Ep. 007 - 扼杀 AI 热潮的3个瓶颈（Core Research）| Nick Doyle、Nigel Chiang、Konrad Wang、Jordan Nanos

All right. Hello, everyone. Welcome back to SemiAnalysis Weekly, episode 7. We're here with the Core Research boys, Nick, Konrad, and Nigel. We're going to chat about what Core Research is, talk a little bit about our latest newsletter note, which we covered a little bit last week, and then go into some of the research that these guys do. Welcome to the show, guys.

Hey, thanks for having us, Jordan.

Happy to be here.

Awesome. For those who don't know, let's start with a quick introduction to what Core Research is. This is probably our flagship product. The boys are sitting back-to-back in New York City, in the office, in the pit, heads down and grinding every day, and we're going to try to avoid mic feedback during this podcast. Outside of that, what do you guys work on every day? Give me the high level.

1. Core Research Distills The Data

Yeah, let me jump in there, just because I feel like I explain this a lot to our clients. Even our own clients don't really understand how it works when they're first starting. The way that I explain it, hopefully in a way that makes sense, is that this side of the SemiAnalysis research house is structured into 10 different teams. I call them verticals. They each deliver their own model. These are extremely detailed models supported by very technical drivers, and each of the team members on each of the teams is extremely intelligent. The talent pool here is incredible.

What these different teams deliver with their models and technical updates gets sent out to clients via email. We, the Core Research boys, sit beneath all 10 of these verticals and try to distill this technical information into actionable insights for investors. Our customers, the people we interface with the most, are investors: big funds, hedge funds, long-onlys, and even some quant guys are starting to drip in here. That's a little bit of what we do at a high level. Does that make sense, Jordan?

It makes sense to me because I know what you're explaining. Hopefully, it makes sense to the audience, too. I also know what your backgrounds are. You guys have diverse backgrounds, both working in finance and as investors on the sell side or buy side, but then also working in the industry a little bit. Thinking about you, Konrad, with your software engineering background as well, it's not like you guys are strictly focused on finance. Maybe it's just the fact that the industry vertical lead, or model lead as we would say at SemiAnalysis, takes the lead on that range of research, and then you guys distill it down to the investors who are trying to make decisions.

Let me set the stage here. We put an article on the free-tier newsletter about GPU rental price increases and how much things have changed. Dan and I were joking last week that it feels like trying to get the last flight out because prices are rising, and then they're just gone. There's no more capacity left.

But you guys have been swimming in choke points for the last few weeks or months, right? Everybody is trying to understand what happens when all of this CapEx is poured in by TSMC, by many of the players and hyperscalers on the cloud side, and by all the labs. There are some choke points that bubble up, right? Maybe we can walk through those one at a time when it comes to things that are top of mind, like memory, construction, and wafer fab equipment. Konrad, maybe we can start with you, going all the way back to the beginning of the supply chain at TSMC. You've got your TSMC jacket on.

Yep. Yep.

Very nice. What are they thinking about most right now when it comes to choke points, in terms of how they can deploy their CapEx and expand?

2. The Supply Chain Choke Points

Yeah, absolutely. I think the AI infrastructure supply chain is constrained at almost every layer. Obviously, with the big article today, the conclusion of the choke point is that we're seeing higher GPU rental prices, right?

It's also pretty interesting if we look at the price action for stocks. People don't like the spenders, like the hyperscalers. They're all down big year to date, and then people are trying to buy whatever has the biggest shortage. From my side, and then Nigel and Nick, feel free to jump in and add your coverage as well, I think TSMC N3 is where everything starts. That is the biggest shortage. It's sold out through 2027.

Hyperscalers are all doing internal allocations to offset the trade-offs. That has a lot of implications for clean rooms and wafer fab equipment. I know Nick is doing a lot of big research. We haven't published it yet, so I can't really dig in, but there are lots of implications on wafer supply and demand.

The second big one, at least in my coverage, is lasers—namely, InP-based lasers used in optical transceivers. A big theme this year is co-packaged optics, where they use continuous-wave lasers, these very powerful lasers that use a lot of InP. So there are lots of shortages there as well.

The third one that I see is PCBs and CCL substrates. For our audience that doesn't know what they are, PCB stands for printed circuit board. Every single chip, GPU, or accelerator has a printed circuit board, and the sizes are getting bigger and bigger. We're not seeing a lot of substrate availability. CCL is copper-clad laminate, which is the material used to build PCBs. There are lots of shortages there as well, extending even to the drill bits used to make these PCBs. So at least those are the big 3 that I'm seeing. Nigel and Nick, feel free to jump in, or Jordan, if you have any follow-up questions.

No, all good, man.

Yeah, I want to jump in and ask Konrad more questions, just because it almost takes us a step back and a little bit higher level. I think, first, it speaks to the incredible nature of SemiAnalysis as a firm. All 3 of us come from different backgrounds, generally finance backgrounds. I wasn't a specialist in wafer, raw wafer silicon, or in the hyperscalers. I barely looked deeply at hyperscaler CapEx until I came to this firm.

Konrad wasn't diving into co-packaged optics testing and PCBs until we joined. It's through the incredible talent pool that's sharing ideas and collaborating that we're able to really accelerate our learning and, of course, use Claude Code—we can get into that later—to take us to the next level and to the leading edge of what the bottlenecks are. Everyone cares about bottleneck investing lately.

Going back to the PCB question, what are PCBs, why do they matter, and is the PCB environment changing so much? Why are we seeing this become a bottleneck? What about the technology is so different with GPUs that we need different substrate materials?

Yeah.

Also, just to double-tap on Nick's earlier point about us really going deep on semiconductors, Dylan Patel, our CEO, was on Jim Cramer yesterday. Jim Cramer said they think we're the god of semiconductors, and that when we bless something, we're apparently always right and we're the most honest guys. I don't know if we're always right, but—

I think you're really—

We strive, yeah. We are truth-seeking for sure.

Yeah, we strive to be as objective and go as detailed as possible.

Specifically, Jim referred to the SemiAnalysis research as the gospel, and then Dylan tweeted out asking everybody if he should short himself or how he could short himself.

Short semis?

Yeah.

Please don't. Please don't.

All in good fun. I think I'd probably go find some local San Francisco craft stores that are making these one-off T-shirts, flowy T-shirts, and flowy pants that Dylan uses everywhere. I think we've got to find a way to short those boutique fashion stores where he's getting all his nice clothing.

Yeah.

I don't know any other Dylan Patel-specific stuff. Maybe headphones?

Headphones.

Maybe headphones, yeah. The flowy pants are a big hit in New York, I think.

Yeah, we need a proxy for this guy. Hot pot, headphones—something in the San Francisco area. I think they'll do fine without him anyway.

Let's dig in on PCBs. Let's use that as an example because maybe you could run us through what they are and why people think about them. If people think of a GPU shortage, they think of NVIDIA or Amazon.

Right.

Who's a PCB vendor? What is it used for? Can you take me through that?

Yeah, absolutely. I think the biggest ones are all Taiwan-based local suppliers.

3. Why AI Boards Are Bottlenecks

PCB itself—PCB stands for printed circuit board. Think of it as a board that connects different parts together. If you ever build a laptop, it’s the green or black board that chips are mounted on.

AI server PCBs are extremely complex. Commercial electronics might have fewer than 10 PCB layers, but as things get more intense, you’re looking at 30, 40, or more layers. Those require specialized materials and more precise drilling. Because they’re thicker, they require more drills as well. Broadly speaking, you can think of it as the motherboard of a server.

The raw materials that make PCBs are sheets of fiberglass, per se, with copper foil bonded on each side. Those are what we call copper-clad laminate. If we go even a step deeper, these fiberglass materials have different levels as well, like T-glass, low-Dk glass, and L-glass. These all signify different levels of performance. As you can tell, higher-performance, higher-grade ones have much lower yields, and because of that, they have longer lead times. We’re seeing a big backlog in all of this.

With that, these companies, because it’s only being sold by a handful of suppliers, are able to charge a much higher premium than before, when they were only supplying commercial electronics.

Is this the same style of shortage as memory or NAND flash? It sounds mildly complex, but producing fiberglass with copper at a high level, for a general audience, doesn’t seem quite as complicated as producing a GPU. Is it raw materials or production capacity? Are the PCB companies just being opportunistic and jacking up their prices? That doesn’t quite seem like the Taiwanese cultural way that we’ve seen. What’s the price action in the market?

I think it’s just that there are so few suppliers in this space. For example, if we look at the substrate side, there really is 1 major supplier, which is Ajinomoto Build-up Film, or ABF, the ABF provider. If you look at CCL, there are only maybe 1, 2, or 3 suppliers that can really make the really thick, very hard CCL.

For these materials, I think Doosan is one, and then there are probably a couple more. Those materials also have much lower yields because they’re much thicker. They’re much harder to produce: They need to withstand higher voltages, and the production process is a lot harder than it is for lower-end glass, for example.

I’m going to add that this just goes back to the fact that this didn’t matter a couple of years ago. It all has to do with the end of Moore’s Law and the move to GPU-based compute. Going to the next level of GPU-based compute requires all these new materials, and I think that’s the coolest thing, frankly.

Yeah.

With Nigel writing about construction companies and the power electronics guys, I don’t think they ever thought they would be selling into data centers. Now there’s this massive boom in TAM, which is pretty crazy.

Do you have any favorite higher-level stories about massive price action in some of these companies, or just weird stuff going on across clean rooms, lasers, and PCB construction? What’s a favorite story that you’re seeing right now?

Clean room hasn’t really moved. I think it’s the optics that have been the most chaotic.

Yeah. Optics are crazy. Lumentum is up about 100% year to date. AAOI is up 200%. I think this year is the year of networking. Last year was probably the year of ASIC accelerators. This year, everyone’s talking about CPO and lasers. Substrates are all up as well because of the shortages.

The big 3, right? Unimicron and Kinsus are both up over 100% year to date. In Taiwan, the stock market has a limit, so you can’t go up further than 10% each day. They keep hitting the limits. Quite wild. CCL, Elite Material, is up 70% year to date. It’s a crazy time to be in the market.

4. Modular Construction Accelerates Data Centers

I think in the industrial space, which is where I spend most of my time, the trend of data centers moving to modular construction is going to be big, and it’s still in its early innings.

Call it 10% to 15% of data centers today that are modular. We see no reason why that can’t go north of 50% if you really listen to what the guys at Meta and Google are saying. The reason for that is that SemiAnalysis has been really good at keeping on top of data center delays, and that has partly to do with labor shortages and other constraints.

What modular does is solve for that in a way, because now you don’t have to aggregate your field labor, which is the scarcest, most difficult-to-gather, and most expensive form of labor, and send it out to a site to actually construct the data center. You construct these modular blocks of electrical systems off-site in a factory and then ship them to your project location to be assembled.

We think the modular story is really going to be big. There’s 1 company that’s really good at that: Comfort Systems. What people don’t appreciate is that as a hyperscaler does more modular, it becomes standardized. It doesn’t make sense for, say, Google or Meta to have Comfort Systems and No. 2 doing it, because what’s the point of having 2 guys doing it if it’s going to be modular and standardized?

There’s going to be 1 standardized design that Comfort Systems has done for Google. So there’s going to be lock-in for a certain hyperscaler customer. I think this is really within the construction space, which we’ve recently done more work on. This is something worth paying attention to.

Can you comment at all on modular and its relationship to behind-the-meter power generation? It seems like some of those are intertwined, where, 3 or 4 episodes ago, we had Jeremy on talking about how next year, over 50% of new data center builds are going to be behind the meter for generation. That has to impact not just the shell itself, but also turbines and all sorts of switchgear and equipment that’s actually going to power the thing as well.

Behind-the-meter and modular data centers are all different expressions of the same thing: time to market. The faster these guys can stand up data centers, the better, because we’re in a supply-constrained environment for AI. The moment you bring GPUs online, revenue flows. Behind-the-meter is the same thing: Instead of waiting 3 years for that grid interconnection, you install power plants on-site to power those data centers.

Where it also overlaps with modular construction and Comfort Systems is the labor aspect. When we were looking into construction, we realized that labor shortages are probably most binding in gas power plants. If you talk to GE Vernova, Siemens Energy, and Mitsubishi, and ask them, “Globally, between the 3 of you, you have, I don’t know, 100 gigawatts of capacity. Why can’t you divert your slots from, say, the Middle East or Europe to the US, where your prices are much higher?” what they’ll tell you is that even if they wanted to do that, there’s not enough labor to actually build these power plants.

I think what Jeremy was talking about was how a majority of, or an increasing share of, data centers that are going to be built will be behind the meter. I think there’s also this unappreciated labor angle to it that’s really worth paying attention to. Some of these stocks are really moving on this trend.

What’s a stock exposed to labor in this space? I don’t know if you’re going to give away all this, but I can’t even imagine what 1 of the stocks would be.

Well, memes aside, I think there is only 1 pure-play gas EPC. These companies are called EPCs—engineering, procurement, and construction. That’s Argan, and we’ve been positive on the name for a long time. I think we first called that out when, Nick, you were working on that.

Or was it Konrad?

That was Derek—

Yeah, called that out.

In November.

Okay. Yeah, that makes sense.

The earnings a couple of days ago were interesting. I think the stock reaction on the day was up 9% or something, and the next day it was up 30-plus percent.

Really? Okay. Wow.

Just evidence of higher pricing power.

Because—

Because they have a shortage of labor, and it's coming through into their P&L.

Makes sense. Okay. The last thing that obviously jumps to mind, that we've talked about on what feels like 3 sequential shows, is memory. You guys kind of covered both ends of the spectrum: the inputs to the fab and some raw materials, and then the inputs and raw materials to the data center itself. But as both those things flow together, they actually create some products. Are there different ways in which the different components have been moving most recently when we think across GPU, CPU, memory, networking, and all the other stuff that's actually going to go inside those data centers when TSMC expands and pours in its $56 billion in CapEx, or whatever Shravan said they were going to do on last week's show, and that filters its way through the industry with more chips?

For memory, there are a lot of different angles you can talk about. It's been in investor focus for the last 6 months, and I think Ray, our memory expert, and Mats That Matter were a little early in calling out that capacity isn't coming online as quickly as people think. Even with these CapEx pull-ins from 2027 into 2026, it still doesn't impact supply.

Not to give too much away, but I think incrementally we're even more positive on the theme itself. I think we would defend the pullbacks in memory stocks because there are a couple of different dynamics that make these players even stronger through the cycle than in prior cycles.

Yeah. Let's talk about the stuff that's in the news right now. There's been a lot of discussion about how these OpenAI letters of intent that Sam Altman was signing could be impacting memory stocks and why sentiment is down. What's your impression of public sentiment around memory stocks right now, and why has there been recent price action in those stocks?

I think it makes a lot of sense why there's negativity. Historically, when we're starting to see these LTAs—these long-term commitments that lock in pricing—that has typically been the sign of the top. Everyone gets excited because demand is skyrocketing, supply is capped, and prices go through the roof. That's been happening.

Now we're saying, “Okay, we don't want to see any more price increases. Just give us a price, and we'll lock it in for 3 years, and then we can keep moving on.” When that happens, historically, the stock prices don't work anymore because it's kind of capping the upside. So it makes sense that the public is seeing that, because, like you said, there are hints of this happening.

But I think what Ray and we think is different—there are a couple of different dynamics. I don't want to share too much, but the main thing is the prepayments. We think that this time there'll be a really big prepayment aspect that completely changes the financials of the memory players through the cycle. Also, the pricing that those terms are based on is way higher than everyone expected, and I think that's not understood. That's why we're still bullish on the theme.

Last week, I had the guys that went around the horn. I said, “When we get to the top, there will be signs,” right? Do you have any creative interpretations for how you answer that question from investors—what's the sign of the top in any cycle? It could be the broad AI—

We just saw 2. We just saw Dylan go on CNBC, and we just saw TBPN get bought.

Oh, this is the top. New media selling out.

On my end, I'll offer 2 signals. I know we all have Twitter accounts, et cetera, and one thing I like to look out for is an increase in the frequency of people victory-lapping screenshots of their personal accounts. Usually, if you squint, it's all the same retail names. That's one.

Second, I think, as stock guys like Konrad and Nick, you guys will agree with me: when news comes out unambiguously positive, but the stocks trade in the other direction, I think that's also pretty indicative.

We really saw that with Micron, and that's another thing that was fueling the fire about, “Okay, this round is over. We're selling off an incredible print. Let's move on to the next thing.”

I just want to add that we have one of our employees, Wega, who's based in Taiwan, and he was telling me that his aunt has been getting a lot of alpha just by watching the Taiwanese Jim Cramers pitch stocks.

Back to your earlier point, Jordan, one of my favorite stock stories that I just remembered is these PCB drill-bit companies. They are literally drill-bit companies, and their stocks are up. I just looked at the chart. It's up 92%.

So if Wega's aunt is making easy money, you'd think people would start to get a little jittery. But I think as long as the big hyperscalers are spending, neoclouds are spending, and we're going to do some free-cash-flow analysis on how this all ties back to ROI, I think that'll give us comfort.

Makes sense. Yeah.

I do just want to get out there: I heard those guys talk a week ago, 2 weeks ago, and I think they all agreed—and I totally agree—that it does feel early in terms of the cycles that are going on. I say that just because the incremental use that I get out of Claude Code and similar CLI tools is just—

Okay. Yeah.

Incredible. And I know that hasn't flowed through enterprise. People still don't even know what it is. I've said at the beginning of the year that by the end of the year, everyone will know what it is. I don't know that everyone will be using it, but I'm leaning into the thesis that by the end of the year, my family—my grandparents—will start asking me questions about Claude Code. So my point of view is pretty positive.

5. Claude Code Crosses The Adoption Hurdle

Yeah. Let's talk about Claude Code, though. From personal experience, there was this moment: December 2022, ChatGPT launches. In January, February, or March 2023, something happens, and my grandmother asks me questions about this thing, ChatGPT, right?

I'd been personally using GPT-3 in the OpenAI Playground since 2020. But when they obviously ran the SFT and had a chatbot, it was so different that it made it all the way to my grandma.

What do you think it takes to get Claude Code there? They're crossing over 25 billion in ARR, like 4× in 3 months or something crazy, right? They have Super Bowl ads. They have New York Times articles about them getting into a war with the Department of War. What more does it take to get broad industry recognition—to make Claude a household name like ChatGPT?

Yeah. Go for it.

It's a great question. I honestly don't know the answer because I've pitched it to so many of my friends now, and even some of my smartest friends still don't want to use it. There is a hurdle of understanding—basically understanding what it is and what it can do—and then installing it on your PC. For whatever reason, that hurdle is really high even for some of my very computer-savvy, computer-programming-type friends.

Well, that's—

From my point of view, it's awesome because that just gives me more time to use the tool before everyone else.

Yeah. Good alpha for you.

Right? I'm trying to sell you fire insurance to a burning house.

What did it take to get you over the hurdle, and when did you get over that hurdle to start using Claude Code for the first time?

Right. Totally applicable because I think Doug had mentioned in December, “Hey, this is amazing. You should use it.” And I was like, “Okay, I'm going to try and use it.”

I used just Claude, right? I think this is everyone's experience—or not everyone's, but I would assume a lot of people's experience. Okay, I'm using Claude to build a program, or whatever it is. Then I get back, and I'm all proud: “Hey, I built this IRR calculator.” And he just looks at me, and he's like, “That's not what I'm talking about. That's wrong.” And I was like, “What?”

So I go back in, and I figure it out. Just like that ChatGPT moment that I had in 2023, I typed in a couple of questions and literally my jaw dropped.

Just like that, but even more so this time, I was asking it to manipulate data and give me some outputs. My jaw dropped. So, coming back to your question, I think the reason I jumped over the hump is because my boss is pushing me to do it, right? I think not everyone has that situation.

Yeah. Yeah. Makes sense. Konrad, how about you?

Yeah. On that point, I think the real Claude Code moment for me—I mean, obviously, we have Doug and Dylan really encouraging us to use it—was that I realized generative AI turned from reactive to proactive. It used to really just be a chatbot, and I also had to keep questioning where the source was: Is ChatGPT hallucinating things? With Claude Code, it really brought agentic AI to life.

For me personally, my biggest use case is that I’m building my own personal RAG. I’m constantly having bots scrape certain information so that I can be as up to date as possible. And then it improves a lot of my efficiency. I don’t really draw charts anymore. I was an Excel-heavy user, but now Claude for Excel is just so good. It is so good. There are just a lot of use cases there.

For adoption to really grow from here, enterprise use cases just need to increase. I have a lot of friends who are in banks, in consulting, and in these more traditional industries, and I think the adoption rate is just not there. Obviously, there are a lot of compliance reasons, but very few firms are like SemiAnalysis, where we’re very forward-thinking and very encouraging in terms of AI tool usage. And rightfully so, because this is our bread and butter.

I think very few companies are willing to spend. For us, we have, what, 70 people today working at the firm, and we’re spending $5 million on Claude. Is it each month or each year? I forgot.

It’s a year—an annualized number.

Yeah, but we’ve had some $10,000 days at this point.

Right, right, right, right. So I’m very grateful that I’m able to be part of this curve.

Yeah. And by the way, we have—

So it’s really a good time.

We have 79 people in the general channel. I just checked.

Absolutely crazy.

Almost at 80, whenever the next person starts, which is soon.

Yeah, I was employee 42.

Yeah, yeah, yeah. Now you’re in the 50th percentile almost, man.

Yeah.

Of tenure. Nigel, how about you? Are you using Claude Code?

Yeah. Yeah. I just want to follow up on what Konrad said. Personally, I was in investment banking before this gig, and I still keep in touch with those guys, and they are not using AI at all. It’s kind of crazy to think about, but to me, there’s also an upside to it, which is that we’re probably still in the early innings, and there’s still a long way to go for adoption.

What stocks are these guys buying that are not AI-related? They’re buying the Anthropic basket without using any AI to do that analysis.

Exactly. Yeah, that’s really one way to think about it. Claude for Excel has been incredible. This last week, I was shipping out a note and really using Claude for Excel, and more than once I had this thought in my mind: If someone tells me that this is AGI, I don’t think it would be a statement I would strongly push back against because it’s gotten so good.

To think that it’s still running on current-generation hardware, before the Blackwell-vintage and Rubin-vintage models come in, is frankly crazy to think about.

“Vintage,” man. That’s such a good word for these models trained on this BF16 Hopper model. I’m going to put these weights aside from this 2025 vintage and let them age for a little bit. See what Llama 3 tokens taste like in a few years’ time, after they’ve had some time to age. I love that.

Okay, guys, I think this has been awesome. Great to get to know Core Research in a little bit more detail. Any final closing comments on Claude Code, the memory cycle, or CapEx? Stunned silence. Awesome, man.

Diamond hands. Diamond hands.

Okay, last one.

I think we’re going to have a lot of volatility with the macro stuff and with the war. On our side, we really need to figure out how all this ties in together with supply and demand, the sustainability of investing, and all this high CapEx. For our institutional subscribers, we will have something out soon. For our public audiences, we will have some form of synthesis coming out as well.

Subscribe for more. All right, guys.

Subscribe for more.

Pleasure having you on today. Look forward to the next one and, yeah, good job.