FULL INTERVIEW: Dylan Patel Says We’re Still Underestimating AI

Good to see you again. Good to have you. Data centers in space, what you got?

When are we going to space?

Me, you, the International Space Station. Let's break it down.

You know, the space tourism industry is quite a, quite a fun one, right? Yeah.

Would you go? Would you do the Blue Origin thing where they blast you out past the Kármán line? It's good enough for Katy Perry. It's not good enough for you? What's going on?

It's like, you know, like you're in free fall. You're not actually—

Oh, it doesn't count.

Shots fired.

Kármán line, come here. I want to be like going around for days. I want my bone density To start to atrophy, right? I truly want to feel the negative effects of space.

Yeah, yeah. It's not enough just to go and back. I think I would do it.

It's even if it's 90 seconds, right?

Yeah, but it's better than being hanging out on Earth.

But look at all the cool stuff that astronauts do, right? Like, you know, put water and then like they're bubbling and then you like try and drink the water.

Soon they'll be unplugging the GPU and plugging it back in.

Oh yeah, yeah, yeah, yeah. That's how you pay for your space tourism. You gotta go on the secret ship.

90 seconds of servicing. Yeah, SpaceX satellite.

One 90-second trip at a time. No, but people were wondering, you know, TPUs, NVIDIA going on the Starlink V5 or whatever, something gets up there. It feels like this will be something more like a Tesla silicon chip, an AI chip. Like, do you have any insight into like what the process, if you wind up figuring out how to heat dissipate, if you wind up figuring out the costs, what might the chip look like?

So I think, I think, you know, everyone freaks out, oh my God, putting stuff in space is expensive.

Yeah.

But if you look at like Starship launch costs, and, you know, they keep falling, you're like fine, right? Like, I think that's not, you know, by the end of the decade, the cost of space launch will be fine. The heat dissipation, I mean, it's a challenge, but you just put a massive, massive effectively radiator.

Yep.

And it's fine, right? By the end of the decade, like, you'll be good. I think the big challenge is that chips are just really unreliable, right?

Sure.

And so how do you deal with like, a couple things, right? Satellites can only be so large before they, like, you need— start needing a lot of support and structure before they tear themselves apart. So when you look at, like, the launches, right, these things are shooting out, like, tiny satellites.

Yeah.

And many of them. Okay, so you can't have, like, a big, fully connected cluster of chips. And then, like, on top of that, right, how do you deal with any random error? On Earth, you have techs running around the data center unplugging stuff, Putting in spares, things like that. What do you do in space? You RMA it to the factory where they like might unsolder it and resolder it and then like test it and it works and go back out. Sometimes it is just trashed. But like, you know, that's the challenge to me.

Is that— is that— I feel like maybe the pattern we should be looking at is like, how often do the Tesla self-driving chips need to get serviced? Because that's like the team that would probably be building or like bridging the gap there. Like the Starlink satellites, sure, they go down, but like the service works. Like you're just relying on some sort of like, you know, 90% uptime, stuff's coming down. But most people that are in a Waymo, like the chip keeps working, right? Most people that are in a Tesla Self-Driving, like they're not— like you don't hear about Tesla owners being like, I love FSD, but I'm constantly in the shop getting my custom silicon chips unseated and reseated, right?

Well, I mean, it's also a function of like the complexity of a chip, right?

Sure.

You know, if a chip is twice as fast, and let's say the bit error rate, right, like how often a bit flips, is the same, then it's erroring out twice as often.

Yeah.

But let's say the chip is 10x as big, right? And so when you look at like a Tesla FSD chip, very, very good, very, very efficient, very still like relatively inexpensive and cheap compared to, you know, a big old GPU or TPU or whatever, right? Those things are extremely large. And, you know, again, like if the error rates are the same, Then it fails 10x more. But in fact, the error rates are a bit harder, higher, because they're pushing these things to the absolute limit.

Yeah.

Whereas, you know, Tesla does have some level of like, well, first of all, the Tesla car has 2 chips, sort of redundancy already built in, right?

Maybe you do that on the satellite, but then that's more power, more—

Yeah, right. So the whole allure, right, of it is, is, you know, effectively power is free, right? And solar panels, you look at the cost curve of solar panels, you look at the cost curve of satellite launches, you're like, this is, this is free. This is great. But power is less than 10% of the cost of the cluster, right?

Sure, yeah.

So, so like, it's that 90% you're not saving anything on.

Yeah, yeah.

And insofar as much as—

For potentially 100 times the hassle.

Yes, yes.

Yeah.

There's this whole like, you know, like if you look at NVIDIA GPUs, right?

Yeah.

When you first turn on the cluster, about 10 to 15% of them fail RMA in the first 2 weeks.

Wow.

And then that's fine, like you have to reseat them, whatever. And like, the industry knows how to deal with this, right? And over time, like, Hopper's now at 5%, but Blackwell's still 10 to 15%.

Wow.

Right? Actually started out higher than that.

Sure.

And when a new generation comes out, it's gonna be higher than 10, 15%. It'll have its curve gradually decline down. But, you know, who's gonna— are you gonna test it and burn it in on the ground? Or are you gonna say 5% of my chips, or 10, 15% of my chips are trashed?

Yeah.

Because someone can't go up there and like do these things. Or am I saying, oh, I need robots who can do all this stuff in space, and now that's like an additional engineering problem. When sacks of meat are actually very cheap.

Yeah. Speaking of Nvidia, we haven't talked since the Groq acquisition. What does that look like in the bull case? Like, if it's a good— if the next version of Groq is a great chip, is it sitting next to the H200, H100s in the rack, GP200s? Like, how does it fit into the actual, like, what Nvidia deploys? Is it just a separate chip that's off to the side?

I think it's a big vibe shift from Nvidia, right? Before, they were like, All right, I got this big GPU. Everyone's gonna use this GPU. Software ecosystem of the GPU is so good. It's one size fits all. Everyone's trying to make all these specific point solutions, but we've got the thing that's good at everything. And then they had a vibe shift, right? They launched this thing called CPX, which is a chip made for prefill, prompt processing, creating a KV cache, and also good at video generation and image generation. And that's coming out later this year.

In the blog post release, they were really talking about video generation as well.

So, so yeah, you've got like CPX, you've got like the standard GPU, and now you've got the Groq chips and they all fill a different niche.

Okay.

But really it screams, oh crap, we don't really know exactly where AI's going, which I don't think anyone does. Right. I mean, it's, it's moving so fast. The software is, um, the model architectures, et cetera. So we're just gonna like engineer solutions that are along multiple points of the Pareto optimal curve. And then, you know, if one of them will win.

Yeah.

Right. And I think, I think it's like sort of like a big vibe shift from NVIDIA.

Mm-hmm.

Also, they just knew OpenAI was gonna do the Cerberus deal, so they freaked out.

But. Got it, okay. Yeah, get me up to speed on what makes Cerebras important in the ecosystem right now.

So, you know, you have people thinking like, oh, latency matters in terms of where a data center is. It doesn't matter at all. What matters is, you know, as we've moved from, you know, chat applications, which were like, or search response immediately, chat applications, let's say response takes 10, 20, 30 seconds. You've got agents, you know, I don't know, my Claude codes are working in the background for a long time, right? It doesn't matter where the data center is, But what does matter is that these streams of inference take, you know, 30 minutes versus 10 minutes versus 5 minutes. And for a lot of people, I'm fine to spend 10x the price on something that completes 10x faster.

Yeah.

And so Cerebras sort of just makes a ton of sense there. So OpenAI, you know, they've got these like long horizon— there's like Codex 5.2 extra high thinking or whatever. It's terrible. Can you guys teach them how to market?

Oh yeah.

OpenAI, you have to sponsor this podcast. Absolutely.

Yeah, yeah.

We had DeepMind on yesterday.

And I did actually ask him, like, I had the Codex app pulled up on my desktop and I was like, there are 6 different models and then there's another button that I can pick to—

Well, how many different products are called Codex now?

There's a lot.

Right?

And now there's an app, yeah.

Yeah. We actually had another guy on just to do branding. Lexicon Branding came on the show yesterday talking about all the naming—

Naming architecture.

Naming architecture. It is complicated, but hopefully—

You could tell he's just, blood's boiling because like all the AI companies just have the most chaotic, like Anthropic, Claude, Claude. Claude code, but also you can use Claude code for other stuff.

Yeah. But yeah, I mean, uh, with Cerebras, it seems like there is a value to it, but are they constrained on the supply side? Like, can they actually scale up to, you know, a Colossus-style data center that could actually speed up Codex not just for like one user, but all the users?

So, I mean, Cerebras can speed up multiple users for sure. Yeah. The question is sort of like where you use it, and that's where they have to like Figure out where within Codex, right? Yeah, because there are times where Codex is running for like 10 hours, and sometimes you don't mind, right? Like, yeah, screw it, I've put up this nice prompt, gone, work on it, refactor my code, do this thing, do this task. Other times I want this iteration feedback loop. So how do you expose it to the user without saying, hey, actually there's another toggle, so your permutation is, is 18 times?

Hopefully like a really robust model router, but it feels like that's been a process.

Yeah. So the OpenAI deal is like for 750 megawatts. It's not that much capacity on the order of like what OpenAI has talked about, you know, by the end of '28, they'll be at like 16 gigawatts of that.

So it's just like the absolute cutting edge, the most price insensitive customers in that specific use case of this is the type of prompt that you need to return fast, then you'll get the speed up potentially.

Right, right. And they've got to figure out how to do it from a product exposing it to the user, et cetera. But it's, it's clearly something where there is demand, right? Like, I don't know, like Andrej Karpathy doesn't care if he's spending $1,000 per agent per second or whatever, right? Like, you know, whoever it is, these like super cracked engineers don't care at all. And then obviously, there's like a long tail of like, actually, cost does matter for most people. And so all along that curve, they've got to have solutions, right?

Yeah. When did you first think that xAI might end up at another Elon company.

I mean, this has been rumored for a long time, right? Like, people are saying Tesla, Tesla, Tesla for the longest time.

It's harder with a public company.

Yeah, yeah. And then, and then a few— a bit ago, people were like, oh, SpaceX. I'm like, wait, this makes no sense.

No, but there was a very coordinated, like, narrative pump.

Oh, yeah.

Like, and then the space data center at the end of last year.

No, it was perfect.

It was like almost like perfectly telegraphed.

Well, there's, there's a bet, right, between Basically, the head of compute of XAI and the head of compute of Anthropic. And the bet is what percentage of worldwide data center capacity is in space by the end of '28? And the bar is 1%.

Oh, wow.

And so the XAI guy is like really bullish, and the Anthropic guy's like, eh. Yeah, a little slower. Yeah, yeah. So, but it's a really interesting bet. I take the under on 1% by '28, because that's a gigawatt in space.

Yeah.

But it's actually not that crazy, right?

Yeah.

It's roughly 150 Starship launches.

Yeah.

We'll get them to 100— to get them to a gigawatt in space.

Yeah.

So, you know, Starship hasn't worked yet.

Yeah.

Fully.

I was looking at the energy draw of the current Starlink fleet, and I think they're at like, what is it, 200 kilowatts or something like that? So you get 1,000 of those, 200 megawatts, and like, you're starting to be in the territory of something like that.

Yeah. So the V2 Starlink satellites, I think, are the only ones they've launched. Maybe they've launched a few V3s, but the V3s are coming soon. And those are Those are like 100x more bandwidth each, right? And more power.

And just more power. And so when I'm just thinking of like, can you scale this thing up at all? It's like, are they 2 orders of magnitude off? Are they 3 orders of magnitude off? It feels like they're like 1 order of magnitude off being able to run something that looks like an H100.

I think the metric is like 50, it's either 50 kilowatts a ton or something like this per satellite for V3. If let's say from V3 to whatever the compute thing is, they double it again, get to 100. I think the V2s are like 25. So if you get to 100 kilowatts per ton for launch, it's It's only 150 or so Starship launches. I think that's so reasonable. Maybe not 28, maybe it takes 29, but like, you know, it's so reasonable. The question is cost and reliability and, you know, what happens when the chip fails? How do you service it? That kind of stuff. How do you deal with having clusters be much smaller instead of like, you know, these big clusters? Even for inference, big clusters are useful. Yeah.

Yeah. How do you think about Google's response to Groq's or Iris'? TPUs, obviously very successful, but are they forking that project to eat more of the Pareto curve?

Yeah, so for the longest time, Google's had one main line of TPUs, right? All made by Broadcom, and then sort of next year they've diverged it, right? Where Broadcom makes a TPU and MediaTek makes a TPU. These 2 TPUs are focused at different things.

And they're fabbed at—

They're both fabbed at TSMC. Everything at the end of the day goes to Arrakis, right?

I wanna go there next, but— Everything goes to Arrakis.

So fabbed by TSMC regardless, but both of these TPUs are focused on different things.

Okay.

And they've actually got a third project for another kind of TPU there. They also see this need to proliferate along the curve of like, hey, do I care a lot about super high amounts of flops, not that much memory? Do I care a lot about super fast on-chip memory only? Do I care about 3D stacking memory? Do I care about, you know, this sort of general-purpose middle-ground AI chip, which is what, you know, an H100, a Blackwell, a TPU looks like today? You know, they're sort of like, oh, we need to hit the entire Pareto optimal curve. And it's like, okay, within this there's training versus inference differences and like what numerics you want and all these other things. There's so much complexity there. Everyone sort of is diverging their roadmaps once they're at a sufficient scale, I think.

Yeah. Or is Google still way ahead on cross-data center training?

Yes.

Are the other labs like— is that important to the other labs to catch up there? Or is it something that will just naturally happen because everything sort of commoditizes? Or do the other labs need to sort of marshal some Herculean effort to like crack the code on what it takes and what Google's doing?

Yeah, so, so it's a couple of things, right? In 2023, everyone thought that scaling was pre-training. Yeah, right. Um, you know, more parameters, more data, and that's very difficult to split across data centers.

And has Google been able to do that?

And Google's been able to do that to an extent, right? So what they've done is They've got, you know, they don't have the largest individual data center campus, but what they do is they do these like regions where it's like, hey, each data center is roughly 40 miles apart from each other.

Sure.

So in Nebraska and Iowa and then in Ohio, they've got like these complexes and now they're building one in Oklahoma, Texas. You know, these complexes where there's all these data centers pretty close to each other.

So it's not really cross-data center, like across the world, right? It's just across like region.

Yeah. And then that makes a lot of the difficulties a lot easier. Um, flip side is we've also moved to RL, right?

Yeah.

Um, and majority of the time of the chips is spent generating data, right? Only doing forward passes through the model.

Sure.

Um, and then you only send the final tokens that you verified sort of back to train on to the training, to the train, to train, right?

Yeah.

So then you end up with like, oh, instead of in pre-training scaling, you need to like synchronize all the weights every 10, 20, whatever seconds.

Mm-hmm.

Um, when you're doing these rollouts and especially as things get more and more agentic in training, you might not only need to send not the entire weights but just the tokens that are relevant. So way smaller amount of data and way less frequently, right? Minutes at a time instead of seconds at a time.

Yeah.

So you've got this like now, now it's become like reasonable where, oh, actually multi-data center training is completely reasonable.

Yeah.

People do this. People do multi-data center, multi-chip training.

Sure.

Right? You do your inference on one set of chips and you do your training on another set of chips. So like Anthropic does this.

Yeah.

I don't know if Google does this, but Google's kind of already got the cards. Yeah.

Okay. Got it. Let's go to Arrakis.

Yeah. Talk about Arrakis.

Just there's this debate, TSMC risk. Is that the bottleneck or is energy the bottleneck? I was doing back of the envelope calculations. It seems like we're using maybe like 1% of global energy production or Western energy production on AI specifically workloads. And then we're using like 50% of leading edge fab capacity. on AI workloads. And so that feels like, okay, well, even if we all agree and we say as a society, we're going all in on AI, we can only double the AI chip capacity before we need to build more fabs. That takes years. Whereas we could say, everyone turn off your air conditioning, we're sending the electricity to the data centers, right? Like, we have the ability to generate without creating new— Did I just get some clapping?

Clapping for turning off the AC.

Turn off your flawed Heat strokes for all the grandmothers.

Yes, yes, yes.

I need my cat dancing videos.

You need to feed Claude, right? Um, but, but, but, but seriously, like, there's this debate over, you know, is TSMC the main bottleneck or energy the bottleneck? How are you feeling about that?

Yeah, yeah. So, so sidebar before I answer the question, 'cause I think it's fun.

Yes.

Um, you know, in the US it's insane to say turn off your AC for AI.

Yes.

Right. And the general public hates AI already.

Of course.

Of course.

Um, but in Taiwan they've had droughts before and they've turned off water to entire cities. They're like, oh, you get to, you get water 3 days of the week.

Whoa.

And then the fab still gets supplied water. It's like, this is, you know, you've gotta understand the mindset. We are not ready as weak Americans to do this.

Yeah, that's crazy.

No, but at the end of the day, right, like water and power are certainly less big of constraints. Now, you've gotta imagine like, you know, semiconductor industry's used to, hey, doubling the amount of transistors made every year or 2. Part of that is Moore's Law, part of that is more capacity. Whereas the energy industry in America wasn't. And so like, Initially, people were like not creative. They're like, let's do, let's do these kinds of gas plants. It's like, well, no, now we've realized, you know, yes, there's 3 main manufacturers of turbines, and then you've got for a dual combine cycle, then you've got like IGTs, but you've also got like medium-speed reciprocating engines, right? Like, turns out Cummins can make like a million diesel engines a year, and like those can make electricity. Like, if I don't give a fuck and I put it in West Texas, easy.

Yeah.

So now it's more of like a regulation thing, a supply chain thing. Power is not a constraint in insofar like that much, right? I think it certainly is a constraint still today. It was the biggest constraint in 24-25 data center capacity power because the industry was just not ready. People have woken up; they've like sort of been shocked to the system. Now you've got you know tens of gigawatts being deployed. You know next year, 30 gigawatts are being added, and we think the power is there for it.

Wow! What was this year?

It's it's or this year is like I think it's like.

18-ish, 10-ish. Okay.

15 to 18-ish, sorry.

So almost a doubling.

Yeah, almost a doubling, yeah.

Yeah, wow.

And when you look at TSMC and the crew, right, there is not really, oh, this random, you know, there's 12 people making medium-speed reciprocating engines that you can now convert to make power at some random data center. No, no, no, there's like, there is Arrakis, right? There is one set of spice, like, you know, there's, you know, that's it, right? And so, And then the flip side is like, okay, when you have 12 vendors, everyone's got a little bit of slack capacity. You know, there's more likelihood, you know, you can— people are like, oh, turbines you can't get. You can call a broker and you can get a turbine. You might be paying 50% more, 2x more, but you can get a turbine.

Yeah.

Right? Like it's—

But you can't get a 3-nanometer fab.

You cannot get a 3-nanometer fab. Exactly. And so when you talk about what's the, you know, the baton got passed from semiconductor shortages in '23 to power and data centers in '24, '25. '26, we're still— we're swinging the pendulum, but it will fully be semiconductors again in '27, right? And so we see this across the entire space of the ecosystem. It's not just TSMC, it's also memory, both, because both of them have built at a certain pace. Now, TSMC has been expanding at some rate. The memory makers, in fact, have just not expanded capacity. Basically, they have not built new fabs since 2022 because their cycle is so undulating.

Right.

Yeah, and so when you look at it, it's like, oh, even if they wanted to double capacity, they need to build the fabs, right? And building the fabs, it is the most complex building humans make, right? It's— the entire air of a clean room circulates itself every 1.5 seconds.

What?

And you don't even feel it when you're inside.

Really?

It's like that, and it's like parts per billion of particles, right? Like, it's actually insane how— you could get coughed in the face by someone who has COVID, and not get COVID.

And so— Because it gets circulated so fast it doesn't even hit you?

It's like, it's like—

It's like that meme of like the spraying when someone's talking and then it just gets circulated.

So one, another, another sidebar is, um, everyone knows COVID like really popped off in Wuhan, right? Wuhan also is home to, uh, China's, uh, largest memory company, YMTC. And so when they were like welding people into their homes, the people who worked in the fab still went to work.

Wow.

It was because it's, you know, one, it's a national, like, it's national importance, but two, like, These people aren't getting sick. This fab is like way too clean. Oh, it's crazy. Yeah.

Sorry, Jordy.

I want to talk about Oracle. They put out a post this morning that said our partners' financing for the Doña Ana County, New Mexico, Shackelford County, Texas, and Port Washington, Wisconsin data centers are secured at market standard rates, progressing through final syndication on schedule and consistent with investment-grade deals. Obviously, they were fast following their post from yesterday where they said the NVIDIA OpenAI deal has zero impact on our financial relationship with OpenAI. We remain highly confident in OpenAI's ability to raise funds and meet its commitments. And obviously everyone was looking at this being like, give me a cigarette. Like smoking. Like it's like bank run language. I haven't seen posts like this since like the FTX.

Is it just bad comms or is there something worse?

It's terrible comms. Yeah. Like, like I told my Oracle contacts, like, who the hell is in charge of the Twitter? Like, what are you doing? Nvidia did something similar last year when the whole TPU mania was going on.

Yeah, it was. Yeah, it was.

It was.

It was like we're thrilled with Google's progress with the TPU. That said, Nvidia chips are the only, you know, it's like no one asked you to comment. I mean, like, I'm sure a handful of people in your DMs and random, but that doesn't mean—

doesn't project confidence.

It's sort of the lion shouldn't concern themselves with the sheep. And like, okay, Nvidia is the lion. Maybe Oracle is a little bit more bumpy, but I think Oracle is like fine. People are just freaking out because, you know, OpenAI is is peak, you know, people are peak negative on OpenAI right now because of how good Anthropic's been killing it. Sure, sure, sure. Yeah, I think it's just like kind of silly. Like they need to hire someone to do comms, like a Lulu or something, right? Both NVIDIA and Oracle, because what are you doing?

How did you process yesterday in general? Jensen was clip farming. He was like, I don't know why he does these street interviews, right?

No other CEO does those where they just stick 25 microphones in your face and and the paparazzi's flashing. It's a great vibe.

It's, you know, Jensen's not been as famous as other CEOs for as long, and yet he's so important now. And if you've like, if you know Jensen, how he is in meetings, I feel like there's 2 Jensens, right? There is like PR, like good at PR, just good at talking, good at like making people hyped up and believe what he's doing.

He's great at standing on stage, holding up the chip, delivering like a sermon.

And then there's the real Jensen, which is like, a business killer, and like actually just knows about every like aspect of the supply chain, right? All the way from like niche semiconductor, you know, design and manufacturing stuff, all the way to like energy, power, data center, like, and then doing the business deals too, right? And so like, you've got this whole Pareto, like, of a whole thing, whole range of things that he's good at, and he's a killer in. And clearly he's like, he was in a meeting where he was being a killer, and like negotiating like supply contracts or something. Oh, and he walks out.

And then he walks out.

That's hilarious. Yeah, yeah.

But I like it.

Yeah, that's awesome. And that's why he was like, you know, like he was like still killing it.

Like, no, we never said we committed to $100 billion, you know, like, and it's like, I don't know, where do you even get the $100 billion number from? It's like, well, you did go on CNBC and make a big deal out of it. So people would assume that it was, but, but they did say in the press release, I remember these are early talks, but they just kind of jumped the gun. This was the height of the press release economy. Yeah.

Yeah. What's funny is, um, Oracle stock peaked like just like a week after they announced the OpenAI deal.

Yeah.

And so like the press release of like, hey, OpenAI is gonna do this humongous deal, stock peaks. Yeah. Um, same happened with a couple other vendors who, uh, announced deals with OpenAI or NVIDIA, like sort of a lot of these, like, like they all peaked then. And then it's sort of been like NVIDIA OpenAI trade has been going poorly.

Yeah.

And sort of like the TPU Anthropic, Google Amazon complex has been doing well. It's quite interesting that this happened.

Good, good energy back at home with the roommates.

What?

What? What's going on?

I wanted to. Yeah, I got one more thing. I so yes, over the weekend it was sort of drowned out by all the Justice Department stuff.

But wait, have you guys talked about Elon saying you can smoke a cigar in the fab? No. Yeah, yeah, yeah. Okay, I was gonna say that this is this is this is part of the whole thing.

I didn't realize that was related yet. I mean.

Yeah, indoor heaters. Yeah, we have indoor heater technology. No one's taking advantage.

Yeah, what does the fab look like if you have no humans inside? Like, that's probably his long-term thing is like, yeah, there will be an Optimus.

But no one knows, like, the number of people working in a fab is like irrelevant.

Like, but, but, but is it, is it irrelevant? Because there's all these things you have to do when a human's in there because they sweat and they breathe. And if you don't have to do that because it's a robot walking down, even if it's puppeteered or teleoperated, you, you might be able to have different considerations. I don't know if that actually affects—

Well, it's like a nesting of like, it's a nesting of like cleanliness, right? For example, you've got this wafer, you've put like down, let's say you put down copper.

Yeah.

And now you're moving it from one area to another. Well, it needs to be stored in a vacuum, but the easiest way to store a vacuum, like, or, or an inert gas.

Yeah.

Um, and that's like the thing that's being transported in. But then around that, you want it to be super clean as well. Um, you, you, you, if you don't, then the copper starts getting oxidized, it affects your yields, all this sort of stuff happens. And so like, you kind of want it to be a nested layer, nested layer of like, Well, this thing inside the EUV tool is super clean, and then the thing feeding it is super clean, and then the thing it sits in is super clean, because that's how you get to like there's zero particles. Because like, you know, in the FOUP, in the transportation devices, like parts per trillion, and maybe FOUP, it's called F-O-U-P, front operated, front opening, I don't know, something pod. But it's called a FOUP. It's like the thing that moves and it carries the wafers. And then the fab is like parts per billion, and you know, sort of like, You know, you've got to, like, got this nesting relationship, so everything is super clean. You know, I'm bullish on robots, like, super bullish on robots, but only for, like, not for tasks that have, like, TSMC's Arizona fabs, or, okay, let's say TSMC Tainan, which I think produces, like, you know, indirectly hundreds of billions of dollars of global GDP. Even directly, it's, like, still tens of billions of dollars, has, like, 5,000, 10,000 people in it. Like, it's, like, irrelevant in terms of the number of people who work there.

In terms of the overall economic value.

Right. It's like, it's like, it's like how many people fold laundry or how many people wash dishes or how many people like do construction work? Like these are way bigger markets for robotics.

Yeah.

Yeah.

Makes sense. Speaking of China, what what are you making of the the the Dario essay or he I guess his comments at Davos about you know selling chips to China is equivalent to you know nuclear weapons these days. The the Ben Thompson line was something like. He's okay selling chips because he wants dependency on the NVIDIA ecosystem, CUDA, but he would ban lithography tools from going to China. And I've been wrestling with this idea of, like, I don't know if China would accept this, but wouldn't there be a different world where you want them dependent on American LLM APIs and you don't even send them the chips? And you say, yeah, you can have as much AI as you want as long as you're paying OpenAI and Anthropic API.

Yeah, I think it's I think it's like a curve of like what they will accept. It's it's it's you know one you you you push someone into the corner, they're going to start swinging, right? And and I'm I'm like very concerned that China does this, right? Do they do you do you push them too far into the corner? Do they say screw this? We're going to start being a lot more aggressive. We're going to we're going to you know do more military actions or military actions or or even just invest twice as much in in global supply chains or take over Africa more than they already have. Sure sure sure. LatAm, like, etc. There's, there's, or, or just take over Taiwan. Yeah, right. Because if I can't have the chips, what value is there in Taiwan existing? Sure, sure. In its current state, right? So there's like, there's this like game theory aspect. Yeah. At the same time, you don't want China to be able to like, you know, if you believe AI is going to do what I think many, at least in San Francisco, think it's going to do, which is like completely revolutionize humanity and cause GDP growth to accelerate. Do you want to have China also own that technology? Um, and all, you know, their ability to integrate that into their military and all these other things much faster. You know, so there is like these competing, like, you know, interests. Yeah. Where, where is the, like, right line? And some people think it's like, hey, yeah, sell them AI model. Well, I think Dario would say don't even sell them AI model access.

Don't even sell them tokens.

Yeah, I think so. I think, like, I think Anthropic does not sell AI access to China.

Yeah.

You know, they loop it through, and you can see this in the traffic data. They go through Korea and Japan and other places, but like, and so they get it.

Yeah.

And then the other side is sort of like, I think, like the Ben Thompson view, which is like— and I think I'm more sympathetic to that, although I think I'm not exactly aligned with that, which is like— and we've been saying, like, don't sell them equipment, don't sell them equipment, don't sell them equipment. And my argument is more economic in the sense of, like, if you sell them tens of billions of dollars of equipment, they can make hundreds of billions of dollars in AI value or chips with that equipment. Whereas if you sell them AI model access, then it costs them this much to get the economic— You know, they're not able to—

You're capturing more of the value.

Exactly. And so that's sort of the question that is at foot here, right? Do you want them to capture all this value of the supply chain in equipment or by buying the chips or using the models, right, and services? And we've seen, you know, across many, you know, stacks, China refuses to accept, you know, using American ecosystem and they'll wait many years before they develop their own, whether it was like, hey, they didn't use Windows, they figured out a bootlegging economy, or They didn't use Visa, and eventually they came out with like Alipay and WeChat Pay or whatever it's called. And like, these things are way better than Visa, in fact, right? Lower transaction cost and higher volume.

I've never used Red Star Linux. It's North Korea's Linux distribution.

Wait, really?

Yeah. If you, if you don't, if you put it on a network, it'll immediately call home. So you have to put it on a firewalled network or else it just like steals everything immediately.

I'm a fan of TempleOS, you know?

Yeah. Is Doug O'Loughlin suffering from a case of Claude Code psychosis?

Okay, yes, yes. So I think everyone's like, Claude Code is for coders. It's like, no. Claude Code is for people who don't code now.

Yes.

Right? And that's the big realization this year.

Yeah.

You know, we've got a couple folks now in the firm who have psychosis, but Douglas O'Loughlin, who is like, you know, SemiAnalysis number 2, he's president, you know, he's my boy. He's the one. In fact, he's the one who encouraged me to make a Substack a long time ago. A long time ago. What were you doing before? I had a WordPress blog, and I was like consulting on the side. But I was like, okay, let me do a Substack now because I saw him making money off it. I was like, this is shit. Like, why are you getting paid for this? Like, there were multiple times where he wrote something. I was like, I could do way better. I'll show you. And like, obviously, like, it was good because we both taught each other a lot of things. We've been great friends. And eventually he joined SemiAnalysis. But like, you know, his background is he was a hedge fund analyst, and then he decided to do a Substack slash walk, hike the Continental Divide Trail for like 6 months, walking from Mexico to, you know, and then, you know, came back to doing Substack and tried to do a fund.

6 months of touching grass, and then he was like, I'm ready to lock in on Claude code.

Yeah, yeah. And so now he's, you know, like, he's never been a software developer, right? But he's been on a generational run. Like, he's not coding anything, right? He's just telling Claude to do stuff. And like, it's to the point where it's like our our like head of data, head of IT is like, "Oh, can you send me that?" And he's like, "How do I do that?" And then he like he zips the whole thing and sends it to him. It's like localhost. He sends him a leak once. It's like localhost.

It's like, bro, that's not how this works. But but yeah, no, I I've talked to some folks who vibe code and they'll be like, and I'll be like, "Why'd you choose Node.js?" And they're like, "What's Node.js?" That's a very specific choice.

Someone?

Yeah, Tyler.

No, but like it's it's um we went on we went on a little. Tour of a lot of our clients. Like, you know, roughly like half our business is, or 40% of our business is like hedge funds. So we went to New York a week, two weeks ago, and we went to all of our clients. And like part of it's like them asking me, is OpenAI fucked? And I'm answering like, no, I think they're fine. And then like some like actual ideas. And then like a lot of it's Doug just telling them Claude Code is like, they're like, you don't have to hire any junior hedge fund analysts anymore. And they're like, the junior hedge fund analysts are like, and then he's explaining, you know, what can you do? It's like, well, like You can just do like financial models and pro forma financial models and like everything in Claude Code without ever opening Excel. And you can generate charts and like you don't need to know how to code. You just need to know how like how this stuff generally works and you can just do it.

Are— how many hedge funds are just trying to copy trade situational awareness now?

I mean, I think everyone who's— I think, I think a lot of hedge funds obviously believe in AI. I think there's a lot of them who don't believe in it, right, to be clear. But a lot of them that have done the best believe in AI.

If they don't believe in it, why are they selling software everywhere?

Oh, you mean selling software stocks?

Yeah, yeah.

Oh, yeah. Why the sell-off then?

Yeah, I mean, I mean, of course, it's like an incremental thing, right? But anyway, so, so these hedge funds, like, and then the question is like, okay, if you believe in it, how do you manifest that trade? And so when you look across the ecosystem, I would say Almost all my clients sometimes think our 2-years-out numbers are too high. But like, there's, there's like, Leopold's like, your numbers are too low. And so it's like, it's like, it's like, in general, right? And I think, I think like, if you think about how much do you believe in AI, and what's your access to information of AI, you know, there's not many hedge funds who live in San Francisco and like fully breathe and live and understand it. And then, and then depending on how much you believe in AI, how do you manifest that trait, right?

Are you surprised that more hedge funds Wouldn't, like, even just smaller shops wouldn't say like, hey, this AI thing seems like it's gonna be big, maybe we should set up in San Francisco?

There's a number of people, right? So we're, you know, we're getting an office together. Leo pulled myself, Dwarkesh, and then a client of mine, another hedge fund, and they have one analyst here. And it's like, and there's like a number of other hedge funds that are like hiring analysts here. But, you know, being plugged into the AI ecosystem does not mean you're just in San Francisco, because you can just walk around and talk to like doofus like startups and VCs and like not actually You know, see what's coming down the pipeline.

Yeah.

And you have to combine it with all sorts of information, right? You have to have a good tune with like what's going on in Asia supply chains. You have to have a good tune with what's going on in New York.

Sure.

You have to get tuned with like what's going on like in the financial markets, right? And then like what's going on in credit markets and what's going on in all, you know, the data center energy blah blah blah all these different industries. And so it's it's it's actually not like so simple to like be in tune with what's going on in AI. You can easily get like headfaked, right?

Yeah.

You know, for the longest time, people were thinking, you know, Adobe's an AI company. And like, you know, it's like, for a bit, like, Adobe was going down on AI, and then they like launched a few AI features and the stock skyrocketed. And then now it's going back down again, because people realize, oh, wait, no, actually, it's not an AI company. Like, I think it's the manifestation and thought of like, what is actually gonna— the world gonna look like if Anthropic 3x's its revenue again this year, OpenAI 2x's its revenue again this year, or you know, by the end of the year, how many people even believe by the end of the year, AI startup revenue is over $100 billion? I think that's an insane statement for a lot of people, but that's what it's gonna be.

Yeah.

Right? And who believes that number, right? It's like, very few people. And then you, you draw the continuation. It's like, who believes, you know, and when Anthropic says in their funding, like, hey, we're gonna have $300 billion of revenue by the end of the decade. And it's like, actually, I think that number's too low. Because, because the economic value of what they're going to create is going to be in insane.

Yeah.

Um, and, and you tell people, oh, you know, OpenAI is gonna have 18 gigawatts or 16 gigawatts by the end of '28, and they're gonna be able to pay for it. And that's like, well, that's $300 billion of spend. How are they gonna pay for it? It's like, you, you sweet summer child, don't worry. Sam can raise. They're gonna blow up on revenue. They're fine. Right? Like, it is like a bit of a vibe thing. It's a bit of like, you know, irrational exuberance almost, right? Like, um, Liv holds in his, you know, mid-20s. Like, I'm 29, like, we are irrational, right? Because we have not lived through, you know, you've got these PMs who like—

You've never been that humble.

I don't know, like, we almost— my family almost went bankrupt in 2008, like, you know, because we lived in a motel and we almost foreclosed, and we actually did foreclose on one motel. It's like pretty bad, but like, yeah, I mean, I was still a kid, right? Like, it was like, yeah, I've never been humbled.

No, it's good to— I mean, it's good to live through that and understand how things can go wrong. Uh, what are you expecting out of uh, Zuck and Meta this year. We've been big Zuck defenders, especially— I mean, there's this pressure of like, oh, Meta is spending so much and yet they haven't created, you know, any, any AI product that's super compelling or that's really working. And our stance has, has generally been Meta's making more money from AI than almost any company in the world outside of Nvidia. So it's like, of course Zuck should be justified in saying, hey, this is real, it's big. Like, I'm going to, like, back the truck up and go all in.

Yeah, I mean, it's clear if you look at the most recent earnings, I think their CPM went up 9% when the consumer's weak, which means, like, if you were to, like, try and strip out, like, what is consumer spending increasing for CPM of ads versus what is the effectiveness of their algorithms, their algorithm got better by double digits in one quarter, right? It's, like, actually insane how good the algo's getting, right, at serving you the slop in the ads, right? Um, so, so, so in that sense, like the pig sound, the trough. I love it.

I gotta hear slop for the slop sound.

We're going all in on that.

All in on the farm.

Um, so I love it. Uh, so, so, you know, if you, if you think about it, right? Like, okay, Metas, where are they gonna like win? Right. Um, you know, I think if you have the galaxy brain take, it's like, well, they've got the best, like, Wearables coming down the pipeline. They're gonna put AI on it. Apple won't be able to put good AI on their wearables, so they'll seed it all to like Microsoft, Google, or Microsoft.

Well, the other thing is people, people have had this narrative, oh, uh, as AI gets better, the value of real-world experiences will increase. And I think that's a cool theory, but if you actually play it out, AI getting better means more content that's more, uh, like, uh, effectively crafted for you, more personalized. 100 times more content, 1,000, 1 million times more content, that would imply to me that people will just use digital products more, which means more time on site, more time in the app for Meta.

So I don't know. I mean, I'm with you entirely. But I think, I think like the Galaxy Brain take is that you're just gonna have a wearable and that's gonna have an AI assistant. OpenAI is trying to make wearables, you know. You know, there's, there's, you know, everyone's trying to make wearables. Google is, et cetera, et cetera. I think Meta will actually execute and then they'll have A good AI, and then you stack on like a few things. Right? How do they get users? Well, we've seen at least if you look at the user metric charts, Google's use you know OpenAI's users were growing, growing, growing. They were gonna hit a trillion by the end of the year. They eight hundred billion. Why did they not keep growing in the last quarter? It's because Nano Banana came out and they took all the incremental users. Right? And and likewise, if you go look at like you know Gemini Three didn't actually make Google grow that much. It was Nano Banana and then. Pro or 2 or whatever it's called, right? Those are the ones that made them really grow. Meta's licensed all of Midjourney's code, data, models, right? One. Two, they're like actually just like focusing hardcore on—

Was that a billion-dollar-plus deal?

The number is undisclosed. Midjourney still exists as a company.

No, it felt— it looked to me like effectively a massive exit, but the best-case scenario where they can just keep kind of being artists?

I think, I think if you had me guess, I would bet it's a bill over a billion, right?

Every deal that Meta did was over a billion, basically, like whether it's an employment contract, a licensing deal, an acquisition, everything had a B after it.

Well, so the interesting thing is Meta—

You're missing a zero again. Don't ever miss the zero again.

Yeah, every discussion was how many billions are we spending on hiring this person, buying this company?

Well, Meta interestingly, um, has gone downmarket for compute because they, There's not enough compute in the big size deals, so they've actually gone and like bought like small clusters. Oh, because it's like, well, I want more compute.

From like long-tail NeoClouds?

Yeah, just like, yeah, from a longer tail. Okay, because that's the only place they can get the compute they need. Interesting. Because, you know, they've already like went out and signed big deals with Google and CoreWeave and so on.

Is ClusterMax 3 gonna be a smaller chart because of consolidation in the industry?

No, it's— there's more.

It's gonna be bigger.

It's gonna be bigger, bigger. But, but, you know, so Meadow— That's ominous. It's ominous. So I think Meta will, you know, capture consumers through generative. If there's more content, people are just gonna go to the content marketplace, right? The creator of the content captures less value as there are more content creators and more diversification of content, right? And so I think Meta just wins by being a platform, right? Google does too, and ByteDance does too, right? But like, those 3 win by having the platform. And then the real question is, can they get in the assistant productivity game, right? And I think this is important.

And through that, effectively search. Like, if you're an assistant, it means that you can, like, there's some commerce happening.

Well, they span out and poached a bunch of people from Google. So this wasn't in the media much, but like, they actually poached Google search people with similar-sized deals as like these crazy—

AI researchers.

Yeah, and I always, I always, I, you know, demoing, demoing any of the wearables, you can imagine like Meta wants you to walk around in the world and see like, oh, what are those headphones? And like while we're talking, I just hit my little thing and buy it, right? And it's like you didn't even necessarily know that it happened. Uh, but like of course Meta is going to want to monitor that.

Everyone knows those are the Sony MDR-XZ272 462.

Dude, I've been screaming about them like doing some proper marketing.

Branding so bad.

It's literally like their over-ear is like WH-1000XM5.

It's crazy, dude.

And then their in-ear is like WF-1000XM1000. It's like, dude, just call them like Bravia Buds and Bravia like headphones or some shit.

Well, China just bought Sony, right?

Yeah, yeah, Bravia brand's actually a Chinese company now. Sony sold their TV and PlayStation Buds. Yeah, yeah, yeah, PlayStation Walkman. Oh yeah, Walkman Buds.

Come on, like something, something for sure.

Anyway, uh, anything else, Jordy? No, this is great.

I'm excited for this weekend.

Yeah, yeah, super excited.

You guys, what are some plays? We don't watch a lot of sports.

What are some plays?

You are a football guy, right? Yeah, yeah, yeah. Grew up in rural Georgia, so I like football. High school football was the thing. College football was the thing. I think NFL is a little less soulful.

Sure.

But you know, now college football has the NIL, and so it's also soulless to some extent. It's fine. We enjoy it, you know, primal desire of seeing heads clash.

Yes.

You know, and sometimes that manifests in like, you know, Twitter drama and sometimes that manifests in real football.

Yeah.

All I can say is fuck the Patriots.

Okay.

Whoa. Okay.

Okay.

I'm kind of bummed. We're going to, we're going to, since we're going to be at the game, we're not going to really get the great experience seeing the ads. I'm going to be like glued to my phone. I want to see all the AI, the different—

Well, don't worry. I got some more ads for you. Thank you so much for coming.

Thank you so much.

Great segue.