WEBVTT - AI Made Everything Expensive (ft. GamersNexus) 0:00:02.440 --> 0:00:03.560 Also media. 0:00:05.880 --> 0:00:09.320 Hello on, welcome to Better Offline. I'm your host ed Zitron. 0:00:21.680 --> 0:00:24.000 Subscribe to the newsletter, joined the subway and buy a 0:00:24.040 --> 0:00:25.599 T shirt. All the links are in the notes. And 0:00:25.640 --> 0:00:27.400 here's a joke to kick us off. What do you 0:00:27.440 --> 0:00:30.000 call the CEO of a French RAM company? That's right, 0:00:30.320 --> 0:00:33.440 Immanuel Micron, And that's what we're here to talk to 0:00:33.479 --> 0:00:36.400 Steve Burke of Gamers. Next is about today, Steve, Welcome 0:00:36.400 --> 0:00:37.239 back to the octagon. 0:00:37.600 --> 0:00:38.880 Emmanuel Micron. 0:00:39.360 --> 0:00:43.600 Yeah, before we started, yeah, because you would have killed 0:00:43.640 --> 0:00:46.440 me before I read it on the But look, we're 0:00:46.479 --> 0:00:49.000 here to talk about RAM. So we'll start with a 0:00:49.040 --> 0:00:52.640 simple question. Why is RAM so expensive? Right now? What 0:00:52.840 --> 0:00:53.519 is doing this? 0:00:56.080 --> 0:00:59.600 I mean, the reason it's so expensive is not grounded 0:00:59.640 --> 0:01:04.120 and reac I think you know, it's the surface of 0:01:04.160 --> 0:01:06.600 it is that all the data center build out is 0:01:06.600 --> 0:01:10.080 consuming all the memory and therefore, you know, the price 0:01:10.160 --> 0:01:13.280 higher because everyone wants it. But I do think there's 0:01:13.280 --> 0:01:14.160 a little bit more to. 0:01:14.080 --> 0:01:18.039 It than that, right So from what so when you 0:01:18.120 --> 0:01:20.760 build chips We're going to really simplify this one for 0:01:20.800 --> 0:01:23.240 the audience because it's pain in the US to explain otherwise. 0:01:23.520 --> 0:01:26.120 But when you build chips. You have something called a waifer. 0:01:26.200 --> 0:01:29.280 You have the silicon out, and then you entch it 0:01:29.400 --> 0:01:34.119 using a lithograph machine. But you can make less high. 0:01:34.280 --> 0:01:37.520 So there's the high bandwidth memory that they use inside GPUs, 0:01:38.240 --> 0:01:40.280 and then there's the regular DRAM that they use in 0:01:40.319 --> 0:01:43.000 everything else. Right, is that it's a simplification. 0:01:43.040 --> 0:01:46.840 But is that failure? I think like the the most 0:01:46.840 --> 0:01:49.800 base level stuff to know is that memory, some form 0:01:49.800 --> 0:01:54.040 of RAM or memory is in almost everything. A common misconception, 0:01:54.320 --> 0:01:56.800 I'm not sure. I think your audience is like fairly technical, 0:01:56.840 --> 0:02:01.120 but a layperson misconception is that storage is the same 0:02:01.120 --> 0:02:04.600 as memory. They'll use it interchangeably. If storage is like 0:02:04.640 --> 0:02:07.320 your drive that you you know, you put data on, 0:02:07.400 --> 0:02:10.880 like photos or whatever, and then memory is is a 0:02:10.919 --> 0:02:13.799 temporary form of storage of volatile storage, like it's wiped 0:02:13.840 --> 0:02:17.000 when it's not actively being used when you restart and. 0:02:16.960 --> 0:02:19.320 It's and when they put stuff in memory, that's where 0:02:19.880 --> 0:02:23.600 that's like RAM and such where they like, that's where 0:02:23.600 --> 0:02:25.160 you put the stuff that you gonna need quickly. 0:02:25.560 --> 0:02:28.880 Right, Yeah, so yeah, memory is much faster than storage, 0:02:28.919 --> 0:02:31.600 permanent storage, and so it acts as the sort of 0:02:31.639 --> 0:02:35.000 transactional buffer to help get things to you quickly, So 0:02:35.280 --> 0:02:39.560 memories and phones, refrigerators these days, cars, there's all kinds 0:02:39.600 --> 0:02:43.760 of different grades of memory, and it's it's used in 0:02:43.760 --> 0:02:46.720 pretty much everything to do some kind of quick like 0:02:46.760 --> 0:02:50.120 in a car, Uh, maybe you need to retrieve something 0:02:50.240 --> 0:02:53.520 related to like if there's a built in map on 0:02:53.600 --> 0:02:55.919 the screen or something like that, or in a refrigerator. 0:02:56.080 --> 0:02:58.480 Who the fuck knows, honestly, but I think that's I 0:02:58.560 --> 0:02:59.200 think so that. 0:02:59.080 --> 0:03:01.679 You can have a giant, horrible screen on your Samsung 0:03:01.760 --> 0:03:05.840 fridge that displays you adds occasionally, yes, exactly, yeah, and 0:03:05.880 --> 0:03:08.679 so so yeah, to go to your wafer count. Yeah, 0:03:08.720 --> 0:03:12.560 there's different types of sort of production for memory, and 0:03:14.440 --> 0:03:17.200 you'll find DRAM is used in all of it in 0:03:17.240 --> 0:03:19.920 some capacity for the most part. So you'll find DRAM 0:03:20.639 --> 0:03:24.600 dies in HBM like inside of an HBM package. But 0:03:25.280 --> 0:03:29.440 data center high end GPUs will use high bandwidth memory, 0:03:30.040 --> 0:03:34.760 and basically every consumer device uses something called GDDR, which 0:03:34.800 --> 0:03:39.320 is just a lower bandwidth kind of more normal solution 0:03:39.880 --> 0:03:43.040 that's not as overboard on the performance for a consumer application. 0:03:43.440 --> 0:03:47.040 But they all kind of come from the same place, 0:03:47.080 --> 0:03:50.240 which is a wafer. They're all cut from right so 0:03:51.320 --> 0:03:55.400 right now, they're taking up more, they are using more 0:03:55.440 --> 0:03:57.240 of the how much of the compacity are they actually 0:03:57.320 --> 0:03:59.400 using to build high bandwidth memory? So the stuff that's 0:03:59.400 --> 0:04:00.440 going to the GP use. 0:04:01.560 --> 0:04:04.680 It appears to be an awful lot. I mean, so, 0:04:06.280 --> 0:04:10.760 I know that the recent class action lawsuit or complaint 0:04:10.760 --> 0:04:15.360 I should say that was filed notes that the profitability 0:04:15.520 --> 0:04:21.200 the plaintiffs attorneys claimed that the profitability of traditional conventional 0:04:21.240 --> 0:04:25.760 DRAM is greater than the profitability of HBM. That was 0:04:25.800 --> 0:04:30.000 one of their claims. And yet they sort of argue 0:04:30.080 --> 0:04:32.400 that it seems like their evidence is good, but they 0:04:32.520 --> 0:04:37.200 argue that why would you shift more and more capacity 0:04:37.279 --> 0:04:40.839 to producing this lower margin thing HBM? Right of like 0:04:40.920 --> 0:04:45.159 the easier thing to make is potentially higher margin, And 0:04:45.200 --> 0:04:47.120 so it does seem like more of it's shifting over there. 0:04:47.440 --> 0:04:47.640 You know. 0:04:47.680 --> 0:04:51.039 Open AI cause that sort of big kerfuffle previously when 0:04:51.360 --> 0:04:56.640 they claimed forty of just all DRAMs globally though yeah 0:04:56.640 --> 0:04:58.480 it didn't have it was all like basically. 0:04:58.320 --> 0:05:00.840 One of the most insane just I mean sane story. 0:05:00.880 --> 0:05:02.680 By the way, just last year, for those of you 0:05:02.720 --> 0:05:05.240 who don't remember, open A story came out the open 0:05:05.279 --> 0:05:08.360 A I was buying thirty five to forty percent of 0:05:08.560 --> 0:05:12.560 all the world's dram wafers. It never happened, Yeah, but 0:05:12.600 --> 0:05:14.960 we just talked about it like like we did, like 0:05:15.000 --> 0:05:15.440 it did, but. 0:05:16.000 --> 0:05:17.320 Now it did affect the price. 0:05:17.920 --> 0:05:21.719 It did. But here's the question, So why is everything 0:05:21.800 --> 0:05:25.000 so expensive that like is it? Is it the because 0:05:25.880 --> 0:05:27.719 are they just jacking the price up? 0:05:29.320 --> 0:05:31.680 I think so? I mean so, yeah, there's a couple 0:05:31.640 --> 0:05:35.159 of things here a layout like some of the highlights, 0:05:35.160 --> 0:05:37.240 and we can you can figure out how you want 0:05:37.240 --> 0:05:39.040 to go through them. But sure one of them is 0:05:39.160 --> 0:05:43.880 the memory industry has convictions actually in its past of 0:05:44.000 --> 0:05:48.120 price fixing, collusion and as the DOJ dubbed it, cartel 0:05:48.200 --> 0:05:51.360 like behavior. So this goes the early two thousands, there's 0:05:51.360 --> 0:05:53.479 a bus to twenty sixteen, twenty eighteen. There was an 0:05:53.480 --> 0:05:57.520 attempt at a similar lawsuit in the early nineties. The 0:05:57.560 --> 0:06:02.560 memory manufacturers were talking about how they thought their competitors 0:06:02.600 --> 0:06:05.279 were price fixing and colluding and so this is a 0:06:05.360 --> 0:06:08.400 long and story to history where they and now there's 0:06:08.400 --> 0:06:12.599 an active lawsuit, putative class action for price fixing. So 0:06:12.800 --> 0:06:16.560 there's that antitrust aspect, which that obviously you know, prices 0:06:16.560 --> 0:06:19.800 would be higher if you're just doing illegal things to 0:06:19.800 --> 0:06:22.440 make the prices higher. There's also the sort of physical 0:06:22.440 --> 0:06:29.240 manufacturing side, which is that to make memory, it's not 0:06:29.400 --> 0:06:32.559 quick to get another fab online, and the companies seem 0:06:33.279 --> 0:06:35.679 to sort of signal to each other through earnings reports 0:06:36.480 --> 0:06:41.120 whether they're going to expand capacity or maintain capacity of 0:06:41.160 --> 0:06:45.320 their fabs, which are factories for memory. So there's that aspect, 0:06:45.360 --> 0:06:49.080 which is if there's a real or not a sudden 0:06:49.080 --> 0:06:51.840 increase in demand, there's going to be a long latency 0:06:51.880 --> 0:06:54.200 to react to that with more supply, if they even 0:06:54.279 --> 0:06:57.160 want to. So the current projections are twenty twenty eight 0:06:57.240 --> 0:07:01.120 or twenty thirty for getting like more fabs online. The 0:07:01.200 --> 0:07:03.440 other reason it goes up, though, is really simple, which 0:07:03.480 --> 0:07:06.599 is that in a server that goes in a data center, 0:07:07.279 --> 0:07:09.600 some of Nvidia's newer solutions will use one and a 0:07:09.640 --> 0:07:13.840 half terabytes of system memory. So that's d RAM they'll use. 0:07:14.240 --> 0:07:17.280 Well they use this, they don't. They also use the 0:07:17.320 --> 0:07:20.640 same RAM as we find in smartphones as well. 0:07:20.960 --> 0:07:26.040 Yeah, LPDDR typically Yeah, yeah, so so Grace will use 0:07:26.200 --> 0:07:28.160 I think it's five X I think is what they're on. 0:07:28.640 --> 0:07:31.960 Yeah, I pulled the numbers and it's something like twenty 0:07:32.080 --> 0:07:34.880 terabytes for an NVL seventy two, which is seventy two 0:07:34.920 --> 0:07:35.600 GPU rack. 0:07:36.560 --> 0:07:42.280 Yeah, so they do. The single blade for the Big 0:07:42.320 --> 0:07:46.680 Boy is one and a half terabytes of system memory alone. 0:07:46.800 --> 0:07:51.680 And you know, for perspective, a consumer desktop computer might 0:07:51.720 --> 0:07:54.680 have sixteen gigabytes with Windows these days, you probably want 0:07:54.720 --> 0:07:57.440 like thirty two, and they're they're taking one and a 0:07:57.440 --> 0:08:01.680 half terabytes for one, you know, server solution. But that's 0:08:01.760 --> 0:08:05.320 just the system mary. They also have other processors on there, 0:08:05.400 --> 0:08:08.360 like the Blue Point or blue Field I think it's 0:08:08.400 --> 0:08:10.600 called DPU. It takes one hundred and twenty eight gigabytes 0:08:10.600 --> 0:08:13.240 if I remember correctly. The GPUs take hundreds of gigabytes 0:08:13.280 --> 0:08:17.280 to like close to half a terabyte per GPU of HBM, 0:08:17.840 --> 0:08:22.800 and the HBM has a there's kind of a yield 0:08:22.840 --> 0:08:26.040 factor at play where you know you can when you've 0:08:26.080 --> 0:08:29.040 talked about manufacturing, that means you're also talking yield how 0:08:29.080 --> 0:08:31.960 many good products do you get for every bad And 0:08:33.040 --> 0:08:36.040 HBM is more difficult to make. If you have a 0:08:36.120 --> 0:08:41.400 failure in manufacturing, then you're also potentially losing more silicon 0:08:41.520 --> 0:08:42.760 dyes that could have gone. 0:08:43.920 --> 0:08:47.920 And HBM is made by stacking the memory, right, so 0:08:47.960 --> 0:08:51.880 it requires just physically more memory. That's kind of an 0:08:51.920 --> 0:08:52.720 APES description. 0:08:53.440 --> 0:08:57.600 No, No, it's fine. Yeah, So HBM high bandwidth memory 0:08:58.040 --> 0:09:01.600 is literally just really high bandwidth. It's kind of like 0:09:01.679 --> 0:09:04.480 if you were to if you were a city and 0:09:04.520 --> 0:09:07.200 you wanted to solve a public transit problem and you 0:09:07.360 --> 0:09:10.240 just built a shitload of I don't know, let's. 0:09:10.080 --> 0:09:11.280 Take a massive tunnel. 0:09:11.960 --> 0:09:14.040 You build like a bunch of trains. You build a 0:09:14.040 --> 0:09:17.000 ton of trains for passengers, but you don't have enough 0:09:17.040 --> 0:09:19.320 tracks to put those trains on. They can't actually get 0:09:19.320 --> 0:09:22.360 on the track without blocking each other. You haven't really 0:09:22.400 --> 0:09:24.600 solved your problem. And so high bandwidth is trying to 0:09:24.640 --> 0:09:27.000 solve that by saying, Okay, we've got all this capacity 0:09:27.040 --> 0:09:30.200 for passengers, which is the memory itself. Now we need 0:09:30.240 --> 0:09:34.160 something like a bigger road system to actually feed that 0:09:34.240 --> 0:09:37.320 capacity through to access it. And so HBM tries to 0:09:37.320 --> 0:09:41.800 solve that, especially because you look at GPUs that might 0:09:41.880 --> 0:09:46.200 have like twelve plus times the amount of capacity that 0:09:46.240 --> 0:09:49.719 a consumer GPU would have, and so you don't want 0:09:49.760 --> 0:09:54.679 to starve the GPU core of the data that sits 0:09:54.720 --> 0:09:56.680 in the memory by not being able to get it 0:09:56.800 --> 0:10:01.160 fast enough. And so HBM tries to solve that problem. So, yeah, 0:10:01.200 --> 0:10:03.880 to go back to your point of stacking, high bandwidth 0:10:03.920 --> 0:10:08.840 memory is physically vertically stacked, as in, they take silicon 0:10:08.960 --> 0:10:12.600 dyes for memory and they stack them atop each other 0:10:12.640 --> 0:10:15.800 and then they connect them all together, as opposed to 0:10:16.000 --> 0:10:21.040 consumer memory, where it's closer to there's still layers, but 0:10:21.040 --> 0:10:22.560 it's closer to a flat plane. 0:10:23.559 --> 0:10:25.959 Right. But it just means that by virtue of that, 0:10:26.200 --> 0:10:28.440 it takes up more space in the place where they 0:10:28.480 --> 0:10:29.160 build the chips. 0:10:29.400 --> 0:10:33.360 Yeah. Yeah, if you're on a consumer device, then the 0:10:33.640 --> 0:10:37.440 consumer memory you'll need more physical area to put the 0:10:37.440 --> 0:10:41.280 same amount of capacity, whereas with the high end server 0:10:41.360 --> 0:10:44.679 stuff you just stack that capacity vertically. And then it's 0:10:44.720 --> 0:10:50.640 also hbm's really significant advantage is that it's very low 0:10:50.679 --> 0:10:53.760 power comparatively, right, Yeah. 0:10:54.040 --> 0:10:55.880 But just to be clear, it takes up more space 0:10:55.920 --> 0:10:57.320 on the fabricator itself. 0:10:57.960 --> 0:11:00.920 Oh yeah, yeah, if you're talking about the space, then yes, 0:11:01.080 --> 0:11:05.160 HBM requires a lot more space for a few reasons. 0:11:06.080 --> 0:11:09.120 In terms of manufacturing area. In terms of like footprint 0:11:09.240 --> 0:11:11.839 on a product, it's less, but in a fab it's more. 0:11:12.040 --> 0:11:17.559 And why is that in the fab One of the reasons. 0:11:17.720 --> 0:11:21.120 So there's a few, but one is that HBM requires 0:11:21.120 --> 0:11:27.000 additional silicon pieces to be able to function. So with 0:11:27.280 --> 0:11:31.840 GDDR consumer GPU memory, you pretty much you have the 0:11:31.880 --> 0:11:34.959 memory module. It looks like a black rectangle, and then 0:11:35.080 --> 0:11:38.280 if you were to use like acid or something to 0:11:38.280 --> 0:11:40.960 etch away the black plastic part, you'd be left with 0:11:41.440 --> 0:11:44.360 pretty much a singular silicon die underneath. That's all it 0:11:44.440 --> 0:11:48.760 is for consumer. For HBM, you also need a logic dye, 0:11:48.920 --> 0:11:52.240 So there's another piece of silicon that interfaces with the 0:11:52.280 --> 0:11:55.600 memory and the stack of memory, because you need something 0:11:55.640 --> 0:12:01.520 to kind of to command and control and understand, you know, 0:12:01.559 --> 0:12:03.840 the vertical tower you've got of memory. And then you 0:12:03.880 --> 0:12:08.480 also need an interposer die, which sits underneath everything. So 0:12:09.040 --> 0:12:12.320 this is a solution that will sit under your GPU core, 0:12:12.440 --> 0:12:15.360 which is a very large piece of silicon and is 0:12:15.440 --> 0:12:19.680 shared with the HBM stacks. So all of your your 0:12:19.720 --> 0:12:22.760 core and all your HBM is stacked on top of 0:12:22.800 --> 0:12:25.280 the same it's called an interposer. It looks like a 0:12:25.480 --> 0:12:28.640 PCB and they all communicate through that. So that's another 0:12:29.040 --> 0:12:32.880 piece of silicon. And so for this these reasons alone, 0:12:33.080 --> 0:12:35.559 there's also others, but you know these alone. Now you're 0:12:35.600 --> 0:12:39.400 involving more fabs. There's more logistics shipping things around, and 0:12:39.440 --> 0:12:44.120 there's more fab or fabrication plant area allocated. Even if 0:12:44.160 --> 0:12:46.400 you're just talking about making someone's got to make the 0:12:46.400 --> 0:12:50.480 logic dies TSMC makes the interposers, right. Yes, there's just. 0:12:50.480 --> 0:12:52.720 More moving parts and that kind of takes up more 0:12:52.720 --> 0:12:58.319 space exactly. Yeah, So why So to my early question, 0:12:58.720 --> 0:13:01.640 so this stuff is obviously lower margin, but they seem 0:13:01.679 --> 0:13:04.160 to be able to sell an absolute shit ton of it. 0:13:04.240 --> 0:13:07.120 So is that potentially why they're doing this that just 0:13:07.679 --> 0:13:11.000 they know there is insane demand on a level that 0:13:11.640 --> 0:13:13.560 for a high well, I guess they can charge what 0:13:13.600 --> 0:13:15.600 they want for it as well. I'm just having trouble 0:13:16.520 --> 0:13:19.000 trying to work out what it is they're doing other 0:13:19.040 --> 0:13:20.480 than just filling their boots. 0:13:20.760 --> 0:13:22.800 Have you try to asking the chat GPT. 0:13:23.920 --> 0:13:27.400 No, I could ask my cat and get an answer 0:13:27.400 --> 0:13:30.800 of respect, but it's genuine. But it's genuinely because it 0:13:30.920 --> 0:13:35.240 seems the unless I'm misunderstanding here, that these micro and 0:13:35.360 --> 0:13:38.760 eske High Inex and Samsung they're just kind of cranking 0:13:38.840 --> 0:13:41.720 up the price of regular RAM because they can. 0:13:43.040 --> 0:13:45.440 I mean to some extent. Yeah, So I think this 0:13:45.480 --> 0:13:49.439 is where the price fixing allegations come into I don't know, 0:13:49.480 --> 0:13:51.480 if you try it, if you try to take their side, 0:13:51.520 --> 0:13:52.920 which I really don't want it. 0:13:52.960 --> 0:13:54.200 I'm trying. I'm trying. 0:13:54.360 --> 0:13:58.720 Yeah, if you try to, I think their argument, whether 0:13:58.720 --> 0:14:02.240 it's for truthful or not, but would be We've got 0:14:02.280 --> 0:14:05.960 this crazy high demand for HBM. These are guaranteed customers. 0:14:06.080 --> 0:14:12.360 I can sell terabytes of it to Jensen, and it's 0:14:12.480 --> 0:14:15.640 it's one point of contact. It all moves quickly and easily. 0:14:16.280 --> 0:14:18.920 And this there's this data center demand right now, and 0:14:18.960 --> 0:14:20.400 we don't know if it's going to be here forever. 0:14:20.440 --> 0:14:22.640 But consumers, you know, they're they're always going to have 0:14:22.720 --> 0:14:25.400 to buy electronics or whatever. So I'm not exactly sure 0:14:25.440 --> 0:14:28.720 if that's to me, it seems like maybe you make 0:14:28.800 --> 0:14:33.640 some sort of you assign some value to the B 0:14:33.680 --> 0:14:37.800 two B relationship over B two C or something maybe, 0:14:37.840 --> 0:14:43.920 But also I guess if they yeah, I'm not sure, 0:14:43.960 --> 0:14:44.640 if you make. 0:14:44.840 --> 0:14:47.760 Why would that mean that they charge more for regular RAN. 0:14:47.840 --> 0:14:50.000 Oh As far as why more for regularly, that part's easy. 0:14:50.080 --> 0:14:56.000 That's just because the HBM itself takes the same a 0:14:56.000 --> 0:14:59.600 lot of the same machines as consumer memory. And so 0:14:59.600 --> 0:15:02.880 if you're made more HBM, you're naturally making less consumer, 0:15:03.880 --> 0:15:06.280 which I think is where you start seeing the supply 0:15:06.400 --> 0:15:11.480 squeeze and the price go up. Likewise, a single piece 0:15:11.520 --> 0:15:14.480 of HBM can use up to something I think it's 0:15:14.520 --> 0:15:16.920 like it might be like thirty two dies max or 0:15:16.960 --> 0:15:21.480 something of DRAM you'll commonly see like four referenced. But 0:15:22.320 --> 0:15:27.520 single piece of HBM will take more dram dies to build. 0:15:27.560 --> 0:15:30.840 And so one of the problems is if let's just 0:15:30.920 --> 0:15:33.000 say four, If you have a piece of HBM that's 0:15:33.000 --> 0:15:35.600 got four dram dies in it and you have one 0:15:35.680 --> 0:15:40.680 failure somewhere in that stack or in packaging an assembly, 0:15:41.760 --> 0:15:46.840 now you've destroyed you know, four individual pieces, rather than 0:15:46.880 --> 0:15:49.080 if you have a yield failure with a consumer device, 0:15:49.600 --> 0:15:50.280 you destroy one. 0:15:50.360 --> 0:15:53.280 Sorry to ask a stupid question, what is a die exactly? 0:15:53.360 --> 0:15:56.120 Is that just the what we would think of as 0:15:56.160 --> 0:15:57.240 a stick of ram. 0:15:57.800 --> 0:16:02.840 Uh No, a die is so a wafer we'll start 0:16:02.880 --> 0:16:06.240 the wavefer level. A wafer is just a disc, right, 0:16:06.280 --> 0:16:09.400 it looks like a CD that typically eight or twelve inches, 0:16:11.000 --> 0:16:15.360 and they they process it and they pattern it where 0:16:15.360 --> 0:16:19.560 they literally project like a pattern or they lithographically etch 0:16:19.600 --> 0:16:24.640 a pattern onto the surface of this disc. Once that's done, 0:16:25.000 --> 0:16:28.520 you know, these these patterns, they're called patterns because it's 0:16:28.520 --> 0:16:35.240 a repeating sort of set of shapes that ultimately form 0:16:35.280 --> 0:16:40.080 the logic that makes the silicon. And the word dye 0:16:40.160 --> 0:16:43.320 comes from the process dicing. So with a wafer, when 0:16:43.360 --> 0:16:48.000 they're done producing it and patterning it and it's basically 0:16:48.040 --> 0:16:50.680 ready to become a product, they need to cut that 0:16:50.800 --> 0:16:52.600 disc into a bunch of small squares. 0:16:53.520 --> 0:16:55.600 So it's like a big slab that they caught top. 0:16:56.080 --> 0:16:58.320 Yeah, it's I mean, it's like making a pizza with 0:16:58.360 --> 0:17:02.000 a pizza cutter. You know, they're they're doing a pattern 0:17:02.040 --> 0:17:04.200 that only psychopaths would do on a pizza, which is 0:17:04.240 --> 0:17:07.800 making a bunch of squares and uh, and so they 0:17:07.880 --> 0:17:10.720 cut it into the squares. That's called dicing. And then 0:17:10.760 --> 0:17:14.240 each of those pieces, the squares that yield from that, 0:17:14.280 --> 0:17:17.439 those are the dyes. And so if you were to 0:17:17.440 --> 0:17:21.360 pull up an image of any GPU, typically you'll see 0:17:21.400 --> 0:17:24.720 one large it's called the monolithic dye, which means one 0:17:24.760 --> 0:17:30.600 singular sort of piece of shiny reflective silicon mirror like finish. 0:17:30.600 --> 0:17:33.520 That's your die. And then for memory, same exact concept, 0:17:33.600 --> 0:17:35.040 just a lot smaller. 0:17:35.080 --> 0:17:40.240 Right, So right, okay, so there's less space to make 0:17:40.320 --> 0:17:43.440 regular rams, so they just crank up the price. That's 0:17:43.520 --> 0:17:47.440 really that really is it? I think that's their entire thing, right, 0:17:47.600 --> 0:17:50.400 But I saw Micron zernings and they had eighty four 0:17:50.440 --> 0:17:53.240 point nine percent margins. So it just kind of feels 0:17:53.280 --> 0:17:55.000 like they're setting the price to whatever they want it 0:17:55.000 --> 0:17:55.199 to be. 0:17:55.800 --> 0:17:59.359 You don't think there, you don't think that is? Is 0:17:59.359 --> 0:18:00.840 that not an I think they need more? 0:18:00.920 --> 0:18:03.399 Don't They should get ninety five percent? Well, it's just. 0:18:03.440 --> 0:18:06.200 Kind of why are you setting the setting it so low? 0:18:06.240 --> 0:18:09.000 They should get one hundred and ten percent margins? We 0:18:09.000 --> 0:18:13.120 should No, It's just it's very frustrating as well, because 0:18:13.800 --> 0:18:16.920 I've read now arguments in favor and against Micron where 0:18:16.960 --> 0:18:20.960 it's like, oh, Micron got screwed by Apple and consumer 0:18:21.000 --> 0:18:25.680 electronics companies during the last during the last supply chain 0:18:25.800 --> 0:18:29.120 crisis because they they kind of made a bunch more 0:18:29.200 --> 0:18:33.280 ram than they needed to from the river. You know, like, no, 0:18:33.440 --> 0:18:37.639 it really it really is. But what and it seems 0:18:37.680 --> 0:18:39.919 to be there's this boom and bus cycle in memory 0:18:40.280 --> 0:18:43.560 where they get a bunch of demand and then so 0:18:43.680 --> 0:18:46.359 do they build more capacity? Is it that they just 0:18:47.400 --> 0:18:50.119 ordered too much RAM and then people cancel their orders 0:18:50.160 --> 0:18:52.040 like what happened? What happened last time? 0:18:54.880 --> 0:18:58.200 Yeah, so it's cyclical, I will say right now, as 0:18:58.359 --> 0:19:00.240 nothing like I've ever seen. And we suppoke to a 0:19:00.280 --> 0:19:03.480 bunch of companies purchasers of memory like g Skill and 0:19:03.520 --> 0:19:05.960 course Air, who sell the sticks of RAM that you 0:19:05.960 --> 0:19:08.640 put in a computer, and those guys are telling us 0:19:08.880 --> 0:19:12.639 on record that likewise they've never this is not like 0:19:12.680 --> 0:19:17.040 part of the cycle. This is special. And as for 0:19:17.600 --> 0:19:20.920 how is it different, how the difference is, how much 0:19:21.040 --> 0:19:24.040 the difference is the percentage increase it is this time? 0:19:24.160 --> 0:19:27.840 So you know, you're looking at like seven hundred percent 0:19:27.840 --> 0:19:30.200 increase in price in some situations in a very short 0:19:30.200 --> 0:19:35.600 period of time. Normally, cyclically you'll see some ebbs and flows, 0:19:35.640 --> 0:19:38.920 but it might be like back in the day, a 0:19:38.960 --> 0:19:41.840 fifty dollars kit of DDR four became eighty five or 0:19:41.920 --> 0:19:45.080 ninety dollars, maybe even one hundred and twenty if it's 0:19:45.119 --> 0:19:47.080 a really bad year, you know, but you're still at 0:19:47.440 --> 0:19:51.200 like a little worse than double, Whereas now a two 0:19:51.240 --> 0:19:53.679 hundred dollars kit of memory is suddenly a thousand to 0:19:53.680 --> 0:19:58.119 twelve hundred dollars. Uh, And so it's just the the 0:19:58.160 --> 0:20:01.200 amount of the price in is what's different. And then 0:20:01.240 --> 0:20:07.679 also this time, the sort of the contest for the 0:20:07.800 --> 0:20:15.119 supply is it's coming from within the own industry, where 0:20:15.840 --> 0:20:19.840 the very companies that would be the cause for an 0:20:19.840 --> 0:20:23.000 increase in sales of consumer supply are now the cause 0:20:23.119 --> 0:20:37.360 of these like insane unprecedented volume of sales for AI compontents. 0:20:40.680 --> 0:20:42.959 But it isn't based like so the twenty twenty one 0:20:43.119 --> 0:20:45.800 supply chain crisis chip crisis was because there was a 0:20:45.800 --> 0:20:48.719 flood of consumer spending off the COVID, Yeah, and that 0:20:48.800 --> 0:20:51.879 led to a bunch of consumer demand for products and 0:20:52.280 --> 0:20:55.600 overall spending across the board. But like consumer demand drove, 0:20:56.160 --> 0:20:59.919 this one seems different because it's not demand outside of 0:21:00.160 --> 0:21:04.920 the outside of whatever like psychosis, the various hype sets. 0:21:05.280 --> 0:21:09.879 They're not spending a trullion dollars in CAPEX really feel like, 0:21:09.880 --> 0:21:13.960 it's not like consumers can't access whatever micros COVID. 0:21:14.040 --> 0:21:16.920 I think you know what's interesting is the big three 0:21:16.960 --> 0:21:20.000 memory makers decided not I think it was twenty twenty 0:21:20.040 --> 0:21:24.320 two or three, to bring on more capacity in some 0:21:24.400 --> 0:21:26.919 of these situations because they were talking about how they 0:21:27.480 --> 0:21:30.760 got burned by oversupply and sort of fellow in COVID. 0:21:30.840 --> 0:21:31.000 Right. 0:21:31.320 --> 0:21:33.359 But what's interesting is if you go and look at 0:21:33.760 --> 0:21:36.240 some of the small suppliers's not many of them, but 0:21:36.920 --> 0:21:39.479 there's a couple of small memory makers that make them 0:21:39.520 --> 0:21:41.399 like less than five percent of the market, Wind Bond, 0:21:41.880 --> 0:21:45.240 Nanya or two of them. Nanya's in uh Well, headquartered 0:21:45.280 --> 0:21:47.639 at least in Taiwana of where their fab is, but 0:21:48.240 --> 0:21:52.560 they like these two companies as far as I'm aware. 0:21:52.560 --> 0:21:56.200 I was just looking at this this morning. They increased 0:21:57.119 --> 0:22:00.240 their their production capabilities at the time when the big 0:22:00.280 --> 0:22:04.280 three we're saying that they weren't gonna do it because 0:22:04.640 --> 0:22:08.639 you know, they got burned, right. It's interesting to me 0:22:08.680 --> 0:22:11.160 because if you're this is the argument the lawsuit makes too. 0:22:11.680 --> 0:22:15.159 If you're a small company, I would think you're at 0:22:15.160 --> 0:22:18.919 the biggest disadvantage, the biggest risk to increase your production 0:22:19.000 --> 0:22:23.160 capabilities with this massive amount of spend. If your biggest 0:22:23.400 --> 0:22:27.119 competitors are saying like this, it looks like bad times ahead. 0:22:27.160 --> 0:22:29.119 You know, we don't want to spend into this. But 0:22:29.160 --> 0:22:31.159 if you're a small guy, like you're at the greatest 0:22:31.240 --> 0:22:34.520 risk of going under. If you fuck up your your. 0:22:34.440 --> 0:22:36.680 Cap back right? Why didn't that this is maybe this 0:22:36.760 --> 0:22:40.320 is a dumb question. Why didn't them increasing supply allow 0:22:40.359 --> 0:22:42.000 them to kind of on the cut the microns of 0:22:42.000 --> 0:22:42.359 the world. 0:22:43.000 --> 0:22:47.880 Yeah, that one is just because the there's a micron 0:22:47.960 --> 0:22:52.040 does have a technology advantage, so they can make memory 0:22:52.040 --> 0:22:53.919 that these other guys can't. HBM is one of them. 0:22:53.920 --> 0:22:55.919 Maybe this is I didn't think about this till now, 0:22:55.960 --> 0:23:01.840 but maybe there is a very a small part, I think, 0:23:01.960 --> 0:23:05.600 but possibly part of an answer to your question of 0:23:05.640 --> 0:23:10.159 why HBM that is something that's that other companies are 0:23:10.160 --> 0:23:12.919 struggling to make successfully in current generations. 0:23:12.920 --> 0:23:15.760