MLCC，会是下一个“内存”吗？村田制作 VS 三星电机 VS 太阳诱电，谁是真正的 MLCC 之王？万字拆解 MLCC 赛道

1
00:00:00,000 --> 00:00:01,532
自从三年多以前

2
00:00:01,532 --> 00:00:03,200
ChatGPT 横空出世

3
00:00:03,266 --> 00:00:05,065
AI大模型浪潮爆发

4
00:00:05,200 --> 00:00:08,199
美股也随之进入了轰轰烈烈的AI牛市

5
00:00:08,365 --> 00:00:09,766
诞生了不少十倍股

6
00:00:09,766 --> 00:00:11,032
甚至是百倍股

7
00:00:11,333 --> 00:00:13,066
最早是AI训练时代

8
00:00:13,066 --> 00:00:15,532
海量参数的大模型并行计算

9
00:00:15,566 --> 00:00:17,033
需要海量的算力

10
00:00:17,233 --> 00:00:18,966
GPU是绝对的主角

11
00:00:19,233 --> 00:00:20,800
英伟达过去三年时间里

12
00:00:20,800 --> 00:00:22,265
股价涨了十几倍

13
00:00:22,332 --> 00:00:24,199
市值占上了5万亿美元

14
00:00:24,199 --> 00:00:26,533
刷新了人类历史上所有的记录

15
00:00:26,832 --> 00:00:29,132
紧接着进入到AI推理时代

16
00:00:29,199 --> 00:00:30,432
AI大模型的瓶颈

17
00:00:30,432 --> 00:00:33,066
从GPU转向了内存和存储

18
00:00:33,432 --> 00:00:34,932
这个领域的几家巨头

19
00:00:34,932 --> 00:00:37,732
包括三星、海力士、美光、闪迪

20
00:00:37,966 --> 00:00:39,966
他们的股价在短短一年时间里

21
00:00:39,966 --> 00:00:41,033
集体狂飙

22
00:00:41,066 --> 00:00:43,432
碾压了几乎所有的美股上市公司

23
00:00:43,832 --> 00:00:46,432
然而就在最近两三个月的时间里

24
00:00:46,533 --> 00:00:47,832
又有一个新的东西

25
00:00:47,832 --> 00:00:50,365
悄悄登上了AI主线的历史舞台

26
00:00:50,665 --> 00:00:52,832
它被投行机构称为新内存

27
00:00:52,832 --> 00:00:54,399
名字叫做 MLCC

28
00:00:54,399 --> 00:00:56,000
多层陶瓷电容器

29
00:00:56,365 --> 00:00:57,365
今年5月底

30
00:00:57,365 --> 00:00:59,432
高盛连续发了好几篇研报

31
00:00:59,466 --> 00:01:02,765
认为当前这一轮AI驱动的 MLCC 周期

32
00:01:02,832 --> 00:01:04,299
将是史上规模最大

33
00:01:04,299 --> 00:01:06,066
持续时间最长的一次

34
00:01:06,233 --> 00:01:08,200
而且仍然处于早期的阶段

35
00:01:08,500 --> 00:01:09,965
他原话是这样说的

36
00:01:10,033 --> 00:01:12,165
MLCC是AI服务器里面

37
00:01:12,165 --> 00:01:15,200
继GPU和内存之后的第三大成本项

38
00:01:15,665 --> 00:01:18,965
他把 MLCC 从一个传统的周期性元件

39
00:01:19,066 --> 00:01:19,965
正式归入到了

40
00:01:19,965 --> 00:01:22,433
与HBM同一层级的AI算力

41
00:01:22,433 --> 00:01:24,165
关键瓶颈叙事

42
00:01:24,299 --> 00:01:25,365
那么问题来了

43
00:01:25,365 --> 00:01:28,365
MLCC 真的会成为下一个内存吗

44
00:01:28,632 --> 00:01:31,400
村田制作、三星电机和太阳诱电

45
00:01:31,400 --> 00:01:32,900
这三大核心玩家

46
00:01:33,033 --> 00:01:35,266
谁才是真正的 MLCC 之王

47
00:01:35,700 --> 00:01:37,365
对于我们普通散户来说

48
00:01:37,365 --> 00:01:40,599
MLCC这个赛道现在还能不能投资

49
00:01:40,965 --> 00:01:41,900
今天这期视频

50
00:01:41,900 --> 00:01:43,766
我会用40多分钟的时间

51
00:01:43,865 --> 00:01:46,400
一次性把这条产业链拆解清楚

52
00:01:46,832 --> 00:01:47,665
hello大家好

53
00:01:47,665 --> 00:01:49,400
这里是Nico前沿Alpha

54
00:01:49,500 --> 00:01:51,233
力求用最通俗易懂的语言

55
00:01:51,233 --> 00:01:53,200
拆解分析最前沿的赛道

56
00:01:53,200 --> 00:01:54,665
挖掘下一个Alpha

57
00:01:54,865 --> 00:01:57,066
如果大家想要获取今天视频的原稿

58
00:01:57,066 --> 00:01:58,465
视频大纲PPT

59
00:01:58,533 --> 00:02:00,233
或者想要在第一时间

60
00:02:00,233 --> 00:02:02,266
看到我的个股买入卖出动态

61
00:02:02,266 --> 00:02:03,832
深度个股投研内容

62
00:02:03,832 --> 00:02:06,400
美股行情观察、财报季解读的话

63
00:02:06,465 --> 00:02:08,265
欢迎订阅我的 patreon 社群

64
00:02:08,265 --> 00:02:10,165
加入 discord 会员聊天室

65
00:02:10,466 --> 00:02:12,733
订阅费用是每周一杯咖啡的价格

66
00:02:12,733 --> 00:02:14,199
可以免费试用7天

67
00:02:14,599 --> 00:02:17,199
关于 Patreon 会员社群的具体权益介绍

68
00:02:17,199 --> 00:02:18,332
以及订阅链接

69
00:02:18,332 --> 00:02:19,599
我放到视频简介区了

70
00:02:19,599 --> 00:02:21,400
欢迎志同道合的朋友加入

71
00:02:21,400 --> 00:02:21,900
接着

72
00:02:21,900 --> 00:02:25,165
我们就正式进入今天视频的主题MLCC

73
00:02:25,366 --> 00:02:26,733
在具体聊投资之前

74
00:02:26,733 --> 00:02:27,800
我们需要先学习

75
00:02:27,800 --> 00:02:29,599
了解一些最基础的概念

76
00:02:29,800 --> 00:02:31,665
MLCC到底是什么

77
00:02:31,765 --> 00:02:35,133
它的英文全名叫做Multilayer Ceramic Capacitor

78
00:02:35,300 --> 00:02:37,765
中文名字叫做多层陶瓷电容器

79
00:02:37,866 --> 00:02:39,000
是电子设备里面

80
00:02:39,000 --> 00:02:40,800
最常见的被动元件之一

81
00:02:41,099 --> 00:02:42,400
所谓被动元件

82
00:02:42,400 --> 00:02:43,900
就是它不会像GPU

83
00:02:43,900 --> 00:02:46,000
CPU那样主动参与计算

84
00:02:46,165 --> 00:02:48,099
也不承担存储数据的功能

85
00:02:48,366 --> 00:02:49,765
它在电路里面的角色

86
00:02:49,765 --> 00:02:51,400
更像是一个稳定器

87
00:02:51,532 --> 00:02:54,599
负责储存电荷、稳定电压、过滤噪声

88
00:02:54,599 --> 00:02:56,466
改善信号和供电的质量

89
00:02:56,800 --> 00:02:58,733
相比起GPU、CPU

90
00:02:58,733 --> 00:03:01,165
HBM这些耳熟能详的玩家

91
00:03:01,432 --> 00:03:03,500
MLCC并不是AI服务器里面

92
00:03:03,500 --> 00:03:04,765
最显眼的主角

93
00:03:05,000 --> 00:03:07,099
它不会让模型运算的速度更快

94
00:03:07,099 --> 00:03:09,400
也没办法存储搬运更多的数据

95
00:03:09,733 --> 00:03:12,699
但是它能够决定GPU、HBM这些主角

96
00:03:12,699 --> 00:03:14,366
能不能够吃到足够干净

97
00:03:14,366 --> 00:03:16,032
稳定及时的电荷

98
00:03:16,432 --> 00:03:18,599
MLCC这个东西看起来非常小

99
00:03:18,599 --> 00:03:19,699
非常不起眼

100
00:03:19,800 --> 00:03:21,532
概念听起来也比较无聊

101
00:03:21,699 --> 00:03:24,032
但是我们身边几乎所有的电子设备

102
00:03:24,032 --> 00:03:27,032
包括手机、电脑、汽车、服务器

103
00:03:27,099 --> 00:03:29,165
里面都有大量的MLCC

104
00:03:29,400 --> 00:03:32,300
那为什么它被叫做多层陶瓷电容器呢

105
00:03:32,633 --> 00:03:34,099
你可以把MLCC

106
00:03:34,099 --> 00:03:37,432
想象成一块被切得极小的千层夹心饼干

107
00:03:37,432 --> 00:03:39,432
它的内部不是一整块陶瓷

108
00:03:39,432 --> 00:03:40,966
而是一层陶瓷介质

109
00:03:41,000 --> 00:03:42,300
盖一层金属电极

110
00:03:42,300 --> 00:03:44,000
再加上一层陶瓷介质

111
00:03:44,066 --> 00:03:46,000
再继续堆一层金属电极

112
00:03:46,300 --> 00:03:47,566
这样不断地重复

113
00:03:47,566 --> 00:03:48,699
最后压在一起

114
00:03:48,699 --> 00:03:50,633
高温烧结成一个小方块

115
00:03:50,900 --> 00:03:53,032
外面再接上两个金属端头

116
00:03:53,066 --> 00:03:53,665
就变成了

117
00:03:53,665 --> 00:03:56,466
我们在电路板上面看到的那颗MLCC

118
00:03:56,866 --> 00:03:58,966
对于一颗高端的MLCC来说

119
00:03:58,966 --> 00:04:00,699
可能比一粒芝麻还要小

120
00:04:00,866 --> 00:04:01,699
但它的内部

121
00:04:01,699 --> 00:04:03,066
可能堆叠了几百层

122
00:04:03,066 --> 00:04:04,400
甚至是上千层

123
00:04:04,665 --> 00:04:06,265
这就是它的核心结构

124
00:04:06,699 --> 00:04:08,366
很多人第一次看到这个东西

125
00:04:08,366 --> 00:04:10,633
会觉得这不就是一个小电容吗

126
00:04:10,866 --> 00:04:12,532
但它真正的难点在于

127
00:04:12,532 --> 00:04:14,566
要在一个非常小的体积里面

128
00:04:14,633 --> 00:04:17,065
把这么多层的陶瓷介质和金属电极

129
00:04:17,065 --> 00:04:17,966
堆叠进去

130
00:04:18,100 --> 00:04:20,833
并且每一层都要足够薄、足够均匀

131
00:04:20,865 --> 00:04:22,432
可靠性要足够高

132
00:04:22,500 --> 00:04:24,533
任何一层出现微小的缺陷

133
00:04:24,533 --> 00:04:27,065
都可能导致整颗产品漏电、击穿

134
00:04:27,065 --> 00:04:28,132
寿命下降

135
00:04:28,132 --> 00:04:29,533
甚至是直接报废

136
00:04:30,033 --> 00:04:33,100
MLCC最常用的核心材料是钛酸钡

137
00:04:33,333 --> 00:04:34,432
这个材料决定了

138
00:04:34,432 --> 00:04:36,500
电容能不能在很小的体积里面

139
00:04:36,500 --> 00:04:38,132
存下足够多的电荷

140
00:04:38,466 --> 00:04:39,899
越高端的MLCC

141
00:04:39,899 --> 00:04:42,065
越考验钛酸钡粉体的细度

142
00:04:42,065 --> 00:04:43,665
陶瓷薄层的均匀度

143
00:04:43,665 --> 00:04:46,833
金属电极的精度、层压工艺、烧结曲线

144
00:04:46,966 --> 00:04:48,833
以及最关键的量率控制

145
00:04:49,333 --> 00:04:52,500
接下来我们再说说MLCC具体能干什么

146
00:04:52,500 --> 00:04:54,100
它主要有四个基础作用

147
00:04:54,100 --> 00:04:56,632
耦合、去耦、平滑和滤波

148
00:04:56,899 --> 00:04:59,333
耦合的意思就是让交流信号通过

149
00:04:59,333 --> 00:05:00,966
同时隔离直流信号

150
00:05:01,233 --> 00:05:01,800
去耦

151
00:05:01,800 --> 00:05:04,399
就是把电源线上的噪声和瞬态波动

152
00:05:04,399 --> 00:05:05,800
从芯片旁边吸走

153
00:05:06,333 --> 00:05:07,365
平滑的意思

154
00:05:07,365 --> 00:05:09,500
就是把电源输出的纹波压平

155
00:05:09,500 --> 00:05:10,966
让电压更加稳定

156
00:05:11,365 --> 00:05:14,466
滤波就是把不想要的频率成分衰减掉

157
00:05:14,766 --> 00:05:16,632
听起来可能还是有点抽象

158
00:05:16,800 --> 00:05:18,865
我给大家做一个更简单的类比

159
00:05:19,165 --> 00:05:20,000
你可以把GPU

160
00:05:20,000 --> 00:05:22,766
想象成一个疯狂耗电的超级发动机

161
00:05:23,100 --> 00:05:24,766
这个发动机有个特点

162
00:05:24,833 --> 00:05:26,065
它不是匀速运转

163
00:05:26,065 --> 00:05:27,833
而是会突然猛踩油门

164
00:05:28,132 --> 00:05:29,899
比如说训练大模型的时候

165
00:05:29,899 --> 00:05:32,300
GPU可能突然进入高负载状态

166
00:05:32,533 --> 00:05:33,632
推理的时候

167
00:05:33,632 --> 00:05:35,932
GPU可能突然处理大量的token

168
00:05:36,300 --> 00:05:38,565
HBM 访问、NVLink 通信

169
00:05:38,565 --> 00:05:39,766
都可能让芯片负载

170
00:05:39,766 --> 00:05:41,899
在很短的时间里快速变化

171
00:05:42,165 --> 00:05:45,199
当GPU从低负载切换到高负载的时候

172
00:05:45,300 --> 00:05:47,199
电流需求会瞬间暴涨

173
00:05:47,500 --> 00:05:49,833
但电源模块离GPU还有些距离

174
00:05:50,000 --> 00:05:53,300
中间有 PCB 走线、过孔、焊点、电源平面

175
00:05:53,300 --> 00:05:54,533
封装、基板

176
00:05:54,699 --> 00:05:56,699
这些东西都不是理想的导体

177
00:05:56,699 --> 00:05:58,165
都有电阻和电感

178
00:05:58,600 --> 00:05:59,399
这个时候

179
00:05:59,399 --> 00:06:02,500
贴在GPU旁边的MLCC就派上用场了

180
00:06:02,766 --> 00:06:05,233
它就像是一排小型的应急电池

181
00:06:05,365 --> 00:06:07,033
在微秒级时间内响应

182
00:06:07,100 --> 00:06:08,932
并进行局部放电补偿

183
00:06:08,966 --> 00:06:11,632
帮GPU顶住这一瞬间的电流需求

184
00:06:11,766 --> 00:06:13,632
等远处的电源慢慢补上

185
00:06:13,833 --> 00:06:15,533
当电压产生波动的时候

186
00:06:15,533 --> 00:06:17,365
MLCC可以吸收波动

187
00:06:17,365 --> 00:06:19,132
当高频噪当高频噪声出现的时候

188
00:06:19,132 --> 00:06:21,632
MLCC能够直接把噪声过滤掉

189
00:06:21,833 --> 00:06:23,665
如果没有MLCC的话

190
00:06:23,665 --> 00:06:25,565
GPU虽然能够正常开机

191
00:06:25,733 --> 00:06:27,632
但是在高负载的状态下

192
00:06:27,632 --> 00:06:30,466
可能会出现电压下限、降频、误码

193
00:06:30,699 --> 00:06:33,365
甚至让整台设备瞬间崩溃重启

194
00:06:33,833 --> 00:06:34,632
讲到这里

195
00:06:34,632 --> 00:06:37,266
MLCC的基本概念我们就大概了解了

196
00:06:37,533 --> 00:06:39,766
但真正值得我们深挖的问题是

197
00:06:39,833 --> 00:06:42,365
MLCC并不是一个新诞生的东西

198
00:06:42,632 --> 00:06:44,500
过去几十年一直都存在

199
00:06:44,500 --> 00:06:47,300
手机汽车传统的服务器里面都有

200
00:06:47,833 --> 00:06:50,100
但为什么进入AI Agent时代之后

201
00:06:50,100 --> 00:06:52,565
它突然被投行机构称为新内存

202
00:06:52,699 --> 00:06:55,632
甚至被重新纳入AI算力瓶颈的趋势

203
00:06:55,632 --> 00:06:57,132
要理解清楚这一点

204
00:06:57,132 --> 00:06:59,500
我们必须得先回到AI服务器本身

205
00:06:59,800 --> 00:07:02,132
我们先看一条最基础的供电链路

206
00:07:02,399 --> 00:07:05,199
数据中心的电力会首先进入机架

207
00:07:05,365 --> 00:07:06,932
然后经过Power Shelf

208
00:07:06,932 --> 00:07:08,565
也就是机架及电源架

209
00:07:08,632 --> 00:07:12,533
再经过 PSU、电源模块、VRM、DC 转换器

210
00:07:12,665 --> 00:07:14,165
最后才被一步步降压

211
00:07:14,165 --> 00:07:19,132
送到 GPU、HBM、CPU、交换芯片这些核心负载的附近

212
00:07:19,565 --> 00:07:20,833
在这条链路里面

213
00:07:20,833 --> 00:07:23,733
MLCC不只是负责GPU供电的稳定

214
00:07:23,932 --> 00:07:24,932
它几乎分布在

215
00:07:24,932 --> 00:07:27,300
整个AI服务器的供电网络里面

216
00:07:27,500 --> 00:07:29,932
从数据中心供电到机架供电

217
00:07:29,932 --> 00:07:31,665
再到芯片近端供电

218
00:07:31,899 --> 00:07:33,665
MLCC都有涉及到

219
00:07:33,665 --> 00:07:35,632
这是一个系统级别的需求

220
00:07:36,199 --> 00:07:40,033
当然越靠近 GPU、HBM、ASIC 芯片

221
00:07:40,132 --> 00:07:42,399
用的往往是高容值、低 ESL

222
00:07:42,399 --> 00:07:45,233
小尺寸、高可靠的高端 MLCC

223
00:07:45,399 --> 00:07:46,966
它们的价值量就越大

224
00:07:46,966 --> 00:07:48,199
越难被替代

225
00:07:48,432 --> 00:07:49,766
理解清楚这一点之后

226
00:07:49,766 --> 00:07:51,600
我们再看看当下的现状

227
00:07:51,800 --> 00:07:54,833
MLCC元件出现结构性的供需失衡

228
00:07:55,065 --> 00:07:56,233
有很大一部分原因

229
00:07:56,233 --> 00:07:58,065
是因为AI服务器的升级

230
00:07:58,399 --> 00:07:59,632
它从单机逻辑

231
00:07:59,632 --> 00:08:02,833
比如说传统的英伟达A100、H100服务器

232
00:08:02,966 --> 00:08:04,865
已经走向了机架级的系统

233
00:08:05,065 --> 00:08:09,766
比如说最近英伟达出货量最大的 GB200 NVL72 机柜

234
00:08:09,766 --> 00:08:12,500
对于传统的A100 H100这些服务器

235
00:08:12,500 --> 00:08:13,833
就是一个铁盒子里面

236
00:08:13,833 --> 00:08:16,065
放几块GPU几块CPU

237
00:08:16,199 --> 00:08:18,665
再加上内存主板电源网卡

238
00:08:18,665 --> 00:08:20,466
散热风扇这些零部件

239
00:08:20,833 --> 00:08:22,800
虽然说英伟达的A100、H100

240
00:08:22,800 --> 00:08:24,632
已经是高性能AI服务器

241
00:08:24,699 --> 00:08:26,500
但它仍然是单机的形态

242
00:08:26,966 --> 00:08:30,365
而现在英伟达主流的 GB200 NVL72

243
00:08:30,365 --> 00:08:32,332
它是 36 个 Grace CPU

244
00:08:32,332 --> 00:08:36,466
加上72个Blackwell GPU组成的液冷机架级系统

245
00:08:36,466 --> 00:08:38,666
供电网络的复杂度急剧上升

246
00:08:39,100 --> 00:08:42,432
过去MLCC 只是给独立的芯片个体供电

247
00:08:42,633 --> 00:08:45,500
它密密麻麻地贴在单张H100算力卡

248
00:08:45,500 --> 00:08:47,332
或者单颗CPU的旁边

249
00:08:47,466 --> 00:08:48,500
它们各管各的

250
00:08:48,500 --> 00:08:50,600
属于局部的点对点供电

251
00:08:50,799 --> 00:08:53,265
而在进入机架级系统时代之后

252
00:08:53,265 --> 00:08:54,732
MLCC要服务的

253
00:08:54,732 --> 00:08:57,732
已经不只是芯片旁边的一小块供电区域

254
00:08:57,732 --> 00:09:01,899
而是整个机柜的高速互联网络和共享供电系统

255
00:09:01,899 --> 00:09:02,732
一方面

256
00:09:02,732 --> 00:09:04,399
NVSwitch 铜缆背板

257
00:09:04,399 --> 00:09:07,033
会在机柜内部产生大量的高频噪声

258
00:09:07,399 --> 00:09:10,365
MLCC需要分布在交换芯片、网卡

259
00:09:10,365 --> 00:09:11,865
高速通信链路周围

260
00:09:11,966 --> 00:09:13,533
帮助降低电源的噪声

261
00:09:13,533 --> 00:09:15,133
维持信号的完整性

262
00:09:15,600 --> 00:09:16,500
另一方面

263
00:09:16,500 --> 00:09:19,799
整个机柜的供电从过去分散的小电源

264
00:09:19,832 --> 00:09:21,000
变成Power Shelf

265
00:09:21,000 --> 00:09:23,765
加48伏直流母线的集中式供电

266
00:09:24,066 --> 00:09:26,432
电源先在机柜里转换成48伏

267
00:09:26,432 --> 00:09:29,166
再分发给各个计算托盘和电源模块

268
00:09:29,500 --> 00:09:30,466
这个过程中

269
00:09:30,466 --> 00:09:31,799
需要大量高压

270
00:09:31,799 --> 00:09:34,832
大容量、高可靠的MLCC来平滑电压

271
00:09:34,832 --> 00:09:35,966
抑制纹波

272
00:09:35,966 --> 00:09:38,566
缓冲枕柜级别的动态电流冲击

273
00:09:38,966 --> 00:09:41,732
AI服务器从单机走向机架时代

274
00:09:41,799 --> 00:09:43,633
对于 MLCC 行业的影响

275
00:09:43,633 --> 00:09:45,000
就是用量不断增加

276
00:09:45,000 --> 00:09:46,200
规格不断升级

277
00:09:46,200 --> 00:09:50,033
传统单台A100、H100服务器里面的MLCC用量

278
00:09:50,033 --> 00:09:51,566
都还在千颗级别

279
00:09:51,732 --> 00:09:55,332
而一个 GB200/GB300 NVL72 的机柜

280
00:09:55,332 --> 00:09:58,365
成了 72 颗 GPU、36 颗 CPU

281
00:09:58,500 --> 00:10:01,799
大量的 NVSwitch 芯片以及电源分配系统

282
00:10:02,100 --> 00:10:04,200
单机柜的 MLCC 总消耗量

283
00:10:04,200 --> 00:10:05,832
直接拉升到了数万颗

284
00:10:05,832 --> 00:10:07,566
甚至是十万颗的级别

285
00:10:08,033 --> 00:10:10,299
当然除了需求端的压力之外

286
00:10:10,299 --> 00:10:12,966
还有就是供给侧的产能完全跟不上

287
00:10:13,299 --> 00:10:15,966
高盛估计AI服务器的 MLCC 需求

288
00:10:15,966 --> 00:10:19,432
会在2025-2030年间增长4倍以上

289
00:10:19,500 --> 00:10:20,533
而行业产能

290
00:10:20,533 --> 00:10:23,633
预计每年最多小幅增长10%出头

291
00:10:23,966 --> 00:10:26,899
而且AI所需要的高端MLCC

292
00:10:26,899 --> 00:10:29,600
还会受制于材料、设备、良率

293
00:10:29,600 --> 00:10:31,232
以及客户认证周期

294
00:10:31,332 --> 00:10:32,633
导致供给端的产能

295
00:10:32,633 --> 00:10:34,633
完全跟不上需求端的增速

296
00:10:35,432 --> 00:10:37,500
这里我们还需要提到一个概念

297
00:10:37,700 --> 00:10:41,000
AI服务器真正需要的不是普通的MLCC

298
00:10:41,000 --> 00:10:42,765
而是高端的MLCC

299
00:10:42,899 --> 00:10:43,899
更具体来说

300
00:10:43,899 --> 00:10:45,799
就是高容值、低 ESL

301
00:10:45,799 --> 00:10:48,600
小尺寸、高压、高可靠的MLCC

302
00:10:49,133 --> 00:10:50,966
首先第一点高容值

303
00:10:50,966 --> 00:10:52,133
电容值越高

304
00:10:52,133 --> 00:10:53,799
意味着它在有限空间里面

305
00:10:53,799 --> 00:10:55,566
能够储存更多的电荷

306
00:10:55,765 --> 00:10:57,533
在新一代的AI服务器机架

307
00:10:57,533 --> 00:11:01,166
GPU、CPU、HBM、电源模块、交换芯片

308
00:11:01,166 --> 00:11:02,932
网卡的数量都在增加

309
00:11:02,932 --> 00:11:05,365
PCB板子上的空间非常紧张

310
00:11:05,566 --> 00:11:06,732
我们不可能无限制的

311
00:11:06,732 --> 00:11:09,166
往这些芯片旁边堆积MLCC

312
00:11:09,500 --> 00:11:11,432
所以说从工程的角度来看

313
00:11:11,432 --> 00:11:13,365
我们希望在有限的面积下

314
00:11:13,365 --> 00:11:15,133
放置更大的有效电容

315
00:11:15,365 --> 00:11:17,899
这样MLCC的数量就可以减少

316
00:11:18,000 --> 00:11:19,832
高容值的价值就在这里

317
00:11:20,299 --> 00:11:22,700
接下来第二点就是低 ESL

318
00:11:22,966 --> 00:11:25,365
先给大家解释一下ESL的概念

319
00:11:25,533 --> 00:11:28,033
它的中文名字叫做等效串联电感

320
00:11:28,232 --> 00:11:30,600
简单理解就是ESL越低

321
00:11:30,600 --> 00:11:33,832
电容越能处理高频快速的瞬态电流变化

322
00:11:34,200 --> 00:11:36,299
AI GPU的负载变化非常快

323
00:11:36,332 --> 00:11:38,399
电流需求可能在极短的时间里面

324
00:11:38,399 --> 00:11:39,432
突然变化

325
00:11:39,600 --> 00:11:41,332
如果电容响应不够快

326
00:11:41,432 --> 00:11:43,732
GPU的核心电压可能已经掉下去了

327
00:11:43,899 --> 00:11:45,700
电容还没来得及发挥作用

328
00:11:46,000 --> 00:11:48,432
所以我们希望ESL越低越好

329
00:11:48,832 --> 00:11:50,399
第三点是小尺寸

330
00:11:50,566 --> 00:11:52,299
这样做的核心优势就是

331
00:11:52,332 --> 00:11:54,700
MLCC可以离芯片更近一些

332
00:11:54,932 --> 00:11:56,399
这样寄生电感越低

333
00:11:56,399 --> 00:11:57,932
响应的速度就越快

334
00:11:58,232 --> 00:11:59,000
除此之外

335
00:11:59,000 --> 00:12:01,000
PCB板的空间也非常紧张

336
00:12:01,133 --> 00:12:04,100
尺寸越小就越能装下更多的MLCC

337
00:12:04,633 --> 00:12:06,100
第四点就是高压

338
00:12:06,100 --> 00:12:08,432
这一点得从AI机架的供电说起

339
00:12:08,899 --> 00:12:11,365
这几年AI机架的功耗涨得特别猛

340
00:12:11,600 --> 00:12:14,000
早期一个机架也就是几十千瓦

341
00:12:14,200 --> 00:12:17,232
现在的旗舰机架已经冲到一两百千瓦

342
00:12:17,232 --> 00:12:19,000
未来还要奔着一兆瓦去

343
00:12:19,332 --> 00:12:20,232
功率这么大

344
00:12:20,232 --> 00:12:22,500
怎么送进机架里就成了大问题

345
00:12:22,500 --> 00:12:25,932
这里我们需要提到一个很基础的初中物理公式

346
00:12:25,932 --> 00:12:27,899
功率等于电压乘以电流

347
00:12:28,133 --> 00:12:29,700
如果还用以前的低电压

348
00:12:29,700 --> 00:12:30,966
比如说48伏

349
00:12:31,133 --> 00:12:32,865
但是功率不变的情况下

350
00:12:32,865 --> 00:12:34,265
电流就会非常大

351
00:12:34,500 --> 00:12:35,332
举个例子

352
00:12:35,332 --> 00:12:36,500
48伏电压

353
00:12:36,500 --> 00:12:37,566
功率 1 兆瓦

354
00:12:37,666 --> 00:12:40,500
那么对应的电流要冲到上万安培

355
00:12:40,500 --> 00:12:42,000
而当电流变大之后

356
00:12:42,000 --> 00:12:43,633
铜缆就得做的特别粗

357
00:12:43,832 --> 00:12:46,200
对应的发热和损耗也会跟着暴涨

358
00:12:46,332 --> 00:12:48,200
这条路径很难一直走下去

359
00:12:48,533 --> 00:12:49,033
所以

360
00:12:49,033 --> 00:12:51,533
对应的解决方案就是把电压拉上去

361
00:12:51,732 --> 00:12:52,566
同样的功率

362
00:12:52,566 --> 00:12:53,332
电压越高

363
00:12:53,332 --> 00:12:54,832
需要的电流就越小

364
00:12:55,100 --> 00:12:57,265
这跟国家电网为什么要搞高压输电

365
00:12:57,265 --> 00:12:58,332
是同一个道理

366
00:12:58,633 --> 00:13:00,533
所以从新一代的架构开始

367
00:13:00,533 --> 00:13:02,100
英伟达把机架的主供电

368
00:13:02,100 --> 00:13:04,299
一路拉到了800伏直流高压

369
00:13:04,500 --> 00:13:06,133
但是问题也随之而来

370
00:13:06,365 --> 00:13:08,466
供电电压升级到800伏之后

371
00:13:08,466 --> 00:13:10,533
挂在这条线路上的所有元件

372
00:13:10,566 --> 00:13:13,500
包括MLCC都得扛得住这么高的电压

373
00:13:13,666 --> 00:13:15,332
否则就会被瞬间击穿

374
00:13:15,700 --> 00:13:18,033
所以说在800伏的新架构之下

375
00:13:18,033 --> 00:13:20,133
必须得用耐高压的MLCC

376
00:13:20,832 --> 00:13:22,765
第五点高可靠就不必多说了

377
00:13:22,765 --> 00:13:25,765
AI服务器涉及到了最核心的模型推理输出

378
00:13:25,765 --> 00:13:28,666
数据搬运、网络互联这些核心环节需要在满

379
00:13:28,700 --> 00:13:30,899
需要在满负荷、高温的条件下

380
00:13:30,899 --> 00:13:32,566
连续运转好几年时间

381
00:13:32,865 --> 00:13:35,732
MLCC 必须在这种复杂艰苦的环境下

382
00:13:35,732 --> 00:13:37,500
不失效持续地工作

383
00:13:37,533 --> 00:13:39,533
确保AI服务器供电的稳定

384
00:13:39,932 --> 00:13:42,133
在我们了解了MLCC是什么

385
00:13:42,133 --> 00:13:45,066
MLCC为什么出现了结构性供需失衡

386
00:13:45,299 --> 00:13:47,033
再回过头来看高盛的研报

387
00:13:47,033 --> 00:13:48,299
就会顺畅很多

388
00:13:48,533 --> 00:13:50,466
高盛所谓新内存的意思

389
00:13:50,500 --> 00:13:52,432
就是指高端MLCC

390
00:13:52,432 --> 00:13:55,100
正在从一个传统周期性被动元件

391
00:13:55,100 --> 00:13:58,232
被重新定价为AI算力系统中的关键瓶颈

392
00:13:58,533 --> 00:14:00,232
市场不再把 MLCC

393
00:14:00,232 --> 00:14:03,033
当做一个便宜普通的周期性电子零部件

394
00:14:03,100 --> 00:14:04,299
而是开始意识到

395
00:14:04,299 --> 00:14:06,765
它在AI服务器供电系统里的重要性

396
00:14:06,765 --> 00:14:07,966
已经明显提升

397
00:14:08,466 --> 00:14:09,332
高盛认为

398
00:14:09,332 --> 00:14:12,166
AI驱动了这一轮高端 MLCC 周期

399
00:14:12,232 --> 00:14:13,865
可能会是史上规模最大

400
00:14:13,865 --> 00:14:15,466
持续时间最长的一轮

401
00:14:15,533 --> 00:14:17,466
而且仍然处于早期阶段

402
00:14:17,700 --> 00:14:18,732
在他的研报中

403
00:14:18,732 --> 00:14:20,365
有这么几个核心的依据

404
00:14:20,899 --> 00:14:21,966
第一个依据是

405
00:14:21,966 --> 00:14:24,432
MLCC的周期性进一步延长

406
00:14:24,700 --> 00:14:27,133
在过去市场对于MLCC的认知

407
00:14:27,133 --> 00:14:28,966
是一个典型的周期商品

408
00:14:29,299 --> 00:14:31,066
背后会存在一个循环

409
00:14:31,066 --> 00:14:32,299
需求变得旺盛

410
00:14:32,299 --> 00:14:33,500
商品开始涨价

411
00:14:33,500 --> 00:14:34,666
厂商开始扩展

412
00:14:34,932 --> 00:14:37,133
而下游的客户担心继续涨价

413
00:14:37,133 --> 00:14:38,566
于是提前下订单

414
00:14:38,832 --> 00:14:40,432
各个渠道也开始囤货

415
00:14:40,666 --> 00:14:42,466
短期看起来需求非常好

416
00:14:42,466 --> 00:14:44,466
商品的价格也在继续往上走

417
00:14:44,466 --> 00:14:46,265
源头厂商公司的业绩释放

418
00:14:46,265 --> 00:14:47,732
股价也会开始上涨

419
00:14:47,732 --> 00:14:49,899
但是等到产能慢慢扩大之后

420
00:14:49,899 --> 00:14:52,166
如果真实的需求没有持续增长

421
00:14:52,332 --> 00:14:54,332
前面积累的库存就会滞销

422
00:14:54,399 --> 00:14:55,500
客户开始砍单

423
00:14:55,500 --> 00:14:56,932
渠道开始去库存

424
00:14:57,100 --> 00:14:58,332
价格开始下跌

425
00:14:58,332 --> 00:14:59,899
厂商的利润被压缩

426
00:14:59,932 --> 00:15:01,399
股价也就跟着回落了

427
00:15:01,600 --> 00:15:03,365
这就是典型的周期性商品

428
00:15:03,365 --> 00:15:04,299
很好理解

429
00:15:04,600 --> 00:15:07,399
所以这轮大家一听到MLCC涨价

430
00:15:07,500 --> 00:15:10,399
第一反应会不会又是2017-2018年

431
00:15:10,399 --> 00:15:11,832
那种常规的周期

432
00:15:12,200 --> 00:15:13,899
但从高盛的研报上来看

433
00:15:13,899 --> 00:15:15,500
这一轮并不完全一样

434
00:15:15,832 --> 00:15:17,299
MLCC 的需求

435
00:15:17,299 --> 00:15:19,399
不是单纯由消费电子拉动的

436
00:15:19,500 --> 00:15:22,166
更多是来自于AI 数据中心、AI 服务器

437
00:15:22,166 --> 00:15:24,899
高端汽车、边缘 AI 设备的共同推动

438
00:15:25,200 --> 00:15:26,600
AI服务器的建设

439
00:15:26,600 --> 00:15:28,600
并不是一年两年的短周期

440
00:15:28,666 --> 00:15:30,832
而是一个与AI推理需求扩张

441
00:15:30,832 --> 00:15:32,500
AI大模型厂商增长

442
00:15:32,633 --> 00:15:35,832
hyperscalar capex投入长期绑定的大周期

443
00:15:36,133 --> 00:15:37,299
所以高盛认为

444
00:15:37,299 --> 00:15:39,899
这一轮高端MLCC的需求周期

445
00:15:39,932 --> 00:15:42,066
可能会延续到2030年左右

446
00:15:42,133 --> 00:15:44,732
而不是市场之前预期的2028年见顶

447
00:15:45,232 --> 00:15:47,533
第二个依据是价格重置机制

448
00:15:47,732 --> 00:15:50,332
这一点是高盛逻辑里面最重要的地方

449
00:15:50,533 --> 00:15:51,832
传统消费电子里

450
00:15:51,832 --> 00:15:54,299
很多零部件的价格是逐年下滑的

451
00:15:54,600 --> 00:15:55,466
今年100块

452
00:15:55,466 --> 00:15:56,399
明年90块

453
00:15:56,399 --> 00:15:57,533
后年80块

454
00:15:57,732 --> 00:15:59,633
因为产品变得越来越成熟

455
00:15:59,633 --> 00:16:00,832
竞争越来越激烈

456
00:16:00,832 --> 00:16:02,100
客户持续压价

457
00:16:02,466 --> 00:16:05,500
但是AI服务器高端 MLCC 的逻辑不一样

458
00:16:05,633 --> 00:16:08,799
在英伟达平台从 GB200、GB300

459
00:16:08,799 --> 00:16:12,432
继续升级到 Rubin、Rubin Ultra 架构的演变过程中

460
00:16:12,432 --> 00:16:14,100
系统的复杂度不断提升

461
00:16:14,100 --> 00:16:15,299
BOM 大幅增加

462
00:16:15,466 --> 00:16:18,566
MLCC的数量、规格都在持续升级

463
00:16:18,865 --> 00:16:21,133
新平台需要更高规格的新产品

464
00:16:21,265 --> 00:16:23,899
新产品就有机会设定更高的价格

465
00:16:24,166 --> 00:16:26,633
高盛把这个机制称为价格重置

466
00:16:26,865 --> 00:16:27,666
表面上看

467
00:16:27,666 --> 00:16:29,299
这是产品结构优化

468
00:16:29,365 --> 00:16:31,100
但从最终的效果上看

469
00:16:31,100 --> 00:16:32,566
基本等同于涨价

470
00:16:32,765 --> 00:16:34,765
因为下游每一代AI服务器平台

471
00:16:34,765 --> 00:16:38,066
都需要更高端更复杂更可靠的MLCC

472
00:16:38,365 --> 00:16:40,966
所以这些MLCC相关的制造公司

473
00:16:40,966 --> 00:16:43,066
并不是对单一产品硬性涨价

474
00:16:43,066 --> 00:16:44,200
而是靠新平台

475
00:16:44,200 --> 00:16:46,566
新规格、新料号来重置价格

476
00:16:46,765 --> 00:16:49,033
这就和消费电子逻辑非常不同

477
00:16:49,033 --> 00:16:50,533
在这里顺便引用一下

478
00:16:50,533 --> 00:16:52,633
摩根士丹利自己的渠道测算

479
00:16:53,000 --> 00:16:55,732
目前每个GB 300机柜里面的MLCC

480
00:16:55,732 --> 00:16:58,200
价值量大概是在1,500美元左右

481
00:16:58,432 --> 00:17:01,066
而到了Vera Ruben VR200的机柜里面

482
00:17:01,066 --> 00:17:04,633
MLCC 的价值量可能提升到 4300 美元左右

483
00:17:04,865 --> 00:17:06,865
这个数据如果最终被验证

484
00:17:06,932 --> 00:17:09,633
那么说明每一代高端AI机柜的升级

485
00:17:09,700 --> 00:17:12,400
都在抬高 MLCC 单机柜的价值量

486
00:17:12,732 --> 00:17:13,633
关于这一点

487
00:17:13,633 --> 00:17:16,032
我们刚刚也有详细讲过背后的逻辑

488
00:17:16,333 --> 00:17:17,833
高盛的第三个依据

489
00:17:17,833 --> 00:17:19,432
就是涨价和出口数据

490
00:17:19,432 --> 00:17:20,865
已经开始给出验证

491
00:17:21,066 --> 00:17:22,599
他引用的数据里面

492
00:17:22,633 --> 00:17:24,965
MLCC 的核心制造公司村田

493
00:17:24,965 --> 00:17:27,465
从今年2026年4月1号开始

494
00:17:27,633 --> 00:17:30,465
对AI服务器和高端汽车应用的MLCC

495
00:17:30,465 --> 00:17:33,032
涨价15%到35%

496
00:17:33,465 --> 00:17:36,432
另一家公司太阳诱电也从5月份开始

497
00:17:36,500 --> 00:17:38,066
对MLCC、电感

498
00:17:38,066 --> 00:17:40,965
射频器件等多条产品线通知调价

499
00:17:40,965 --> 00:17:42,865
三星电机也被市场传出

500
00:17:42,865 --> 00:17:45,000
考虑上调MLCC的价格

501
00:17:45,400 --> 00:17:48,000
同时根据日本财务省5月底公布的

502
00:17:48,000 --> 00:17:49,232
贸易统计数据

503
00:17:49,400 --> 00:17:51,633
4月份MLCC的平均出口价格

504
00:17:51,633 --> 00:17:53,232
环比上涨了3%

505
00:17:53,400 --> 00:17:55,099
同比上涨了16%

506
00:17:55,299 --> 00:17:57,133
出口量同比增长10%

507
00:17:57,133 --> 00:17:59,165
出口金额同比增长28%

508
00:17:59,500 --> 00:18:00,432
以上就是高盛

509
00:18:00,432 --> 00:18:03,599
对于MLCC成为新内存的核心判断依据

510
00:18:03,833 --> 00:18:04,766
我看完之后

511
00:18:04,766 --> 00:18:07,000
感觉高盛的态度确实非常乐观

512
00:18:07,133 --> 00:18:10,200
他的定调就是MLCC仍然处于早期阶段

513
00:18:10,333 --> 00:18:11,932
认为这条路线还没有走完

514
00:18:11,932 --> 00:18:13,532
不是过度透支的状态

515
00:18:13,700 --> 00:18:15,599
他确实是一个纯粹的多头

516
00:18:15,932 --> 00:18:16,799
我个人觉得

517
00:18:16,799 --> 00:18:18,700
不能把所有投行机构的测算

518
00:18:18,700 --> 00:18:20,799
都当做已经发生的确定事实

519
00:18:20,799 --> 00:18:21,900
去无脑all in

520
00:18:22,165 --> 00:18:23,900
但至少他的这些数据论点

521
00:18:23,900 --> 00:18:26,432
给我们提供了一个非常清晰的验证方向

522
00:18:26,799 --> 00:18:28,133
如果未来几个季度

523
00:18:28,133 --> 00:18:29,599
村田、三星电机

524
00:18:29,599 --> 00:18:32,232
太阳诱电的MLCC出货继续增长

525
00:18:32,299 --> 00:18:33,566
价格继续上涨

526
00:18:33,566 --> 00:18:35,465
产能的利用率继续改善

527
00:18:35,566 --> 00:18:37,700
那么MLCC新内存的叙事

528
00:18:37,732 --> 00:18:39,465
就会得到进一步的强化

529
00:18:39,633 --> 00:18:40,333
反过来

530
00:18:40,333 --> 00:18:42,900
这个叙事本身也一定会被市场纠正

531
00:18:43,299 --> 00:18:44,732
对于我们投资者来说

532
00:18:44,732 --> 00:18:47,266
如果要投资MLCC这一赛道的话

533
00:18:47,266 --> 00:18:49,232
不能只听投行机构怎么说

534
00:18:49,266 --> 00:18:51,432
我们自己必须得进一步做功课

535
00:18:51,432 --> 00:18:53,965
深入了解整个MLCC产业链

536
00:18:54,032 --> 00:18:56,232
很多人第一次接触MLCC

537
00:18:56,232 --> 00:18:58,633
会觉得这玩意儿不就是一颗小电容吗

538
00:18:58,732 --> 00:19:02,799
我们日常用的手机、电脑、电动车里面都有MLCC

539
00:19:02,799 --> 00:19:04,500
应该没什么技术含量吧

540
00:19:04,833 --> 00:19:06,633
这个理解其实并不准确

541
00:19:06,799 --> 00:19:09,665
普通的 MLCC 确实有很多厂商可以做

542
00:19:09,865 --> 00:19:12,500
但是对于高端的AI服务器 MLCC

543
00:19:12,500 --> 00:19:13,732
又是另外一回事

544
00:19:14,000 --> 00:19:15,665
它的规格要求非常高

545
00:19:15,700 --> 00:19:17,665
在固定的封装尺寸之下

546
00:19:17,665 --> 00:19:20,133
如果你想要把 MLCC 的容值做大

547
00:19:20,200 --> 00:19:21,465
主要有三条路径

548
00:19:21,700 --> 00:19:24,400
第一条用更高介电常数的材料

549
00:19:24,400 --> 00:19:27,532
第二条把每一层陶瓷戒指做的更薄

550
00:19:27,665 --> 00:19:30,133
第三条需要堆叠更多的层数

551
00:19:30,465 --> 00:19:32,500
想要把这三点全部做到位

552
00:19:32,500 --> 00:19:33,700
要求非常高

553
00:19:33,865 --> 00:19:35,299
因为介电常数越高

554
00:19:35,299 --> 00:19:37,566
对于材料的稳定性、温度特性

555
00:19:37,566 --> 00:19:40,465
直流偏压特性这些因素就越难控制

556
00:19:40,700 --> 00:19:43,099
很多电容的标称看起来容量很高

557
00:19:43,133 --> 00:19:45,066
但是在真实工作电压之下

558
00:19:45,066 --> 00:19:46,900
有效电容会明显下降

559
00:19:47,099 --> 00:19:48,465
这就是直流偏压

560
00:19:48,799 --> 00:19:51,465
AI服务器的真实工作环境是高温

561
00:19:51,465 --> 00:19:53,500
高功率、高密度、高负载

562
00:19:53,500 --> 00:19:54,766
条件非常苛刻

563
00:19:54,900 --> 00:19:57,299
这就意味着MLCC的飙升参数

564
00:19:57,299 --> 00:19:58,432
好看没什么用

565
00:19:58,599 --> 00:20:01,232
必须得在真实工作状态下稳定才行

566
00:20:01,599 --> 00:20:04,766
而且高端MLCC的陶瓷介质层非常薄

567
00:20:05,000 --> 00:20:05,665
这样一来

568
00:20:05,665 --> 00:20:08,700
击穿、漏电以及寿命的风险就非常高

569
00:20:09,000 --> 00:20:09,799
除此之外

570
00:20:09,799 --> 00:20:11,200
堆叠的层数越来越多

571
00:20:11,200 --> 00:20:13,965
内部缺陷、良率问题的风险就很高

572
00:20:14,299 --> 00:20:16,900
一颗MLCC上千层堆叠在一起

573
00:20:17,000 --> 00:20:18,865
只要其中任何一层有问题

574
00:20:18,865 --> 00:20:20,799
整颗电容都可能被废掉

575
00:20:21,032 --> 00:20:21,865
总体来看

576
00:20:21,865 --> 00:20:24,200
高端MLCC最大的考验就是

577
00:20:24,266 --> 00:20:26,133
能不能在极小的体积里面

578
00:20:26,133 --> 00:20:26,965
把几百层

579
00:20:26,965 --> 00:20:29,599
甚至上千层陶瓷介质和金属电极

580
00:20:29,599 --> 00:20:30,599
稳定的堆叠

581
00:20:30,732 --> 00:20:33,200
并且能够确保长期高良率量产

582
00:20:33,532 --> 00:20:35,833
它是材料、工艺、设备、良率

583
00:20:35,833 --> 00:20:38,032
客户认证、系统设计支持

584
00:20:38,333 --> 00:20:40,799
这些因素共同叠加出来的行业壁垒

585
00:20:41,333 --> 00:20:44,665
接下来我们具体看看MLCC的制作流程

586
00:20:44,665 --> 00:20:45,833
前面我们说过

587
00:20:45,833 --> 00:20:47,000
MLCC看起来

588
00:20:47,000 --> 00:20:49,133
只是一个芝麻粒大小的小方块

589
00:20:49,232 --> 00:20:50,865
但它不是一整块陶瓷

590
00:20:51,066 --> 00:20:53,965
而是由很多陶瓷介质层和金属电极层

591
00:20:53,965 --> 00:20:55,333
反复堆叠起来的

592
00:20:55,599 --> 00:20:57,032
所以它的制作过程

593
00:20:57,165 --> 00:20:59,932
本质上就是把陶瓷层和金属电极层

594
00:21:00,133 --> 00:21:01,165
一层一层做薄

595
00:21:01,165 --> 00:21:02,665
做均匀、做精准

596
00:21:02,799 --> 00:21:05,200
最后再烧结成一个稳定的小方块

597
00:21:05,566 --> 00:21:07,633
具体来看就是这么几个步骤

598
00:21:07,665 --> 00:21:08,865
大家可以暂停一下

599
00:21:08,865 --> 00:21:09,833
截个图

600
00:21:09,833 --> 00:21:13,165
首先制作的第一步就是准备介电粉体

601
00:21:13,400 --> 00:21:16,266
高端MLCC最常用到的强介电材料

602
00:21:16,266 --> 00:21:17,333
就是钛酸钡

603
00:21:17,432 --> 00:21:19,465
它具有非常高的介电常数

604
00:21:19,599 --> 00:21:22,333
能够大幅缩小电容体积并提升容量

605
00:21:22,532 --> 00:21:24,365
而钛酸钡的核心质量

606
00:21:24,532 --> 00:21:26,032
主要是围绕粉体粒径

607
00:21:26,032 --> 00:21:29,099
晶体结构、纯度和杂质控制展开的

608
00:21:29,365 --> 00:21:31,165
比如说粒径是否足够小

609
00:21:31,165 --> 00:21:32,965
粒径分布是否足够窄

610
00:21:33,066 --> 00:21:35,232
高端MLCC要求钛酸钡的粒径

611
00:21:35,232 --> 00:21:36,665
在100纳米以下

612
00:21:36,833 --> 00:21:39,700
最大颗粒不得超过平均粒径的1.5倍

613
00:21:39,965 --> 00:21:40,700
这样一来

614
00:21:40,700 --> 00:21:42,266
既可以在极小的体积里面

615
00:21:42,266 --> 00:21:43,532
实现超大的容量

616
00:21:43,732 --> 00:21:45,099
也可以消除大颗粒

617
00:21:45,099 --> 00:21:46,900
防止在极薄的介质层中

618
00:21:46,900 --> 00:21:48,932
出现局部电场集中的现象

619
00:21:49,165 --> 00:21:50,932
导致器件击穿和漏电

620
00:21:51,333 --> 00:21:53,465
再比如说纯度是否足够高

621
00:21:53,465 --> 00:21:55,066
杂质控制的好不好

622
00:21:55,266 --> 00:21:57,365
纯度要求一般是在99.9%

623
00:21:57,365 --> 00:21:59,766
甚至是99.99%以上

624
00:21:59,965 --> 00:22:01,566
这样就可以防止漏电

625
00:22:01,566 --> 00:22:03,333
延长材料的使用寿命

626
00:22:03,633 --> 00:22:06,665
再比如说钛酸钡的结晶度是否足够高

627
00:22:06,665 --> 00:22:08,333
晶相是否足够稳定

628
00:22:08,432 --> 00:22:10,532
这样最后制作出来的MLCC

629
00:22:10,532 --> 00:22:12,266
就不会出现漏电的情况

630
00:22:12,400 --> 00:22:13,833
电容容量更加稳定

631
00:22:13,833 --> 00:22:14,799
更加耐用

632
00:22:15,032 --> 00:22:17,833
还有就是钛酸钡的核心参杂配方

633
00:22:17,833 --> 00:22:20,865
高端MLCC使用的并非纯粹的钛酸钡

634
00:22:20,932 --> 00:22:22,865
而是根据不同的温度特性

635
00:22:22,900 --> 00:22:24,465
定制参杂的配方粉

636
00:22:24,900 --> 00:22:27,532
掺杂的东西就是一些微量的稀土元素

637
00:22:27,732 --> 00:22:30,432
这样可以让介电温度曲线更加平滑

638
00:22:30,432 --> 00:22:32,099
确保存电能力的稳定

639
00:22:32,165 --> 00:22:34,133
抑制高温下的漏电情况

640
00:22:34,432 --> 00:22:35,665
钛酸钡搞定之后

641
00:22:35,665 --> 00:22:36,766
就是第二步

642
00:22:36,865 --> 00:22:39,365
把钛酸钡粉体做成陶瓷浆料

643
00:22:39,633 --> 00:22:42,032
粉体本身只是很细的陶瓷颗粒

644
00:22:42,099 --> 00:22:43,133
有点像面粉

645
00:22:43,133 --> 00:22:45,032
不能直接一层一层堆起来

646
00:22:45,232 --> 00:22:47,500
下一步要把粉体、溶剂

647
00:22:47,500 --> 00:22:48,965
粘结剂、分散剂

648
00:22:48,965 --> 00:22:51,932
增塑剂以及各种添加剂混合在一起

649
00:22:51,932 --> 00:22:54,299
做成足够均匀稳定的陶瓷浆料

650
00:22:54,432 --> 00:22:56,000
这就像你揉面团一样

651
00:22:56,000 --> 00:22:57,633
钛酸钡就相当于面粉

652
00:22:57,766 --> 00:23:00,000
这些添加剂就相当于附加的水

653
00:23:00,000 --> 00:23:02,432
小苏打、盐、糖之类的东西

654
00:23:02,766 --> 00:23:05,032
这一步听起来是简单的添加搅拌

655
00:23:05,032 --> 00:23:06,400
但是难度非常高

656
00:23:06,700 --> 00:23:08,299
如果浆料分散不均匀

657
00:23:08,299 --> 00:23:09,900
里面有气泡杂质

658
00:23:09,932 --> 00:23:12,232
后面做出来的陶瓷片就会有缺陷

659
00:23:12,599 --> 00:23:13,833
如果粘度控制不好

660
00:23:13,833 --> 00:23:16,000
后面流延的时候厚度就不稳定

661
00:23:16,333 --> 00:23:17,900
同样是钛酸钡粉体

662
00:23:18,032 --> 00:23:19,900
不同厂商做出来的陶瓷浆料

663
00:23:19,900 --> 00:23:20,900
稳定性不同

664
00:23:21,032 --> 00:23:24,099
最后MLCC的良率和性能也会完全不同

665
00:23:24,500 --> 00:23:25,566
接下来第三步

666
00:23:25,566 --> 00:23:28,500
把陶瓷浆料流延成超薄的陶瓷片

667
00:23:28,500 --> 00:23:30,133
留言的大概意思就是

668
00:23:30,133 --> 00:23:31,099
把陶瓷浆料

669
00:23:31,099 --> 00:23:33,333
均匀的涂布在一层载体膜上面

670
00:23:33,333 --> 00:23:36,232
然后烘干形成一张非常薄的陶瓷片

671
00:23:36,532 --> 00:23:37,465
这个陶瓷片

672
00:23:37,465 --> 00:23:39,365
在行业里面通常叫做绿片

673
00:23:39,665 --> 00:23:41,165
这个过程可以理解为

674
00:23:41,165 --> 00:23:43,865
把面糊摊成一张非常薄的煎饼

675
00:23:43,865 --> 00:23:45,965
这个饼的要求也非常苛刻

676
00:23:46,099 --> 00:23:48,266
不能有气泡、裂纹、杂质

677
00:23:48,266 --> 00:23:49,833
厚度必须非常均匀

678
00:23:50,066 --> 00:23:52,633
因为后面MLCC的每一层陶瓷介质

679
00:23:52,633 --> 00:23:53,965
都来自这张绿片

680
00:23:54,299 --> 00:23:57,232
高端MLCC要做高容值、小尺寸

681
00:23:57,299 --> 00:23:59,865
就必须把陶瓷介质层做的非常薄

682
00:24:00,165 --> 00:24:01,833
但是我们刚刚也提到过

683
00:24:01,833 --> 00:24:02,799
陶瓷层越薄

684
00:24:02,799 --> 00:24:04,400
漏电、击穿、破裂

685
00:24:04,400 --> 00:24:06,365
良率下降的风险也就越高

686
00:24:06,500 --> 00:24:08,165
所以这一步非常关键

687
00:24:08,500 --> 00:24:09,665
接下来第四步

688
00:24:09,665 --> 00:24:12,066
就是在刚刚那张超薄的绿片上面

689
00:24:12,066 --> 00:24:13,599
印刷一层金属电极

690
00:24:14,032 --> 00:24:14,900
这层电极

691
00:24:14,900 --> 00:24:17,099
才是MLCC真正用来存电

692
00:24:17,099 --> 00:24:18,400
导电的关键部分

693
00:24:18,400 --> 00:24:20,400
MLCC不是只有陶瓷层

694
00:24:20,432 --> 00:24:22,099
它还需要金属电极层

695
00:24:22,232 --> 00:24:23,633
因为电容的本质就是

696
00:24:23,633 --> 00:24:26,165
两片电极中间夹着陶瓷介质

697
00:24:26,400 --> 00:24:27,766
陶瓷负责存电

698
00:24:27,766 --> 00:24:30,200
电极负责把电荷导进去导出来

699
00:24:30,333 --> 00:24:32,766
两者一层一层交替叠加在一起

700
00:24:32,799 --> 00:24:34,066
才能形成电容

701
00:24:34,465 --> 00:24:36,200
如果用刚才煎饼的比喻

702
00:24:36,200 --> 00:24:37,532
印刷电极这一步

703
00:24:37,532 --> 00:24:40,400
就相当于在煎饼上面刷上一层酱料

704
00:24:40,500 --> 00:24:43,865
目前高端的MLCC主要用的电极材料是镍

705
00:24:44,165 --> 00:24:48,400
过去很多 MLCC 会用钯、银这类贵金属电极

706
00:24:48,400 --> 00:24:49,932
但是它们的价格很贵

707
00:24:49,932 --> 00:24:51,133
用量一旦上去

708
00:24:51,133 --> 00:24:52,200
成本压不住

709
00:24:52,432 --> 00:24:55,465
所以行业慢慢转向使用镍这种贱金属

710
00:24:55,532 --> 00:24:56,665
成本低很多

711
00:24:56,833 --> 00:24:58,900
不过它仍然有一定的规格要求

712
00:24:59,032 --> 00:25:01,232
镍粉的粒径同样必须非常细

713
00:25:01,266 --> 00:25:02,532
要做到纳米级别

714
00:25:02,566 --> 00:25:04,566
而且分布要窄、要均匀

715
00:25:04,566 --> 00:25:06,432
因为电极层也要做的非常薄

716
00:25:06,432 --> 00:25:07,932
能够跟陶瓷层匹配

717
00:25:08,232 --> 00:25:10,099
这样最后做出来的MLCC

718
00:25:10,099 --> 00:25:11,865
就可以确保性能的稳定

719
00:25:12,232 --> 00:25:14,299
接着第五步是多层堆叠

720
00:25:14,500 --> 00:25:16,465
把前面那些印好电极的绿片

721
00:25:16,465 --> 00:25:18,400
一层一层精准的堆叠起来

722
00:25:18,599 --> 00:25:21,333
高端MLCC的层数可能达到几百层

723
00:25:21,333 --> 00:25:22,599
甚至是上千层

724
00:25:23,000 --> 00:25:24,299
这里的难点在于

725
00:25:24,299 --> 00:25:25,965
每一层都要对的非常准

726
00:25:26,200 --> 00:25:27,700
如果堆叠的位置偏了

727
00:25:27,700 --> 00:25:29,799
内部电极重叠面积就会变化

728
00:25:29,799 --> 00:25:31,165
电容值会不稳定

729
00:25:31,400 --> 00:25:34,165
如果层与层之间有灰尘、气泡、褶皱

730
00:25:34,299 --> 00:25:37,000
之后烧结的时候可能会产生内部缺陷

731
00:25:37,099 --> 00:25:38,833
如果绿片太薄、太脆

732
00:25:38,833 --> 00:25:41,865
还可能在搬运、剥离、堆叠的过程中破裂

733
00:25:42,266 --> 00:25:43,365
在堆叠之后

734
00:25:43,365 --> 00:25:45,200
这一摞绿片还是松散的

735
00:25:45,266 --> 00:25:46,700
下一步需要加压

736
00:25:46,700 --> 00:25:48,833
把它们压成结结实实的一整块

737
00:25:49,032 --> 00:25:51,099
这一步叫做层压或者压合

738
00:25:51,333 --> 00:25:53,465
需要在一定的温度和压力之下

739
00:25:53,465 --> 00:25:56,099
把这几百上千层紧紧的压在一起

740
00:25:56,266 --> 00:25:58,299
把层与层之间的空隙和气泡

741
00:25:58,400 --> 00:25:59,566
全部挤出去

742
00:25:59,900 --> 00:26:01,200
这一步看似简单

743
00:26:01,200 --> 00:26:02,500
但是温度、压力

744
00:26:02,500 --> 00:26:05,000
时间这些参数控制仍然很重要

745
00:26:05,165 --> 00:26:08,766
否则会影响后面的烧结良率和长期可靠性

746
00:26:08,965 --> 00:26:09,766
压好之后

747
00:26:09,766 --> 00:26:11,400
下一步就是把这一大块

748
00:26:11,400 --> 00:26:13,000
切成一颗颗小方块

749
00:26:13,299 --> 00:26:15,732
这一步就像把一整块千层蛋糕

750
00:26:15,766 --> 00:26:17,732
切成一块块平整的小蛋糕

751
00:26:17,932 --> 00:26:19,465
要求同样非常精准

752
00:26:19,465 --> 00:26:21,099
切割的位置如果偏了

753
00:26:21,099 --> 00:26:23,799
会影响内部电极和外部端头的连接

754
00:26:24,099 --> 00:26:26,532
切割时如果产生裂纹和毛刺

755
00:26:26,599 --> 00:26:28,465
后面烧结和使用的过程中

756
00:26:28,465 --> 00:26:29,732
可能就会出现问题

757
00:26:30,133 --> 00:26:32,665
随着高端MLCC的尺寸越来越小

758
00:26:32,732 --> 00:26:35,232
对于切割精度的要求也就越来越高

759
00:26:35,500 --> 00:26:38,232
接下来第八步是重头戏高温烧结

760
00:26:38,599 --> 00:26:42,000
通过高温让陶瓷颗粒彼此结合致密化

761
00:26:42,133 --> 00:26:42,532
同时

762
00:26:42,532 --> 00:26:45,266
让内部金属电极形成连续的导电层

763
00:26:45,732 --> 00:26:47,665
这一步是MLCC制造里面

764
00:26:47,665 --> 00:26:49,932
最关键也是最难的环节之一

765
00:26:50,333 --> 00:26:52,232
难点并不在于单纯加热

766
00:26:52,232 --> 00:26:54,900
而是需要严密控制温度和氧气的参数

767
00:26:55,232 --> 00:26:56,032
温度太低

768
00:26:56,032 --> 00:26:57,200
陶瓷不够致密

769
00:26:57,266 --> 00:26:59,200
介电性能和机械强度不够

770
00:26:59,400 --> 00:27:00,432
温度太高的话

771
00:27:00,432 --> 00:27:01,900
晶粒可能异常长大

772
00:27:01,900 --> 00:27:03,066
导致漏电、击穿

773
00:27:03,066 --> 00:27:04,032
寿命下降

774
00:27:04,500 --> 00:27:05,299
氧气太多

775
00:27:05,299 --> 00:27:06,732
电极可能会被氧化

776
00:27:06,833 --> 00:27:07,833
氧气太少的话

777
00:27:07,833 --> 00:27:09,766
钛酸钡陶瓷可能会被还原

778
00:27:09,766 --> 00:27:10,900
电性能变差

779
00:27:11,333 --> 00:27:12,432
烧结之后

780
00:27:12,432 --> 00:27:15,232
下一步需要做的就是外部电极和电镀

781
00:27:15,465 --> 00:27:18,032
MLCC内部虽然已经有很多层电极

782
00:27:18,333 --> 00:27:20,833
但还是需要在两端做出外部电极

783
00:27:20,932 --> 00:27:23,133
方便焊接到PCB主板上面

784
00:27:23,133 --> 00:27:25,566
之后还要进行镀镍、镀锡的处理

785
00:27:25,700 --> 00:27:27,365
提高焊接性和可靠性

786
00:27:27,799 --> 00:27:30,532
最后一步就是测试、分选和包装

787
00:27:30,700 --> 00:27:32,665
每一颗MLCC都要测试

788
00:27:32,665 --> 00:27:34,799
比如说容量、耐压性、漏电

789
00:27:34,799 --> 00:27:38,200
绝缘电阻、ESL、温度特性、直流偏压特性

790
00:27:38,200 --> 00:27:39,200
以及可靠性

791
00:27:39,566 --> 00:27:41,500
高端AI服务器用的 MLCC

792
00:27:41,500 --> 00:27:44,200
还要经过更严格的筛选和客户验证

793
00:27:44,400 --> 00:27:45,599
同样非常重要

794
00:27:45,932 --> 00:27:46,732
到此为止

795
00:27:46,732 --> 00:27:49,665
MLCC 的完整制作流程就全部讲完了

796
00:27:49,932 --> 00:27:52,000
接下来我们顺着制作流程

797
00:27:52,000 --> 00:27:54,365
看看产业链中有哪些瓶颈环节

798
00:27:54,799 --> 00:27:56,365
按照制作流程来看

799
00:27:56,365 --> 00:27:59,400
MLCC产业链上有三个核心的环节

800
00:27:59,766 --> 00:28:01,865
第一个是上游的原材料

801
00:28:01,900 --> 00:28:03,299
包括钛酸钡粉体

802
00:28:03,299 --> 00:28:04,566
镍粉、铜粉

803
00:28:04,566 --> 00:28:08,000
稀土添加剂、溶剂、粘结剂、分散剂等等

804
00:28:08,099 --> 00:28:10,232
其中最关键的是介电粉体

805
00:28:10,232 --> 00:28:12,032
也就是前面讲到的钛酸钡

806
00:28:12,333 --> 00:28:15,665
高端 MLCC 越往小尺寸、高容值方向走

807
00:28:15,833 --> 00:28:18,032
对于钛酸钡粉体的要求就越高

808
00:28:18,266 --> 00:28:20,665
这一环节有一些独立材料供应商

809
00:28:20,865 --> 00:28:24,732
比如堺化学工业、日本化学工业、富士钛工业等等

810
00:28:24,865 --> 00:28:27,099
第二个环节是专用制造设备

811
00:28:27,333 --> 00:28:28,900
MLCC的制造过程中

812
00:28:28,900 --> 00:28:30,633
需要大量的专用设备

813
00:28:30,700 --> 00:28:32,232
包括浆料制造设备

814
00:28:32,232 --> 00:28:33,099
流延设备

815
00:28:33,099 --> 00:28:34,232
精密印刷设备

816
00:28:34,232 --> 00:28:37,032
层压设备、切割设备、烧结设备

817
00:28:37,032 --> 00:28:39,333
端电极设备、测试分选设备

818
00:28:39,700 --> 00:28:40,865
这些设备公司

819
00:28:40,865 --> 00:28:43,865
必定会受益于高端MLCC的扩展周期

820
00:28:44,266 --> 00:28:45,232
第三个环节

821
00:28:45,232 --> 00:28:47,500
也是整个产业链最核心的位置

822
00:28:47,700 --> 00:28:49,432
MLCC本体制造

823
00:28:49,532 --> 00:28:50,099
这个环节

824
00:28:50,099 --> 00:28:51,365
要把前面的材料

825
00:28:51,365 --> 00:28:53,865
浆料设备、工艺全部打通

826
00:28:53,965 --> 00:28:56,133
最终做出可量产、可交付

827
00:28:56,133 --> 00:28:58,200
可以认证的MLCC

828
00:28:58,200 --> 00:29:01,299
高端AI服务器 MLCC 最难的地方在于

829
00:29:01,299 --> 00:29:03,365
既要小尺寸又要高容值

830
00:29:03,365 --> 00:29:05,365
还要低ESL、高可靠

831
00:29:05,432 --> 00:29:07,365
并且能够长期稳定量产

832
00:29:07,732 --> 00:29:10,200
所以高端MLCC制造厂商

833
00:29:10,200 --> 00:29:12,266
就成为了整条产业链的核心

834
00:29:12,599 --> 00:29:13,232
这也对应了

835
00:29:13,232 --> 00:29:16,099
我们今天视频要重点讲到的三家公司

836
00:29:16,232 --> 00:29:19,066
村田制作所、三星电机和太阳诱电

837
00:29:19,400 --> 00:29:19,932
接下来

838
00:29:19,932 --> 00:29:22,500
我就一家一家详细给大家介绍一遍

839
00:29:22,500 --> 00:29:23,732
首先是村田

840
00:29:23,732 --> 00:29:25,766
它是全球MLCC行业里面

841
00:29:25,766 --> 00:29:27,465
最核心的龙头公司

842
00:29:27,732 --> 00:29:30,365
它的MLCC份额全球排名第一

843
00:29:30,365 --> 00:29:31,833
客户覆盖非常广

844
00:29:32,066 --> 00:29:34,133
不同机构的口径可能会有差异

845
00:29:34,165 --> 00:29:35,266
但总体来看

846
00:29:35,266 --> 00:29:38,365
村田长期都是全球MLCC行业的第一名

847
00:29:38,732 --> 00:29:41,333
市场份额大概在40%左右

848
00:29:41,432 --> 00:29:44,000
在汽车MLCC这种高可靠领域

849
00:29:44,000 --> 00:29:45,333
份额也非常高

850
00:29:45,532 --> 00:29:47,766
这说明村田不只是消费电子里面

851
00:29:47,766 --> 00:29:49,032
有规模化的优势

852
00:29:49,200 --> 00:29:51,965
在汽车、工业、服务器、数据中心

853
00:29:51,965 --> 00:29:54,333
这些对可靠性要求更高的应用里面

854
00:29:54,400 --> 00:29:56,066
也有非常深的客户基础

855
00:29:56,566 --> 00:29:57,965
对于AI服务器来说

856
00:29:57,965 --> 00:29:59,932
客户最看重的不只是价格

857
00:29:59,932 --> 00:30:01,965
而是长期稳定供货的能力

858
00:30:01,965 --> 00:30:02,932
可靠性数据

859
00:30:02,932 --> 00:30:05,232
客户认证经验和工程支持能力

860
00:30:05,432 --> 00:30:07,633
AI服务器里面的 MLCC

861
00:30:07,633 --> 00:30:11,732
尤其是 GPU、HBM、ASIC 近端的高端 MLCC

862
00:30:11,833 --> 00:30:14,532
不是随便找一家供应商就能够替代的

863
00:30:14,799 --> 00:30:16,965
因为这些电容不是孤立存在的

864
00:30:17,000 --> 00:30:18,032
而是需要融合到

865
00:30:18,032 --> 00:30:20,000
整套供电网络的设计里面

866
00:30:20,333 --> 00:30:22,000
客户一旦基于村田的料号

867
00:30:22,000 --> 00:30:26,266
完成 PDN 仿真、PCB 打板、整机测试和量产认证

868
00:30:26,266 --> 00:30:28,799
后续切换供应商的成本非常高

869
00:30:29,000 --> 00:30:31,066
这就是村田的客户锁定能力

870
00:30:31,465 --> 00:30:33,400
其次村田在高容值

871
00:30:33,400 --> 00:30:36,232
小尺寸的产品上面有非常明显的优势

872
00:30:36,665 --> 00:30:39,566
AI服务器最缺的不是普通的 MLCC

873
00:30:39,566 --> 00:30:41,165
而是在有限空间里面

874
00:30:41,165 --> 00:30:44,133
提供更大有效电容的高端MLCC

875
00:30:44,133 --> 00:30:47,432
GPU、HBM、ASIC 周围的空间非常紧张

876
00:30:47,566 --> 00:30:50,133
工程师希望电容尽可能靠近芯片

877
00:30:50,333 --> 00:30:51,566
电容越靠近芯片

878
00:30:51,566 --> 00:30:52,665
寄生电感越低

879
00:30:52,665 --> 00:30:54,032
响应速度就越快

880
00:30:54,266 --> 00:30:58,865
村田在 2024 年发布了 0603 英寸 100μF MLCC

881
00:30:59,000 --> 00:30:59,932
官方明确说

882
00:30:59,932 --> 00:31:01,066
这是面向服务器

883
00:31:01,066 --> 00:31:04,932
数据中心和 IT 应用持续小型化需求的产品

884
00:31:04,932 --> 00:31:06,066
到了2025年

885
00:31:06,066 --> 00:31:10,400
村田又开始量产 0402 英寸 47μF MLCC

886
00:31:10,633 --> 00:31:12,700
并强调MLCC小型化

887
00:31:12,700 --> 00:31:15,432
高容值、高可靠的需求正在不断提升

888
00:31:15,833 --> 00:31:18,865
还有就是村田的材料和工艺壁垒很深

889
00:31:19,200 --> 00:31:20,333
前面我们讲过

890
00:31:20,333 --> 00:31:22,799
MLCC制造涉及到介电粉体

891
00:31:22,799 --> 00:31:24,500
陶瓷浆料、流延绿片

892
00:31:24,500 --> 00:31:28,066
内部电极印刷、堆叠压合、切割、烧结

893
00:31:28,133 --> 00:31:31,599
外部电极、电镀、测试分选等多种环节

894
00:31:31,799 --> 00:31:32,532
任何一个环节

895
00:31:32,532 --> 00:31:35,333
出现问题都会影响最终性能和良率

896
00:31:35,599 --> 00:31:36,865
村田真正的壁垒

897
00:31:36,865 --> 00:31:38,700
就是把这些环节全部打通

898
00:31:38,833 --> 00:31:40,465
并且长期大规模量产

899
00:31:40,865 --> 00:31:43,066
他不是只会做某一个工艺步骤

900
00:31:43,099 --> 00:31:44,665
而是从材料到制造

901
00:31:44,665 --> 00:31:45,833
再到客户导入

902
00:31:45,865 --> 00:31:48,032
形成了一整套垂直整合闭环

903
00:31:48,333 --> 00:31:49,099
这就是为什么

904
00:31:49,099 --> 00:31:50,932
很多公司买了一整套设备

905
00:31:51,032 --> 00:31:54,099
也很难快速复制村田的高端MLCC产能

906
00:31:54,099 --> 00:31:56,400
最后村田的 PDN 设计支持

907
00:31:56,532 --> 00:31:59,833
是一个非常容易被市场忽略的隐藏加分项

908
00:31:59,833 --> 00:32:02,500
AI服务器客户买高端的 MLCC

909
00:32:02,500 --> 00:32:04,099
不只是买一颗小电容

910
00:32:04,200 --> 00:32:05,165
他们真正需要的

911
00:32:05,165 --> 00:32:08,032
是一颗可以进入 PDN 设计体系的器件

912
00:32:08,432 --> 00:32:11,633
PDN 的英文全称是 Power Delivery Network

913
00:32:11,633 --> 00:32:13,000
也就是供电网络

914
00:32:13,032 --> 00:32:14,333
在AI服务器里面

915
00:32:14,333 --> 00:32:15,932
供电网络是一个多频段

916
00:32:15,932 --> 00:32:18,200
多层级、多约束的复杂系统

917
00:32:18,599 --> 00:32:20,066
同样100克MLCC

918
00:32:20,066 --> 00:32:21,000
不同的位置

919
00:32:21,000 --> 00:32:22,833
不同的分装、不同的走线

920
00:32:22,865 --> 00:32:24,165
不同的电源平面

921
00:32:24,232 --> 00:32:26,566
最后产生的效果可能完全不一样

922
00:32:26,865 --> 00:32:29,299
村田在这种背景下的优势就是

923
00:32:29,299 --> 00:32:31,299
它除了制造MLCC之外

924
00:32:31,299 --> 00:32:33,833
它会给客户提供高精度仿真数据

925
00:32:33,833 --> 00:32:35,400
免费的设计辅助软件

926
00:32:35,400 --> 00:32:37,599
以及定制化的调优咨询服务

927
00:32:37,900 --> 00:32:38,799
综合来看

928
00:32:38,799 --> 00:32:41,766
村田在MLCC赛道上最强大的护城河

929
00:32:41,932 --> 00:32:44,200
就是它深度掌握了从材料体系

930
00:32:44,200 --> 00:32:48,700
陶瓷绿片、内部电极印刷、层压、烧结、测试筛选

931
00:32:48,865 --> 00:32:52,165
到客户认证和系统设计支持的一整套能力

932
00:32:52,165 --> 00:32:55,333
是MLCC产业链上确定性最高的公司

933
00:32:55,865 --> 00:32:58,133
如果AI服务器继续向更高功耗

934
00:32:58,133 --> 00:32:59,799
更复杂的 PDN 演进

935
00:32:59,965 --> 00:33:02,532
村田大概率是最稳定的核心受益者

936
00:33:02,532 --> 00:33:05,365
接下来我们继续看三星电机这家公司

937
00:33:05,500 --> 00:33:07,566
它是AI服务器 MLCC 里面

938
00:33:07,566 --> 00:33:10,165
最值得重视的规模化进攻型玩家

939
00:33:10,465 --> 00:33:12,000
整体 MLCC 市场里

940
00:33:12,000 --> 00:33:14,365
它是全球第二阵营的核心厂商

941
00:33:14,532 --> 00:33:16,200
市占率超过了20%

942
00:33:16,566 --> 00:33:18,732
在AI服务器这个细分市场里面

943
00:33:18,732 --> 00:33:20,365
三星电机官方宣称

944
00:33:20,365 --> 00:33:23,232
自己已经拥有大约40%的市场份额

945
00:33:23,500 --> 00:33:26,900
三星电机的第一大优势是规模制造能力很强

946
00:33:27,165 --> 00:33:29,200
MLCC 是一个非常吃良率

947
00:33:29,200 --> 00:33:31,099
产能稳定交付的行业

948
00:33:31,500 --> 00:33:33,965
AI服务器客户需要长期、大批量

949
00:33:33,965 --> 00:33:35,532
高一致性的供货能力

950
00:33:35,665 --> 00:33:38,599
而三星电机背靠韩国的电子产业链

951
00:33:38,633 --> 00:33:40,932
在消费电子、IT、服务器、

952
00:33:40,932 --> 00:33:43,066
汽车电子这些领域的积累很深

953
00:33:43,165 --> 00:33:44,532
量产能力非常强

954
00:33:44,932 --> 00:33:47,732
这一点在AI服务器放量阶段非常重要

955
00:33:47,932 --> 00:33:49,200
公开的资料显示

956
00:33:49,200 --> 00:33:52,000
三星电机负责MLCC的工厂利用率

957
00:33:52,032 --> 00:33:54,833
已经从2023年第一季度的59%

958
00:33:54,965 --> 00:33:57,333
上升到2024年超过80%

959
00:33:57,665 --> 00:34:00,833
并且在2025年第三季度达到了99%

960
00:34:01,000 --> 00:34:02,465
接近满产的状态

961
00:34:02,732 --> 00:34:03,965
它的第二大优势

962
00:34:03,965 --> 00:34:06,532
就是AI服务器产品线的扩张很快

963
00:34:06,932 --> 00:34:08,733
AI服务器需要的MLCC

964
00:34:08,733 --> 00:34:10,766
是一整套围绕电源、计算

965
00:34:10,766 --> 00:34:12,400
网络通信的产品组合

966
00:34:12,733 --> 00:34:16,565
比如说电源侧需要高压、高容量、高可靠的MLCC

967
00:34:16,865 --> 00:34:18,800
GPU CPU ASIC附近

968
00:34:18,800 --> 00:34:21,932
需要低压大容量、低 ESL 的 MLCC

969
00:34:22,166 --> 00:34:23,865
网卡交换芯片附近

970
00:34:23,932 --> 00:34:25,800
需要能够抑制高频噪声

971
00:34:25,800 --> 00:34:27,932
维持信号完整性的MLCC

972
00:34:28,233 --> 00:34:29,032
三星电机

973
00:34:29,032 --> 00:34:31,532
现在明显在往这些方向上扩张

974
00:34:31,666 --> 00:34:32,733
它的产品逻辑

975
00:34:32,733 --> 00:34:35,032
非常贴合AI服务器的真实需求

976
00:34:35,432 --> 00:34:36,233
除此之外

977
00:34:36,233 --> 00:34:38,500
三星电机不只做 MLCC

978
00:34:38,500 --> 00:34:41,266
也做分装机板、摄像头模组这些业务

979
00:34:41,432 --> 00:34:44,365
所以说它不是最纯正的 MLCC 标的

980
00:34:44,865 --> 00:34:47,865
但是从AI服务器 MLCC 规模化放量

981
00:34:47,865 --> 00:34:48,900
这个角度来看

982
00:34:49,099 --> 00:34:51,800
三星电机必须进入重点的观察名单

983
00:34:51,800 --> 00:34:53,800
最后是太阳诱电这家公司

984
00:34:54,000 --> 00:34:55,833
如果说村田代表确定性

985
00:34:55,833 --> 00:34:59,766
三星电机代表 AI 服务器 MLCC 的规模化进攻

986
00:34:59,766 --> 00:35:01,900
那么太阳诱电代表的就是弹性

987
00:35:02,500 --> 00:35:03,565
它的第一个优势

988
00:35:03,565 --> 00:35:06,065
是高端MLCC业务占比很高

989
00:35:06,065 --> 00:35:08,199
周期向上的利润弹性更大

990
00:35:08,432 --> 00:35:10,233
相比起村田和三星电机

991
00:35:10,233 --> 00:35:12,599
这种业务更加综合的平台型公司

992
00:35:12,865 --> 00:35:13,500
太阳诱电

993
00:35:13,500 --> 00:35:15,666
对于电容业务的依赖程度更高

994
00:35:15,833 --> 00:35:18,400
这意味着如果MLCC的价格上涨

995
00:35:18,400 --> 00:35:19,666
产品结构升级

996
00:35:19,666 --> 00:35:21,000
产能利用率提升

997
00:35:21,233 --> 00:35:23,632
太阳诱电的利润改善会更加明显

998
00:35:23,900 --> 00:35:25,032
这也是为什么

999
00:35:25,032 --> 00:35:27,766
市场一旦开始炒作MLCC周期

1000
00:35:27,833 --> 00:35:29,300
太阳诱电的股价弹性

1001
00:35:29,300 --> 00:35:31,000
往往会比村田更大

1002
00:35:31,166 --> 00:35:33,500
因为太阳浴电的体量相对更小

1003
00:35:33,500 --> 00:35:34,766
业务更加集中

1004
00:35:34,965 --> 00:35:36,632
所以周期向上的时候

1005
00:35:36,632 --> 00:35:37,632
他更容易被市场

1006
00:35:37,632 --> 00:35:39,699
当成高弹性的标的来交易

1007
00:35:40,032 --> 00:35:41,900
当然这也是一把双刃剑

1008
00:35:42,365 --> 00:35:43,400
它的第二个优势

1009
00:35:43,400 --> 00:35:47,300
是低 ESL 产品非常符合 AI GPU 的去耦需求

1010
00:35:47,632 --> 00:35:48,900
前面我们讲到过

1011
00:35:48,900 --> 00:35:51,833
AI GPU的供电难点是低电压、大电流

1012
00:35:51,833 --> 00:35:53,465
快速的负载变化

1013
00:35:53,465 --> 00:35:55,465
GPU不是一直稳定的耗电

1014
00:35:55,599 --> 00:35:59,632
而是会随着训练、推理、HBM 访问、NVLink 通信

1015
00:35:59,632 --> 00:36:02,099
在极短时间里出现负载的变化

1016
00:36:02,233 --> 00:36:06,132
这就要求靠近GPU HBM ASIC的MLCC

1017
00:36:06,233 --> 00:36:07,865
必须能够快速响应高频

1018
00:36:07,865 --> 00:36:09,233
瞬态电流需求

1019
00:36:09,632 --> 00:36:11,099
这时候ESL越低

1020
00:36:11,099 --> 00:36:14,099
电容就越能够快速响应高频电流的变化

1021
00:36:14,333 --> 00:36:17,000
AI服务器里面GPU附近的驱有电容

1022
00:36:17,000 --> 00:36:18,532
最怕的就是反应慢

1023
00:36:18,699 --> 00:36:20,266
太阳诱电在低 ESL

1024
00:36:20,266 --> 00:36:23,666
反向电极 MLCC 这些方向上的布局比较积极

1025
00:36:24,000 --> 00:36:27,132
它的第三个优势就是 Embedded MLCC

1026
00:36:27,132 --> 00:36:29,233
也就是机板内置 MLCC

1027
00:36:29,300 --> 00:36:32,099
是我觉得最值得重点跟踪的一个方向

1028
00:36:32,099 --> 00:36:34,900
传统 MLCC 是贴在 PCB 表面的

1029
00:36:35,000 --> 00:36:37,400
但AI芯片周围的空间越来越紧张

1030
00:36:37,400 --> 00:36:39,699
同时去耦电容离芯片越远

1031
00:36:39,733 --> 00:36:40,800
寄生电感越高

1032
00:36:40,800 --> 00:36:42,233
响应速度就越慢

1033
00:36:42,500 --> 00:36:43,632
在这种情况下

1034
00:36:43,632 --> 00:36:45,333
行业的趋势非常明确

1035
00:36:45,400 --> 00:36:47,099
就是把电容放的更近

1036
00:36:47,099 --> 00:36:48,400
从PCB的正面

1037
00:36:48,400 --> 00:36:49,599
到PCB的背面

1038
00:36:49,599 --> 00:36:51,233
再到封装基板的附近

1039
00:36:51,266 --> 00:36:53,266
甚至直接嵌入基板的内部

1040
00:36:53,365 --> 00:36:54,632
核心目的只有一个

1041
00:36:54,632 --> 00:36:55,965
降低寄生电感

1042
00:36:56,000 --> 00:36:57,632
让电容能够更快地响应

1043
00:36:57,632 --> 00:36:59,465
芯片瞬态电流的需求

1044
00:36:59,465 --> 00:37:02,400
这个逻辑其实和光互联里面的CPU非常像

1045
00:37:02,599 --> 00:37:04,599
CPU是让光靠近芯片

1046
00:37:04,833 --> 00:37:07,599
Embedded MLCC 是让电容靠近芯片

1047
00:37:07,800 --> 00:37:10,932
二者的核心逻辑就是越高频、越高速

1048
00:37:10,932 --> 00:37:11,932
功耗越大

1049
00:37:11,932 --> 00:37:14,932
越要让关键的部件尽可能离芯片更近

1050
00:37:15,166 --> 00:37:16,833
太阳诱电已经开始量产

1051
00:37:16,833 --> 00:37:21,565
面向AI服务器这些高性能计算芯片的基板内置型MLCC

1052
00:37:21,565 --> 00:37:25,099
其中代表性产品0402 英寸 22μF

1053
00:37:25,132 --> 00:37:27,266
以及 0805 英寸 100μF

1054
00:37:27,500 --> 00:37:29,532
这个方向非常值得

1055
00:37:29,532 --> 00:37:33,500
因为未来 AI ASIC、GPU、HBM 的封装和基板设计

1056
00:37:33,500 --> 00:37:34,733
会越来越复杂

1057
00:37:34,800 --> 00:37:37,065
芯片周围的空间会越来越紧张

1058
00:37:37,166 --> 00:37:38,199
在这种趋势下

1059
00:37:38,199 --> 00:37:39,500
内置型 MLCC

1060
00:37:39,500 --> 00:37:41,465
很可能成为高端 MLCC 里面

1061
00:37:41,465 --> 00:37:43,965
最具瓶颈属性的细分方向之一

1062
00:37:44,532 --> 00:37:46,599
最后我们把三家公司放在一起

1063
00:37:46,599 --> 00:37:48,166
对比一下它们的优缺点

1064
00:37:48,166 --> 00:37:49,532
做个简单的总结

1065
00:37:49,932 --> 00:37:52,166
村田、三星电机、太阳诱电

1066
00:37:52,199 --> 00:37:54,532
代表了MLCC赛道里面三种风格

1067
00:37:54,532 --> 00:37:55,865
不同的投资逻辑

1068
00:37:56,166 --> 00:37:57,865
村田强在确定性

1069
00:37:57,965 --> 00:38:01,166
它是全球MLCC行业里面最核心的龙头

1070
00:38:01,266 --> 00:38:04,400
优势在于全球份额、客户覆盖、材料工艺

1071
00:38:04,400 --> 00:38:07,333
产能稳定性和 PDN 设计支持

1072
00:38:07,333 --> 00:38:09,532
它不一定是 MLCC 纯度最高

1073
00:38:09,532 --> 00:38:10,800
弹性最大的标的

1074
00:38:10,965 --> 00:38:13,000
但它的确定性一定是最高的

1075
00:38:13,333 --> 00:38:14,432
三星电机强在

1076
00:38:14,432 --> 00:38:17,532
AI服务器 MLCC 产品上的规模化进攻

1077
00:38:17,865 --> 00:38:19,865
它在整体 MLCC 市场里面

1078
00:38:19,865 --> 00:38:21,800
是全球第二的核心厂商

1079
00:38:22,000 --> 00:38:25,666
在 AI 服务器高端 MLCC 方向上非常积极

1080
00:38:25,666 --> 00:38:27,599
产品线围绕电源、计算

1081
00:38:27,599 --> 00:38:29,833
网络通信这些真实的需求展开

1082
00:38:30,065 --> 00:38:32,065
它的问题在于业务比较综合

1083
00:38:32,199 --> 00:38:34,166
不是最纯的MLCC标的

1084
00:38:34,300 --> 00:38:37,065
所以股价不一定完全跟MLCC对应

1085
00:38:37,300 --> 00:38:39,000
太阳诱电强在弹性

1086
00:38:39,166 --> 00:38:41,965
它的优势是MLCC业务占比非常高

1087
00:38:42,000 --> 00:38:44,300
所以MLCC周期向上的时候

1088
00:38:44,300 --> 00:38:45,965
利润弹性更加明显

1089
00:38:46,266 --> 00:38:48,032
同时它在低 ESL

1090
00:38:48,032 --> 00:38:52,233
基板内置型 MLCC 这些高端细分方向上有先发优势

1091
00:38:52,565 --> 00:38:53,965
潜在的问题就是

1092
00:38:53,965 --> 00:38:55,766
它目前的份额以及规模

1093
00:38:55,766 --> 00:38:58,432
都不如村田和三星电机这两家公司

1094
00:38:58,833 --> 00:38:59,432
所以

1095
00:38:59,432 --> 00:39:02,166
如果你准备投资 MLCC 这个赛道的话

1096
00:39:02,166 --> 00:39:04,632
需要提前想清楚自己的收益预期

1097
00:39:04,632 --> 00:39:06,365
和可以承担的风险波动

1098
00:39:06,632 --> 00:39:07,800
如果你追求稳健

1099
00:39:07,800 --> 00:39:08,833
追求确定性

1100
00:39:08,833 --> 00:39:10,199
村田是最合适的

1101
00:39:10,500 --> 00:39:12,233
如果想要押注AI服务器

1102
00:39:12,233 --> 00:39:14,166
MLCC 的大规模放量

1103
00:39:14,233 --> 00:39:16,666
三星电机更值得重点跟踪

1104
00:39:16,833 --> 00:39:18,632
如果想要追求更高的弹性

1105
00:39:18,632 --> 00:39:20,233
可以承担更高的风险

1106
00:39:20,266 --> 00:39:21,699
太阳诱电更合适

1107
00:39:21,833 --> 00:39:24,432
讲完这三家公司以及投资框架之后

1108
00:39:24,432 --> 00:39:25,833
在视频的最后一部分

1109
00:39:25,833 --> 00:39:27,699
我们回到开头的那个问题

1110
00:39:27,900 --> 00:39:30,400
MLCC到底是不是下一个内存

1111
00:39:30,632 --> 00:39:33,032
说实话我觉得现在很难判断

1112
00:39:33,099 --> 00:39:35,365
毕竟内存和存储这一轮主声浪

1113
00:39:35,365 --> 00:39:37,132
已经走了接近一年的时间

1114
00:39:37,233 --> 00:39:39,365
到现在似乎并没有结束的迹象

1115
00:39:39,666 --> 00:39:40,865
而且大家别忘了

1116
00:39:40,865 --> 00:39:42,965
在内存和存储上涨的过程中

1117
00:39:43,032 --> 00:39:45,266
市场上的质疑声音也一直非常多

1118
00:39:45,266 --> 00:39:46,500
但这些并没有妨碍

1119
00:39:46,500 --> 00:39:48,065
相关的股票越涨越高

1120
00:39:48,065 --> 00:39:49,766
一年时间翻了十几倍

1121
00:39:50,032 --> 00:39:51,500
所以回到 MLCC

1122
00:39:51,500 --> 00:39:52,865
如果只看炒作热度

1123
00:39:52,865 --> 00:39:55,032
市场认知和资金的关注程度

1124
00:39:55,199 --> 00:39:56,266
我觉得 MLCC

1125
00:39:56,266 --> 00:39:58,766
现在确实有点像早期的内存叙事

1126
00:39:59,032 --> 00:40:01,166
它刚刚从一个传统的被动元件

1127
00:40:01,166 --> 00:40:02,632
被市场重新放进AI

1128
00:40:02,632 --> 00:40:04,233
算力瓶颈的框架里面

1129
00:40:04,565 --> 00:40:07,099
高盛这些机构开始把它叫做新内存

1130
00:40:07,365 --> 00:40:10,032
市场开始讨论高端MLCC涨价

1131
00:40:10,032 --> 00:40:12,666
讨论下一代 Rubin 机柜价值量提升

1132
00:40:12,833 --> 00:40:14,865
说明MLCC这个赛道

1133
00:40:14,865 --> 00:40:16,965
已经开始被资金重新定价了

1134
00:40:17,166 --> 00:40:18,132
问题就在于

1135
00:40:18,132 --> 00:40:20,266
这轮炒作到底能够持续多久

1136
00:40:20,365 --> 00:40:22,032
它有可能走出一条真正的

1137
00:40:22,032 --> 00:40:23,666
大级别AI硬件主线

1138
00:40:23,766 --> 00:40:25,266
也有可能只是阶段性的

1139
00:40:25,266 --> 00:40:25,965
主题扩散

1140
00:40:25,965 --> 00:40:26,932
情绪扩散

1141
00:40:27,266 --> 00:40:28,632
所以面对这种赛道

1142
00:40:28,632 --> 00:40:30,166
最舒服的参与方式

1143
00:40:30,333 --> 00:40:30,865
永远不是

1144
00:40:30,865 --> 00:40:32,699
等到市场所有人都相信的时候

1145
00:40:32,699 --> 00:40:33,632
再追进去

1146
00:40:33,865 --> 00:40:35,065
要么先信先布局

1147
00:40:35,065 --> 00:40:36,500
小仓位试一试

1148
00:40:36,699 --> 00:40:38,266
要么从头到尾就不要信

1149
00:40:38,266 --> 00:40:39,333
永远不要买入

1150
00:40:39,532 --> 00:40:41,166
最危险的就是嘴上不信

1151
00:40:41,166 --> 00:40:42,733
心里一直害怕错过

1152
00:40:42,766 --> 00:40:44,666
等到最后市场极度疯狂的情况下

1153
00:40:44,666 --> 00:40:45,465
冲进去

1154
00:40:45,632 --> 00:40:48,300
这才是普通投资者最容易亏钱的地方

1155
00:40:48,666 --> 00:40:49,800
如果大家想要知道

1156
00:40:49,800 --> 00:40:52,465
我关于MLCC赛道的实时投资操作

1157
00:40:52,465 --> 00:40:53,932
以及最新观点的话

1158
00:40:53,932 --> 00:40:55,965
欢迎加入我的 Patreon 会员社群

1159
00:40:56,000 --> 00:40:57,766
我会持续通过文字的形式

1160
00:40:57,766 --> 00:41:00,166
分享最新最及时的深度内容

1161
00:41:00,333 --> 00:41:00,800
好了

1162
00:41:00,800 --> 00:41:02,599
以上就是今天视频的全部内容了

1163
00:41:02,599 --> 00:41:05,166
欢迎大家点赞留言订阅并开小铃铛

1164
00:41:05,300 --> 00:41:06,599
这是Nike前沿Alpha

1165
00:41:06,599 --> 00:41:08,199
力求用最通俗易懂的语言

1166
00:41:08,199 --> 00:41:10,000
拆解分析最前沿的赛道

1167
00:41:10,032 --> 00:41:11,333
挖掘下一个Alpha

1168
00:41:11,432 --> 00:41:12,532
新频道更新不易

1169
00:41:12,532 --> 00:41:13,833
还请大家多多支持

1170
00:41:13,833 --> 00:41:15,032
我们下周再见

共 4 条，按点赞数排序。

采集完成，该视频无弹幕