Why Compute Prices Are Still Rising | Steve Hou of Silicon Data on GPU Rental Indices, Fed Meeting, and Anthropic S-1 (LIVE)

This episode is brought to you by Sarmaya Partners who manages a return to Tangible's ETF ticker LENS. The Sarmaya thematic ETF built for a return to Tangibles Learn More at Sarmayaetf.com and see the show notes for disclosures. Let's get into it, Max. - All right, thanks for kicking it off, Jack. And we are joined today by Steve Ho, head of research at Silicon Data. Steve, thank you so much for joining us on this livestream where it's Fed Day, but really hyperscaler earnings are sort of what the market is really focused on. I mean, if you just look at the move from both Google and Microsoft, pretty high expectations, look, I personally think that the biggest question in markets right now is, are data centers profitable and are they so profitable that the CAPEX boom is going to continue? That's really what everybody wants to know. And that's why we wanted to have you on, Steve, because you're out there trying to build futures, talking to all of these data center providers to understand what the pricing of compute is going to be. So that's why we wanted to have you on today, Steve. I want to start with that big question. Are data centers profitable and are they so profitable that people are going to continue to borrow money to build them, even if the Fed continues to hydrate? - I mean, first of all, they are profitable today. The clearly, I mean, that's why you're seeing there is this famous chart that, by this point, we look at essentially how much each GPU is able to produce in terms of tokens, how much you're able to sell for the tokens, and then how much you cost the rent, the GPUs, right? So I think the data center is also a little bit of an interesting slide. So this thing is some ways complicating that there are many layers, right? The data centers, there is the people that are selling the tokens, some data centers are doing both. Some companies that are originally selling tokens being post printing of inference platforms are also now in the space of by only an on-compute and selling whatever excess capacity that they don't use. So that being said, we used to ask the question of, whether or not the centers are profitable, is a bit like looking at this rolling of the ability of the labs. On the one hand, the existing activity, the margins are positive. On the other hand, you're making continuous new investments in the new capacity, which of course, any given moment, you're looking at that very snapshot, your cash flow is actually not positive, right? You're having the borrow to keep investing. So the question then becomes, how do you assess that combination on an ongoing basis? - Right, Max, Steve, excuse me. And I think that EBITDA, for data centers, like SOXI Core, we've is massively positive, three billion, four billion. But of course, depreciation is a very real expense. And so that eats away. I've never seen like a bigger gap between EBITDA and a net income, just because the capital required to invest is so huge. Steve, if you could send me on apps or wherever, you know, text me, email me just that slide that you were referencing, we can put it up on screen, 'cause I understand, yes, data centers now, like EBITDA, they are profitable. - Can you see this? - But yeah, what are the key drivers? Oh yeah, here we go, boom. All right, Steve, what are we looking at right here? - So the key chart here, right? This is the one that, you know, someone has referred to as the most important chart in, you know, today, right, is the AI economy. And we are looking at, on the one hand, you've got the token index that's telling you the usage weighted price index of all the models that are being used out there, between closed and open. And that has been trending downward since the end of May. - Which is basically bear porn. The bears are saying, hey, the tokens are going down in price, this is deflation. Race to the bottom, commoditization, bear market, boom. - Yes, it's been very interesting to watch everybody, you know, using the same chart to tell, bullish and bearish stories, right? You can tell a bear story. I can also tell you that this is exactly what AI proliferation should look like, right? And prices of tokens coming down and people using the supermodels to do a lot more stuff. And if you plot it, the total token usage, you know, that's going way, way up, right? And in the meantime, part of the reason why you know that's the case is because the rental rates of the GPUs are surging, right? If you look at the other three lines, so there's the white line, and there's the other three lines where the blue line, that's the rental price for H100. And then you've got the rental price for P200, and you've got the rental price for P300. They are all surging. And in most recent weeks, you know, they've been actually almost going vertical. And there's almost a sense in which the most advanced chips are seeing the greatest amount of rental rate increase. And that has to do with the fact that these are the chips that produce the most amount of tokens, right? - And so, sort of the Goldilocks of the models was the second quarter, you know, Anthropic released its S1 and optically an simply exceptional quarter in Q2. And that was when token pricing was high. And before that these rental rates had really surged that much. So since then, token price is down and rental rates up. So their output price is down and their input price is up, which doesn't sound great, but just explain why the bears aren't 100% right here. - Well, so-- - Where are they? - I mean, there's a certain sense in which, so if you break down, right, we have the ability to break down the token index into two components. There is the overall index that's the white line and then there is the green line that's overmodels and the red line that's the proprietary models, right? And what's interesting is that over the last few months, you are seeing not that the underlying two lines have not been going down, right? Even though the overall index has been trading side with colliding ever so slightly downward. And that has to do with the fact that there is still a, within the sample that we are observing, a substitution away from expensive proprietary models towards the open models, right? That being said, if you look at what enterprises are doing and the S1 is showing an extraordinary amount of enterprise adoption of anthropic models and the open, or opus 5.45 is very popular. So it is true. I think, once we get a Q3 numbers, I think I think was leaked, right? Recently, there was an investor day that anthropic did reveal their Q3 revenue and there was essentially a slowdown in growth, but there is also evidence from even ramping itself that there's been a reacceleration because I guess how popular opus has been. So it remains to be seen whether or not we can actually see a inflection in the other direction. - Steve, I want to, what are we looking? Do you have comments on another chart before we ask you a question? No, no, go ahead. - No, I just wanted to ask about the market's reaction to both Google and Microsoft's earnings. Obviously, cloud growth for Google huge, that's partially because they were sort of late to the cloud party. They're catching up to Microsoft and Amazon, but Microsoft likewise, incredible cloud growth. We saw improving margins, as well as the revenue growth, but both of them came in under the market's expectations on CAPEX. When so much focus has been on CAPEX before, it does feel like we're starting to switch to that question about ROI and adoption. When you're looking at hyper-scaler earnings, which of those metrics are you most focused on? - So obviously for the AI trade, you want to look at the CAPEX itself, right? That's what funds are, a lot of the picks and shovels, they are revenue and a host of sustainable, it is and so on, but that's obviously also related to how well these companies are doing themselves. It's hard to imagine that they will keep raising CAPEX if they are not seeing that activity being rewarded by the stock market, and you certainly will not be raising CAPEX like there's much more if your stock has been cutting half, every time you release earnings. So I think those two things are intimately related. So you are bringing news to me. I haven't obviously since I've been along with you, I've not really, I think, checked the earnings from these two companies in terms of what they've been releasing on CAPEX plans, but so I didn't remember that matter itself also did not raise CAPEX plans relative expectations, but there is a sensing which I think people are just sort of guiding exactly what they've been saying. And I don't, did they actually cut CAPEX? They didn't, did they? - They didn't cut CAPEX. They just came in under the markets expectations for Q3. So Microsoft. call has not started yet. Jack, you're muted. Yeah, no guidance yet. I think that they're probably saying it, you know, in a few moments. But the actual capex that they did for the third quarter, both Microsoft and Google were below expectations. And Oracle before earnings, they had an investor day and forecasted to 35% revenue growth through 2030. So that's very bullish. But I think that they may be not officially lower their capex guy, but hinted that they were going to, you know, have tried to have more positive free cash flow. So perhaps good for Oracle, but not good for the AI hardware trade that that by itself. I mean, certainly coming to a stage where everyone is feeling a little nervous about, you know, with all the different narratives going about financing, about RRI and so on and so forth, that I'm not surprised that, you know, probably there's going to be, you know, some degree of caution of discipline, you know, on the side. So I don't have guide guide on the guidance part. Steve, how much are prices going up for the different compute rates? And what does that indicate to you? What shall we, what shall we interpret from this? You can share it. I'll share it screen again. Yeah, so if you look at this chart, right, where we assemble all the GPU rental rates, you know, across the different type of chips we are we're tracking, you know, from the most advanced, you know, B300 to B200, you know, H200, H200, H100, like across the board, GPU rental rates are on demand is going up, right? And that speaks to your shortage, at least as far as GPU rental is concerned, there's no sense in which, you know, there's any sort of slowdown or lack of demand. Now, that being said, the biggest question people are all asking is, you know, what is the source of the quality of demand, right? How much of it does it come from a diverse set of different AI users, enterprises, you know, always how much is it coming from this sort of AI native and AI adjacent companies that are directly either air labs themselves or companies that are directly working on implementing AI and it doesn't create some sort of a circular loop of demand. I think, you know, error from RAM has been running with that narrative, right? So that remains to be seen. At the moment, all we can see is that demand for GPU rental is extremely strong and if anything, over the last few weeks, if you look at this chart, they're showing a sign of acceleration. Yeah, that's a strong sign of demand. Max, go ahead. No, it's going to say a real quick shoe shiner capital is listening to the call and said that Google basically held guidance unchanged on their capex. Can you respect our capex to increase significantly in 2027 and we'll provide more info in the next quarter, but no major changes to guidance. Thank you very much, shoe shiner capital. But Steve, Steve, to your question or to your point about the the sort of circular or circular demand that's coming through, I mean, if data centers are continuing to be profitable, rental rates continue to go up. Yes, maybe the market in the short term is cautious about seeing capex go up. But if we continue to see that, I mean, won't management, maybe Zuck is the king of this of saying, I don't really care what the market has to say, I'm going to keep doing this. Are we going to start to see that maybe from other CEOs saying, I'm just getting paid so much, or what I have, why am I not going to build more? Well, I mean, they are building more, I mean, you know, right now the constraint is on the physical side of actually, can you actually find that power and land and whatnot? I mean, this is talking about, we're talking about projections about what's going to happen, you know, well, in the future, years into the future, when you have very poor visibility, even over the next few months, you know, sometimes in terms of how much you can actually find. So it's not, I mean, we are in this current a bit of a fog, right? A bit of uncertainty as to some of these very, very big questions that numbers are getting so enormous that you are literally talking about, you know, what are the fractions of order of GDP? And I think some of these answers are not really knowable by even the most powerful executives, you know, at the largest companies in the world, right? I think that, you know, your guess is as good as theirs in some sense. Uh oh, that's, that's the trouble. Well, Steve, I also think that the, there's the executive saying, I want twice as many GPUs, I want twice as many data centers as I wanted last year, that's a factor. But it also is, oh, actually, I just want the same amount of data centers. It's just that it costs so much more because memory, prices, because it has gone up, the price of power, the prices of building it out has gone up so much. So I think that, you know, the huge surge and CapEx we've seen has been both of those, has been, yes, we need a lot more data centers, but also the price has gone up. So I think to, to the extent that we've seen a moderation in the increase of, you know, high bandwidth memory and a DRM and other other, other forms of memory, that, that maybe is factoring into the guidance that, that we see. I mean, clearly, right, the underlying components are seeing extraordinary inflation, you know, it's not just, you know, chips for memory and aspects that, you know, be yourself, just to build the same amount of data center, right, if you've got to increase CapEx in all our terms to keep up. So, there's no, no disagreements there. Again, goes back to the question that Max was asking, you know, what is actually, you know, the thing that would drive this forward is that we're going to start seeing, you know, signs of definitive, you know, RRI returns on investment in this form of a higher productivity, higher returns and getting something concrete out of it, right. And yeah, from point that shows up in the macro level, whether it's the economy growing faster, you know, are there concrete signs across the different sectors that are adopting, you know, AI, are they actually showing concrete signs of higher productivity, higher profitability and revenue growth. And we do see some signs of that, right, ramp I've seen as shown data, you know, for example, on how companies that are actually actively adopting AI are also seeing faster growth and, you know, I'm sure it's from the profitability and so on. So, the question is just, you know, what's the pace of that happening? And everyone is eager to have the answer right now when some of these things may take a while to play out. Steve, several things. First, can you please retweet some what I DMed you? Just on X. Thank you. Thank you very much. And I'll just going to fire off a couple of statistics, Max, which is, so Google their search business, which of course is their core business grew 12% and their cloud revenue, which is where they're investing so much in AI, grew over 100%. So I think that you just to put us in the head of some of the executives at Google, they say, hmm, search is growing 12% and this AI thing is growing 100%. You know, it's not crazy to think why they're pretty excited about AI. And by the way, they did that profitably that their operating margins last year were 23% and this year, in this quarter, it was 36%. So they're growing and they're growing profitably in their cloud business. And that's exactly what we see with Microsoft's cloud business, Azure. And of course, who were the clients extraordinarily concentrated in open and eye and in and through. And actually, I think in in Google clouds business, I think that also includes TPU sales. They're actually selling their TPUs and to whom are they selling them? They are selling them to and through a Broadcom deal. So, you know, the the Anthropic S1 came out and there's several interesting financing structures and one of them is a is a Broadcom lease. So Broadcom of course helps Google makes the chips, the TPUs, Google sells them to Broadcom and that really the end customer is is is an Anthropic. So I think that the even though the overall weakness, it was definitely a miss for alphabet, but the miss was in the non AI parts of the business search. And the huge beat was in the growth. But I yeah, Max. No, I was I was just going to ask to tie it back to the Fed, right? The big a big question is can the the people who need to borrow money, can they afford the higher rates? Chairman Worsh was actually directly asked about the sort of blowout in high yield spreads. It was kind of like, yeah, that you know, we're trying to tighten financial conditions. We're happy to see financial conditions tighten on their own. I guess I want to ask you that question, like can can the people who are borrowing sustain these higher rates? And is what we're seeing in terms of of credit spreads widening what you would expect to see in this environment? So I so by the way, I didn't you know, watch the FMC pressure today. You guys were praying cause of tension, but I did a quick summary of what I mean on it sounds like they're not much happened and there was some discussion about taking a monetary medicine. Was that what it was called? He said there was them. It took time for the medicine to cook in. I actually don't really think the interest rate itself is that you're going to stop anyone here per se. Go back to the question we were arguing and talking about earlier, right. If these things are really, you know, as a high, high return, you know, as we people are expecting, if this is really genuinely, you know, going to perfect produce the type of returns that people are projecting, and it's in a stable fashion, the fluctuation, the small movements in the interest rate, this company is not really going to, you know, I suppose to all anyone. You're only in a case in which if the fact really comes to see that the AI economy, to the extent that is strong and is holding up the economy and creating another level of aggregate demand and creating inflation that it becomes directly an enemy, right, obviously, you know, dual mandate, at which point, you know, the interest rate reaction function becomes where it has to be whatever it takes to slow down this train, right. And that, you know, obviously, is going to be very, very impactful for the AI economy in terms of the build out, because if you know that the fact is going to keep raising rates and keep doing it until the every economy slows down, specifically until the AI economy slows down, then, you know, you're going to obviously change your investment plans. On the other hand, I mean, you know, it also comes down to this question of whether or not we were just talking about the aggregate investments, AI investment is going to actually produce the table returns and other things are just going to be sustainably profitable. And that's clearly what's the, first of all, the driver behind a lot of the credit spreads and so on. It's not because of the fact that we're with high tighten and clearly, right, or changing the underlying, you know, macro economy condition. So in a way, everything comes back to the AI being itself, the driver. And so you think that the surge in AI data center bonds, your core we've yielding up to 15% that that's basically supply. Just there's too much demand for capital and there's so much issuance or is there something else? No, I think basically the market, the bond market never believed even once that AI was going to actually produce, you know, miraculous productivity, right? You know, the bond market was never AI-filled, much less AGI-filled. So when people start questioning, you know, whether or not these investments are going to actually pay off, especially when the size gets larger, I think that's really what's been driving up the credit spreads is not because you somehow, you know, the perception of higher rates, you know, what I already wanted to know about, right? That's, you know, the way I see it. Yeah, that makes sense. Steve, how do you square the very high rental rates that you put on screen, perhaps we can put on screen again, with the abysmal performance over the past three weeks of the so-called Neo Clouds core weave and nebius, as well as other, you know, lesser-known companies whose business model is particularly core weave, you know, core business model is borrowing money to buy GPUs and then sell it, it seems like now would be a very good time to be in that business, and yet the stock market is not, not re-waring the stock at all. Again, I mean, I think it's the same, for the same reason why, you know, the last forever long period of time, memory stocks, you know, storage stocks, and really itself have been trading at ever lower valuation, right? You know, the fee ratio has been going lower and lower, and the only reason that's happening, that would happen, right, is if people believe that the earnings cannot be sustained, that eventually this is going to come to an end, right? And that's clearly what the market, you know, is, is, uh, uh, same, and they don't believe whatever that's happening in terms of the compute rental prices, uh, to the extent that it's going up, is going to, you know, be sustained. And again, uh, this is why we are bringing futures out to the compute market so that people have an instrument to which which they can express such views, and if you are in the market for supplying compute, you can use the instrument to hedge, you know, the risk of compute prices going down, and which means in turn the revenues becoming, uh, you know, affected. So, uh, I think they're all, I think, expressing the same set of skepticism, uh, that is spreading throughout the market, that, uh, people don't believe that, uh, what the numbers who are seeing come to sustain, like, whatever the latest, uh, quarterly earnings from, uh, S.K. Heinrichs, what, what was it? Like some crazy $60 billion for a quarter or whatever? Yes. I mean, uh, uh, that is like more than the combination of the hyperscalers, you know, or, or something, right? Uh, and, uh, and clearly, and nevertheless, if you look at their earnings multiple, you know, it's lower. Uh, so, so clearly there is a skepticism that this could be sustained years, years out. Um, yeah, I just look at what was S.K. Heinrichs earnings in USD, uh, this quarter. I literally, I looked at this morning, um, it's not, it was a ginormous number. Yeah. Yeah. Um, Jack, why don't I look that up and while you do, while I do, um, yeah, well, yeah, the, I think they did a very large, uh, uh, six to two billion dollars and it's a six hundred and thirty two percent year on year increase. Yeah. That's, uh, staggering. That's just, I know. Also, also, here's the crazy thing. The revenue is only $78 billion or $79 billion, you know, so in other words, it's all profit. Yeah, that, that operating profit of 62.8 billion. That's, that is simply massive. See, I got, I got another question, but first, I've got a question for our audience, which is how much of your portfolio actually depends on the physical world. If you only have the S&P 500, more than a third of it is tech, energy materials and precious metals about 5%. So Maya partners believes that the next cycle belongs to tangibles. The energy and metal sectors have been under invested and under loved for over a decade. They call it a return to tangibles and the lens ETF ticker L E and S is an actively managed ETF by Sarmaya that they managed to express this view and active is the point. The team follows the theme across sectors, market caps and borders as the cycle evolves, adding where they see opportunity and trimming themes that look overextended. Miner's energy, precious metals, critical metals, all in one actively managed ticker L E and S. When inflation runs, stocks and bonds can fall together, lens adds real acid exposure with the potential to diversify both the stock side and the bond side against inflation and geopolitical risk. Learn more about lens ETF. They're full data in comparison, including performance at SarmayaETF.com, investing involves risk, including possible loss of principle, the funds, investment objectives, risk charges and expenses must be considered carefully before investing. This and other important information is contained in the perspective, which may be obtained at SarmayaETF.com. Please read the perspectives carefully before investing. Steve, I want to ask you about compute futures, because a key thing about Cory of course is that they have locked in so much of their compute before the surge in compute pricing. So the way that they were able to borrow these 10 to billions of dollars was by being into an agreement to sell compute for three, five, even like seven plus years. And they did so before compute prices ran. So it was kind of like Silicon Valley bank or all these banks, Bank of America, which bought interest rate, you know, bought the 10 year and mortgage back securities in 2020 and 2021 when they yielded one and a half percent or two percent. But obviously it would have been a great time if they would have waited to buy it when the 10 years now above 5 percent. It would have been great actually with the benefit of hindsight. Of course, we could have just operated completely in the spot market to harvest this massive premium that we're having right now. I just want to ask you, what is the compute forward market pricing, whether it's Blackwell, Rubin or, you know, H-100s, A-100s, whatever, all Nvidia chips. What's the compute market, you know, compute futures pricing for just how high, excuse me, how long compute pricing is expected to remain just that this high? Well, that's a long question. For example, great job. I love how you did it without, you know, breaking a bath. And you guys shared a screen again, by the way. Yeah, yeah. I shared a screen. And so this is our term curve implied four rates for rental four rates. And we have this for different types of chips. So if you go to our website and you get a subscribe, become a subscriber to our data, you'll be able to see, you know, different rental curves of different maturity starting with on demand all the way up to 36 months for a different chipset. So H-100 and so on. So these are rental term curves that tells you how much you cost to run the GPU per hour of a different term length. And from here, you can infer what the market is currently implying about four rates. We don't currently obviously have, right, traded futures contracts. Those are still subjected to CFTC approval. But if you look at the market implied rates, it seems at least directionally clear that we've gone from an environment where, where there's a typically a downward sloping, you know, fork curve or, you know, guessing commodity which are my lingo, you've got it here. backwardation where people expect future supply to come online and they expect, you know, it discount for long-tenancy contract. That's exactly the type of stuff that, for example, Corrieve has been, you know, selling, right? And they will actually, for a long time applaud it for that type of practice, right? Because they structure it in such a way that they effectively hatched by selling forward, right? At the point of sale, saying, "Okay, I'm going to build this data, I'm going to have all this capacity, I'm going to actually already sell it out so that the revenue is relatively guaranteed, you know, subject to the credibility of the tenant or the off-taker, so to speak." But more recently, as you can see, since October of last year, that's the yellow line through the latest date, which I pulled 25th of October of 1996. The entire fork has risen in level, right? So every term, line, where every maturity forward is actually gone up. And in fact, the shape of the curve has also become, you know, up or sloping, especially, so if you look at the longer horizon, there's the increasing unwillingness of clouds being locked into these super long contracts and they want to have the flexibility and optionality of raising prices as spot or on demand or short maturity contracts for renewal. So indeed, I mean, this is typical of our typical cases we've seen again and again in other types of industries that face a similar type of commodity trade-offs, right? Airlines famously should an airline hatch, you know, Jeff, you are not. Every time you hatch, you know, the Jeff, you know, goes down, and then if you don't hatch, you know, Jeff, you know, go through the sky unintended. And similarly, I think there was this company, Japanese, memory maker Kyosya, all of them. Yeah, yeah, yeah. And they did the same thing, right? Where they essentially locked in, you know, some revenue, right? And then so forward, and then they couldn't participate in the surging memory price and which point at which point they were punished, similarly for, for not being able to participate in that cycle. So, I don't know, I don't know what to tell you. I mean, you know, if things have transpired in a different way, you know, you could very well look smart to have hatched, but in this case, this is what the customer is telling us is that the market clearly still expects. So this is a different type of curve, by the way, then you would expect from a more classical commodity, you know, that's like the storeable, right? So oil typically is, you know, rarely in Contango, when it's in Contango, it's typically fairly bearish. And the reason is because you can just sort of bite oil and hold it and then deliver in the future. You can obviously, cannot do that with compute, you know, it's not a storeable commodity, but so to that extent, the upper sloping curve is actually indicative of expectations of rising future rental rates. This is fascinating stuff, Steve. So this is a H100 Hopper, so more of an old school chip, and even though it was the old school chip a year ago, and people were worried about depreciation, the price has actually gone up, knocked down. So over the past year, the depreciation argument has not played out. That, of course, does not mean that, you know, it won't play out eventually, depreciation is a real expense. So that's Hopper. Steve, what about Blackwell as well as maybe Rubin? Maybe we don't even have Rubin, because it's. So I don't have Rubin here, but I do have Blackwell, have the B200 right here, you found it in? Yeah, if you're just saying you'll be able to see a similar type of a story, right? You know, the yellow line being from a year ago, October 2025, and the blue line, 2026, obviously, this is a relatively nascent market, right? So the curve, you know, is a little bit, so I think, you know, twisty turning, you know, at different points, by generally speaking, you know, the qualitative feature is remained the same. And in terms of depreciation, you were talking of depreciation, we also did this interesting chart that, which by the way, Nvidia also used it in their investor road show. We estimate, going concern market fair value or research value of these chips by assuming like a maximum physical life span of eight years at most, but then, you know, we only discount the cash flow as we observe it, you know, based on market, based on on demand as far as market imply for for rental rates, and then beyond which we apply a pretty conservative decay and discounting to obtain these research values. And if you look at the three generation of chips that we were just looking at, and compare that to something like an accelerated, you know, depreciation schedule, this is actually all line well above what will have been implied, right, by such depreciation schedule. So there is clearly the, you know, depreciation, you know, you can see how, over time, the A100 is actually depreciated, but even six years after market national availability, this is still well of zero access. And wow, that is a number one, a really cool chart. It's cool that Nvidia used it themselves. I will say, Steve, isn't it, isn't it kind of an economic law that either the old chips have to decay in value or Nvidia's new chips aren't even that good? We all know Nvidia's new chips are incredibly innovative and powerful. So I mean, is it, is it crazy to say that eventually these chips depreciation is going to pick up and that in the same way that container shipping was in the shortage in 2021 and tanker, you know, oil shipping tankers are in shortage now. It's like, you can't really expect to this incredibly tight market to be enormous to put it mildly. Yeah, I mean, sure, I mean, this is even in that picture, A100 was depreciating over time. It's just, you know, it's not zero, right? Yeah. The question is just like, you know, what is the rate at which things are depreciating? I show you also the rental rates, right? You know, the rental rates are going up, right? Even with new chips being released, you know, we have not seen a scenario which the H100 just secular goes down with A100 secular going down because new chips are going to be continuously released and scaled out and put into production. Now, part of this is obviously reflecting the fact that a reflection of the fact that there is very strong demand for AI and tight supply such that everything's being pushed up and you have got this, you know, a creative rather than, you know, substantive, right? So impact. So the new chips are not necessarily making the old chips right away obsolete. A100 chips are still being rented out, right? I'll be at the price is probably lower than it would it would have been if the old new chips hadn't, you know, been scaled out and made available. So, yeah, depending on how long it looks at it, but again, depreciation, you know, is happening just maybe not as quickly as some people I thought. So, Steve, I want to jump in. We did get a gift response to this stream with somebody saying, you keep saying that the market for compute is in contango and it had the classic, like, you keep using that word. It doesn't mean what you think it means. Can you definitively say, based off of your data that the compute market is in contango? Would you say that that's fair and accurate or is the curve a little bit more complicated than that? Now, the curve is clearly complicated. More complicated than that, right? It's not a straight line to the right, you know, it goes down and then it goes back up and the market is relatively then on the longer end, this is, you know, again, a market that is not, nothing is, you know, I'm not making any sort of definitive statement of whether a market is in a income tangle or not income tangle, like, the market front of the curve is going down. You are still getting a discount for renting reserve. It's just that it seems the sex signal is pointing in the direction that there is a reluctance of being locked into super long contracts without actually being compensated with a higher rental rate and that is what is all that's meant by it. Okay, and that's super thin market. I mean, that's sort of like, sometimes what you see with the 20-year, like, nobody trades the 20-year, and so oftentimes it actually trades at a higher yield. It can sometimes trade at a higher yield than the 30-year, because it's kind of, like, an orphaned agency. Yeah, possibly. I mean, the exact number, you know, you could be that is, you know, a bit on the thinner side, right? But at the same time, the direction of travel is clear, right? Okay. Now, we got two questions from the same person, Ellie Preta 423, he said, "Isn't the token cost a produced greater than the token price to users?" You know, related question-blended cost to produce a million tokens is still higher than the price charge to the average user. So if you were in AIBARE, one of the things you'd be pointing to is just that this whole thing is subsidized. All of the demand is being subsidized by VC money, by debt. It used to be cheap debt, perhaps not cheap debt anymore. And so that all of this demand that we see, all this revenue, can't really be trusted, because people are not paying the proper price. Is that a fair argument at this stage? I didn't really catch those two questions, but I don't know. They actually mapped the ones, you seem to have added a bunch more of your own ad-lib, but is the total cost greater than the token price they use is? No, I don't think so. I mean, you actually do see, I mean, even in the anthropic S1, they reported their gross margin. So I don't know how you actually, someone came to this conclusion, the total cost produces greater than total cost. Like that's not what this chart means, by the way. Like if you look at this thing, right? - Yeah. - Maybe this is what it meant. Like this chart does not say, right? Like these are, first of all, different Y-axis, dual-axis on the left. You've got dollar per million token normalized, right? So that you're not just sort of seeing something that's going to the vertically to the right. On the right hand side, you've got a per hour rental rate that's normalized to 100 S of that 31st of May date, just so that you can see things on a percentage increase perspective. So this by itself certainly does not say that cost of producing tokens is higher than cost of selling those tokens. If in fact, newer chips are so much better at producing tokens that you're certainly producing, those tokens are certainly selling well above cost. The question is about whether or not, whatever we're looking at now is actually greater sustain. If let's say people are earning enormous margins and earning enormous profits from running GPUs or owning GPUs and selling tokens today, and more people are going to enter, is that just going to necessarily drive down the cost and this spread such that there is no more profits? Or maybe there is certain degree of sustainable mode right to the way in which the frontier models and existing sort of operators are operating that this is actually going to sustain. And that is the trillion dollar question. - So I guess you're saying this statement is wrong. The blended cost to produce a million tokens is still higher than the price charge the average yield. That's just not true. - Yeah, of course it's wrong. I mean, we've seen again and again, right? But margins for selling tokens is very much positive. - Yeah, and Steve, I can kind of help out here, which is I think that that statement may have been true in the second quarter of 2025 for perhaps opening I who knows or maybe some Chinese models, but we saw the Anthropic S1 and I would say it is a radicalizing experience. The gross margin of Anthropic in the Q1 and Q2 was 87%. Excuse me, 88%. Actually, that was Q2 was implied at 88% gross margin and just to give a percentage, that's a higher gross margin than Microsoft, Adobe Salesforce service now. So the claim that Anthropic is selling tokens at a loss is categorically not true. I did some further work on this and also if you take into account basically the cloud service providers and the fact that they get a cut, the gross margin would be more like 82%, 83%, but well, regardless, it's not negative. - Yeah, it's extremely not negative. It's extremely positive and by the way, that doesn't include training costs, but operating margin in Q2 was 8%. And there's been some extremely misleading reporting about the so-called 72 billion dollar loss or the 80 billion dollar loss. That was just basically a convertible note that Anthropic issued to Amazon. And because it's a derivative as Anthropic's market cap has ballooned that liability increases and it shows on a generally accepted accounting principle, it shows a massive loss. But if Anthropic issued equity instead of a convertible note which they can convert into preferred equity, it wouldn't be so. So I'm not very, I don't think that the reporting about that has been accurate. And I would just say, I'm gonna share my opinion that, yeah, Anthropics S1 showed a very strong quarter in the second quarter admittedly. It's, that was the time of token maxing. So the question is, what is the operating margin in Q3? And the bear case, of course, always remains. And yes, there's gonna be token deflation and the Chinese models are to come competition. The fact that to remain competitive, you have to basically spend trillions of dollars. And that could be, you know, bankrupt your company, perhaps other companies, perhaps the US government. Who knows, of course. But I'm just saying that like, I think as someone who, you know, Max and I on this channel try and call balls and strikes and be intellectually honest. Like, I just, it's my view that the, the Anthropics S1 showed not just obviously growth, whichever one, everyone knew, but profitability. And that is a key point. Steve, your thoughts? - Yeah, I mean, the whole debate, so revolve around not what has already happened and what has happened in now, about the projection forward, right? And, you know, the balls believe that demand, you know, for AI and demand for intelligence is, you know, practically endless or any event, an enormous amount and it's going to show up, you know, in strong cadence and it's going to show up, you know, soon, right? And that's the basis on which people are forecasting in a shortage as well. A lot of, by the way, the discussion around, you know, AI supply shortage also surrounds a forecast about the future, right? 2027, 2098 and beyond, what demand will be and what's relatively to supply. So, all of the things we're talking about here, including, you know, whether or not, you know, Delta Sun is going to be a good business, whether or not, you know, sort of memory, and, you know, or even in this case, you know, Anthropics, you know, future performance, it's not about what has happened already, in Q2, clearly, in extremely impressive, like a company grew 14,000%, you know, yeah. The question is, well, they'd be able to maintain, right, this pace of growth, which they themselves, by the way, in the S1 warned that it's not going to sustain that this kind of rate is this only making sense, right? You know, yes, you get bigger, you know, you hit that sort of concave, well, it's a certain event, it's certainly like peer to hear, you're going to slow down somewhat, but also, more importantly, structurally, are you going to be able to maintain that type of competitive edge? And what actually surprised me, my personal experience, is that I find myself, despite all the options of just persistently using Claude, right, you know, it is good user experience for a while, they were like, the model was talking funny, it was like really annoying, how naggy it was, right, you know, giving back all these naggy answers, and 5.5 completely fixed that, and these days, I use only 5.5, and there's a sense in which they have very strong enterprise adoption, and they're very strong retention, right, 900% retention, right? They are clearly, once you are in the enterprise, you know, once you, you know, on the inside, you can start growing, and that's very, that's extremely sticky, right? And this idea of there being, there being threatened by open models, you know, overcast, I think it's probably in a short run, overstated in that they have their own supermodels, right? Instead of, you know, finding some way of adopting, you know, post-trained Chinese model from somewhere, and overcoming whatever issue you have overcome with the compliance and safety and whatnot, it's just much easier just click over to a slightly supermodel on the drop down menu from either open area or Claude. So, on that front, I think they're not, probably threatened externally by, you know, cheap open models as much as just internally. They do suffer the potential risk that open, open 5.5 is just so good that nobody ever thinks about using Fable, right? Or whatever, the equivalent they're off, you know, open eye. - So Steve, you brought up adoption. Gavin Baker pointed something out that he saw from, from the S1 that he thought was pretty staggering, which is, I just cannot believe that half of the Fortune 500 has adopted Anthropic, and 0.1% of their expenses. So people are talking about the cost, consciousness of this, but it represents such a small fraction of the expenses of these companies, and they're already onboarded with the platform. I mean, is that a bullish stat for you when you talk about like the strong adoption? Like I actually think adoption has been, it's been strong, but at the same time, it's the idea that we are close to peak adoption when you see stats like that does seem pretty far-fetched. Again, goes back to your question of forecasting, right? You can read the same exact number in both veins, right? You can read it as a bullish statement, you can read it as kind of a bearish, you know, facts. The balls will read that number saying, there's so much more to go, right? You were barely scratching the surface. Anybody and everybody can just start using a lot more AI, and it's already at the fingertips at the disposal. Now, on the other hand, are they going to? And that is the open question, right? Are people going to be motivated to become more productive? Are people going to be motivated to become more curious and create a lot more ways of doing the same job more efficiently and doing a lot more of it with it? So, they're in lies the debate, and I think that's the entire debate, really. Steve, how dare you bring all that nuance into this? Yeah, really. So we'll say we have three, we're, we have, we have our company of the three. of us and we got employees and let's say we're spending a million dollars a year on tokens but that's with we're not doing on fable we're very cost conscious. Wouldn't we be able to save a lot more money by switching to Chinese models and open models? I mean you hear that these open models are charging five to ten cents of what and theropic charges. Why wouldn't we switch if the capability is almost as good? I mean how do you know the capability to almost go do you guys use Chinese models? I never have used a Chinese model but I've seen the benchmark. Are you saying they're benchmark vaccine? I'm not actually good. I just find it very interesting that like I've not really used the Chinese models either. I mean I didn't the early days when DFC first came out I was sort of playing around with the chatbot I mean I haven't really been using it for coding agents or maybe I've been told right you know that it works just as well but then again I think it's one of those things where I think it's very telling the fact that we're having a discussion about potentially what Chinese open models can do but you know you're not doing it you know why not I think there's a certain degree of inconvenience right it's already so easy for you to click a button and it's not that expensive for you to be using the American models and if you feel like you're a little bit cost-pinched you just you know sort of scroll down one more and use a cheaper model for a little while I mean and the model is getting better all the time anyway so sure Steve but that's you know I'm individual within a small company and like certainly using a Chinese model is a lot more convenient than like making a shoe in China and then paddling across the ocean and bringing it to America but obviously you know that's what Nike does or it's honestly let's be honest it's like Vietnam now but you know I'm talking about at scale if companies are saving hundreds of millions of dollars to do something they probably will just as they did with outsourcing so I'm talking about what is what is the most other than being at the frontier and what's your level of confidence that you know the frontier really matters once the the labyrinth the commoditized versions have an IQ of 170 like what is that even me? I mean how many times have you you guys use iPhones or or Android these iPhones right yeah I've never been a show I've never been told that you can buy a Android phone with much better specs at a lower price yes of course yes how come you haven't done it I mean because the brand but just to say the graphic really have a brand like people really going to be like oh my god I got a text on the graphic and the blue chat into blue it's not blue not it's not just the brand it's it is the user experience right is that the seamless experience of being able to use your MacBook your your your iPhone be able to text your friends with blue bubbles and what not and I think there isn't analogy going on with that like if you were telling me a small company of your size is not finding it or that persuasive to switch over to a much cheaper Chinese overmodel now put yourself in the studio of a much bigger company that has to you know contending with the question of you know compliance and onboarding and so on and so forth I mean there's a reason why right I think part of it was being impressive with Anthropic is the fact that they've managed to I think I think you know convince so many enterprises to adopt right to go through this onboarding process I don't if Max is threatening me because he's finding what I'm saying ridiculous no no I think look is because with the with the the iPhone like there's the joke about how like girls don't like it when they get a text with a green message I'm just like can you can you can you believe he planned our date with a Chinese personal agent like instead of using instead of using Meta's agent to plan our date he used a Chinese agent like I don't know if I want to go on that date anymore like that's what we're debating I hardly think that I mean the funny thing about agent is that agent is actually I'm making you almost model agnostic like when I'm using instinct or when I'm using mues or rock bottom I have no idea what model is powering yet right or I care about it's whether or not is actually delivering an outcome see isn't that isn't that a quite bearish because of course Apple does have the lock-in and the brand and you're basically saying intelligence is a commodity where what's the mode I'm not here to make it give you a bullish or bearish you know argument I'm just here to present you the data and the facts and people can come to their own conclusions I mean I don't know I mean I I've been you know honestly surprised by you know exactly everyone has heard all the same arguments right about how there's no mode there's no mode nevertheless right I mean I find myself using you know and Claude all the time even when Claude was being annoying I just switched it over to GROC or you know begrudge me you know chat GBC I don't know they seem there seems to be some mode I think the mode lies in the fact that over time your AI is going to know get you know you better the longer you chat with that the more information it has about you the more it's going to come back as king right but I've heard that you can just tell Anthropic or Cell Claude excuse me to export all my context and then you can just import that into a Chinese model I haven't verified that I haven't done myself I'm just saying whatever on the streets okay okay you heard a bunch of stuff would you do it I'm asking you would you do it I wouldn't see but but it does it now I mean you're only spending like ten thousand dollars a year whatever but if I was paying ten million dollars a year for a company then maybe I would you know I think something might I think some people genuinely I think some like yeah I don't know if it's going to be a Chinese or the model is going to be a American or the model maybe it's going to be a cheaper version of a flash model where people you know find some ways of tweaking fine tuning to suit their needs I'm sure some are going to do it right the if the economics is there especially if you use a such large scale the question is just you know does that you know actually really scale out right is that going to be most people is that going to be most companies and I think the jury is very much out and I think we're still very much early days for now if AI becomes a meaningful part of opax and you know literally on the same order but not quite the same size of employees or payroll I suspect the incentives are going to get a lot stronger for people to ask these questions um Jack I was going to say you're like uh Tom Townsend in um in metropolitan you're like I haven't read Mansfield Park but Trilling's essays are good enough for me yeah I love a nice literary critique yes yes uh definitely well yeah I mean I'm I'm trying to push hard on the bare argument in the future about pricing power while I want to recognize earlier that like the bare arguments over the past year just have not aged well um to put it mildly and I just think the uh you know we we want to uh communicate that with with the audience um you know that eight percent gross margin for for inthropic um yeah I want to take it back just to the market's reaction uh we did see a pretty negative reaction from the socks initially um after hours it was down 2% essentially giving back everything it had gained today it's now only down 1% um so we're seeing some bounce back uh Google's still down pretty significantly um after hours and Microsoft uh modestly um but there was some other news um we we have shoe shiner capital has once again pointed out um some news for us thank you shoe shiner for for diligently watching and sharing this but yeah I like the shoe shiner guy never I never heard of him but he's yeah we're everybody gets you know after this I'm actually on jumping over to have a chat with him in Chinese so if you guys want to like tune me to that okay open source or closer yeah but uh marvel is said to advanced talks to buy lumentum holding so um yeah and it's it's at I pulled up a Bloomberg article here a deal could be announced as soon as next week said the people who asked not to be identified because the discussions are private lumentum as a market value about 83 billion while marvel is 221 they said uh lumentum was up about 17% marvel down 8% I mean it's hard to believe with those market caps it's not gonna be an all stock deal um maybe there'll be some cash component but uh yeah marvel buying that would be a light yeah marvel buying an optics maker um I mean yeah well yeah what would it say max no it's gonna say jack did does that make sense to you well I think it's it clearly shows that it's a continual of of a trend of semiconductor and the semiconductor supply chain companies um joining forces getting bigger uh you know beefing themselves up and I think as you know this sounds topi but it is true that the pricing power and of the semiconductor supply chain is is higher and it's more you know oligopolized not only because like it's more it's more advanced and it's a lot harder to do you know it's not just like putting people you know in a factory in Taiwan and just like um you know maybe making chips oh what are you saying about Taiwan well I'm saying that I'm saying that you know in the 1970s it was a much more simple process that they still do it in Taiwan it's extremely advanced uh process I respect the semiconductor supply chain in all countries but especially Taiwan but my point is that they they're a fewer of them now you know there's one company company that does um EUV there's you know only three companies really that do like opposites uh excuse me uh etch and and deposition depending on what you're talking about and they have um quite strong pricing power so I expect that to uh to continue um it is uh it is man it is really interesting I guess max we we don't have that microsoft guide yet that I will say max like so I think the enthropy guess one was phenomenally bullish for semi-conductors and made sense to me why it took us out of the doldrums of the Extreme declines one point day. We've had it in October for three days for three days. Let's be honest Three days deep, but I will say if if Microsoft CapEx Um, if they don't raise CapEx, then Oracle will have you know not raise CapEx of their yesterday and actually probably reversed it and then Google Not raising CapEx and then Microsoft not raising CapEx might meta not raising CapEx. This is a different environment Um, well, I mean, you know People have been watching, you know CapEx being raised caught up the quarter Uh, I don't know. I mean depends on what people are actually, you know, expecting about the future These are enormous numbers. I really don't have much to add there. I do think it's interesting that we're seeing this type of consolidation Even with the current multi-environment and optics clearly, you know, it's a whole area Uh, but it is I think it's it's When it comes to hardware, right? So you know these suppliers How big you are as a share of the overall be you know bomb right that's the word um actually matters, right? You know memory and logic chips are obviously big, you know deals and when they go up in price Uh, with with very extremely strong pricing power Uh, that's going to affect the CapEx much more than if you have a relatively small component that's uh, you know Uh, going up be did you some sort of shortage? Well, I mean, the interesting thing about Lumentum is it's not like it was one of those Names that was up like so high in June is like now down 50% and you would say oh wow This is an opportunity to scoop in and buy this thing like it was it's relatively close to its high and assuming that You know the pop that we're seeing after hours is indicative of the price that Marvel is going to pay They're basically going to pay top dollar in terms of what it's traded at For that shows con shows you uh, somebody has having conviction in the future Being optics, right? Yeah, I mean, I mean, it could be a huge mistake like you never know. There's it's not like there's never been a bad acquisition But you know, it shows a lot of confidence to pay top dollar For sure, I always think about when I first You know, I personally invested in Lumentum back in May of 2023 Good. Wow. That was good because someone had told me about a number of random tickers I still never really figured out and certainly at the time had no idea why even faintly why you would be remotely relevant for the AI trade, right? But then again, I also didn't have the foresight of believing in why memory and storage would become as big of a deal as he eventually did Yes, pretty pretty crazy Well, look, we're we're at the hour. It sounds like you've got an important conversation with shoe shiner capital to get to So Steve, we will let you go there jack. You have the final message for us Yes, thank you everyone for watching hope you enjoyed that want to say a big thank you to today's sponsor the lend ETF The Sarmaya thematic ETF ticker LENS which gives real-acid exposure and the potential to diversify a 60 40 portfolio Read the prospectus at Sarmaya ETF calm disclosures in the show notes Thanks for watching leave a rating and review and let us know what you actually thought Not just of the the content but of the format the live format Thank you so much until next time I think you're happy me All right