Compute Capital Markets with Lucas Tcheyan

20:42Alex Thorn:We got more. This was a good one on my friend, Bimnet Obibi from Galaxy Trading. Thank you so much. Thanks for having me. Let's go now to our guest, Lucas Cheyenne, Vice President at Galaxy Research. Lucas, welcome back to Galaxy Brains. GM, GM. Thanks for having me. You've been hosting our other great podcast, Galaxy Grid. How's that been going? Galaxy Grid Monday. You know, trying to learn from you. Trying to speed up how fast I say the disclaimer. No, it's been fun. I mean, it's awesome to cover what's happening every week. We've been doing it only in a bear market so far. So I tweeted this week, I'm super pumped to do it in a bull market.

21:14But yeah, I mean, we're looking at kind of more of the, what I would say, crypto native stories that are happening every week and trying to keep track of what's going to be exciting in the years to come.

21:23Alex Thorn:Yeah, it is fun. It's a four box style pod. You guys move quickly through topics. I join sometimes, but mostly with the Galaxy Research team. We had one external guest, right? Yeah, we had Kareem, former Galaxy Research. Kareem Helmy. We'll Probably have some guests on in the future, too. That'll be cool. It's fun. It's an entertaining podcast. Right now, I think we still just put it on YouTube and X, right? Spotify. No, you can get it on the audio feature. Yeah. Good job, Phineas and team. Let's talk about your new report. That's why I wanted you to come on. Yeah. Your report was called Inference Capital Markets.

21:54Alex Thorn:That sounded like the overlap between crypto and AI when I hear that, or I guess AI and financialization. Yeah. What was the overall report about? Yes, definitely. So I would say it's part of this broader coverage that we've been doing here, just looking at crypto and AI and how they are integrating. This one's a little bit broader. And what it's trying to see is, all right, if AI is becoming a part of everyday use for everyone around the world and there's this huge input for it, which is, you know, there's GPUs, there's energy, there's memory. Are these things going to become financialized? Because they're becoming a massive part of our economies and they're going to continue to grow.

22:31And so it started, the idea originally started because I saw a lot of cool on-chain primitives starting to pop up, things that were tokenizing access to inference, tokenizing inference production, tokenizing credit so that you could lend to data centers and GPUs. But as I dug deep into that, I also saw that off-chain there is this much bigger market that's already emerged, which is on GPU futures, GPU price indexes. And so the report gives an overview of first kind of what's happening off-chain on the GPU side. Then it dives deep into kind of the on-chain primitives that are starting to emerge.

23:03For me, crypto is always a really good place for frontier markets to emerge. And if those frontier markets are successful, they usually get adopted in the TradFi world. And so especially during times like this in the market, I think it's really good to look at like what those markets might be. And so those are the three that I identified. I would guess in a year from now there's a lot more.

23:21Alex Thorn:Yeah, it's super interesting. When you say inference capital markets and you were mentioning the GPU forwards and indexes, what exactly is being financialized? Is it the full GPU, the full machine? Is it like the GPU hours, the tokens and the models? There's a lot of different models that are coming out there. But the primary ones that are live right now by companies like Orn, Galaxy's investor in them, Silicon Data, is basically tokenizing actual GPU hours. And so you're saying for the price of rent, how much is it to rent a GPU for one hour? I see. And so it's an index of that price. There's a lot of issues with that right now, right?

23:58Because like depending on what the GPU specifics are, there can be a lot of changes, right? Where is the GPU located? What's the interconnects for the GPUs? What is the memory the GPU has? So right now they're providing index. They also have partnered with like ICE, with CME. So we should see like actual futures go live and trade it at some point. And if you look at some like on-chain exchanges, you can actually do that. Koushi has a prediction market where you can trade a GPU price. But yeah, right now it's mostly just an index that's aggregating across all of the order flow that these companies see and trying to create like one price point for different GPUs.

24:34So H100s, H200s, B300s.

24:37Alex Thorn:Yeah, it's interesting because I was going to ask that. I feel like obviously in commodity markets, there's different types of oil. You got WTI and other refined versions, Brent crude, right? But there's not that many. And like gold, most people are pricing like a troy ounce or something. Like there seems like there's a lot more differentiation in the GPU market. So like can they – I guess the fungibility is a question I have. And then also, you know, for these futures, whether it's on GPU hours or just like the, you know, just the price of an H100, like, don't we need a big available, I guess, on the GPU hour tokenization, financialization, don't we need like an available spot market to actually settle this?

25:21Alex Thorn:Is there like neoclads where, I mean, I guess there must be where you can just like rent a GPU hour? For sure. There are spot markets. So, like, you can just go on like aggregators and providers and rent a spot hour. Okay. A lot of this is actually happening like the big deals are happening OTC, right? And like there's a great quote by this guy, the founder of Base10, which is a really large inference provider, saying like for actually like getting GPUs in size, like it's kind of like doing a drug deal right now. Like you have a guy that you know and just has like access because they're so supply constrained right now.

25:49Honestly, it's kind of like ASICs on Bitcoin. It feels similar. If you're a broker in that market right now, it's like an incredibly competitive but lucrative market to be in right now. On the differentiation side, yes. So I think that is kind of where the issue is right now in terms of the – besides just the market being young but in terms of getting broader adoption of these types of products is figuring out like how do you handle the differentiation. And they have looked at markets like oil and stuff where there's different types of oil. And then you kind of like have an aggregated price that's representative of a basket of them.

26:20And I think where the real issue will come in the future is on like having to actually deliver GPUs. So right now, a lot of the focus is on the cash settled futures, right? So like there's not an underlying you need to deliver. So you can aggregate pricing. And then you can say like this H100 index is representative of like these categories of H100s. But in the future, if you want to buy that as an actual hedge to then get a delivery of that GPU, you have to probably be a lot more specific.

26:48Alex Thorn:Yeah, that makes a lot of sense. And you just referenced the cash settled nature. But in the report, you also noted that the indices that exist today are so thin. Like they're not liquid enough to be resistant to manipulation or other dislocations. Well, because most of the actual deals are not happening. You know, like most of the actual deals are just happening OTC bilaterally. And that's kind of what I mean because, I mean, again, I don't know a lot about that. I learned a lot from the report, and so I encourage people to read it. I still need to learn a lot more about this market. I think everybody does.

27:20Alex Thorn:Yeah, for sure. You know, the big deals that we read about in the news on the Bloomberg Terminal, like, these are not open platforms for anyone to go and price a GPU hour. They're multi-year deals between, you know, power providers like a Galaxy and a NeoCloud or a Mag7 firm and, like, you know, SpaceX or whatever. And, like, those GPU hours are not, like, public market GPU hours. Exactly. There's a huge pricing discrepancy. Like a spot market hour of a GPU versus a long-term contract is going to be very different, right? So that will all be right. And like is the Mag7 though and like the big hyperscalers, are they going to not – is there still going to be a public accessible one or are they going to gobble it all up?

28:03Because they need so much. Like they're always – because there's always going to be some capacity that something goes wrong and excess capacity frees up and you want to monetize that. So you're going to want to have spot markets. They've been around in crypto for a long time. You know, you go to Amazon, Azure, any of that. They do have spot markets. Yeah, yeah. I think like the real issue – or not the real issue, but the question is like is there going to be demand that materializes to actually use these types of hedging products for real actual like hedging mechanisms instead of just speculation, right?

28:33And so we are starting to see some early inklings of that. There was a Reuters report actually yesterday that talked about how CoreWeave is looking into potential hedging mechanisms. And so they haven't done anything yet, and it didn't say exactly what it would be. But you would assume, you know, with CME, ICE all coming online with GPU features, that's probably going to be one very viable mechanism.

28:53Alex Thorn:Yeah, and there's so many big, you know, whether it's sort of the intermediary providers or the Mag7 themselves or the open AIs and Anthropics and whatever. It's like this is a very big market. Like hedging would be useful. Yes, for sure. Sure. And like the very interesting thing is everyone, basically the price of GPUs has gone the opposite of what most people thought. A lot of people thought, you know, a lot of new, every year new GPUs come out, they're much better and efficient than the old ones. So you think the old ones become obsolete. But actually there's such a supply constraint right now and there's so much demand that, you know, GPUs that came out a year or two ago are either maintaining price or going up in price.

29:32So that actually makes it a better market, right? Because it allows you to have a two-side market. If the price is just going down, it's hard to find a lot of buyers.

29:40Alex Thorn:This is like something the Bitcoin miners faced consistently. Yeah, and it was the same thought, though. Like, why would you run the S19 from Bitmain when you could run the S21? Yep. But then it was like, well, I can't get my hands on an S21. Exactly. And the S19s might have traded cheap, so then you bought a bunch of those. And then difficulty actually goes down at some point, so they become more viable again. Yep. That was probably much more constrained than this is because hashing does get a lot, lot better. Yeah. But anyway, similar dynamic. Similar dynamic. And, like, there is kind of a similar dynamic, I guess, to, like, the difficulty adjustment, which would be, like, even though older GPUs can't process the new models as fast, there is still a lot of, like, the shift to open source models that have smaller parameters and can run as fast on these older GPUs means that there is continued demand.

30:32demand for it. So, you know, it's not necessary that even though new ones are coming out, older ones are going to go obsolete as long as demand overall just continues to pick up quite a bit.

30:42Alex Thorn:And let's talk about that because one of the reasons the older models are being used is the newer models are so expensive in tokens. You know, I know a lot of companies have started defaulting to like Sonnet 4.6. Try not to get them, don't use, you know, Opus 4.8 all the time. You know, Especially if you're not coding something essential, like if you're asking it questions. So the older models are still getting used, let alone the smaller open source. Where do you see tokens? How do tokens fit into this? Are we tokenizing tokens? I guess – and it's so funny that tokens are now this giant international tech term and they're not talking about crypto tokens.

31:24Alex Thorn:talking about, you know, inference tokens, I guess, model usage tokens. And also, where does something like Venice play in? Yeah, definitely. So actually, it's funny, like when I was writing the report, like our editor, Mark, kept being like, you need to differentiate between crypto tokens and AI tokens because the reader is not going to understand the difference. Especially if you're talking about both. Yeah, exactly. Which it does. Yeah, so that I think is going to be the next evolution probably is that we start to see financialization of the actual tokens themselves. Once again, this is a really challenging issue because no two tokens are really the same, right?

31:58Like these models are very different in terms of how they price tokens, whether it's a frontier lab, open source, the size of the model. So that's a challenge. What we are starting to see emerge on chain is, yes, the tokenization of inference access. So it's not yet at the token level, but something like Venice, for example, which is an inference provider. You can use it for private inference. They have a specific product, basically, where you can buy a token called Diem. And what that does is - A crypto token. A crypto, yes, a crypto token called Diem. Yeah. And what that does is it gives you access to$1 of inference on their platform perpetually.

32:38So it's a totally different way about thinking of how to financializing inference access. Now, you know, there's a lot of questions about, all right, like, is that the optimal instrument? Like, for them, it's a liability, right? Like they're now having to – one dollar that previously would have been bought on spot every day, they're now giving away essentially for free. In the future. Yeah, in the future. It's like a gift card of inference. Exactly. But for them, it was a very good way to basically like bootstrap early demand and get people interested. And I would expect in the future they probably like change some of the mechanics around it or introduce other ways of tokenizing the inference.

33:13The cool thing about it now though is like let's say you own Diem. So every day you get one API dollar access to Venice models.

33:23Alex Thorn:So they drip it to you as well? It's just like you hold a token and then every day you're like account resets. Any day you could use it. Yeah, every day your account resets and you have$1 that day. So if you hold it 100, DM. So DM gives you one hour. One dollar. One dollar of token for every day. Every day. I see. So if I own$10 ,000 of DM or$10 ,000 DM, then I have$10 ,000 of DM of Venice spend per day. Exactly. Per day. Yes. And I think right now it's priced at basically almost like four years worth of – Well, that's the other thing that I'm thinking here is, one, are people thinking that the dollar cost of inference will go down?

34:04Alex Thorn:So actually their access will go up? Kind of like it's deflationary. Exactly. But also frontier models use a lot more tokens, so they don't go down. And, yeah. And then separately, like, does – it's$1 of – to your point about tokens, like, the prices of the tokens vary. For sure. But is it actually of tokens or is it of access to Venice's platform? It's of access to Venice's platform. So, like, they could change how much the price is. That's exactly. That's one of the issues I bring up is, right? So, like, you're not actually really, like, hedging anything. You're more just like saying, but here's like how they say why it's maybe an advantage.

34:39So let's say it costs$1 ,200 or so. So like you're almost getting like four years of inference. And after you use, after four years, if you're still holding it. Four years of$1 a day inference. A day. So after that, every$1 a day is basically free, right? Now, obviously, there's like huge, there's like capital costs and everything. So it's not that simple of a calculation. Yeah. The other thing here is, this is an example they like to give a lot. They say, all right, let's say you buy one DM and DM costs$100 that day. You use it and the next day DM costs$105 and you sell it back to the market.

35:11You just got a dollar of inference for free and you made$5 by holding the DM, right? Now, I'm always like, well, the opposite is true. Let's say you buy it and it drops$5. You've now lost – you've paid more for that$1 for inference than you would. So like it introduces this financial element. If you look at the actual usage of Diem right now, I think like on a daily basis, less than 50 % of it actually gets used, right? So if they have like$40 ,000 of inference credits they have to give out for Diem, like less than$20 ,000 is being used every day. And so it shows I think for now it's mostly being seen as like a speculative vehicle for betting on Venice itself.

35:49I see. But I think beyond just Diem, like the cool thing is there are these like DeFi platforms that are starting to come out on which Diem can be used, right? So let's say you have Diem. You bought it because you actually use it for inference. But you now know for the next week you don't need to use that Diem. You don't want to resell it back in the market because you're worried about what happened to the price. You can go onto a platform and you can resell your inference you have through Diem on that platform and recoup some of those costs. You hold the Diem.

36:18Alex Thorn:but you resell the inference hours or time or whatever. So if you go on like some of these like marketplaces that are emerging, you could buy like Fable access or like GLM 5.2 access for like 40 % discount because the DM guy is just saying like, I'm just trying to recoup some of my costs. Like I would get zero. Yeah, exactly. So like you're seeing a discount market emerge there. That's actually quite interesting. It's kind of cool. We haven't yet seen like lending markets emerge, but I imagine very soon we'll start to see like – Like lending out your inference? Take your DM, like borrow against it.

36:48Alex Thorn:This is one reason that this is a great example of your point you made earlier about how the frontier markets can emerge in crypto. Because the – I guess because the composability of the financial – the ledgers are so composable. That's interesting that the idea – because I'm using Fable 5 on my Claude Code Max plan. And right now it's like you have concession and weekly limits. And then they have the – you can use 50 percent of your total weekly limit on Fable 5. currently. They're, you know, dripping it to testers in the wire. But you can't, I can't trade it. If I don't use it, I literally lose it.

37:28Alex Thorn:Exactly. So of course, me and everyone else, we all try to use it. Exactly. But it'd be interesting if I could sell it or rent it, or I guess lease it to you. Exactly. Some other way to monetize it. And if you expect that, like everyone's going to be using inference for everything every day, like you kind of want to have more efficient markets where you can start to do this, right? Especially if you have a supply constraint where like people can't get access to inference because there's just not enough GPUs and things like that. So I don't know if the Venice form factor is the perfect form factor, to be honest.

37:58I think it's the first sort of big one. Yes, but it's a good experiment. And I think it points to where we're going to move in the future. I did see there's this exchange architect that has GPU futures, and they actually put out a post today talking about how they also want to have token futures in the future. So trade like a GLM 5.2 token. And let's say there's tons of demand for GLM 5.2, but there's not enough GPUs that are like actually servicing it, suddenly like that token price goes up in value, which is the opposite of what we would expect. We would expect that they all just go down.

38:26Alex Thorn:I love these words, like we're going to have crypto tokenized AI token futures, derivatives futures. It's going to get wild. In the future. Yeah, exactly. I want to ask you about GPU-backed loans. Yeah, yeah, yeah. I know a lot of people own, some people, but also, you know, certainly the big companies and data centers own a lot of GPUs. and you talked about excess capacity. Maybe they can fill those with pit tents or training or Venice, I don't know, or running a local model and leasing it to somebody else or whatever. But maybe they, even if they can or can't, while they're spinning, can they borrow off the value of them?

39:05Alex Thorn:You talked about how this is like already commercially viable and people are doing GPU-backed loans. Like, what does that look like? Yeah, well, so, I mean, the GPU financing complex is just insane. Is it really? I can imagine. Outside of crypto, just in general, right? That's what I mean, yeah. We've seen it explode. What's happening on crypto is, you know, there's projects that are basically saying like, all right, well, like there's, it's still a slow process outside of crypto and there's not enough capital, I think, yet to service all of the demand. And so the crypto markets are going after, I would say like the smaller to medium-sized data center build-outs where they might not have the same access or partnerships to get the type of capital they need from banks.

39:46And so the crypto products are basically saying, all right, well, we have this great product of stablecoin and we have a global audience where anyone with a stablecoin can lend to us. So we're going to aggregate all of that capital and we're going to lend to these GPU markets.

39:59Alex Thorn:So it's like tokenizing the GPU as like an RWA. And then they like get it onboard into like an Aave pool. Exactly. Now there's obviously like, this is all off chain ultimately. So like there's issues like, you know, what happens if there's legal or like they go bankrupt. And we haven't yet seen an incident like that emerge. We'll be very interested to see its stress test wins. It does. Presumably with the right legal structures. It should work. Yeah. USDAI is like a very robust one. It shouldn't be much different than like a tokenized money market fund except that, of course, in this instance, the underlying is a physical object rather than a financial instrument.

40:31Alex Thorn:But, I mean, I can see it being possible. And you can also see like a very interesting like, I don't know, not partnership, but I guess like ecosystems start to emerge where what do these lenders need really well to underwrite a lot of these loans? They need really good accurate pricing of like what GPUs cost now and what they're going to cost in the future. Like how much is a GPU going to provide in terms of cash flow for a data center in the future? And how do you get that? With GPU indices, GPU future curves, GPU curves, GPU futures. And so you could see how like the development of this separate GPU indices and futures is probably going to help create more efficiency in the lending markets for companies that are trying to do this type of stuff.

41:12Alex Thorn:Before we wrap, you have a broader conclusion in the report. You talked about how crypto may not win the model application or sort of inference layer, you know, whether it's the training or the inference, you know, a Bitensor or a Venice or something like that. But it might have a chance to win the financialization layer. Yeah. What would, I don't know, fast forward a few years, like what would you see that tells you that thesis is correct? I think you would want to see like Morpho and Aave and these big lending markets being like huge sources for people to borrow against like GPUs or to also like be able to borrow against like tokens and stuff, you know?

41:49And I think you want to be seeing a lot of these interesting primitives pop up that enable you to take your access to inference, which you rent right now, right? You can't own that. Own it and then monetize it when you're not using it. And so I would expect in two to three years we see that really emerge. It's going to be really dependent on companies like Venice, to be honest, which like being successful. Because ultimately you need to have a successful underlying inference business to make this work. We haven't seen a lot of those pop up in crypto yet. Venice, I think, has been the best example.

42:22The only other thing, and this is just one of the other aspects I talk about, is like tokenizing inference production. And basically would be it's like Bitcoin, but instead of like mining a block, what you're doing is you're providing inference. And if you provide inference, you get a reward. And so I think that's like maybe another area where we'll start to see growth because ultimately what it does is it just like that token subsidy pushes down the price of inference. And so it makes, that's what drives demand for those types of products. You're just providing the cheapest inference.

42:51Alex Thorn:I like that. That's very interesting. Well, check out the report on galaxy.com slash research. Lucas Cheyenne, VP of Research at Galaxy. Thank you so much, my friend. Thanks for having me.

43:05Alex Thorn:Thank you for listening to Galaxy Brains, the weekly podcast from Galaxy Research. I'm Alex Thorne, head of firmwide research at Galaxy. Follow me on X at Intangible Coins. Follow Galaxy Research on X at glxyresearch. Read our written reports at galaxy.com slash research. And don't forget, if you like Galaxy Brains, to like and subscribe on your favorite podcast platforms like YouTube, Spotify, Apple Podcasts, and more. We'll see you next time.