, I had this idea of how you could create an interest rate swap futures contract without this convexity bias problem. And that is what I focused my time on there. What was the letter? So back to the letter, the letter was about after a really unpleasant period in the euro dollar option pit where all the market makers lost tons of money because the shape of the skew shifted dramatically as the Fed started hiking in a very predictable manner. And nobody had really developed a measure for linear skew.

5:33And so during the week, I said, well, this isn't that much fun. We're losing a lot of money every day. But the good news is that that means we have something to learn. And so I spent the weekend working with the quants and we came up with, you know, kind of a measure of the linear skew between the calls and puts and decided to use the Greek letter psi to describe it. And, you know, so by Monday morning, we had put it into the risk and onto the sheets. And before the open, I explained to the traders, how to talk about it, how to use language around it. And before you know it, we had made the money back because we were able to trade, manage this risk better than anybody else because we had a whole language around it.

6:19Amazing. So we've established your street cred when it comes to solving problems in contracts for financial instruments. If I think about a GPU future or something like that, the first problem that comes to my mind is standardization. Because, of course, you know, all different types of chips, different types of memory, different latency, I guess. How do you go about addressing that? So that's a great question. And right now, so what we've done is we set up two companies. One is called Compute Exchange, not very creatively named. We have a tendency to do that. DRW is your initials, right? That was my trading badge.

6:57And yes, also my initials. Yeah. I mean, we did better later on with Cumberland, our crypto trading arm. That was actually a reference to the Grateful Dead song about the Cumberland Mines. Oh, I didn't know that. I didn't know that either. Yeah. One of my partners who does the more creative naming came up with that one. He's a dead fan. Anyway, the other company is called Silicon Data. And Silicon Data's job is to create indices that will become tradable, will be viable to have futures contracts listed on them. And right now they've created a number of different ones, but one is the H100 index.

7:38Another one is the A100 index. And believe it or not, those indices are both available on Bloomberg. Oh, amazing. That's a love hearing that. If we were in the studio, I would already be looking up the chart as you were talking about it. Who are the natural participants? Because when I think about AI or training, you know, imagine someone goes to one of the big cloud vendors and they sign a long-term contract or whatever. Who are the participants who would be better off in an environment where there was a liquid market for compute? So what we found, and DRW actually uses compute exchange to source compute, and we find that because there are something like 70 different cloud providers that participate, you can often get better pricing.

8:27And one of the things that you can do is you can specify if let's say that you're an AI company and you know roughly what kind of cluster you want, you can specify that. You can even say, you know what, I'm indifferent between locations or, you know, if it's in the Middle East, I'm still OK with it, but I want to pay 20 cents per GPU or less. Whatever it is, you can kind of express your preference curve. Compute Exchange can conduct an auction and then, you know, find the kind of best price compute that matches your needs. So that's kind of the idea of how it works. And, you know, it probably doesn't work if you want a 10 ,000 cluster monster for doing a huge training run.

9:11But for inference, it works great. Or for smaller training runs, it works really well. Is the broader impact the idea that once you establish a liquid market where people can presumably hedge their exposure, that that would bring down the cost of capital? So that's right. So once you have a liquid market, then you have much more confidence in the indices and you can then list futures contracts. And so what does that do? It enables the neoclouds that are going out, raising capital, buying a bunch of GPUs, putting them in data centers and kind of hoping that they can rent them out and not really knowing what they're going to be able to rent them out for six months from now, let alone two years from now.

9:57So a neocloud could buy the GPUs, sell a strip of futures contracts. And I envision that these will be traded kind of like electricity futures where there's one for every month. And if you want to hedge the next three years, you sell 36 of them. And now you've locked in your pricing. Obviously, their cost of capital is going to go down, which in turn should make GPUs more readily available. And then on the flip side, if you're running an AI company and you raise a finite amount of dollars and you kind of know how much training you're going to do, but you don't know exactly what configuration, you can go ahead, buy the compute in the derivatives market.

10:39And then once you have a clear view on exactly what configuration you want, then you can swap those derivatives for actual compute. Talk to us a little bit more about the cell side. So like we have these like big clouds, right? The ones that everybody knows. And then you mentioned the neoclouds. Do you see that changing? Like what do you see as the future mix of cloud vendors in the future? So that is a great question. I think that the whole space is going to grow, but that the AWS GCPs of the world will make up a smaller percentage of the whole. Okay. That's my guess. How come? Because there is such proliferation of other companies buying GPUs and deploying them.

11:29Okay. That's a good answer. You know, Joe asked you who would be the natural market participants for this. I'm going to ask you the opposite question. Who wouldn't want this? Because I think of some of the hyperscalers, They seem to like controlling the GPU supply and maybe squeezing some of their competitors. Would you expect resistance from them? Yeah, I mean, I think the hyperscalers benefit from opaque pricing and kind of bundled pricing. And of course, they would prefer to have all the GPUs. But NVIDIA... I would also prefer to have all the GPUs. Yeah, yeah, that's always a good thing. But I think NVIDIA wants the GPUs to be widely distributed.

12:07And they're really the ones that make the call. This isn't the first time that there's been an attempt to create futures markets out of technology. I think there's been multiple efforts decades ago to DRAM futures. It doesn't seem that fundamentally different, although maybe it is. Why did those fail? When you think about what's going to be different at this time, what was the failure that caused? Why didn't DRAM futures take off? So the thing about DRAM was that the price just kept on going down. So in a very predictable way. And so why would you want to buy a futures contract if you know the price in the future is going to be lower?

12:46Whereas GPUs, we've certainly gone through periods where GPU demand was super high. And then we've gone through a period where there was kind of some excess supply. So there's not a consistent trajectory of pricing. I think that there will be a consistent trajectory lower in terms of, I don't know, however you want to measure it, dollars per flop or dollars per token. I think that that's going to continue to decline. But, you know, an H100 is going to be a useful GPU for a very long time. And over its life, I think there will be periods where there's more demand, less demand and, you know, a little bit more cyclicality and less predictability.

13:28So I know that the Trump administration has said that they want this market to happen. Right. So you seem to have some regulatory, I guess, tailwind behind you. Yeah. I mean, I don't think that this is a controversial thing. I think that it's pretty clear that once we figure out the right index construction and have kind of sufficient data that I don't think the CFTC would complain about the product. Thank you.

14:25How do you make every data center like your best data center? The answer is Ecolab. Better performance, better outcomes, better impact. Ecolab. Now every location is your best location. How many vendors does it take to meet all your organization's food needs? Just one. EasyCater, the workplace food platform that lets teams order from a huge variety of restaurants, over 100 ,000 nationwide, all through a single vendor. In addition to all that variety, Easy Cater also gives you full visibility of your organization's food spend with invoicing, centralized reporting, and seamless integration with expense management systems, all on one platform.

15:08Easy Cater, your business tool for food. To learn more, visit easycater.com slash podcast. You're thoughtful about where your money goes. You've got your core holdings, Some recurring crypto buys, maybe even a few strategic options plays on the side. The point is, you're engaged with your investments, and Public gets that. That's why they built an investing platform for those who take it seriously. On Public, you can put together a multi-asset portfolio for the long haul. Stocks, bonds, options, crypto, it's all there. Plus an industry-leading 3.8 % APY high-yield cash account. Switch to the platform built for those who take investing seriously.

15:52Go to public.com and earn an uncapped 1 % bonus when you transfer your portfolio. That's public.com. Paid for by Public Investing. All investing involves the risk of loss, including loss of principal. Brokerage services for U.S.-listed registered securities, options and bonds, and a self-directed account are offered by Public Investing, Inc., member FINRA and SIPC. Crypto trading provided by Backed Crypto Solutions, LLC. Complete disclosures available at public.com. This is a little bit of a sideways question, from your attempt to build this market. But speaking of the cloud, in your main business at DRW, I assume you're sort of major customers or users of the CME.

16:28Are you excited about the CME's migration of its backend to Google Cloud? Because they tout it, they talk about their partnership with Google, et cetera. As a client or customer, are you enthusiastic about this move? We interviewed Terry earlier today and he was excited for sure. Yeah, so it depends on what you put into the cloud. And it's totally fine to put a lot of things into the cloud. But the thing that you don't want to put into the cloud is a matching engine. And the reason for that is you want the matching engine to be as deterministic as possible. So that means that if you send two orders into the matching engine, one, let's say, a couple of microseconds behind the other one, You want the one that gets there first to be filled every time.

17:20And if you put stuff into the cloud, it's very hard to make that happen. You wind up getting a wide distribution around which order will be filled first. And even as you kind of stretch those times out, you could have an order that comes in maybe a couple milliseconds later be filled first. that is super disruptive for liquidity providers. And it means that the liquidity in the market is going to suffer. But this is, you say it's not ideal for them to have a matching engine in the cloud, but this is the direction it's going in. Yeah. And it's unclear exactly which part of the matching engine will be in the cloud.

18:03Is it some kind of a dual structure? I don't know, but that's what matters is a deterministic matching engine. I mean, if Google can figure out how to make a matching engine in the cloud deterministic, go for it. I'm very skeptical that that's even possible. Can you just describe the sort of theoretical problem? What is it about cloud computing that makes this particular problem, the deterministic aspect, difficult as opposed to traditional infrastructure? Well, when you have on-prem computers, it's all right there. You can control where the wires go. And so when it's in the cloud, it's a little bit more, well, nebulous, I guess.

18:43It's just harder to do. That's a good pun. I admire it. So you mentioned that you have this long and storied career in the trading industry, starting from old school trading. And now we're here talking about GPU trading and what's in the cloud and what works and what doesn't. Tell us what your company, what DRW is actually doing when it comes to practical application of AI? This is a question we're asking everyone. We ask all companies to spill all their proprietary secrets about AI. Excluding the engineers. We know that they're generating code. We know people are coding. Yes, we know that they're using clogged code or whatever.

19:21So besides the engineers. Yeah, yeah, yeah. You're right. That's kind of the boring answer. Yeah, that's the boring. And then the other thing is then when we ask this question, people cite a bunch of machine learning things that have been around forever. So let's talk about actual AI. Yeah. So I think that the way that we make trading decisions is going to change dramatically. And it already is. You can use AI to interact with your proprietary data, your proprietary models, and suggest trades. That's pretty cool. Are you doing that right now? Yeah. So we're starting to do that. But we have some tools that kind of do that now.

20:02And And the other thing that's really interesting is to fiddle around with agents and have different agents interact. And so you could kind of think about maybe you have a couple different analysts, AI analysts that both work on some stock. And then you have kind of a risk taking agent or maybe a couple different risk taking agents that interact with those analysts and then come up with trades based on that. So, I mean, these are, you know, that's a little bit of a theoretical concept, but I don't think we're that far away from things like that. Just on the cloud trading a little bit more, I am really interested in this topic.

20:41What is the current state today? Just so that we understand where you're at. Like, what is today's snapshot of usage of the platforms? I mean, as far as where the matching engines are? No, no, no. Oh, sorry. On the GPU trading. Oh, the GPU. How active is it right now? Like, where is the state of the business? Oh, you know, I think last month we conducted five or six auctions. So it's early, but it's happening. So when I think about how futures contracts are born, it's usually bespoke options. And then you get the index, I guess. And then you get a forward and then a future. That's kind of how I think about it in my head.

21:19Is that the process that you imagined for this? Not necessarily. I think that the simplest, I mean, yeah, I suppose you could do some privately negotiated compute swap or something, and maybe that will happen first. But no, I think the first thing is a futures contract that settles to an index. If the spot market becomes really liquid and you have very standardized auctions, and one of the things that you asked about was, well, how do you deal with the lack of standardized? you know and and so one thing is you go to a certain type of gpu you know h100 for instance but even within that you can configure them in different ways you could use infiniband you could use some other way of connecting them and so what's important is you need to decide on some benchmark and one of the things that silicon data has done is they've actually built some measurement tools that measure how fast a GPU cluster is.

22:20And so you can then say, okay, well, in order for this GPU to be kind of eligible to be in the index, it needs to meet a certain standard. And you can, there are a couple of different vectors you can measure by. So I think that that's kind of how you would do it. And then if you got very liquid auctions, you could actually have a futures contract that cash settles to the auction price. And then people could have the option of either essentially just cash settling their derivative and walking away or cash selling their derivative and participating in the auction. And they would know that price would transfer from one thing to another.

23:02That might be a future state of the world. And the initial state is probably just a generic index and the futures cash settle to the index. What would a market failure look like in GPU trading? Because your analogy is the oil market and, you know, weird stuff happens in the oil market. Could we get negative GPU prices? Or if everyone wakes up one day and decides they want to use ChatGPT as their psychotherapist or whatever, which some people are doing, could you have a GPU shortage where maybe people can't deliver into the contract? There are lots of ways that markets can break and go wrong. And I remember to this day that when oil futures went negative, it was during COVID.

23:42I was sitting at home. I was trading oil futures and I bought oil futures for negative prices. You were one of the ones who actually got us. Amazing. My then, what was that? 2021. So my then 14-year-old said to me, please, please, please, I want to buy negative priced futures contracts. And I said, well, you have no way of taking delivery of the oil. And he said, I will go to Cushing, Oklahoma and figure out how to do it. You've really raised a son, daughter, son. Son, yeah. You've really, he's been learning. We have an episode about taking physical possession of oil. I do not recommend it. Turns out if you keep it on your desk for long enough, it evaporates into the atmosphere and poisons your colleagues.

24:30Yeah. Anyway, a little bit of a tangent. So I think on the upward trajectory, if there's tons of demand, you know, that's something that commodity markets are really good at dealing with. The price will go up and more supply will come in. And I think that's all good. On the downward side, you know, you can always just turn the GPUs off. So I don't think they trade negative. How much of the volatility that do you when you anticipate market volatility in the price of GPUs? How much is that like embedded electricity cost? So when you buy compute, right, you're buying the chip, but also the power, like how much of that volatility will be the power?

25:08So the industry lingo that's used is total cost of ownership. And, you know, what percentage of the total cost of ownership is the power price? And for an H100, it's less than 15 percent. Less than 15. Yeah. OK. So GPU trading, obviously one of the things you're working on, but you're a busy guy and you've got other stuff up your sleeve. What are you doing in the realms of tokenized trading? So that is an area that we're super excited about. And we've been thinking about this for a very long time. So in 2012, when we started talking about Bitcoin at DRW, and there were a number of traders at DRW that were very excited about Bitcoin.

25:49You were very early into it. 2012 was still pretty early. Very early. Yeah. So we were having these discussions of why is this interesting? Is it interesting? What about it is interesting? And we came away with the following thesis. There's some small chance that Bitcoin could be digital gold. I don't know, you know, call it 1%. It's kind of an interesting product. So we should probably make markets in it. So we set up Cumberland as the, and, you know, we didn't call it DRW because at the time everybody knew that anybody trading crypto was obviously a crook. So, you know, we wanted to kind of separate the brand a little bit.

26:24But the other thing was this idea that you could move value instantaneously in a trustless ecosystem was super interesting to me. And I said, wow, if you could do that in traditional financial markets, that would make the market so much better, so much more resilient. And so we should really figure out how to do that. So we started a company called, again, not very creatively named, Digital Asset Holdings, which created the Canton blockchain. Initially, the Canton blockchain was a private permissioned chain, but last summer it actually became a public chain. And that chain was designed specifically with tokenization of traditional financial instruments in mind.

27:07So it has a couple of characteristics. One is it has configurable privacy. And believe it or not, for people who are in the finance business, They don't want to broadcast to the entire world when they are buying or selling something. I mean, obviously, if it's above the reporting thresholds, you do. So that was kind of a fundamental characteristic of this chain. It's different than Ethereum or Solana or any of these other things where if you tokenize something and put it on top and you move it around, everybody sees it move around. So that's kind of something we've been working on for quite a while.