1
00:00:02,720 --> 00:00:16,520
Speaker 1: Bloomberg Audio Studios, Podcasts, Radio News.

2
00:00:18,280 --> 00:00:21,799
Speaker 2: Hello and welcome to another episode of the Odd Lots Podcast.

3
00:00:21,880 --> 00:00:24,200
Speaker 3: I'm Joll Wisenthal and I'm Tracy Alloway.

4
00:00:24,400 --> 00:00:26,119
Speaker 2: Tercy, you know what I'm worried about. I don't know

5
00:00:26,160 --> 00:00:28,200
Speaker 2: if I'm worried about it, but it kind of feels

6
00:00:28,280 --> 00:00:30,280
Speaker 2: like the direction things are going, which is that like

7
00:00:30,880 --> 00:00:36,720
Speaker 2: all energy, industrial commodities, anything that we use for any

8
00:00:36,720 --> 00:00:40,720
Speaker 2: purpose is just going to like feed the AI beast

9
00:00:41,080 --> 00:00:43,760
Speaker 2: and us humans We're just gonna get left out in

10
00:00:43,800 --> 00:00:46,040
Speaker 2: the cold. They're like nothing for you. We gotta like feed,

11
00:00:46,120 --> 00:00:48,760
Speaker 2: We gotta feed it all to the AI. And and

12
00:00:48,840 --> 00:00:51,880
Speaker 2: maybe in ten or twenty or fifty years, the AI

13
00:00:52,040 --> 00:00:53,960
Speaker 2: is it's so powerful. They just decide, why why do

14
00:00:54,080 --> 00:00:56,440
Speaker 2: humans get any of this? We just thought we should

15
00:00:56,520 --> 00:00:57,480
Speaker 2: keep it all to ourselves.

16
00:00:57,760 --> 00:01:00,880
Speaker 3: A legitimate concern, I would say, I mean, to some extent,

17
00:01:00,920 --> 00:01:04,120
Speaker 3: we are already seeing this crowding out effect. Right, So

18
00:01:04,360 --> 00:01:07,640
Speaker 3: energy prices in certain areas have been going up, and

19
00:01:07,920 --> 00:01:12,040
Speaker 3: most recently memory chip prices. So this has been all

20
00:01:12,080 --> 00:01:16,000
Speaker 3: over earnings. Recently. You had companies like Apple saying that

21
00:01:16,160 --> 00:01:19,520
Speaker 3: because of a crunch in memory chip supply, they might

22
00:01:19,560 --> 00:01:21,839
Speaker 3: have to either raise their prices or like cut down

23
00:01:21,840 --> 00:01:25,440
Speaker 3: on the amount of phones they use. Nintendo, Like, if

24
00:01:25,440 --> 00:01:27,480
Speaker 3: you pull up a chart of Nintendo shares right now,

25
00:01:27,560 --> 00:01:30,920
Speaker 3: they are just getting hammered, and supposedly it is all

26
00:01:30,959 --> 00:01:35,080
Speaker 3: because of this memory chip shortage. By the way, do

27
00:01:35,160 --> 00:01:38,959
Speaker 3: you remember do you remember like buying your first PC

28
00:01:39,840 --> 00:01:42,200
Speaker 3: in like you bought one before I did, I think,

29
00:01:42,240 --> 00:01:44,360
Speaker 3: But mine was in the mid nineties, ye, I think,

30
00:01:44,880 --> 00:01:48,000
Speaker 3: I go. Do you remember how much memory those things had?

31
00:01:48,120 --> 00:01:48,280
Speaker 4: Oh?

32
00:01:48,320 --> 00:01:48,840
Speaker 2: Like nothing?

33
00:01:49,080 --> 00:01:52,200
Speaker 3: Yeah, Like I looked it up, something like less than

34
00:01:52,200 --> 00:01:54,440
Speaker 3: ten megabytes for a bunch of them. You know how

35
00:01:54,520 --> 00:01:55,639
Speaker 3: much PCs have nowadays?

36
00:01:55,680 --> 00:01:57,120
Speaker 2: Actually, I have no idea.

37
00:01:57,160 --> 00:02:01,360
Speaker 3: I think it's let's see sixteen gig by Yeah, isn't

38
00:02:01,400 --> 00:02:02,040
Speaker 3: that crazy?

39
00:02:02,240 --> 00:02:02,520
Speaker 5: It is?

40
00:02:02,680 --> 00:02:04,800
Speaker 2: It is really crazy. I remember, like, do you remember

41
00:02:04,880 --> 00:02:08,320
Speaker 2: zip drives? Yeah, oh my god, or like, and there

42
00:02:08,320 --> 00:02:12,359
Speaker 2: were all these special you know, obviously various memory peripherals

43
00:02:12,400 --> 00:02:14,200
Speaker 2: that you could buy and they came in sort of

44
00:02:14,200 --> 00:02:17,280
Speaker 2: dedicated cartridges and stuff like that. I think they're called

45
00:02:17,360 --> 00:02:19,680
Speaker 2: zip drive. Yeah, I think right, there was zip drive.

46
00:02:19,760 --> 00:02:22,160
Speaker 2: I don't remember. Yeah, yeah it was. Yeah, that's wrying,

47
00:02:22,200 --> 00:02:24,960
Speaker 2: additional memory storage, etcetera. It used to be this big

48
00:02:25,000 --> 00:02:27,880
Speaker 2: part of the consumer process. I'm actually glad. You know,

49
00:02:27,960 --> 00:02:30,480
Speaker 2: my son really wants a Nintendo. But if they can't

50
00:02:30,480 --> 00:02:33,200
Speaker 2: produce them anymore because they can't get the memory, then

51
00:02:33,240 --> 00:02:34,880
Speaker 2: I say, sorry, no video games.

52
00:02:34,960 --> 00:02:36,560
Speaker 3: That's so mean. You're gonna have to sit them down

53
00:02:36,600 --> 00:02:39,119
Speaker 3: and be like, I'm so sorry, son. There's this thing

54
00:02:39,200 --> 00:02:42,000
Speaker 3: called the supply shortage. I'm going to use this as

55
00:02:42,000 --> 00:02:45,760
Speaker 3: a learning opportunity to explain how global supply chains actually work.

56
00:02:45,960 --> 00:02:48,119
Speaker 2: Totally, you know what. The other the one interesting thing

57
00:02:48,320 --> 00:02:51,120
Speaker 2: is if you look at spot dram prices, which we

58
00:02:51,160 --> 00:02:53,480
Speaker 2: have on the terminal, this is I think maybe the

59
00:02:53,520 --> 00:02:56,800
Speaker 2: interesting dynamic. There's a lot of AI related things on

60
00:02:56,800 --> 00:02:59,520
Speaker 2: the terminal that have been surging for several years, most

61
00:02:59,560 --> 00:03:02,160
Speaker 2: notably some of the big chip companies and video et cetera.

62
00:03:02,360 --> 00:03:04,799
Speaker 2: If you look at a chart of Spot dram prices,

63
00:03:05,240 --> 00:03:07,720
Speaker 2: the big surge wasn't until late last year, and then

64
00:03:07,840 --> 00:03:10,240
Speaker 2: since then it's gone nuts. So for a long time,

65
00:03:10,280 --> 00:03:13,640
Speaker 2: even as this AI story, the AI meme, the AI

66
00:03:13,680 --> 00:03:16,720
Speaker 2: industry like propagated, people were really paying attention to that

67
00:03:17,280 --> 00:03:19,960
Speaker 2: area and then suddenly it's just gone like you know,

68
00:03:20,000 --> 00:03:24,320
Speaker 2: the haywire. Yeah, and so obviously huge supply demand and balances.

69
00:03:24,440 --> 00:03:26,440
Speaker 2: We need to understand what's going on because I think

70
00:03:26,480 --> 00:03:29,120
Speaker 2: the stakes are you know, it's getting so much. As

71
00:03:29,120 --> 00:03:31,520
Speaker 2: you mentioned, a lot of companies are losing out pretty big.

72
00:03:31,560 --> 00:03:33,520
Speaker 2: Others are making a lot of money. Yeah, we started

73
00:03:33,600 --> 00:03:35,480
Speaker 2: need to get a handle on what's going on with

74
00:03:35,480 --> 00:03:35,920
Speaker 2: the memory.

75
00:03:36,320 --> 00:03:36,840
Speaker 1: Let's do it.

76
00:03:36,960 --> 00:03:39,400
Speaker 2: I'm really excited to say we really do have the

77
00:03:39,400 --> 00:03:41,600
Speaker 2: perfect guest. We're going to be speaking with Ray Wong.

78
00:03:41,720 --> 00:03:45,320
Speaker 2: He's an analyst at Semi Analysis, and he's recently published

79
00:03:45,360 --> 00:03:49,600
Speaker 2: a brand new report on exactly this topic, talk about

80
00:03:49,600 --> 00:03:52,880
Speaker 2: a memory super cycle and so forth. So truly the

81
00:03:52,880 --> 00:03:56,839
Speaker 2: perfect guest. As I mentioned, he's comes to us from Korea. Ray,

82
00:03:57,000 --> 00:03:59,360
Speaker 2: thank you so much for coming on odlots.

83
00:03:59,680 --> 00:04:02,320
Speaker 4: Nice to me in Bove and thanksful having me so

84
00:04:02,360 --> 00:04:03,080
Speaker 4: many questions.

85
00:04:03,160 --> 00:04:06,520
Speaker 2: Really appreciate you joining us. Let's just start with the basics,

86
00:04:06,600 --> 00:04:09,960
Speaker 2: like what is it about AI or what is going on?

87
00:04:10,040 --> 00:04:13,560
Speaker 2: What is the core idea such that demand for memory

88
00:04:13,840 --> 00:04:17,440
Speaker 2: in various forms is absolutely seem to be booming right

89
00:04:17,480 --> 00:04:20,000
Speaker 2: now in a way that must have caught the supply

90
00:04:20,080 --> 00:04:21,440
Speaker 2: side of the equation very off guard.

91
00:04:21,839 --> 00:04:23,880
Speaker 4: Yeah, I think you know, to really do get kind

92
00:04:23,880 --> 00:04:26,680
Speaker 4: of the some of imbalanced supply TMO dynamics. Really we

93
00:04:26,760 --> 00:04:28,400
Speaker 4: wanted to get like you know what happened a few

94
00:04:28,480 --> 00:04:31,120
Speaker 4: years ago, right, because you know your capacity to check,

95
00:04:31,240 --> 00:04:34,480
Speaker 4: like your investment few years ago, and what's happening you know,

96
00:04:34,680 --> 00:04:37,280
Speaker 4: like in twenty twenty twenty four, it's it's sort of

97
00:04:37,320 --> 00:04:40,480
Speaker 4: a down cycle, right because remember doing the COVID, all

98
00:04:40,560 --> 00:04:43,000
Speaker 4: the companies so try to flug in to buy tons

99
00:04:43,040 --> 00:04:46,000
Speaker 4: of DRAM, right, and you know there's a lot of

100
00:04:46,320 --> 00:04:48,880
Speaker 4: purchases there, right, and there's sort of like upcycle, right,

101
00:04:48,880 --> 00:04:51,080
Speaker 4: but it's like a very very short up cycle because

102
00:04:51,120 --> 00:04:54,600
Speaker 4: like when the COVID gradually getting better, right, you actually

103
00:04:54,600 --> 00:04:57,839
Speaker 4: don't need that much DRAM And back then the company

104
00:04:57,839 --> 00:05:00,520
Speaker 4: will be like, hey, you actually don't have a sustainable right,

105
00:05:00,560 --> 00:05:04,200
Speaker 4: So they don't want to over invest to expand their capacity.

106
00:05:04,680 --> 00:05:07,279
Speaker 4: And when you go into the down cycle, your capet

107
00:05:07,360 --> 00:05:10,440
Speaker 4: typically be Americans trained Meri like, you know company from

108
00:05:10,440 --> 00:05:13,400
Speaker 4: their perspective, they try to be conservative. So that leads

109
00:05:13,440 --> 00:05:17,000
Speaker 4: to twenty twenty four and twenty five that you're incremental

110
00:05:17,000 --> 00:05:20,599
Speaker 4: waver for capacity that goes So DRAM is actually quite limited.

111
00:05:21,200 --> 00:05:23,719
Speaker 4: But what's happening on the demand and it's absolutely the

112
00:05:23,720 --> 00:05:26,159
Speaker 4: demand was asserrating so fast, right, you know, it's you know,

113
00:05:26,200 --> 00:05:27,599
Speaker 4: one way to look at it is to look at

114
00:05:27,600 --> 00:05:30,240
Speaker 4: a media's earnings, right, and if you when you when

115
00:05:30,279 --> 00:05:33,320
Speaker 4: you're on one side, your demand asserting so fast. On

116
00:05:33,360 --> 00:05:36,520
Speaker 4: the other hand, on the supply side, your supply just

117
00:05:36,560 --> 00:05:40,279
Speaker 4: couldn't catch up. And there's more nuances behind the supply,

118
00:05:40,440 --> 00:05:43,720
Speaker 4: right because before they are just supplying all the commodity

119
00:05:43,760 --> 00:05:47,920
Speaker 4: DRM like DDR LPDDR, right, the all the memory that

120
00:05:47,960 --> 00:05:51,640
Speaker 4: goes to PCs, goes to laptops, goes to your mobiles,

121
00:05:51,720 --> 00:05:55,200
Speaker 4: goes to Nintendo switches. Right. But you know what's happening,

122
00:05:55,560 --> 00:05:57,560
Speaker 4: you know, in this AI, it's like there's a new

123
00:05:57,600 --> 00:06:00,000
Speaker 4: thing called SBN. The think about. It's like a special

124
00:06:00,520 --> 00:06:04,920
Speaker 4: memory for directly for airs arter and that memory is

125
00:06:04,920 --> 00:06:08,680
Speaker 4: extremely drent wayfar intensive to give you a sense, right,

126
00:06:08,839 --> 00:06:11,280
Speaker 4: you know, on the same way, for a basis, you

127
00:06:11,320 --> 00:06:15,200
Speaker 4: can produce three more bits if you do commodity DRM,

128
00:06:15,279 --> 00:06:18,360
Speaker 4: but you can only produce one bit of HBM, right,

129
00:06:18,400 --> 00:06:21,120
Speaker 4: and that ratio will actually going higher when you go

130
00:06:21,200 --> 00:06:24,040
Speaker 4: to HPM four HBM five in the next few years. Right.

131
00:06:24,360 --> 00:06:26,560
Speaker 4: So in that way, on the same way, for basis

132
00:06:26,600 --> 00:06:29,159
Speaker 4: you can you can only produce that much HBM, and

133
00:06:29,240 --> 00:06:31,599
Speaker 4: right now what's happening is like, hey, this HPN is

134
00:06:31,640 --> 00:06:34,760
Speaker 4: super super profitable, and why not we that? Okay, more

135
00:06:34,800 --> 00:06:38,320
Speaker 4: wafare for HBM. But so here's the problem, right, you

136
00:06:38,440 --> 00:06:40,400
Speaker 4: only have that much wafer and you need to try

137
00:06:40,480 --> 00:06:43,919
Speaker 4: to do everything. So that's funding out a lot of

138
00:06:43,960 --> 00:06:46,680
Speaker 4: the supply going to the commodity dram. Right, so we

139
00:06:46,760 --> 00:06:48,919
Speaker 4: have we have sort of both things. That's happening and

140
00:06:49,080 --> 00:06:51,080
Speaker 4: essentially comes down to like I think, really, you know,

141
00:06:51,120 --> 00:06:53,359
Speaker 4: things start really kind of people are wearer. It's like

142
00:06:53,400 --> 00:06:56,120
Speaker 4: probably Q two T twenty five that people realize, hey,

143
00:06:56,200 --> 00:06:59,440
Speaker 4: like you know, we actually couldn't find enough like enough DURAN,

144
00:07:00,360 --> 00:07:04,040
Speaker 4: enough commodity d REM and also HPN for their demand.

145
00:07:04,560 --> 00:07:07,440
Speaker 3: So you mentioned the sea word just then oh wait,

146
00:07:07,680 --> 00:07:09,640
Speaker 3: not that seaword commodities.

147
00:07:10,240 --> 00:07:11,280
Speaker 2: You mentioned commodities.

148
00:07:11,320 --> 00:07:14,720
Speaker 3: And when I'm reading these articles about a memory chip shortage,

149
00:07:14,800 --> 00:07:19,040
Speaker 3: people use commodity type language. Right, So you hear super

150
00:07:19,120 --> 00:07:23,600
Speaker 3: cycle a lot, you hear commoditized chips versus I don't

151
00:07:23,640 --> 00:07:25,960
Speaker 3: know other types of chips. Are they customized? I don't

152
00:07:26,000 --> 00:07:29,920
Speaker 3: think so. But anyway, how much does this particular industry

153
00:07:29,920 --> 00:07:33,560
Speaker 3: actually resemble a commodities like industry, and then maybe explain

154
00:07:33,600 --> 00:07:37,160
Speaker 3: a little bit more why historically it has been really

155
00:07:37,200 --> 00:07:38,040
Speaker 3: really cyclical.

156
00:07:38,800 --> 00:07:40,240
Speaker 4: Yeah, so I think a couple of reasons.

157
00:07:40,320 --> 00:07:40,520
Speaker 3: Right.

158
00:07:40,880 --> 00:07:44,080
Speaker 4: You know, Fundamentally, the sort of the cost of bits

159
00:07:44,120 --> 00:07:48,520
Speaker 4: for DRM was actually coming down every single year significantly, right,

160
00:07:48,560 --> 00:07:50,720
Speaker 4: even you know in most recent years was also coming down.

161
00:07:50,760 --> 00:07:53,280
Speaker 4: But like the extent that they are coming down for

162
00:07:53,600 --> 00:07:56,000
Speaker 4: every single year is actually slowing down, like you know,

163
00:07:56,080 --> 00:07:58,640
Speaker 4: over the past like ten twenty years. Right, So the

164
00:07:58,680 --> 00:08:02,000
Speaker 4: cause was key going down, right, So because it's hard

165
00:08:02,040 --> 00:08:05,160
Speaker 4: to become of like you know, the most competitive part. Right.

166
00:08:05,320 --> 00:08:08,720
Speaker 4: Another thing is there's the nest of the committee that

167
00:08:08,800 --> 00:08:11,680
Speaker 4: let's set up the standard for d REM right. So, well,

168
00:08:11,720 --> 00:08:14,360
Speaker 4: you're doing the similar products, right, you're just having you know,

169
00:08:14,360 --> 00:08:16,520
Speaker 4: oh this is kind of a little better power efficiency,

170
00:08:17,160 --> 00:08:19,760
Speaker 4: the performance was a little better right here and there.

171
00:08:20,080 --> 00:08:22,320
Speaker 4: So like it's hard for you know, a memory supplier

172
00:08:22,400 --> 00:08:27,520
Speaker 4: to significantly differentiate their products compared to their competitor. Right,

173
00:08:27,560 --> 00:08:30,280
Speaker 4: So essentially, what's the what's the things that will compete

174
00:08:30,440 --> 00:08:32,640
Speaker 4: in M will be in the market prices? Right, And

175
00:08:32,679 --> 00:08:35,440
Speaker 4: then if you are selling the similar products and you

176
00:08:35,480 --> 00:08:38,760
Speaker 4: are talking in similar audience. Right, that's sort of like

177
00:08:39,160 --> 00:08:42,160
Speaker 4: other commodity you know, sort of semic adopt the commodity

178
00:08:42,200 --> 00:08:44,400
Speaker 4: products to me, and I think those are the two

179
00:08:44,520 --> 00:08:47,839
Speaker 4: man characteristics I'm saying for DRAM, and I think that's

180
00:08:47,840 --> 00:08:50,160
Speaker 4: a little a little different for HBM.

181
00:08:50,520 --> 00:08:54,560
Speaker 2: Tell me explain this. So there's commodity d RAM and

182
00:08:54,600 --> 00:08:58,320
Speaker 2: then there's HBM. What HBM is a type of d RAM?

183
00:08:58,400 --> 00:09:01,480
Speaker 2: Is that what you're saying? Yes, can you explain this

184
00:09:01,640 --> 00:09:04,120
Speaker 2: further and like sort of yeah, walk us through what

185
00:09:04,200 --> 00:09:06,720
Speaker 2: the commodity version and why there is this thing called

186
00:09:06,880 --> 00:09:09,079
Speaker 2: HBM High bandwidth memory what that's all about.

187
00:09:09,720 --> 00:09:13,439
Speaker 4: Yeah, Yeah, So essentially the emergence of HPN is really

188
00:09:13,520 --> 00:09:16,280
Speaker 4: go down to the kind of model layer, because you know,

189
00:09:16,440 --> 00:09:18,600
Speaker 4: this is when you like try to scale and stary

190
00:09:18,679 --> 00:09:23,920
Speaker 4: scaling your AM model. Right, your thoughts, your followsing increasing

191
00:09:23,920 --> 00:09:26,600
Speaker 4: significantly every single year, right, because your compute was increased

192
00:09:27,160 --> 00:09:30,280
Speaker 4: lat faster, but your memory bandwidth was actually sort of

193
00:09:30,320 --> 00:09:33,160
Speaker 4: it's increasing, but it's very limited, especially if you are

194
00:09:33,240 --> 00:09:36,640
Speaker 4: using the memory from kind of traditional commodity dan that

195
00:09:36,760 --> 00:09:39,760
Speaker 4: the bandwidth is actually mare limited. So from a memory supply,

196
00:09:39,800 --> 00:09:42,199
Speaker 4: they're thinking, hey, like how can we expand the bandwidth

197
00:09:42,480 --> 00:09:47,160
Speaker 4: to supporting the memory. They're going to scale the AI, right,

198
00:09:47,640 --> 00:09:50,320
Speaker 4: so let's consoles HBM, so the way they do hBN.

199
00:09:50,400 --> 00:09:55,040
Speaker 4: It's so they are stacking multiply eight were told for

200
00:09:55,160 --> 00:09:58,720
Speaker 4: even in the future sixteen DR and dies together, right,

201
00:09:59,040 --> 00:10:02,920
Speaker 4: So that making you you have like essentially having higher

202
00:10:03,040 --> 00:10:07,400
Speaker 4: bandwidth and have enough capacity to support AI developments over

203
00:10:07,440 --> 00:10:09,439
Speaker 4: the past few years and in the next few years.

204
00:10:09,559 --> 00:10:11,480
Speaker 4: And let's very different, right, you are stacking all this

205
00:10:11,640 --> 00:10:13,320
Speaker 4: dies together compared to a Hey, I would just we

206
00:10:13,440 --> 00:10:15,720
Speaker 4: just have like one DRM products as a chip.

207
00:10:15,800 --> 00:10:16,000
Speaker 3: Right.

208
00:10:16,240 --> 00:10:18,360
Speaker 4: And also that also comes down to the point that

209
00:10:18,880 --> 00:10:22,480
Speaker 4: HP and manufacturing also the packaging, both front end and

210
00:10:22,559 --> 00:10:25,679
Speaker 4: back in there's a lot more complex compared to commodity

211
00:10:25,720 --> 00:10:28,360
Speaker 4: d REM. And that's why I say commodity DRM is

212
00:10:28,400 --> 00:10:30,680
Speaker 4: more like commoditize, but I think HP is less so

213
00:10:30,840 --> 00:10:34,000
Speaker 4: because it's lumbo complex, right. And for supplier you can

214
00:10:34,040 --> 00:10:37,480
Speaker 4: actually differentiate either from the front end or back end technology.

215
00:10:37,800 --> 00:10:39,920
Speaker 4: And when you you know it's better, right, it gives

216
00:10:39,920 --> 00:10:43,080
Speaker 4: you a better margin, better ability to compare compete with

217
00:10:43,080 --> 00:10:46,240
Speaker 4: your competitor. Right, you know, stuff like thermal things like

218
00:10:46,280 --> 00:10:48,440
Speaker 4: that is the thing that will really stand out right,

219
00:10:48,440 --> 00:10:50,320
Speaker 4: and let's marry different commodity.

220
00:10:50,000 --> 00:10:52,760
Speaker 3: D rem getting more and more layers on HBM. It

221
00:10:52,840 --> 00:10:55,120
Speaker 3: kind of reminds me of the Razor Wars, where like,

222
00:10:55,400 --> 00:10:58,360
Speaker 3: you know, Gillett would unveil a three blade thing and

223
00:10:58,400 --> 00:11:01,120
Speaker 3: then the next day someone has like twelve or whatever.

224
00:11:16,840 --> 00:11:19,880
Speaker 3: So if I'm a company like Nintendo, how am I

225
00:11:20,040 --> 00:11:23,920
Speaker 3: actually purchasing chips? Because I think this is going to

226
00:11:23,960 --> 00:11:26,560
Speaker 3: be important when we talk about what's going on right now.

227
00:11:26,600 --> 00:11:30,160
Speaker 3: But do I have like a forward contract with a

228
00:11:30,200 --> 00:11:32,400
Speaker 3: major supplier, or do I have I don't know, a

229
00:11:32,440 --> 00:11:35,280
Speaker 3: warehouse full of chips that I bought previously. How do

230
00:11:35,360 --> 00:11:36,320
Speaker 3: they actually manage that?

231
00:11:37,280 --> 00:11:39,080
Speaker 4: So there's a couple of ways they do it, right.

232
00:11:39,200 --> 00:11:41,000
Speaker 4: They will they will do it in in terms of

233
00:11:41,080 --> 00:11:42,439
Speaker 4: like kind of they will do a different way like

234
00:11:42,600 --> 00:11:46,240
Speaker 4: different environments and also different vendors, right, So it's hard

235
00:11:46,240 --> 00:11:48,800
Speaker 4: to say like, hey, this is exactly exactly the way

236
00:11:48,840 --> 00:11:50,600
Speaker 4: they do it. But typically I think you can think

237
00:11:50,640 --> 00:11:53,199
Speaker 4: about like, hey, this is amount of the chips we're

238
00:11:53,200 --> 00:11:55,760
Speaker 4: gonna purchase in the next four months, and this is

239
00:11:55,760 --> 00:11:57,600
Speaker 4: a month we're going to commit, right, and we're gonna

240
00:11:57,600 --> 00:12:00,000
Speaker 4: sign a counter partly saying whether this term will be

241
00:12:00,200 --> 00:12:02,800
Speaker 4: four months or for the LTA, you can go to

242
00:12:02,840 --> 00:12:04,400
Speaker 4: like as long as like to a year or even

243
00:12:04,440 --> 00:12:06,600
Speaker 4: two years. Right, So let's see the people the way

244
00:12:06,640 --> 00:12:08,480
Speaker 4: you think about it. But people you will be like, hey,

245
00:12:08,480 --> 00:12:11,400
Speaker 4: we signed this country for four months, right and by

246
00:12:11,400 --> 00:12:13,360
Speaker 4: the end, sorry for three months for a quarter. So

247
00:12:13,400 --> 00:12:15,640
Speaker 4: by the end of the quarter, you negotiate a new

248
00:12:15,760 --> 00:12:18,640
Speaker 4: country pricing for the next quarter. So that's typically the

249
00:12:18,679 --> 00:12:21,640
Speaker 4: way it works. But you know the problem for right

250
00:12:21,679 --> 00:12:24,600
Speaker 4: now is for consumer and toronic, it's like the pricing

251
00:12:24,600 --> 00:12:27,559
Speaker 4: prob is sort of I would say secondary, because if

252
00:12:27,559 --> 00:12:30,079
Speaker 4: you couldn't secure value, you couldn't even make any money, right,

253
00:12:30,240 --> 00:12:33,400
Speaker 4: So number one is secure in value, whether that you know,

254
00:12:33,520 --> 00:12:36,199
Speaker 4: make sure that fits to the demand all look, you

255
00:12:36,240 --> 00:12:38,920
Speaker 4: are same two. It's whether you can negotia the best

256
00:12:38,920 --> 00:12:43,679
Speaker 4: pricing you can. But given the overall pricing environment was slow, volatile,

257
00:12:43,920 --> 00:12:46,680
Speaker 4: was so ascanting so much, it's hard to like not

258
00:12:46,840 --> 00:12:51,000
Speaker 4: taking some kinds of marching impacts because essentially, you know,

259
00:12:51,760 --> 00:12:53,800
Speaker 4: you try to get the chips right and you cannot

260
00:12:53,840 --> 00:12:56,679
Speaker 4: just you know, get too much away from your vendors.

261
00:12:57,240 --> 00:12:59,720
Speaker 2: Yeah, I want to get more into the demand destruction

262
00:13:00,200 --> 00:13:02,199
Speaker 2: because I think This is very important to think about

263
00:13:02,200 --> 00:13:04,920
Speaker 2: where the e globrium is going, just to clarify what

264
00:13:05,120 --> 00:13:09,400
Speaker 2: is it about AI specifically? Is it in the training,

265
00:13:09,640 --> 00:13:12,920
Speaker 2: is it in the inference? Where specifically in the process

266
00:13:12,960 --> 00:13:16,840
Speaker 2: of building out AI services does this voracious demand for

267
00:13:17,200 --> 00:13:18,200
Speaker 2: DRAM come from?

268
00:13:18,559 --> 00:13:21,120
Speaker 4: So honestly, is to be up, very honest, I think

269
00:13:21,160 --> 00:13:24,439
Speaker 4: is everywhere, Like, right, you know AI training you need

270
00:13:24,440 --> 00:13:27,000
Speaker 4: to council HPM. That's just everyone, no, right, you also

271
00:13:27,080 --> 00:13:29,480
Speaker 4: need a lot of the CPU d RM and that's

272
00:13:29,600 --> 00:13:31,520
Speaker 4: LP d d R and a DDR that going to

273
00:13:31,600 --> 00:13:34,079
Speaker 4: the server. You also need HPN, right, So let's on

274
00:13:34,120 --> 00:13:38,000
Speaker 4: a training part. In an inferencing you also need HPM, right,

275
00:13:38,080 --> 00:13:40,840
Speaker 4: and you also need the CPU DR and the ghost

276
00:13:41,000 --> 00:13:43,520
Speaker 4: loss vocal right, But you know the main one will

277
00:13:43,559 --> 00:13:46,240
Speaker 4: be the HPM and even the inferencing right right now

278
00:13:46,240 --> 00:13:49,240
Speaker 4: they are separating to pre fill and dec and you know,

279
00:13:49,320 --> 00:13:51,680
Speaker 4: I think eighty percent I think right now, like the

280
00:13:51,679 --> 00:13:53,720
Speaker 4: most important thing will be to code, right, and the

281
00:13:53,800 --> 00:13:57,119
Speaker 4: CO is super super like memory bound and memory intensive

282
00:13:57,679 --> 00:14:01,120
Speaker 4: and HPN and memory importance of their will just will

283
00:14:01,120 --> 00:14:04,160
Speaker 4: only increase to my understanding, especially given the long contest

284
00:14:04,160 --> 00:14:07,599
Speaker 4: window continue to expand and the inference usage. You know,

285
00:14:07,679 --> 00:14:10,679
Speaker 4: douctions continue to going up, right, So I think those

286
00:14:10,720 --> 00:14:12,840
Speaker 4: are typical way I will think about those, right. And

287
00:14:12,880 --> 00:14:14,120
Speaker 4: the last thing is I think a lot of people

288
00:14:14,200 --> 00:14:17,520
Speaker 4: talk about these dates is about argentic AI, right, agentic AI.

289
00:14:17,920 --> 00:14:20,360
Speaker 4: You want to power those? You need a lot of CPU.

290
00:14:20,520 --> 00:14:22,360
Speaker 4: You're not a lot of CPU based server, right, So

291
00:14:22,400 --> 00:14:25,280
Speaker 4: what's what's including in the CPU server? There's a lot

292
00:14:25,280 --> 00:14:27,680
Speaker 4: of d R in there, right, and a lot of

293
00:14:27,720 --> 00:14:29,680
Speaker 4: urgentic world law also mean you need a lot of

294
00:14:29,680 --> 00:14:32,160
Speaker 4: inference as well, right, And then go back to my

295
00:14:32,200 --> 00:14:35,080
Speaker 4: previous point, you need a lot of XP and d REN. Right,

296
00:14:35,120 --> 00:14:36,680
Speaker 4: So I would say like kind of d RN and

297
00:14:36,720 --> 00:14:38,680
Speaker 4: I get HP and d R and kind of everywhere,

298
00:14:38,720 --> 00:14:40,720
Speaker 4: and we have to talk about storage bout storage or

299
00:14:40,720 --> 00:14:43,960
Speaker 4: so everywhere. You know, in doing this process.

300
00:14:43,760 --> 00:14:47,360
Speaker 3: I have a basic question, but why does AI need

301
00:14:47,680 --> 00:14:50,920
Speaker 3: a lot of memory at all? So you just explained,

302
00:14:50,960 --> 00:14:54,400
Speaker 3: you know, it's important for inference, it's important for training whatever,

303
00:14:54,520 --> 00:14:57,880
Speaker 3: but like why Yeah, I think.

304
00:14:57,680 --> 00:14:59,680
Speaker 4: I think this this needs to like goes down to

305
00:14:59,720 --> 00:15:02,400
Speaker 4: the to the model later, right, because essentially when you

306
00:15:02,560 --> 00:15:05,160
Speaker 4: try to train, uh, let's say you want to train

307
00:15:05,240 --> 00:15:07,600
Speaker 4: the model, right, and then you need. You just need

308
00:15:07,640 --> 00:15:09,760
Speaker 4: the pure tons of data. You need to write tons

309
00:15:09,760 --> 00:15:11,680
Speaker 4: of data there to train your model. And then for

310
00:15:11,800 --> 00:15:14,760
Speaker 4: the influencing partline, you need one data after another to

311
00:15:14,840 --> 00:15:16,760
Speaker 4: kind of compute one after another, sort of the chain

312
00:15:16,840 --> 00:15:19,360
Speaker 4: of salts. Right, yeah, So to to do that process,

313
00:15:19,440 --> 00:15:21,600
Speaker 4: how can you keep the previous data on your process

314
00:15:21,680 --> 00:15:23,400
Speaker 4: to keep it in there and the process to another

315
00:15:23,480 --> 00:15:25,880
Speaker 4: data that's coming out after right, So that's why you

316
00:15:25,920 --> 00:15:28,280
Speaker 4: need not of memory and go back to my previous point. Right.

317
00:15:28,440 --> 00:15:32,160
Speaker 4: The colon context window basically means like when you're when

318
00:15:32,280 --> 00:15:34,960
Speaker 4: the things that your process is longer and longer, right, like,

319
00:15:35,000 --> 00:15:36,840
Speaker 4: how can you how can you have enough memory to

320
00:15:36,920 --> 00:15:39,280
Speaker 4: process those? Right? And you know, one for a good example,

321
00:15:39,360 --> 00:15:42,320
Speaker 4: if you use TGBT, you guys definitely remember right, you

322
00:15:42,320 --> 00:15:46,000
Speaker 4: know back togbts, which starts coming out people just like hey,

323
00:15:46,160 --> 00:15:46,960
Speaker 4: like what's the weather?

324
00:15:47,080 --> 00:15:47,240
Speaker 5: Right?

325
00:15:47,280 --> 00:15:51,480
Speaker 4: Like marriage poems people right now? Is or nastier people

326
00:15:51,480 --> 00:15:54,720
Speaker 4: are doing like hey, can you write a twenty page

327
00:15:54,760 --> 00:15:57,320
Speaker 4: of poor? How canser it can happen on people?

328
00:15:57,360 --> 00:15:57,440
Speaker 3: Right?

329
00:15:57,520 --> 00:16:00,000
Speaker 4: For example? Then will give twenty page of poor and

330
00:16:00,000 --> 00:16:02,120
Speaker 4: they were wait like five minutes, right, And so if

331
00:16:02,120 --> 00:16:04,520
Speaker 4: you think about the process like they do all the calculation,

332
00:16:04,680 --> 00:16:08,040
Speaker 4: all the researches, and also give you the output tokens

333
00:16:08,040 --> 00:16:12,720
Speaker 4: that the answer you get from your prompts, that's significant,

334
00:16:12,800 --> 00:16:15,400
Speaker 4: way longer than the proms that you are given, right,

335
00:16:15,400 --> 00:16:17,920
Speaker 4: And that's also a lot longer response that you are

336
00:16:17,960 --> 00:16:20,880
Speaker 4: getting when you call back to the GPD three. And

337
00:16:21,040 --> 00:16:24,320
Speaker 4: remember the usage we know back to GPD three was

338
00:16:24,440 --> 00:16:27,920
Speaker 4: it was the U kind of the monthly daily users,

339
00:16:28,200 --> 00:16:29,800
Speaker 4: but right now it was I think the user for

340
00:16:29,840 --> 00:16:32,440
Speaker 4: GGPT was eight hundred million, right, and we are not

341
00:16:32,600 --> 00:16:37,240
Speaker 4: included the users for the Germini, for clud for XAI. Right,

342
00:16:37,280 --> 00:16:39,720
Speaker 4: So you think you multiply those effects. I think the

343
00:16:39,760 --> 00:16:43,680
Speaker 4: memory they management significant and it's very clear to me.

344
00:16:43,920 --> 00:16:46,480
Speaker 2: Yeah, it's really interesting. I saw someone people were talking

345
00:16:46,520 --> 00:16:48,520
Speaker 2: about this, but it really strikes me in my own

346
00:16:48,680 --> 00:16:51,920
Speaker 2: usage of AI just basically how much more use it

347
00:16:52,040 --> 00:16:55,200
Speaker 2: as capabilities have grown. Like it was sort of cute

348
00:16:55,240 --> 00:16:57,480
Speaker 2: in late twenty twenty two. It's like, oh, get it

349
00:16:57,520 --> 00:16:59,720
Speaker 2: to write a poem about this, and it's like okay.

350
00:16:59,800 --> 00:17:02,200
Speaker 2: Then I would go, you know, several weeks sometimes without

351
00:17:02,440 --> 00:17:05,840
Speaker 2: having anything to query, and then you know, these days

352
00:17:05,880 --> 00:17:08,159
Speaker 2: like I'm using it all the time, like trying to

353
00:17:08,240 --> 00:17:11,000
Speaker 2: do things with code or scrape data, et cetera. So

354
00:17:11,800 --> 00:17:17,800
Speaker 2: overall token consumption has grown massively with increased capabilities. All right,

355
00:17:17,880 --> 00:17:24,199
Speaker 2: let's talk about the demand destruction element. Any commodity is

356
00:17:24,280 --> 00:17:25,639
Speaker 2: going to be the price is going to be a

357
00:17:25,680 --> 00:17:29,000
Speaker 2: function of supply and demand, and eventually prices get high

358
00:17:29,119 --> 00:17:31,320
Speaker 2: enough such that certain forms of demand just are no

359
00:17:31,359 --> 00:17:33,639
Speaker 2: longer economical. And it's like, you know what, We're just

360
00:17:33,680 --> 00:17:36,440
Speaker 2: not going to make as many video games or cameras

361
00:17:36,560 --> 00:17:40,240
Speaker 2: or whatever else uses a memory. Are we starting to

362
00:17:40,280 --> 00:17:44,320
Speaker 2: see any of that yet such that certain uses and

363
00:17:44,520 --> 00:17:45,679
Speaker 2: consumption of memories.

364
00:17:45,920 --> 00:17:48,320
Speaker 3: Joe doesn't have to buy switch for his son, Yeah,

365
00:17:48,320 --> 00:17:48,720
Speaker 3: that's right.

366
00:17:50,440 --> 00:17:53,240
Speaker 4: Yeah, I actually both have switched RAC So the price

367
00:17:53,359 --> 00:17:54,159
Speaker 4: you should resell it?

368
00:17:54,480 --> 00:17:55,399
Speaker 2: Are you going to resell it?

369
00:17:57,160 --> 00:17:59,240
Speaker 4: I don't know. We'll see it up hol the opts

370
00:17:59,280 --> 00:18:02,080
Speaker 4: rocks there works. But yeah, I think I think we're

371
00:18:02,119 --> 00:18:04,240
Speaker 4: already seeing a lot of impacts. Actually, but kind of

372
00:18:04,480 --> 00:18:06,720
Speaker 4: you kind of vary in different types of products, right

373
00:18:06,800 --> 00:18:10,320
Speaker 4: right in p you are showing seeing that thell non

374
00:18:10,320 --> 00:18:13,360
Speaker 4: noble answer sous. They are all having a price tip

375
00:18:13,440 --> 00:18:16,840
Speaker 4: for kind of different their problems, right, and you know,

376
00:18:16,960 --> 00:18:19,240
Speaker 4: on the kind of demand side. You're also seeing that.

377
00:18:19,640 --> 00:18:22,560
Speaker 4: I think for the Chinese small phone market. Right, a

378
00:18:22,600 --> 00:18:24,720
Speaker 4: lot of the analysts, a lot of the research firm, right,

379
00:18:24,800 --> 00:18:26,360
Speaker 4: a lot of the company was saying, like they are

380
00:18:26,359 --> 00:18:29,600
Speaker 4: cutting off their small phone outook a full example. I

381
00:18:29,600 --> 00:18:31,880
Speaker 4: think media attacked recently left for Arniest. Right, they're saying

382
00:18:31,880 --> 00:18:34,919
Speaker 4: they're cutting off the mobile to I think it was

383
00:18:34,960 --> 00:18:37,880
Speaker 4: twelve for ten to fifteen percent of the twenty twenty

384
00:18:37,920 --> 00:18:39,960
Speaker 4: six outlook. That's mere significant.

385
00:18:40,040 --> 00:18:40,159
Speaker 2: Right.

386
00:18:40,200 --> 00:18:43,200
Speaker 4: You're saying this year is supposed to be a hundred million,

387
00:18:43,640 --> 00:18:46,720
Speaker 4: but you will essentially go to a nineteen million. Right,

388
00:18:46,840 --> 00:18:49,520
Speaker 4: that's mere significant. So I think we are already seeing that,

389
00:18:49,560 --> 00:18:52,000
Speaker 4: but you kind of vary different places. Right. Apple will

390
00:18:52,040 --> 00:18:54,920
Speaker 4: be a great example here because they're saying, hey, like, hey,

391
00:18:54,960 --> 00:18:56,359
Speaker 4: we have all the price TI like, right, but we

392
00:18:56,440 --> 00:18:58,960
Speaker 4: actually managed to put it well. Uh, they are saying

393
00:18:59,000 --> 00:19:01,800
Speaker 4: the real sort of for marketing impact will actually show

394
00:19:01,920 --> 00:19:04,240
Speaker 4: up in the second half twenty twenty six instead of

395
00:19:04,280 --> 00:19:07,120
Speaker 4: like the first quarter or second quarter. I think those

396
00:19:07,119 --> 00:19:09,119
Speaker 4: of at of the things I'm seeing, So it's definitely

397
00:19:09,200 --> 00:19:11,320
Speaker 4: already showing out in the market. Right. I think what

398
00:19:11,440 --> 00:19:14,640
Speaker 4: you can expect this sphere will be two things one

399
00:19:14,760 --> 00:19:17,879
Speaker 4: is sort of demand destruction in PCs right, the border

400
00:19:17,880 --> 00:19:22,280
Speaker 4: consumer and astronics or the dispect whether it's the memory

401
00:19:22,280 --> 00:19:25,560
Speaker 4: displce Le's merri straightforward. But also the camera right, the

402
00:19:25,640 --> 00:19:27,760
Speaker 4: camera is also very important. I'm hearing in some of

403
00:19:27,800 --> 00:19:30,480
Speaker 4: the display on the camera as well. And you know

404
00:19:30,560 --> 00:19:33,840
Speaker 4: also on the sort of what we call it delayed launch,

405
00:19:33,920 --> 00:19:36,240
Speaker 4: right because you don't want to launch your new products

406
00:19:36,240 --> 00:19:40,320
Speaker 4: that have higher price than initially stalled, which will impact

407
00:19:40,359 --> 00:19:43,640
Speaker 4: your new launch performance. Who's usually the new new lunch

408
00:19:43,640 --> 00:19:46,240
Speaker 4: product have better asp, better margin compared to your previous

409
00:19:46,320 --> 00:19:47,080
Speaker 4: generation products.

410
00:19:47,160 --> 00:19:47,320
Speaker 3: Right.

411
00:19:47,359 --> 00:19:49,359
Speaker 4: I think those other things we're going to say in

412
00:19:49,400 --> 00:19:52,480
Speaker 4: the coming months, in the coming quarters, I think will

413
00:19:52,520 --> 00:19:54,359
Speaker 4: show impacts and people will start feeling it.

414
00:19:54,960 --> 00:19:56,760
Speaker 3: Joe, do you think we're going to get people like

415
00:19:57,080 --> 00:20:01,680
Speaker 3: stripping all newtendos for memory hips? Like remember how during

416
00:20:01,680 --> 00:20:05,200
Speaker 3: the commodity super cycle in like the early twenty tens.

417
00:20:05,280 --> 00:20:07,439
Speaker 3: I guess there are all those stories about people like

418
00:20:08,000 --> 00:20:10,840
Speaker 3: stealing electricity wires going to get.

419
00:20:10,680 --> 00:20:13,119
Speaker 2: Twenty twenty one. Yeah, I think, yeah, we might. We

420
00:20:13,160 --> 00:20:13,840
Speaker 2: might get some of that.

421
00:20:14,040 --> 00:20:14,200
Speaker 1: Well.

422
00:20:14,240 --> 00:20:17,520
Speaker 3: Actually, on that note, is there like a short term

423
00:20:17,680 --> 00:20:21,000
Speaker 3: fix for this problem either in terms of design, maybe

424
00:20:21,000 --> 00:20:23,520
Speaker 3: you can make the products more efficient in some ways

425
00:20:23,520 --> 00:20:28,720
Speaker 3: so that it needs less memory or recycling old memory chips.

426
00:20:28,720 --> 00:20:31,040
Speaker 4: I don't know, yeah, I think you know from the

427
00:20:31,080 --> 00:20:34,000
Speaker 4: dement side, it's a lot difficult, right from an in perspective, right,

428
00:20:34,400 --> 00:20:37,280
Speaker 4: there's only that much you can do, right. You can dispect, right,

429
00:20:37,280 --> 00:20:39,399
Speaker 4: but I don't think that addresses to the fundamental issues. Right,

430
00:20:39,440 --> 00:20:42,160
Speaker 4: you are just dispect your products and so you can ship.

431
00:20:42,800 --> 00:20:45,720
Speaker 4: But that there are this double age word right when

432
00:20:45,760 --> 00:20:49,439
Speaker 4: you downgade your products that actually look back compared to

433
00:20:49,480 --> 00:20:52,600
Speaker 4: some of your peers who doesn't dispect right, and Mengro

434
00:20:52,640 --> 00:20:55,879
Speaker 4: product less competitive that impact your final sort of the

435
00:20:56,680 --> 00:20:59,639
Speaker 4: totally or annual shipment. I think what's need to be

436
00:20:59,640 --> 00:21:01,960
Speaker 4: address where what can be done is probably on the

437
00:21:01,960 --> 00:21:05,000
Speaker 4: supply side, right, So supply side we mainly talk about

438
00:21:05,000 --> 00:21:07,480
Speaker 4: a couple of vendors right, the micro and high in

439
00:21:07,520 --> 00:21:10,280
Speaker 4: eight Samson and for legacy ones. Right you have windbound

440
00:21:10,800 --> 00:21:14,440
Speaker 4: and China like CSMT and JHIICC right most people, So

441
00:21:14,560 --> 00:21:16,760
Speaker 4: I think the most right way right now this year

442
00:21:16,840 --> 00:21:20,520
Speaker 4: because this year the really challenges there's a clean room constraint.

443
00:21:20,840 --> 00:21:23,760
Speaker 4: So clean room country is basically you have limited space

444
00:21:24,040 --> 00:21:26,800
Speaker 4: that you can put those increments and start marifaction chips

445
00:21:26,880 --> 00:21:30,120
Speaker 4: in the vat right, so you have clean room shortage

446
00:21:30,640 --> 00:21:34,800
Speaker 4: in all three major memory makers. Right, so your incremental

447
00:21:34,840 --> 00:21:38,240
Speaker 4: waver capacity coming online for twenty twenty six will actually

448
00:21:38,280 --> 00:21:41,480
Speaker 4: be quite limited. So in that sense, how can you

449
00:21:41,520 --> 00:21:45,879
Speaker 4: produce more bits, more different bits while having only limited

450
00:21:45,920 --> 00:21:48,760
Speaker 4: incremental waver capacity. The only thing you can do it

451
00:21:48,880 --> 00:21:52,920
Speaker 4: is non migration. So in deer processes, now you have

452
00:21:53,080 --> 00:21:55,720
Speaker 4: the most advances right now it's one C right, going down,

453
00:21:55,800 --> 00:21:58,560
Speaker 4: have one B, one A and one X ball. But

454
00:21:58,840 --> 00:22:01,679
Speaker 4: right know, I think why now they are trying to

455
00:22:01,680 --> 00:22:04,720
Speaker 4: do its migra as much as they can to one

456
00:22:04,760 --> 00:22:07,159
Speaker 4: B and one C, which on the same way for

457
00:22:07,280 --> 00:22:10,120
Speaker 4: basis you have produced a lot more chip, a lot

458
00:22:10,160 --> 00:22:13,239
Speaker 4: more bits of compared to the legacy not right, so

459
00:22:13,320 --> 00:22:15,200
Speaker 4: do doing that way on the same wave for pass

460
00:22:15,240 --> 00:22:18,520
Speaker 4: you can produce more bits. So that should produce more

461
00:22:19,760 --> 00:22:23,920
Speaker 4: supply even though your wave capacities constrained. But the challenge

462
00:22:23,960 --> 00:22:26,359
Speaker 4: is two things. One how fast you can m about

463
00:22:26,359 --> 00:22:30,240
Speaker 4: this non migration? Right? No migration is typically difficult. Well

464
00:22:30,240 --> 00:22:32,720
Speaker 4: I always say difficult is you know it takes time

465
00:22:32,920 --> 00:22:36,320
Speaker 4: right for to go into the most events now, right,

466
00:22:36,359 --> 00:22:38,760
Speaker 4: you need EUV machine, right, you need all different kinds

467
00:22:38,760 --> 00:22:42,000
Speaker 4: of manufaction processes, right, and not every fact can already

468
00:22:42,040 --> 00:22:44,320
Speaker 4: prepare to doing that, right, so it will take time

469
00:22:44,400 --> 00:22:47,040
Speaker 4: to m about all those productions for the new events

470
00:22:47,080 --> 00:22:51,320
Speaker 4: not Also you are having the challenge to like, hey, okay,

471
00:22:51,320 --> 00:22:52,840
Speaker 4: I want to do that migration, right, But how we're

472
00:22:52,840 --> 00:22:56,320
Speaker 4: going to balance the capacity with HPN Again, that's another issue.

473
00:22:56,320 --> 00:22:58,280
Speaker 4: You want to kind of think about it, right, because

474
00:22:58,320 --> 00:23:01,359
Speaker 4: some of the HPN processed not uh it's actually using

475
00:23:01,600 --> 00:23:03,760
Speaker 4: for hidings, are using one BEE so that's actually the

476
00:23:03,760 --> 00:23:06,280
Speaker 4: events processing all right, So even you're including one bee

477
00:23:06,400 --> 00:23:08,679
Speaker 4: by Edison, I want to do more ESPN, So like

478
00:23:08,760 --> 00:23:11,639
Speaker 4: that increase years actually quite limited. So I think also

479
00:23:11,920 --> 00:23:14,800
Speaker 4: the way it can potentially fix the issue. But you know,

480
00:23:15,040 --> 00:23:18,760
Speaker 4: from our house view, right, even we factor all those in,

481
00:23:19,040 --> 00:23:20,480
Speaker 4: I think this year we're still going to have ce

482
00:23:20,640 --> 00:23:21,760
Speaker 4: press significant shortage.

483
00:23:22,240 --> 00:23:23,880
Speaker 2: One of the themes that comes up when we talk

484
00:23:23,920 --> 00:23:28,640
Speaker 2: about any commodities supercycle is that sure, higher prices will

485
00:23:28,680 --> 00:23:32,840
Speaker 2: eventually elicit more production more supply, but in the meantime,

486
00:23:33,119 --> 00:23:36,080
Speaker 2: you know, there's a sweet spot for the existing producers,

487
00:23:36,359 --> 00:23:39,040
Speaker 2: especially when you have just a small number of producers

488
00:23:39,040 --> 00:23:42,360
Speaker 2: that can produce at scale, they enjoy massive profits, right,

489
00:23:42,400 --> 00:23:45,040
Speaker 2: and they're sort of going to be reluctant to build

490
00:23:45,040 --> 00:23:47,880
Speaker 2: out new fabs because that's money going out the door.

491
00:23:47,920 --> 00:23:50,520
Speaker 2: It also means lower prices, So there's always this sort

492
00:23:50,520 --> 00:23:54,520
Speaker 2: of goldilocks period for them before the supply comes online

493
00:23:54,520 --> 00:23:57,480
Speaker 2: where there's just raking in profits. What do we see

494
00:23:57,840 --> 00:24:02,040
Speaker 2: is the industries in paul to invest And I'm curious,

495
00:24:02,040 --> 00:24:04,600
Speaker 2: like you know, whether it's Chinese producers. Are they in

496
00:24:04,720 --> 00:24:08,480
Speaker 2: a position to potentially undercut the profits of the Korean makers?

497
00:24:08,760 --> 00:24:11,720
Speaker 2: Are the Korean makers seeing themselves as having a long

498
00:24:11,800 --> 00:24:14,960
Speaker 2: window where they can enjoy large profits before they have

499
00:24:15,040 --> 00:24:17,439
Speaker 2: to invest? Like what is the thinking on the supply

500
00:24:17,560 --> 00:24:19,760
Speaker 2: side about those big capital outlets?

501
00:24:20,200 --> 00:24:22,560
Speaker 4: Yeah, and that's a very good like strategic question. And

502
00:24:22,640 --> 00:24:24,560
Speaker 4: I think you know, every management or public give you

503
00:24:24,680 --> 00:24:26,240
Speaker 4: so no different answer or and kind of on the

504
00:24:26,280 --> 00:24:28,960
Speaker 4: high level, probably same by the in general, right, Like

505
00:24:29,040 --> 00:24:31,560
Speaker 4: you know, you know, look at like historical cycle. It's like,

506
00:24:32,400 --> 00:24:34,920
Speaker 4: you know, if we are seeing quite sustainable demand in

507
00:24:34,960 --> 00:24:37,919
Speaker 4: the next few years and then clearly the capacity couldn't

508
00:24:37,960 --> 00:24:40,119
Speaker 4: catch on the demand, they are going to expand the

509
00:24:40,160 --> 00:24:42,680
Speaker 4: capacity in some time. Right, It's probably not going to

510
00:24:42,760 --> 00:24:45,359
Speaker 4: be like hey were announcing twenty twenty six the topacit

511
00:24:45,359 --> 00:24:47,560
Speaker 4: they do it currently online and meaningfully it's probably twenty

512
00:24:47,560 --> 00:24:50,479
Speaker 4: twenty eight, so they will I think they will still announce, right,

513
00:24:50,520 --> 00:24:52,960
Speaker 4: and we are seeing that from micro Right. Micron is

514
00:24:52,960 --> 00:24:55,119
Speaker 4: announcing a new non fab in Singapore.

515
00:24:55,280 --> 00:24:55,440
Speaker 3: Right.

516
00:24:55,480 --> 00:24:58,120
Speaker 4: They are expanding they and they have two fabs currently

517
00:24:58,200 --> 00:25:01,400
Speaker 4: it's on the construction in you right, also doing tons

518
00:25:01,440 --> 00:25:04,640
Speaker 4: of a non migration and recently acquire the new fat

519
00:25:04,680 --> 00:25:07,800
Speaker 4: from PSMC, which is a Taiwanese stream maker, and then

520
00:25:07,880 --> 00:25:10,000
Speaker 4: all those movie You are already seeing some side of

521
00:25:10,040 --> 00:25:13,120
Speaker 4: that and another side you are saying. It's the Capex, Right,

522
00:25:13,600 --> 00:25:17,480
Speaker 4: both the Heene's, Samsung and Microne there during Capex is

523
00:25:17,520 --> 00:25:21,320
Speaker 4: actually increased quite significant this year and expects similar train

524
00:25:21,760 --> 00:25:24,119
Speaker 4: in twenty twenty and seven. Given I think you know,

525
00:25:24,160 --> 00:25:27,760
Speaker 4: they're gonna probably try to expand more capacity and invests

526
00:25:27,800 --> 00:25:30,359
Speaker 4: more in both HPN as well as the equipments that

527
00:25:30,400 --> 00:25:32,800
Speaker 4: go to events equipment. Right, that goes back to the

528
00:25:32,880 --> 00:25:34,920
Speaker 4: point I was talking about the non migration. If you

529
00:25:34,960 --> 00:25:37,240
Speaker 4: want non migration, you need to use more advanced equipment

530
00:25:37,480 --> 00:25:39,359
Speaker 4: to produce more advanced chips.

531
00:25:39,800 --> 00:25:42,600
Speaker 3: Yeah, it really does remind me of the oil industry, right,

532
00:25:42,600 --> 00:25:45,680
Speaker 3: where like there's a bust in the underlying commodity price

533
00:25:45,760 --> 00:25:50,560
Speaker 3: and everyone starts talking about being disciplined, punkapex bend and

534
00:25:50,600 --> 00:25:53,960
Speaker 3: so it takes a while to ramp up. I wanted

535
00:25:54,000 --> 00:25:57,320
Speaker 3: to ask about, I guess the balance between HBM and

536
00:25:57,440 --> 00:26:00,640
Speaker 3: other types of dram How are people or how are

537
00:26:00,720 --> 00:26:04,000
Speaker 3: the actual chip makers thinking about that? And is there

538
00:26:04,359 --> 00:26:08,879
Speaker 3: is there a chance that because HBM my understanding is

539
00:26:08,920 --> 00:26:12,879
Speaker 3: that it has higher profit margins, is there a chance

540
00:26:12,920 --> 00:26:16,679
Speaker 3: that everyone just pours into HBM and kind of leaves

541
00:26:16,760 --> 00:26:19,720
Speaker 3: the more basic stuff in the dust.

542
00:26:20,080 --> 00:26:22,680
Speaker 4: Yeah. So, actually, you know, it's very interesting. I think

543
00:26:22,680 --> 00:26:25,560
Speaker 4: what probably things folk when you ruled about this for

544
00:26:25,760 --> 00:26:28,919
Speaker 4: our institutional client, it's that, you know, not the margin. Well,

545
00:26:28,920 --> 00:26:31,359
Speaker 4: typically we think hBN has higher margin, right, which is

546
00:26:31,440 --> 00:26:35,199
Speaker 4: which is definitely true. But what's happening you know, we

547
00:26:35,240 --> 00:26:37,440
Speaker 4: see all this like spot prize country price going out

548
00:26:37,480 --> 00:26:39,919
Speaker 4: so much, right, the margin of the commodity d RN

549
00:26:40,000 --> 00:26:43,880
Speaker 4: right now is actually higher than hBN. So here that

550
00:26:43,920 --> 00:26:47,240
Speaker 4: created a real dynamo that tracy that you talk about

551
00:26:47,240 --> 00:26:50,840
Speaker 4: it because when your marginal command is actually going higher,

552
00:26:51,000 --> 00:26:52,960
Speaker 4: why would you make more hBN? Like why?

553
00:26:53,119 --> 00:26:53,280
Speaker 3: Right?

554
00:26:53,640 --> 00:26:55,760
Speaker 4: But I think if you really think about like a

555
00:26:55,800 --> 00:26:58,840
Speaker 4: long term perspective, form a memory supplier right before the

556
00:26:58,920 --> 00:27:02,960
Speaker 4: old air bloom, your demand is really driving from couple

557
00:27:03,000 --> 00:27:07,560
Speaker 4: a market right, PC, mobile, automotive, industrial, blah blah blah. Right,

558
00:27:08,200 --> 00:27:11,840
Speaker 4: hBN right now is search as a new growth driver

559
00:27:12,160 --> 00:27:15,879
Speaker 4: for the company, which we believe will last for the

560
00:27:15,880 --> 00:27:16,840
Speaker 4: next few years.

561
00:27:16,920 --> 00:27:17,080
Speaker 3: Right.

562
00:27:17,480 --> 00:27:19,720
Speaker 4: And for the memory supported I'm thinking the same. And

563
00:27:19,840 --> 00:27:22,680
Speaker 4: if you are now continue to push your capacity, push

564
00:27:22,680 --> 00:27:26,200
Speaker 4: your technology. When you're lacking behind, it's hard to coming back.

565
00:27:26,560 --> 00:27:28,760
Speaker 4: And they go back to my previous point that this

566
00:27:28,800 --> 00:27:32,840
Speaker 4: is an area that's relatively easier to differentiate your products

567
00:27:32,920 --> 00:27:35,800
Speaker 4: to get more market share more meaningfully. Right, So I

568
00:27:35,800 --> 00:27:39,520
Speaker 4: think three memory makers are mermer value HPM even though

569
00:27:39,600 --> 00:27:41,959
Speaker 4: right now on the margin of commodity r is higher. Right,

570
00:27:42,000 --> 00:27:44,040
Speaker 4: they will, I think they will still invest quite a

571
00:27:44,080 --> 00:27:46,560
Speaker 4: lot on HPM, uh, and they will do their best

572
00:27:46,600 --> 00:27:49,760
Speaker 4: to sort of balance the commodity dream market. But I think,

573
00:27:50,440 --> 00:27:52,000
Speaker 4: you know, I think this year, at least the number

574
00:27:52,080 --> 00:27:54,400
Speaker 4: I'm saying, it's still quiet significant.

575
00:27:54,440 --> 00:27:59,919
Speaker 2: Short, how do the Chinese producers compare to the Korean producers.

576
00:28:00,080 --> 00:28:01,840
Speaker 2: Is there a Is there a gap?

577
00:28:02,720 --> 00:28:04,640
Speaker 4: Oh yeah, I think there's definitely the gap. I think,

578
00:28:04,720 --> 00:28:06,800
Speaker 4: like you know, if you look at the memory in general,

579
00:28:07,240 --> 00:28:11,280
Speaker 4: I think it's probably three years, say yeah, it's probably

580
00:28:11,359 --> 00:28:14,320
Speaker 4: four years. I will say that this is merely like

581
00:28:14,359 --> 00:28:16,760
Speaker 4: you know, time list that kind of today. You know,

582
00:28:16,760 --> 00:28:18,600
Speaker 4: who knows about happen in the future. But I think,

583
00:28:18,640 --> 00:28:20,680
Speaker 4: you know, there's definitely gap. You know, when I think

584
00:28:20,680 --> 00:28:24,200
Speaker 4: about the competition pressure from Chinese memory makers, we always

585
00:28:24,240 --> 00:28:27,280
Speaker 4: want to kind of separate the Chinese market and also

586
00:28:27,320 --> 00:28:29,639
Speaker 4: the x China market. Let's also the memory supply was

587
00:28:29,640 --> 00:28:32,320
Speaker 4: looking at it, right, So Chinese market is probably about

588
00:28:32,359 --> 00:28:35,160
Speaker 4: twenty five percent of the global deuring dement. So really

589
00:28:35,320 --> 00:28:38,920
Speaker 4: the competition is happening in China because for the leading

590
00:28:39,000 --> 00:28:42,480
Speaker 4: Chinese memory supplier like CSNT for example, I believe like

591
00:28:43,440 --> 00:28:46,240
Speaker 4: or ninety five percent of revenue is coming from China

592
00:28:46,320 --> 00:28:48,360
Speaker 4: and Hong Kong. So meaning you know, most of the

593
00:28:48,440 --> 00:28:52,240
Speaker 4: DRA is really compaying Chinese market, competing with micro Hinings

594
00:28:52,280 --> 00:28:54,360
Speaker 4: and Samsung. I think I won't say that right now.

595
00:28:54,400 --> 00:28:56,520
Speaker 4: On the events on the high end, right, I think

596
00:28:56,520 --> 00:28:59,720
Speaker 4: it was still dominated by three memory makers, but CSMD

597
00:29:00,080 --> 00:29:02,360
Speaker 4: getting to momento and why, and they are probably getting

598
00:29:02,600 --> 00:29:05,520
Speaker 4: sort of law and mem products. And they are also

599
00:29:05,520 --> 00:29:08,680
Speaker 4: benefited by Chinese Commons policy in terms of the self

600
00:29:08,720 --> 00:29:12,640
Speaker 4: sufficiency drive to you know, whether it's for the commodity there.

601
00:29:12,680 --> 00:29:15,480
Speaker 4: And also right now they're pushing for hBN developments that

602
00:29:15,600 --> 00:29:17,160
Speaker 4: support Chinese AI harware.

603
00:29:17,920 --> 00:29:19,960
Speaker 3: So if there's a shortage, and it seems like the

604
00:29:19,960 --> 00:29:21,960
Speaker 3: shortage is going to be with us for some time,

605
00:29:22,440 --> 00:29:26,880
Speaker 3: how are the chip makers actually allocating what supply they have.

606
00:29:27,120 --> 00:29:29,719
Speaker 3: Does it go to, you know, the company that I

607
00:29:29,760 --> 00:29:32,360
Speaker 3: think is going to be really big and important in

608
00:29:32,400 --> 00:29:34,680
Speaker 3: the future, like I don't know, an Open AI or something,

609
00:29:34,760 --> 00:29:38,000
Speaker 3: or does it go to an existing customer that's been

610
00:29:38,160 --> 00:29:40,840
Speaker 3: buying in volume from me for years?

611
00:29:42,160 --> 00:29:44,280
Speaker 4: Yeah, I think there's no doubt to me that you know,

612
00:29:44,480 --> 00:29:48,280
Speaker 4: sounds the highest highest tier customer will get a value

613
00:29:48,320 --> 00:29:51,760
Speaker 4: for sure. Of course there are the more complex pricing

614
00:29:51,800 --> 00:29:54,360
Speaker 4: negotiation behind. But I will I will say, like the

615
00:29:54,440 --> 00:29:58,719
Speaker 4: value would be still be the most important thing, and uh,

616
00:29:58,800 --> 00:30:00,760
Speaker 4: I think that's probably what they gonna do, right And

617
00:30:00,800 --> 00:30:03,720
Speaker 4: you know, besides thinking about like a buy customer, I

618
00:30:03,720 --> 00:30:06,440
Speaker 4: will think about kind of bisector. I will say the

619
00:30:06,520 --> 00:30:10,720
Speaker 4: server durent and HPN two things together will be the top,

620
00:30:10,800 --> 00:30:13,680
Speaker 4: top priority because you know, if for the whole a

621
00:30:13,840 --> 00:30:17,720
Speaker 4: marketing deal we are seeing actually HPN and Server duren together,

622
00:30:18,000 --> 00:30:20,000
Speaker 4: it's probably like more it's more than half of the

623
00:30:20,040 --> 00:30:24,320
Speaker 4: durrent market, right, and that's important, and it's going so fast, right,

624
00:30:24,360 --> 00:30:27,240
Speaker 4: and it's unlike the mobile, right, mobile is kind of flat.

625
00:30:27,560 --> 00:30:31,320
Speaker 4: The mobile demand is really driven by the increased different

626
00:30:31,360 --> 00:30:33,880
Speaker 4: contents in the mobile, which is Crian Dimater right, if

627
00:30:33,920 --> 00:30:35,880
Speaker 4: you buy iPhone over the past, you know, like they

628
00:30:35,880 --> 00:30:39,120
Speaker 4: are actually crane demating on their deferent content increase. So

629
00:30:39,280 --> 00:30:40,760
Speaker 4: I think, I think, you know, loads are the top

630
00:30:40,800 --> 00:30:43,240
Speaker 4: two parody for the memory company to focus on in

631
00:30:43,240 --> 00:30:45,400
Speaker 4: the next two two years. And we'll see what happened,

632
00:30:45,720 --> 00:30:48,600
Speaker 4: you know, after like second half of twenty twenty seven

633
00:30:48,760 --> 00:30:51,840
Speaker 4: or twenty twenty if there is some structural change. But

634
00:30:51,840 --> 00:30:54,080
Speaker 4: I think Server Durent HPM will be the top parties.

635
00:31:09,560 --> 00:31:12,600
Speaker 2: You know, the other day, I was very lazy, it

636
00:31:12,760 --> 00:31:18,160
Speaker 2: was not just one day, and I used claud code.

637
00:31:18,480 --> 00:31:20,120
Speaker 2: You know, I have a bunch of screenshots on the

638
00:31:20,160 --> 00:31:23,240
Speaker 2: desktop of my computer and it sort of makes my

639
00:31:23,280 --> 00:31:25,280
Speaker 2: desktop look like a mass and I can't see folders,

640
00:31:25,440 --> 00:31:27,560
Speaker 2: so I like drag and drop them and put them

641
00:31:27,560 --> 00:31:30,719
Speaker 2: in the recycling then and it was very lazy, and

642
00:31:30,760 --> 00:31:33,120
Speaker 2: I put it to claud Code. I said, clean up

643
00:31:33,160 --> 00:31:35,560
Speaker 2: my desktop. Clean up my desktop. And I was like,

644
00:31:35,600 --> 00:31:40,360
Speaker 2: this is so insane. I'm using you know, anthropics computers

645
00:31:40,400 --> 00:31:42,560
Speaker 2: wherever they are on the other side of the country

646
00:31:42,800 --> 00:31:45,160
Speaker 2: to clean up my own computer's desktop. Why do I

647
00:31:45,200 --> 00:31:47,680
Speaker 2: even have my own A good job, of course, yeah,

648
00:31:47,680 --> 00:31:49,400
Speaker 2: it did great, And I was like, why do I

649
00:31:49,440 --> 00:31:52,479
Speaker 2: even have my own computer? At this point I started wondering,

650
00:31:52,560 --> 00:31:54,560
Speaker 2: you know, I'm like, use, what why do I even

651
00:31:54,600 --> 00:31:58,240
Speaker 2: have a computer? Could it so related to this sort

652
00:31:58,280 --> 00:32:01,960
Speaker 2: of dram question. Could we just get into a situation

653
00:32:02,200 --> 00:32:06,120
Speaker 2: in which people don't really have their own compute on

654
00:32:06,240 --> 00:32:10,920
Speaker 2: their own device, because more that, like anthropics, brain is

655
00:32:10,960 --> 00:32:13,280
Speaker 2: sort of or open the eyes, brain is sort of

656
00:32:13,320 --> 00:32:17,000
Speaker 2: controlling the things we use. Why not just have a

657
00:32:17,160 --> 00:32:20,640
Speaker 2: very low spec computer or low spec phone or whatever,

658
00:32:20,920 --> 00:32:24,640
Speaker 2: or you know, low memory whatever, because I'm already using

659
00:32:24,800 --> 00:32:26,880
Speaker 2: their computers, et cetera. And are we just going to

660
00:32:27,160 --> 00:32:31,000
Speaker 2: see whether memory or elsewhere the sort of migration or

661
00:32:31,080 --> 00:32:33,360
Speaker 2: all of the interesting stuff and all of the memory

662
00:32:33,360 --> 00:32:36,560
Speaker 2: and all of the compute happens elsewhere and my phone

663
00:32:36,600 --> 00:32:38,840
Speaker 2: just becomes a sort of you know, or my computer

664
00:32:38,960 --> 00:32:42,080
Speaker 2: just becomes a screen with an Internet access.

665
00:32:42,400 --> 00:32:45,080
Speaker 4: Yeah, I mean I do. I don't know of all

666
00:32:45,120 --> 00:32:47,840
Speaker 4: the idea, like you know, not having the personal devices.

667
00:32:47,920 --> 00:32:50,000
Speaker 2: Yeah, I might like have a physical device, but it

668
00:32:50,120 --> 00:32:53,160
Speaker 2: just sort of seems similar, like over time, does it

669
00:32:53,240 --> 00:32:57,000
Speaker 2: make sense for me to have all of this actual

670
00:32:57,080 --> 00:32:59,280
Speaker 2: compute inside my house when I'm just like I might

671
00:32:59,360 --> 00:33:01,000
Speaker 2: as well let them have it all.

672
00:33:01,800 --> 00:33:04,280
Speaker 4: Yeah, I will say it depends on what's your purpose

673
00:33:04,400 --> 00:33:06,400
Speaker 4: y you know, for example, laptop for example, right, Like,

674
00:33:06,400 --> 00:33:08,880
Speaker 4: if you're doing like media editing merment heavy, right, you

675
00:33:08,960 --> 00:33:10,800
Speaker 4: probably still a loft. You probably still need quite a

676
00:33:10,840 --> 00:33:13,360
Speaker 4: lofty DRN right. If you're just you know, doing like

677
00:33:13,360 --> 00:33:16,400
Speaker 4: you know, document work right talk cold depends on how

678
00:33:16,440 --> 00:33:17,800
Speaker 4: are you going to use it. If I'm using like

679
00:33:17,840 --> 00:33:20,480
Speaker 4: super intensively, I think you still need a quite good DRENT,

680
00:33:20,840 --> 00:33:23,920
Speaker 4: especially you're like pulling all different API from different places, right.

681
00:33:24,480 --> 00:33:26,120
Speaker 4: So I will say it just really depends on the

682
00:33:26,200 --> 00:33:28,280
Speaker 4: m purpose. I don't think like you know, at least

683
00:33:28,280 --> 00:33:29,680
Speaker 4: at least I haven't seen like you know, there will

684
00:33:29,720 --> 00:33:33,240
Speaker 4: be like kind of structural uh sort of theorem destruction

685
00:33:33,440 --> 00:33:35,840
Speaker 4: in the devices from five genting AI.

686
00:33:36,680 --> 00:33:39,440
Speaker 3: So, going back to the beginning of this conversation and

687
00:33:39,480 --> 00:33:43,880
Speaker 3: the cyclicality of the industry, people have been throwing around

688
00:33:44,000 --> 00:33:48,640
Speaker 3: the term super cycle again very commodities ask is this

689
00:33:48,960 --> 00:33:52,360
Speaker 3: just a super cycle, you know, maybe bigger than a

690
00:33:52,960 --> 00:33:56,120
Speaker 3: previous turns that we've seen throughout history, or do you

691
00:33:56,160 --> 00:34:00,719
Speaker 3: think something structural has shifted here? Perhaps given and the

692
00:34:00,800 --> 00:34:03,680
Speaker 3: rise of AI and the fact that Joe is you know,

693
00:34:04,080 --> 00:34:07,920
Speaker 3: using Claude to clean up his desktop, this.

694
00:34:07,960 --> 00:34:10,480
Speaker 4: Is an actually very dangerous question to real You are

695
00:34:10,520 --> 00:34:13,040
Speaker 4: basically asking, right, whether this time is different?

696
00:34:13,120 --> 00:34:13,319
Speaker 1: Right?

697
00:34:13,560 --> 00:34:13,799
Speaker 3: Right?

698
00:34:15,800 --> 00:34:17,200
Speaker 4: I think, like, you know, I think there's a couple

699
00:34:17,200 --> 00:34:19,560
Speaker 4: of things that it actually runs with the history, right,

700
00:34:19,600 --> 00:34:22,640
Speaker 4: the non migration, fair capacity, right, those things. But I

701
00:34:22,800 --> 00:34:25,759
Speaker 4: was like, you know, there's some kind of difference I'm seeing, right,

702
00:34:25,800 --> 00:34:29,400
Speaker 4: because we really see in a sort of supercycle, that

703
00:34:29,840 --> 00:34:33,319
Speaker 4: there's a new demand driver coming online. It's not only

704
00:34:33,360 --> 00:34:36,439
Speaker 4: constrain the demand, this thing is also constrained to supply, right,

705
00:34:37,040 --> 00:34:39,160
Speaker 4: because we know, for example, in twenty twenty ten to

706
00:34:39,200 --> 00:34:42,200
Speaker 4: twenty twelve, there's a supercycle driven by mobile. Right. So

707
00:34:42,280 --> 00:34:44,200
Speaker 4: mobile is basically, hey, we have a new part of

708
00:34:44,280 --> 00:34:46,040
Speaker 4: coming online and it require a lot of DRA and

709
00:34:46,120 --> 00:34:49,680
Speaker 4: let's let's and our capacity couldn't catch up. But right

710
00:34:49,680 --> 00:34:51,799
Speaker 4: now for h men, and it's like, hey, okay, we

711
00:34:51,840 --> 00:34:53,960
Speaker 4: need to do HPN, but your SHPN is actually constraining

712
00:34:53,960 --> 00:34:56,920
Speaker 4: your commodity de ran capacity. So I think this is

713
00:34:56,960 --> 00:34:59,319
Speaker 4: the key difference that we are saying for this, right,

714
00:34:59,360 --> 00:35:03,640
Speaker 4: and apparently the demand accelerations it's also last quite long, right,

715
00:35:03,840 --> 00:35:06,120
Speaker 4: if you really starting counting the sort of I would

716
00:35:06,120 --> 00:35:09,040
Speaker 4: say the air acceleration in terms of demand probably let's

717
00:35:09,080 --> 00:35:10,879
Speaker 4: say second half of twenty twenty straight when a VIDIA

718
00:35:11,000 --> 00:35:14,800
Speaker 4: really start kind of emerging in people's attention. Let's to

719
00:35:14,920 --> 00:35:17,120
Speaker 4: now it's almost probably like let's say three years, right,

720
00:35:17,160 --> 00:35:18,759
Speaker 4: two year a half, two three years, and we are

721
00:35:18,760 --> 00:35:22,200
Speaker 4: looking at this cycle to last until let's say safely

722
00:35:22,239 --> 00:35:24,200
Speaker 4: to say like second half of twenty seven, right, So

723
00:35:24,239 --> 00:35:26,440
Speaker 4: we are looking at like kind of four year cycle

724
00:35:26,520 --> 00:35:30,120
Speaker 4: that's quite rare in the history. The usually like the

725
00:35:30,200 --> 00:35:32,320
Speaker 4: cycle if you look at historically for the memory is

726
00:35:32,360 --> 00:35:35,640
Speaker 4: probably like fifteen months, like the probably longest like eighteen

727
00:35:35,640 --> 00:35:38,560
Speaker 4: months from start to its peak. But right now we

728
00:35:38,600 --> 00:35:40,880
Speaker 4: are going to an area that was like, hey, the

729
00:35:40,920 --> 00:35:44,520
Speaker 4: demand was going up, not a price directly. But the

730
00:35:44,560 --> 00:35:47,160
Speaker 4: demand at least was growing quite significantly over the past

731
00:35:47,239 --> 00:35:50,080
Speaker 4: few years, right, and the pricing impact is happening right

732
00:35:50,120 --> 00:35:53,080
Speaker 4: now starting from probably Q two twenty twenty five.

733
00:35:53,560 --> 00:35:55,920
Speaker 2: Well, what is it when you say, okay, the cycle

734
00:35:55,960 --> 00:35:58,799
Speaker 2: could end by twenty twenty seven or the second half

735
00:35:58,840 --> 00:36:02,160
Speaker 2: of twenty twenty seven, what are the ingredients for it

736
00:36:02,239 --> 00:36:05,400
Speaker 2: to slow down? Is it the more production capacity? Is

737
00:36:05,440 --> 00:36:09,440
Speaker 2: it a slow down in act total demand? Like what

738
00:36:09,440 --> 00:36:12,399
Speaker 2: what do you see potentially happening in twenty twenty seven

739
00:36:12,440 --> 00:36:14,240
Speaker 2: that brings applying demand more into balance?

740
00:36:14,880 --> 00:36:16,920
Speaker 4: Yeah, so I will say that it will end. I

741
00:36:17,000 --> 00:36:21,320
Speaker 4: will say, like, it's a more tricky area to because

742
00:36:21,360 --> 00:36:23,560
Speaker 4: you know, based on the number you know I'm modeling,

743
00:36:23,640 --> 00:36:26,359
Speaker 4: we are seeing here the demand is actually going higher

744
00:36:26,400 --> 00:36:28,799
Speaker 4: and the shortage is actually going to worse. And the

745
00:36:28,840 --> 00:36:32,319
Speaker 4: power of that is because a media ad server, their

746
00:36:32,400 --> 00:36:36,279
Speaker 4: dealer in contents is going to three acts compared to Gregwell, right,

747
00:36:36,680 --> 00:36:42,080
Speaker 4: and the video server lubin two hundred. Yeah, okay, I'm sorry,

748
00:36:42,160 --> 00:36:44,760
Speaker 4: looping out, Sorry looping out our looping three hundred okay?

749
00:36:45,120 --> 00:36:47,960
Speaker 4: And if that's okay, and the demand will continue to

750
00:36:47,960 --> 00:36:51,640
Speaker 4: be that strong, right, you're actually being shortage to even

751
00:36:51,760 --> 00:36:55,160
Speaker 4: be worse compared to twenty six, and your dealer should

752
00:36:55,200 --> 00:36:57,000
Speaker 4: go to the go worse, right because the Murmer record

753
00:36:57,040 --> 00:37:00,720
Speaker 4: is gonna try to make more HPN waifer right, and

754
00:37:00,800 --> 00:37:04,680
Speaker 4: that kind of conting out more on commodity dram. I

755
00:37:04,719 --> 00:37:06,920
Speaker 4: think the reason I say it's more tricky because two

756
00:37:06,960 --> 00:37:09,920
Speaker 4: things I mentioned about a lot of migration. It's actually

757
00:37:09,960 --> 00:37:11,560
Speaker 4: happened this year, It's also going to happen in twenty

758
00:37:11,600 --> 00:37:13,359
Speaker 4: ninety seven, right, So towns of la migration are going

759
00:37:13,400 --> 00:37:15,640
Speaker 4: to come in online and being completed, so then will

760
00:37:15,680 --> 00:37:20,399
Speaker 4: add a lot of additional bits for the memory supplier. Additionally,

761
00:37:20,880 --> 00:37:24,440
Speaker 4: the wafer should have a lot more coming online by

762
00:37:24,480 --> 00:37:27,680
Speaker 4: the end of twenty six and for you know, throughout

763
00:37:27,680 --> 00:37:30,040
Speaker 4: twenty ninety seven. Right. So I think those are two

764
00:37:30,120 --> 00:37:32,880
Speaker 4: variables that you want to factor it in from what

765
00:37:33,000 --> 00:37:35,520
Speaker 4: I'm saying, at least I think like twenty twenty seven

766
00:37:35,719 --> 00:37:37,160
Speaker 4: therein going to still be storateaged.

767
00:37:37,520 --> 00:37:43,319
Speaker 3: Yeah, we need a strategic memory to preserve put a

768
00:37:43,360 --> 00:37:47,120
Speaker 3: floor on prices in the cycle. Actually, on that note,

769
00:37:47,360 --> 00:37:49,840
Speaker 3: do you see a lot of stockpiling in the market

770
00:37:49,920 --> 00:37:52,600
Speaker 3: or panic buying right now where people are seeing these

771
00:37:52,640 --> 00:37:55,879
Speaker 3: forecasts and they're freaking out and just buying whatever they can.

772
00:37:56,040 --> 00:37:57,759
Speaker 2: How you get a ball whip perfect, that's right.

773
00:37:58,960 --> 00:38:01,840
Speaker 4: Yeah, it's hard to say no, right, Like one of

774
00:38:01,920 --> 00:38:04,680
Speaker 4: the good I think the signal you've seen a tiny

775
00:38:04,840 --> 00:38:09,520
Speaker 4: sampsone and microt during an inventory was dropping every single

776
00:38:09,640 --> 00:38:13,600
Speaker 4: quarter see setting Q Street twenty twenty five. And that's

777
00:38:13,480 --> 00:38:15,160
Speaker 4: that's one of the things that you are not just

778
00:38:15,280 --> 00:38:19,319
Speaker 4: you know, buying the sort of products on the shelf, right,

779
00:38:19,360 --> 00:38:21,480
Speaker 4: and the products are coming online in that quarter, You're

780
00:38:21,480 --> 00:38:25,840
Speaker 4: also buying the products that in their inventory, right, You'll

781
00:38:25,920 --> 00:38:28,080
Speaker 4: try to get as much of the product as you can,

782
00:38:28,640 --> 00:38:30,640
Speaker 4: and you know, and you know, given the older demands

783
00:38:30,680 --> 00:38:34,000
Speaker 4: and how AI has developed so fast, like by month

784
00:38:34,120 --> 00:38:36,880
Speaker 4: monthly basis, it's hard to say, like, hey, there's no

785
00:38:37,719 --> 00:38:41,680
Speaker 4: pre entitive purchasing behavior happening, right, And especially when the

786
00:38:41,760 --> 00:38:44,640
Speaker 4: hyperskillers and the leading labs, they are competing each with

787
00:38:44,760 --> 00:38:49,360
Speaker 4: each other very intensively. Securing capacity on the hardware is

788
00:38:49,400 --> 00:38:51,759
Speaker 4: sort of the baseline, right. You want to get the

789
00:38:51,800 --> 00:38:53,520
Speaker 4: best equipment that you can compete, right.

790
00:38:54,239 --> 00:38:57,319
Speaker 2: So obviously if I'm buying a Nintendo Switch, if I'm

791
00:38:57,320 --> 00:39:00,400
Speaker 2: buying a PC or maybe some sort of consumer device,

792
00:39:00,920 --> 00:39:04,560
Speaker 2: then the increased memory cost is a meaningful driver of that.

793
00:39:04,760 --> 00:39:06,800
Speaker 2: And maybe you know, the company you have to raise prices,

794
00:39:06,920 --> 00:39:08,719
Speaker 2: maybe I won't buy it or whatever, and you might

795
00:39:08,800 --> 00:39:13,719
Speaker 2: get this. What about like for the big hyperscalers, when

796
00:39:13,760 --> 00:39:17,319
Speaker 2: we see these huge capital expenditure numbers from the likes

797
00:39:17,320 --> 00:39:21,600
Speaker 2: of a Microsoft or an Amazon, do these prices changes

798
00:39:21,719 --> 00:39:25,040
Speaker 2: move the dial at these levels in terms of do

799
00:39:25,120 --> 00:39:27,680
Speaker 2: they affect anything when we're talking about these buyers.

800
00:39:28,400 --> 00:39:30,160
Speaker 4: Yeah, I don't think you know number one, for sure,

801
00:39:30,280 --> 00:39:33,640
Speaker 4: I think the tapex is not really well, the tapex

802
00:39:33,719 --> 00:39:37,879
Speaker 4: increases not because of the dealerent price increase, Sure, deer

803
00:39:38,120 --> 00:39:42,279
Speaker 4: price increase does gonna have some kinds of impacts to

804
00:39:42,440 --> 00:39:46,560
Speaker 4: their purchases of memory, right, assuming the hyperscialers are a

805
00:39:46,640 --> 00:39:48,840
Speaker 4: direct buyer of the memory, which you know is not

806
00:39:48,880 --> 00:39:52,120
Speaker 4: necessarily the case in a different locations. Right, So if

807
00:39:52,160 --> 00:39:54,759
Speaker 4: they assumeing they're the direct buyers of your memory for sure? Right,

808
00:39:54,800 --> 00:39:56,920
Speaker 4: if you know, if your initial call was using a

809
00:39:57,280 --> 00:39:59,840
Speaker 4: hundred million to buyas set amount of memory, right now,

810
00:40:00,000 --> 00:40:02,319
Speaker 4: probably you need to do some this calendar for sure,

811
00:40:02,680 --> 00:40:04,840
Speaker 4: especially this thing that's probably gonna that there's going to

812
00:40:04,880 --> 00:40:07,080
Speaker 4: be more and more in the coming quarter, right, So

813
00:40:07,200 --> 00:40:08,440
Speaker 4: one of the things that I'll try it, and they

814
00:40:08,480 --> 00:40:10,600
Speaker 4: are hopefully they are trying to do, but they it's

815
00:40:10,640 --> 00:40:12,360
Speaker 4: still very difficult. It's like they try to have a

816
00:40:12,480 --> 00:40:16,200
Speaker 4: long term agreement, right to commit large value for the

817
00:40:16,320 --> 00:40:19,840
Speaker 4: year basis and to hopefully to have bene pricing, but

818
00:40:20,160 --> 00:40:24,200
Speaker 4: it's really hard to you know, to achieve that because

819
00:40:24,760 --> 00:40:26,400
Speaker 4: why the memory supply. I want to do that, right,

820
00:40:26,520 --> 00:40:29,400
Speaker 4: we can you know, ne negotiat of pricing in a

821
00:40:29,480 --> 00:40:31,719
Speaker 4: coding basis. We can actually get more money, right, and

822
00:40:32,120 --> 00:40:35,960
Speaker 4: especially given the current pricing and supply demand environments.

823
00:40:36,800 --> 00:40:39,440
Speaker 2: Wong, thank you so much for coming on Odd Lives.

824
00:40:39,480 --> 00:40:42,120
Speaker 2: Really the perfect guest. Let's us stay in touch and

825
00:40:42,800 --> 00:40:45,480
Speaker 2: maybe we'll revisit it in twenty twenty seven to see

826
00:40:45,480 --> 00:40:46,320
Speaker 2: if things have eased.

827
00:40:46,840 --> 00:40:50,400
Speaker 4: Yeah, hopefully all right, Thanks than care, take care of it.

828
00:40:50,520 --> 00:41:05,720
Speaker 5: That was great, Take care, Thank you, Tracy.

829
00:41:05,840 --> 00:41:08,480
Speaker 2: That was fun. I do feel like there's a lot

830
00:41:08,560 --> 00:41:11,560
Speaker 2: of moving pieces there, but just from AI in general,

831
00:41:11,680 --> 00:41:14,840
Speaker 2: the crowding out effect is such a big part of

832
00:41:14,920 --> 00:41:18,400
Speaker 2: the story. When when you saw those big capital expenditure

833
00:41:18,480 --> 00:41:21,880
Speaker 2: plans for twenty twenty seven, like these are real macro drivers.

834
00:41:21,920 --> 00:41:24,560
Speaker 2: They're going to show up at CPI, et cetera like that,

835
00:41:25,000 --> 00:41:27,000
Speaker 2: and because it's like, you know, maybe we'll get a

836
00:41:27,000 --> 00:41:30,160
Speaker 2: lot of productivity gains in the future, but right now

837
00:41:30,280 --> 00:41:32,680
Speaker 2: that pace of spending is so furious. It is like

838
00:41:33,080 --> 00:41:35,360
Speaker 2: a fiscal boom. Yeah, I meant stimulus.

839
00:41:35,520 --> 00:41:37,760
Speaker 3: This is what happens when some of the world's biggest

840
00:41:37,840 --> 00:41:42,040
Speaker 3: companies and most cash rich companies decide that this is

841
00:41:42,239 --> 00:41:46,440
Speaker 3: an existential threat. Right there's no upward limit on how

842
00:41:46,560 --> 00:41:49,600
Speaker 3: much they're willing to spend in order to survive, and

843
00:41:49,719 --> 00:41:53,359
Speaker 3: so you could just throw money at HBM chips, I guess,

844
00:41:53,600 --> 00:41:56,360
Speaker 3: and then you get the chip makers going well, actually,

845
00:41:56,560 --> 00:41:59,080
Speaker 3: like we want to produce a bunch of HBM like

846
00:41:59,160 --> 00:42:01,120
Speaker 3: forget about track all that stuff.

847
00:42:01,320 --> 00:42:03,920
Speaker 2: I do think like, yeah, you know, when it comes

848
00:42:03,960 --> 00:42:07,200
Speaker 2: to some of the data center and energy stuff, it's

849
00:42:07,200 --> 00:42:10,879
Speaker 2: a little ambiguous how much it's affecting energy prices right now,

850
00:42:11,000 --> 00:42:13,960
Speaker 2: but already the politics of that, yeah, very fraud. And

851
00:42:14,080 --> 00:42:17,440
Speaker 2: then you're going to start upsetting the gamer community because

852
00:42:17,600 --> 00:42:20,840
Speaker 2: they can't get act. You know, Nvidia has talked about

853
00:42:21,200 --> 00:42:24,759
Speaker 2: critiquing mistake, and then they're going to upset my son

854
00:42:25,200 --> 00:42:28,919
Speaker 2: because he can't get a Nintendo switch et cetera. People

855
00:42:28,960 --> 00:42:31,279
Speaker 2: are just going to start feeling it more and more,

856
00:42:31,360 --> 00:42:35,120
Speaker 2: the sort of visceral reality that various resources that they

857
00:42:35,200 --> 00:42:37,760
Speaker 2: thought they could get abundantly. It's like, nope, we've switched

858
00:42:37,800 --> 00:42:41,360
Speaker 2: this line over to the data centers, over to AI,

859
00:42:41,640 --> 00:42:44,520
Speaker 2: and it's going to become more This crowding out effect

860
00:42:44,560 --> 00:42:47,640
Speaker 2: is going to become more and more real to people directly, which.

861
00:42:47,480 --> 00:42:49,880
Speaker 3: Is so ironic because when you think about AI, you know,

862
00:42:49,960 --> 00:42:53,000
Speaker 3: it's this thing that exists in the ether, right, but

863
00:42:53,480 --> 00:42:57,120
Speaker 3: at the same time it has this huge physical impact

864
00:42:57,440 --> 00:43:00,920
Speaker 3: in the real world. It's kind of funny, and I guess, like,

865
00:43:01,360 --> 00:43:03,120
Speaker 3: not what a lot of people would have expected.

866
00:43:03,600 --> 00:43:06,880
Speaker 2: No, I suppose not the DRAM people. If anyone who

867
00:43:07,000 --> 00:43:09,920
Speaker 2: has a terminal or just should just look up DRAM prices,

868
00:43:10,040 --> 00:43:10,680
Speaker 2: because this is.

869
00:43:11,200 --> 00:43:12,680
Speaker 3: I thought you were going to say, anyone who has

870
00:43:12,719 --> 00:43:14,479
Speaker 3: a terminal should start stripping out.

871
00:43:15,239 --> 00:43:18,840
Speaker 2: Terminal, should just unscrew the back and pull out the DRAM. No,

872
00:43:18,920 --> 00:43:21,040
Speaker 2: but if you had people should just look at the chart,

873
00:43:21,200 --> 00:43:23,960
Speaker 2: because really, here's this thing that's very sleepy. I mean,

874
00:43:24,000 --> 00:43:26,440
Speaker 2: this was true commoditized to you know, this was the

875
00:43:26,560 --> 00:43:29,440
Speaker 2: low end in terms of what people were excited about

876
00:43:29,840 --> 00:43:32,640
Speaker 2: in chips and so forth, and it is great mentioned

877
00:43:32,680 --> 00:43:35,760
Speaker 2: in the beginning, you know, costs generally around this downward

878
00:43:35,880 --> 00:43:39,000
Speaker 2: trend and so forth, because growth was modest and technology

879
00:43:39,080 --> 00:43:41,800
Speaker 2: continues to improve. And then it literally just looks like

880
00:43:41,840 --> 00:43:45,360
Speaker 2: an l really in like the last four months, just

881
00:43:45,480 --> 00:43:49,640
Speaker 2: gone completely nuts. And so yeah, kind of a fascinating

882
00:43:49,800 --> 00:43:51,560
Speaker 2: spot to watch it. And it's just interesting to the

883
00:43:51,600 --> 00:43:54,200
Speaker 2: point that you were making how much it really is

884
00:43:54,360 --> 00:43:57,600
Speaker 2: like a commodity supercycle. Yeah, commodity cycle.

885
00:43:58,360 --> 00:44:01,000
Speaker 3: Well, I guess in twenty twenty seven, maybe we'll find out.

886
00:44:01,080 --> 00:44:02,200
Speaker 2: We'll see if it balances out.

887
00:44:02,280 --> 00:44:03,000
Speaker 3: Shall we leave it there?

888
00:44:03,120 --> 00:44:03,759
Speaker 2: Let's leave it there?

889
00:44:03,840 --> 00:44:05,960
Speaker 3: All right? This has been another episode of the Odd

890
00:44:06,000 --> 00:44:08,719
Speaker 3: Lots Podcast. I'm Tracy Alloway. You can follow me at

891
00:44:08,800 --> 00:44:09,560
Speaker 3: Tracy Alloway.

892
00:44:09,760 --> 00:44:12,480
Speaker 2: And I'm Jill Wisenthal. You can follow me at the Stalwart.

893
00:44:12,560 --> 00:44:15,480
Speaker 2: Follow our guest Bray Wong. He's at r Wong seven.

894
00:44:15,840 --> 00:44:19,040
Speaker 2: Follow our producers Carmen Rodriguez at Carman armand dash Ol

895
00:44:19,080 --> 00:44:22,200
Speaker 2: Bennett at Dashbot at Kelbrooks at Kelbrooks. And from our

896
00:44:22,239 --> 00:44:24,759
Speaker 2: Odd Laws content. Go to Bloomberg dot com slash odd

897
00:44:24,840 --> 00:44:27,520
Speaker 2: lotsch for a the daily newsletter and all of our episodes,

898
00:44:27,800 --> 00:44:29,719
Speaker 2: and you can chat about all of these topics twenty

899
00:44:29,760 --> 00:44:33,840
Speaker 2: four to seven in our discord Discord dot gg, slash odlogs.

900
00:44:34,040 --> 00:44:36,520
Speaker 3: And if you enjoy Oddlots, if you like it when

901
00:44:36,560 --> 00:44:39,920
Speaker 3: we talk about d RAM and HBM and other acronyms

902
00:44:40,000 --> 00:44:42,560
Speaker 3: of the memory chip industry, then please leave us a

903
00:44:42,600 --> 00:44:46,160
Speaker 3: positive review on your favorite podcast platform. And remember, if

904
00:44:46,200 --> 00:44:48,839
Speaker 3: you are a Bloomberg subscriber, you can listen to all

905
00:44:48,880 --> 00:44:50,960
Speaker 3: of our episodes absolutely ad free.

906
00:44:51,360 --> 00:44:52,279
Speaker 4: All you need to do is.

907
00:44:52,320 --> 00:44:54,960
Speaker 3: Find the Bloomberg channel on Apple Podcasts and follow the

908
00:44:55,000 --> 00:45:00,680
Speaker 3: instructions there. Thanks for listening it