SemiAnalysis：AI 基建扩张的需求、回报与风险

Content

How big does AI have to get to justify the trillions in infrastructure spending? Let's talk about it with the leading authorities from semi analysis right after this. Welcome to Big Technology Podcast, a show for coolheaded and nuanced conversation of the tech world and beyond. The AI infrastructure numbers just seem to be getting bigger with no end in sight. So, we're going to talk about what the physics of this entire buildout looks like and whether it can continue to sustain itself without not maybe not destroying the world, but destroying the economy. Uh, cuz we have some very interesting guests to speak with us about it. Today, we're joined today by Dylan Patel, the CEO and founder of Semi analysis, and Jordan Anos, member of the technical staff at Semi analysis to talk all about everything uh going on within the infrastructure buildout.

Uh Dylan Jordan, welcome to the show.

Host: Thanks. Great to be here.

Guest: Let's just start by talking about the scale of what we're seeing uh in this AI infrastructure buildout if we can. Uh Dylan, there's like a chart going around X right now that compares uh the AI infrastructure buildout to previous buildouts in the past like railroads, um like highways. And the chart is showing that the percent of GDP for the AI buildout today is about 3.6% 6% or sorry the P the median is about 3.6% uh the average year compared to 2.2% 2% for the railroads. And basically the commentary is that this is far bigger, right? That would put it basically 50% bigger than railroads even in terms of its share of GDP. And the commentary is this is bigger than anything we've ever seen before on a scale that is, you know, effectively unprecedented. Um I'd love to hear your perspective on just how big this is, whether that's the right way to look at it.

Um and then what it means if we're building out something this big.

Host: Yeah, I mean uh the first thing I when that chart was going around, I was like, well, people's numbers are too low. It's actually much higher than that already. Um,

Guest: really? Okay. So, share the real numbers.

Host: Yeah. Yeah. So, so like I mean it's it's if you look at like next year um capex across the US will be on the order of $2 trillion uh include not just data centers and chips but also all the rest of the supply chain that people are investing in. Um so you're close to $2 trillion of capex. Um and then you stack on top of that uh GDP is like close to 30 30 and some change trillion. So you're at like you know five or 6% right not not 3% or 2%. Um so so sort of my thought process is actually like oh wow these numbers are lower than the actual because people are understating the amount and size of capex that's happening in the US and and and North America and the world. Um, and I think like part of this is because, you know, with railroads, your capex is in the country that you're deploying it in.

With fiber, your capex is in the country you're deploying it in. But for AI infrastructure, your capex in America is actually a lot of it's being used to serve outside of America uses, right? So, if you look at, for example, Europe, most of Europe's inference for anthropic is in America and most of OpenAI's inference in Europe is in America. Um and and similarly if you look at a lot of Asian countries um and so there is a bit of this like this this you know capex in America is actually like much higher as a result um because one you're doing all the training here and then two you're actually serving other countries from data centers in America um and so ultimately you're sort of at this stage where capacity is much um the economic impact of it is much larger um than any other sort of physical infrastructure.

investment cycle ever before.

Guest: All right, Jordan. And you are the author of Clustermax 3.0. This is semi analysis rankings of the GPU uh cloud systems. Uh basically putting a lot of the hyperscalers and the neoclouds uh in tiers talking a little bit about their performance and we're going to get to that in a bit. But I I you know you do look across the board in your report and this is something that I I'd like you to comment on. Uh basically we're looking at um where the uh the hyperscaler quarterly yearon-year capex growth is going and in the third quarter of 2026 uh the growth is at 116%. Year-over-year which is astonishing.

um the consensus estimates so what Wall Street is expecting to come after that you know they they drop down in the first quarter well in the fourth quarter to below 100% year-over-year growth and then the first quarter next year uh in the 70% and it continues to drop from there and I think this is a cycle that we're used to right where it's like well you know we're we're on a bigger base so it can't possibly exceed expectations again and it can't possibly double again um so I'm curious to hear from your perspective in your research, uh, do you think that we're actually going to see that trend like the growth is going to decelerate or are, you know, are we again falling into this trap or is Wall Street again falling into this trap where it's like, all right, you've you're basically at the biggest numbers you could possibly imagine, but you still need to think bigger.

[sighs] Yeah, I think we're definitely going to hit limits towards the end of next year and into

Host: Wait, Jordan, before you answer seriously, I got I gota I got, you know, it's like when you ask a kid what's the biggest number they can think of and they say like,00 right and and now and it's like no no no we're at a trillion babe.

Guest: Yeah.

Host: Yeah. Four commas club. Um

Guest: so wait in in Dylan in in this example the kid is Wall Street. Not not making any any statements that people are children. Just uh yeah, go ahead. Sorry.

Host: Dylan may have an a habit of referring to his employees and his customers as kids sometimes, but that aside, um

Guest: I think I think I use a much more mean word.

Host: Okay. Yeah,

Guest: I like where this is going already.

Host: We're we're running out of big numbers. There are there are things to uh look at when it comes to the buildout. This could be everything from like industrial equipment to firms that are prepared to do EPC work like engineering, procurement, construction to chips to cooling systems to you know all the components that go outside of that and um an IT you know estate basically. Uh there's a lot of work being done to expand this supply chain in order to be able to accommodate for all of this capex to to go into it. Um, and it's a lot of it's working and so we're going to see massive growth, but yeah, you you can't just have tripledigit growth and accelerate the second derivative, you know, even further from where it is right now. Um, you run out of things and I think the biggest thing right now is uh twofold.

One is like it it becomes really hard for the biggest NeoClouds in hyperscalers to just organize and marshall the resources required to build another gigawatt on top of what they're already trying to build, for example. and it becomes pretty hard for them to raise the capital. Even though there's incredible returns on a lot of these investments already, you still just need to find the capital to pour into these projects and that's uh you know a difficult thing to do when you're when we're talking about numbers in the trillions of dollars.

Guest: I would I would say the other thing is like by definition the numbers have to shrink. Um you know this year we have 10x growth entropic going from sub10 billion of revenue to north of 10 uh north of 100 billion. So, they had 10x growth. Um, definitionally, they're just not going to have a trillion dollars of revenue by the end of next year. Um, so yes, there's some decelerating growth there. Um, you know, Nvidia's revenue, uh, growth is still 2x year, but it's, you know, it's not as much as it was at an earlier point. Um, just because it is definitionally harder to grow 2x off the same base. So, 3x off the same base. So, it's sort of like, um, same applies to like other firms. So it's it is it is definitionally like definitely decelerating.

Host: Okay. So the growth on spending is going to need to decelerate at some point. But the growth on revenue and profit for that matter is going to have to accelerate you know beyond uh it's basically going to have to trail this big infrastructure uh buildout and accelerate probably beyond where we've seen uh the the dollars being spent. I mean you would you would imagine uh to facilitate all the AI action. And there's been uh further further conversations that have circulated about this. And Dylan, let me just throw these numbers out to you. Right. So this is a Goldman Sachs uh report. Um and sorry, so Goldman Sachs, they say Meta, this is a a tweet, right? Meta's committed corporate suicide after taking on insurmountable debt that the company will struggle to pay back. There's Bane. Bane says hyperscalers will need 6 trillion in revenue, not ever.

This is per year by 2020 by 2031 uh to break even. And there's a Colombia professor uh Stinvan Newberg who said that America will need to spend roughly 3.5 trillion annually on AI services by 2032 8.8% of GDP to justify today's massive data center investment. Um what do you think of these these sort of assessments of where the spending and where the revenue needs to be to pay off the investments and do they have a chance at making good on those investments?

Guest: Yes. I mean I think I think everyone's numbers for revenue requirements are quite low um given the pace of where this buildout is going to go. I think I think it's going to go like I think revenue will be much higher than that uh by 2031 for example. Um I mean there's a couple points here right? One is like you know how much is people are people spending on capex and then revenue and the revenue always trails the capex. These infrastructure investments are six years on an accounting basis at least depreciable assets although there's quite a bit of evidence that they're actually going to last longer than that but you know six years is let's just continue to use six years. So if I spend a trillion dollars today on you know AI data center and GPUs um the GPUs are lasting six years the data centers are lasting 15.

Um but that means definitionally in terms of revenue um either I could assume it's a straight line right so then over the 6 years I need to call it 1.5 trillion of revenue so $250 billion of revenue a year for that trillion dollars of capex or it's going to be something sloped where the first year is much less and the last year is much higher um to get the payback um as far as like you know like statements on meta it's like this is this is just kind of silly because meta meta at any point can just stop buying new stuff and their cash flow can easily pay for all the stuff that they've signed, right? They just don't have to, you know, they don't have to build a hundred billion dollars of data centers next year. They could pull the plug whenever they feel like it. Um there's obviously contracts that they've committed to, but there's also a lot of stuff they haven't committed to yet.

Um and so as far as like the solveny of Meta, like they're completely solvent. They're completely fine. Um and then the question is like what revenue do they make on top of that? Um and as far as like AI's impact in the in the US economy, um yeah, I mean if if AI revenue is only $3 trillion in 2031, then yeah, we'll have big problems. Um but you know, I I'm I'm like of the belief that we can get to like 600 700 million across openthropic alone by the by December of next year. Um I'm of the opinion that you can uh get to, you know, 1 and a.5 trillion maybe even by 2020 end of 2027 uh with just those two companies. Um so that's end of 27. So then the payback is clear, right? Um you know where they're getting that from. Um ultimately um the the capacity that people are building out is matched to what they think business is coming to.

Um, and so there's always going to be a mismatch if we're in the building phase of how much infrastructure are you spending on versus what's the AI revenue because that AI revenue for today is really based on the infrastructure you spent in prior years. Um, and right now capex is not going up at 10x a year but inference just revenue just went up 10x year on year um for the AI labs. And so um we've seen entropic is now you know profitable in terms of um compute cost minus revenue or revenue minus compute cost. um both training and inference combined um and OpenAI gets there you know within the next couple quarters uh maybe two or three quarters um so based on at least what like we expect in the tokconomics model so so I think you know those companies are already uh breaking even if you look at like Amazon for example the biggest builder of AI infrastructure they all of their AI infrastructure investments are profitable today um now you know the customers that for them could go insolvent those customers being openthropic etc Um but right now they are solvent on those um you know making very strong gross margins.

In fact AI has made their gross margin go up in the most recent quarter um from bedrock and AI revenue from Bedrock. Um so so yes AI revenue needs to grow a lot. Um, but I would posit that like being able to automate all the stuff that we've been able to and being able to build all the stuff that the world has been able to um is going to generate meaningful revenue and surplus for the consumer ultimately.

Host: Yeah, I think we need to stretch it a little bit in our minds to understand what we are going to measure as AI revenue going into the future. If these companies really do have the best models in the world that are capable of effectively all knowledge work, you can imagine them going into a whole bunch of different industries beyond coding and having returns on their investments similar to what they've seen in coding. So you can look at things like robotics and self-driving vehicles. You can look at drug discovery. You can look at just health and life sciences generally. And we can take this across all these industries and say if they want to bring a product to market like a cancer drug or like a self-driving vehicle, you can just imagine how much revenue they can potentially return based on the investments they have today.

And if they have the AI, the system to be able to produce the knowledge work required to build that product incredibly well, I think they're the only ones that can really compete on the frontier of being able to do these things. And so I think people lack the creativity to understand how returns can be viewed beyond the like claude code subscription and per token pricing regime of today.

Guest: Yeah, there's there's this chart that looks at how much we're spending as a share of GDP. This is just 2025. So a share of GDP health is 18% that's the top category. Food is 9.1%. AI services is 8.8%. It's the third category in 2025. I imagine it's going to lap or it will pass food in 2026. So guys, I I'm just curious where's this money coming from? Like it, you know, let's say OpenAI and Anthropic combined get to $700 billion in revenue next year. Um, is the money coming from these companies labor pools? Is it just additive because they can sell more? Uh, where is the money that these companies are the customers of these companies? uh where's that money coming from that they can funnel it into the AI companies?

Host: It it's a few factors, right? First off is um obviously companies are going to be able to accelerate what they're building um and so they can invest more. Um increasing share of R&D dollars will go towards AI. Increasing share of cost of goods sold dollars will go towards AI. Um but also you'll have a significant amount of money sort of going coming from um you know labor as well, right? labor and I don't think labor shrinks per se. I think the rest of the economy grows so fast, at least in the short term, that labor labor sort of stays, you know, flattish. Um and and so like um you know, in the past, we've had different cycles across the economy, but one long-standing one has been labor share of GDP has fallen over the decades, and it's been a steady slow since the 70s. And so a lot of the like strife that people say, "Oh, America's economy sucks."

And yet, oh look, the economy is bigger is actually well like labor share of economy is shrank and that's why um in reality a lot more of the labor the value is getting acrewing to capital. What AI does is it's going to supercharge that trend, long-term trend of labor share versus capital share because AI tokens are gener basically just capital um you know in the form of chips and data centers that are sitting there and creating labor um and so that labor share is going to fall even faster um and and and so you know I I I would hope and and for society that that that's my belief you know and maybe it's like a irrational belief but I believe that we will just grow the economy so fast that labor in terms of dollars will not shrink. Um, and it might even grow slowly. It's just that the rest of the economy grows so fast.

Um, and that's where that's sort of where it comes from, right? So, all these companies that do uh offer various services, their services are going to get better. They're going to be able to charge more. They're going to be able to sell more. Um, and they're going to be able to do it with fewer people. Um, or another way to look at it is they're going to have the same number of people and their business just going to get better. Um and that's that's and and and there's a number of industries where this is possible, right? Um you know, obviously coding is an area where people have done it quite a bit. Um obviously general white collar work is getting rapidly automated, but also companies are just able to do a shitload more than they were able to do in the past. Um but there's a lot of industries like pharma and uh trading um and chip design and such, right?

Like chip design is a great example, right? chips. The number of American R&D engineers working in chip design has been basically flat for 20 years. And yet the economy for it has exploded. Now, obviously manufacturing isn't, we're not counting manufacturing. We're just counting the designers. Um, but the designers are like, you know, two of the most valuable companies in the world, right? Nvidia, Broadcom, um, as well as like Google and all these other companies that are making their own chips and many others, right? These designers have not really grown that much in terms of number of people. It's been flat as an industry for us and yet the value that's accured has exploded. And why is that? There have been a lot of tools. Now what is what does AI assisted chip design now do? Well, now all of the chips that are across the economy get better and better.

Your phones are going to get better. Your you know AirPods are going to get better, the chips in them. Chips for robots are going to get better. Chips for obviously AI um are going to get better. Chips for all sorts of things. Bluetooth and GPS and all these other things. Sensors um for gas leaks or whatever that Bosch makes. these are all just going to get better because AI can help you design chips. You do the same number of people, revenue grows, margins grow, um everything looks fantastic, right? And that's sort of like there's many industries where this can get replicated. Um in my view,

Guest: Dylan, you've said that uh I think in the past you've said something to the extent of knowledge work is effed. So do you no longer believe that? Um I think most knowledge work okay like let's separate like my belief in like Looney Tunes 2035 versus like you know 2028 um I think in 20 I think for the next you know X number of years as long as AI is very powerful but it is not better than humans in most every way the most the most high agency people are going to be able to do the coolest stuff they're going to be able to make way bigger impacts on the world Um now you know in like the case where AI is just better than everyone um that that could change drastically right um you know then then like knowledge work is screwed and actually like you know what matters is like people being able to like play games and have fun people being able to act people being able to like you know do the things that humans do that humans only appreciate in each other if the AI doesn't kill us all right but that's sort of like you know [snorts]

Host: we we won't go there today. Um Jordan, you know, this is kind of in your your area of expertise. I want to get started talking a little bit about the risks to like what might derail this story. Um and you know, we'd already talked about the fact that uh there are some contracts that a company like Meta uh could, you know, may not have to go through with. Um you're obviously looking a lot at the businesses of the Neoclouds which are dependent on uh companies like Meta actually making good uh on the on the promises that they've they've put forth to build with these companies. Um is this sort of an underappreciated risk that like everything looks like [snorts] fairly fine right now if you speak to one of these Neoclouds because they'll tell you that they have the best customers in the world.

uh they have the Metas and Meta will make good on the on the um on the promise and they're raising debt based off of this. Uh but in reality the risk might be that a company like Meta might just decide yeah we're not going to do that and that the contracts might not be as binding as uh you know the Neoclouds tell us they are.

Guest: Okay. So I I want to be clear about one thing Dylan was talking about there in terms of termination versus what's going on with meta and the neoclouds which is that generally speaking these hyperscalers or you know large frontier labs who are signing up with contracts for the Neoclouds do not have arbitrary cancellation rights on like a monthly basis where they can just pull the plug and the contract's done. Uh generally speaking, they're signing five years, six-year term lengths and they're off taking the full capacity and they have a set, you know, hourly rate and total contract value that's very very attractive to the NeoCloud providers because those guys have gone out and secured land and power and built a site and even in some cases purchased the chips respectively before they've actually gone and done that contract.

what Dylan's referring to with the the potential to slow down or pause or cancel projects, it's more with respect to the self-build sites, which is to say like Meta has massive campuses that are going to serve as more of a base load of capacity going into the future, which they can either choose to buy chips for or just leave as an empty shell for years if they want to slow down and not go as hard as they can. And generally speaking, construction costs while very expensive are a small percentage of the total cost of a new chip uh project. Like in other words, the concrete and the steel and the walls is quite you know it's important and big and and you know very a real challenge to get it done but on a dollar basis like the HBM and the optics and the GPUs themselves are the the bigger uh portion of the bill.

Um, and so you know what it's what it's looking like right now is that all of these big companies are really needing to contend with the massive growth that they're expected to pursue by their demand from their customer base. either that's meta or openai or enthropic who have demand on their consumerf facing applications like uh chat gpt or cloud code or uh codeex or uh even muse from meta things like that in order to serve that demand they need GPUs and it's profitable s to serve the demand so they need to ramp up and be prepared to roll this thing out to the billions of monthly active users on the meta uh different you know uh surfaces that they serve and approaching a billion of monthly active users for chatpt and hundreds of millions for uh cloud code and things like that. So anyway, um with that distinction in mind, what's the the risk with project cancellation or or a downturn or things like that?

I mean, in order to for for us to um forecast some sort of change in the demand supply curve there, you'd really need to see demand slow. Um, there are hard constraints on how many chips you can produce, how many data centers you can build, and how fast this stuff is going to come online. And as of right now, all of the modeling we can do basically shows that demand outstrips supply. And so, is demand 10% more than supply right now? Is it 50%, is it infinite? I mean, we we don't get to answer that question because we are not able to catch up to demand with the supply of GPUs that you can bring online right now. And so when that question gets answered, if it gets answered, if we find limits to the demand for knowledge work, to the demand for intelligence from the economy, then we'll we'll be able to figure out who needs to slow at what time, where, and how that plays out yet.

But so far, I guess those limits haven't been found. So, um, all right. Uh, before we get into Cluster Max, one last topic I want to cover, uh, big IPO on it way, uh, which is, of course, the anthropic IPO. Uh Dylan, you had a very interesting tweet about it. I'm just gonna read it. You said, "You can't invest in the anthropic IPO at$ two trillion dollars because it could be zero and I'm or it could be 20 trillion, but at 20 trillion, we're all fucked." Uh can you explain that and and tell us a little bit about what you mean by that?

Host: Um to be clear, it's it's entirely a post. Um I was I was like

Guest: You know what? I I get it, but I think there's some truth to to what you're saying. So talk to us about it. I I'm I'm generally extremely bullish Anthropic. I think Anthropic will make you make, you know, be a very successful company in the world. Anthropic could even be the first company to be a 10 trillion valuation company in the world. Like I'm like that bullish entropic. But at the same time, um for them to be a 10 or 20 trillion valuation company, especially 20, um they're first of all, their multiples on revenue are going to have to collapse quite a bit. Um, you know, so they used to, you know, they used to raise at 30x and now they're, you know, IPO is going to be, you know, let's call it they're at hundred billion dollars of revenue and they're going to maybe be at 20x. Um, you know, if they get to 10 trillion, I bet you they would need to be at like a trillion dollars of revenue um or something crazy, right, like that, right?

Um, which could happen at some point in 28 um or 29. um you know it depends of course like the stock market can also go crazy and like make multiples go crazy but you know in the world where they're doing you know where they're worth 20 trillion you know I don't think their revenue I don't think they're even at a 10x sales multiple now they're probably at like a 5x or even less um in which case they're you know if they're at $20 trillion of revenue they're at are they at like 2 3 4 trillion $20 trillion valuation 2 three 4 trillion revenue and at that point you know what is the state of AI if on [snorts] a 150 $50 billion global economy or trillion dollar economy, they're generating multiple trillion dollars of revenue and by far the largest company in the world. Like AI's capabilities are freaking insane.

Um, in which case like you know all of this worry of like whether it's AI kill us all or AI so good that everyone loses their jobs and because distribution of that earnings is going to be heavily leveraged to the people who own anthropic stock or um constituent supply chains of it and not distributed well then people are going to have dist dis distress and like overthrow governments because all of the growth is accumulating to a very small number of people. So like these things could like tear the fabric of society apart and therefore like if you believe anthropic will get to 20 trillion, you know, society as a whole could be

Host: Jordan, do you have similar worries?

Guest: I think Yeah, I think I think uh there's a lot to do to get to that point. Like um I I actually don't think I'm necessarily as worried uh as Dylan.

Host: I'm optimistic, by the way. I just painted the pessimistic picture. I'll go optimistic.

Guest: Yeah. I mean like the the rosy optimistic view is kind of what I described earlier which is like if they're getting to $20 trillion of revenue. This is not done on a per token metered basis for people using cla code as a product and I don't think you know you go and see cloud code go 20x better from where it is right now in terms of like quality and that drives the revenue. I think that they get into other areas of business and I think I would trust Anthropic quite a bit to bring like I said earlier a cancer drug to market or something like that and we've seen similar stuff happen recently with GLP1s just be such a meaningful thing for so many people in the economy and as a result you know companies like Lily have have had massive uh you know returns on on that incredible research that led to those uh incredible drugs that help people and so I think um I would be more optimistic for the future where theoretically anthropics at a $20 trillion valuation.

Um, but I do think that there are a lot of uh going to be a lot of bumps along the road along the way. Um, which is to say going back to your to your question previously, like how can they possibly serve all of this demand and train these models and uh, you know, have intelligence diffused to all of these different parts of the economy so that people can realize the benefits. And I think the way you do that is to build compute and it so depends on you know the nitty-gritty details of construction and installing GPUs and running them reliably and all of that. Like to me the center of the world is the NeoClouds on that road from a $2 trillion IPO to a 2020 20 trillion valuation in the public markets. Yeah, anthropic uh kind of an interesting business case, right? They're spending all this money on infrastructure.

Uh even in the S1, you saw big losses accompanying their revenue. Although um Dylan, as you mentioned, it looks like what they've done is they've turned the tide a bit and they're making more money uh right now than they're spending on compute. Um talk a little bit because I think the the um the belief here is that a company like Anthropic is a low margin business or at least much less than uh than software. But that doesn't actually seem to be the case. And you guys have run the numbers. I would say like if you look at like Salesforce which is probably like one of the greatest SAS companies out there um you know I mean I think like at least like consistently grown to become one of the biggest SAS companies um their gross margins are like 75%.

and Enthropic on their inference is doing 75% gross margins a little bit higher even um you know especially on the frontier models so it's sort of like well and it took it took Salesforce quite some time to get there right they were not at 75% when they IPOed um you know they they took quite a bit of time actually they were but then they like you know the mid in the late 2000s like they were you know I think like it took they were like not quite there it took you know anthropic quite a bit less time to get to that strong levels of growth margin. Um, and if you look at many other SAS companies, they're not even that high. Um, and then the other thing is like their customer acquisition cost is insane, right? If you look at like a a SAS company, it's like, you know, fine, my uh my COGS is really low, but then my customer acquisition cost can be as high as like 30%.

Um, and Anthropic has near no customer acquisition cost. Right now, obviously, they have a humongous R&D spend. um but their R&D spend has grown slower than their revenue. Um and so now they've reached profitability and there's no reason why that trend doesn't continue, right? And they continue to grow more profitable. Um so at least uh operating margin wise and then and then like scalability is the other thing, right? Like the the beautiful thing about a lot of SAS companies even if your customer acquisition cost is so high, you know, as long as your margins are high enough to pay for it initially, eventually you don't have to reacquire that customer. Now you've got this super profitable business. But with like the AI labs, their business seems somewhat sticky. Yes, there's all this like messaging of like, oh, you know, Meta's cutting spend or Microsoft cutting spend or what have you, but ultimately the revenue has only been a straight line up on the chart and almost every customer's only grown on revenue for them.

And uh switching cost even when you know some other Chinese model comes out or comes out with a better model doesn't really shift the spend share that much. um for existing customers um you end up with like this like pretty sticky customer potentially who is growing revenue at some pace and there's no customer acquisition cost and the margins are high so it's arguably a better business than software is is an AI lab

Host: if you can successfully

Guest: but how much is the risk the fact that like there this pacing the frontier um mo movement movement is coming out um

Host: I think that's just signaling I think it's completely

Guest: talk about that you don't think that's going to that's true at all.

Host: Um, I think pacing the frontier is a responsible thing to do if you believe that we could kill ourselves. Um, I don't necessarily think that um they will pace the frontier too much though. I think like there's there's like very little pacing they can do given the race that they're in. Um, and people have made skits about this, but like you know, you know, Daario can't slow down because he doesn't trust Sam and Sam's not gonna slow down and Sam can't slow down much because same with Daario. Obviously, they're going to slow down a little bit when they have these massive security in incidents. Um, and neither of them can slow down because China's not too far behind. So, it's like sort of like very difficult.

Guest: That's reassuring. All right. Um, look, we're going to take a quick break and then when we come back, we're going to talk a little bit about, um, the state of the AI clouds, uh, cluster max, of course, uh, and then we'll continue on with some more questions about the technology and the business. So, we'll do that when we're back right after this. And we're back here on Big Technology podcast with Dylan Patel and Jordan Anos from Semi analysis. Semi analysis, I'm sure you're familiar. It's the Bible for anything AI infrastructure. Uh, these guys are the best. very grateful to have your time both of you and um you know definitely recommend you at least uh check out the blog. You can also get semi analysis on subscanalysis.com. It's the number one technology newsletter uh and welld deserved.

So congratulations guys.

Host: In the newsletter you can uh find the clustermax 3.0. Uh it's it's actually even if you're not technical I was talking with Jordan about this. It's a fun read. It goes into the state of the buildout. Um, and the subhead tells you, you know, it's already kicked off in the right way in the gory detail, the reliability, performance, support, and pricing and security of the GPU cloud providers. There's a lot in there. Um, there's a lot of areas we can go in this uh Jordan and Dylan, but let's begin with the role of Nvidia here. Um the report you know though it spends most of the time on the technical side of things does go into the fact that uh Nvidia is playing a ma major role in keeping these clustries providing uh infrastructure AI infrastructure maybe to some of the frontier labs but really across the board uh and making this whole thing work and Nvidia has been backstopping uh a lot of their um their businesses with uh as you write um revenue floors landlord guarantees and leases that it plans to transfer to third parties helping the Neo Labs and AI labs in obtaining investment grade financing without a hyperscaler being involved.

Um Jordan to you on this one. Uh you know there's the the criticism sort of like the Edotron criticism of this uh of this industry is that Nvidia is basically funding uh the demand for all these Neoclouds and without that uh backs stop um you know this whole thing would basically evaporate and make Nvidia look a bit smaller and make these neoclouds um you know non nonviable. What's your thought on that after going so deep into this?

Guest: Uh yeah, well I mean both in Clustermax and in another article we did recently called Nvidia's backs stop universe, we go into a little bit of the dynamics of what backs stops are and how they help the Neoclouds raise the money they need to deploy the GPUs which then customers are renting pretty strongly at this point. And uh the number that we have is that currently we've got tracking of $588 billion of offbalance sheet backs stops that Nvidia is supporting right now, which is one of yet another unbelievable number in the list that we've been going through on this podcast so far. Um I mean, but Nvidia's revenue is approaching 400 about to cross 500 billion. So we're talking about really big numbers there. For context, that's like 2.5% of the M2 money supply in the US. So obviously there's massive massive amounts of money going on uh in this um universe as we refer to it and uh I mean the only thing I can say about this is that it just seems really pragmatic from Nvidia.

They really uh want a wide diversity of customers. They want don't want to be dependent on the hyperscalers for all of their revenue. Um they also don't want their customers to be dependent on getting their chips from a select set of three or four different hyperscalers. And so they have every incentive to have a a nice healthy market with a diversity of options and uh lots of selection for all of the customers to go uh work with. And in order to deliver on that, people have more problems than just uh raising the money required to buy the GPUs. Well, first of all, they need the guarantee that if they raise the money, those creditors are going to see returns. And that requires you know some sort of investment grade offtaker which Nvidia can serve in these cases in many in many of these cases.

But the way it's worked so far is that Nvidia signs up as that offtaker and then when it comes time to actually you know take those GPUs for themselves for their own internal research work. Somebody else comes along and wants them more at a higher price. And so then they sell them to their to those guys. They don't take them for their internal research and they roll on to the next one. And for everybody who says that Nvidia, you know, is doing this for some nefarious purpose or to to keep some bubble going, I'd encourage you to look at the quality of Nvidia's research. They publish more models on Hugging Face than anybody else. They have incredible CI across the suite of software that they they support on their GPUs and they train models like Neotron. We're not the biggest fan of these models or this software.

you know we think there should be a diverse ecosystem but the idea that Nvidia does not use their own GPUs for research that serves the purposes of their company is you know itself wrong and then the dynamic that we've seen is that this amount of compute is actually not a meaningful portion of Nvidia's revenue it's it's a very small percentage of Nvidia's total revenue that is you know going back for their own internal research workloads so what we're seeing is just that they are supporting Neocloud's build out the capacity for the demand from the customers which is so incredibly strong right now and hard to serve. Um, and then maybe the the last thing to say about this is that it's not just the chips. Like we we we track a whole wide range of things that they're backstopping. This is everything from like, you know, the construction costs, the memory suppliers, the people building fabs, the people building industrial equipment.

I mean, it's not just like I need help raising money for my GPUs. is I need help raising money for all of the other things in the supply chain which is just the long pole in the tent for that specific project.

Host: Yep. Let's talk about security because um you know the first image in this report uh is a quote tweet of an Ilia subsever post which was made right after uh the hugging face incident. Uh Ilia says neoclouds have limited cyber security. Next time agents successfully go rogue, uh they'll try taking over a Neocloud to run more copies. This is bad. Thus, Neocloud should greatly strengthen their cyber security and every company with strong cyber models uh should help with that. This is a rare tweet uh from Ilia. Dylan, what do you make of this? Uh is this sounds to me like if he's out saying this right, remember he's building this stuff himself. Uh I'd say super intelligence that we might have a problem here. Yeah. So, we we posted an article titled most Neoclouds suck at security. Um documenting,

Guest: by the way, the art for that uh is is incredible. We'll put it on the screen uh for those who are watching this guy getting

Host: Yeah, exactly. getting stabbed by everybody and it's like it's SOC2 compliant.

Guest: Um but yeah, so we we posted that report and then uh that you know like maybe eight or eight hours later or whatever, Ilia uh posted his tweet. um because they are very much um you [snorts] know using various clouds um they're obviously super secretive um and care a lot about the secrecy of like what they're working on but also they are big believers in super intelligence um you know I guess safe super intelligence SSI they were the first you know Trump actually copied Ilia Ilia's first with another one right um first to reasoning first to uh LLM scaling and now first to uh super intelligence branding. Um, but Ilia sort of tweeted that out. Um, and and it sort of echoed what we wrote, right? Which is like, look, most of these clouds are terrible at security. You can get open- source models to just hack Neoclouds or get access to stuff within other within Neoclouds.

Um, you know, GLM 5.3 especially is like really good at this. So, open source Chinese models can just hack dozens of the Neoclouds that exist out there and you can go look at what other people are doing. And so we've seen stuff like this um at other clouds. There was a cloud that you know we we didn't even intend to, but we just accidentally ended up like seeing all of the other clients on that cluster, seeing what stuff they had in storage, see what stuff they had in slurm, etc. And there were like intelligence agencies, right? Like on this. So we immediately like reported it to them. We're like, yo,

Host: which intelligence agencies?

Guest: We we cannot say I don't think that's a valid thing to say. national like like federal well security agencies governments

Host: they were not American

Guest: they were not American but we said it was for a country that was in a has a top 10 GDP it's not like some tiny country

Host: I'm I'm losing my mind here that that was possible

Guest: yeah and it's like it's like if we were malicious and by the way there were probably a dozen other customers and maybe one of them was malicious we could have just like seen what they were training right and they were training models and running models for their themselves presumably um because this was a GPU infrastructure. So what else would you do with a bunch of Blackwell GPUs?

Host: Yeah, we we throw around this this word sovereign AI and it it gets entangled with the concept of a neocloud many times and and people even when they're raising money, they use this word a lot because they view compute as infrastructure. And I think that's exactly right. This is just like the telecommunications buildout where this is this infrastructure is viewed by the sovereign nations that are deploying it and are involved in this as a asset that the country would like to control and have sovereignty over and they believe that similar to what you're saying at the beginning of the show uh they want to be a net token importer or exporter and that is driven by the infrastructure they build which means data centers with GPUs inside of it and so I think it's actually no surprise that uh federal government research and intelligence agencies are on some of these clusters sharing infrastructure with others.

You would like them to have an air gap data center, but if there's exactly one facility in the entire country with GB300, NVL72, I mean, it's a sellers market.

Guest: Goodness. Uh all right. Well, there goes my sleep tonight. So um let's let you know you you you do in this report also rank uh you know a lot of these hyperscalers um into tiers and and the neoclouds into tiers and the at the top tier is coreweave and nebus um I'm more interested in the second tier and that is Oracle and Google cloud um mostly Jordan because you know I I you know I have like in the document that I put together for you know our conversation today I have like a lot of like topics and sentences um that I've like pulled out of the report with questions and then with Oracle, I just have Oracle and then a bunch of question marks because that's kind of how I feel about that company. I don't know how to feel about about them there. Obviously, their debt has been rated um you know, not well by the rating agencies.

They're making this all in bet. Um they rate pretty high and they actually in your report have been credited with pioneering some techniques. Um, but I would love to get your readout on on Oracle and you know whether whether this has been a smart bet for them and whether you know they're they're going to to rise in glory or or sort of fall in flames like a lot of people anticipate. I think the key thing about cluster max is it is not stock recommendations, right? Um it is not anything to do with hey this Neil cloud is going to build a gigawatt next year or 200 megawws u which is far more relevant for the stock than you know what how good are they at operating a cluster right um has nothing to do with hey this data center that they said they're going to deliver with this much capacity next year got delayed by 3 months that's not what cluster max measures cluster max is explicitly measuring sort of how how good is the cluster for someone to use if you rent the GPUs um how reliable is it?

How secure is it? How's the storage? How's the network? How's the life cycle? How's the slurm on Kubernetes? How are all these things? Um in which case, you know, it's just sort of I we try to be more objective. Obviously, making a rating pyramid is like, you know, it is it is something you kind of have to do. It would be nice if you could just make a long report and no one cared if there were no rankings. I mean, that's what would happen. So, you kind of have to do the rankings. Um and and set certain boundaries and you can always litigate the hell out of why and what is in what tier. Um but like it is an objective fact that we think Oracle is one of the top Neoclouds in terms of quality of service. Now that doesn't mean on the other hand, hey at semi analysis in April, we published to our clients that their New Mexico data center was going to be delayed because they weren't going to get the uh pipelines.

And this was a big issue for OpenAI because OpenAI has contracted a huge data center with them in New Mexico. But if these pipelines don't get built, this data center can't go up. Um and then later you know we said it again in uh May um and June and then most recently Oracle announced a force measure on it and their stock tanked because of it right so so we're disentangling two things right yes Oracle is an amazing cloud at managed for their managed cluster one of the top Neoclouds but they also at the same time may have execution issues whether it's regulatory in the case of this New Mexico data center where they can't get the pipeline there or other issues you know on their balance sheet right and so that disentangling like what is good or bad for the stock versus what are we trying to test here.

The goal in first cluster max was

Host: create something that people renting GPUs can use. Um and it turns out now like there's a lot of other uses for it that people try to hamfist in whether it's like you know um hey we're going to write debt for this NeoCloud. If they're new if they if they upgrade rankings from bronze to silver their rate will actually go down. Um you know things like this or like hey I'm going to invest in stocks that are at the top of cluster max. Um so obvious these are not the decision these are not what cluster max is designed for. Um it's a data point that you should use if you're doing anything with these companies. But um as far as Oracle as a company um there's sort of like you know two things to go here, right? Like one is they signed a ton of AI data center deals um with OpenAI. Um was it last year?

Yeah, it was last year. Holy crap, it was only last year. Feels like eons ago. Um I think it might have been exactly a year ago. Um but the interesting thing is they signed these massive deals. Um and there's a number of issu risks embedded with them. One is they signed them with a fixed margin essentially. Um so now that the price of GPUs has gone up a lot for clouds, they can't take advantage of this higher margin. So they're kind of hammed in [clears throat] there. Number two, um they raised a bunch of debt and then the market started getting freaked out about how much debt companies can have and so they spooked out even though Oracle is completely solvent. And then number three, they have some execution issues. You know, I mentioned the regulatory one. There's some other data centers that have some delays and such there as well, but the main one being this, you know, such as in in Wisconsin, but the main one being this New Mexico site that is delayed potentially multiple years because of this pipeline issue.

Um, and so, you know, these things can have very negative impacts on the stock, but I think the company is still fine. Now, you know, they're not Nebius. Nebus gets to turn around and they had the they have the shortest average contract length of the top clouds in cluster max. Um, and so Nebius has been able to turn around and really jack up pricing and take advantage of this, right? Whereas coreweave and Oracle despite having also good cluster um you know management and clust and reliability and security and all this stuff were not able to take as much advantage because they lock themselves into long-term contracts primarily.

Host: Oh that is fascinating. All right Jordan let me ask you a technology question then. um you know as Dylan put out uh put out there it is a pyramid um and there's like these top tier uh groups like we talked about coreweave and nebius Oracle Google cloud is up there but in the notrecommended section there's a lot of logos uh there's underperforming and unavailable um I think there's more something like more logos in the unavailable than you know top four tiers combined What do you think that says about the fact that I mean, what do you think that says about this technology that there's so many that are out there uh that just aren't meeting the bar for you?

Guest: Look, as people who are trying to represent the interests of buyers, I think it just makes us sad. It's sad to see the uh

Host: look it's sad to see that like a a year and a half almost you know since um the first version of cluster max was published actually approaching 2 years now um and since we coined the term neocloud at semi analysis that a lot of people have just not made progress towards the things that we were writing about and the things we were writing about is just exactly what we talk to the customers of these neoclouds about what they Um, and so I think it's just like the dynamic of it being a sellers market means that the priority for these companies is how can they raise as much money as they can, how can they build data centers as fast as they can, and how can they deploy chips as fast as they can rather than working on the true quality of the underlying service. And that's that's a little bit of the dynamic that you see represented there where people can have these great businesses without actually worrying too much about delivering a great service that their customers really enjoy and recommend to other people.

The other thing is that I think there's like this constant dynamic that we'll continue to see over time which is that there can only be so many leaders because there's only so many people who truly innovate and then everybody else copies them. And when you have a market that by our tracking is actually 323 providers as of this morning uh which we track in our AI cloud TCO model for people who are subscribers to semi analysis they have this full list. there's only a certain amount of them that are even relevant, like large enough to put on the list and then a certain number of those that can pass the bar for what we believe is what we need to see to to recommend them. And you know, in some ways, we think that the bar for being recommended is quite low. like the um the basic things that fail with respect to security and reliability being the two top ones are just um in our view dead simple and it's it's just sad to see that others aren't paying attention enough or caring enough to actually implement this stuff.

Guest: Okay, two quick ones before we go. Um just more broad question. Uh you know Google obviously they're they've announced they have this Gemini uh Argon coming out. Gemini for Argon. Um, is it your view that that you know there have been some people who have called upon Google to just be, you know, an basically a NeoCloud and rent out uh the TPUs from Google Cloud and the computing. Um, do you see them going that way or do you think that they're still going to use like a large percentage of their compute for model training and inference? I mean definitionally they have allocated less compute to uh model training and deployment um in the last 6 months than they had two or three years ago. Um and that that percentage allocation has fallen over time. Um this is sort of you know one of the things that we made a joke about like uh maybe not even joke but just a statement that Thomas Curran won and Dennis lost right?

Um now I guess Karai um Google is producing significantly more compute. um you know, year on year, but the amount that's going to deep mind is not rising as fast on a you know, it's not like it's not like oh Deep Mind had a third of Google's compute, now it still has a third of Google's compute. Um it's actually falling as a percentage. Uh which is really interesting because now if I look at just DeepMind versus Anthropic and OpenAI, DeepMind has the least compute of the three um by the end of this year. So like it's it's a really it's really a big deal. Um, and so Google, you know, whether or not there's a Neil cloud, what the word Neocloud even means, uh, you know, does selling tokens of other models count as Neocloud? Um, you know, does does, you know, in a sense that's what the new cloud business is, is selling AI model tokens, right?

So maybe that's a neocloud, right? Um, the definition is

Host: Blackstone and going and doing new construction projects to build your GPUs to deploy directly for customers outside of your core platform

Guest: or selling chips, right? Is that neocloud? No. I mean, they're selling a lot of chips to anthropic. So it's, you know, physically selling the chips rather than like just renting them out. So I would say like Google's business model is evolving and they want to make money at every layer of the stack. Um, but by definition, you know, helping Enthropic this much with TPUs, um, and helping them this much with these other things is hurting Google Deep Mind's business. Um, okay.

Host: I think they need they need to find a niche, man. Like for Gemini, I mean, nobody wants the third or fourth or fifth best coding model. So they need to find a niche that drives Gemini token sales if they want Deep Mind to continue competing on the frontier of AI research and because right now it's I mean even with this Gemini 4 argon release it's just not there on coding compared to the top two or three

Guest: and and and the risk here is what happens next year when Open Entropic both attempt to add 5 to 10 gawatt of additional compute for their labs and you know for open eye andropic each and then Um, Google only adds a few gigawatts. Now their compute deficit, it it was Google was way ahead in compute and they were kind of behind slash equivalent. Um, mostly just a little bit behind. Um, and now they've fallen further behind, but their compute is the same. So what happens when they have less compute? Are they do they somehow magically catch up? It's possible, right? They have the talent, they have the people, they need to reorganize and really focus up. Um but also like more and more of the best people just keep getting poached to anthropic to open AI and it it's it meta

Host: discovery loop.

Guest: Yeah.

Host: Discovery loop.

Guest: Discovery loop. Right.

Host: Okay. I I want to ask one more before we go if that's okay. Um you know the I think Dylan you've said in the past that um the the people who are getting most of the value out of the AI models that are created are the end users, right? The people that are using the clouds for instance at the you know at the base. Um, and of course, Anthropic makes that model available via API and to go into all these various uses. You know, it's it's going to have to do more licensing and and who knows if they'll be able to do products for all this. I'm curious if you think if you see a scenario where given the big numbers that we talked about in the past um or they keep their frontier models at home for longer, much longer, and use them to build their own products. It's like we know they have the wet lab, we know they have cloud code uh and co-work um and then they don't make them available because they want to get more margin out of the infrastructure that they've they've bought and built.

Guest: Yeah. So, I mean I think I think we've already seen this, right? Um Anthropic had Mythos ready in February. Uh they've their model safety cards not noted uh noted that they have two other models since Methos that have released and publicly we still don't have full Methos. We have a neutered version of it called Fable. Um, and it it took many many months before Fable even came out. They had the glass wing which was like mythos but for only some subsectors and it was neutered in some areas, right? Not all areas. And so it's clear that Anthropic has not released new models. Open eye. Likewise, Astra was actually done a couple months ago and it only came out like, you know, it took a couple months before they released Astra. Um and the new model bell um they even showed a blog where they showed the performance of it in math um and how it you know solved all this stuff um and they have not released it.

So clearly they are just using it internally not externally and and so you know the labs have already moved along this line right where it's like oh yeah yeah the gap between them and and Chinese open source is still 6 months um they're kind of keeping that but in reality they've kept you know months of progress internal only um because the moment you make it external people can distill it um you know and all these other things so I think you know enthropic will likely release a new model in the coming weeks you know for their IPO Um I think open I will likely have to release a new model in the coming weeks as well just because it's such a competitive environment but at the same time they are like keeping a lot of their stuff internal only. Yeah, we were discussing internally some of the uh capitalistic incentives.

If if the incentives are actually going to motivate people to release more models and get them in the hands of consumers because of the incredible gross margins on the per token API pricing or if the incentive is to keep it internally so that you can pursue the end user applications which are even more uh high value than selling the access to the model per token. It's going to be interesting to see how it plays out. But in one of those scenarios, the capitalist approach will be to pace the frontier for yourself, right?

Host: And which what's your gut on where it's going to go? Which one of those scenarios you think plays out?

Guest: I think Anthropic has publicly stated that they are pacing the frontier by keeping themselves uh restricted. And that's the only way that their blog posts have manifest in reality. Nobody else is slowing down. They're only slowing themselves down. I think that's true. I think that's what's happening. And uh I think the incentives to both improve the status of AI by solving things that are, you know, like cancer that are just unequivocally good that everybody can agree is a good thing would be, you know, a positive incentive. And then I think there's also plenty of money to be made if you do something like that. So I my my bet is that we continue to see the minimum viable models released by both openai and anthropic to keep a lead over Chinese and the rest of the open source ecosystem. The website semianalysis.com uh sign up for the blog as Dylan says, you know, sign up for the newsletter as I have.

Uh it's a wealth of insight and analysis and I think you'll love it if you're not there already. Uh and this conversation has just, you know, been the manifestation of that. So, uh, Dylan, great to see you again. Jordan, really nice to meet you. You guys are welcome anytime, and we appreciate your time today.

Host: Thanks for having us, Alex.

Guest: Thanks, Alex.

Host: All right, good stuff. Thanks everybody for listening and watching, and we'll see you next time on Big Technology Podcast.