MBI interviews Asianometry

Before we begin, a disclaimer. This

podcast is forformational and entertainment purposes only and should not be relied upon as the basis for any investment decision. Nothing you hear is

investment decision. Nothing you hear is an offer or solicitation to buy or sell any security. The securities discussed

any security. The securities discussed on this podcast may be owned by persons being interviewed. Before making any

being interviewed. Before making any investment decision, please consult an investment adviser.

So, Never Sale has so far been about uh just Scuttleb and me talking about like all sorts of things and uh we're going to switch it up uh just a little bit this month. Today I'm going to do like

this month. Today I'm going to do like an interview and the person I'm interviewing is John and we're going to talk about the topic dour

semiconductors. John obviously runs the

semiconductors. John obviously runs the famous Asianry channel uh where I spent myself countless hours in the last couple of year a couple of years since

I'm one of those generalists who basically woke up post chat GPD and realized uh it is not quite an option uh

to uh to to not pay attention to uh semiconductors. So, I I'm thrilled uh to

semiconductors. So, I I'm thrilled uh to directly thank and welcome John uh to Never Sell podcast. John, thank thanks thank you so much for joining us today.

>> No, it's a pleasure. I'm really excited to be on. I'm a huge fan. I listen all the time and uh excited to uh let's talk let's talk some let's talk bananas. Just

kidding. We'll talk chips.

>> Right. Right. Right. No, I know one of the fun facts that I have kind of learned about Asian uh from your interview uh with strategic

like I think few months ago. Uh so as started as a tourism channel, right? And

>> uh yeah and and obviously I also came to know that semiconductor is in your genes. Your father used to be analog

genes. Your father used to be analog chip designer. Uh so it makes sense that

chip designer. Uh so it makes sense that you uh your channel has become this kind of super nerdy with dry sense of humor and you've been covering semiconductor

long before it was cool. Uh so

uh so yeah like you know there's there's obviously you know semiconductor has been uh like I said topic dour over the last like you know 3 four years. So we

can start this conversation from probably multiple angles but I'm just going to go like along the value chain uh for this conversation. So let's start

from like semicap companies uh uh like the company I would pick probably talk more granularly uh is ASML. Uh and you

know it's funny because ASML was one of those used to be one of those companies within the semiconductor industry uh which almost appeared bit of a loser up

until like mid 2025. There's a lot of you know risks people used to talk about like peak lithography you know China will be able to figure it out all sorts of stuff right and as kind of you know

the stock has kind of doubled more than doubled I think from mid 205 to you know now uh those bare cases appear to be

somewhat you know disappearing I'm pretty sure when the stock goes down those will come back again so uh maybe they're still relevant you know but I'm I'm curious about a particular risk that

you kind of alluded to. I'm not sure whether you would consider it risk but I'm I'm curious to kind of you know know more about you said I think in in one of your interviews with someone uh that

ASML machines are sort of overengineered and I've noticed Ben Thomson also uh appears to have a lot of sympathy towards the argument because he also

mentioned in one of his sharp tech I think uh episodes that uh he also feels the real worry is not just that China figures it out but they figure it out in

an kind of Elon Musk way which is basically a system the system fails more and is less reliable but it just engenders a better overall process that

scales better and avoids astronomical upfront cost. So do you think there's a

upfront cost. So do you think there's a real uh case there or it's just more of a kind of a general common sense bare case but may not be as relevant in this

particular case for this like in incredibly complex machines that ASML sells to you know TSMC's and the uh and and the you know other other fabs of the world.

You know, it's kind of interesting. I

think, you know, ASML, and when we say ASML, we could think about kind of the photoiththography paradigm, which uses light, increasingly

smaller wavelengths of light to transfer patterns to a wafer. I think that paradigm has been challenged several times in several decades. I think Ben

Ben talks a lot about kind of the concept of maybe a disrupting force but like you have to kind of look at it from ground state principles in my opinion

and I think fundamentally if you gen what what I've seen is that next generation lithography that is not photo lithography tends to have gamebreaking

traits related to not necessarily the it's not it's not like it's physically impossible to do these things without using a different type of lithography. But I've always

felt that it cannot be done in a way that is more economically viable than the current system. So what I mean by that is that like the UV the low NA UV

system >> is probably not that overengineered if you think about it. M

>> the high ina UV machine is probably pretty over engineered and that's um that's defined that is you can actually see uh you can see that in the fact that

TSMC won't buy it >> right they they refuse to buy it >> um and I think you know high AEV's most competitive the most fierce competitor

is its own product um >> I I think the low AEV system is is wonderful I think it's uh it works really well the ecosystem is really

strong and I think with the forthcoming light source like don't look at the sticker price you look at the throughput you look at the economic productivity

you look at how much that the VAB is going to make from running this machine and you can see you and with you know ASML recently announced that they're

going to put in the threeshot light source um and that's going to be supposedly according to them going to raise output by maybe 50%. And I think

that's uh I mean that's a that's substantial and that makes the tool much more productive. And then I think you

more productive. And then I think you know what are you what does a disruption what does disruption look like in that sense.

>> Um I don't think that's necessarily something anyone has answered yet. They

just say disruption.

>> Right. Right. What is the ASP differential between the low NA UV and high NA UV approximately?

>> I mean it's hard to say. I think if we can say the low NA tool is maybe now 120 100 million. The high tool is maybe 300

100 million. The high tool is maybe 300 350. I mean they're never going to tell

350. I mean they're never going to tell us but they're trying to slim down the high EUV machine. And I think uh it they'll probably get there. I mean TSMC

will eventually buy this machine but like when we don't know. Uh I'm hearing a while from now. The low machine works great and I think it's um it's going to

work even greater and I think not a lot of the layers within the chip need high EUV and that's not going to change for a substantial amount of time.

>> Interesting. So you're saying the low NA EUV cost like you know let's say roughly $100 million versus high NA cost like 350. So that's a substantial ASP

350. So that's a substantial ASP differences and >> we don't have any compelling evidence so far at least yet that not opting for high NAV

uh hasn't really cost much uh for TSMC in terms of losing market share uh to anybody else or like you know or you know fe fearing about their own

disruption if someone else makes a great use out of like you know high AUV. So,

so that ASP differential at least appears to be quite substantial. So, if

KSMC kind of drags their feet, uh, that that might be a bit of a headwind for ASML's kind of continued revenue uh, growth, don't you think?

>> No, I don't think so. I think like the low tool is probably very optimized and the ecosystem probably works pretty well. They're building this at scale and

well. They're building this at scale and I think with AI with the AI boom that's going on they're going to buy a lot of low NA EUV machines and to be frank that's kind of what you want to sell a lot of because that's probably where

you're going to have the highest margins the most ability to kind of put this together and get this out the door high machine is big it's heavy it's not ready yet the R&D I mean it prints but the

ecosystem I don't think is ready and I think >> more than just the price than the sticker price of the tool It compro there are technical

compromises that make it unacceptable for TSMC and even Samsung and any fab because the field is twice the size or is half the size. So it's three times

cost it costs three times as much. It

prints a smaller field half the size. So

it's less productive.

>> It gives you a better print. No one

denies that. But the the compromises on the economics just make it really difficult for a fab to say, "Look, I want to spend this much for it." So like

they'll keep it in R&D. They'll keep

working on it, but this is kind of like a this is going to be a very long-term play. I think people need to understand

play. I think people need to understand that high AUV is not going to be some savior that comes down ramping down.

what will be the savior is this AI boom which will mean TSMC is going to buy you know maybe that many more EUV tools >> going down the line.

>> I see. I see. So you're saying even though there's like a substantial ASP differential between high and low NA EUV, the volume uh of low NA and EUV and

how mature the ecosystem around it is uh kind of compensates for the price differential and the margin that ASML gets is still pretty good for for low

AUV. So even if it takes a while to uh

AUV. So even if it takes a while to uh for for high AUV to you know gain adoption, it's totally fine for ASML.

And in any case, it's not like ASML itself is facing some sort of competition for even even for the low EUV uh machines. So even if their

customers kind of drag their feet, uh ASML is still sort of like somewhat agnostic about the adoption rate. Is

that is that a fair characterization?

>> They should be. I mean I mean they should continue pushing the boundaries with high EUV and TSMC owns two machines I think in their R&D fab. So, like

they're still working on it. They're

still working on it. I think the thing that people should uh I think the thing about ASML is that they need to >> is that they're they're scaling up to build more and then they're going to

keep I think they need to keep shoring up their low UV line as well as their DUV line. I think the DUV line is the

DUV line. I think the DUV line is the one that is is especially challenged because there's actual real competition there.

>> There's real competition from Canon, uh Nikon, though. So those are real

Nikon, though. So those are real companies and they have real products.

So like um the margins I mean I I would urge the company not to leave that space. I mean the financial investor would be like oh you should leave that space leave that to the lower

maybe the margins are not as good but like um that's that's kind of where you need to you need to shore that flank to make sure if disruption does happen whatever it may be

>> or maybe price disruption doesn't happen you want to make sure that you have a particular place there. So, um, that's that's generally I don't want to give advice to ASML, but like it's they're

multi-billion dollar company. They don't

listen to me, but like I I think their DV lines are probably where the the bulk of the of the conflict should be. They

would be focusing really hard on making that cheaper because there's a lot of competition there.

>> Interesting. Interesting. Uh, let's move on to SML's customers. Uh, you know, well, one of the customers, major customers, which is TSMC. Uh so we kind

of all know that you know Intel fell behind TSMC in early to mid2010s and has never been able to quite catch up uh in being on the leading edge nodes. Maybe

can you just kind of you know explain uh the process nodes and how they work and why it has been so incredibly challenging for anyone else to not just Intel but also like Samsung to catch up

Samsung uh to to catch up with TSMC. I

don't necessarily think that TSMC surged ahead or anything or I don't necessarily think that they sort of, you know, blew people out of the water. I think the thing about process nodes, if you break

it down, the process node is essentially there's a much more technical phrase, but like it's trying to build a certain size of a transistor. Mhm.

>> And I think that like um along with connecting them, putting them in metal layers and stuff like that and I think what happened with Intel and with the

Intel was that they kind of they did not they they were too ambitious especially at their 10 nanmter node what they called their 10 nmter node. They were

too ambitious and it didn't work. Um

they it didn't yield and it didn't work.

TSMC is has because they have so much more volume. They have Apple volume, Qualcomm

volume. They have Apple volume, Qualcomm volume, Broadcom volume. They have all that coming through. they're able to fund smaller half nodes which are not as

ambitious but um they basically kind of get you there closely cuz if you think about the process node it's a whole production line and you're just squeezing budgets out of each particular

part of the line to make sure in the end you have a you have a chip with tighter tolerances or a better formed transistors or better

formed wires or something like that to benefit fit that cuz in the end the node number hasn't been doesn't convey anything other than just >> it's just a marketing tool. It's

supposed to tell you it's better >> but underneath that number what actually tells you is that that performance differential can come from a variety of different ways. I think the idea of

different ways. I think the idea of thinking just making the transistor smaller is very outdated >> because it's not exactly it's it's far from the best way to give you to give

gains. So like there's all sorts of

gains. So like there's all sorts of different things that are done and it's not being conveyed in that number. So I

think like um that's that whole holistic view and um you know the semiconductor industry is really good at that at really at finding those new options and ideas and integrating them into the

process node. Your transistor might be

process node. Your transistor might be the same size but it performs better and that is actually more important than having a necessarily smaller transistor.

Um I hope that makes sense. And then to answer your question as to why it's been really difficult to for intel and >> it's a it's a it's like a Japanese the Japanese have that phrase right the

where you have like a mastery at your job kind of thing.

>> If you do something for like 30 50 years you kind of really really know it.

>> Yeah. And I would say that like um you can I would say that like when you do something for 30 40 years, you're always understanding and they're always

planning ahead to get these nodes on on track. So I think the hardest part is

track. So I think the hardest part is for Intel and any other company that's trying to catch up is that you're trying to move ahead. You're trying to leaprog

a moving target. So by the time you know a whatever A8A comes out, TSMC already has N2 and then when

>> uh they're going to if you want to just jump ahead of that then you need to TSMC already has A14 and A16 coming out. So

like it's >> that's the hardest part of moving forward because like um the the target is forever moving for now.

So when you said Intel was too ambitious, could you kind of elaborate like what do you mean by they were too ambitious and which was partly the reason why they fell behind?

>> Intel is a very complicated company and I think like one of the things to best understand is that they had two ways of leveraging of squeezing out more

performance gains from their CPUs throughout the 1980s and the 1990s. And

that was a combination of architectural improvements on the CPU as well as improving the node itself meaning shrinking the the Moors Moore's law as they say. But for a long time they talk

they say. But for a long time they talk a lot about Moors law and Moors law is important but >> there was a lot of gains came out from just making a better CPU. Those gains I

would say ended with the in 1995 or so 1995 1996 with the Pentium Pro and then from then on Intel had no choice but to

kind of go hard at Mo's law which they did and they were very successful in doing so for the next 10 years or so.

But after that, after you had the implementation of 193 nanometer lithography, the only option was to say wait for EUV, right? Because they were building up UV

right? Because they were building up UV there, but UV wasn't ready then.

>> So what happened was that they were going to say we're going to multi-attern. We're going to multiatter.

multi-attern. We're going to multiatter.

We're going to rerun these.

>> We're going to run multiple masks and all that. And I think at that time

all that. And I think at that time >> the technology was just not ready. uh

ASML wasn't ready.

>> Um the Nikon definitely wasn't ready.

And then not only that, because they were Moore's law, they had to say, "We need to do a doubling. We need to shrink this. We need to shrink this at a at a

this. We need to shrink this at a at a really high pace. At the same time, they're adding all these new materials in." Right? So like um they did this new

in." Right? So like um they did this new uh half oxide for bringing all these new material. And the things about new

material. And the things about new materials is that TSMC gets really scared >> when you add in new materials into the fab because you don't know how they interact with the rest of the materials.

>> Right?

>> My best guess is that when they jump to 10 nanometers, they said, "Look, we're going to shrink this. Do a full shrink.

>> We're going to add these new materials."

I heard cobalt was a problem. And then

in the end, it's gonna in the end it just didn't yield. And when it doesn't yield, they won't say it, of course, but the products won't arrive and then >> yeah, >> it just keeps going. It's going it's going and they keep extending the old

products. And I think that was one of

products. And I think that was one of the um the major problems. And I think maybe one of the things that made it so long that kept it so extensive, like why

it was such an extensive >> backup was that >> because they were part designer and part manufacturer, they also had to redesign the chip

because >> so like that's the sort of thing that like takes two years. So you need to

these things pile up and um yeah it was it was a it was a raging dumpster fire as my father likes to say.

No, when I think about like you know both ASML and TSMC and kind of you know let's say competitive advantages uh I almost sense like an incredible

similarity in the sense that uh you can't really get to let's say EUV or high AUV without uh accomplishing the

intermediate steps to get there right similarly like TSMC like you can't really get to 2 nanometer uh node without really accomplishing the intermediate steps in the leadup to that

is that like do you think there's this similarity like you know it's really actually very hard to compete against these comp companies for their most advanced

machines or nodes and which makes it like just super hard for anyone else to actually catch up with either of these companies.

M I think it's just the way the semiconductor industry works that they're very incremental and I actually say Intel is sort of the different company is different in the sense that they do tend to really go for

breakthroughs. They really tend to like

breakthroughs. They really tend to like they really swing hard at big things. I

think I just finished a video on Itanium and Itanium was a was a debacle, but it was ambitious and it was wonderful in that in this ambition and I think people should appreciate that that in the

semiconductor industry where everyone only takes one nanometer at a time.

Intel comes in and like we're going to try this whole new thing.

>> It's going to cost $6 billion, but we're going to fail with it. But like we tried.

>> We tried.

>> So I think that's sort of like we shouldn't laugh at them for that. No,

you can tell it's an American company, you know. It's a it's ambitious.

you know. It's a it's ambitious.

>> It's ambitious. It's ambitious. I mean,

their uh their abilities were not up to the ambition, but the um or I mean, their prior there's a better way to say this, but like, you know, they they I like that they are ambitious, and I

think that's something that TSMC isn't necessarily good at. I think like um and ASML as well. You know, hardware is hard. So like um when you're working

hard. So like um when you're working with something that has real physical you can't just recompile it and then it's done in 30 minutes or Yeah. So I

think that has that inflicts years of trauma on people and I think software people in Silicon Valley don't understand that well enough.

>> Yeah. I mean so some of the ambitious people are arguing now that TSMC is not investing aggressive enough. So

obviously you know their their TSMC's capex was lower than it was in 2023. I

mean it's like 30 billion versus 29 billions but still lower 1 billion lower in 2024 versus 203. Uh since then they did increase it to like 40 billion in

205 and I think they got it to like 52 to 56 billion in 206. Obviously

it's really hard to you know say with a straight face that that's not a humongous number but obviously their customers customers which is like you know the hyperscalers they're spending like $700 billion of capex and constant

is saying in the earnings calls that the demand is outstripping supply demand is outstripping supply right so uh do you do you sense that TSMC is actually not you know being aggressive enough or

investing enough uh and kind of not capitalizing on on on these moment.

>> I know Ben Ben's really on to this uh really on this. Um I mean I've I've talked about it before and I think the there's a lot of reasons and I've spoken

to TSMC people who basically say look >> we do things our way like what will happen their MO is generally they'll say we'll never use this we will never do this. They'll say nothing until and then

this. They'll say nothing until and then meanwhile but behind the scenes they're putting together information. They're

thinking about stuff and then it goes up and up the chain until one guy usually it's Morris but someone now it's maybe CC or someone high enough is saying okay we're going to do this and then suddenly

they'll flip 180 and suddenly they're like all in >> like tool vendors will be like they will never buy this. TSNC will tell them we'll never buy it. We'll never buy it.

We'll never buy it. And then one day we'll say we needed yesterday. we need

10 tools, we need it yesterday.

>> Um, that's how they operate and like I think that they're >> they're generally it's like it's something flipped and I think like they take time to cut to understand they do things at their own pace. They need to

figure things out. It's a massive supply chain. It's not just TSMC. It's like

chain. It's not just TSMC. It's like

everything needs to come together at the same time. So, it's just tools,

same time. So, it's just tools, equipment, labor, land, all water, electricity, all that needs to come together in one place to get ready. And

that just takes time to set up.

>> At the same time, there's like you need to recognize that this is uh this is this is a cyclical business. People get

smashed all the time. I think uh TSMC got smashed pretty hard. They people

told them that everyone needed chips, automotive MCUs for COVID and then suddenly it was gone >> and um you know it was suddenly >> it's crazy. It feels just just the other other year you know like kind of couple

of years ago basically and uh all of a sudden now everyone is saying why are you not investing enough uh even though in this kind of cyclical uh scenario already played out like you know 3 four years ago

>> literally like three years ago. Yeah.

Yeah. Literally like the company was turning a loss right it was the losing money on its 7 nmter nodes >> which were very relatively advanced and they were I mean it was it was dumb. Um,

it's dumb that people have such short goldfish memories, but like >> those TSMC people remember and the their suppliers remember.

>> I think um I did a video about testing, right? Pterodine and Advantest, right?

right? Pterodine and Advantest, right?

>> And during the com bubble and the telecom bubble, they they really got big on like optical tools and all that. And

they I mean not just the optical not just the dotcom bubble but prior to that they would have these amazing swings where up 30% up 40 down 40% up 30% down

40%. Because that's just how because

40%. Because that's just how because that's how like the the supply is that's how it's done. And like it abused people like you would hire hundreds and thousands of contractors and then fire

them maybe the next year.

>> And um >> that's rough. And I think that's that's something that these companies, these supplier companies, they they're they're very cautious and like you can, it's not

just TSMC. You can go to all of them

just TSMC. You can go to all of them except maybe Intel and then you can be like, "Look, are you going to can you scale up in like tomorrow?" And

they're going to be like, "We'll see."

So like it's it's just how it is.

>> Yeah. I mean, you're right. Like maybe

TSMC uh shouldn't be investing investing aggressively as they will be the ones holding the bag. uh if AI proves to be a bubble and if it all crashes down uh for all sorts of reasons a AI itself the

technology itself the business models themselves don't need to be bubble like there can be all sorts of like you know scenarios political instability like you know obviously there's a war >> or a deepseek situation right like

somehow we find out >> there you have cloud code using millions of tokens well what happens if a model comes out and suddenly cuts token use by 20 30%.

>> Right like that's that stuff matters right so like um Absolutely.

>> I just remind you, yeah, back in the 70s, we thought we were going to use like 50,000 like like they they drew a line up to the right and then to say we're going to build power plants to provide nuclear power plants to meet all

these power demands and then like it never happened cuz people stop >> people stopped just people became more diligent in how they use their resources and that's how it is. That's pretty

cool. But like um that's how feedback effect works.

>> Yep. Yep. But it's

>> I mean do you do you use cloud code to that extent? Like are you burning

that extent? Like are you burning millions of tokens every day or something like that?

>> I I'm yet to be one of those people. I'm

I'm still kind of you know keeping the AI bubble in check.

But maybe maybe eventually who knows.

>> Yeah.

>> Yeah. But but maybe you know is there a scenario where Samsung and Intel uh who might be able to take advantage of TSMC's uh relative conservatism

conservative posture here or do you think it's it's just still too hard because of all the reasons you kind of laid out before?

It's just really depends on how much you want and how much you want to take that risk right?

>> Uh has like as the chip as the fabulous company, right? Think about from if you

company, right? Think about from if you running maybe a large fab company. Let's

say you're running a startup. Let's say

you're running an AI chip startup and you want to make an AI chip to sell to your customer and your customer needs it very soon.

>> Mhm. And it's a 2year turnaround time from beginning to end. And you're

spending hundreds of millions of dollars on EDA tools, chip talent, all of that, right? If the chip is late, if the chip

>> is bad, comes out and is bad, your business is done, right?

>> Yep.

>> So, if you're Apple, right, you need all these chips to come together. The chips

are an increasingly important part of Apple's competitive advantage.

>> Sure. Like yeah, like just look at the work that they did to get a modem done, right? Like they actually build a whole

right? Like they actually build a whole iPhone around that modem. I think like >> um >> when you have when you're putting that sort of risk on yourself and you're

taking on that kind of risk to outsource the manufacturing to someone generally you don't want to take any any sort of unnecessary risk. And so like I think a

unnecessary risk. And so like I think a lot of people will say oh the strategic people will be like you should diversify you should do all this but you have to

do it in the context of what you have what resources you have and what risks you want to take.

>> If you and this is a game-breaking risk if your fab if your foundry doesn't deliver you're done. And there are examples in history of fabulous

companies that do not um those companies that depend on let's say a particular large company foundry and got absolutely hosed

>> by delays, bad product.

>> Absolutely. No, none of their none of their fault. So you want to go for

their fault. So you want to go for someone who's reliable. Um

yeah I mean that's why I I I had this idea and I think uh some other people also kind of mentioned this idea later

is it is very risky uh to your point uh for any individual customer to go to uh any other fab uh not named TSMC for the

advanced nodes right for the most advanced chips and take that risk. But

if it were more like a consortium. So

let's say obviously the cheap buyers at this point are also very you know extremely concentrated or you know you know has become overly concentrated. So

let's say the top five or seven customers who are the majority of TSM's like you know uh uh chips buyers if they kind of you know coales together and

kind of create this consortium to uh say okay we're going to like you know uh put an order of $2 billion Intel uh chips from Intel uh next year. So if Intel

doesn't deliver we all lose $2 billion each. So it's the competitives like you

each. So it's the competitives like you know dynamic doesn't really change much.

Uh obviously it's hard to kind of you know uh you know create this kind of consortium I guess in the first place but I I feel like it makes an awful lot

sense from this ambitious American company's perspective to kind of at least give it a shot. Uh and if you if they all do it together TSMC can't really punish all of them I guess. Uh,

do you think that's a crazy idea or you think there's there's some uh there's some >> Well, can you name a consortium in history that act an American consortium in history that has worked out?

I would like you to try to figure one out. Uh, I've not heard one. I study

out. Uh, I've not heard one. I study

history, a lot of economic history.

>> Yeah, >> I have countless examples of consortiums that get together and don't work out.

Um, but uh I've not seen one that has that has five or six of the most powerful tech giants who are very used to getting their way. And um and like

also some of them don't like each other, right? Like u I I let's say uh Apple and

right? Like u I I let's say uh Apple and Facebook, right? Or Meta Apple and Meta,

Facebook, right? Or Meta Apple and Meta, they don't like each other and like they they don't necessarily have the best uh uh thing to work together. I would say, you know, Ben Ben's right to say that

there is no economic incentive like like there's no economic reason to like even if like five of the big tech or even two or three of the tech

companies go like go off and create a consortium right?

>> Well, then you have to plan that out a couple more steps. Let's think about it.

>> TSMC says, "Well, we have all this capacity. What we're going to do now?

capacity. What we're going to do now?

Let's offer it to someone else. Let's

offer it to >> let's offer it to one of the other tech giants who hasn't joined the consortium who's most likely to defect and suddenly you know you have it a bu you have a

situation where this consortium which needs at least three to four years to get done like let's think about how fast can you get up a fab and build a fab

>> rapid is the fastest I think I've ever seen any company do it right they took four years they're not going to be in high production till 2027 and end of

2027 and their full speed their full chip wafer capacity is in the low tens of thousands of wafers right that's like

TSMC TSMC has 15 20 phases that do much more than that so like um I appreciate the thought I really do

I I think it's I think it's a very kind thought but we need to plan this out more because you need to plan out more how this things acts in real world because TSMC is very good at

um they're very good at diversifying their customer base, right?

>> Yeah.

>> Nvidia is only 20% of their capacity. So

like they're very clear. It's not if they really wanted to they could say we're going to raise that can double that immediately >> and we can make like then all the Nvidia will get all like we will get all the

Nvidia chips we want but like TSM is very disciplined on saying no we're not going to become utterly dependent on one customer

>> I guess one of the assumptions that uh I'm kind of implying is the demand is outstripping supply right so

>> so that That's why uh TSMC doesn't have spare cap capacity even if some of the demand is routing towards let's say their competitors but also I but your

your point is well taken that it is hard to imagine it's there's all these like five seven companies uh you know the major customers of TSMC

that they will just get along really well and and come together to fund this kind of you know kind of >> wait five years to wait five years in the market.

>> Your point is well taken. Uh so maybe that's why I think uh it only can potentially work if US government forces

them to right like hey you we have we need homegrown alternatives to uh TSMC and yes TSMC has has been building their

fabs in Arizona as well. Maybe uh US government they are now shareholder of Intel as well. So uh maybe they can nudge directly or indirectly that uh

this is the path you guys should take.

>> Well these the US government is big.

Great.

>> Yeah that's true.

>> Yeah >> five tech giants are also very big as well and they have a lot of influence >> and I think the Biden administration has showed what happens when you kind of try to piss off Silicon Valley. So I think

like um but I think what I'll say is that like Intel has real technical reasons for not being a chosen

company, right? I think like um the

company, right? I think like um the company's manufacturing still doesn't have the sort of capacity that is able to accommodate. Think about

a GPU. A GPU is very different from a CPU right?

>> Sure.

>> A but TSMC doesn't just fab GPUs. They

fab A6, right, which are simplified GPUs essentially. They they fab modems.

essentially. They they fab modems. They'll fab >> microcontrollers. They fab like all

>> microcontrollers. They fab like all types of semiconductors, right?

>> It's not just and there's the router chips. There's every types of chips and

chips. There's every types of chips and they're all being done. This all

requires immense ecosystem of all different IPs and and and and different technical techniques to build. Intel has

only really made CPUs like for decades, right? That's their thing. That's their

right? That's their thing. That's their

capacity. That's their manufacturing. So

now you're asking So you're it's kind of like asking five pizza companies to go or >> you have a you have an Italian restaurant, you have a French restaurant, you have a Mexican

restaurant, and you have a uh uh American burger restaurant. They're all

asking them to go. They say you're only allowed to use one kitchen and the kitchen only has uh walks or like chopsticks or something. I don't I'm trying to think of something that's not

that's not culturally insensitive, but it's like it's you only have one piece of equipment. And I think that's

of equipment. And I think that's something that >> like if you're a cook, like you hear this and you're saying, "What the heck? Like what? like I have

KPIs I need to hit otherwise I'm out of a job, >> right? Like like are you willing to risk

>> right? Like like are you willing to risk your job to say to fulfill some vague jew some vague strategic diversification? Um I don't I don't

diversification? Um I don't I don't think a lot of people will put their jobs on the line for that.

>> I don't think they will.

>> What do you think about uh the idea of space uh data centers? you think

building fabs uh in space is is realistic or anything imminent? I mean

there there's a supposedly a company that's coming to market this year supposedly worth 1.7 trillion. So what

do you think?

>> I think it could be done. Why not? Sure.

Why not? Let's do it.

>> Let's do it. Ambitious.

>> Let's do it.

>> Let's do it. I'm It's not my money.

>> Do it. Let's do it. I want to see it.

>> Yeah.

>> Let's see it.

>> I don't care. I mean, like, if he's going to do it, yeah, I'm not investing in that. I'm not touching that with like

in that. I'm not touching that with like a 10-ft pole, but I'm going to I'm not going to watch it. I'm going to be like, "Hey, that's cool. I'm going to watch the YouTube video." But like I think like um

>> I mean I think like Doresh did some interview, I think, with Elon. And I

think >> that part that part I'm like, "Sure, why not?" Yeah, it's there's some things

not?" Yeah, it's there's some things that can happen, but who cares? It'll

probably be five years late, but like who cares? But like it'll be cool.

who cares? But like it'll be cool.

They're trying to figure out some way to make money from from higher utilization of rockets. I get it.

of rockets. I get it.

>> I think the thing that I kind of was not so clear on was like at the very end he kind of asked Elon about the terrafab and even when I'm listening to Elon talk

about terraab. I'm like he's not really

about terraab. I'm like he's not really thought this out yet. Um

>> I think like >> he's not he's gonna he's that type of guy I guess to kind of wade his way through feeling the rocks in the river but like They haven't really thought about terra fab yet. Um, but like uh

space fabs or not space fabs, space data centers, sure, why not? I mean, you stick it in a >> you stick it in a box and set it up in space and then like I'm I'm all for it.

Let's do it.

>> Yeah. Yeah. I mean, it definitely follow it. It definitely tracks uh the

it. It definitely tracks uh the ambitious nature of American companies.

unfortunately without much of a result when it comes to this particular subject which is you know uh you know uh manufacturing the most most advanced

chips but we'll see you know uh you say you say 5 years Elon promised lot sooner than that so we'll see you know and if you're coming in the market with $1 you

know 7 trillion enterprise value and like $15 billion of revenue I mean again you know not >> that SOUNDS LIKE A I I'LL BE VERY HONEST. I I I first uh

you know read this report in Wall Street Journal that u uh SpaceX has like $15 billion revenue and uh they're raising money or something like that like at $800 billion. I almost fell out of my

$800 billion. I almost fell out of my chair like seriously $15 billion revenue and $800 billion of you know enterprise value. Okay. And then like few weeks

value. Okay. And then like few weeks later I read I read that uh well they're coming to this coming to IPO potentially this year at 1.7 trillion. And then I

was like okay I didn't fall off my chair. I I just became numb to it you

chair. I I just became numb to it you know. So so yeah I

know. So so yeah I >> isn't TSMC's market cap 1.9 trillion.

>> Yeah I know I >> that's got wild meta is like Meta has like $200 billion revenue and they're like $ 1.7 trillion

enterprise value. So uh yeah and I I

enterprise value. So uh yeah and I I don't know. So so if $1.7 trillion is

don't know. So so if $1.7 trillion is actually their kind of you know market cap or enterprise value when they come to IPO I don't think market will allow

them to wait for 5 years to make uh the space data center thing a reality. It

probably needs to be sooner than that.

But then again, who knows? It's Elon

Musk's company. Market can actually give them, you know, a longer leash than, you know, they typically tend to do for other like, you know, XYZ company. So

maybe we'll see.

>> I was just doing a video, I'm working on a video about um uh Lum, the fiber optics company, >> and they have a wild history. And like

back in 2000, they had revenue of 1.4 4 billion and their market cap was 156 billion 181 at the peak.

>> I was like, >> "Wow, >> that's what it reminds me." I was like, "That's that's a wild they're and they were losing 100 million in rev uh in profit." So, like that's that's good

profit." So, like that's that's good times there. That's that's fiber optic

times there. That's that's fiber optic fiber optic telecom bubble peak valuations there. I'm excited to I I

valuations there. I'm excited to I I can't wait. Can't wait.

can't wait. Can't wait.

>> I'll buy one share just to see what happens. 150 billion. Okay. 100 million

happens. 150 billion. Okay. 100 million

dollar in losses. Like what's that? Like

it sounds almost nothing at this point, right? Like

right? Like >> and then they bought >> and then they bought another company for $40 billion. And I was like and had 25

$40 billion. And I was like and had 25 million in profit. And I was like, "All right, 40 billion. That's cool. Let's do

it." Yeah. Let's see.

>> There's some similarities there as well.

like SpaceX just bought Noox AAI for I think pretty you know significant amount of money. I'm not I exactly for forgot

of money. I'm not I exactly for forgot the number but it's it's it's it's definitely sizable acquisition. So maybe

there's a lot more similarity than we think.

>> We'll find out.

>> But you know we all know in the end like telecom bubble is not going to be the same as the AI bubble like it will always be something different but like yeah >> that's it for sure. For sure. Uh so

let's go to like you know now TSMC's customers well one of the major customers which is Nvidia and you know Nvidia is one of those companies uh like when you look at their chart

it's kind of mindboggling like they used to be like $20 billion market cap in 2016 and even when Chad Jeopardy was launched it was like $300 billion company >> and now it's like four and a half

trillion five trillion like depending on the day uh So as someone who kind of have been following semiconductors for much longer than it has been cool, I'm

just curious like from your vantage point uh which aspects were somewhat more predictable pre-hat GPT uh that you

you could kind of sense that this company uh is on the cusp uh of being something uh much bigger than you know just a gaming company or something like

that. and and which ones were more

that. and and which ones were more surprising to to you who have been following this company for uh again much longer than like you know let's say three to five years uh I'm just curious

how you kind of you know think about those two aspects like and which ones kind of you feel like you kind of expected to expected them to do and which ones are like even you are kind of blown away like they have been able to

uh accomplish this >> you know I mean I'm the I'm the worst person to ask about this cuz then I should have bought the stock if I really knew what is going to happen.

>> I basically was like, I I I own I own the S&P 500 and that's it. So then I was but I was watching this I was watching going on and I was thinking I was thinking hey I think I think uh maybe

they'll be okay. Um,

>> I think like, you know, Nvidia is this really weird I think like all the mega companies are really weird in their own way. And I think like

way. And I think like >> I think Nvidia is probably the one company that kind of looks more like a cult of personality than an actual company. I think what I've heard is

company. I think what I've heard is basically that they're the whole company rolls up to Jensen and Jensen basically goes hard at ideas and he's he's really

fast at responding to stuff. And I think that's sort of like um that's something that they really work on. I think I generally you've not seen a company move

this fast with certain things. And like

I've was really surprised to see I wasn't necessarily surprised to see that Nvidia has really benefited from this.

They had the A100, the H100 for years, for one or two years before ChachiD came out >> and it was doing fairly well. It was

like it wasn't like a small part of the business. I think like the thing that um

business. I think like the thing that um really surprised me is that they sort of retained that edge throughout years throughout the two three years. Like you

haven't really seen a substantial competitor kind of emerge um to be like a default

like other provider. And I think I think like that's partially and like like it's not exactly like the market doesn't want one. And so like

I think it's it's kind of curious to me how it's kept on going. Um

>> I think you tend to see front runners not retain their front running position over decades over years. And I've I uh I

mean it's something special. It says

something special about Nvidia Jensen to say that they have remained like really big and they really still growing at huge huge speeds going on. Um what do you think about like all their

investments into other companies like up and down the stack?

>> Yeah, I mean obviously they they're kind of doing this like almost every week now. Uh and

now. Uh and there's like you know the one way I think it makes sort of rational sense to

me is so obviously they don't have as much cap capex that let's say the hyperscalers have. They are not capex heavy

have. They are not capex heavy companies. They don't have periods to

companies. They don't have periods to deploy like uh all the cach that they're generating right. So, so they're doing

generating right. So, so they're doing basically two things at this point, which is just buying back stocks, which is like they bought like 40% of their cash that they generated. They used it to buy back stocks. And what do they do

with the rest of the money? Well,

they're investing in OpenAI, Anthropic, XAI, Lum, and like you know, all all the like, you know, uh, bits and pieces across the, you know, semiconductor

value chain. And the way I think it

value chain. And the way I think it makes sort of rational sense to me is ultimately if you know AI is let's say

bubble right then Nvidia's chips demand will also go down so even like you know there's almost no point in buying back stocks uh uh if it's a bubble right then

the stock itself is that's overvalued anyway right so what do you do with with with the cash you're generating well you can actually invest across the value chain to keep uh uh the health of your

customers uh you know uh improving. So the open AI, the anthropic, these are like you know obviously they're super you know fast growing companies. They're in a

definitely hyperrowth mode. Uh but their economics is still somewhat speculative right like we still don't know the end state of the frontier you know model developers where the economics will

shake out and obviously Alphabet is this kind of you know 800 lb gorilla here who don't need to use Nvidia's chips. Right.

Right. Uh right. So if for whatever reason uh Alphabet's competitors in the frontier model developer race uh you

know fall off the cliff that directly you know uh undermines Nvidia's uh cheap demand right so so to an extent it makes

sense why Jensen is so deeply interested in keeping the the health of the ecosystem you know stable to like

improving as these companies or as their the major customers of Nvidia find the kind of endstate economics like you know we there still there's still kind of a lot of question marks uh how how some of

the debates will be settled so so in that sense like you know I I can see where what obviously this will look like they will look like complete fool if AI

turns out to be like you know a bubble or like let's say the way it I mean so how how uh this still be a bubble. So for

example, we now know uh based on different media reports what uh sorry anthropic and open AI are pro forecasting their revenue to be next

year this year then thereafter that 29 to30 we have these data points right so for example let's say if this number if the actual numbers consistently fall way

below uh these estimates that's already out there it will be very hard for open and anthropic you know, go back to investors and say, "Oh, we're now going to raise money at like one $1.5 trillion

valuation." Uh, even though we haven't

valuation." Uh, even though we haven't been able to hit the numbers that we kind of said that uh we we're forecasting. So why are your uh

forecasting. So why are your uh valuation is increasing if you are not being able to hit the numbers that you said you know you're forecasting like let's say in 2025, right? So that that's

a real concern. uh you know and we don't so we don't know exactly how those debates will be settled but for the time being I think uh and the numbers are

becoming so big I that on Nvidia has the cash to invest like even the VCs they don't have the if Nvidia says no I'm not going to invest who has the money like

the hyperscalers Microsoft and a uh sorry Amazon they're still participating but honestly speaking they don't have much capacity left to participate in this mega rounds uh if this continue for

like you know next year or two because they are also in the business of investing in a in in the capex right they're and they're they're investing like hundreds of billions of dollars so

so it's it's it's Nvidia who has like you know a lot of cash uh and uh uh are wondering what to do with that cash so and so it makes that's you know kind of

in a roundabout way some sense to me but like I said obviously they would look like the biggest fools in town uh if if openai and anthropic do not or or are

not able to uh match the expectations that they themselves have laid out uh you know uh uh for their company in the next like 3 to 5 years. So we we'll see how about that. But you know you you

made so you made an interesting point that I also wonder about and I would love your thoughts on it like so semi analysis for example uh does this great

analysis on what TPU this version is uh kind of you know uh you know the specs of this version and the next version right and they kind of compare and

contrast trainium uh uh and and and ver rubin uh and TPUs and do all these analysts right so that's like you know okay what's what's happening ing now and

what's happening the next uh for for the next one. But as an as an as an investor

next one. But as an as an as an investor obviously I'm not buying Nvidia or or or or any of these companies for what they're going to deliver for the next

product, right? Whether Nvidia will be a

product, right? Whether Nvidia will be a will be able to stay ahead of uh next versions of of TPU or trrenium. That's

not what I'm buying Nvidia for. I'm also

trying to get a sense why Nvidia's GPUs will stay ahead of uh TPUs and traniums let's say in 5 years right in 10 years

>> what how do you articulate that like what is for someone who has been following this industry for a while like how how would you articulate why uh

Nvidia will be able to stay ahead of uh traniums and and and TPUs in let's say in 2030 and 2035 right? Not just the next next version.

>> I mean, it's it's kind of hard to understand. I mean, I like it's it's

understand. I mean, I like it's it's it's it's moving so fast and there's so many different things with these models.

These models are not >> they're not very easy. They're not very they're not very like transparent either. No,

either. No, >> I would say that's like if I were to draw like a historical metaphor, I would be looking at like

um like CPUs during the risk wars of the 1980s when Intel had this older architecture, but they had the volume and they had Moore's law basically and

they had Mo's law to basically say we're going to we're going to yes these other competitive CPUs might be faster and like they you know it IBM had Power PC,

Sun had Spark and I think they were all very like strong competitors, very fast.

They looked good on the benchmarks and stuff like that, >> but in the end I think Intel out competed them all simply because they had immense volume and they had the

ability to roll in new, you know, they had an ecosystem that they could rely on to kind of really um keep them keep them ahead of their customers, right? And I

think like Nvidia envisions something like that as well.

Like they're building systems that expand beyond simply the GPU, beyond simply the chip. Looking they like I said they're they're not investing two billion into lamentum simply because you

know it's um because it's not it's it's a very important part of the system. It's

critical. Um, so I think like there those parts of that is very underappreciated uh and underappreciated part of what they were going to find and in the end

we're going to have I think they're going to simply lean on the manufacturing capability of these companies to of TSMC and other suppliers

to out compete. And I think like um if you're able to say you're going to get 30 40 50% like think about Apple during the 20s 2020s right where they say we're

going to get first dips on the latest mode.

>> You come out maybe maybe TPU follows a year after AMD follows two years after because they're slow and then like >> you're going to say like you get one or two years where you get really really

high profit and then by then the next chip's already out. So like

>> that if you think about it that is a leverage point to say like we're going to we're going to be number one obviously then TSMC is incentivized to say Google you come in like you'll be

number two like then but like we'll see how that sort of turns out for a while because being on the absolute absolute absolute leading edge like Apple is is a

has its own costs that like are not easily you can't just wave them away but like that could be a leverage point to say like we're always going to be on the leading edge and we're always going to take stuff. We're always going to bring

take stuff. We're always going to bring out the best product.

>> What are the costs of being on the leading edge from the customer's point of view? Like for example, Apple being

of view? Like for example, Apple being having the first tips and being on the uh leading edge. Uh what are the cost from Apple's side?

>> You just need to be really really effing good. Like you just like there's

good. Like you just like there's absolutely you're building the airplane before it flies. So like um think about it as like you're putting together the API. You don't know what's the API is

API. You don't know what's the API is going to look like. You need to like really build together and have experience with how this thing's going.

Which is why >> I've always wondered why AMD or or like other companies didn't go up to the leading O's cuz it used to be Apple and Huawei.

>> Huawei got kicked out for whatever reason and then like um it was just Apple for the longest time >> and like I think that was a problem.

That's a problem for TSMC cuz they don't like that. They don't they can't accept

like that. They don't they can't accept that because that does that's not enough to fund the next node. Secondly, also

like but for I would say for the other companies, Apple is that one pipe cleaner that says we're going to accept a 30 or a 60% yield or 50% yield simply

because like we want to be absolutely first.

>> You need to have the margins for that.

You need to have the balls for that. You

have the technical capability to to actually say this is what I want to build every new year. And I think that's the kind of thing that very few companies have the stomach for. But

Nvidia seems to be doing now. I think

they're they're going to be 50 they're going to take 50% of the 2nmter.

They're doing it right now. They're

taking 2% of the inaugural node and uh we'll see how they do with A16, A14, all the others to come up.

>> Yeah. So, so in that case it appears TSMC again they they are the bottleneck.

If you have the capacity of the leading edge node and if the let's say Nvidia uh does not fall off the cliff in terms of

like being able to uh utilize that capacity to deliver like a uh superior chip compared to the you know next best alternatives that that itself should

make it really hard for TPUs and traniums to be able to catch up. No, I

mean what like I said TSMC has the incentive to bring Broadcom to say to go to Broadcom and say let's let's bring TPU in. They want to have as many people

TPU in. They want to have as many people on the node as possible.

>> Um because it diversifies the customer base and also makes sure that TSMC gets paid gets an investment back for this N2 node. I think

node. I think >> TSMC early on I would say two years ago as early as 2 three years ago didn't think N2 was viable. um

>> their first N2 fab was very small was actually sub I wouldn't even say it's it's sube economical and I would say that's a direct >> that's a direct consequence of Huawei getting kicked out. So like I think like

um >> it's not you it's I make this argument before you will if you were to force >> the leading edge is so tight now you need as many people on it as you can

otherwise it will not you will not get it you will not get the note at all you'll get something worse >> right right so from TSM's perspective

they also don't want sort of uh a monopoly in the in in in in the leading being a leading edge capacity being dominated by one customer. They would

rather have like multiple uh customers willing to uh want that capacity so that like there's they have some leverage over their customers. Okay. Interesting.

>> Yeah.

>> Makes sense.

>> And there's real costs like I said there's real costs for being on the absolute alpha node. Like it's not this is not a done product. So like you you need to don't think of it as like

something that's useful. Like it's it really is it's it's really like it's like working in a it's like living in a house that's like not built that literally was just built. So it's like

like there's still plastic on the walls or something. The paint's still not dry.

or something. The paint's still not dry.

You have to think about it that way.

>> Interesting. Interesting. So if you think like from the entire you know obviously we can talk about the whole value chain for I guess for hours and hours but if you think about like the

entire semiconductor value chain where do you think uh the profit pools are the most defensible and obviously like I feel like almost all the

components of the entire semiconductor value chain are have this either monopoly or like duopolies or at best like igopies right so every every part of the value chain seems you know this

highly optimized industry structure at this point uh so to go back to my question which ones are like the most defensible and where do

you still see some vulnerability >> I would say that like just I mean I generally stick closest to the end user I would say the hyperscalers are probably doing the best um they're the

probably the most invincible I think then you keep going it's like a it's like it's like uh it's like crumbs. The

crumbs will eventually get down to the semi I think the semicap people have it the worst or the suppliers to the semicap people >> but like um they're the ones who have to take the most volatility like I talked

about pterodine right up 40% down 40%. I

think there's not really profits to be made um for these which is why you look at that structure and everyone's thinking oh this is this is this is bad monopoly and all that but it's like se

60 or 702010 has generally been the market share for these semic semicap companies for decades like since the 80s7s actually the 70s so like that's for a reason that's because when the

winter hits and it hits like every three years like like everyone gets hosed and then everyone like no one gets No one survives. So then if you holding the stock when the winter comes like

you're screwed.

>> The stock collapses. Like I don't think none of these like if you look at all these like semicapped founders name one who's rich.

>> I mean you can name recently you can name them now. You can name them now but like >> the ones that are started in the 70s and the 60s like they're none of them are rich. Like they're all

rich. Like they're all >> they're like comfortable but they're not like they're not like Elon Musk. I mean

they're not like they like this. It's a

capital destroying space. They call

that's why they call AMD advanced money destroyer. I mean this this company's up

destroyer. I mean this this company's up and down like these companies just destroy each other. That's why we have these uh that's why this is that's why I love seeing I mean China is going to be

even hilarious. Like they have so much

even hilarious. Like they have so much over capacity that like I don't even know what they're gonna do. I mean this I don't I can't I I want to have I mean it's a state secret now but like someday

we said I would love to send people to see these abandoned fabs like creating a 20% utilization. It's like and like it's

20% utilization. It's like and like it's like they're making a chip that was old 20 years ago. So it's like this is it's it's capital destruction. It's

wonderful. Semiconductors is is wonderful. It's the best way to lose

wonderful. It's the best way to lose money. Even though on the whole like the

money. Even though on the whole like the market's gone up for a while, but like if you try to invest like being in that like you you're just going to lose your shirt. It's hilarious. It's wonderful. I

shirt. It's hilarious. It's wonderful. I

love it.

>> I I can tell you you index your money because any anyone who's actively managing their money, they have all fallen in love with semiconductor industry, you know.

I'm like, if you try to buy any of these things, like you're making you're making the dumbest bet because like no one knows more than like like nobody knows how this stuff works. Like if you can't

explain how this stuff works, why are you trying to invest in it?

>> Yeah.

>> And then if you don't know how it works, then there's a second layer of that you don't know how the dynamics of how it works with the rest of the ecosystem.

You don't know how that company is working with that company. What's the

dynamics of that? Are you gonna get specked in? If you don't get specked in,

specked in? If you don't get specked in, you're screwed. So, like it's it's all

you're screwed. So, like it's it's all sort of it's all it's like three layers of of like dumbasserie when you're trying to invest. It's like when you're making you're investing in a semi

semicap company or like you're making an ass out of you and me like it's it's dumb.

No, I mean so you mentioned uh hyperscalers whose profit pool you think will be more protected than the rest of the kind of compute fo chain let's say

uh uh over the course of the cycle but one one concern that I have uh or one one concern I increasingly have some appreciation for is even the hyperscalers obviously hyperscaler when

I say hyperscaler they're like you know uh people usually include meta uh and alphabet as well and and I mean they should be part part of that group. But

let's for the moment uh take Alphabet and Meta out of this and think from like the uh AWS and Azure's perspective and Oracle's perspective. So they're they're

Oracle's perspective. So they're they're also hyperscalers who are primarily supporting uh you know 3P customer.

>> Is Oracle a hyperscaler? I I heard they're they're entertainment company now.

>> They're part of the uh they're part of the Warner Brothers conglomerate.

May May maybe you know they will they'll >> their biggest shareholder has pledged hundred billion dollars to a uh to a to a company that makes what five billion a

streamer I don't know that sounds I don't know if that's Oracle doesn't sound like a >> maybe Oracle will be fine you know because they will have massive you know

first party demand to uh you know for the streaming services so >> to generate new videos on Warner Brothers new Harry Potter videos.

>> Right. Right. Right. No, but I mean so if you think okay uh forget Oracle, you know, but let's say AWS or Azure, right?

So uh >> so now these companies have this massive kind of customers OpenAI and Anthropic.

So they have this concentrate customer concentration which was wasn't was never the case in the in the entire 2010s or even like you know uh uh up until chat

GPT uh came to the scene uh their customers were exceptionally fragmented uh that's why like even though compute prices have been falling but it it has

fallen at the at the pace uh that keeps their margin intact right so these business are incredibly profitable AWS has like mid-30s margins you Azure doesn't report but probably in the

similar even higher. So the question that I think about is now that they have customer concentration risk open AI and

anthrop anthropic and they their customers are not reporting 30 40% margins or like you know uh uh they they still have arguably somewhat speculative

economics and how their you know fate will be settled like in 5 to 10 years.

Uh so if something happens to their customer base >> then hyperscalers will have this enormous over capacity. Uh they're

they're spending $700 billion this year.

They'll probably keep doing that next year. Right. Uh and and year after that

year. Right. Uh and and year after that and if you if you keep doing that for a while and then some all of a sudden let's say open and anthropic are not able to keep up with let's say Alphabet.

Right. Right. Right. uh and then you know they're not being able to raise money and let's say these companies have become almost shadow of what we we we

currently think of them uh to be like in 5 to 10 years now people are thinking open AAI anthropic will be a trillion dollar companies like maybe in 2035 or something or maybe a lot sooner so if

something like you know if hyperscalers these two major customers turn out to be not so great businesses don't you think that also makes hyperscaler business uh

bit of a question mark in the interim period as well until the this fragmented you know customer enterprise customer base eventually catch up but eventually

can take a long time uh do you think that's a fair concern uh to entertain >> I mean I I mean it's tough to say I think like >> I think

there's two options that I that scenarios that just occurred to me one is yes that that open AAI and anthropic ditch. Right. A second one is that they

ditch. Right. A second one is that they actually the two those two companies vertically integrate. They decide that

vertically integrate. They decide that they're not going to decide they're gonna >> Exactly. Right.

>> Exactly. Right.

>> Yep.

>> The models stop advancing and there are indications that maybe the models do stop advance. They don't go forever

stop advance. They don't go forever >> and like they they stay up advancing to such a degree that you're now able to say we can predict where we can we can target we can target the chip design. We

can target all that. We then I don't know Something that comes off the top of my mind is like maybe those companies just buy the data centers. They'll buy the data centers off AWS. They'll

>> they'll buy something uh or they'll lease it there'll be a lease back or >> whatever happens. And then I think like AWS and Google and all these hyperscalers, they have other businesses

that can keep them flushed with cash and they'll just stop investing. And I think that's a concern that people do have in the industry is that maybe it's not forever growth, maybe it's not a crash, but it's going to be a stasis, right?

You're going to have it's all just pull forward and it's pull forward and then suddenly we're going to see if the models keep advancing. But if the models stop advancing, >> I mean, we thought pre-training was going to be a thing forever, right? We

thought we're going to pre-train for the world and then now everything is just RL.

>> Well, RL, is RL going to keep working forever?

>> I don't know. We don't know. So then

what happens after that? No one's really thought about that yet. So like I think there's people to say that like once these stop uh

uh stop addressing so fast then then like we have a problem and then but I think the hyperscalers will be fine. I mean that's not their main

fine. I mean that's not their main business. I mean what what Oracle it's

business. I mean what what Oracle it's their main business but like for these other companies is it really their main business? No, it's not.

business? No, it's not.

>> No, I I I I don't worry about, you know, uh hyperscalers being out of business or like even like being in a spot of bother uh deeply from financial perspective. Uh

but obviously like you know uh stocks inherently you know are about like you know the future, right? So it's

not just being able to protect your profit but also being able to grow your profit, right? So let's say if you're

profit, right? So let's say if you're assuming like you're going to generate $100 billion profit this year and let's say you are we are expecting you to generate $200 billion profit in 5 years

and it turns out you didn't generate $200 billion profit but generated like $100 billion profit which is still plenty good like you know you'll be fine you your business is totally fine uh but

your stocks are like going to be down a lot more because people expected you to generate $200 billion profit right so that that's where my kind of you know concerns come from like if there's Yeah, but like do they care? Do these does

does Larry Page care?

>> I mean like didn't you say on another that like if you bought Google at the IPO and you held it like up until like recently like you probably lost you underperformed the S&P or something like that

>> like so that that's like you know I think sometime in mid 2025 when Google stock was like 150.

>> Yeah. That's when I I noticed like it was a crazy kind of you know thing to realize that even if you bought Google at their IPO and you kind of held held it till like some I think mid25 or

something like that like uh you you underperformed uh NASDAQ 100 which was crazy.

>> Yeah. I mean there's a I mean that it's actually you can consider the fact that these hyperscalers probably underperformed the market because they're such terrible conglomerate businesses. You get a huge conglomerate

businesses. You get a huge conglomerate discount. I mean, you could simply argue

discount. I mean, you could simply argue that this is a return to the norm.

These these businesses are >> these businesses are like, you know, >> I mean, pre2022, you could argue that a lot of these businesses had were were

probably in need of some sort of reform.

Like they're I would say like the you should probably expect that to happen more than less than than >> than than uh to not happen. continue to

expect continued growth in my opinion.

>> Yeah.

>> Like these these these businesses are weird.

>> No, I mean uh I think that's fair. Uh if

if I were Larry Page, uh I would not go I would not lose sleep if Google stock is like 0% return for the next 5 years.

I I I don't think I Yeah. Yeah. I I

>> would on my yah my mega y >> I cannot imagine being Larry Page and worrying that you know the stock's profit like you know a company's profit didn't double in five years right I I

can't imagine myself being worried about that yeah >> I'd be like hey you know you want to you want to argue about it here's my class C shares or whatever I'm super voting shar

>> right right >> going to try to Yeah. So, don't worry about them.

>> Maybe that actually does mean that then hyperscalers are kind of sucky investments. So, like I don't Maybe it

investments. So, like I don't Maybe it just means that we should start investing in the semicap stuff cuz that's insane. It's going to go insane

that's insane. It's going to go insane and then it'll turn around and then we'll be screwed later. But like just dance while the music is playing.

>> That's my opinion. Yeah, maybe maybe those those Dutch guys in ASML, you know, they they're laying off people when they are kind of, you know, generating revenue growth like you know

uh much faster than they used to before and now even then they are laying off people. So they are probably in a much

people. So they are probably in a much more you know shareholder friendly situation than the hyperscalers are.

>> Yeah. I mean I they must really be working hard because do you know how hard it is to lay people off in the Netherlands? It's really

hard. Oh, like they must really really it's better to do it now in my opinion.

>> Not to say bad things about the uh about those people, but like I feel sorry that they're losing their jobs, but like uh they probably it's probably the right time to do it. I would do it now because

like it takes like a year or something to to lay someone off in the in Europe.

So could it's a shareholder conscious call. That's the best way to say.

call. That's the best way to say.

>> Okay.

shareholder consc uh it's been pleasure talking to you and let's you know yeah let's talk talk again maybe in a couple of years to see

what Larry Page might be thinking in his yard you know and whether the Dutch guys were actually right to laying up people during good times >> hopefully you have me on better sooner than that I would love to do a call we

can do something totally unrelated to chips we talk we can talk we talk farming let's talk farming I don't know thing. All right, have a good one.

thing. All right, have a good one.

>> Okay, awesome. Thank you.