Rating AI supply chain bottlenecks: Interview with @bubbleboi | Edited Transcript

A professionally copyedited transcript of Chris Barber's interview with @bubbleboi on AI supply chain bottlenecks.

This is a professionally copyedited transcript of Chris Barber’s interview with @bubbleboi about AI supply chain bottlenecks: HBM and DRAM, co-packaged optics, advanced packaging, substrates, PCBs, power delivery, OSAT, alternative accelerators, and the company-level exposure behind each constraint.

Chapter Timestamps

00:00 HBM and DRAM: why the bottleneck starts with wafer allocation

07:24 Silicon photonics, CPO, and the parts of optics that may still stay scarce

11:35 Advanced packaging: why TSMC is tight and Intel may have room

13:49 ABF substrates, HBF, and where substrate monopoly windows can appear

17:02 Memory accelerators, TurboQuant, and ways around bandwidth constraints

22:35 PCBs, Unitika, and why low-glamour materials can still become chokepoints

27:27 Power delivery and the shift from 48V toward higher-voltage data centers

31:32 Assembly, test, fiber coupling, and the practical bottlenecks after design

36:23 HBM versus DRAM versus NAND over the next few years

42:03 Where AI hardware founders are likely to come from

44:10 Alternative accelerators: Etched, Taalas, MatX, and specialization

48:43 CPO skepticism, hyperclouds, HBF timelines, and CXL

51:48 Voltage, cooling, and the next wall in power density

56:30 Rapid-fire company ratings: Intel, Nvidia, TSMC, Alphabet

1:05:25 ASML, Hynix, Lumentum, Wolfspeed, and what changes the ratings

1:13:25 Building conviction in Intel and separating thesis from fandom

1:19:37 How the Intel pitch lands with funds and why distribution matters

1:22:39 X accounts, analysts, conferences, and why the supply chain matters

Made with: The Transcript Desk Chrome Extension: https://thetranscriptdesk.two-lambda-ai.com/

Full video: Rating AI supply chain bottlenecks: Interview with @bubbleboi:

The useful frame is not simply whether “AI infrastructure is scarce.” The conversation breaks the stack into specific physical and industrial constraints: memory wafers, packaging capacity, optical components, substrates, voltage conversion, cooling, assembly/test, and the companies positioned around each one.

Transcript

00:00-00:05

Chris Barber: HBM?

00:03-00:05

@bubbleboi: Medium to low, not tight.

00:08-00:12

Chris Barber: And your view, what’s the delta between your view and others who view it as tight?

00:14-00:17

@bubbleboi: Well, I think the people who view it as tight don’t understand how it’s made. HBM is just stacked DRAM. So as long as you have extra DRAM wafers—and we obviously do, you can get a DRAM stick, it’s very expensive—you can make HBM. And that’s a big reason why people are very careful with expanding the supply, because 60%, I think, of DRAM wafers right now are going to AI, which is HBM, more or less, right? That’s what’s actually causing this DRAM pricing, like commodity DRAM, to spike so much, because you’re stealing wafers from one share of the market to give it to the other. And that’s why HBM prices are probably middle to low.

01:00-01:10

@bubbleboi: I mean, also, there’s a couple vendors. Samsung and SK are obviously the highest quality. Micron makes some HBM, but people will source from whoever they can get. You need it if you’re doing AI, right? So that’s why I think HBM has less pricing power than DRAM.

01:10-01:29

@bubbleboi: If you look at Samsung and SK in particular, if you look at their profits, almost all of the incremental profit gains is not from HBM. I mean, there’s a little bit, but it’s like 60% of their increase in profits is from DRAM, DDR memory, right? Just basic DDR sticks. And it’s because we’re eating that supply to make HBM. So HBM is medium to low, like we can expand it still. But when you expand it, you steal from someone else, more or less. It’s a fixed pie for now. We’ll see if anyone brings on supply.

01:59-02:03

Chris Barber: So HBM medium to low, DRAM would be high then?

02:02-02:13

@bubbleboi: Oh, DRAM’s the worst. Yeah, DRAM as far as like some kind of—maybe there’s other new things—DRAM’s really bad right now. And DRAM’s probably gonna get worse. So that’s the scary aspect because of the dynamics of the fair. If you want more AI accelerators, you need more HBM. Every time you create HBM, you take away from DRAM supply. So it’s gonna create more and more of a shortage.

02:31-02:39

Chris Barber: Why are they making more incremental profit from DRAM than HBM? Are they just selling a lot more DRAM than they sell HBM?

02:39-02:45

@bubbleboi: Priced out. It’s literally that stupid. It’s literally like the price of your 16 gigabyte stick went from a hundred bucks to like a thousand dollars. And HBM prices are pretty flat or somewhat up.

02:53-03:03

Chris Barber: Why has the price—like if they both kind of utilize DRAM, why has HBM gone up less proportional to DRAM itself?

03:03-03:25

@bubbleboi: Well, think of it like this: to make HBM, you just have to take DRAM wafers, cut them up and stack them. So there’s some packaging that creates a premium for the price, but you always know you can expand. We’ve only eaten up about 60% of the DRAM wafers. We still have a little bit more we can chop down, right? So if you do get a new order for, I don’t know, Amazon or Google says, “Hey, I need some more HBM, Samsung,” well, Samsung says, “Well, I don’t need to charge you more because I can just take from this pile and just pop, pop, pop. There you go, there’s the price.” Meanwhile, with DRAM, it’s just a structural shortage. Like you come to me and you ask for it and I’m like, “Well, sorry, I can’t meet your demand.” So it’s whoever is the highest bidder. With HBM, I can meet your demand.

03:57-04:05

Chris Barber: And then does making more DRAM capacity—that takes a long time and not many players kind of thing?

04:05-04:22

@bubbleboi: No, no, it’s a cartel. I don’t—sadly, like, yes, there is some timeframe for expanding capacity. DRAM’s easier to make than HBM for obvious reasons. DRAM’s a component in HBM. But no, I think the big cartel, which is Samsung, SK, and somewhat Micron, they don’t want to. They like this. They like having high prices. They like having control of this market.

04:38-05:07

@bubbleboi: And they are scared because, again, when you make a new—like what I think a lot of people don’t understand in semiconductors is, like, you want to expand supply. Okay, what does that mean? $100 million machine, one of them, an ASML machine, right? Who knows what other—you might need like some deposition tools. You might need other tools for, like, you know, testing. So all that capex, once you do it, it sounds like it’s a one-time cost, but running those machines, those machines need to run 24/7 or else they’re not economical. So you think, okay, it’s just some upfront money. Well, what they’re thinking in their head is, I have to keep this running 24/7 to get back my money. And if I’m running this 24/7, I’m flooding the market with supply.

05:25-05:54

@bubbleboi: So they are very, very scared of adding supply to the point where I read that the big technology companies—let’s call it like Google, Amazon, Microsoft—they’ve went to SK hynix and they’ve said, “SK, we will finance the lithography machine for you. We’ll put up the money as long as you give us a contract price for the DRAM you’re going to make with it.” And SK said, “We’ll think about it,” because they want to keep the price high. They’re not dumb. They know 60% of their new profits have come from DRAM. So DRAM is like—we’re literally in this shortage.

06:08-06:24

@bubbleboi: I think it’s obvious now DRAM costs more than a CPU, almost any CPU, unless you’re going like crazy EPYC server. But even then it’s getting close because those servers, you need like 768 gigabytes of RAM to make it worth your while, right? So it’s literally crazy. It won’t last forever.

06:29-06:45

@bubbleboi: And another thing, which—I mean, we mentioned it earlier off camera—but another thing that I’m thinking is if anyone does bite the bullet and expand supply, we will start seeing signs of that towards July and the beginning of fall. My view is just supply is tight right there.

06:50-07:08

Chris Barber: Sorry for the whole rant. Multi-year period before the supply becomes not tight?

06:57-07:15

@bubbleboi: It’d be a couple of months. I think six—I think the start of next year, we’d have good signs of, is supply gonna weaken? Maybe in the summer we start getting signals where I would be scared to just be long SK and Samsung. Because again, that’s where they’re making their money. So we’ll see. These timelines take a while in lithography.

07:23-07:26

Chris Barber: Next one. Silicon photonics, optics, CPO.

07:28-07:39

@bubbleboi: Not tight anymore. Honestly, it was tight probably at the beginning of the year because there was a procurement cycle for these CPO switches, which is Nvidia Spectrum. And I think Broadcom has some Tomahawk switch that also does CPO and they were procuring a lot. I think right now you can procure parts. You can go to Coherent or Corning and get fiber.

07:55-08:04

@bubbleboi: The interesting thing about optics, which I feel like is a bit different than DRAM, is like, it sounds like it’s a commodity. And also you mentioned quite a few things there. You mentioned CPO and I think a couple other things. It sounds like it’s a commodity, but the pricing difference between like a quality part—which we can mention a little Lumentum here as an example—like Lumentum’s lasers are the most highest quality. They have the least amount of noise or narrowest wavelength. That’s ideal if you’re going to be doing this silicon photonics and co-packaged optics. Honestly, if you can’t get those kinds of lasers, you really can’t pull it off.

08:44-09:00

@bubbleboi: So in a way, for high quality lasers, supply is medium, right? But people have already kind of done procurement and we’re going to see at the beginning of the next year how it goes when they actually start deploying these switches in the data centers. And we’ll see.

09:00-09:11

@bubbleboi: My view though is silicon photonics has—not just silicon photonics, but let’s just say co-packaged optics and everything around that—it’s kind of known to be not as reliable as pluggables. There’s other benefits. Pluggables are commodities. There’s like all kinds of pluggable companies, right? And they all are pushing prices down.

09:20-09:49

@bubbleboi: But with these lasers, like Lumentum’s lasers for co-packaged optics, silicon photonics, the laser is very likely to fail when you start delivering these systems in a co-packaged optics and they need to be replaced. So in a weird way, because they make those high quality lasers and the laser is one of the first things to fail in photonics and co-packaged optics, it’s kind of like we will see if they end up being bid up because there’s not really many other competitors in the game right now.

10:00-10:09

@bubbleboi: ...with their quality, and they need some reliability. So there may be a recurring bid on some of those components.

10:09-10:41

@bubbleboi: Now, when you mention co-packaged optics, it’s harder for me to give you that view because there are a few foundries now. It’s not just TSMC with COOP—that’s their offering. There are others now, like Tower Semi, which is the big one. They have an indium phosphide node that you can use. So I think over there, the supply isn’t that tight. And I think it’s actually easier to fab some of these nodes, like indium phosphide nodes that you would need for photonics, rather than logic, right? So we’ll get a bit of competition there.

10:41-11:15

@bubbleboi: The real question will be the lasers, right? Are those going to fail? Any kind of small components. I’m not a photonics expert, but there’s ring modulators and all this weird, wacky stuff. So we’ll see—are those bid up? But the actual fabs, they can expand capacity quite a bit, and they are. So it’s a good place to be. They are ramping up. We will see if the lasers are good and the testing and co-packaged capacity is good, right? So we’ll see there.

11:15-11:24

Chris Barber: Next one, advanced packaging.

11:24-11:42

@bubbleboi: Ooh, incredibly tight on TSMC, not tight on Intel.

11:42-11:47

Chris Barber: Nice.

11:47-12:02

@bubbleboi: Yeah, I think Intel’s ramping it up, and we know there are a few customers that are definitely interested. I don’t have any insider info, right? People might think I’m so close to Intel that they’re telling me stuff. They’re not. They don’t talk to me at all.

12:02-12:14

@bubbleboi: But my guess is, you know, we’ve already heard the Apple announcement. There’s probably a Google TPU announcement coming soon. So we will see. And I know the Google folks are very, very into Intel.

12:14-12:30

@bubbleboi: And I think Intel’s packaging is—I always try to make this point to people—comparing TSMC’s CoWoS to Intel’s EMIB right now is like comparing a semi-truck to a Ferrari at this point. They’re so far ahead.

12:30-12:40

@bubbleboi: And if packaging is really the next Moore’s Law, which I think it really is, and it’s really how we’re going to be scaling compute from Rubin to so on, right? I mean, today’s the Cerebras IPO, right? So, not saying Cerebras is going to win the game, but we need more silicon area because we can’t trick the transistor. So they’re definitely right directionally.

12:40-13:02

@bubbleboi: So yeah, I think on TSMC, you’re absolutely cooked. It’s already pre-booked out. Most—60% of it’s on NVIDIA. The rest might be AMD and others. So yeah, it’s tight there. You gotta wait till probably 2028 to start getting any kind of capacity.

13:02-13:23

@bubbleboi: If you want to get some capacity on Intel, call me. I’ll make a call and I’ll get you in touch so you can get all the capacity you need.

13:23-13:28

Chris Barber: Nice. Next one, ABF substrates.

13:28-13:40

Chris Barber: ABF substrates. This is the substrates for packaging or something else?

13:40-13:44

@bubbleboi: It’s like bottom of the chip, I believe. ABF. I couldn’t tell you exactly.

13:44-13:47

Chris Barber: Cool.

13:47-13:50

@bubbleboi: I feel like more commoditized at this point. Next one.

13:50-13:54

Chris Barber: Now this is obviously a future thing, but High Bandwidth Flash.

13:54-14:06

@bubbleboi: Well, it’s probably the worst because we don’t have it yet. I think when it comes to market, SanDisk and SK will probably have a one or two year monopoly. And it’s a very interesting thing to see who comes in.

14:06-14:39

@bubbleboi: My mental model is that High Bandwidth Flash is a lot easier to make than High Bandwidth Memory, and you don’t need the same amount of quality. So there probably will be a few people who try to come in if it’s successful. I know Kioxia, which is in, I believe, Taiwan—they’re looking at it and they have an idea, but I don’t think it’s quite the High Bandwidth Flash that SanDisk and SK are working on. And I’m very excited to see what comes of it.

14:39-14:50

Chris Barber: And other people’s view is that there’s reliability concerns with HBF, but your view is there’ll be hardware and other solutions for that over time, and it just is not going to be as big a concern as people feel, basically.

14:50-15:01

@bubbleboi: I think the concern that people have is write reliability specifically. And if you start thinking of this as like an HBM on the GPU accelerator, that is a fair critique. You’re packaged there, but eventually your memory fails, right? Because you’ve written too much.

15:01-15:19

@bubbleboi: Now, if you think inference is more important, that doesn’t really matter as much, right? Write reliability—we’ll just load all your weights there or whatever we need and we’ll just read it, right? Also, I think there’s this misconception that it needs to be packaged next to the compute dies. I mean, at the end of the day, even though that’s ideal and will get you the best bandwidth, that’s probably not what’s coming immediately.

15:19-15:46

@bubbleboi: Okay, wait. And here’s like the big leak maybe, right? Google—I mean, everyone knows—Google dropped TurboQuant, right? And everyone’s been asking me, why did Google drop TurboQuant? Well, it’s because DRAM prices are insane and Flash is 55 times cheaper than DRAM right now, despite the Flash price increasing. So Flash could probably price higher and higher and still offset DRAM.

15:46-16:19

@bubbleboi: So Google—it’s leaked that for the next TPU, they’re not just making a TPU, they’re making a memory accelerator. Now, we don’t know who SanDisk’s customers are—they won’t tell you, right? But it’s definitely the big hyperscalers. I don’t think they would have embarked on this HBF journey and this investment if they didn’t have a customer in mind. I think it’s interesting that Google is so against these, because they’re going to get hit the hardest on these DRAM prices, that they’re willing to make a memory accelerator. I think that memory accelerator will probably be a high bandwidth Flash or an array of Flash memory packaged very neatly and offloading a lot of AI to it.

16:19-17:01

Chris Barber: What do you mean, memory accelerator?

17:01-17:30

@bubbleboi: Oh, yeah. So it’s kind of a weird concept. I didn’t hear about a memory accelerator until this year, too. What it is—like, we have AI accelerators, but AI needs a lot of memory. And also, the way you interact with memory in AI is a little different than a generic computer running an OS, right? There are different tiers: there’s your queries, your weights, files, context. So what Google is probably going to do is really offload a lot of this memory that would have gone to DDR to Flash, and do something very smart with it to get a similar amount of speedup as they would have gotten on DDR or faster, right?

17:30-18:03

@bubbleboi: Do you know, Google was the first to really fight near these long context models. I think Gemini still has the longest context—2 million, right? You can’t fit 2 million tokens per user on these GPUs. I mean, you’d need a huge number of GPUs and aggregate across. If you think about it, if Flash is 55 times cheaper than DRAM, how much cheaper is it than HBM? It’s probably like a thousand times or two thousand times, right? Okay, it’s cheaper.

18:03-18:34

@bubbleboi: Now, if you can get the bandwidth up to HBM2 levels or equivalent with some very smart tricks—I think the people at Google who are on that team are very smart, maybe even smarter than me sometimes, right? So I think they definitely know what they’re doing, and they’re not into this DDR cartel. They don’t want to negotiate with these—for lack of a better term—these Korean terrorists, right, who are making this DRAM. So I think you will be shocked when it comes out what Google makes. I have no insider info, but I can read the tea leaves and tell you with good confidence what they’re probably going to do.

18:34-19:10

Chris Barber: Thick. And I know obviously users want infinite context windows.

19:10-19:23

@bubbleboi: Well, users probably do want infinite context windows. I mean, I don’t want to go into this too much because I might be doing my own thing on the side, but yeah, if you really look at it, right, why do we need 288 gigabytes of HBM? That sounds kind of crazy. Like, bro, we could play Counter-Strike on 150 gigabytes of storage, and all of that is context, right? And context is exploding.

19:23-20:00

@bubbleboi: I’ve had this argument with a lot of people when they say, well, agentic workloads, they don’t really care for context because you can just make the tool call. But I’m like, well, a large code base, you’re doing a complicated model in Excel or something—probably a compute task.

20:00-20:32

@bubbleboi: It’s like, yeah, you can make a call to a tool, but that’s going to add latency. The context is going to be faster. And also, context right now—this is what’s led to the growth of DDR and Flash—because it’s growing so much. We need to serve so many users at long context, and they all want to do these coding and energetic workloads that are spilling to DDR. It’s the only economical way. And then they’re spilling into Flash. And now we have VLLM, right, to even make this manageable. So I feel like what people aren’t understanding is that we’re getting a third tier of memory.

20:32-20:57

@bubbleboi: We had—like when I was in school, they taught me L1 cache, L2 cache, L3, DDR, disk. Okay, that makes sense on one CPU or one laptop, but now we’re at the multi-rack scale. And now it’s going to become a new tier of memory, and maybe even a new tier of networking to communicate that memory across all these racks. I mean, you have thousands of racks with GPUs, and they all look like one computer to someone, right? That’s kind of crazy. That’s a real testament to the engineering being done here, right?

20:57-21:16

Chris Barber: Yep, sick. Next one, broad category—there’s multiple in there—so maybe pick the one that is high, if any: wafers.

21:16-21:54

@bubbleboi: Hmm, I don’t think we have much of a shortage of wafers. I think wafers are interesting—reticles, which make the patterns of the chips. I think we don’t even have a shortage there, but what we have a shortage of is quality, right? When people go to these higher-end, like high-NA EUV nodes, the reticles need to be completely redesigned. So it sounds like this basic thing. I think the wafers are basic, right? We have our 12-inch wafers, and we’ve been using them for years. It’s kind of served us well. No one really wants to redesign it right now, but the reticles—every new generation of litho is like a completely new reticle design. And getting those reticles right is huge.

21:54-22:31

@bubbleboi: Well, I just know this from my own experience—taping out a chip is like the biggest cost. When you tape out a chip, the biggest cost is the reticle, is getting that reticle made. It costs millions before you get it perfect, and then, voila, you start doing your tape out. Sorry for the rant.

22:31-22:35

Chris Barber: No, this is good. The next one: PCBs.

22:35-23:13

@bubbleboi: Ooh, interesting, interesting one. So again, the PCB that you would use for your Arduino? No, we got that down. But what we’re seeing is, because packaging—like advanced packaging—is tight, and also they can’t scale it past... Well, right now, I mean, Vera Rubin Ultra was supposed to be four reticle dies pitched together, and now it’s two reticle dies and another two reticle dies communicating over a PCB. Those PCBs, which need to be multi-layer and have very good signal integrity, those are definitely in shortage because it’s a quality problem.

23:13-24:11

@bubbleboi: My guess is, if NVIDIA is still stuck on TSMC co-ops and they can’t figure out how to do advanced packaging better on there, it’s going to lead to even worse scenarios on PCBs, where NVIDIA might need all those high-quality PCBs. And it’s not just the PCBs. It’s literally like—I could find some of the companies, if you give me a sec—but even the materials in the PCB, like the glass fibers that they use for shielding different layers, those have gone parabolic. There’s like one or two companies in Japan who can do it, and there’s no one else. That is an actual structural shortage, and everyone’s trying to find someone else that they can get these data center, or these PCB high-quality glass fibers for PCBs, right? And there’s like three or four really good PCB companies in Taiwan and Japan, and that’s really it, who can do this mass manufacturing of hundreds of layers of PCBs. So I would put PCBs as pretty tight, and the materials for them, even worse. And if we don’t solve these packaging issues, it’s going to go from bad to good. So if you want to make money, make a PCB company right now, maybe.

24:11-24:43

Chris Barber: Do you want to—yeah. I’m curious which those companies are.

24:43-25:19

@bubbleboi: So I can give you a few right now. Unimicron is the biggest one in Taiwan. There’s Gold Circuit Electronics in Taiwan, who makes all these high-speed PCBs. There’s a couple—Nanya PCBs in Taiwan is very good as well. In Japan, there’s a couple of these other ones. I think Unimicron—not Unimicron, but Unitika. Unitika is like the funniest story of all time. It’s literally like a loom—I don’t even know what a loom is, because it’s like back from the 1700s.

25:19-26:34

Chris Barber: Yeah, yeah, yeah.

@bubbleboi: They make fibers, right?

Chris Barber: Weavers.

@bubbleboi: Yeah, they’re weavers. They’re weavers, right? And it’s basically been a zombie—it’s a zombie company. It’s this old 300-year-old Japanese company. And they make fibers for textiles, plastics, and now they’re weaving glass. Well, they already did do it, right? They already weaved glass, but it wasn’t a big part of their business, as you can imagine, right? Now, all the hyperscalers and NVIDIA are going to them and they’re like, “Yo, you got to expand this and get qualified so we can use you for these PCBs, so we can source these glass fibers from you.” So it’s like the funniest story of all time. You could write a movie about it. It’s literally like a 300-year-old dying Japanese weaving company drowning in debt, and if they get qualified to make these glass fibers for PCBs, they will save themselves. They will become an AI trade company, and they’ll probably be good enough that they can exit most of these other businesses, which are very commodity businesses, right?

26:34-26:50

Chris Barber: Makes me think of Tatar and these other companies where they have some skill that they’ve developed for a different kind of product, and now all of a sudden they get pulled into the AI space because they have what AI compute needs.

26:50-27:21

@bubbleboi: Oh yeah. So Unitika—looked at their stock movement over the past year, it’s insanity. We’re talking like 100% moves in days, up and down, because no one knows if they’re going to get qualified. And if they do, you really have no idea what that’s going to do to their revenue and the impact. So it’s a very exciting time to be a Japanese investor, I think.

27:21-27:27

Chris Barber: Yeah. Next one would be just generally power delivery.

27:27-29:15

@bubbleboi: Power delivery is tough. Honestly, I’ve already seen Elon was willing to go to Louisiana to tap into the biggest gas pipeline in America to use it, right? I mean, that’s a smart move, but I feel like not everyone can do that. Power is tight. Doing deals with these utility companies is going to become way more common. Now, as we try—it’s kind of been telegraphed—and the big trade now is we’re transitioning from 48 volt to 800 volt DC architectures. And that whole supply chain is also—I mean, we’re not there yet, but people are now rushing to figure out how to solve that. You need all kinds of transformers, inverters to deliver 800 volt DC between racks. So I think there it’s pretty tight on the real power side—just raw power, it’s tight. People are thankfully working hard to figure out 800 volt before it comes online, but it’s just going to get worse. So power is going bad probably in the future. And the power that the GPUs are using is not increasing by 10, 20% a year—it’s doubling. So, yeah, I think my biggest bets right now are on power, but not directly on small modular reactors and stuff like that. Even though I think this stuff is very cool, it’s more like, how do you deal with this power? I don’t think anyone knows. This is more of the, how do you move 800 volt power between racks? How do you move it more? Also, once it’s there, how do you not burn the data center down? How do you keep things cool and calm? I’m not going to spell it out, but I think one of my—we did talk about this already. So one of the companies that I was very bullish on, other than Intel, but I didn’t want to give it too much credit because it was a small company—I didn’t want to manipulate the price—but it’s this company called Flex. And Flex makes data center equipment, cooling equipment for data centers.

30:00-30:11

@bubbleboi: But they do it for everything, right? From hospitals to anybody who needs huge amounts of power. But what Flex has within that company is these very, very smart cooling solutions for chips. It’s not just normal water cooling blocks or cooling; even the transformers that you’re going to be using need to be cooled. Even the power rails need to be cooled at 800 volts, right?

30:12-30:37

@bubbleboi: So I think they’re doing a really good job of being that future. But I think they only have a roadmap up to like Feynman. After that, we actually have no idea how we’re going to cool things. No one has an answer. There are some views, but again, this is a two-year-away problem. And hopefully by then, you know, me and you are both just chilling on the beach and we don’t have to worry about it too much. But I feel like no one really has an idea of how to cool these chips. Even at Feynman, it’s going to be tough to cool them.

31:02-31:04

Chris Barber: Next one, optical fiber.

31:05-31:07

@bubbleboi: No, it’s not in a shortage.

31:07-31:09

Chris Barber: Cool. Next one, liquid cooling.

31:11-31:18

@bubbleboi: Quality liquid cooling—no, it’s not in a shortage either right now, but it kind of goes back to my prior point, which is we have no idea how we’re going to cool things in the future. So yeah, it’s going to be a problem in a year or two, or whenever Feynman comes online.

31:30-31:32

Chris Barber: Yep. Outsourced assembly and test.

31:32-31:54

@bubbleboi: Pretty big bottleneck, I think. I think it can definitely be expanded. There are a few companies, but I think it probably needs to be quadrupled if you think advanced packaging and photonics is coming online. All those need insane testing, and not just testing at the wafer, but in the package, right? So once it’s packaged, that’s real testing.

32:03-32:09

@bubbleboi: Forgot the name. This is like Rigaku. This is like what an insane, irrational analysis. Yeah, yeah. Aehr Test is the other one that a lot of people point to. ACMR does testing, but they have a big China exposure.

32:19-32:24

@bubbleboi: There are a couple of companies, although like—

32:24-32:28

Chris Barber: Aehr Test, how do you spell that one?

32:28-32:29

@bubbleboi: A-E-H-R is the ticker.

32:29-32:30

Chris Barber: Okay, cool, yep.

32:30-32:34

@bubbleboi: Aehr Test. Yeah, so burn-in testing, which is like when you have all the packages up and you’re pushing these crazy voltages, you want to make sure nothing blows or explodes.

32:41-32:46

@bubbleboi: So you’re going to do testing in the package. I think one thing I haven’t heard about—a good reason why I haven’t—is because in photonics, fiber coupling, which is like connecting the fiber to the die, is a big problem because you get coupling loss.

32:58-33:14

@bubbleboi: Now, all right, so free to your viewers, we could ask Claude later, but what are the companies that are good at fiber coupling? I don’t think there are any, because it doesn’t really exist right now. It’s like being hand-done by some experts in a university lab, or the one or two photonics-only fabs, and those are really just university labs at the end of the day, right? So fiber coupling is going to be a big problem if you think photonics is going to scale. I don’t think there’s an automated way to do that.

33:29-33:31

@bubbleboi: Next one would be CPUs.

33:35-33:39

@bubbleboi: CPUs are probably medium.

33:39-33:53

@bubbleboi: I think we will see if advanced logic gets expanded. I think CPUs—it’s already been telegraphed that AMD might have just been like, “F3 nanometer, we’re just going to move some stuff to five nanometer right now just to meet this demand.” And I think people will eat it. They don’t really care. They just need more CPUs.

34:02-34:06

Chris Barber: By advanced logic, do you mean lower nanometer scale?

34:05-34:06

@bubbleboi: Yes, yes.

34:06-34:13

@bubbleboi: So like anything below, get some C5 nanometer, sort of advanced, right?

34:11-34:15

Chris Barber: Anything below N5, is it?

34:14-34:15

@bubbleboi: Yeah, correct.

34:15-34:21

@bubbleboi: Whereas they’re just going to push it up to the higher nanometers and it’s good enough, kind of thing. So most of these servers that we’re using here, right? But they don’t use it for everything. Like, you don’t need it for everything. For compute, you need it, but for IO and other things, you can get away with even six nanometer if you wanted.

34:38-34:39

@bubbleboi: Next on hard drives.

34:40-34:47

@bubbleboi: Ooh, pretty bad, I think. If you’re getting spinning disk hard drives, I don’t think there’s any way they’re really expanding capacity on that. The only thing that they can probably do is add more bits to the hard drive, which I think there’s a way you could do that when you’re manufacturing. It’s almost like more voltage levels in flash. Although I haven’t done much research of exactly how they do it. But I think a spinning disk is completely in shortage.

35:08-35:17

@bubbleboi: The question is though, spinning disk kind of competes with flash, right? So we will see what people decide to use.

35:19-35:24

@bubbleboi: Data center land and shelves.

35:21-35:27

@bubbleboi: Thankfully in America, we have a lot of land. And so I think power and water are the constraints for the land. We have a lot of land. I think we can just fill up Texas with every data center imaginable, you know? Like make barges in the water of data centers if people have a problem.

35:45-35:46

Chris Barber: Yep. Power semis.

35:49-36:18

@bubbleboi: They’re not in a shortage right now. It’s like the most hyped trade that everyone’s doing. There’s not a shortage and they’re very mature nodes that anyone could make. And my guess is, if it does become profitable, China’s just going to flood the market with tons of power. I mean, they already do with knockoffs, but yeah, I’m not in the power semi trade right now. Actually, technically was, but it’s not really my focus.

36:19-36:24

Chris Barber: Yep. How do you—let’s say two or three years from now, what do you think will be most—well, how would you rank the three of HBM, DRAM, NAND in terms of pricing power?

36:33-37:17

@bubbleboi: Pricing power is an interesting question. I think pricing power kind of follows usefulness. The list is ranked by usefulness. I think NAND is number one. I think there’s not much more supply. There’s only a handful of fabs. Toshiba is a big one, Kioxia is the other. And they’re not really into expanding more right now. There’s not really a need to because people are willing to pay another 100%, 200% because it’s still cheaper than DRAM. If you also think NAND is going to become way higher bandwidth and way more useful—voila, its pricing power is amazing. If it’s just as fast as DRAM and cheaper, it’s a great thing to have.

37:19-37:50

@bubbleboi: Two would probably be HBM. And the reason I say this, even though I said before HBM prices are kind of stabilized or somewhat down, it’s because they keep wanting to push more and more stacks of HBM. So I think it’s like 16 or 18 high now and they want to go higher. And now they’re trying to do co-designed HBM with NVIDIA. So they’re going even off of the standard for HBM, which is set by JEDEC. So I think whoever cracks that nut is going to make a lot of money and people will pay a premium if they crack it, right? I don’t know. I think from our whole convo from now, you can kind of understand that we’re hitting thermal limits of physics right now, right? And it’s probably like, Feynman’s like, where do we go next, right?

38:15-38:23

@bubbleboi: And third, yeah, DRAM. I mean, I just think NAND is going to flood the market. CXMT is going to flood the market with DRAM. I think in July, TSMC might expand logic nodes, which is going to lead to even more HBM needed, and it’s going to just make the price of DRAM go from bad to worse and people are not going to take it anymore. I think we could literally see the big cloud companies team together and make their own DRAM.

38:47-39:28

@bubbleboi: I think it’ll get that bad. If they don’t expand capacity, they’re going to lose.

38:53-39:00

Chris Barber: If they don’t expand capacity, they’ll make it, you said?

38:57-39:22

@bubbleboi: Oh yeah, like if Samsung and SK don’t expand capacity, the customers are just going to be so pissed off, they’ll get it from CXMT. And if China tries to block them—I mean, if the US tried to block them from buying from China—they’ll just say, “F it, we’ll make our own, we’ll do a joint venture.” And there is precedent for this, believe it or not. So in Japan, I think this happened once. So let’s see how it goes.

39:31-39:32

Chris Barber: Interesting. Next one would be—

39:35-39:49

Chris Barber: Technologies that have become much more important over the past 12 months that you expect to continue increasing in importance.

39:49-39:58

@bubbleboi: Absolutely, packaging is number one. Packaging is the new Moore’s law. It’s our only way to scale compute today and there’s no other option. So packaging is number one. Two—

40:00-40:21

@bubbleboi: It hasn’t really gotten important yet, but new cooling techniques—whether it’s direct die cooling, like microfluidic-style cooling, or maybe two-phase cooling—could be used. This is going to become extremely important.

40:21-40:39

@bubbleboi: I always say this to people as a joke, but I really mean it: the way you’re going to change the world is by making a new kind of memory or curing cancer. Memory—and specifically, I think flash—I’m so bullish on flash. I think it’s a memory technology that scales and is relatively cheap. I think that’s going to become important. We’re going to see all kinds of advancements in memory from here. I think it’s going to look like a different world.

40:39-41:12

Chris Barber: Sick. We are probably—like, there’s the big John Hennessy computer architecture book.

@bubbleboi: I have it right here on the side of me.

Chris Barber: Throw it out, because new things are coming. None of that stuff is going to matter anymore. The line between compute and memory is going to get blurred.

@bubbleboi: Yep, and that’s partly because if we want faster tokens and longer context, it’s just a memory bottleneck.

Chris Barber: Yeah, and all—

@bubbleboi: —is ruling memory, brewing also—oh, your mic is going out again.

Chris Barber: Yeah, apologies.

@bubbleboi: So it is a memory bottleneck, also.

41:12-41:47

@bubbleboi: We haven’t had innovation in the space. HBM and stacking is one way, but we’re going to see crazy stuff in memory. I’m very excited to see what people do. If I was a VC, I would just go to anyone who has a good memory idea and try to give them money and just hope that they pull it off, because they will become—you’ll go down as a legend. If you’re that guy, you’re winning Nobel Prizes for this stuff.

41:47-42:03

Chris Barber: So, yeah, where do you think hardware founders will come from? What is the—like, you know what San Francisco is to AI, enterprise startups, or consumer startups—what is the equivalent in the U.S., especially for hardware companies like that?

42:03-42:38

@bubbleboi: It’s such a good question, man. I’m in Palo Alto right now, right? So I’m not in SF, and I probably spend most of my time in Palo Alto, Sunnyvale, Cupertino, San Jose, and I meet all these engineers. So that is like South Bay and San Jose—I mean, this is Silicon Valley. And thank God, I feel like a lot of people think America’s lost the chip race because Taiwan and these people can manufacture chips cheaper. What people don’t get is we have some of the best chip designers—not some of—the best chip designers in the world. No one compares to us. The only people who come second is the UK.

42:38-43:12

@bubbleboi: But even then, they don’t have the industry we have. They don’t have the amount of startups, the amount of big companies. They have Arm and maybe a couple others, but that’s really it, and that’s the only reason why the UK kind of has this chip legacy. It’s going to come from these engineers who are working at these big behemoth organizations. Counterintuitively, I don’t think they’re going to be Stanford undergrad dropouts. They’ll be like PhDs who got kicked out of ID. Similar to the Groq kind of founding team—that kind of stuff. They’re going to be guys who are just like, “Bro, I built my career for so many years. That’s it, man, I’m going to do my own thing.”

43:12-44:04

@bubbleboi: It’s going to be archetypical. I think they’re going to be these guys where, if you talk to them, you’re like, “Yo, what’s wrong with this person?” It’s going to be like a Steve Jobs kind of character, because I think making hardware is hard, and the only people who are still doing it are doing it because they love it. Hardware engineers are like basketball players—they’re like, “I’m the best, man. You’re not on my level.” I think it’ll be a very fun time to back hardware.

44:04-44:18

Chris Barber: Any chip companies you like? Like, there was Fractal—I think it was a UK one—and there’s Etched, MatX, Taalas. Any you think could be doing work at scale in five years?

44:18-44:53

@bubbleboi: So, I can’t say too much because I know all of them and I’ve talked to all of them. I will say, two years ago I was definitely of the opinion that we need a general-purpose AI accelerator, like a GPU or something even more general than that. I’ve completely changed my tune on that, and I think specialized, specialized, specialized is the way to go. The market’s big enough that if you can specialize even in one niche of AI compute, you’re going to win so much. So I’m very proud of my friends at Etched. I think a lot of people have maligned them online, but I’ve looked at what they’re doing very closely, and from what I’ve gathered—and I don’t know if they’re going to like this when I say this out when this comes out—but I feel like a lot of Nvidia’s innovations are just Etched ripoffs at this point.

44:53-45:40

@bubbleboi: Like, even Nvidia’s tensor cores—they’re just taking stuff that Etched was doing maybe a year or two ago and just being like, “Yeah, maybe a specialized attention kernel makes sense,” right? So I’m definitely of the feeling that the more specialized you go for your niche, the better. So I’m very proud to see what they’re going to do. I’m actually meeting them later today, so I want to see how things go there. As far as the others, I mean, there’s not much info out. I think Taalas, from a scientific standpoint, is very interesting.

45:40-46:19

@bubbleboi: I think Taalas maybe will need to partner with someone, though. I don’t think they should stand up their own compute. They should kind of license that IP, because it is very interesting and I think it has uses way far beyond AI. So Taalas, for maybe the listeners who may not know, is like—they basically found a way where, when you have a reticle with your chip pattern, they can burn in the weights. But when you want to make a new chip or change the model, you don’t need to change the whole reticle. You can just change a small amount. And remember, I said reticles are the most expensive part, and you—voila—new model comes out on a chip.

46:19-46:56

@bubbleboi: Now, do I think burning the weights into a chip, even if you want to think inference is the future, is the way to go? Maybe now, but burning some stuff is probably useful. I think having that flexibility is very important in the future.

46:57-47:12

Chris Barber: Yeah, I know. So you mean they should partner with someone as like an inference cloud, or you mean they should partner with someone on the creation of the chips themselves?

@bubbleboi: The creation of the chips. I think that IP is interesting and that know-how is interesting, and they should be licensing that to every hyperscaler—Broadcom, Marvell. Those guys can definitely find where to use it better than they can.

47:12-47:59

@bubbleboi: I’m also of the opinion that selling chips may not always be the best idea for these hardware models. If you’re trying to sell into a data center, there’s two ways you can do it: you could be niche, or something that Nvidia hasn’t fully captured. If you’re going after compute, it’s hard. But if you’re going after networking or memory offload, people are willing to work with you, but you have to integrate into their ecosystem. I think Google’s done that—I mean, every major cloud has done this, but Google’s probably done it the best, where they want to own that stack. Maybe the thing with these new hardware startups is partnering with someone who doesn’t have that stack, right? And those are the big AI labs—that’s Anthropic and OpenAI. That’s smarter, because they don’t have—I mean, I’ve heard that OpenAI has a hardware team, but I haven’t seen much come out yet. But they are more willing to work with you if you design for that. That might be a smarter use case.

47:59-48:38

@bubbleboi: Or the other use case, which I think helped Groq a lot, was stand up an inference API. If your hardware is definitely better on TCL, start selling tokens, right?

48:39-48:53

Chris Barber: Yep, yep. Next one would be emerging or, you know, technologies that are going to be very important in the coming years. So, haven’t seen it as much yet—so maybe HBF is one here. What else?

48:53-49:19

@bubbleboi: Even though I’m pretty bearish on co-packaged optics or photonics, I think if they crack it, it would be great. I really expect problems with the rollout.

Chris Barber: Oh, yeah.

@bubbleboi: Oh, yeah. I expect next year, most of the rollout—most of these things will probably fail, because of the economics of having to buy these expensive-ass Lumentum lasers all the time.

49:20-49:48

@bubbleboi: And any other thing in there—the hypercloud or the large cloud companies are just going to say, “Nah, we’re good. We’re just going to go back to cheaper switches.” Right now, hypercloud—that would actually be a good name for the hypercloud. Hypercloud is probably the better category. If people say that, I personally just made it up right now. Yes, that would be better. I’m on like four hours of sleep, so it should be “your clouds” and “hyperclouds.” Obviously, I like “hypercloud.” Maybe we’ll make it stick. I think that could be good.

49:49-49:59

Chris Barber: When’s HBF—like, when do you think HBF kind of comes online? We’re getting samples of the—

50:00-50:12

Chris Barber: We’ll probably figure out in Q2 of next year when the production starts. I’m very excited to see what comes out. Again, as an EE myself, I just want to see this field pushed forward.

50:13-50:26

Chris Barber: So anyone who’s really investing and taking a chance, taking a risk to push computing forward, I love it. I’m not a big bottleneck investor. Like I told you, the PCB and these fibers—I’m not investing in that. I want to invest in innovation. It’s a very VC mindset applied to public markets.

50:32-50:38

Chris Barber: You’re excited about Compute Express Link?

50:39-50:45

@bubbleboi: CXL? Oh, yeah. I think NVIDIA is not excited about CXL, but I know Google is, and everyone else is.

50:47-50:59

@bubbleboi: I think CXL and HBF go hand in hand. Because, again, if you believe there’s going to be a memory accelerator, it’s going to use CXL. To anyone who doesn’t know, CXL just makes all of your peripherals that would have been on PCI just look like memory, just look like RAM.

51:05-51:08

@bubbleboi: You don’t have to worry too much about it. The programming model is great.

51:10-51:15

@bubbleboi: Way less latency, and it makes all these things communicate a lot easier versus PCI.

51:16-51:19

Chris Barber: Would that be like Astera? Is that kind of the—

51:19-51:34

@bubbleboi: Oh, yeah. I definitely was very long Astera, yeah. But it could not be Astera. It could be Broadcom. I mean, we will see. CXL is an open standard, right? So we will see. But I think people who are building on CXL—people making a CXL switch, people making CXL cards—they’re going to be the winners. CXL is a standard.

51:44-51:49

Chris Barber: Can you say more about the voltage of packages for new chips?

51:50-52:05

@bubbleboi: OK, let me find my—oh, because I want to get this right for you. So, typically in data centers, we were distributing 48 volt DC, and now we’re going to 800 volt. Yeah, I have the numbers of the estimated power.

52:12-52:19

@bubbleboi: Power, right? So for Rubin, the power draw is 2,200 watts, which is 2.2 kilowatts. It’s estimated for Feynman, 4.4 kilowatts, right? So it’s basically double.

52:20-52:34

@bubbleboi: I know right now that in the cooling market, I don’t think anyone really knows how to go past like four and a half for cooling. For post-Feynman, we’re at like 5.9 kilowatts. And for whatever comes after—15, we’re fucked.

52:43-52:58

@bubbleboi: And these numbers come from my friends at the Korean Institute of Advanced Technology—KAIST. Best school in Korea. Let’s go KAIST. But yeah, I mean, I look at this and I just think, is someone telling me a joke? Is someone trying to pull the wool over my eyes?

53:02-53:10

Chris Barber: You feel like the people that are increasing the voltage are not talking with the people that are having to cool the chips, and that’s why you kind of care about both. You care about cooling.

53:10-53:21

@bubbleboi: Basically, it’s like everything that can get more density on a chip, which increases volt, increases power, and then everything that can cool that power.

53:21-53:33

@bubbleboi: So, in my eyes, it’s just like, yo, we’re hitting the limits of physics. Jensen’s telling us we’re going to double, double compute. And he’s done it so far, right? 800 to 2,200—roughly a double. Rubin should roughly be a double. Feynman should be more than a double. But then after that, I’m just like, wait, what’s going on?

53:43-53:48

Chris Barber: What’s the most important ratio, by the way, for cooling? Is it amount of power per square foot, or is it just amount of power? Is it amount of power per something?

53:48-54:21

@bubbleboi: It’s flux. So it’s like heat—it’s flux. Flux. Flux. Yeah. Flux is like heat traveling through an area. It’s been a while since I’ve been out of physics, but flux is basically the rate of change through an area—the surface area, more or less. So heat flux, which is basically how much heat is moving within a unit of area. That’s the thing that just makes— and that’s just derivative of how many watts you’re pushing into it, too. The more watts you’re pushing in, the worse the flux gets. So, yeah, I think this is a big unknown.

54:42-54:53

@bubbleboi: Honestly, if any VCs listen to this, look into some cooling companies. Look into anyone who can solve this problem, because this is a real problem.

54:54-55:05

Chris Barber: You want to see new memory companies? You want to see new cooling companies? What other kinds of things should engineers at all these South Bay chip companies be starting?

55:06-55:21

@bubbleboi: They should work on getting the reliability of the mix better.

Chris Barber: Reliability of what?

@bubbleboi: Photonics and optics. Yep. And I think also just working—maybe this isn’t an engineering thing, but maybe the math and AI geeks can figure out some unique ways so we don’t need to use as much compute. Like sparsity. TurboQuant is an amazing innovation. A lot of people say, hey, it’s nothing special, we’ve been looking at ways to compress KV, but I think that’s kind of more of a leap than the other techniques we were doing in the past.

55:54-56:06

@bubbleboi: So I think we need innovation on the algorithmic side somewhat, which should lead to the right hardware innovation. I think TurboQuant opens up the way for memory accelerators and all these things.

56:06-56:20

Chris Barber: Two things we can do next and then we’ll do the other one afterwards. We could do favorite X accounts and this kind of stuff, or we could do rapid fire opinions on different companies.

@bubbleboi: Let’s do rapid fire and then we’ll do the favorite X accounts.

56:20-56:36

Chris Barber: Perfect. I’ll go through some of your ones first. Let’s see. Obviously, you’re still very bullish Intel at this point.

@bubbleboi: Of course. Yeah.

Chris Barber: NVIDIA.

56:38-56:59

@bubbleboi: Overpromise and underdeliver is the slogan for NVIDIA. I think Jensen is the best semi-strategist of all time. He knows where all the bottlenecks are. Just follow Jensen and you’ll find out. So that’s my view on NVIDIA. Also, somewhat innovative, but the innovation at this point is by having a lot of these closed standards. Anyone who does anything innovative, they won’t adopt it for their architecture anymore because they don’t own it.

57:12-57:18

Chris Barber: If you were Jensen, what area would you try and lock up next? Or has he kind of done all the important ones at this point?

57:18-57:38

@bubbleboi: I would lock up Intel’s EMIB capacity. Yep. That was Jensen.

Chris Barber: TSMC.

57:38-58:04

@bubbleboi: Too conservative, will lose market share. Too conservative on investing in improved advanced packaging mainly. Very much so. A little set in their ways there. I think also too conservative with not expanding capacity fast enough for advanced logic. I think a lot of these memory issues we’re seeing here would have probably been solved if TSMC expanded it earlier. I think they will expand it again in July. I think we just need them to double the fab right now.

Chris Barber: Alphabet.

58:11-58:45

@bubbleboi: So I don’t know about all of Alphabet, but Google is definitely—I feel like a lot of people hate on them because Gemini lags or whatever. But I mean, Gemini’s job isn’t to chase benchmarks, it’s to set what the new benchmark is, in my view. I think the hardware, most of the innovation right now is coming out of the TPU team and Google. Other than like agentic workloads, we’ll see when they want something now. And I think they’ll make a lot of money and they’ll be fine. They’re probably one of the few big tech companies which I have no worries about.

58:45-59:01

@bubbleboi: And it’s weird because SaaS has completely fallen off a cliff. All that’s left now is like hyperclouds and ads.

Chris Barber: What’s falling off the cliff?

@bubbleboi: SaaS. B2B SaaS.

Chris Barber: Yeah. Hyperclouds and chips.

@bubbleboi: Exactly.

59:01-59:12

@bubbleboi: Hyperclouds and ads. That’s where you want to be in the software landscape. You want to be in the cloud and you want to be selling ads, which is why Meta’s still doing it.

59:12-59:27

Chris Barber: Five years from now, how much of Google’s TPUs are they using internally versus selling or renting?

@bubbleboi: Like 25-30% internally.

59:28-59:46

Chris Barber: So heavy external. Is your view that Google’s in an advantageous position? Maybe actually a better question is, Google, OpenAI, Anthropic—five years from now, do you feel like they all are similar in terms of frontierness of their models, or has one broken away significantly?

59:54-59:59

@bubbleboi: That’s such a good question. I think Anthropic is just making so...

1:00:00-1:00:24

@bubbleboi: There’s so much money right now, and money solves a lot of problems, sadly, that they may break away and it might turn into an Anthropic versus Google thing. I think OpenAI, even though they were definitely the first to come out—and I’m not saying to count them out—but at this point, we’re transitioning from the aspect of high burn to high capital appreciation, right? And I don’t think they’ve appreciated that. They still think they’re trying to capture market share at this point, which maybe we are, but you also need to start now making money so you can ensure you have compute and resources for the future. You can’t just burn capital. You actually have to have a plan for accumulating it, right? We’re going from burn to accumulate, and if you can accumulate faster than you can invest it, you’re going to be in a very strong position.

1:00:24-1:01:24

@bubbleboi: Anthropic and Google are both there. They’re just crazy capital accumulators, cash flow accumulators, right? I think OpenAI is still just trying to invest in compute, capture market share, burn at a loss, and I feel like we’re transitioning from there. I think the future is these tokens are very profitable to serve, and you have to make money eventually.

1:01:25-1:01:28

Chris Barber: What’s the point where you are no longer bullish on SanDisk?

1:01:33-1:02:47

@bubbleboi: When SanDisk management makes mistakes, when they’re not adding value to their customer, that’s when I’m not bullish. I will always be bullish on Flash as a technology that could scale in the future, right? But the good thing is, SanDisk is really the only game in town right now who can serve the needs in Flash. We have Micron and technically Samsung, but they’re not really pushing the needle there. SanDisk is only Flash, so that’s their main business. I will not be bullish when we get a CEO in there who says, “Hey, I’m actually returning money to shareholders. We’re going to go longer-term contracts on Flash. I’m just going to cut staff and make things cheaper so we can get better margins.” I do not care about that. I want them to innovate. I want them to push it. And I think they know that Flash could be innovative. I don’t think they’re dumb about this. They’re not playing the game that SK and Samsung are doing, where they just want to price gouge DDR. They’re playing the game of, “Let’s be innovative. We made this money. Let’s take a risk now.”

1:02:47-1:02:50

Chris Barber: Yep. Amazon.

1:02:52-1:03:02

@bubbleboi: My view on Amazon is a little all over the place, honestly. They have a lot of business lines, but I will say—

1:02:59-1:03:02

Chris Barber: You can also just do their chips as well if you wanted.

1:03:02-1:03:40

@bubbleboi: Well, Tranium has kind of not shipped too well. I think they’ve had problems ramping up their next generation of chips. But I will say, if your view is that tokens are like the new electricity or oil, the money is going to accrue to whoever has the infrastructure to serve at the cheapest. And it’s interesting—Anthropic, all their cloud models are basically hosted on AWS, AWS Bedrock, and all these other things. OpenAI made a deal to go onto Bedrock. They have this institutional access point where, if you’re on AWS Bedrock, institutions will just use you. The big banks just trust you. So I think there could be an angle where Amazon—and maybe Microsoft, I don’t know about the others—but Amazon, I definitely think they’re thinking, “Yeah, we might be burning money and not growing too much, but we’re getting people addicted to our cloud and we’re cheap for hosting.”

1:04:06-1:04:49

@bubbleboi: All of that big enterprise market share that Anthropic is winning is because they’re on AWS, right? And I think that gives them a lot of power in the future. Right now, it’s a little iffy, but they may turn around and just say, “Hey, you know what?” And Andrew Jassy’s already mentioned this in his earnings calls—if I spun out Tranium, it’d be this big of a business. This is a big thing with these huge companies: there are little kernels in there that are just so valuable, but it gets diluted by everything else. Flex was like that story too—this little cooling block that’s so important, but the rest of the company kind of dilutes it. We could see, five years from now, that a lot of the value approval might go to the people who own infra and can—just like AT&T charges you for internet and your phone—they’re going to be charging you for your token bill.

1:04:49-1:05:04

Chris Barber: Yeah. It’s funny. The Amazon thing makes you think that right now they’re going to distribute OpenAI and Anthropic tokens, and then eventually they’ll do Amazon Basics tokens.

1:05:11-1:05:18

@bubbleboi: Maybe, yeah. Or maybe they just go to OpenAI and Anthropic, where they didn’t scale out their own compute, and just say, “Hey, guess what? We’re taking a portion now.”

1:05:19-1:05:28

Chris Barber: Yep. The distributor. Next one, ASML. When do you anticipate that we would see ASML—it seems like memory makers have exercised their pricing power. ASML hasn’t as much, but presumably has a lot of pricing power. Do you anticipate that it would exercise pricing power?

1:05:44-1:06:12

@bubbleboi: They don’t have any pricing power, sadly. I don’t think so, because the economics of a fab are like, when you buy this machine, you have to run it all day to get your money back. People are very cautious because it’s a huge capital expenditure, and people are very cautious about spending capacity right now. What pricing power does ASML or the other wafer equipment tools really have? It’s just like the customer is going, “Your machines are too expensive, and we don’t need to buy that much right now. I’m not trying to flood the market today.”

1:06:12-1:07:21

@bubbleboi: ASML—China was a big part of their business, the growth in their business, and now they can’t sell EUV tools to China. Let’s see if they get locked from selling DUV tools or anything else to China, which could happen. I don’t think they’re in a good place. I don’t think there could ever be really a structural shortage because who holds the cards? It’s TSMC and now Intel and Samsung to an extent. They’re not aggressive growth maxis. They like high prices. They like heavy control. They like setting the price. It’s tough for ASML. I will say ASML is trying to expand into advanced packaging now, which I think is very smart. We’ll see how they do, but if they actually do a good job of expanding into it, I can see them taking a lot of market share. We’ll see, but right now, litho is just not expanding in line.

1:07:22-1:07:24

Chris Barber: Yup. Hynix.

1:07:29-1:08:04

@bubbleboi: I think Hynix is probably going to continue being the best HBM supplier and the best advanced memory supplier if they’re doing HPF. I think they definitely have a niche there. If you think memories and innovation in that space are important, SK is investing and taking risks to do it. I think they’ll still maintain their position, but without expanding DRAM supply, they might actually shoot themselves in the foot and cause a bunch of pushback that they don’t want in the long term.

1:08:04-1:08:08

Chris Barber: Because they’ll create competition for themselves, basically?

1:08:08-1:08:13

@bubbleboi: They’ll create competition, but also they’ll just lose goodwill. After a certain point where things are in a very, very, very bad shortage, you will get demand destruction. 32 gigabytes can’t cost me $10,000. It becomes uneconomical to use.

1:08:13-1:08:31

Chris Barber: Are you still long Lumentum at this point or rotating out?

1:08:31-1:08:41

@bubbleboi: No, no, no, no. Definitely sold out of Lumentum a month or two ago.

1:08:41-1:08:45

Chris Barber: Importantly, everybody kind of gets it, basically?

1:08:45-1:09:00

@bubbleboi: Well, also, the procurement cycle is kind of done. We’re going to see these switches come online at the end of the year. Let’s see how they do. If they’re a screaming success, then definitely own Lumentum. If you see that people are willing to replace these lasers and replace these switches, that’s not my view. But if it is true, I mean, yeah, Lumentum’s great. Lumentum has the best lasers for this use case. So we will see.

1:09:05-1:09:08

Chris Barber: You feel the same way then about AXTI and AIXA?

1:09:10-1:09:23

@bubbleboi: For now, yes. But again, they’re just making indium phosphide, as I understand. So that will eventually get competed away.

1:09:23-1:09:29

Chris Barber: Nokia and Infineon in particular.

1:09:30-1:09:59

@bubbleboi: Nokia is interesting because everyone knows Nokia is a telecom company, right? But telecoms actually have to do crazy high-bandwidth data manipulation. It’s the same stuff you’ve got to do now for data centers. So all these data center interconnects are probably going to be using Nokia or competitors. But Nokia is the best at this high-bandwidth data movement, right? And they need to for telecoms if you’re doing 5G, 6G, low-latency communication. So I think data...

1:10:00-1:10:17

@bubbleboi: The data center business is going to grow quite a bit. We’ll see how they do. I don’t have a pick as far as who wins what, but I think they can become a very important data center company with all the networking infrastructure there.

1:10:19-1:10:27

Chris Barber: Wolf.

@bubbleboi: I sold Wolf before joining this meeting today. Yeah, literally, right before. It’s too frothy now. Wolf—so, Intel is my biggest position in my portfolio right now, and Wolf today surpassed it. I was like, okay, something weird is happening with Wolf, right? It just went up so much that it surpassed Intel. I thought, let me rebalance.

1:10:44-1:10:50

Chris Barber: I mean, we’ll see. I think Wolf is basically in a short squeeze right now. We will see.

Chris Barber: You exited it completely or you kept a bit?

@bubbleboi: No, no, no. I had to rebalance. I had to rebalance because it was just too big.

Chris Barber: Okay, so you still hold some, just much smaller.

@bubbleboi: So for context, Wolf was probably up over 100% for me. So it was just too much.

1:11:07-1:11:10

Chris Barber: Enphase and SolarEdge.

@bubbleboi: I don’t have an opinion. I’ve heard of the names. I know why people are maybe a bit bullish because of the inverters for DC, but I don’t have an opinion as far as who’s good or who’s not.

Chris Barber: Yep. Semtech.

@bubbleboi: Again, I can’t give you an opinion.

1:11:30-1:11:58

Chris Barber: Rigaku. That’s the x-ray inspection stuff.

@bubbleboi: I haven’t looked at Rigaku terribly closely, but inspection—and I forgot, there’s a specific term for this, where you test things with light, like for inspection of wafers. It’s definitely an interesting space to be in. I think it’s an interesting product offering.

Chris Barber: Sumimoto and IQE, you would kind of feel similarly—wait for the CPO rollout?

@bubbleboi: Yeah. Yeah, probably.

Chris Barber: Navitas and Texas Instruments.

@bubbleboi: Again, I can’t tell you which. I mean, if you’re a trader and everyone just wants to pick which one to buy and make money off of, it’s probably going to be Navitas because they’re just like a shittier company. And when good news comes to shittier companies, it just makes them do—it saves their lives, right? So Texas Instruments is a very mature and fine company.

Chris Barber: What about Gallium Nitride?

@bubbleboi: Yeah. I mean, I’m in EE, right? But even as you’re going to find out, I’m the digital chip design kind of person. When it comes to power and signal processing and all this stuff, I will defer to people who are smarter than me. But I mean, it’s important, but the question is just, okay, they’re not the only people who can do it. So what are you betting on? What’s your point here? I just want to buy companies that, if they disappeared, it would break your heart if you’re trying to stay in the middle and center, right? That’s how you know someone’s adding value in your life. What if this guy just took a vacation for three months? How would I feel? I don’t think anyone would really care if Navitas took a vacation for three months.

1:13:00-1:13:22

Chris Barber: What was your process for building conviction in Intel? You went to New York, pitched a few hedge funds on it. When did you first see it? What was the thing that made you go from, “Oh, this is interesting,” to, “Oh, I’m actually going to put a large portion of my portfolio in”? How much time did you spend trying to convince other people? What were their reactions? That kind of stuff.

@bubbleboi: It’s a great question. So what led me to build conviction was—the reason why Intel had previously crashed quite a bit was that they were investing. And what were they investing to do? To overtake TSMC. Now, from our whole time up to now, you know that I kind of like when companies are taking a risk and trying to add value to the marketplace, right? But to the finance community, that just tells you negative free cash flow for two years. And I’ve been evaluating this company for like five years as an analyst. I just care about free cash flow and EPS, right?

Chris Barber: So you’re like a technology VC perspective on public?

@bubbleboi: Oh yeah. I almost never—like, I will look at valuation if it’s really bad or really high to understand what’s going on. But my main view is, what’s my view of the world, of technology, and what’s going to happen? And it was clear we needed an alternative to TSMC. They could do it, right? I mean, they fabbed their own chips up to now pretty successfully. Most of the semiconductor industry up to this point has kind of come out or spun out of Intel, right? Including silicon photonics, which people don’t know—they were the first to really invest in that.

So I’m looking at this, I’m looking at a company basically that’s valued less than their assets, right? So I could have bought the whole company, sold their assets, and theoretically made like 10 or 20%. And I was like, okay, this technology—like high NA EUV, 18A—it doesn’t need to be world-beating day one. And I think the market doesn’t get that. You don’t need to beat TSMC day one. You just need to be second best maybe at first and have good quality and good customer relationships. So that was where I was like, this is an obvious long, there’s no risk. Even if they became bankrupt, I would have probably made 10 or 20% on my money, right? But I was also just bullish.

I knew what high NA, what the benefits would be and what maybe some of the drawbacks could be. I knew their innovations in backside power delivery would lead to total efficiency dominance over TSMC as far as wattage. And again, I think wattage matters more—power matters more now than even flops at this point, right? And that’s why Apple went onto Intel for their laptops and stuff. Because they may not be better as far as density right now in 18A, but they’re definitely better on power and no one can take that away from them. And that matters quite a bit.

Then you’re looking at, okay, were they executing? They were. The 18A PDK came out. It wasn’t perfect, but it was good. It was decent. 18AP, which just came out recently, was announced at some conference recently.

1:16:48-1:17:41

Chris Barber: What was the process like where you were kind of going through all this stuff? This was like, you’re doing a lot of reading, talking with friends of yours that are technical, talking with friends of yours that play in the public markets?

@bubbleboi: Yeah. So my process was literally reading everything I could find on lithography, especially high NA versus just low NA. And how many layers of EUV would they need to compete with TSMC? So it was all my own reading and research. And at the time, I honestly wasn’t talking to any friends. One thing I think people don’t know is I actually don’t try to talk to people for investment ideas. I try to avoid that as much as possible because I really don’t want anyone clouding my judgment. I’ll talk to them about technology though. So it was kind of obvious to me that they were positioning themselves to be strong and they just had a balance sheet issue.

Chris Barber: What was the time period between when you first kind of started diving in to when you developed enough conviction to have a big position—like weeks or months or shorter, longer?

@bubbleboi: It was like a month maybe. So I think January of 2025 was when Intel was really at their lowest probably. They fired their old CEO. There wasn’t a CEO in place. Then Louboutin took over. So right then is when I was like, okay, well, I definitely know the company’s undervalued completely. It’s basically valued like it’s going bankrupt. It’s not. I know their technology roadmap is good. And then I knew that they didn’t have to just be the best, best fab. I mean, Samsung’s shittier than Intel right now and they’re still alive as a logic fab, right? So I just knew the technicals were there. And to me, it’s like, when you know you’re right—when you absolutely know you’re right—like most of the time we don’t know that we’re right. We don’t know what the demand for GaN is going to be or if silicon photonics is going to work. I arguably do not know these things. I’m very aware of what I don’t know. But I knew for sure that Intel was being valued below the value that they are making and that they would not go bankrupt. So it was very easy for me to just put all my money into it. And from there, the thesis played out perfectly.

At the time, packaging wasn’t in my thesis, but soon I started understanding about the packaging and that’s when I quadrupled down on packaging. And 18A shipped fine. 18AP is now amazingly like 30% better than 18A. 14A is basically almost done. 1.0 PDK probably next quarter. And it’ll probably be volume production sometime next year. So it was a no-brainer for me.

Chris Barber: What made you go and pitch it to funds?

@bubbleboi: Well, I was just like, people thought I’d changed so much about it. People were like, what’s wrong with this guy? Why does he like Intel so much? So people would chat. I was in New York at the time. So friends in New York would be like, “Yo, my friend’s really into Intel. You should talk—”

1:20:00-1:20:13

@bubbleboi: And I would tell them my idea and they were like, “All right.” But it was when Trump—when the government took a stake with the CHIPS Act money, when they asked for equity instead of a loan—that’s when they started listening to me a little bit.

1:20:13-1:20:27

@bubbleboi: And then once I got really right—like once the target announcements came out—everyone was like, “How did you see this?” So a lot of people passed, by the way. There’s only one guy who heard me out and immediately made it the biggest position in his fund.

1:20:27-1:20:41

@bubbleboi: I can’t tell you his name because he’s anonymous, but he’s a really good friend of mine. He’s, you know, done—what’s the word for it—like economically... Yeah, like economically independent. No, it’s not even that. It’s like post-economic. That’s the word. Yeah, he’s post-economic at this point, off of Intel’s move.

1:21:00-1:21:12

@bubbleboi: So shout out to him. And then, of course, the big funds ended up talking to me later. They’re like, “What do you know?” They thought I might have had insider info. I’m like, “No, actually, you can just read these papers that come out of these conferences.”

1:21:13-1:21:23

@bubbleboi: The next one—I think your answer to this might be none—but do you pay for any news or info sources?

1:21:24-1:21:29

@bubbleboi: I pay for Bloomberg Terminal. Yep. I don’t ever use it. I actually just use it to message my friends in finance sometimes for fun and get them in trouble at their job.

1:21:29-1:21:42

@bubbleboi: Yeah, I use it sometimes to look at nice little graphs. Other than that, I really just use Twitter. But even the Bloomberg Terminal, right—I kind of got it just because if you get a meeting with someone at a big bank and they want to lend you money, they’re like, “Who is this guy?” They don’t really know who you are.

1:21:42-1:22:06

@bubbleboi: So it kind of got me a little legitimacy, right? Sadly, that’s how that world works. But the main value is being on Twitter and talking to people who are engineers. I just want to talk to engineers who are experts and who can teach me. That’s the value. Don’t read research—I mean, SemiAnalysis is great, it’s fun to read their research—but I’m trying to understand the technology and what happens when something changes.

1:22:06-1:22:29

@bubbleboi: I’m not trying to predict who’s going to win what, or when DDR is going to roll over, yada yada. There are smarter people out there than me who can figure that out. I’m trying to find what’s the next big thing.

1:22:29-1:22:43

Chris Barber: What are some of the X accounts that you most often read for this stuff?

1:22:44-1:23:13

@bubbleboi: So I like Yukon quite a bit. Drew Cars is just a friend—I like chatting with him. I run a chat on Twitter, so he chats with the best electrical engineers in the world—like 200 of them—and they’re all the most cracked people ever. So you can just go there if you ever have a technical question on photonics or EE, just chat with some people in there, they’ll tell you. Lucas Grozon—he’s very good in photonics.

1:23:13-1:23:28

@bubbleboi: Outside Five Sigma is by far my favorite account for technical info. By far. I’m smart—Outside Five Sigma. Yeah, Outside—he’s a good friend of mine, one of my favorite people to talk to. There are a couple others that will come up, but those are the main ones. I’m trying to think who else.

1:23:29-1:23:44

@bubbleboi: Vikram—I think everyone knows Vikram’s kind of been doing a good job. So Vikram Sarkar. I really like the other Krums’ Substack. Some small accounts—Lithos, oh, so Lithos Graphene or something. Yeah, he’s a real fab engineer doing research on this stuff.

1:23:44-1:24:16

@bubbleboi: So he’ll give you info that you would never think of. And then Laurie Wired was a very big account. I really like reading her tweets because she’s extremely technical and knows what she’s talking about. Those are the main people I like to follow.

1:24:16-1:24:32

Chris Barber: Any other Substacks that you tend to read? Most of the stuff that goes in your inbox?

1:24:32-1:24:55

@bubbleboi: No, no. Oh, I will say there’s one analyst that I really like, though. He just got on Twitter. His name is Jeff Pu. He’s an analyst in Hong Kong or something—like a sell-side analyst. He works for a bank or some kind of brokerage, but he was one of the few people who was really bullish on Intel.

1:24:55-1:25:05

Chris Barber: How do you spell his last name?

1:25:05-1:25:15

@bubbleboi: It’s P-U, I think. Jeff Pu. If you look up—I think I found it, GF Securities. Yes, yes, yes. So Jeff Pu, to me, is a semiconductor analyst, and he’s going to see things other people don’t because he’s in Hong Kong. He can help with these Asian companies all day.

1:25:17-1:25:31

@bubbleboi: There’s another small account—Mohammed Zuhair. Mohammed was really early on Intel as well, and I think he has the best coverage on Intel specifically.

1:25:31-1:25:33

Chris Barber: Cool.

1:25:33-1:25:37

Chris Barber: Any Discords or conferences or things like that?

1:25:39-1:25:51

@bubbleboi: ISCC, maybe—the International Solid-State Circuits Conference—every February. And then SPIE’s in September, so if you want to learn about optics and photonics, SPIE is great. And Computex, which is coming up in June, but that’s more of an industry conference, not academic.

1:26:08-1:26:18

Chris Barber: The answer to this is probably no as well, based on your style, but are there any investors where, when they take a position in something, you’re like, “Okay, I’m going to give this a serious look”? Just the fact that they have a position makes it worth my time to dig in.

1:26:18-1:26:41

@bubbleboi: Absolutely. I think Druckenmiller, maybe in the past—Stanley Druckenmiller. But Stanley’s a generalist. I would probably look at what he does maybe in macro trades, not really single stocks.

1:26:41-1:26:44

Chris Barber: Any publications or projects in your future? And a related question to this is: what was Bubble Computing?

1:26:45-1:27:03

@bubbleboi: Mmm, so it might still be happening right now. I really like flash, so—yep, I want to get in the flash game.

1:27:03-1:27:06

Chris Barber: Last question: anything I should have asked that I did not ask?

1:27:07-1:27:24

@bubbleboi: That’s the best question. Maybe the question is: why care about any of this, right? Why are we expending so much energy on this and researching it and killing ourselves? I think part of the answer a lot of people might want me to give is, “I can make money off of this.” But honestly, if money was your only motivation, you’d go sling SPVs in San Francisco or something. That’s a way quicker way to get some cash.

1:27:24-1:27:54

@bubbleboi: I mean, I just think we’re living in a new computer revolution. We lived in this in the ‘80s, probably in the early ‘70s. That’s when Intel was the growth—Intel was the OpenAI. Cisco was the big company, now Cisco’s going crazy. So we’re in a new computer revolution, and I would just tell people: CS was the big thing—computer science, programming. It’s still very much important, but I think now we’re realizing that the problems we’re hitting are on the edge of physics. We’re bumping our heads on it.

1:27:54-1:28:26

@bubbleboi: We need to go back to first principles and learn this stuff. I think now is the most exciting time to be doing engineering because we’re at the limit—we are at the edge. You’re not redesigning the little chip in your iPhone just for fun, right? You’re not doing incremental changes. We have to rethink computers and architecture and how we’ve been doing things from first principles now, and it creates so much excitement.

1:28:26-1:28:55

@bubbleboi: I think you said people follow me like it’s a TV show at this point. Well, I follow some of the computer stuff like it’s a movie, right? So it’s very exciting. I think a lot of people are like, “Oh, yeah.” Let’s not get lost from the fact that we really are in this computer revolution, thanks to AI.

1:28:55-1:29:07

Chris Barber: Yep. Sick. Yeah, it makes me think—the demand is so strong that now we have the need for genuine new innovations, and it sounds like that’s the thing that gets you.

1:29:07-1:29:45

@bubbleboi: Oh, yeah. Because all this stuff about new memory architectures, photonics—photonics has been researched in academic environments since the ‘90s. People have been designing this stuff for years. It’s just now we’ve reached the point where that’s it, man—give it a shot. Let’s try to do it. So I think it’s so exciting for me as someone who’s always loved electrical engineering and the stuff on the limit. It’s very exciting.

Made with The Transcript Desk Chrome Extension: https://thetranscriptdesk.two-lambda-ai.com/.