算力是一个在群聊里交易的万亿美元市场 | Brett Harrison & Andrawes Bahou

Welcome to Token 2049. Token 2049 returns on October 7th and 8th, bringing together 25,000 attendees, 300 speakers, and 500 exhibitors at the world's largest cryptocurrency event. Token 2049 is taking place in parallel and in partnership with our own Asia Digital Asset Summit. So you can attend both conferences in Singapore within one week. There will be over 1,000 side events during Token 2049 week, culminating in the post-2049 and Formula 1 weekend, and the speaker list is very packed. Shane Kopan of Poly Market, Jeff Yang of Hyperlid, Arthur Hayes, NASDAQ CEO Adena Freidman, and many others. Join us in Singapore on October 7th and 8th for Token 2049 and Digital Asset Summit Asia. Nothing said on Empire constitutes a recommendation to buy or sell any investment or product.

1. Why Does Compute Need Markets?

Okay, everyone. I'm really happy about this. I've wanted to do this episode for a long time. I feel like I don't understand the computing markets well enough, so I'd like to invite two of the smartest people in those markets that I know: Brett Harrison, founder of Architect, and Andrawes Bahou, founder and CEO of Compute Desk.

I'm glad to be here. Thank you very much.

Nice to see you. Glad to be here.

So, guys, I think Brett, I could ask you to start by saying that you two are very knowledgeable about the computing markets. We have derivatives, indexes, and markets, and I think that for those who haven't been following us, a lot of computing markets will be launching in the relatively near future. What market problem do computing technology markets solve? Why do we build markets around computing technologies?

The main unit of expenditure in the AI economy is companies renting or buying GPUs to perform various types of training or inference tasks. As I'm sure people know, the amount of money we're talking about here is extremely large. People are spending $100 million to $1 billion, or even $10 billion, at a time on large-scale computing contracts.

The companies that create them are sometimes called neoclouds or hyperscalers. Many different companies in this field are involved in selling computing technology to other companies. This raises the question of what will happen if the prices of computing technologies decrease over time. What if they grow significantly over time?

That difference in potential risk, and the price at which they could offer these computing technologies, is a commodity that cannot be hedged today. We're trying to collectively track the prices of computing technology and create derivatives based on computing technology so that people can hedge or speculate and have full transparency about the future value of computing technology.

Good. Who are these people? Can we identify them? We have labs, hyperscalers, neoclouds, and cloud companies. There are many of them, and they are synonymous with one another. Can you, perhaps, Andrawes, tell us who the players are here?

Once upon a time, there were hyperscalers. That's Google, AWS, whatever. They bought GPUs, accumulated hardware, and then made it available to you on AWS, GCP, or Azure.

Then people started needing a lot more GPUs. The buyers of hyperscale capacity were, of course, small research labs, then larger research labs, AI companies, and developers who needed to train new models, perform inference, and do everything related to machine learning workloads.

You probably know the hard lesson, which is that you really should increase the amount of compute and the amount of data. If you increase the amount of compute, your models will continue to increase their capabilities. Along came OpenAI, and this started the race to create something new, until the consensus was, “Let’s just create better algorithms.” But then, at that point, they said, “Let’s just build bigger computers.” So they started making them much, much bigger.

In the process, because of how capital-intensive this new infrastructure was, a new type of participant started to emerge: lenders. Lenders are large private credit firms—not necessarily just private credit firms, but credit firms of all kinds.

Eventually, the bond markets started coming in because hyperscalers needed a lot more capital to create these clusters. They came in and started guaranteeing free cash flow for hyperscalers. The idea is that Google can generate a lot of money from advertising, but it has a big, expensive data center project. The lender is happy to guarantee this because they think, “Google will always pay its dues. We’re going to lend it the capital to build these clusters.”

This led to the emergence of three types of participants in this market: lenders, infrastructure providers, and then buyers. These are businesses and AI labs. Now things started to get a little more fractalized.

Right. So, out of these three types of players, just to be clear, the players today are Google, Google Cloud, Amazon, OpenAI, Anthropic, JPMorgan, and Goldman Sachs, for example.

Yes, for example. Now, the world dominated by these big hyperscalers is starting to come to an end. Previously, if you needed compute, you went to one of them.

In the 2010s, new, smaller companies that were just providing GPUs—not a full cloud experience, just GPUs as a service—came along. Companies like FluidStack and Lambda saw a need for machine learning engineers and AI companies that needed to run machine learning workloads or train machine learning models. They said, “We’ll just have a cloud of GPUs for you. You don’t need to bother with all this AWS stuff. We just have a cloud of GPUs for you.”

Now fast-forward to today, and what has happened is the hyperproliferation of these types of GPU-as-a-service providers. Some call them neoclouds. I don’t like the term. Some people call them data centers with GPUs. Some are neo-hyperscalers, such as CoreWeave.

The thing is that there are many more new participants in this market, and this will be an important moment. When there are a very large number of participants in a market, it is usually a sign that something is being commoditized, and that is exactly what is happening now.

Many people see a low barrier to entry into the computing services market. This is a product. There’s not much differentiation here. There’s crazy demand for computing equipment. Everyone wants to serve it, and a lot of people can. That’s the important part: all you need is some capital to buy the equipment, a place to put the equipment, and demand for it. That’s all you need to play this market.

So suddenly, with this hyperproliferation of new GPU-as-a-service providers, we’re tracking something like 500 of them now. Previously, GPUs could be obtained from 3 companies. There are 500 of them now.

So it just so happens that you can get GPUs.

Yes. Or you can get compute on a GPU, and then rent the compute for anywhere from 3 to 500.

Yes.

2. Compute Still Trades In Group Chats

What emerged from that is now a very interesting category. This is a phenomenon of the last 2 years: a whole bunch of brokers for computation. There are a lot of people who need computation, so the cardinality of the number of people who need computation and the cardinality of the people who supply it both increased. Brokers came in to essentially match supply and demand, and a very inefficient market was formed.

Well said. Yes. One of my friends was the chief of staff at one of the big AI companies, and I asked, “What do you do all day?” He said, “I’m on the phone with brokers trying to get compute.” And he said, “This is the worst job. They all quote different prices at different times.”

That was a few months ago—probably over a year ago—so I think things have probably changed now. But the market is evolving every day in all directions.

And so, first, to Andrawes’s earlier point: if you have access to any source of electricity, to the power grid, or to the ability to buy GPUs, people are creating these little neoclouds all over the place, in every country, through various kinds of renewable energy sources, in abandoned gas stations, old car depots, and refurbished Bitcoin mining rigs. It’s everywhere.

Now this spread is happening. The question is, first of all, what is the actual fair price for any of these things? You have this giant, opaque market. Most of the actual buying and selling of compute happens through text messages, phone calls, and opaque channels.

There is a complete price gap between what hyperscalers offer for compute and what neoclouds offer for compute. It is a very heterogeneous market in terms of what compute actually is, and it can vary depending on whom you get it from. Regardless, there’s a critical need right now to bring this pricing out into the open and standardize it, and to have a liquid tool that allows you to track and hedge what’s actually happening in the computing market.

And that extends from computing outward in both directions from a supply-chain perspective, because there are a lot of things that go into creating compute. There is electricity and energy. There are foreign currencies so that you can purchase GPUs. There are metals that can be used in memory production, on the one hand. On the other hand, there are things such as inference.

So, there are many different aspects that affect the cost of computing and what computing is used for. That’s the industry we’re in together: how we hedge the prices of everything related to computing and everything that’s part of the computing supply chain.

It’s shockingly reminiscent to me of Bitcoin in 2017. There were a few years of peer-to-peer trading, and there was sometimes a 20% price dislocation between Asia and the U.S., but it didn’t matter because nobody really cared about the markets until 2017. Bitcoin was skyrocketing, going from 1,000 to 20,000, and then, oh my God, we have to create really efficient markets around these things. This is just 100 times larger in scale, maybe 1,000 times larger in scale, than that.

Is this it? What are the similarities, and maybe what are the differences?

Yes, it’s probably more than 1,000 times. I don’t know—if I do a quick count, will it be more than 1,000 times? That’s more than 1,000 times, and it has some similarities and differences.

As far as similarities go, I’ve heard people talk about what the OTC markets looked like in digital assets about 10 years ago: very broker-to-broker, peer-to-peer, opaque pricing, and large dislocations and arbitrages between different geographic locations and markets.

One of the big differences here is that it is absolutely a commodity, but it’s not a perfectly interchangeable good. If you have 1 bitcoin, 1 bitcoin is 1 bitcoin. That’s true. It’s in the address; it’s in the wallet. It is not tied to a specific location or provider. It exists on its own.

Computation does not exist on its own. It’s not something you pull out of one place and carry to another. It’s a service provided by a data center that determines everything from the amount of memory on each chip to the type of network, the cooling—liquid or air—whether there’s good software for managing GPUs, and the utilization rate.

There are so many variables that make computing a heterogeneous commodity, which is precisely what makes it difficult to immediately value and bring to market in the form of an exchange.

Good. Can you tell me, maybe going back to Andrawes for a second? By the way, is your name Andy or Andre?

3. Why Finding GPUs Takes So Long

Both. My name is a bit difficult to pronounce, but my own mother calls me Andy.

I think I understand. I think I’m close. Okay, let’s go back to Andy for a second.

I was listening to a podcast with the founder of Instinct on Invest Like the Best. Have you been at Meta before?

I was at Meta.

You were at Meta. There’s a lot of buzz around Instinct. The founder of Instinct said that 40% of his time is now spent trying to get computing resources.

I’m wondering what that really means. Who is he calling? Who is he writing to? What company is he calling? Does he contact a broker? Do you just call someone? Are you calling the data center? Are you calling a hyperscaler? Who is this? What does it look like?

We said earlier that there are now hundreds of computing-power providers, and there are probably just as many brokers of that power, if not more. What he is probably doing is calling his top 10 friends who run AI companies and asking, “Who’s your guy who provides you with computing resources?”

They’ll send him a list of 10 people, and he calls them. They say, “He’s a good guy. You should talk to him.” So he picks up the phone and says, “Hey, I need compute.” They reply, “Great. It’ll be available in 16 weeks.” He says, “Okay, I need it today. We’re growing very quickly. I just raised $1 billion. I need a lot more, and I need it a lot faster.”

He’s going to bypass all these brokers and all these end-to-end compute providers, from the hyperscalers all the way down to the random broker who promises they’ll have 40,000 GPUs in 2 weeks. Because of the opaque market, you don’t know how trustworthy someone is behind the scenes. They could be a broker for another broker for someone who actually has the compute.

He’ll probably spend a lot of time trying to verify whether that compute exists. There’s a big problem now. I call it phantom power, so it’s probably worth spending 30 seconds on.

Many of these GPU clusters are presented in such a way that the developer of the project—which could be a data center, a hyperscaler, or a neocloud—wants to raise debt capital to buy the equipment. To do this, the lender tells them, “Of course, we’ll do it, provided that you find someone to sign an agreement with you.”

This is called a buyout agreement: someone who will sign a long-term lease on this equipment for the next 3 years, even though the equipment doesn’t exist yet. They have to look for an imaginary cluster that only comes into existence when a customer signs up and says, “I’ll buy it for the next 3 years.”

Then they take that agreement to a lender and say, “I have a buyout agreement. Give me the capital.” They unlock the debt capital, spend it on Dell, Supermicro, and Nvidia, buy the hardware, store it in a data center somewhere, and then make it available.

Why? It takes 4 to 16 weeks to complete an order for a medium-sized cluster, if not a little longer these days. So that’s what he spends his time on. This is what he spends 40% of his time on.

It works primarily through email, Slack, and WhatsApp—something like a request for quotation, or RFQ—where he has to separate the real capacity from the fake capacity and then get a competitive offer.

Do you think that, at his scale, with Instinct being very popular and having just raised a lot of money, he’ll turn to brokers? Does he need to go to brokers, or can he just go directly to computing-service providers?

I’m sure it combines all of the above. He probably knows, and we know this from our own experience, because even in the process of building an exchange for computing derivatives, we naturally have computing sellers coming to us looking for buyers and computing buyers coming to us looking for sellers in the spot-computing market.

We have relationships with several non-Aklar companies that have capacity or are building capacity, so we know we can act as a broker in this market. He’ll approach brokers and maybe look for suppliers he has already gotten compute from in the past.

As Andy said, he’ll turn to his friends who run other AI companies and have their own verified sources. He’ll probably put it out on Twitter, saying, “Hey, I’m looking for compute. I just raised $1 billion. Contact me if you have these GPUs next week.” Whatever means are necessary, it will come to him.

Again, it’s such a strange market because there’s such a large volume of capital expenditure and such a large volume of debt being issued to finance these things. Yet there is no standardized market for either spot computing or the actual buying and selling of computing, nor any instrument that could hedge or lock in that price, including things like insurance.

The market for cluster provisioning, providing residual value for chips, or establishing a lower bound on the cost of computing is extremely volatile.

4. How Much Compute Should You Buy?

Good. Staying with this intuitive story a little longer, in the podcast, the founder of Manus said that they’re growing at 10% per day—let’s say 5% to 8% per day, but maybe it’s actually 10% per day. Those numbers add up very quickly if he stays at that pace.

He said, essentially, “How much leverage do I need to increase? Should I buy computing resources? Should I buy enough computing resources to serve 100 million users, or 1 billion users?”

I don’t know how many users they have today, but if he buys enough compute resources just to serve his current users, he’ll run out of compute in a week. But if he buys enough to make it to, I don’t know, the third quarter or something, he’ll actually be done in 3 weeks.

Since their growth rate is so extreme, how would you think about this problem if you were in his shoes? How far out should you buy? I’ll turn it over to Andy to answer that question, and I’ll answer the question of what should be done in this particular market.

Yes. He was stuck between a rock and a hard place. The reason is that it’s not an easy decision. He’ll have to compromise between several options.

Either he moves fast and compromises on the price of the compute he gets, or he compromises on the price by 10% to 20%, or by 2 or 3 times—probably 2 times. You can compromise up to 2 times. We’ve seen this in transactions from some very large inference providers that are growing very rapidly.

For them, the opportunity cost of lost revenue outweighs some degradation in unit economics. He has to get the compute very quickly, because otherwise it’s a revenue wipeout.

He knows what his demand will be in the short and medium term, maybe in the near future, but not in the long term. So if he wants a good price and buys a lot of compute from the other side, whoever is selling it to him will say, “Hey, take a 5-year commitment or a 3-year commitment.”

That’s the problem. There are several dimensions you need to consider when buying compute: the length of the commitment and the volume of compute you’re buying.

For such a term of commitment, the inflexibility here is something that can be untangled or broken if you have a working futures market.

Good. And enter Brett, stage left.

Yeah, maybe another step back is that anyone who buys compute resources in these absolutely astronomical quantities—or sells compute resources because they're building data centers and trying to get contracts with buyers—has to combine 2 different things. One is buying or selling a product, and the other is managing all the risks. In all traditional commodity markets, these 2 things are separate.

Generally speaking, when Hershey buys cocoa beans to make chocolate bars, they don't have to retain all the risk of what happens with the price of cocoa beans going up and down or with supply. In general, this is something they can pass on to the other party. Obviously, there are financial companies whose entire job is to model risk, store that risk, hedge it, and create a diversified portfolio of risks. That's their whole reason for existing, while producers and consumers of goods do not have to do so.

The problem with compute is that there's no way for a company like Instinct, or any other company that's grown so quickly in the last week or 2 weeks and is now thinking about the next 5 years of compute it's buying, to think, “What if, in 1 year's time, nobody cares about AI anymore? What happens if, a year later, I discover that I've overbought compute resources by a very large amount? What happens if, a year later, I discover that I've under-purchased compute resources by a very large amount?”

No matter which of these cases happens, I can't say, “No matter what, I can sell the last part of this contract if I don't need all of it,” or, “No matter what, I'm guaranteed a floor price or a ceiling price.” That's why we have financialization. It's not just a vehicle for speculation, gambling, or betting on the prices of things. It's a way of transferring risk and addressing things like duration mismatches.

I have a specific need right now, and it might be related to what my compute resources are going to be over the next year. Someone really wants to sell compute resources in 5 years. There is a complete mismatch in duration, and who will take the risk from the second to the fifth year? There are probably many firms that would be happy to hold that risk on their own behalf if there were a liquid derivatives market or even an over-the-counter derivatives market, which is obviously what we're trying to do.

I have a stupid question about hedging. Do Fortune 500 companies do this on their own? Does Hershey, for example, hedge all of this in-house with the help of a risk department and trading department? Do they send it to the trading department at Goldman Sachs? Do they send it to a specialized firm?

Both. This is exactly the combination. The canonical example is when airlines used to hedge their own fuel prices through the crude oil markets, and then they stopped doing that, which was a bad move given what's happened to oil over the last year. There were also a lot of internal trading divisions in big energy companies and commodity companies.

They either do it themselves, outsource the risk-management function to a large bank or FCM that will handle it for them, or do both. There are internal analysts who figure out how to properly manage treasury, from interest rates to commodities, and then they work with the trading department to figure out how to make or close those deals in the most efficient way.

Do you think a startup should hedge risks? You both build startups. Shouldn't you just put everything on the line if you're a startup founder? For example, if AI assistants don't take off, then there's no incentive.

I think it's not about being as protected as possible or as vulnerable as possible. Here's a very simple example. Let's say you're going to spend the next 2 months looking for this compute, and you're just worried that by the time you finally find the person who will sell you the compute, the price will have doubled. So fix the price. You can lock in a price right now using a futures contract.

It's not about being as hedged as possible or avoiding risks. You simply know that there is a price movement in the market, and you can hedge against it and fix the price today before you can find a supply contract for yourself. This is just one example of how this can help even startups.

5. Why Are Compute Prices Rising?

Good. Let's look at what's happening in the compute markets themselves with pricing. I have a few questions, but maybe the first thing is: what's happening with the pricing of compute over the last couple of years? Obviously, we know that demand has increased. Has it stabilized? Has it decreased at all?

Let's back up a bit. One interesting thing is that, before I get to the price of compute, there's something interesting that's happened to compute margins. If you look at the reports of hyperscalers, the margins they get on their GPU compute have shrunk. This is an encouraging signal that this is indeed a product. In commodity goods, you don't have the luxury of nice, big profits, so those profits are shrinking.

There's a narrative—and it's an absolutely true narrative, by the way—that this year, or since the very end of 2025, the price of compute technology has gone up. We see this in the prices we track. On average, the price of compute for every type of chip and every type of GPU for data centers has gone up. The reason is that there's high demand and not enough supply to meet that demand in a timely manner, so demand is growing a little faster than supply. That drives prices up.

You don't think about compute prices the way you think about, “What's the price of crude oil right now?” You can check what WTI is showing or what Brent is today, and prices around the world will be trading with a very tight spread. Today, in compute, you can get 2 very different quotes for practically the same thing, but the spread is very large.

The reason it's so large is simply because the market is very opaque and poorly informed about compute pricing. It's essentially an illiquid market. There is no single global standard price that everyone looks at and says, “Here's the price. Let me see how much the purchase price of my compute differs from this price, or how differently I sell it.”

We're trying to change that. My company publishes price indices for H100, H200, B200, and B300 compute on Bloomberg, oddly enough, so people can get an actual reference price for what it costs to buy compute today. But that average price is very different from the spreads that are being realized.

We expect this market to become much smarter and more efficient. Those spreads will tighten, and you'll have market-maker input and more efficient price discovery on exchanges. All of this will contribute to the compression of these spreads, or rather, the pricing inefficiencies that exist now.

So, probably a stupid question: you can't determine one price for compute, right? There is no single compute price. There is a compute price for B200 or H100. You need to set a price for each of them. Is that correct?

Of course. Crude oil brokers will privately negotiate a price. Now there are so-called price-reporting agencies. They take all the crude oil transactions—these are technically midstream companies—and check the prices at which those transactions were completed that day. They combine them into what you see as the price of crude oil, whether that's WTI or Brent. That's the price that actually exists.

Somewhere, there is a private negotiation between supply and demand, and 2 people agree on a price. There is a price-reporting agency that collects this price, and it becomes an index and a reference price. That's what is happening in compute right now. People can negotiate privately, and we try to report how those private negotiations ended and publish a reference price.

So how do you get this data? We're trying to think about where compute operations are being performed. There is compute that you can get on demand through AWS. There are prices listed online. You could scrape websites and take data from people's APIs.

Overall, that's useless. The reason it's useless is that most compute is not purchased on a website. Most compute is privately agreed upon through large contract agreements.

So where does this price exist? We thought it was in the minds of the person who bought it, the person who sold it, and the billing system that recorded it. We have exclusive agreements with the biggest neoclouds to have exclusive access to their billing systems, where we get the real, agreed-upon transaction price. We take all these prices, compile them, and then publish an index.

And why would they give it to you?

We pay them a lot of money.

You pay them money. Good. And then you monetize on the back end by charging an exchange?

The exchange pays us.

Us?

Yes. It's very typical for derivatives exchanges. When you build a futures contract, option, or perpetual based on another company's index data, you want to split the economics between the index provider and the exchange.

6. Can Architect Take On CME?

There is a combination of a minimum commission plus a revenue share. Andrawes will benefit from the growth in volume and open interest on our exchange, and that's how this partnership works. Again, this is very typical if you buy index data from S&P, MarketVector Indexes, or some other index company. This is a very standard relationship.

For anyone listening, since we skipped all the introductions at the beginning, Brett, you build an exchange, and Andrawes, you're building an index business. What does the computing market look like today? We have examined the players in the computing markets. What is the computing space like with hyperscalers, neoclouds, lenders, and all that stuff? What does the market look like for the markets you guys are building?

When I think about futures trading, CME, ICE, and people like that come to mind. So I'm curious how you see them being your direct competitors.

Architect acquired a designated contract market license a few months ago—a DCM, which is the CFTC's designation for the ability to organize an exchange of futures and options on commodities and related products. We are in a small group of potential players who can offer futures and options to U.S. clients.

The market for derivatives on computing technologies does not yet exist. This is what we are trying to create, and this is what we have various proposals to the CFTC to launch. There is a public comment period going on right now that the CFTC has announced to get comments on how to create a reliable index, how to prevent manipulation, and how to actually track the price of computing technology.

Which is better: cash-settled futures or physically settled futures? All the issues related to the creation of a completely new derivatives market are being raised now, and hopefully the process will be completed soon. After that, we will be able to launch these products for the first time.

So, is CME your main competitor here?

Yes, absolutely.

And why would anyone trade on Architect and not CME?

There are several reasons, everything from the contract design to the user experience—how you actually trade and sign up. We could talk about a few examples.

This is a better platform. This is the best technology platform. One hundred percent. Good. Let's talk about a few of them.

First of all, one of the reasons we work with Andrawes as our index provider is because it's really important what the actual index tracks in terms of creating a good future. If you have a future that tracks the price but doesn't actually provide a good hedging product—that is, it doesn't hedge enough of the variance in the prices of your actual computing systems that you see in the spot market every day—then it's not going to be a useful future, and nobody's going to want to trade it.

For example, there are several futures offerings that track the on-demand rate for computing systems. This is, as Andrawes said, not how people trade computing systems in the real market. They don't go out and buy 7 GPUs for 2 days. The market is not on demand: they buy 3-year contracts for 10,000 or 5,000 GPUs. So this index that Andrawes and his team created is much more similar to actual contract rates for computing. That's the first thing.

Secondly, you can't go to CMEGroup.com, register a trading account, enter your information, and then start trading through their API. It's not that they don't provide it—they do not distribute their own products. They work through intermediaries, brokers, and futures commission merchants. You have to pay a lot of fees for the data itself, and it is often not cloud-based. The same thing applies to the API.

We built our exchange much more in the image of a modern exchange, similar to how crypto exchanges were built, where you can register an account directly and start trading almost immediately after signing up. You can use a GUI. You can use the API. There's no advantage for a big player over a small one. We provide open access to everyone in the cloud, and these are the things that we think are necessary to create completely new products in the modern era.

The last thing I'll say is that, especially for commercial-computing hedgers, who we hope to make among the first customers of these products, these are not people who have traded futures before in their lives. A lot of them don't even know what futures are, but they intuitively understand the idea of futures because they live and breathe these sales contracts. They understand what it means to fix prices.

7. How Do You Hedge Compute?

These are not customers with existing loyalty to a particular exchange. They will be looking for people who live and breathe computing and artificial intelligence every day, can speak their language, and can engage them with a product that specializes in computing. No Silicon Valley AI firm wants to look through soybeans and crude oil to get to computing on the front end. They want something tailored specifically to their needs.

Speaking of adapting to those needs, what does someone actually come up to you and say? Do they say, “Hey, I just got this big contract. Can you show me what it actually looks like?”

Yes, of course. Here's an example. Someone bought, say, a 3-year contract for the availability of computing resources and sold someone an 18-month contract for compute. A data center—a neocloud—has locked in someone who bought 18 months of compute resources, and let's say they bought 100% of the capacity in their data center.

Now let's say they have the option to renew this contract in 18 months at whatever the current rate is. You don't know what that rate will be in 18 months. Let's say today it's $5 an hour for a GPU. What if in 18 months it's $2 per hour per GPU, and you don't have the ability to hedge that price and lock it in now?

Here's what they come to us with and say: “How can I protect myself against the risk of prices going down? When I sell this contract again in 18 months, I won't even be able to pay my debt anymore because it's just not that valuable. Can I set a minimum price for computing power at $4? Can I set a minimum price for computing power at $3.50?”

It will be some combination of futures or options to lock in that price.

I remember when you launched Architect. If I remember correctly, the thesis was this: people come to trade, essentially.

That's true.

Now you're saying that futures and options are more like a standard future. Or is it something like a perp, which I think most people listening to this are familiar with, as opposed to a standard future that expires? So why not a perp?

With a perpetual contract, it is impossible to fix a specific cash flow tied to a specific term in the future. Perpetual contracts are great when you need perpetual exposure to the price of something. For example, single-stock perpetuals were a great product because most people don't say, “I really want to hedge my March exposure to Apple.” People just say, “I want to be long Apple, and I want to be long Apple until I don't want to be long Apple anymore.”

That doesn't apply to computing, where people say, “I have a very specific contract that ends on a certain date,” because it's completely based on a contract. The same goes for interest rates. I think perpetual interest-rate contracts will be relevant, and we have some proposals for that as well. But mostly, people have loans, and loans have a specific term, so people need to hedge the duration of that particular loan.

This is why derivatives in the form of standard futures or standard options are so important. That's why our exchange chooses the right tool for the job. Is this a perpetual contract? Is this a future or an option? Or some combination of all 3?

What other deals are there that people are making? I'm trying to imagine what some of the other ones look like.

There is an example. It's really very simple. There was an AI company that came in and said, “Hey, we want 4 nodes.” A node is a group of 8 GPUs. They said, “We want 4 H200 nodes per year, but in 60 days it may turn out that we actually need 8 instead of 4, and they all have to be in the same data center.”

How do you give them that flexibility? People will only want to sell them a 4-node annual contract or an 8-node annual contract. We solved this problem for them by structuring a 1-year contract for 4 nodes, a put option on 4 nodes expiring in 60 days, and a call option on 8 nodes at the same time.

This gave them the flexibility they needed not to commit in advance to something where there was a certain level of uncertainty. They paid a premium, and that premium was worth it to them so they wouldn't spend the extra money on 8 extra nodes if it turned out that they didn't need them. It was a way to manage a risk that made sense.

That makes sense. They had a risk, and we were able to move it away from them and give it to someone else who was willing to take that risk.

Yes.

8. The Risks Behind GPU Lending

Okay, now I understand. It became clear to me. Can you guys do one thing we haven't done? At the beginning, you talked about players and lenders, and we didn't talk too much about how they're borrowing and lending GPUs now.

What happened in this market?

This is what we hear from all the different clients in this area. The lenders themselves will also become derivatives customers, because one of the ways they hedge the fixed-income contracts they issue is through derivatives.

Obviously, people want to borrow as much money as possible to build as much capacity as possible and try to meet the growing demand for computing technology. The problem, of course, is whether you're willing to guarantee anyone who comes up with the idea, “I want to buy some computing technology.” And, of course, it is very difficult.

First, most people don't have years of history to prove their creditworthiness. Second, people don't know what GPUs are going to cost in a few years, whether the capacity is going to mean the same thing in a couple of years, or whether the algorithms are going to be different. So unless you're an investment-grade company, good luck getting a loan from a lender at less than 20% or 25%, and so on.

We hear that sometimes people can get these loans, but they are at exorbitant interest rates, or they just can't get loans at all. People don't want to guarantee them. So we start to see a few different things.

First, we see companies emerging that do things like subprime lending, which in itself is perhaps important for getting these small companies off the ground, but it's also dangerous, and it reminds us very much of the subprime mortgage problems of about 20 years ago. At the same time, we're starting to see lenders trying to figure out how they can secure their contracts, either through actual insurance or eventually through a derivative product, to reduce some of the risk and be able to lend more.

Finally, we're seeing a lot of other companies entering this market, especially if you think about it as a parallel to the private-lending boom. Large-cap trading firms that understand computing and maybe can implement it themselves, along with many existing large private-credit issuers, are entering this market and combining venture capital and debt to finance these transactions. This is what we see.

Okay, Andy, what about you? And one more question: maybe you can take this if you want, Andy, but who are the biggest players in the computing-lending markets today? Are these banks or private-credit funds?

There are several layers to the world of debt. There are big, very big banks and financial institutions like Apollo, KKR, top-tier Blackstone, Brookfield, and so on. There are maybe about 10 such firms, and then you can add JPMorgan and Goldman Sachs to them.

These firms provide loans to large operators whose clients are very large and very creditworthy, so-called offtakers. So it's an operator like CoreWeave, for example, that sells a contract to Microsoft or Meta, all of whom have incredible balance sheets and creditworthiness. This is one segment of the lending world.

In this world, nobody considers loans smaller than $1 billion, $2 billion, or $3 billion. This is the starting amount. Now there's a sort of middle market of computing, and then there's a very long tail of computing, where there are very different types of credit firms.

That's where some of these underwriting shops come in, like Blue Owl, for example. They underwrite these loans for GPU data centers—GPU deals, mostly—ranging from, I don't know, hundreds of millions to $1 billion.

And then there's a very long tail of asset-leasing companies making loans to everyone else. Therefore, the risk profile in each of these segments is very different. The type of risk is very different, the type of operator is very different, and what is fundamentally underwritten is different.

There is one thing that is happening right now. I think a lot of people are wondering why we didn't hear about this financialization of computing 9 months ago, and suddenly it's so trendy now. There is one thing that seemed to catalyze this moment.

I don't know if your viewers saw this, but there's a chart that was circulating that showed the free cash flows of hyperscalers. It showed free cash flow dropping to zero or even becoming negative. The idea is that hyperscalers' ability to generate free cash flow is being affected by how much they spend on AI development, and that capital expenditure on AI development has caused it to drop to zero or become negative.

So the lenders who guaranteed Google's creditworthiness guaranteed free cash flow. Now there is no free cash flow. Suddenly, there's a big shift happening where, instead of capital markets underwriting Google's ability to make money from advertising and then pay the GPU bills, capital markets are thinking: How do we underwrite computing technology itself as an asset class?

This raises several questions: How do we evaluate this? How do we hedge this? How can we insure this? Previously, it was believed that all these risks would be completely absorbed by Google or Amazon. But not anymore, because the economy is now much larger than any single player can absorb on their own.

Hence financialization, if you think about the role of capital markets in transferring risks. That's why this is happening now.

These hyperscaler free cash flows going to zero or becoming negative are a very real issue right now. It just happened.

Yes. Good. Brett, are there any concerns? It seems like you spent some time in Washington. Are there any concerns that if you create markets around something, it will mean that the markets will become more manipulated?

Of course, there are concerns. I think historically, when there was publicly available data and tools to track something, it helped normalize and smooth out risk over time. In all cases where there was a real financial crisis, it happened because everything that was happening was completely opaque and over-the-counter, because there was no market.

That's right. So, do you know what happened? What if there were publicly disclosed swap markets that represented something like credit default swaps and all these crazy financial instruments that were created during the financial crisis, but not during the implosion?

I think this is precisely one of the main questions being addressed in the current CFTC comment period: How do we ensure that indices are not subject to manipulation, that they reflect the fair price of computing, and that they are properly administered and monitored?

But we believe that this will only help the market. I think another turning point for me was Nvidia's announcement of its 25% discount on these contracts. What Nvidia essentially meant was that since there's no forward compute curve that tells everyone how much these things cost, we'll build one for everyone and set a floor price on it in the meantime.

9. What Really Constrains The AI Market?

This really shouldn't be Nvidia's job. This should be the job of the market. So you're absolutely right that this is a problem, but we think this is the wrong way to look at the market.

10. What Is Everyone Missing About Compute?

Yes. Good. Guys, maybe I want to take a little time while we think about wrapping this up with what you see that other people don't see yet. Andy, you have data that no one else in the world has. Maybe let's talk about where exactly the main constraint in the market is.

I think if you ask people on Twitter, or anyone who's not really involved, maybe a third would say we're computationally constrained, a third would say we're memory constrained, and a third would say the whole network would grind to a halt. What do you see, and what is the real limitation over the next few years?

Bringing the power of the GPUs online depends on 3 or 4 things. The first is access to capital. The second is access to power inside the data center. The third is access to equipment.

Depending on the scale you're working at, one of these three paths becomes the critical path, or the bottleneck. In a sense, it's true that all three things are bottlenecks. But by the definition of a bottleneck, only one of them can be true at a time.

There's a hardware bottleneck, and there's a memory bottleneck. If you talk to Supermicro or Dell, or any of the GPU hardware vendors right now, you're going to have a very real conversation of this sort: “Hey, we have 2 configurations of this equipment, with a large amount of memory and with a smaller amount of memory.” The lower-memory configuration will be delivered in a much shorter period of time.

There's a second type of conversation where it's something like: If you need this specialized network fabric called InfiniBand, if you want to build training clusters, as opposed to using some other general-purpose network fabric, then it's going to take a lot more time and much longer lead times to acquire the specialized fabric.

Now, if you don't have these restrictions, your access to electricity could become a limitation. Access to electricity is not the same depending on scale. If you're trying to build gigawatt-scale campuses, interconnection to the grid is a multiyear nightmare.

People are resorting to building gas turbines onsite and building their own, essentially, natural-gas power plants behind the meter.

This is another type of restriction. If you’re operating on a small scale, the amount of data center capacity that’s ready to go and running in the United States, and then in other parts of the world—but it’s particularly acute in the United States—exists for small deployments but not for large ones. So we see even the biggest players—hyperscale companies, OpenAI, AI-powered companies, and so on—moving downmarket. You couldn’t call them before unless you had 100 megawatts. Now they’ll pick up the phone if you have 30 megawatts, and it’s getting smaller and smaller because of all this extra capacity, which, by the way, is related to the fact that a lot of this demand is based on inference. But that’s another one of these phenomena that we’re seeing.

Your ability to find data center capacity in the United States is interesting. I have a funny story. There’s a class of people—you can think of them as Miami-style real estate brokers—who happen to know a bunch of data center vendors with a megawatt here and 5 megawatts there for a small deployment. These guys take a commission for introducing you to their buddy who runs a data center that has some space. They’re making crazy money right now by essentially hooking up someone who wants to host some equipment to a data center in some random location.

This is the second limitation. The third constraint, probably the most systemically important, is capital. There’s a kind of paradox: There is currently huge demand for computing technologies, but at the same time, lenders are hesitant to underwrite computing technologies themselves. Their hesitation means that these projects take longer to underwrite, and therefore to finance, and therefore to deploy, and so on.

These 2 things at the same time don’t make sense to me, because if lenders knew about the state of demand, they wouldn’t have to be so hesitant. They would know that there’s a market for this compute, and this stems from their inability to guarantee, price, and hedge compute. So I have a little analogy, which is this: For these lenders, a buyout agreement, which kind of forces the person they’re lending to to enter into a very long-term contract with the client, is a kind of hedge. It’s something like, “I need the whole project committed for the next 3 years.”

It’s kind of like if you have a car. It’s like a handbrake. What’s funny is that if you have a car and you only have a gas pedal and a handbrake, you’re not going to go very fast. If you have a small brake, you can go a lot faster because you can stop pretty quickly. That’s what hedging provides. Hedging instead of a full buyout. If you provide hedging, they will feel much more confident.

By that I mean that lenders and capital markets in general will feel much more comfortable making loans because they know they have a place to hedge. This is the third and largest systematic constraint, which is a systemic constraint: access to capital. That’s it.

From our position, I want to give a very different answer to what we see as limitations. Regardless of the limitations, this market is so powerful, so important, and so in demand that it will find a way around them. This is perhaps a naive answer, but I actually hardly believe in any of these restrictions.

Look at what happened in China. Do you think people just find their way around a constraint? 100%. Look at what happened in China. They were told that they couldn’t have access to any of the latest and greatest chips. What they did instead was create extremely cost-effective algorithms that can run on older-generation hardware and are now almost completely competitive with cutting-edge models. They understood it.

Whatever the constraint is—whether it’s not enough capacity, politics forcing people to put a moratorium on building new data centers, or not enough credit or capital—it will somehow find a way. Because if companies are just greedy for tokens and they need them, logically, they will solve it somehow. Maybe if all 3 of these things are limited, people will just create better algorithms that can run more cheaply on the available electricity.

So, if I had to rank the limitations, let’s look at power, chips, memory, and local permitting. How would you rank them?

It’s impossible to rank them because this set of constraints is different in each geographical location, and the market is becoming increasingly global. We worked with neoclouds that said, “Oh, we have a new data center in the Philippines. We have one in Iceland. We have American companies that are now very happy to do these weekend contracts from locations outside the U.S., even though they could never do that before.”

So suddenly, there’s no single answer to the global ordering of all these constraints.

How do you think the U.S. midterm elections will impact data centers, compute, and these markets? If data centers become this huge political topic that leads to more regulation, is that something that will perhaps put— you talked about a pause and a break—or is that a moment of pause?

The simple answer to that is that it’s the end of September, and the elections are in November. There’s not much time left. We have 5 weeks. When this is over, I think the floodgates will open again.

I think we’ve temporarily stalled because nobody knows how the general public feels about these big data center contracts when they go to the polls. Once this is over, we’ll get back to business.

Do you have an opinion, Andy? This is a bit of a spicy question that I think might be a good closing point.

I think the average NIMBY-type data center opponent will finally be very happy to trade compute futures. You’re bothering me by building a data center in my backyard, but I know it’s profitable, and I want you to know that I’m not stupid. I want to have some involvement in this activity—not in my yard, but maybe in my brokerage account.

I don’t disagree with you. As we think about wrapping this up, maybe we could ask one final question that seems very obvious to you. It’s a very broad, general question and may refer to the markets for compute, compute technology, or artificial intelligence in general. From your perspective, what seems very obvious to you that you think the general public is currently overlooking?

I think what seems obvious to me is that the pricing power and dominance of cutting-edge labs like Anthropic and OpenAI are bound to decrease. The availability, power, and efficiency of the remaining models, like open-source models and so on, are just going to continue to increase.

Better tools will come along, and we’re launching something very soon that will help with that. But there will be greater accessibility for people to draw conclusions and apply them to all sorts of models that don’t necessarily limit them to just a few of the biggest companies.

I don’t think so. I think in niche AI circles, on X, and in places like that, people are certainly talking about open-source models and how good they are. But I don’t think the general public—the several billion users of artificial intelligence—has a real understanding of this.

There’s one thing that I think everyone has heard but not understood. There’s this banal saying that compute is the new oil. But in reality, it’s not that trivial. I want to put this number in perspective.

Physical trading of crude oil is a market of about $2.8 trillion, possibly $3 trillion. Physical trading of compute is approximately $1 trillion. We’re actually halfway there.

Now, if you consider that this is the order of magnitude of physical trading of compute, and then look at other markets that have both a physical market and a derivatives market, the derivatives market for compute on a notional basis will be several orders of magnitude larger—perhaps 1.5 orders of magnitude larger.

Crude oil, for example, has a paper market, or a crude oil derivatives market, that’s 40 times larger than the physical trading of crude oil. I expect the same thing to happen with compute, and that’s at its current scale.

What if compute grew another 10 times? We have the potential for 2, maybe 3, orders of magnitude of growth, given the notional transaction volume in compute. People just don’t realize how big it is or how quickly we’ve achieved this. How fast.

Okay, guys, thank you for that. Thank you. Congratulations to everyone.

Thank you very much. Let’s do it again in a year and see where the compute markets are.

Thank you, and see you later.