# EP55 Token Prices Dropping on Model Shifts with Carmen Li, CEO of Silicon Data & Compute Exchange Pitch The PM · 2026-10-16 · 57 min · https://www.youtube.com/watch?v=hwYMbEAwaxQ ## Transcript Carmen Li We see a lot of people shifting from premium models to cheaper models post token-mixing era. That’s why you see token prices start dropping. Most people think they can predict the future. If everyone thinks there’s going to be no cycle, or one cycle and that’s it, I keep telling people there are cycles. ### From energy options market maker at DRW to Bloomberg Data, and why AI startups' swinging GPU costs made compute an industry-wide hedging problem Doug The only thing I know about the future curve is that it’s always wrong. Carmen Li The curve is the market consensus view today. If you map our token indices on top of the bar chart, you’ll see something really interesting: they’re almost a leading indicator. When our token indices ramp up, that’s when you see revenue dropping and ramping up. Just like American Airlines can’t run an airline without hedging its oil prices. It’s not American Airlines’ job to speculate on oil prices. It’s their job to run an airline company. Similarly, it’s not AI companies’ job to speculate on GPU cycles. You want to lock in your cost of production, which is compute, and focus on products. Silicon Data is the referee layer for compute. The flip side of that is Compute Exchange. It’s the liquidity layer for compute resources. Doug Hello and welcome to Pitch the PM. My name is Doug Garber. After a decade at Citadel and Millennium, I interview top PMs on their highest conviction ideas. Be sure to subscribe to our Substack for more research updates. Today, I’m here with Carmen Li, the CEO of Silicon Data and Compute Exchange. Welcome. Carmen Li Thank you. It’s good to be here. Doug I’m glad you’re here. I know you’re very busy on the media circuit. I’d love to hear more about both your companies and how you came to run them. ### Where to invest in the AI stack, lessons from the internet era and why "AI is a gigantic baby" Carmen Li You’re right about the media circuit. I never realized how much media engagement I’d need to do, though we were just talking about this. I should have expected it. We’re essentially establishing a brand-new asset class: compute. By definition, compute is gigantic in scale. Everything we do will be impactful in terms of transparency and getting people the right products and benchmarks. We need to do a lot of education for the industry. How did I get into compute? I started my career as an energy options market maker in Chicago. I worked for DRW Trading. In a way, compute is another form of commodity, another form of resource, just like oil and gas. I started my career in traditional commodities, in energy. That was really fun, and the experience gave me strong intuition about market making, what the index volatility is, and how to manage risk. After that, I moved to different banks. My last job was at Bloomberg Data, where I managed global data integrations for Bloomberg. Any data leaving the Bloomberg Terminal and going into a third-party ecosystem, including back offices, cloud, and AI applications, was governed by me and my team. That included technical integrations, go-to-market, compliance, and everything else. In 2024, after the ChatGPT moment, there were already a lot of AI companies. They were SaaS companies creating solutions and services, some of them aimed at financial services firms. I was excited for them. They were integrating, or looking to integrate, with Bloomberg Data, and I thought, “Great, you’re a young company. You’re doing really well. You’ve got big-name clients. That’s amazing.” ### How DRW founder Don Wilson pre-seeded Silicon Data, and building products with no product or sales team They were complaining that even though they looked like they were doing well in terms of client numbers, their margins could actually be negative. Think about SaaS: the customer pays them monthly fixed revenue. But at the time, in 2024, there were two new clouds, and most of the time they had to rely on hyperscaler on-demand GPU rates. They were paying $3 per GPU hour, sometimes up to $9 or $16. It was a big swing every single day. The market contracts they brought on didn’t really help them control margins. There was no scale, and they couldn’t really sign long-term contracts because they didn’t know what the future would look like. They were startups. In my mind, this was obviously an industry issue, not an individual company’s issue. It’s an industry issue. Just like American Airlines can’t run an airline without hedging its oil prices, it’s not American Airlines’ job to speculate on oil prices. It’s their job to run an airline company. Similarly, it’s not AI companies’ job to speculate on GPU cycles. You want to lock in your cost of production, which is compute, GPUs, and tokens, and focus on products. At that moment in 2024, I realized that if the whole industry grew at the speed it was supposed to, it would absolutely need all the traditional risk management tools, benchmarks, and platforms that every other asset class enjoyed. To me, it was a no-brainer. I left Bloomberg right away and started thinking about data. ### Two sides of the coin: Silicon Data as "the Bloomberg for compute" and Compute Exchange as "the Nasdaq for compute" Doug What was the first thing you did to take Silicon Data from zero to one? You’re affiliated with your former employer. What’s the relationship there, and how did you get your first customer? ### Why token prices are falling: workflows moving to open-source and open-weight models Carmen Li It’s really fascinating. Don Wilson is the former CEO of DRW Trading. When I was at DRW, I was a junior trader. I was definitely not one of the top-tier traders by any means, but I worked on a few initiatives at DRW back then. He knew me from those initiatives. We engaged again, and he saw the world even bigger than I did. He saw it not only from an index-linked perspective, but also through spot market auction processes and compute marketplaces. We got connected, and for us it was a no-brainer. He said, “Great. Why don’t you do Bloomberg, and we will seed you, and then the rest is history.” This is really interesting: we’re in the process of hiring our first product manager for Silicon Data. For the past 2.5 years, there was no product person. We’re hiring our first salesperson, too. There hasn’t been a salesperson to this day. All the clients and products we have at Silicon Data were created by engineers and me, I guess, because everything is very intuitive. We’ve seen every other asset class, and I know exactly what an asset class would need: great benchmarking, indices, alternative data, and things that help people have transparency and efficiency and trust in the system. ### Licensing indices to CME for H100 and B200 compute futures, and four ETFs filed before the futures even trade Again, we’re not reinventing the wheel. We’re looking at other asset classes, following best practices, and creating relevant products. But now we’re reaching a scale where I can’t possibly do all that myself, so we’re hiring new people. We have about 30 people at Silicon Data and about 10 at Compute Exchange. I took over Compute Exchange, as you mentioned in the intro, late last year. Compute Exchange is fascinating. Think of them as two sides of a coin. Silicon Data is like Bloomberg for compute. We give you market intelligence and tools for risk management, but you can’t get anything real from Silicon Data. You can’t get GPUs or API access for tokens. It’s the referee layer for compute. The flip side is Compute Exchange, which is the liquidity layer for compute resources. You can procure long-term contracts and forward contracts on compute, GPUs, and tokens through Compute Exchange. You can also do secondary transactions on servers and physically get servers through Compute Exchange. We’re like Nasdaq for compute. So you’re managing two sides of a coin. Doug So one company sells the data and is the industry benchmark, and the other company sells the GPUs. Carmen Li The physical asset. Doug Exactly, the physical asset. The spot market; the other one is for hedging. Carmen Li That’s right. Doug What are the actual products you sell to customers at Silicon Data, and who do you sell to? Carmen Li We like to say licensing. The data world is about licensing. Think about S&P or Bloomberg index models. We license our indices to CME, one of our most famous products. CME is going to launch compute futures in the next few weeks based on our indices: H100 and B200 to start. We think that’s an excellent choice. H100 is one of the most popular chips for inference and some training, and B200 is the later-generation chip. It’s getting extremely popular globally for inference and training. Those are the first two chips on which futures will launch. I’m sure they’ll eventually launch options on top of the futures. ### What makes an index the source of truth: tracking supply-demand, resisting manipulation, staying neutral Another exciting product is a licensing agreement. It’s very similar to any other index. Usually, it’s a combination of different things. It can be a revenue-share model, for example. That’s the indices layer. We can also license indices for OTC transactions. If someone wants to do a bilateral trade with you, a block trade, or create any sort of vehicle based on my indices, such as swaps, there’s usually a fee involved for that particular creation. Carmen Li You can also license for ETF creation. There are already 4 ETFs filed on CME futures based on my indices. This is the first time I’ve seen this. Usually, ETFs are registered only after futures have been trading for a year, once you’ve seen how liquid both sides are and can decide what kind of product to create. I thought it was extremely ambitious and exciting for ETF issuers to see the potential of compute as a marketplace. There are 4 ETFs already filed. There’s even a 1x inverse one, which means you have to have a lot of confidence in the liquidity on both sides. We see a lot of activity. Obviously, there’s usually a one-time base fee for the index. Doug And these ETFs would try to track your— Carmen Li They use futures. Doug —your data as the benchmark. Carmen Li The ones already registered use CME futures based on my indices. There can also be ETFs based on some composition of assets using my GPU index. There are different mechanisms, depending on the angle and the exposures the asset manager wants to create. ### Who hedges first: neoclouds, hyperscalers and the financiers holding GPUs on their balance sheets Doug Creating the indices and having the name brand can be a great asset. If you look at what S&P is doing, those are very high-margin, highly leverageable assets. I think they’re talking about spinning off Capital IQ, a more competitive, lower-margin business. Who else is doing this? How do you define yourself as the source of truth for this market? How do you know you’re aggregating enough of what’s happening in the industry so people are confident that what you’re creating reflects what’s really going on? Carmen Li I love that question. You’re basically saying I should care about competition. Even if there were no competition, if my indices stopped reflecting the true demand and supply curve, no one would trade. With or without competition, your point is critical. I’m only as good as my last product to the world, which makes sense. ### Normalizing 50 B200 server configurations, and what earns a premium or a discount to the benchmark We have to continuously improve our methodology and underlying data sets to make sure our indices continue to reflect the demand and supply curve. There are a few golden rules for index providers. Number one, you have to make sure your index reflects the market’s demand and supply curve, which you can do in a few different ways. The most important thing is that it’s not manipulable. If someone else can manipulate it, or even we can, that creates a conflict of interest, and no one will trust the indices. This is all about trust. We’re the only neutral player. That means we don’t have any exposure to the assets we’re supposed to measure. You can’t participate and be the referee at the same time, for obvious reasons. Doug Are any of your investors also participating? I know DRW trades a lot in commodities, and they’re your investor. Carmen Li They’re not involved, by design. In terms of methodology and governance, I’m actually not on the index calculation committee. I’m too busy, for one thing, and there are much better people for the job: our head of research and our head of indices. I’m going to work with external academic advisers. Doug Do you allow those people to go on expert call networks? I know index arbitrage and passive index investors can try to front-run or guess what assets are going in and out of the S&P. I mean, that’s— Carmen Li Let’s take the B200 index, for example. Even though you and I might think the B200 is just one of the NVIDIA chips, it’s really a machine. The B200 is a GPU name, but every server can look different. They can have different CPUs, RAM, disks, and geolocations. They’re essentially different machines. If I have a different CPU from yours, or a different machine, I can’t compare apples to oranges. We have to normalize all the B200 servers. You can easily have 50 configurations, and we have to normalize them to one base case before we can calculate the index. You can’t just say, “You’re in the U.S. East, I’m in the U.S. West. Let’s compare.” You can’t do that. Normalization matters a lot. Doug Of course. What are some of the key variables that would put you at a premium to the B200 benchmark or at a discount? Better memory? What are some of the other technical variables where you’d say, “Mine’s better. I get a 20% premium to your benchmark,” or I get a discount for some reason? Carmen Li Sometimes it can be specific to the machine configuration. You might have better CPUs, better RAM, or better connectivity. You might have better throughput and perform better at your task, so you should price higher. ### Sizing the market: from 1–3 trillion in spot compute to a potential 30 trillion in derivatives notional Sometimes it’s region-based. Maybe you have the same machine, but more people want to use mine in U.S. cities for latency reasons, while yours is somewhere more remote. It can be based on geolocation, or on regional supply and demand. It can be machine-based, so you really have to take a look. We do calibrate. Once a relationship is formed, that doesn’t mean it will never change. We monitor every day to see how the basis deviates. Obviously, we normalize everything to a global rate. We want to make sure it makes sense for every other region. Doug How do you get access to everyone’s data on what they’re charging, their configurations, and their specs? Are all the hyperscalers giving you the data and already trusting you to make a benchmark? Carmen Li They give you the data through a licensing agreement. We call it licensing. There can be many arrangements. It can be exclusive rights, meaning I’m the only provider you’re allowed to license to for creating financial products. Or it can be a paying client relationship. Let’s say you’re a new cloud provider. You probably have 2 transactions every day—or maybe 2 transactions per week. I have 250,000 rows of data coming in every single day, so I still need your data. Don’t get me wrong, you might grow, but the impact you’ll have on my data collection probably won’t be that much. You might want my data more because you want to see how the market is trending and how the term structure changes. So there can also be a paying client relationship. Doug They give it to you, and you give them the average back, kind of like consensus. How big do you think this market is in terms of licensable revenue? For example, I’ve heard consensus data for earnings is worth $800 million in revenue. Then you have someone like Visible Alpha that gives you a little more granularity. That business might be $75 million, $100 million, or $150 million in revenue, maybe, because it’s a better version of what’s out there, but still small compared with the overall consensus everyone is benchmarked against. Carmen Li I think the market is growing. The comparison I usually make is with energy products. The more people have exposure to underlying assets, the greater the need for hedging and risk management. Obviously, you can always go long or short single names. Doug Is it the WTI or Brent market? Carmen Li They’re pretty sizable. Doug Probably the biggest one in the world is WTI, right? Carmen Li It’s pretty sizable. But they’re supplied by different exchanges. WTI and Brent are owned by different people. For compute, if you look at NVIDIA’s spot market notional transactions, they could be $1 trillion to $3 trillion annually today. Derivatives markets are usually 10 times the size of spot markets—10 to 15 times. But we think more conservatively and say 10 times. If this is a mature market, we’re looking at potentially $30 trillion in compute derivatives activity notional every year. Doug And what’s the value of the licensable data, then? Carmen Li We’re looking at $30 trillion traded for compute. Then you can compare that with the notional traded for oil and look at how much data sells for oil. If you trade compute futures heavily with CME, you probably want a lot more data for benchmarking and understanding the granularity, not just the global number. You might want regional exposure, and there are other things you might need. We think the market will grow as the financial services layer grows. Doug How much of the market do you think will be financial players, versus corporate players who are actually using it to smooth out operating costs, or creditors who want to lock things in? Carmen Li I think the first wave of people will definitely be the natural hedgers, especially people who are naturally long GPUs. Think about neoclouds or hyperscalers, or anyone with thousands of GPUs—2.5 million GPUs on the balance sheet. They have concentrated risk. If you’re a tech AI startup, your risk may be more distributed. The neoclouds have more concentrated exposure. It depends where the balance is and whose balance sheet the servers are sitting on. If they’re sitting on a financial services company’s balance sheet, then that company has the exposure and wants to hedge it. Usually, the financing terms can be 5 to 7 years. Let’s say I have GB300s and I’m financing the deal for 7 years, and at the same time I get really lucky and get a reserve contract for the next 3, 4, or 5 years. ### Michael Burry's depreciation bear case vs. the physical market: A100s still signing three-year contracts after six years Doug That’s really the duration we’re looking at. Either way, I have exposure to years 4, 5, 6, and 7. Those are the longer-term risks. That’s usually when the natural hedgers come in and say, “Hey, who can help me hedge out the rental rate for years 4, 5, and 6?” Is it deep enough in those out years for people to find liquidity? I know that happens in oil sometimes, where going out that far gets a little tricky. Carmen Li At this point, even for oil, which is a very mature market, it’s not there. CME futures only cover the first 36 months, so even if it’s liquid, it only helps with the first 3 years. Usually, those longer-term, longer-tail hedges happen through block trades and OTC trading. But you’re leveraging the forward curve from the futures market to better understand the curve for years 4, 5, and 6. Doug The utility of what you’re doing is that someone who maybe doesn’t have a contract but wants credit financing can lock that in. Then the creditor says, “I feel comfortable with this,” and that gives you the security to grow. Carmen Li 100%. There are a few things we can really help de-risk. Number one, to your point, is that financiers have a way to manage risk by hedging longer-term exposures, so they should feel more comfortable financing those deals. Number two is that ideally, they should have a good understanding of what those assets are worth. Right now, everyone is relying on 5-year straight-line accounting depreciation for servers. They assume everything depreciates to zero by the end of year 3, or in years 4 and 5. It’s straight-line accounting depreciation. ### Why older chips stay productive: cooling, floor weight, wait times and the second-hand market Doug Talk about that, because earlier this year there was a huge bear case that the accountants were wrong and companies were over-earning because they were extending the useful life of their servers. It’s a blended asset, not just the GPU, of course, and different companies are doing different things. Michael Burry was the poster child for this, and then all of a sudden everyone was saying, “Actually, the market’s going up. There’s a tick. The useful life is actually longer. Things are rolling off after 3 years at higher prices.” That’s a constant, and it’s the best they can do. But now we have what’s really going on in the physical market. What are you seeing today for the actual useful life of whatever the standard chip is? Is it an NVIDIA H200? Carmen Li I love this topic. One example is the A100, which has been around for 4 to 6 years at this point. We still see 3-year contracts being signed for A100s. I’m not saying every A100 lasts 9 years. For the ones that last, if you do a good job managing thermal degradation and maintenance and make sure they’re well maintained, some A100s have a minimum life expectancy of 6 to 9 years. If you sign 6-year and 3-year contracts, that gives you at least 9 years of earning cash. Doug Whoa. Doesn’t the life of the chip have more to do with supply and demand in the current earnings environment than actual useful life? You can always maintain it and extend its life. It’s just a question of whether it’s worth buying a new one that’s so much more productive and worth the energy cost and all of that. Carmen Li Yeah, totally. But there are also constraints. New architectures are a lot more efficient, and you can potentially serve bigger models because they have more memory than the A100s. But at the same time, any new cloud provider will tell you, “I have energy, I have space.” Their decision is which chips to get. Number one, the data center has to have the right cooling. If you can’t do direct liquid cooling, you can’t get the latest and greatest by design. The laws of physics prevent you from having the latest and greatest GPU if you can’t do that. Number two, is your whole floor strong enough to support the latest and greatest servers? They’re very heavy machines. Physics can prevent you from getting the latest and greatest sometimes. Let’s say you design something from scratch and get everything you need, including liquid cooling. Obviously, you can potentially get the latest and greatest. But if you have an older GPU or CPU data center somewhere and don’t have enough budget, or don’t want to wait—sometimes you have to wait a few months, if not longer, to get the latest and greatest; it can take the same 6 months—or you can get secondhand chips tomorrow. You may want those. Even the 4090s and 5090s, some of those consumer graphics cards, can be productive depending on the workflow. My point is, it depends on what you’re looking for. If you go to a hyperscaler today, you can rent L40s, which are ancient chips at this point. You still have to pay 20 to 40 cents per GPU per hour. Those are really old chips, so it depends on the workflow. ### 300+ neoclouds and why more data sources make an index harder to manipulate I think when people say that, it’s because Nvidia keeps pushing out new chips, and AMD keeps pushing out new chips. Obviously, they’re always better than the older chips, but that doesn’t mean people don’t use older chips anymore. Another nuance is that the residual value of the H800 has been decreasing because the chips are older. But compared with the accounting standard people use—5-year straight-line depreciation—they’re still hovering above that straight-line depreciation, and they still have a longer tail. Doug Is that a liquid market where you could sell them at those prices and book an accounting gain? Carmen Li It’s trading—it depends on what you mean by a market for A100. A100s are trading between $4,000 and $9,000 per GPU. Another interesting thing is that there are many layers to the cake, as Jensen put it. The value of the GPU layer can change based on the other layers. For example, every model is getting larger, so some GPUs can no longer support the large models we love today. But between the model layer and the GPU layer, there’s another layer: the software layer. You can think of optimization and quantization. If that layer gets really good, maybe you can deploy large models back onto older chips. Then all of a sudden, the supply and demand curve for older chips might change. There are many nuances and factors that can affect GPU pricing. Doug Where do you see the current supply and demand for GPU pricing? ### B200s at ~$5 per GPU hour, the implied forward curve and the shift into contango Carmen Li It depends. The majority of GPU consumption still happens in the hyperscaler ecosystem. However, neoclouds have become an extremely vibrant ecosystem. Two years ago, there were probably 2, 3, 4, or 5 neoclouds. Now there are over 300 neoclouds globally, and every single day there’s one new neocloud provider. Many are popping up in Australia. Many traditional telecom companies are becoming neoclouds, which is amazing because as consumers, we want to have options. We see a big landscape change toward the neocloud environment. That’s good: number one, we have more choices. Number two, from a data point of view, the more data sources we have, the harder it is for people to manipulate the data. Think about the hyperscaler industry. We have 3, 4, or 5 providers, maximum. You can argue that all 5 of them could manipulate the data together. But if the neocloud industry has 300 data sources globally, it would be very hard to manipulate 300 sources at the same time. That’s what we want as data providers: many data sources. As consumers, we want many neocloud providers as well. Doug Mm-hmm. Who’s the largest player? Who owns the most chips? Is it Microsoft or Google? Carmen Li Who owns the data? That’s a really good question. If you’re a provider, you don’t have to own the servers. The servers might be sitting on some financier’s balance sheet, but you’re operating the servers and selling or leasing them. You have the transaction data. You own the transaction data, for example—not whoever owns the balance sheet and the assets themselves. Obviously, you have the transaction data. I can license it from you. Or I can license it from your customer base, because they pay the bill. They know how much they pay you, and they also own their own bill—that transaction data. You can license it from them as well, and then you can verify it. There are different sources you can gather. Ideally, it’s the same data based on the same transaction, essentially. Doug Mm-hmm. So where do you see supply and demand and current prices for the most popular chip, call it a B200? Carmen Li It’s a very healthy range. Over the past 30 days, most chip prices have stabilized. You’re looking at around $5 for B200s on demand, at the global neocloud rate. What’s interesting is— Doug Is that a spot rate or a 3-year contract rate? Carmen Li Rates. Doug Spot rate? Carmen Li We do have an implied forward curve, because futures haven’t started trading yet. ### Why an index provider can't give forward guidance, or own the assets it measures So we create implied forward rates based on the reserved contracts we see transacted. This can be different once CME starts trading. It may differ from CME’s forward curve because we use a different methodology. Doug Is the curve in contango or backwardation? I’m very curious. Carmen Li Yes. Last year, it was a downward-sloping curve because prices were coming down. People expected prices to come down. The most common conversation I had last year was, “Carmen, why do people need to hedge? Isn’t the price always going to come down?” It’s fascinating to me that most people think they can predict the future. Everyone thinks there won’t be a cycle—there’ll be one cycle, and that’s it. I keep telling people there are cycles. Doug Well, the only thing I know about the futures curve is that it’s always wrong, right? It’ll go like this or like this, and the actual price will just— Carmen Li So— Doug —with the spot rate. Carmen Li You’re 100% right. The curve is a market consensus view today of what people think about the future, so it changes every day. Last year, there was a downward-sloping curve for every single chip. This year, everything went into contango. People weren’t getting any discount for signing long-term contracts because they expected prices to increase. Prices did come up quite a bit across all the chips, not just the B200s. Recently, we’ve seen the A100 and H100 curves flatten. B200s are in contango. I don’t know what the future holds, or whether it’ll flatten or go into contango again. Doug How do you create an implied futures curve? I’ve never heard of one. I know it doesn’t exist yet, until we have a real one. Carmen Li It’s based on reserved contracts. Doug Based on everyone else’s contracts, you’re just— Carmen Li —for 36 months, five years. Doug You’re just aggregating— Carmen Li —data. We aggregate it. That’s— Doug Okay, so it’s real data. It’s not implied. It’s like— Carmen Li It’s implied based on the reserved contracts we have. The forward curve is supposed to be created from futures prices. This isn’t futures pricing. It’s based on longer-term reserved contracts—that’s where we get our data. Doug Very interesting. Carmen Li Yeah. Doug If you had to guess, by the end of the year, will prices be above or below $5 per GPU per hour? Carmen Li Based on the forward curve, we’re still in contango. As an index provider, I’m not supposed to give forward guidance. Doug You’re not making bets anymore? You’re just— ### Token indices on the Bloomberg Terminal, and a leading indicator for AI revenue Carmen Li I actually don’t know. That’s one thing. Number one, I don’t know. Number two, I’m not supposed to give forward guidance. Number three, I’m not going to trade it. We don’t know, and we shouldn’t have any exposure to the underlying assets. That’s one of the things I really want to highlight. I thought this was table stakes: as an index provider, you’re supposed to be neutral. You’re not supposed to own GPUs like H100s. Imagine if I were a neocloud and I were long H100s. What would my incentive be when I create H100 indices? Doug Mm-hmm. Exactly. Carmen Li I thought that was table stakes. But computing is such a new asset class, and there are a lot of people who want to apply the rules of other industries. They’re not competitors, but other market participants—not traditional data providers or financial service providers. They don’t see things the way we do. I believe rules are there for a reason. They were created because we’ve seen bad things happen before, and we want to avoid that happening again. We want to avoid conflicts of interest. That’s why index providers aren’t supposed to trade their own indices or have financial exposure to the assets they measure. Doug Cool. Carmen Li You’re supposed to have a high-quality compliance program audited by a third party. Someone should be separately calculating the indices. But surprisingly, only a few participants do those things. Doug No one else creates a GPU index? Carmen Li People are creating GPU indices. Anyone can create indices and call themselves an index provider. But you should look closely: do they actually have exposure to the things they’re trying to measure? Do they have a natural conflict of interest, meaning they own the things they’re supposed to measure, or they’re owned by people who do? Doug Why doesn’t Nvidia make the index? They’re biased. They’d keep the price high so everyone— Carmen Li Exactly. You can’t, right? Nvidia is a good client, obviously, looking at data. We measure AMD too, and we don’t care whether the price goes up or down. It’s not our job to control the price. It’s our job to reflect the supply and demand curve. Doug The actual one. What about forecasting the supply and demand curve? Is this something I should look at in SemiAnalysis, understand their assumptions, and tweak them? Carmen Li You should look at everybody if you’re on the buy side or sell side, as an analyst. What we do is give you what we see based on our data—for example, our implied forward curve and our intrinsic value. We tell you our assumptions and our hypothesis. You should tweak them based on your own information; maybe you have an edge. We’re very transparent about our limitations. One of our most popular products on Token Data is our token indices. It’s fascinating: the token indices went viral before we even launched them. We connect with the Bloomberg Terminal and test to make sure the data pipe works before a product launches. We usually integrate with them before officially announcing anything, just to make sure things work. We piped into the Bloomberg Terminal without telling anyone, to make sure everything worked for the token indices while we were testing everything else. Doug You piped in, meaning you’re distributing your indices on Bloomberg? Carmen Li People could search for them, but no one knew who we were. We hadn’t marketed them. Doug You’re on Bloomberg now? Carmen Li Yes, on Bloomberg. Doug Are you on other data terminals? Carmen Li Refinitiv, Reuters as well. The five GPU and token indices are distributed to them. It’s a terminal-only use case, though. You can’t download the data or get it through their data licensing model. We have 12 indices on Bloomberg and Refinitiv, and we’re delayed by design. But anyway, the token indices are fascinating. The altcoin revenue chart alone— Doug You see that? Carmen Li It went up, but it’s stagnant, right? Remember that chart? The bar chart? Doug Derivative. Yeah, it’s getting a little bit slower. Carmen Li The bar chart, right? Doug Yep. ### Token pricing has no transparency, and why customers will demand control Carmen Li If you map our token indices on top of the bar chart, you’ll see something really interesting: they’re almost a leading indicator. When our token indices ramp up, that’s when altcoin revenue ramps up. Then our token indices crashed to about a dollar today, I think. That’s when you see revenue start to stagnate a little. Our index measures willingness to pay. For that particular product, I don’t need to tell you what Anthropic charges for a million tokens of input and output, or how much you’re supposed to pay them. Input and output usually have very different prices and very different consumption volumes. We don’t care about model quality; we only care about how people use the models. We track 300 open-source and open-weight models, plus closed-source models, to figure out how much society would pay if it were a person. You can use any models you want. We see a lot of people shifting from premium models to cheaper models after the token-mixing era. That’s why you see token prices dropping. It’s not because people are cutting prices in different ways; people are moving at least some of their workflows to open-source and open-weight models. Doug And why are they going to cheaper models? They just don’t need that much compute at the leading edge for everything they’re doing. There’s some sort of harness that says, “Use the cheaper stuff for these workflows.” Carmen Li That could be. We don’t really have information about customer use cases. All we know is how many tokens were consumed by which model. You see people saying, “Hey, we can use token factories,” right? Most token factories host open-source and open-weight models, rather than proprietary models, because they really can’t do otherwise. Some models are getting close enough that you might not need the smartest models. It’s more economical. For summarization, you might not need the best cloud models. Maybe open-source models are fine. Doug The big difference here is that tokens are like the internet with megabytes and gigabytes. But instead of getting $100 a month for all-you-can-eat, they’re charging for every megabyte uploaded and downloaded. Carmen Li Input and output, that’s correct. They charge different prices for input tokens and output tokens. What’s interesting is that with GPUs, there are on-demand rates and reserved rates. You can reserve a GPU at a rate that you determine. Carmen Li For tokens, the world is almost entirely on demand. If I were a token factory, I could change a lot of things on you without telling you. Price is one; precision is another. You might not even know I’d changed something unless I told you, which I usually wouldn’t. Customers have absolutely no transparency or control, unfortunately, when they use these services. And forget about security, latency, and all those things. I think control can be the next frontier people demand. It’s not only about pricing; there are other factors, too. As a customer, you can decide what you care about. Maybe you don’t care. Maybe you want to go fast and you’re happy with a lower-quality model. Or maybe you want the model in 2 days. Doug So you need a control plane for your agent, so you can control how much it’s using. You don’t want to say, “Hey, go do this task 24/7,” and then suddenly get a huge bill. Carmen Li Exactly. You want to set a budget. Maybe you don’t want that high a quality. Maybe it’s something random you’re trying to do for fun. You don’t need the best model for that. It depends on your expectations for quality, your budget, and your time frame. Maybe you don’t care when it’s finished; it can take a whole week or a day, up to you. How secure do you want your encryption to be? Do you care about using GPUs in other parts of the world? Ideally, a great token factory will deliver a solution according to your needs. Doug The big difference here, in what Nvidia changed, is the structure of the market. All the telecom or compute we’ve consumed before has essentially been all you can eat for a set price, right? Now we’re paying for uploads and downloads based on usage. It’s a different market structure. Do you think that lasts forever? Are we always going to pay for uploads and downloads in tokens, as opposed to a flat fee like we do for cell phones? Carmen Li If you own your own GPU rack or your company does, then you can basically run anything you want on that server and not pay per use. You might pay someone else— Doug Depreciation’s real quick here. Carmen Li Depreciation’s real, too. But you can run any model. You’re not paying on a per-model basis or per-token basis. You’re paying for the GPU, so you try to squeeze as many tokens as possible out of that GPU. We can consider that all you can eat, but based on the GPU as the hardware asset itself. Doug When I think about the entire AI ecosystem and all the places you can invest, this is something I talk about with portfolio managers, whether at Sands or Janus. They’re invested across the entire AI cycle. Where do you want to be? Do you want to be in the fabricator, the person selling the tokens, the semiconductor company? Now there’s also data licensing, or the software layer where you’re building the use case. Where do you think is the best place in that entire stack to invest? Carmen Li I’ve been thinking about this as well. If I compare it to the internet era, you could invest in the fiber layer. You could invest in Global Crossing and companies like that. Meta just put in a huge pipeline. Doug You could invest in the hardware, the fiber. I’d argue it’s the basis of the internet. Or you could invest in Meta or Apple, or in app stores and applications. Carmen Li Yes, I think many things can capture value. So far, the biggest value capture has been with the platforms: Google, Meta, and other traditional tech companies. It depends. Doug Are you talking about the hyperscaler part, or the ecosystem? Google—the actual part where they— Carmen Li The actual Google part. Doug The advertising. ### Homegrown silicon, fab capacity and why a hundred chip designers can't create a bubble Carmen Li Pre-AI, in the LLM phase. If we compare this to the development of the internet, the internet drove a lot of economic activity. From the beginning, you had the fiber layer. You had people creating computers and browsers, companies like Yahoo, and then Google and Meta creating search and social platforms. Then you had more specialized companies doing e-commerce. So which companies got the biggest share of the upside from the internet as an economy? You know the answer. Doug It’s the ones that created the ecosystem and owned the consumer: Google and Meta. Carmen Li Exactly. So who is going to own— Doug Or Amazon, right? They have distribution. Carmen Li Yes. When we look at the AI era, I don’t know. We’re still at a super early stage in the AI cycle. I consider AI a gigantic baby. It’s still a baby, right? I don’t know how that baby is going to grow. We’re looking at the Nvidia layer right now, and the fiber isn’t even fully built out. Doug Well, Nvidia is getting a large share of the economic rent right now. Carmen Li You have to have that; it’s the base. On top of that, you need the LLM layer, the harness layer, as everyone is building now. Then you need applications: how we can actually use it in the end, how you and I can consume it. Will another layer of winners emerge that we haven’t seen yet? Maybe it’s over there. Maybe it’s a different topic. I don’t know. Doug I want to go back to the index and what’s at a premium and what’s at a discount. Over the last year, there’s been a lot of talk about homegrown silicon. The hyperscalers are saying, “Jensen’s making way too much money. We’re going to build our own.” Everyone’s doing it now, right? Carmen Li Right, right. Doug Which ones are at a premium? Which ones are at a discount that you can share? Carmen Li Last year, AWS made a big announcement about a price cut. I don’t even remember that. Doug On Trainium? Carmen Li On Trainium and other models, not just Trainium. If you own your own GPUs, you can design pricing yourself. Azure is designed and built with other partners. Meta is doing something on its own. Everyone is designing something. But they’re not the same as Nvidia GPUs. Some inferencing chips are meant to supplement Nvidia’s use cases, not replace them. Some are designed just for inferencing: “Don’t worry about Nvidia chips at all. Just use us.” We’re seeing all varieties of those chips. What’s interesting now is that it doesn’t matter who designs the chips; at the end, look at the fabrication. One interesting conversation is about bubbles. People assume, “Hey, now you have 100 design houses.” But does that mean a bubble will form? Doug Mm-hmm. Carmen Li You can have 100 designers, but only one company can make the chips. It’s still constrained. In the past, I was making 100 chips for Nvidia. Now maybe I’m making 80 for Nvidia, 10 for you, and 10 for someone else. The total capacity is still fixed. Doug But they’re growing, right? Carmen Li They’re growing, but generally they’re not shooting up. The design houses can multiply every other day if they want to. You can start a new design house tomorrow, but that doesn’t mean capacity can increase at the same speed. Doug So the bottleneck is fabrication capacity. But they’re adding it, right? In the U.S. and— Carmen Li In the U.S. and everywhere in the world, 100%. But— Doug The people who make the fabrication machines are in the Netherlands, right? ### Why compute futures enable price discovery, not a bubble Carmen Li Totally. That’s the really interesting part: capacity can go up everywhere, but can it all go up at the same time? If anything doesn’t grow at the same rate, we’re still capped by the slowest-growing part. Say Nvidia and TSMC can grow, but the packaging steps can’t be delivered at the rate they’re growing. Then we’re capped at the packaging layer. There are so many steps. Doug The supply chain is slower to react than this gigantic demand, really. ### Running a data company and a marketplace at the same time Carmen Li Everything has to be aligned. A GPU can be turned on by you and me, but even if everything is aligned, regulation might say, “Hey, guess what? No more GPUs, no more design houses.” To get a whole GPU project done— Doug You’re talking about the moratoriums that are delaying them? Carmen Li It could be, right? Or something could be paused. Everything literally has to be aligned across a hundred moving parts. Doug I haven’t heard of this. I mean, I’ve heard about moratoriums and delays, and certain states catching up on “not in my backyard.” I know there’s some security stuff in Washington. I haven’t heard of this sort of scenario where— Carmen Li I mean, New York State said no data centers for a few years, right? New York State. Doug Okay. Carmen Li Obviously, they can change it. It’s not that easy to form a bubble in GPUs and AI, particularly because of the way the system is designed. But people always say, “Hey, now you’re launching futures. This is a signal of a bubble.” I’d say it’s the contrary. Without a financial product, the market is only long. With futures, people with a strong opinion can short, so they can bring the price down. Carmen Li So this is one of the things I think we would create: a mechanism for price discovery. Doug I want to come back to your companies, Silicon Data and Compute Exchange. What has been the hardest thing you’ve had to overcome in your entrepreneurial journey over the last couple of years, compared with working as either a trader or— Carmen Li We could spend another whole hour on this. It’s really fascinating. I started Silicon Data first, and then I took over Compute Exchange. I thought I could carry over some lessons I’d learned, but by definition, they’re very different companies. One is a data company, and the other is a marketplace company. They have very different cultures by design, as well as different margins, ways of thinking, and go-to-market strategies. You really have to think differently. It’s almost like raising daughters. You can’t say, “I raised one daughter and she likes swimming. My second daughter likes swimming, too.” It doesn’t work like that. They have different approaches and personalities, and you have to think of them as completely different people. That’s only fair. Similarly, even when you think there are a lot of lessons you can apply from one company to another, that’s not really the case. You have to be thoughtful about it. I’ve made so many mistakes. But it’s completely fun. I think it’s great. I love being an entrepreneur. The best career experience I’ve ever had in my life was starting my own company in 2024. My learning curve shot through the roof. I’ve grown so much as a person. It’s really fun. At the same time, I wish I’d known how much public speaking and media we’d have to do, and how complicated our world could get. But with computing and AI being such high-profile businesses, our indices matter to the world. ### Carmen's lesson for entrepreneurs: pick something you love Doug You can help the market grow by providing that stability for creditors and users. Carmen Li So, people— Doug This is what Ray Dalio did in the ’70s for McDonald’s, for the agricultural business. He provided that stability for them to create the— Carmen Li I love that analogy. I think it’s great. I love the idea of a market running efficiently, fairly, and transparently. That’s good for everybody: participants, customers, you and me, the government, and society in general. You want risk allocation and efficient capital systems. You don’t want to penalize everybody equally. You want to reward the people who are doing a good job and penalize the ones who aren’t. You want to be fair. That’s where the exciting part is. Doug What is one lesson you’ve learned as an entrepreneur that you want to share with other entrepreneurs, and what did you learn from that lesson? ### The vision: the liquidity provider and the referee for every layer of the cake Carmen Li Do something you’re really excited about, because it’s a continuously hard, high-pressure endeavor. You have to be passionate enough to overcome all the random obstacles you might never have expected. Because you’re passionate and excited about it, you just want to learn and do it, right? Anytime, any day. So definitely, if you want to be an entrepreneur, pick something you’re really excited about. You’re going to dedicate literally 24/7 to that endeavor. It takes that much time. The second thing is, you want to enjoy it. I really love what I do. Every time something happens, every little bit of it gets me excited. I’m excited when new chips come out because it means I get to create new indices. I think it’s really fun. If you’re someone who doesn’t want to be challenged on an hourly basis, you might not want to go through the pain. So pick something you love. Doug I love that. Where do you see the company in 1 year and 5 years? Carmen Li For Compute Exchange, I want to be the liquidity provider for every single GPU. Instead of people telling me, “Hey, it took me 2 months to procure GPUs,” I want people to say, “Hey, it actually took me 3 hours. I’m done.” Then I can focus on what I do best. Doug So you’re the market maker. They don’t go to NVIDIA, they don’t go to AMD in the future? Carmen Li They can. They can. But it’s more like today: when you’re buying and selling stocks, you can go to Nasdaq or call brokers at one of the banks. It’s your option; it depends on the workflow. At least you have the option to say, “Hey, guess what? I’m just going to go to an open exchange and get what I want.” I want to create that kind of liquidity platform for everybody, not just the big players. You want a window? Great. Go to my exchange and get a window for a year or for months. You want 20,000 nodes? Great. Go to my exchange. For Compute Exchange, I want to be a liquidity provider for every layer of compute. For Silicon Data, I want it to be a referee for every layer of compute, so people have a trusted source for indices, valuations, and benchmarks. Then people can have confidence saying, “This is my risk exposure. I mark to market every day. I know what interest rate I’m charged and how I would hedge.” I think that would be a really good place to be. Doug Great. It’s exciting to see your enthusiasm. Running two companies in a fast-growing space, and creating that neutral environment to give companies predictability and enable credit to be provided with that predictability, is quite amazing. Thank you for taking the time to join us here on Pitch The PM. Carmen Li Thank you so much for having me. This is great. This is fun. Doug Yeah. Carmen Li Glad we had fun.