All right. Hello everyone. Welcome to episode number one of the SemiAnalysis weekly podcast. We don't know exactly what the name is or if this is ever going to be public, but we're testing things out and we're going to talk about what happened in the week that was. Today's Thursday, February 12th. We had three articles that came out on the free tier newsletter: Claude Code Rising, CPUs Are Back, Memory Mania. We'll try to talk about all three and other stuff that's going on. By way of introduction, I'm Jordan Nanos. I'm here with Doug O'Laughlin and Myin. I don't even know how to pronounce your last name properly.
Have you ever... Thank you in Chinese.
My last name is literally...
I'm a very grateful person.
Thank you for coming on this podcast today.
Yeah. Yeah. So, is it like thank or thank you thank you? What would the... No, it's...
If you say Xie, that's thank you.
Yeah.
Okay, let's keep it on topic. This is not transistor radio light. So, yeah, let's go through each of the things. I think first, Claude Code, bro. I don't want to talk about Claude Code. I've talked so much about Claude Code. But I do want to talk about the fact that we're on the press release. I think it's kind of a big deal. It's very rare to get linked on a press release anyways, and we're on the Anthropic raise press release. It's kind of cool.
Yeah. We meaning you. So Doug, you made a chart that made it first into our core research product a few weeks ago and then made it into the free tier newsletter that tracks how many contributions to public GitHub repos are, how many commits are being made authored by Claude Code because it signs off as a contributor when it, you know, sends a PR. So that number, you know, went from like 2% to 4%. And then they included that in the announcement release. So like walk us through that. Like how did you get the idea to do that in the first place?
I read a tweet where it said, "You idiots, it all says it's committed by Claude Code. You could turn this off." And I was like, "Wait, can you scrape GitHub?" And then I just asked Claude Code and that was like, that was it. I was like, "Oh my god, can we make a daily series?" And then effectively I realized like, wow, this is like, I've definitely been rushing inside to make sure we were the first to put that chart out because I was like, dude, anyone who's paying attention can figure this one out, like especially with Claude Code, but I was like, we need to be the first. And so I mean like, I don't think this is like exactly news. We're going to definitely be publishing the data for, you know, our paid subs or like, you know, our institutional research tier, but we'll probably have some kind of public dashboard as well or some kind of public facing information. It's like already almost 5%. Right. At least according to the most recent one we did, it's like 4.7 or 4.8 or something like that. It's still growing, but yeah, that's how I got to the idea, just doom scrolling and I was like, wait, wait, wait. And then yeah, me and Claude Code made the chart and then obviously wrote the article and yeah.
How much of the article did Claude write in addition to making the chart about itself?
I don't think I personally still want to be writing with a human voice, right? We got to train these models on something. We can't just be sloshing all this stuff back around. I do like it for outlining more than anything else. I think the very first round of it, I was still very excited. I was like very peaky on my excitement about Claude Code. So I dictated most of it. Then I had the dictation get cleaned up by AI and then I kind of wrote like, okay, make a slop outline based on this, and then I wrote over it using my voice, just like manual tokens, you know, with the fingers. And then afterward, we sent it to the powers that be, aka the internal SemiAnalysis brain squad, and then like people wrote in sections, people gave feedback. At some point, I think the most heavy lifting that Claude Code did do, and specifically did this all on Claude Code, like screw vanilla chat or anything anymore, was mostly the outline stuff, I think. Because when you write something so much, you just get lost as hell and you're like, "Okay, this all makes sense in your weird little brain." But I'm like, "Hey, just do a fresh go through each kind of piece and walk me through this." And I do sometimes have it do some editing to take like, what's the least relevant portion of this paragraph? I like to usually rip stuff out. But that's just me. Sometimes SemiAnalysis loves to add stuff in.
Are you playing with Claude or with ChatGPT or Gemini?
No, I agree with Doug. I like having a human voice. Although sometimes I think about like, how do I express this concept? I use an LLM and it's like, okay, that's a really good way to express it. But I mostly use it, to Doug's point, like human voice. I wouldn't generate like a whole piece from an LLM. It's just like specific ideas I'd flesh out with an LLM, like make this sound more elegant. I'd use an LLM.
Maybe I'm too much of a boomer.
I think it's our last artisanal point, you know, like it's our hand-churned butter, you know, that's all we have.
That's all we have. This hand-churned butter is all we got. And also just like for context, I don't think people realize Myin writes so much of the bangers analysis, like essentially anything accelerated, you know, like do you think Dylan's writing these articles? No, brother. Like it's a team.
[clears throat] I just love... people still believe, like pretty deep into, like I want to say even 2024, it's like, "Dylan, I just love everything you write." And I'm like, you know, it's just like very hard for you to believe. It's like, okay, so we just put out three 10k reports, you know, like 5k word reports this week. It's like, no, there's like a lot of people on these author bylines and there's definitely a lot of heavy lifting. Like obviously Dylan's giant brain behind so much of everything we do here. But yeah, Myin's like an OG SemiAnalysis person and he's written some of the greats.
When you read something like one of our big 20k, 10k word bangers that obviously has a lot of contribution from various people, you can tell. Can you tell whose voice it is?
Like who wrote this? Like did...
Did Dan write this? Did Jeremy write this?
Yes, I can tell. I can tell at least, I can kind of tell.
I'm not at Doug's level. I can tell certain voices or certain people use certain jargon or choose to use the ampersand instead of saying A-N-D or choose to, you know...
There's also British spelling too, like the British spelling definitely pops up, that's a thing, yeah, that's a giveaway. Like Jeremy... kind of British spelling.
Yeah. Yeah. Most of it. Yeah. Actually, our prime minister, Mark Carney, is like going on a crusade right now, making all of his employees use the British spelling of many common words. So it's interesting. It's a return back to the...
So that's like the reverse of Dylan.
Yeah. The data centre centra or whatever. And then my favorite is I think Jeremy has the most typos because it's very clear everything in his entire keyboard is focused for French. Like dude, because it's like you think his autocorrect is going to catch the English typo? No way, bro. So like he definitely slips the most typos I think.
But yeah, for Jeremy, it's if an entire sentence is highlighted with a link to an article followed by the next sentence also being highlighted with a link to the article so that the whole of text is just links to previous articles. That's a Jeremy play for sure.
Yeah. Yeah, that's definitely a Jeremy play. Yeah. So, yeah, but there's so many people who go into all these things. It really does help. I mean, and actually I think that's how all these articles start. Most of the article is like a 2k word thing, 1k word, and then we're like, "Hey, you know, it'd be really nice if you help on this section." And someone helps on a section and they add like 1k words and then like each incremental thing gets added like another like whatever 700 words and then all of a sudden we're like, wait, this is 6,000 words.
Whenever someone's like, "Oh, let's do a brief article on this..."
Frankenstein.
Dude, I think I'm proud. I feel like the Claude Code one was really short because I was the quarterback and I kind of just rammed because I did. I was like, "We need to send this chart out. Like, we need everyone to use our number." I needed that so badly. So, definitely rammed that one out.
So let me try something here and show a bit of the backstory. So just because we're talking inside baseball about how SemiAnalysis works right now. Here's the first chart which went out on January 20th in Core Research. And then here's the one that went out public. So just two weeks later, this one's February 5th.
We're all the way up at around 4%. And here's the Anthropic press release we were referring to where they announced their 30 billion raise at a 380 billion valuation and a link to the article that makes that estimate. So yeah, just for reference there. But maybe like... to be clear, there's some articles that are going to the free tier and not going to subscribers first. In other words, like the CPU article that we just kind of put out about the return of CPUs, I guess, and their importance. That is like a very different article, you know? It's like a bit of a history as opposed to a comment really about what's happening right now or at least it leads in with history or something like that.
Yeah. And I think, I mean, it's mostly because, you know, CPUs are history, right? I mean, sorry...
Not anymore. They're back. Yeah.
No, I definitely think that people, and it's kind of like unfair to be like, oh, all we do is focus on this like paid thing. Yes, that's our jobs, but I definitely feel like we do so much content for free. And I think the CPU article, I don't know if it ever got mentioned or written about in any kind of paid, like higher paid tier. It just went straight out of, you know, straight out of the authors, straight to the newsletter. Same with I think the TSMC Apple piece that was like straight to the sub, to the newsletter. Yeah, honestly, Myin, do you have any kind of takes on the article other than that the CPUs are so back?
Yeah, I mean, I think it ties into what's happening with Claude Code, right? So we need more CPUs for basically RL. They're involved in the RL process. And then also with all these agents running around, you need CPUs to invoke things that are like outside the model, right? Like if you want to do a search, if you want to manipulate something on your computer, that's all CPU. And then we're going to create all this software and what are they going to run on? They're going to run on CPUs, right? So that's what's really driving CPUs coming back. And then also there's the traditional, you know, like what CPUs have been used for. There's, you know, host processes that run GPUs and of course we're all deploying more GPUs, so that's going up. So everything's happening with CPUs. Even, you know, even Intel is sold out, right? So that tells you how many CPUs we really need.
Yeah. Intel things are tight.
Yeah. Yeah. That's actually, I mean, that's been kind of the... and we see this in like the entire accelerator space, right? The marginal supplier actually benefits massively and I just don't think anyone would have ever guessed that, you know, the marginal supply, like essentially that we would hit what is the largest, most mature install base of compute, CPUs, and be like, yeah, actually we're going to run out of supply and the marginal supplier being Intel is actually going to catch a bit. I mean, it's kind of been incredible. And yeah, I think part of this story that's interesting too in terms of like, you know, all this stuff is some of the... I think this round though, ARM is such a bigger threat than it's ever been and that's kind of been like the other part of the story that's like very interesting. AMD continues to crush. This is kind of fragmenting, but the Fenix CPUs are definitely kind of an interesting change, especially in terms of how ARM's business model is going to adapt. And that's... but even then, even with more competition, doesn't matter. Intel's still sold out or probably will be sold out.
So I don't know. Just in terms of the article, did you guys see they pulled up that picture of Pat Gelsinger getting a massive tattoo of VMware on his arm article?
Yeah, dude. I did see that. I think it's sick. I also think it's so like, you know how like gutted VMware is after post-Hock Tan, like the company he has been CEO of is like, it's dead.
There's Pat on display. Yeah.
Yeah. I mean, I don't know if we want to really do the eulogy of VMware right now. I'm sure it's going to continue to be used in the future, but like all of the growth in AI is about people consuming compute through an API via tokens, not by every single Fortune 500 company spinning up a bunch of virtual machines in the cloud the way it was in the cloud era. Like the growth is in the form of CPU data centers supporting cheap GPU... [clears throat] ...data centers. They had a great picture in that article about the Fairwater campus from Microsoft showing the GPU building versus a CPU building. Let me throw that up on screen as well just for the people who are watching online.
I like how this is going to be the dummy version of these articles. You don't have to read the whole thing. You can just listen to our weekly. I mean, that's actually probably the best thing we could possibly do. And it would force me to actually read all of our articles. Boom.
Are you saying you didn't read this article and I'm just showing you this picture for the first time?
I super skimmed the hell out of it. If I'm being honest with you, there's too much content.
[laughter] I guess this is the concept, right?
So much we can't even... Yeah, we can't even read it.
Yeah, seriously. Okay. I think there's like three articles going out today, like smaller ones, and like I mean I don't know...
You know, it's actually becoming such a constraint for the core because like the core research weekly is like the definitive register of everything that we write. It's [snorts] like becoming time constraining to make the table. The table is like what everyone hates the most now of the weekly is they're like... because also because up until I think they figured this out, like literally today I watched them figure it out, you had to manually do it. Like there was no way in WordPress to essentially not manually create this table. And so in the beginning it's like whatever seven or eight things a week and now we're like 15, 20 things or whatever. And then it's like dude, someone is essentially just manually typing all this stuff out and they're like, I hate the table, but the table's essential. Anyways, yeah, that's... keeping up with this is becoming a real deal.
Okay. Burning question, Doug. How many agents do you have running right now while we're recording this?
Um, I have some... Okay, so I actually have an agent team going right now. I have seven in one task and then I have some idle. I have like one, two, three, four, five, six. So 13 windows open, but I would say like maybe... no, seven windows open. One of them with seven sub-agents or whatever, the agent team, and then like two... well, but like three of the seven windows are really the hot ones. Some of them are like I'm kind of like waiting for later...
Star players on the team.
Yeah, the star players. I kind of want to get into teams but I just hate the control-B arrow thing. It's a pain. So I'm a boomer.
Control-B is asking too much.
It's, well no, it's command. No, it's just like... but then it's like it just gets like really... and then I like... it's the scroll. I don't know. I'm too much of a boomer on that one. But yeah, just 13 right now. Okay, let's move on to the next one. We obviously, the topic that everybody's been talking about recently is the increase in memory prices, like just the absolute shortage. And I think it's kind of like been this game of whack-a-mole where you figure out the GPU supply chain just to ship like whatever tens of millions of H100 equivalents per year and then HBM, you know, and CoWoS capacity is like what drove that. Now where's the next like long pull in the tent and it's kind of like everybody's turned to either SSDs like the drives or even hard drives or now memory. I think the question on people's mind is like one, is this happening and when will it let up? And then two, what's next after memory? You guys have thoughts on that?
Yeah. So basically, again, it started with GPUs being the main driver of memory, sort of incremental memory demand, and that's mostly HBM. I think what's really different about this cycle is in the past we haven't seen a new memory product that is so much more difficult and more intensive to manufacture. You know, with HBM, the amount of bits you get per wafer is far less than say conventional DDR. A few reasons, which is number one, you need to dedicate more area to your TSV keep-out zone, which is the TSV, the wires that are kind of built through each stack so that you can have a vertical stack where you deliver power and signal. And then also because of the performance price of HBM, the yield is terrible or much worse, less than conventional DDR RAM. And then also you have to stack these things in an 8-high, 12-high stack and you lose yield on the package. So all in all, like you know, you get like several times more bits out of a conventional DDR wafer than a HBM wafer. And then the more the HBM band grows as you ship more GPUs or XPUs, then the more wafers that the industry has to dedicate to HBM, and then that sort of reduces total industry bits. And then conventional DRAM is picking up because again with all CPUs you need DRAM, so that's driving conventional to be tight. And then you combine those with very little incremental sort of wafer starts because the memory manufacturers after the CoWoS hangover, they've decided, okay, we're not going to invest more. So there's literally not much clean room space that's available to actually put equipment. So you can't... so like we're adding virtually very, very little supply. Both land is moving and that's what's causing things to be so tight. And people just realized this about, was it five, six months ago? And now, you know, that's driving things to be pretty whack in the DRAM market. And then similar for SSDs and hard drives. I mean, SSD, like NAND supply has been... I mean the NAND industry has been much worse than DRAM. I peaked like half a year ago, all the industry players were barely making a profit or I don't think they were profitable. No one added supply of course. And then all of a sudden you have a lot of that incremental AI demand hitting and now things have just become tight. And the problem with like NAND is because it's less profitable than DRAM, the DRAM manufacturers are probably going to add more incremental capacity to DRAM first before thinking about... it's crazy for the next year, for a few years at least.
Yeah, for a few years. I think one aspect I wanted to add on the historical side is it really was... it was literally the worst cycle since 1996. It's like the worst cycle of all time in memory. It literally was like number two. 1996 bankrupted like a third or a third of the industry pretty much. So it's pretty much the worst of all time and then it whipsaws to the best of all time. And like there was an example of last cycle that this was something that they did historically different is they literally shuttered machines, meaning that they're like, hey, we're just going to stop at like, you know, not only are we going to... like usually what happens is bit demand growth rips the entire time and they're like, yeah, actually no, we're not even going to do... we're just going to turn off supply. We're going to take the... we're going to just take the charges. Usually they just have underutilization charges. They were like, we're going to save the clean room space. Like that is historic. No one has ever been, you know, when you have something fully deployed, usually it's cheaper to just keep it on. But NAND was so bad. They're like, no, no, no, no. Let's try to convince everyone to turn it off because it's so much supply. So we go from that market of being like we're turning off the spigot in order for us to fix the problem to we don't, you know, there's not even a place for us to turn the spigots on anymore. And so it's kind of crazy like there's a giant clean room shortage and it's... I mean, there's no other way to put it. It's historic. It's crazy to watch. I think we mentioned this in our Slack the other day, but like I want to say that memory NVMe prices for SSDs went up like 20% in a day recently. Like the prices are also going up 100% like, you know, like [snorts] like month and there's a little bit of panic buying, just a wee bit. But I think it's kind of crazy, you know, this is... I mean it's definitely historic for sure.
So I talked to some other guys about this and somebody made the point about like expansion, right? How do you actually add more capacity? It requires the memory manufacturers to pour in some capex to build some new fabs in order to build more memory and lighten up the shortages here. You said that previously when these cycles happen they get caught because they pour in too much capex, they build too much capacity, and then suddenly there's no buyers and the prices drop and then they go bankrupt. So then people start asking for a call of like, when's the top? When are we going to see the top in the current memory cycle? And it seems like based on history, the top is going to be as soon as Samsung capitulates and decides to start building out some more capacity because they are historically the last one to do it after SK Hynix and Micron who are currently building out capacity and Samsung is still kind of staying with the plan and not going to take the bait. At what point do you think they capitulate and decide to actually start building out capacity here?
I think... how high do we need to go? I think they're going to build capacity. It's still not going to be enough. I mean, it's going to be like... I think the supply-demand gap is like the largest it's ever been and it gets worse from the work that we see. What's interesting is if you want to talk specifically on the Samsung side, in the memory down cycle everyone was like, please Samsung, stop making memory. Like SK and Micron on every earnings call was like, we really think the industry needs to stop making memory. And Samsung would be like, no, we are going to keep making memory. And when they finally stopped making memory everyone's like, oh, that was the bottom, that was like the absolute bottom, the pico bottom. And what was crazy is that I think Samsung did it for three extra quarters than they historically have. And so I expect this to happen this cycle, but I still think Samsung's going to add... it's like...
It's not enough. It's just not enough.
Yeah, I think they want to add capacity. It's just there are constraints like they don't have many clean rooms. They need to build some clean rooms and that's going to take time. Then also just on the tooling side, I mean, ASML can only output so many EUV tools every year. And you know, one of the debates we're having internally is like how many — between TSMC needing a lot, the memory guys needing a lot — like how many EUV tools can the industry get and how is it allocated, because logic is also tight.
Yeah. Um, and you made a comment earlier that logic is effectively less complicated to manufacture than memory in the current state.
Uh, no. No. Logic is probably more complicated to manufacture, especially with — yeah — than DRAM. Doug's on mute, but I mean I was maybe referring to the HBM part of the process where you're doing advanced packaging with the logic and the memory.
Oh yeah, it's — I think — yeah. I think it depends. I'd say the packaging in terms of like stacking things, you know, like stacking 8 DRAM on top of base or 12 DRAM on top of base, which is what's most common now. I'd say that's much more probably more difficult just because like if you mess up as one layer on the way, you need to — like the whole thing's done. So like the cumulative yield loss of like doing something like 8 or 12 times, that really kills you. And that's something that yeah, TSMC doesn't have to deal with, because they have to do like really big cost modules that are difficult to yield. They do stuff like hybrid bonding, which is also you know very complicated. But in terms of just like doing such a high stack, like a large 3D IC assembly, yeah, it's really hard for guys to do it.
But okay, so I want to put some hype on the logic side. Let's not forget another way to think about this. And I don't actually have — I mean actually dang, I think I came up with this on the fly. So the capex per wafer starts, right? Per 100k wafer starts, that's one way to keep it levelized. I think it's two or three times higher for leading edge logic versus DRAM, right? But so that would mean hey, leading edge logic is, you know, let's say two and a half times — I'm vibing it out. I think we have this number internally. I have no idea off the top of my head. It's like, you know, on me. It would be two and a half times more, but you can argue the three times trade ratio of DRAM means that logic or HBM is the most expensive per 100k wafer starts in the entire world. So yeah, I guess I remain defeated. I think HBM's the hardest.
Right. Does that affect the allocation of the EUV tools to the buyers? In other words, who's willing to pay more is the people who can make more money out of the output from the tool.
It's a good question. I think the most sustainable thing is like because logic needs memory and memory needs logic within a system. So you want to make sure that everyone gets their fill so that the industry as a whole can maximize chips that go into your completed systems, otherwise you'd get some weird balances. So that's a question for like ASML planning their strategy team and they should probably buy a box to help them make that decision.
Yeah, to help them make the decision they need to buy our models. That's the right answer. Yeah. I mean, I think the industry itself is not — I would — I mean like one of my favorite parts of the semiconductor industry is I don't perceive it to be massively zero sum. Meaning that like one random person is like, 'Yeah, screw everyone else. We're just going to gouge the [__] out of whatever.' Usually what happens is a technological innovation. You give them great margins, but at the end of the day, it is about making more chips. And this memory problem is becoming such an issue where it's like, hey, we're not going to be able to sell the right ratio of GPUs to DRAM. So for ASML, the strategy plan is they're, you know, their maximizing is essentially to sell the most chips possible. And so they're going to try to make the right trade ratio, but the trade ratio itself is honestly a very hard question to answer. And I don't think many people know. And all in all, like even the hyperscalers, Nvidia, everyone's aware of this memory problem. Like we got to get some, you know, we got to get some EUV machines into the DRAM processes. And actually my galaxy brain take is it's going to pull forward 3D DRAM. I mean, it's not going to be before 2030 to be clear, but like 3D RAM is supposed to be 2035 or whatever, right? But I think it might actually be closer to 2030 because 3D RAM doesn't use any EUV. And so that's like a galaxy brain way to increase total throughput of the entire system. But we're pretty far from there.
How about alternative accelerators that use no DRAM or HBM and instead go all SRAM and get the trade of six times more money for two and a half times the performance? How about that segue to talk about Opus Fast and like saying eggs are a bit more expensive, I'm going to eat caviar for breakfast instead.
