AI has a GPU pricing problem. Silicon Data wants to fix it.

0:00This episode of Equity is brought to you by Schneider Electric. Schneider Electric, advancing energy tech. Hello and welcome back to Equity TechCrunch's flagship podcast about the business of startups. I'm Rebecca Balan and this is the episode where we bring on an industry expert to help us explore a trend in the tech world and dive deep. The AI build out shows no signs of slowing. With billions of dollars a year going into data centers and GPUs, compute has become the single biggest cost for anyone building AI products. But what happens to those billions if the price of renting a GPU crashes or spikes?

0:38Today we're joined by Steve Howe, Head of Research at Silicon Data, the company trying to become the reference price for GPU rental and the index that a Wall Street futures contract would settle against. Steve, welcome to the show. Thank you very much for having me. So you are new to Silicon Data. You joined in May this year as Head of Research. So it's only been a few months, but I'd love to hear just like a little bit of your background. I think you did quant research at Bloomberg, was it? Yeah, Bloomberg Indices. So I was previously a senior researcher in charge of our systematic equity index research benchmarks at Bloomberg Indices, which is actually, I think, new to people.

1:17Most people think of Bloomberg as either the news business or the terminal business, the data platform. And the index business is actually a separate business, one of the many vertical pillars that is a diversifying attempt from the terminal itself. So you and I met at a party in New York and a soiree. And I was really interested because your friend that you were with described what you're doing at Silicon Data as data centers futures, right? And so for me, I'm thinking futures trading, that's like gold and corn and grain and things like that. So it's, you know, commodities. So I would love to hear, before we kind of get into the nitty gritty of what Silicon data is doing, like, for those who aren't aware, can you briefly explain futures contracts simply for listeners who don't trade?

2:05Yeah, futures, actually, it sounds complicated. And indeed, I mean, most average investor, retail investors probably don't use, you know, financial derivative futures. But it's actually one of the most ancient, you know, financial instruments in the world, right? the whole idea is that in the ancient days suppose you were planting crops, right? And let's say your cost of planting crops is a dollar and currently crops sell for$2, you make a profit of a dollar. But you worry that by the time you harvest, let's say six months or a year from now, the price of corn may actually plummet to be below your cost and you may end up with a loss.

2:44So if there exists a futures market which allows you to deliver that bushel or the amount of crop at that point of harvest at the price that's currently trading at, say above your cost, you might be inclined to lock it in. Now, why would that be such a contract? Why would someone offer you such an arrangement? Someone else must have a different point of view. And someone may actually believe that, oh, actually based on my expectation of weather and whatnot, by the time harvest comes around, And corn actually will be more than$2. You know, I can actually make even more by, you know, selling you to buying it from you from 2-2 and selling it on to whatever I think it's going to be, right?

3:28So market is always full of people with different opinions. Some people are bearish, some people are bullish, some people are risk averse. Even if they are, have an expectation of things could be good, but they just want to sort of lock in the safety. So that's how futures come from. I see Doral Clef always worked. It works for shipping. You can have cases where someone goes overseas and they'll bring something back and there's uncertainty about what things are going to be by the time some events materialize. And obviously oil is now one of the biggest futures markets. Same idea, right? Airlines want to lock in prices.

4:02So that's basically futures and you can apply that concept more very broadly. Right. So the CME, which is essentially, as I understand it, the New York Stock Exchange, but for futures trading, or one of the platforms where you can trade futures, plans to launch silicon data linked compute futures on October 5th, pending regulatory approval. What I don't get is like, you mentioned oil. Oil is storable, right? You know, compute's not. NH100 can become technically or technologically obsolete. Data centers have a different networking, they have different power constraints. What makes you confident that this will be a commodity?

4:39So a couple of things, right? I think there is a casual, understandable presumption that there needs to be storability or there needs to be some sort of tangible existence for something to be a commodity. In reality, one of the biggest contracts traded on the CME on futures contracts today is actually just the e-mini, the futures on the S &P 500. You know, obviously it's all cash settled and you don't take physical delivery of a company. No, it's a basket of companies, 500, you know, the S &P 500 index. Right. And all you do is at expiry, you look at the P &L, whether or not the number that you sell or bought the futures at is above or below whatever, you know, the underlying ends up at.

5:20And if it's, you know, favorable, you get a profit and if it's, you know, unfavorable, you get a loss and you settle at the P &L. and in fact if you look at oil the vast majority of oil futures contracts do not you know physical settle right in fact the reason why during covid in march 2020 or whenever the oil futures became negative is it was exactly because people didn't want to take physical delivery so they were paying someone else to take the future contracts away from them at a loss right and that's exactly the reason why because the world had run out of storage facility so people who were about to take physical delivery didn't mean to, they just wanted to P &L.

5:58So with regards to compute, the reason we feel confident that this should be something that exists is because the sheer size of physical exposure this year, if you do some simple tally, just the hyperscalers, right, the Googles, AWS, and Amazons of the world are spending some$750 billion on compute. Next year, they're slated to spend something like a trillion. And that we're not even counting the non-hyperscaler spending in the rest of the world and so on. Within this number, let's say$150 billion worth of compute changing hands because somebody needs compute at any given moment and somebody has access, that already creates a lot of, I think, the demand for trading.

6:41And the people that have exposure with compute that they're bringing onto the market, they may want to actually share some of the exposure or in any event hedge, right? Whatever they may be, you know, the same idea we had discussed, the idea of crops, right? You might want to log in, right? How much you can earn from the compute because while you have a certain expectation about how much compute you're going to need or how the market is going to want, you may actually have some uncertainty or in any event, you may actually want to generate some cash, right? From your holdings of compute. So who's actually trading these then?

7:18Is it so it's compute providers? like hyperscalers and neoclouds who might want to hedge downside versus like AI labs and inference platforms who might want to hedge upside? And then just like anyone in between, retail, futures traders, et cetera? Yeah, so we can sort of think this through, right? So any financial markets that are three sides, typically you think Spire and Salibre is actually three sides. They're the people that are natural loans, right? And they're people that are natural shorts, and then they're people who are in between. I mean, you think about, say, the cloud providers, people who are actually either selling compute, you know, have available compute that's available for rent or, you know, people who are building data centers, you know, small and big and small.

8:01They may have the desire to log in rental income, inference rental compute income today. So they are going to be the ones that will want to naturally sort of sell short, you know, and generate income from the futures contracts. On the other hand, you're going to have AI labs that actually buy compute. Contrary to some people's casual intuition, it's actually not enterprises that will necessarily buy compute. Enterprises may actually just directly buy AI tokens and model tokens. But their supplier, the AI labs, whether it's Anthropic OpenAI, the closed frontier labs, or the inference platforms that will post-strain open source models that will sell on the tokens onto these enterprises, they are the ones who are naturally buying compute and they will have an incentive to lock in prices of our compute if there is expectation that compute prices could actually keep on going up so in fact and then there is everyone in between the so to speak the market makers someone who is going to try to create that double coincidence of want because someone who wants to necessarily you know go long compute futures may not happen at the same time someone who wants to go short compute futures and someone is going to be in between and that will be basically your typical market makers in you know drw uh jane street hrt is a typical market makers including for the banks right that could actually be intermediating uh and then of course there'll be other scenarios in which things may not be cutting so cleanly yeah you can have someone who has exposure they want to hedge uh they may not necessarily align up exactly with one of these three parties suppose you have a large financial institution with a significant exposure to AI, right?

9:45And the underlying AI data center income, they may actually want to hedge, right, some of their existing exposure. So the actual exact composition can actually vary. So is it essentially just like a way for a lot of the companies that have skin in the game to, you know, hedge their bets and earn, you know, a sort of a revenue or potential loss off of something that is not the actual build-out itself, but trading on these futures, trading on something that's really just intangible. That's right. And it doesn't necessarily have to take the shape of a betting. I understand people these days think of financial security.

10:26They immediately think someone's betting on an outcome. It could just be that you want to adjust your exposure. Maybe you have too much, maybe you have too little, and you want to dynamically adjust it. I'll give you a different example. Suppose you have several apartments in New York City and you think there's too much exposure. You worry about a downturn in New York City real estate. Instead of outright selling a portion of your real estate by selling one of the apartments, you can sort of suppose there was an instrument that gives you access to an index on the New York real estate index. And if there was a futures contract on it, you can actually try to short that futures or try to get exposure that way and shave off some of your real estate exposure and take that cash and buy some other exposure, say equities or bonds or commodities or what have you.

11:13Basically, at any given moment, people always have a reason to adjust their portfolio composition that may or may not necessarily mean that they're betting on some outcome. Yeah, like if you have skin in the game, like let's say you were, going back to the landlord example, I have too many Manhattan apartments. What if there's a downturn? What if COVID happens again? Something like that. How can I, so are you essentially, I know you don't like the word betting, but are you essentially betting that the price will go down? You're hedging against the eventuality, yes. You're hedging against it. Sometimes people do that just for psychological comfort.

11:44They may not think that it's going to happen, but it helps them sleep better. And you'd be surprised how often that actually happens. You want to shave off some of that risk, not because you expect it to happen or the worst case to happen, just because you don't want the worst case scenario for you to actually feel terribly. Oh, yeah. No, I have a strong disposition towards anxiety and superstition, so I get it. I would totally do the same. So, okay, so like what other, what new behaviors does this enable otherwise? Like do NeoCloud sell features against GPU fleets to finance data centers? Like do hedge funds start expressing a view on whether the AI boom is real by shorting compute?

12:22Like this could have some serious follow-on effects. Oh, absolutely. So, you know, just like as we have seen with the recent Iran conflict, the oil futures market, you know, is basically the major focal point. Everyone is expressing a point of view, you know, that may or may not exactly even be related to oil itself through oil futures. The existence of a futures contract and financial derivatives in general is to provide liquidity. And this is actually one of the things that this country does best, right? The existence of financial derivatives helps make the underlying physical exposure more liquid and therefore giving you a higher degree of confidence about engaging in the expenditures and the spending.

13:04I remember there was a famous moment when Dariya Amodic went on to the podcast with Dorkash where he was asked, if you are so confident about you're building this country of geniuses, how come we just don't do way, way more compute? And the answer was the difference between getting this wrong can be between having a ton of computers making tons and tons of money or going bankrupt because you misaligned the timing. And there's a certain degree in which you make that risk aversion adjustment by scaling back your real investment. Now imagine that on the contrary, you could actually, if you have liquid financial instruments, you can actually build more and adjust that exposure dynamically subsequently with the help of financial derivatives.

13:46And historically, this is exactly what this country has always done best with every single round of industrialization. Wall Street helps the capitalization and financialization of fiscal investments. Yeah, we've had CoreWeave and Nebius report earnings recently. they've been doing really, you know, they did really well. They reported some good revenue numbers, which we can get into a little bit more later. But a lot of investors have been concerned this year that, you know, there's such a huge capex to build out these data centers. They're essentially landlords for GPUs, right? They're just renting the GPUs out.

14:16They don't really own them. And to build it out, they have to, at least CoreWeave, has to borrow the money to do that. So, you know, if there's a future in which there's, you know, declining rental streams, Like, could futures let neoclouds lock in those declining rental streams and make the infrastructure debt easier to underwrite? What moves our world forward? Energy. History has been shaped by it, revolutions won with it, and progress dependent on it. For 190 years, Schneider Electric has been a leading force in this story, creating the technologies that harness energy and transform the competitive landscape.

14:54And today, in the dawning era of energy and industrial intelligence, Schneider is helping customers build more efficient, resilient, and sustainable operations across every sector, from data centers to homes, from buildings to industry. Schneider Electric, your energy technology partner. Discover more at se.com. I mean, that is the idea. I mean, I think we are all rolling in the same direction, you know, alongside of what Jensen is doing with these, you know, financing platforms with the help of, you know, financial institutions and the rise of futures. We're all trying to find different ways in which we can make the financing of these very expensive capital assets, you know, a bit easier, a bit cheaper.

15:40and the existence of a futures market will just bring about more pricing transparency and everyone can actually observe with greater certainty. And as you exactly should say, right now, I think the data centers, a lot of them, they do different things. Some of them actually proactively, synthetically log in prices by pre-selling. So they only build if there is free commitments of a certain amount of compute. And that way, you're essentially logging it in. You're saying, I'm only building this because so-and-so is committing to buy this much computer at this price from me. At which point you worry about, oh, what if this counterparty risk, part of the market has been worried about is OpenAI or Anthropic has made all these commitments about I'm going to buy all this compute.

16:25But they themselves, are they actually necessarily always, at the end of the day, credible? Will they actually be there by the time the compute capacity arrives? So how do you ensure against their counterparty risk credibility? Is there such a thing as compute default? So I think this is where I like to liken the part of the world we're dealing with is the rise of a huge but young industry, like a giant baby. The thing is very young, but it's also enormous that the things that you need are not here soon enough. Yeah, I'm just picturing a giant baby. Yeah, seriously. Well, it's interesting because we're talking futures trading, but Silicon Data doesn't do forecasting, right?

17:06Why not? Well, I mean, it's not like we cannot do forecasting. It's just that we're not necessarily any better at it. So you say the S &P 500 and the S &P Global publishes an index and is an index provider. Just because you publish an index about what's happening today doesn't make you any better at forecasting what will be in the future. Okay, so tell me a little bit briefly where you get your data from. You source it from two places, right? Yeah. For renting a GPU, just like renting an apartment, we observe contracts from two types of general sources. One is we look at public available, published prices.

17:45People tell you how much you can rent their GPU or how long and so on. Another part is we get it from data partners, including our own sister exchange, Compute Exchange, on which we should observe actual transactions, right, of rental transactions that is being facilitated on the exchange. Yeah, okay. We observe the rental contracts, you know, sort of like of different lengths, different geolocation, different memory, spec, CPU, and so on and so forth. And you've got to sort of normalize them so that they're apples to apples. Just like if you, say, had furnished, unfurnished single-bedroom apartments in different parts of Manhattan City and different square footage, you know, different rental lengths, 3, 6, 12 months, 24 months, whatever.

18:26However, the per day rate is not going to be the same unless you make the appropriate adjustment to come for the fact that they are all different. But there is something that you can do. So we use machine learning statistics. We essentially adjust for these essential features that make them different and create an apples-to-apples comparison and consolidate these observations. And we believe we're one of, I think, by far the most comprehensive look at all the available price points that exist globally. and we create these benchmark indices that we publish daily on our website and some of which you also can find them on data platforms such as the Bloomberg Terminal.

19:04So you've got visibility into the prices across essentially this entire market. What is your data telling you today that you think the conventional AI infrastructure narrative is getting wrong? The conventional AI narrative is always changing. I think it depends on people. The narrative always follows the price. If the stocks are going up, people think You know, there is never enough compute forever. But then, you know, price stock could be going down because of some story coming from China or elsewhere. What we see is that what data informs us is the following. I like to say I look in a cross-section and I look across the term structure.

19:39When I look in the cross-section, I mean that I look at not just the price of a single chip. People like to look at the H100 copper chip, right? And if that thing goes up or down, people think that means more demand, less demand. I look at not just H100, I look at B200, which is a blackwell. I look at A100, which is Ampere. All these are named after famous past physicists. So Ampere 100 was launched and released in May 2020. So this is an ancient chip, right? And if you look at the rental rate on this chip, it was actually going down somewhat about the course of last year. And then since late last year through now, it's been steadily increasing and holding strong.

20:18So what has caused that? That's actually just the rise of agentic AI demand. The fact that you are now able to have AI calling other AI and using tools such that the amount of AI demand just exploded. And the fact that AI also has become more powerful. Every single token has become more intelligent and cheaper. Intelligence has actually become artificial intelligence. Machine intelligence has become cheaper over time. that has actually also led to way more usage. I like to tell people that this year really was the first time that AI has become useful. So in this context, if I track the A100 rental rate, the fact that this has helped strong, despite being such an old chip, that tells me that inference demand is really strong.

21:07Because A100s are good at memory, right? They have a lot of memory storage, is that correct? No, A100 is good at nothing. A100 is just old. It's not very good. It's just old. Okay, so why is it good for inference? Because I can't remember if it was the A100 chip or if it was the H100. So A100 is an older chip, right? And it's the less powerful chip. The fact that it's being used to the maximum and rendered out to the maximum tells you that there is just so much demand that people are willing to use something that is not so powerful. I guess the lesson, though, if we're talking about conventional wisdom, a lot of the conventional wisdom right now around data centers is why is it not going to work?

21:45because they're getting loans out to build data centers off of depreciating assets, right? But are the assets depreciating? Maybe not. Maybe because there's just so much demand, older chips, you find a use for them. I think I was thinking of the H100s that are really good for memory storage. They're really good for inference because they're really good for memory storage. And when you have more agentic tasks, you need to have a longer context window, etc., etc. So every generation, they tend to have more memory and tend to be more powerful. H100 was the frontier chip. At some point, it was a few years ago, and then now we have the Blackwell and the next Bear Rubin and so on.

22:20It's always going to get more memory, and the demand is going to basically be additive, right? And the old narrative was that these chips are all quickly depreciated. Within the two, three years, they all either break or stop working, and they'll lose their value. People have this rough mental image of like an iPhone, right? iPhone 15 comes out, iPhone 14 immediately loses all of its value because everyone moves on to an iPhone 15. It turns out there is so much demand for phone usage that even iPhone 14s are being fully used. That's the A100 analogy for you. Like basically, if everyone is walking around using an iPhone 12, even though there is iPhone 17 released, that tells you there's something going on with demand for phones, right?

23:01There is not enough phones being produced. There is something along that line, right? So part of what I think Martin may be getting wrong is that these things may not depreciate as quickly as people thought and may have actually a longer lifespan as people have previously budgeted for. But obviously, market is not dumb, right? Market catches on to these things and the prices in these things. But then there will be other factors that drive some, you know, equity evaluation, you know, having to do with the fact that equity is pricing not the first derivative, but the second derivative, right? The speed of increase, right?

23:33Maybe things don't increase exponentially, but, you know, just super linearly and then people break out too. Okay, so this really contradicts some narratives that I've heard that maybe the AI build-out is slowing, right? Because there have been reports about data center pullbacks and pauses, like Texas halted new data centers, as the governor called for audits. New York also halted construction of all new data centers. I guess, how are you seeing the market weighing out this pushback against data centers across the country, but it seems like it's not really stopping anything? This is kind of funny, right?

24:04I think there are a few different dimensions, all mixed-mashing all together, And AI has become this all-consuming thing. AI could be stopping and pausing because people think this is not such a good idea. The return on investment is not going to be very high and there won't be very much strong demand. That's one angle that people often worry about. The other angle is that there is a local political pushback. People worry about how AI, with all this buildup, is going to cause electricity prices to go up, cause inflation, cause water pollution, cause noise. And obviously that will feed through the political angle.

24:38and there's also just a hard physical constraint that there is not enough electricity in the country. So whatever people may want to build, they cannot find the land, they cannot find the premise, they cannot find the electrical power. And in order to have that power, you need to bring your own electricity, which means that you have to build your own power plants, which means that you have to go source, generate all power, but then you turn around and realize there are actually no turbines and whatever gas turbines you can strip off of the Polish airplane or Iraqi ship, they've all been stripped out already.

25:10And then you realize you don't have transformers. They're not the model, but the physical transformer that takes structurality from high voltage to lower voltage. Those things are all sold out for years. So there are different dimensions in which things can pause or slow down. At any given moment, someone is worried about one of the channels. I think the Texas news, the pause, that may have something more to do with the fact that there's just been so much built out in Texas. Everything has been built out in Texas because of the more favorable political attitude towards building out, that maybe there is some sort of, oh, let's hold on, let's slow this down.

25:43But obviously that creates, any slowdown is going to cause people to wonder, oh, does this have an impact? Maybe we'll have a whole bunch of ships that are sitting around waiting for a site to be deployed at and they don't find one. So there's endless worries. Yeah. I want to ask you a little bit. So I was reading some of your data curves that you guys posted in July and showed B200 prices were 8 % lower, 36 months out, H100 is 13 % lower, and A100 is 15 % lower. So you kind of touched on this earlier, but the market is pricing compute to get cheaper. So what assumptions are built into that? Is that more supply, better chips?

26:22So that one tells you that, oh, basically, this curve is actually down to slope in a pretty significant way because people were still being given a pretty significant discount for renting long-term reserve contracts, right? Locking in this long-term reservation. And over time, over the subsequent three to six months, we see the entire curve has actually moved upward, which is exactly consistent with what we were saying before about rising agentic AI demand. At every term length, at every contract length, the rental rate has gone up, right? You have to pay more to rent a GPU per hour, right? And if you saw that in the rental market, at every rental length for a single bedroom apartment, New York has gone up, you would immediately know that, oh, this just means that a lot more people are renting, there's a lot more demand.

27:08And what's interesting is that if you track the ensuing four months from March, you know, 2026, and on that chart that you may want to put on that thing later, and, you know, it's actually been monotonically increasing, right? It's been steadily going up and up. So high level, like what are we pulling away from this high level? What can we learn about this the difference between neocloud rental prices versus like what you might pay for a hyperscaler let's separate those two things a little bit right so what we were just looking at that chart only contains the rental prices you will get from neoclouds right neoclouds are so to speak companies that only provide air compute you know as a service hyperscalers tend to actually price their compute at a persistent significant premium maybe two to three times the cost of a neocloud because it's a bundled service, right?

27:57When you come to a hyperscaler, you tend to already have been a customer of AWS, your enterprise, there is trust and compliance, data security, different software analytics, and so on and so forth. So it's a different product, you know? Right, because your H100 index had NeoCloud capacity at$2.70 an hour versus hyperscalers at$7.28. Yeah, okay, so that makes sense because it's all bundled in together. That's right. You wouldn't put Billionaire's Row apartment rental in the same category. Yeah. You know, we talked about this. You guys are, Silicon Data is going to launch their compute futures in October.

28:35Hopefully, if there's no regulatory issues and you'll have an H100 and a B200 rental index futures. Is that correct? Yes. Okay. And so apart from the news that you guys are trading on the CME in October, Silicon Data just raised a Series A round. Can you talk a little bit about that? We announced that we had the initial closing of$30 million. It's a Series A round led by Valor and Atreides AI Fund. So Antonio Garcia and Gavin Baker being the headline leaders of those two. And we have also had participation from many other financial institutions, F-Prime, Fidelity, CME itself as a venture fund participated.

29:17We also had Samsung Next, which is the venture arm of Samsung. There are others that have criminally left out, Tectonic and others. The reason we have actually gathered this group of investors is because we want to build out an ecosystem to actually help create a financial instrument to facilitate a physical AI compute financing. And actually, eventually, we had discussed at the top of the hour how these futures contracts don't have a physical underlying. We are going towards it. We want to have physical delivery of compute. We want to build indices that is with greater compute. We have this thing called Silicon Mark, which actually gives individual GPUs at individual data centers that can visit and give it a grade so that you can actually take instance of on-demand or reserved AI compute at expiration of our futures contract.

30:07That is something we are going towards in the future. we believe that with enterprise adoption of AI and inference in a much, much, much bigger portion of the overall AI demand, inference is going to make compute more interchangeable. You can, in theory, take your inference job to different places, and your inference platforms of the world could pull together compute from different providers of compute also using software. It's such a case that you're going to be able to have a certain degree of fungibility, so to speak. Not perfect fungibility, but fungibility nevertheless. that you can basically say, I can have cheapest deliver provided a point of expiration for futures contracts.

30:48So that is actually what would be a much, much bigger plan for the future. Wow, okay. So we'll all stay tuned. But in the meantime, we'll keep an eye out for October the 5th and see you guys on the CME. Thank you very much. Steve, it was so great chatting with you. We're about out of time. Can our listeners connect with you online anywhere? Yeah, so I'm pretty active online on Twitter, especially so Steve Ho, I'm on X. And people can find me there. You can also email me at shou at silicon data dot com. So look forward to hearing from the audience. Okay, great. And thanks again for joining us. To our listeners, you can find me on X and LinkedIn, and you can find Equity at Equity Pod on X and Threads.

31:28Talk to you next time. Equity is hosted by TechCrunch Senior Reporters and produced by Teresa Locansolo with editing by Cal. Subscribe on YouTube or wherever you get your podcasts and find out what's next at techcrunch.com slash events. Thanks so much for listening and we'll talk to you next time.