暴涨3X只是开始，这三家公司即将爆发，做好准备!揭秘 100 秒量产噩梦背后的真金白银，错过这波再等十年！TER,FORM,KEYS#trading #投資

1
00:00:00,180 --> 00:00:01,459
4875

2
00:00:01,459 --> 00:00:05,450
这是CPU龙头AXTI从低点到高点的涨幅

3
00:00:05,450 --> 00:00:07,209
如果你觉得这只是个例

4
00:00:07,209 --> 00:00:10,669
再看看LITE累计涨幅1330

5
00:00:10,669 --> 00:00:13,009
哪怕是2026年初到现在

6
00:00:13,009 --> 00:00:16,530
短线爆发的AAOI也是直接翻了三倍多

7
00:00:16,530 --> 00:00:17,690
看到这些数字

8
00:00:17,690 --> 00:00:20,269
你的第一反应是不是想立刻冲进去

9
00:00:20,269 --> 00:00:22,329
但我反而建议你先别急

10
00:00:22,329 --> 00:00:25,739
因为现在的CPU赛道不仅是恐高者的噩梦

11
00:00:25,739 --> 00:00:27,420
更是接飞刀的高发区

12
00:00:27,420 --> 00:00:30,620
当全世界都在盯着那些已经长到天上去的

13
00:00:30,620 --> 00:00:32,789
光模块激光器整机厂时

14
00:00:32,789 --> 00:00:35,090
聪明的人已经开始套现离场

15
00:00:35,090 --> 00:00:38,560
去寻找下一个还没有被聚光灯照到的死角

16
00:00:38,560 --> 00:00:39,799
因为当最稀缺

17
00:00:39,799 --> 00:00:42,939
最核心的部件享受完最高的估值和涨幅后

18
00:00:42,939 --> 00:00:46,238
整个系统的瓶颈就会向其他环节转移

19
00:00:46,238 --> 00:00:49,499
而CPU目前正处在这个关键的转折上

20
00:00:49,499 --> 00:00:53,128
真正的下一站很可能是CPU的测试与封装

21
00:00:53,128 --> 00:00:54,189
hi朋友们

22
00:00:54,189 --> 00:00:55,609
欢迎回到ruby投资笔记

23
00:00:55,609 --> 00:00:57,280
我是你们的老朋友ruby

24
00:00:57,280 --> 00:00:58,719
那今天这期视频呢

25
00:00:58,719 --> 00:01:00,200
我们要聊的不是那些

26
00:01:00,200 --> 00:01:02,539
已经让你长到怀疑人生的明星股

27
00:01:02,539 --> 00:01:05,379
而是这些藏在巨头背后的隐形霸主

28
00:01:05,379 --> 00:01:06,739
重点就看两件事

29
00:01:06,739 --> 00:01:08,739
第一这个100秒的测试噩梦

30
00:01:08,739 --> 00:01:11,849
为什么会成为2026年最硬核的财富密码

31
00:01:11,849 --> 00:01:15,409
第二有哪些公司正在接力LITE悄悄进入

32
00:01:15,409 --> 00:01:18,009
像台积电这样的核心供应链体系

33
00:01:18,009 --> 00:01:21,789
那如果你想在CPU的下半场找回那种翻倍的节奏

34
00:01:21,789 --> 00:01:24,849
这一期的内容一定要看到最后哦

35
00:01:25,719 --> 00:01:28,120
那在聊测试之前呢

36
00:01:28,120 --> 00:01:29,840
我们先把大背景讲清楚

37
00:01:29,840 --> 00:01:33,269
不然你很难理解后面的问题为什么会这么棘手

38
00:01:33,269 --> 00:01:34,189
这两年呢

39
00:01:34,189 --> 00:01:35,709
你应该也明显的感觉到

40
00:01:35,709 --> 00:01:38,269
AI数据中心的规模在疯狂的扩张

41
00:01:38,269 --> 00:01:40,209
算力是一层层的往上堆

42
00:01:40,209 --> 00:01:41,090
但问题是

43
00:01:41,090 --> 00:01:44,469
底层的数据传输还在大量依赖传统的同互联

44
00:01:44,469 --> 00:01:46,689
也至于用电缆来跑信号

45
00:01:46,689 --> 00:01:49,650
这套体系其实越来越接近物理极限了

46
00:01:49,650 --> 00:01:51,750
因为铜线传输它本身功耗高

47
00:01:51,750 --> 00:01:52,409
发热大

48
00:01:52,409 --> 00:01:56,150
而且在延迟上已经开始跟不上GPU的迭代节奏

49
00:01:56,150 --> 00:01:57,250
算力在加速

50
00:01:57,250 --> 00:01:59,269
但是数据搬运的速度却在掉队

51
00:01:59,269 --> 00:02:01,549
所以行业里基本上形成了一个共识

52
00:02:01,549 --> 00:02:02,909
要继续往上走

53
00:02:02,909 --> 00:02:03,950
必须换一条路

54
00:02:03,950 --> 00:02:05,280
用光替代电

55
00:02:05,280 --> 00:02:06,879
这就是CPU共封装

56
00:02:06,879 --> 00:02:08,080
光学的核心逻辑

57
00:02:08,080 --> 00:02:09,199
简单理解就是

58
00:02:09,199 --> 00:02:12,769
把原本独立在芯片外部的光模块拆解之后

59
00:02:12,769 --> 00:02:14,669
直接塞进芯片封装里

60
00:02:14,669 --> 00:02:17,819
让数据通过光信号在芯片内部传输

61
00:02:17,819 --> 00:02:19,259
一旦这件事情成立

62
00:02:19,259 --> 00:02:21,199
带来的改变是难以估量的

63
00:02:21,199 --> 00:02:24,319
功耗带宽延迟都会出现结构性的优化

64
00:02:24,319 --> 00:02:27,269
而且这件事情已经不是远期故事了

65
00:02:27,269 --> 00:02:30,209
像台积电它已经给出了非常明确的时间表

66
00:02:30,209 --> 00:02:31,588
就是它的初步平台

67
00:02:31,588 --> 00:02:33,908
预计是在2026年进入量产阶段

68
00:02:33,908 --> 00:02:37,180
这就意味着CPU处处在一个非常关键的节点

69
00:02:37,180 --> 00:02:39,439
从实验室走向规模化的落地

70
00:02:39,439 --> 00:02:41,349
甚至是走向商业兑现

71
00:02:41,349 --> 00:02:44,269
但是问题也正是这个阶段暴露出来的

72
00:02:44,269 --> 00:02:47,348
就在市场还在为量产在即欢呼的时候

73
00:02:47,348 --> 00:02:50,699
产业链里的核心玩家们反而开始极力头疼

74
00:02:50,699 --> 00:02:52,219
因为他们发现一件事

75
00:02:52,219 --> 00:02:54,599
CPU这东西做出来没有那么困难

76
00:02:54,599 --> 00:02:56,159
但是想要把它测清楚

77
00:02:56,159 --> 00:02:58,060
测的准还要测的足够快

78
00:02:58,060 --> 00:03:00,129
才是真正的难题

79
00:03:01,199 --> 00:03:03,639
我先给你个直观感受

80
00:03:03,639 --> 00:03:05,800
你就知道这个问题有多致命了

81
00:03:05,800 --> 00:03:07,159
在传统的电芯片

82
00:03:07,159 --> 00:03:08,960
也就是ESC的测试流程里

83
00:03:08,960 --> 00:03:11,080
整个体系已经高度标准化

84
00:03:11,080 --> 00:03:14,060
基本上就是全自动设备跑流程速度快

85
00:03:14,060 --> 00:03:14,659
效率高

86
00:03:14,659 --> 00:03:15,780
没有什么卡点

87
00:03:15,780 --> 00:03:17,340
但是一旦进入CPU

88
00:03:17,340 --> 00:03:18,939
也就是光电融合的架构

89
00:03:18,939 --> 00:03:20,500
PAIC加EIC

90
00:03:20,500 --> 00:03:23,819
它的复杂程度是直接从三级跳到一级的

91
00:03:23,819 --> 00:03:24,780
先看第一层

92
00:03:24,780 --> 00:03:26,669
就是复杂度的维度变了

93
00:03:26,669 --> 00:03:28,629
以前你只需要测电学参数

94
00:03:28,629 --> 00:03:30,189
比如电压电流这些

95
00:03:30,189 --> 00:03:31,340
但是现在不一样

96
00:03:31,340 --> 00:03:32,520
你不仅要测电

97
00:03:32,520 --> 00:03:33,419
还要测光

98
00:03:33,419 --> 00:03:34,300
更关键的是

99
00:03:34,300 --> 00:03:37,139
还要测光与电之间是怎么协同工作的

100
00:03:37,139 --> 00:03:38,539
也就是光电交互

101
00:03:38,539 --> 00:03:39,939
具体到指标层面

102
00:03:39,939 --> 00:03:43,680
就是你需要看插入损耗IL偏振相关损耗

103
00:03:43,680 --> 00:03:45,039
pd l响应度

104
00:03:45,039 --> 00:03:46,199
波导传播损耗

105
00:03:46,199 --> 00:03:47,599
观绳串扰等等

106
00:03:47,599 --> 00:03:49,838
这样一套完全不同的体系

107
00:03:49,838 --> 00:03:51,098
那更麻烦的是

108
00:03:51,098 --> 00:03:52,098
到目前为止

109
00:03:52,098 --> 00:03:54,590
行业里面没有统一的测试标准

110
00:03:54,590 --> 00:03:57,210
各家厂商用的方案口径不一样

111
00:03:57,210 --> 00:04:00,649
很多环节它只能靠工程师去手动介入

112
00:04:00,649 --> 00:04:02,489
这就直接导致了一个结果

113
00:04:02,489 --> 00:04:04,669
就是测试流程很难自动化

114
00:04:04,669 --> 00:04:06,460
它的效率被严重拖慢

115
00:04:06,460 --> 00:04:08,259
但是这些还不是最难的

116
00:04:08,259 --> 00:04:09,340
真正硬核的瓶颈

117
00:04:09,340 --> 00:04:10,060
在第二点

118
00:04:10,060 --> 00:04:12,139
就是物理层面的穿针引线

119
00:04:12,139 --> 00:04:13,780
光信号要进入芯片

120
00:04:13,780 --> 00:04:15,280
必须通过光纤耦合

121
00:04:15,280 --> 00:04:16,339
但是问题在于

122
00:04:16,339 --> 00:04:20,139
单模光纤的截面积和芯片内部光波导的截面积

123
00:04:20,139 --> 00:04:21,860
相差了大约800倍

124
00:04:21,860 --> 00:04:22,959
这是什么概念

125
00:04:22,959 --> 00:04:25,639
你可以把它想象成站在一公里之外

126
00:04:25,639 --> 00:04:28,389
把一根线精准地穿进一个针眼里面

127
00:04:28,389 --> 00:04:30,970
而且误差必须控制在纳米级

128
00:04:30,970 --> 00:04:32,189
一旦偏一点点

129
00:04:32,189 --> 00:04:34,589
这个光功率基本上就损耗殆尽

130
00:04:34,589 --> 00:04:36,670
所以现在很多的测试动作

131
00:04:36,670 --> 00:04:38,649
还停留在手工艺的阶段

132
00:04:38,649 --> 00:04:40,709
就是工程师一边微调角度

133
00:04:40,709 --> 00:04:41,990
一边盯着功率变化

134
00:04:41,990 --> 00:04:43,009
慢慢对准

135
00:04:43,009 --> 00:04:46,449
那最终带来的结果就是一颗完整的CPU芯片

136
00:04:46,449 --> 00:04:48,430
如果想要做百分之百的测试

137
00:04:48,430 --> 00:04:50,810
他平均耗时是要超过100秒的

138
00:04:50,810 --> 00:04:51,889
注意这个数字

139
00:04:51,889 --> 00:04:53,329
这个数字在实验室里边

140
00:04:53,329 --> 00:04:54,889
你听着可能还能接受

141
00:04:54,889 --> 00:04:56,949
但是一旦进入大规模的量产

142
00:04:56,949 --> 00:04:59,069
这简直就是灾难级别的瓶颈

143
00:04:59,069 --> 00:05:02,689
AI基础设施是成千上万颗芯片堆出来的

144
00:05:02,689 --> 00:05:04,870
如果每一颗都要测100秒

145
00:05:04,870 --> 00:05:08,220
那整条产品线它基本上就是被毒死的状态

146
00:05:08,220 --> 00:05:09,500
所以我的结论就是

147
00:05:09,500 --> 00:05:12,019
如果测试速度和自动化的问题不解决

148
00:05:12,019 --> 00:05:15,459
那CPU的商业化落地基本上就是一个空话

149
00:05:17,319 --> 00:05:20,420
而一颗CPU芯片从出生到真正出场

150
00:05:20,420 --> 00:05:23,149
它中间其实要经历四道测试关卡

151
00:05:23,149 --> 00:05:24,910
这是一个一层层筛选

152
00:05:24,910 --> 00:05:26,569
然后逐步验证的过程

153
00:05:26,569 --> 00:05:29,050
第一道呢是PC的晶圆级测试

154
00:05:29,050 --> 00:05:32,019
ow a t也就是在还没有封装之前

155
00:05:32,019 --> 00:05:34,870
先把光芯片本身的基础性能测一遍

156
00:05:34,870 --> 00:05:37,209
第二步是光电耦合的晶圆级测试

157
00:05:37,209 --> 00:05:39,930
这里是开始把电和光结合起来

158
00:05:39,930 --> 00:05:42,560
看两套系统能不能正常协同工作

159
00:05:42,560 --> 00:05:45,480
第三道呢是OE光学引擎级的测试

160
00:05:45,480 --> 00:05:47,319
这一步的核心目标是

161
00:05:47,319 --> 00:05:49,279
筛选出已知良好的光学引擎

162
00:05:49,279 --> 00:05:50,649
也就是KGOE

163
00:05:50,649 --> 00:05:54,019
这是整个流程里面最烧钱最关键的一环

164
00:05:54,019 --> 00:05:57,379
那最后一步就是先进封装之后的模块级测试

165
00:05:57,379 --> 00:06:00,000
相当于是整机出厂前的总检

166
00:06:00,000 --> 00:06:01,399
这四个阶段看下来

167
00:06:01,399 --> 00:06:04,449
很多人的第一直觉是觉得越往后越重要

168
00:06:04,449 --> 00:06:06,790
但是如果你是站在投资的角度来看

169
00:06:06,790 --> 00:06:08,509
真正值得我们重点盯的

170
00:06:08,509 --> 00:06:11,660
反而是前面的第一部PAIC金元机测试

171
00:06:11,660 --> 00:06:13,420
这个原因呢其实很简单

172
00:06:13,420 --> 00:06:16,500
就是光芯片PAIC大多是基于成熟的制成

173
00:06:16,500 --> 00:06:17,839
它的成本并不高

174
00:06:17,839 --> 00:06:19,759
但是和它配套的电芯片

175
00:06:19,759 --> 00:06:20,500
Esc

176
00:06:20,500 --> 00:06:22,920
往往用的是像台积电这种最先进

177
00:06:22,920 --> 00:06:24,459
也最昂贵的制成工艺

178
00:06:24,459 --> 00:06:25,860
那如果你在第一阶段

179
00:06:25,860 --> 00:06:28,100
没有把有问题的光芯片筛选掉

180
00:06:28,100 --> 00:06:31,300
而是等它和昂贵的电芯片一起完成封装之后

181
00:06:31,300 --> 00:06:32,439
才发现有缺陷

182
00:06:32,439 --> 00:06:34,430
那损失是成倍放大的

183
00:06:34,430 --> 00:06:35,029
换句话说

184
00:06:35,029 --> 00:06:36,649
这就像一条生产线

185
00:06:36,649 --> 00:06:38,810
你一定是先把零件筛选关进

186
00:06:38,810 --> 00:06:40,639
再去组装高价值的整机

187
00:06:40,639 --> 00:06:41,839
越早发现问题

188
00:06:41,839 --> 00:06:42,959
成本控制得越好

189
00:06:42,959 --> 00:06:44,319
利润空间也越大

190
00:06:44,319 --> 00:06:47,519
所以从资金效率和产业价值分配的角度来看

191
00:06:47,519 --> 00:06:51,000
第一道关卡才是真正意义上的金矿

192
00:06:52,800 --> 00:06:54,779
而为了解决掉这个关卡

193
00:06:54,779 --> 00:06:58,439
全球做测试机的玩家其实已经全面动起来了

194
00:06:58,439 --> 00:07:00,360
且他们的路径非常明确

195
00:07:00,360 --> 00:07:02,670
就是各自补短板抢卡位

196
00:07:06,029 --> 00:07:09,180
先看第一类就是走极致精密路线的组合

197
00:07:09,180 --> 00:07:11,420
代表呢就是日本的爱德万测试

198
00:07:11,420 --> 00:07:13,240
和美国的FORM

199
00:07:13,240 --> 00:07:15,259
一个是传统电测试的霸主

200
00:07:15,259 --> 00:07:17,250
一个是全球探针卡龙头

201
00:07:17,250 --> 00:07:18,310
但问题在于

202
00:07:18,310 --> 00:07:19,509
前者强在电

203
00:07:19,509 --> 00:07:20,910
后者补的是光

204
00:07:20,910 --> 00:07:22,269
两边刚好互补

205
00:07:22,269 --> 00:07:23,370
他们的核心武器

206
00:07:23,370 --> 00:07:25,639
也就是所谓的UFO探针卡

207
00:07:25,639 --> 00:07:27,339
这个名字听起来很玄乎

208
00:07:27,339 --> 00:07:30,870
但是关键在于他们做的就是对准容差补偿

209
00:07:30,870 --> 00:07:32,189
我们前面提到了

210
00:07:32,189 --> 00:07:35,069
光纤和波导之间是有800倍的尺寸差

211
00:07:35,069 --> 00:07:36,069
一旦对不准

212
00:07:36,069 --> 00:07:37,819
那信号直接就损耗掉

213
00:07:37,819 --> 00:07:40,319
而它们是通过对光束做整形

214
00:07:40,319 --> 00:07:42,139
即便是存在轻微的偏差

215
00:07:42,139 --> 00:07:44,980
光信号依然能够稳定地耦合进去

216
00:07:44,980 --> 00:07:46,100
在这个基础上

217
00:07:46,100 --> 00:07:48,800
他们又推出了V93000train系统

218
00:07:48,800 --> 00:07:50,439
引入了九轴光子对准

219
00:07:50,439 --> 00:07:51,100
简单理解

220
00:07:51,100 --> 00:07:54,759
就是在传统的三维定位之外叠加旋转微调

221
00:07:54,759 --> 00:07:56,500
再配合机器视觉算法

222
00:07:56,500 --> 00:07:59,250
把原本需要人工慢慢调的对准过程

223
00:07:59,250 --> 00:08:01,540
从分钟级压缩到秒级

224
00:08:01,540 --> 00:08:03,540
这一代玩家的逻辑很清晰

225
00:08:03,540 --> 00:08:04,620
就是不拼产量

226
00:08:04,620 --> 00:08:05,379
拼精度

227
00:08:05,379 --> 00:08:08,920
他的目标就是高端的CPU芯片的大规模量产

228
00:08:08,920 --> 00:08:11,560
再看第二类是典型的效率优先派

229
00:08:11,560 --> 00:08:15,259
代表的就是美国的泰瑞达和韩国的FONTEC

230
00:08:15,259 --> 00:08:17,439
相比第一类慢慢打磨技术

231
00:08:17,439 --> 00:08:20,430
他们的打法更直接就是通过收购和合作

232
00:08:20,430 --> 00:08:22,449
快速补齐自己的能力短板

233
00:08:22,449 --> 00:08:25,149
他们推出的一个关键产品是300mm

234
00:08:25,149 --> 00:08:27,149
双面晶圆探针测试系统

235
00:08:27,149 --> 00:08:29,149
这个系统的核心优势在于

236
00:08:29,149 --> 00:08:31,410
同时作业晶圆顶面测电

237
00:08:31,410 --> 00:08:33,889
底面测光相当于是双线程运行

238
00:08:33,889 --> 00:08:35,879
它的测试吞吐量直接翻倍

239
00:08:35,879 --> 00:08:38,919
再叠加方on t e c本身完整的产品线

240
00:08:38,919 --> 00:08:41,240
从晶圆级到芯片级全覆盖

241
00:08:41,240 --> 00:08:44,600
对于那些追求良率和产能的晶圆厂来说

242
00:08:44,600 --> 00:08:47,179
这种一站式全自动化的解决方案

243
00:08:47,179 --> 00:08:48,789
吸引力是非常强的

244
00:08:48,789 --> 00:08:51,330
所以这一排的核心逻辑就是两个词

245
00:08:51,330 --> 00:08:52,769
自动化和规模化

246
00:08:52,769 --> 00:08:55,470
他们是CPU从实验室走向工业化

247
00:08:55,470 --> 00:08:56,919
落地的关键推动者

248
00:08:56,919 --> 00:09:00,519
而最后李磊其实是更偏细分突破的技术型玩家

249
00:09:00,519 --> 00:09:03,039
代表是包括美国的KEYS的

250
00:09:03,039 --> 00:09:05,500
台湾的科尔马ATE光亮科技

251
00:09:05,500 --> 00:09:06,340
这类公司

252
00:09:06,340 --> 00:09:07,820
它不一定做整套的系统

253
00:09:07,820 --> 00:09:11,279
但是在某些关键环节上有着明显的杀手锏

254
00:09:11,279 --> 00:09:12,840
比如说KEYS

255
00:09:12,840 --> 00:09:15,059
它在测试领域是绝对的龙头

256
00:09:15,059 --> 00:09:17,159
它的核心优势在偏振控制

257
00:09:17,159 --> 00:09:20,318
那光芯片测试最怕的就是偏振态不稳定

258
00:09:20,318 --> 00:09:22,619
而它可以把偏振给锁定住

259
00:09:22,619 --> 00:09:25,318
让整个扫剖长过程保持高度稳定

260
00:09:25,318 --> 00:09:28,250
这对于研发阶段的效率提升非常关键

261
00:09:28,250 --> 00:09:29,929
再比如台湾的科尔马

262
00:09:29,929 --> 00:09:31,129
他在系统级测试

263
00:09:31,129 --> 00:09:34,120
也就是SLT领域经验也是非常深的

264
00:09:34,120 --> 00:09:37,159
现在正在把过去在3D感测和激光器

265
00:09:37,159 --> 00:09:40,000
老化测试上的积累迁移到CPU上

266
00:09:40,000 --> 00:09:42,399
因为一旦芯片内部集成了激光器

267
00:09:42,399 --> 00:09:44,720
那可靠性的测试就变成了刚需

268
00:09:44,720 --> 00:09:47,000
那还有一个非常有意思的新玩家

269
00:09:47,000 --> 00:09:48,578
就是中国的光亮科技

270
00:09:48,578 --> 00:09:50,519
他们推出的net ja平台

271
00:09:50,519 --> 00:09:53,399
本质上是用高光谱成像去看见光

272
00:09:53,399 --> 00:09:56,139
以前工程师呢只能通过间接数据去判断

273
00:09:56,139 --> 00:09:56,960
哪里漏光

274
00:09:56,960 --> 00:09:59,899
现在可以直接在光波导内部定位问题点

275
00:09:59,899 --> 00:10:02,850
就像是给芯片做了一次X光的扫描

276
00:10:02,850 --> 00:10:05,568
这对于提升两率价值非常直接

277
00:10:05,568 --> 00:10:09,009
那总结一下就是这三股力量的分工其实很清楚

278
00:10:09,009 --> 00:10:11,159
就是第一类去解决精度极限

279
00:10:11,159 --> 00:10:13,000
第二类呢是解决产能瓶颈

280
00:10:13,000 --> 00:10:14,940
第三类就是解决关键细节

281
00:10:14,940 --> 00:10:16,480
而真正的行业红利

282
00:10:16,480 --> 00:10:18,899
往往就藏在这三者交汇的地方

283
00:10:18,899 --> 00:10:21,379
但是呢光知道公司的名字没有用

284
00:10:21,379 --> 00:10:23,700
你一样很难把握到赚钱的机会

285
00:10:23,700 --> 00:10:25,179
很多人踩坑的原因

286
00:10:25,179 --> 00:10:27,840
就是因为分不清楚哪些是蹭热点的公司

287
00:10:27,840 --> 00:10:30,690
又有哪些是真正拿到订单的公司

288
00:10:30,690 --> 00:10:32,649
比如只有深度跟踪才知道

289
00:10:32,649 --> 00:10:35,379
泰瑞达的300mm的双面探针系统

290
00:10:35,379 --> 00:10:38,019
已经给某头部经验厂送样测试了

291
00:10:38,019 --> 00:10:40,490
预计第三季度就能拿到批量订单

292
00:10:40,490 --> 00:10:43,570
再比如说FORM的UFO探针卡

293
00:10:43,570 --> 00:10:46,639
它的两率比行业平均水平高出来15%

294
00:10:46,639 --> 00:10:48,000
那这些关键数据呢

295
00:10:48,000 --> 00:10:50,019
我都是花了大量的时间和资金

296
00:10:50,019 --> 00:10:51,519
跑产业链才拿到的

297
00:10:51,519 --> 00:10:54,000
如果你不想自己花几个月的时间调研

298
00:10:54,000 --> 00:10:57,328
也不想错过2026年CPU布局的最后窗口

299
00:10:57,328 --> 00:10:59,269
可以点击评论区的置顶链接

300
00:10:59,269 --> 00:11:00,808
加入我的交流俱乐部

301
00:11:00,808 --> 00:11:02,808
既能节省你寻找资料的时间

302
00:11:02,808 --> 00:11:05,528
也能精准的把握投资机会

303
00:11:07,159 --> 00:11:09,019
那最后呢我再来讲一下

304
00:11:09,019 --> 00:11:12,009
为什么我会觉得现在是一个可以布局的窗口

305
00:11:12,009 --> 00:11:14,750
那如果你从23年一直开始投资半导体

306
00:11:14,750 --> 00:11:17,009
你就会发现一个很有意思的规律

307
00:11:17,009 --> 00:11:20,330
每当半导体行业出现一次底层技术的跃迁

308
00:11:20,330 --> 00:11:23,818
在量产前业测试设备的地位都会明显抬升

309
00:11:23,818 --> 00:11:26,879
不管是当年的EOV还是后来的3D net

310
00:11:26,879 --> 00:11:28,259
本质上都是一样的

311
00:11:28,259 --> 00:11:29,440
逻辑技术

312
00:11:29,440 --> 00:11:32,009
一到跨维度测试就不再是配角

313
00:11:32,009 --> 00:11:34,870
而是决定能不能真正落地的关键环节

314
00:11:34,870 --> 00:11:36,210
那放到CPU上

315
00:11:36,210 --> 00:11:37,889
这个逻辑只会更极端

316
00:11:37,889 --> 00:11:40,250
第一它的复杂度是跃迁式的提升

317
00:11:40,250 --> 00:11:43,029
因为CPU它不是简单多了一个功能模块

318
00:11:43,029 --> 00:11:45,480
而是从电到光的物理层切换

319
00:11:45,480 --> 00:11:48,120
他这就相当于是换了一套游戏规则

320
00:11:48,120 --> 00:11:50,309
测试难度自然是同步放大

321
00:11:50,309 --> 00:11:52,429
第二就是价值量占比在上升

322
00:11:52,429 --> 00:11:54,809
随着单个芯片的测试时间被拉长

323
00:11:54,809 --> 00:11:56,360
测试站数量增加

324
00:11:56,360 --> 00:11:59,320
测试设备在整个半导体资本开支里的占比

325
00:11:59,320 --> 00:12:00,509
会持续抬高

326
00:12:00,509 --> 00:12:01,350
这一点呢

327
00:12:01,350 --> 00:12:04,570
其实已经在先进制程节点里面反复验证过

328
00:12:04,570 --> 00:12:06,350
第三呢也是最关键的一点

329
00:12:06,350 --> 00:12:08,539
就是时间窗口已经非常清晰

330
00:12:08,539 --> 00:12:09,559
像台积电

331
00:12:09,559 --> 00:12:10,059
英特尔

332
00:12:10,059 --> 00:12:12,090
博通这些产业链的核心玩家

333
00:12:12,090 --> 00:12:15,629
基本上都把CPU的量产节点指向了2026年

334
00:12:15,629 --> 00:12:18,399
比之前的2028年又提前了两年

335
00:12:18,399 --> 00:12:20,580
而按照半导体设备的周期规律

336
00:12:20,580 --> 00:12:23,700
设备订单通常会提前12~18个月释放

337
00:12:23,700 --> 00:12:26,299
换句话说就是从现在到2026年末

338
00:12:26,299 --> 00:12:29,379
就是测试设备厂商订单集中兑现的阶段

339
00:12:29,379 --> 00:12:30,539
这也是为什么说

340
00:12:30,539 --> 00:12:33,019
现在其实是一个典型的预期差窗口

341
00:12:33,019 --> 00:12:34,440
市场还在讲故事

342
00:12:34,440 --> 00:12:36,210
但是订单已经在路上

343
00:12:36,210 --> 00:12:37,210
那除此之外呢

344
00:12:37,210 --> 00:12:38,330
我还要提示一下

345
00:12:38,330 --> 00:12:40,250
就是我今天讲的这三家美股

346
00:12:40,250 --> 00:12:41,350
注意是美股

347
00:12:41,350 --> 00:12:43,690
他们的股价对于稳健型的投资者来说

348
00:12:43,690 --> 00:12:45,549
是存在一定的风险的

349
00:12:45,549 --> 00:12:47,110
所以能够等待的人

350
00:12:47,110 --> 00:12:49,379
还是可以等待他的回调再入场

351
00:12:49,379 --> 00:12:50,919
那已经持有的人

352
00:12:50,919 --> 00:12:51,940
我想说的就是

353
00:12:51,940 --> 00:12:54,019
千万不要留恋任何一只过股

354
00:12:54,019 --> 00:12:57,440
尤其是美股测试机的龙头泰瑞达即将发布财报

355
00:12:57,440 --> 00:13:00,539
如果他的财报里面说他的测试机超产量

356
00:13:00,539 --> 00:13:02,379
然后带来的预期收入很高

357
00:13:02,379 --> 00:13:05,980
那么测试机这个行业一定是有一波的带动效果

358
00:13:05,980 --> 00:13:08,659
但是如果他的这个财报不及预期的话

359
00:13:08,659 --> 00:13:10,399
还是存在一定的风险的

360
00:13:10,399 --> 00:13:13,259
那一旦这些个股出现趋势减弱的迹象

361
00:13:13,259 --> 00:13:15,440
一定要酌情调整自己的仓位

362
00:13:15,440 --> 00:13:17,100
那不过据我的估计的话

363
00:13:17,100 --> 00:13:18,820
我觉得泰瑞达这次的财报

364
00:13:18,820 --> 00:13:20,980
应该是给市场一个满意的答案

365
00:13:20,980 --> 00:13:22,620
当然我也有可能打脸

366
00:13:22,620 --> 00:13:23,779
那打脸就打脸吧

367
00:13:23,779 --> 00:13:25,340
打脸就卖了这只股票嘛

368
00:13:25,340 --> 00:13:27,139
那接下来问题就留给你

369
00:13:27,139 --> 00:13:28,460
在这几条路径里面

370
00:13:28,460 --> 00:13:29,879
你更看好哪一类呢

371
00:13:29,879 --> 00:13:32,059
或者你觉得CPU真正量产之前

372
00:13:32,059 --> 00:13:33,820
还有哪些被忽略的坑呢

373
00:13:33,820 --> 00:13:35,980
欢迎在评论区发表您的意见

374
00:13:35,980 --> 00:13:38,500
那如果你觉得这期视频对你有帮助的话

375
00:13:38,500 --> 00:13:40,559
也欢迎您点赞订阅ruby投资笔记

376
00:13:40,559 --> 00:13:42,539
不错过下一期精彩视频

377
00:13:42,539 --> 00:13:43,340
我是ruby

378
00:13:43,340 --> 00:13:44,899
咱们一期一会下期再见

379
00:13:44,899 --> 00:13:45,460
拜

共 2 条，按点赞数排序。

采集完成，该视频无弹幕