美股 AI进入执行时代，真正的赢家是这两家大公司！

1
00:00:00,120 --> 00:00:01,919
我想问大家一个问题

2
00:00:01,919 --> 00:00:06,160
当模型能力和算力投入都已经卷到今天

3
00:00:06,160 --> 00:00:06,980
这个阶段

4
00:00:06,980 --> 00:00:08,199
AI的下半场

5
00:00:08,199 --> 00:00:10,689
我们到底应该关注什么

6
00:00:10,689 --> 00:00:14,349
过去这一轮AI本质上在做一件事

7
00:00:14,349 --> 00:00:15,490
回答问题

8
00:00:15,490 --> 00:00:16,469
你问一句

9
00:00:16,469 --> 00:00:17,429
他答一句

10
00:00:17,429 --> 00:00:18,690
谁的模型更强

11
00:00:18,690 --> 00:00:20,070
谁的算力更充足

12
00:00:20,070 --> 00:00:21,989
谁就更容易胜出

13
00:00:21,989 --> 00:00:24,870
但现在拐点已经出现了

14
00:00:24,870 --> 00:00:28,089
AI正在从回答问题走向执行任务

15
00:00:28,089 --> 00:00:30,809
也就是我们常说的AI agent

16
00:00:30,809 --> 00:00:33,369
他面对的不再是单轮对话

17
00:00:33,369 --> 00:00:35,969
而是一整条完整的任务链

18
00:00:35,969 --> 00:00:37,369
从理解目标

19
00:00:37,369 --> 00:00:39,710
拆解步骤到调用工具

20
00:00:39,710 --> 00:00:40,609
连接系统

21
00:00:40,609 --> 00:00:42,659
读取数据管理状态

22
00:00:42,659 --> 00:00:45,399
最终把任务真正落地执行

23
00:00:45,399 --> 00:00:47,020
一旦进入这个阶段

24
00:00:47,020 --> 00:00:49,929
竞争的核心就彻底变了

25
00:00:49,929 --> 00:00:52,630
彼得已经不只是算得快不快

26
00:00:52,630 --> 00:00:57,179
而是整套系统能不能长期稳定的协同运转

27
00:00:57,179 --> 00:00:59,460
能不能把复杂流程跑通

28
00:00:59,460 --> 00:01:03,170
能不能在真实环境中持续交付结果

29
00:01:03,170 --> 00:01:06,670
这背后考验的其实是基础设施能力

30
00:01:06,670 --> 00:01:09,849
今天要深入聊的这两家公司正式卡位

31
00:01:09,849 --> 00:01:11,269
在这个关键环节

32
00:01:11,269 --> 00:01:13,409
他们提供的不只是算力

33
00:01:13,409 --> 00:01:16,659
而是支撑AI会推理的底层能力

34
00:01:16,659 --> 00:01:18,480
熟悉我们的观众都知道

35
00:01:18,480 --> 00:01:22,280
我们更倾向于把一家公司研究到足够深入

36
00:01:22,280 --> 00:01:24,560
把业务拆解到足够细致

37
00:01:24,560 --> 00:01:28,599
只有这样才能真正看清它的利润究竟来自哪里

38
00:01:28,599 --> 00:01:31,140
这些利润是具备长期可持续性

39
00:01:31,140 --> 00:01:33,180
还是仅仅阶段性释放

40
00:01:33,180 --> 00:01:34,819
也只有在这个基础上

41
00:01:34,819 --> 00:01:38,189
才能进一步识别其上下游供应链中

42
00:01:38,189 --> 00:01:40,010
下一轮周期可能出现的

43
00:01:40,010 --> 00:01:42,780
关键拐点和潜在爆发机会

44
00:01:42,780 --> 00:01:45,620
就像年初我们公开视频推介的

45
00:01:45,620 --> 00:01:49,650
2026必买十只美股上中下三级

46
00:01:49,650 --> 00:01:54,030
GLW和NBIS上涨已经超过90%

47
00:01:54,030 --> 00:01:56,579
DTMI大涨超75%

48
00:01:56,579 --> 00:02:01,489
MURTW和PLAB大涨50%

49
00:02:01,489 --> 00:02:03,230
视频链接我放下方

50
00:02:03,230 --> 00:02:04,489
欢迎大家回顾

51
00:02:04,489 --> 00:02:07,709
另外算力互联龙头迈瑞尔科技

52
00:02:07,709 --> 00:02:10,889
我们VIP社群当时以81美元附近

53
00:02:10,889 --> 00:02:13,310
以30%仓位重点布局

54
00:02:13,310 --> 00:02:14,189
截至今天

55
00:02:14,189 --> 00:02:17,150
这一笔的累计收益也已经达到67%

56
00:02:17,150 --> 00:02:21,150
我们选的这些股背后都是压中了AI基础设施

57
00:02:21,150 --> 00:02:25,930
中光互联和存储芯片等关键上下游供应链

58
00:02:25,930 --> 00:02:27,930
不过在看受益方之前

59
00:02:27,930 --> 00:02:31,338
我们得先看需求到底是从哪里冲出来的

60
00:02:31,338 --> 00:02:33,498
因为只有源头真的开始放量

61
00:02:33,498 --> 00:02:36,269
后面整条链的逻辑才站得住

62
00:02:36,269 --> 00:02:38,908
而在AI进入下半场之后

63
00:02:38,908 --> 00:02:40,449
需求释放最猛

64
00:02:40,449 --> 00:02:44,240
也最具有代表性的公司非ANTHROPIC莫属

65
00:02:44,240 --> 00:02:45,860
先看最核心的数据

66
00:02:45,860 --> 00:02:46,960
2025年底

67
00:02:46,960 --> 00:02:49,990
ATHROPIC的年化收入还只有90亿美元

68
00:02:49,990 --> 00:02:52,449
一个月前升到190亿美元

69
00:02:52,449 --> 00:02:56,079
而现在已经直接冲到了300亿美元

70
00:02:56,079 --> 00:02:57,959
这已经不是普通增长了

71
00:02:57,959 --> 00:03:00,780
而是非常明显的加速放量

72
00:03:00,780 --> 00:03:01,919
更关键的是

73
00:03:01,919 --> 00:03:05,800
每年在cloud上花费超100万美元的企业客户

74
00:03:05,800 --> 00:03:08,949
两个月内从500家翻倍至1000家

75
00:03:08,949 --> 00:03:12,129
这说明企业级AI已经跨过了试一试的阶段

76
00:03:12,129 --> 00:03:14,550
进入了真正的大规模采购期

77
00:03:14,550 --> 00:03:16,509
但这里有一个很现实的矛盾

78
00:03:16,509 --> 00:03:18,289
ANTHROPIC它没上市

79
00:03:18,289 --> 00:03:19,930
公开市场买不到

80
00:03:19,930 --> 00:03:23,030
所以资金想压住这条最强的增长线

81
00:03:23,030 --> 00:03:26,349
最后一定会找那个最直接的受益方

82
00:03:26,349 --> 00:03:28,090
根据我们的深度调研

83
00:03:28,090 --> 00:03:30,750
这个最直接的受益方就是亚马逊

84
00:03:30,750 --> 00:03:32,969
它不是一般意义上的跟着受益

85
00:03:32,969 --> 00:03:35,330
而是通过五层严密的布局

86
00:03:35,330 --> 00:03:37,289
把ANTHROPIC的每一次爆发

87
00:03:37,289 --> 00:03:39,909
都转化成了自己的真金白银

88
00:03:39,909 --> 00:03:41,030
第一层

89
00:03:41,030 --> 00:03:43,169
先把最直接的钱收走

90
00:03:43,169 --> 00:03:44,639
A w s

91
00:03:44,639 --> 00:03:47,400
很多人看到亚马逊累计投了80亿美元

92
00:03:47,400 --> 00:03:48,300
给ANTHROPIC

93
00:03:48,300 --> 00:03:50,269
第一反应呢是财务投资

94
00:03:50,269 --> 00:03:52,389
但如果把时间线拉开看

95
00:03:52,389 --> 00:03:53,169
就会发现

96
00:03:53,169 --> 00:03:54,729
这根本不是普通投资

97
00:03:54,729 --> 00:03:56,699
而是一场一步步加深

98
00:03:56,699 --> 00:04:00,360
而且始终围绕AWS展开的深度绑定

99
00:04:00,360 --> 00:04:02,020
2023年9月

100
00:04:02,020 --> 00:04:04,750
亚马逊先投了12.5亿美元

101
00:04:04,750 --> 00:04:08,330
AWS成为aerobic的主要云合作伙伴

102
00:04:08,330 --> 00:04:10,310
cloud进入bad rock

103
00:04:10,310 --> 00:04:11,750
2024年3月

104
00:04:11,750 --> 00:04:13,650
总投资扩大到40亿美元

105
00:04:13,650 --> 00:04:16,129
绑定进一步深入到底层算力

106
00:04:16,129 --> 00:04:17,949
AERROPIC也长期承诺

107
00:04:17,949 --> 00:04:21,430
会使用aw s的训练芯片和推理芯片

108
00:04:21,430 --> 00:04:23,129
2024年11月

109
00:04:23,129 --> 00:04:25,209
亚马逊再追加40亿美元

110
00:04:25,209 --> 00:04:27,089
总额拉到80亿美元

111
00:04:27,089 --> 00:04:27,970
Athropic

112
00:04:27,970 --> 00:04:31,870
又进一步把AWS定为首选训练合作伙伴

113
00:04:31,870 --> 00:04:32,810
换句话说

114
00:04:32,810 --> 00:04:36,370
AEROPIC当前最主要最深度绑定的基础设施

115
00:04:36,370 --> 00:04:38,170
承接方是AWSS

116
00:04:38,170 --> 00:04:40,629
亚马逊官方甚至直接写到

117
00:04:40,629 --> 00:04:45,459
ATHROPIC计划把大部分工作负载都放在AWSS上

118
00:04:45,459 --> 00:04:48,120
这就形成了一个非常关键的闭环

119
00:04:48,120 --> 00:04:51,079
亚马逊呢先把钱投给ANTHROPIC

120
00:04:51,079 --> 00:04:52,060
AERROPIC呢

121
00:04:52,060 --> 00:04:56,819
再拿着钱回来买AWS算力存储和网络资源

122
00:04:56,819 --> 00:04:59,339
钱表面上是投出去了

123
00:04:59,339 --> 00:05:02,750
但实际上又通过基础设施费用流了回来

124
00:05:02,750 --> 00:05:05,250
而且这已经被数字验证过

125
00:05:05,250 --> 00:05:06,529
2025年9月

126
00:05:06,529 --> 00:05:10,180
ATHROPI的估算收入大约是25.5亿美元

127
00:05:10,180 --> 00:05:12,060
但他支付给AWSS

128
00:05:12,060 --> 00:05:15,250
计算费用却高达26.6亿美元

129
00:05:15,250 --> 00:05:20,000
也就是说他赚到的钱几乎转头就花回了AWS

130
00:05:20,000 --> 00:05:21,399
按ANTHROPIC

131
00:05:21,399 --> 00:05:24,370
当前300余枚的年化收入体量来测算

132
00:05:24,370 --> 00:05:27,370
只要未来大约40%到50%

133
00:05:27,370 --> 00:05:30,449
的计算需求继续留在AWS

134
00:05:30,449 --> 00:05:33,290
亚马逊每年就有机会多拿到100亿美元

135
00:05:33,290 --> 00:05:34,970
以上的直接收入体量

136
00:05:34,970 --> 00:05:36,129
所以这层来看

137
00:05:36,129 --> 00:05:39,209
亚马逊成功的把ANTHROPIC变成一个

138
00:05:39,209 --> 00:05:42,750
会持续给AWSS交钱的超级客户

139
00:05:42,750 --> 00:05:45,430
第二层呢打破芯片霸权

140
00:05:45,430 --> 00:05:46,870
原收入再高呢

141
00:05:46,870 --> 00:05:49,029
如果底层永远依赖英伟达利润

142
00:05:49,029 --> 00:05:50,908
天花板呢就始终在那里

143
00:05:50,908 --> 00:05:52,309
你可以盖很大的楼

144
00:05:52,309 --> 00:05:55,728
但最关键的钢筋水泥吊车都在别人手里

145
00:05:55,728 --> 00:05:58,889
所以亚马逊真正想做的不只是继续做云

146
00:05:58,889 --> 00:06:01,769
而是继ANTHROPIC这种最顶级的模型

147
00:06:01,769 --> 00:06:02,189
客户

148
00:06:02,189 --> 00:06:06,279
把自己的训练芯片TRALIUM和推理芯片INFERATION

149
00:06:06,279 --> 00:06:07,879
真正推上主战场

150
00:06:07,879 --> 00:06:10,199
ATHROPIC在这里扮演的不只是客户

151
00:06:10,199 --> 00:06:11,709
更像一块试金石

152
00:06:11,709 --> 00:06:14,470
他不仅长期承诺使用亚马逊自研芯片

153
00:06:14,470 --> 00:06:17,360
来做训练和推理的一部分核心工作负载

154
00:06:17,360 --> 00:06:21,189
还深度参与了芯片和软件站的协同优化

155
00:06:21,189 --> 00:06:23,170
这意味着ANTHROPIC的存在

156
00:06:23,170 --> 00:06:25,930
等于帮亚马逊向全市场证明一件事

157
00:06:25,930 --> 00:06:29,170
最先进的大模型不一定只能跑在英伟达身上

158
00:06:29,170 --> 00:06:31,629
也可以跑在亚马逊自己的芯片上

159
00:06:31,629 --> 00:06:33,829
而且呢这已经有数据支持

160
00:06:33,829 --> 00:06:34,550
Training um

161
00:06:34,550 --> 00:06:37,750
在训练环节最高可以带来50%的成本

162
00:06:37,750 --> 00:06:41,139
节省并减少29%的能源消耗

163
00:06:41,139 --> 00:06:44,199
inference two在推理环节上性能更强

164
00:06:44,199 --> 00:06:46,300
每瓦性能提升50%

165
00:06:46,300 --> 00:06:48,800
对ATHROPIC这种模型越来越大

166
00:06:48,800 --> 00:06:50,620
调用越来越密的公司来说

167
00:06:50,620 --> 00:06:51,959
这不是小修小补

168
00:06:51,959 --> 00:06:55,329
而是会直接影响扩张速度和成本结构

169
00:06:55,329 --> 00:06:56,699
更关心的是

170
00:06:56,699 --> 00:07:01,180
亚马逊计划在AWSS体系里集成100万科training

171
00:07:01,180 --> 00:07:04,220
用来长期ANTHROPIC以及其他客户的需求

172
00:07:04,220 --> 00:07:07,339
这说明他不是把自研芯片当备胎

173
00:07:07,339 --> 00:07:09,100
而是借ANTHROPIC

174
00:07:09,100 --> 00:07:12,310
把自己的芯片一步一步推成主力

175
00:07:12,589 --> 00:07:15,720
第三层bedrock的平台化收割

176
00:07:15,720 --> 00:07:18,079
如果说前两层赚的是底层的钱

177
00:07:18,079 --> 00:07:18,920
到了第三层

178
00:07:18,920 --> 00:07:21,459
亚马逊赚的就不只是卖算力的钱了

179
00:07:21,459 --> 00:07:22,829
而是入口的钱

180
00:07:22,829 --> 00:07:25,430
目前cod系列模型被公认为

181
00:07:25,430 --> 00:07:27,519
bad rock上的旗舰级产品

182
00:07:27,519 --> 00:07:29,959
企业想要在业务里接入cloud

183
00:07:29,959 --> 00:07:33,230
最安全最方便的路径就是通过bad rock

184
00:07:33,230 --> 00:07:35,350
这里亚马逊藏了一个杀手锏

185
00:07:35,350 --> 00:07:36,788
极高的利润分成

186
00:07:36,788 --> 00:07:38,149
根据行业分析

187
00:07:38,149 --> 00:07:42,468
通过bad rock转售AERROPIC服务所产生的毛利润中

188
00:07:42,468 --> 00:07:45,480
亚马逊的分成比例高达50%

189
00:07:45,480 --> 00:07:48,300
这意味着亚马逊不仅收了算力的过路费

190
00:07:48,300 --> 00:07:50,759
还要收一笔高额的平台税

191
00:07:50,759 --> 00:07:54,339
这种平台化收割的威力和确定性是非常惊人的

192
00:07:54,339 --> 00:07:56,079
分析机构of预计

193
00:07:56,079 --> 00:07:58,860
ANTHROPIC对AWS的收入贡献

194
00:07:58,860 --> 00:08:01,980
将从2025年的39亿美元

195
00:08:01,980 --> 00:08:06,180
爆炸式增长到2027年的250亿美元

196
00:08:06,180 --> 00:08:07,980
翻6.4倍

197
00:08:07,980 --> 00:08:09,180
更关键的是

198
00:08:09,180 --> 00:08:10,259
安全型护城河

199
00:08:10,259 --> 00:08:13,519
ATHROPIC强调的宪法AI安全特性

200
00:08:13,519 --> 00:08:18,069
极大地契合了医疗金融等受监管行业的刚需

201
00:08:18,069 --> 00:08:21,569
这些巨头通过aw2S销售接入cloud

202
00:08:21,569 --> 00:08:24,389
形成一个很难被打破的生态闭环

203
00:08:24,389 --> 00:08:25,129
第四层

204
00:08:25,129 --> 00:08:27,139
从单王到双王

205
00:08:27,139 --> 00:08:29,019
进入2026年2月

206
00:08:29,019 --> 00:08:32,578
亚马逊宣布对OpenAI进行高达500亿美元的投资

207
00:08:32,578 --> 00:08:37,578
并使aw s成为OpenAI企业平台的独家第三方

208
00:08:37,578 --> 00:08:39,690
云分发合作伙伴

209
00:08:39,690 --> 00:08:42,470
这一步的意义不是又投了一家公司

210
00:08:42,470 --> 00:08:45,509
而是亚马逊开始从一家模型阵营的受益者

211
00:08:45,509 --> 00:08:49,340
变成整个AI基础设施世界里的总入口

212
00:08:49,340 --> 00:08:52,679
原本OpenAI和微软是穿同一条裤子的

213
00:08:52,679 --> 00:08:55,068
但随着双方关系出现裂痕

214
00:08:55,068 --> 00:08:58,369
OpenAI转向亚马逊寻求算力和渠道

215
00:08:58,369 --> 00:09:01,589
现在无论企业选cloud还是选gt

216
00:09:01,589 --> 00:09:03,639
亚马逊都能从中获利

217
00:09:03,639 --> 00:09:04,879
更关键的是

218
00:09:04,879 --> 00:09:10,019
OpenAI承诺在AWS上消耗约2千瓦的缺LIM算力

219
00:09:10,019 --> 00:09:13,840
这极大地加速了亚马逊自研芯片的规模化进程

220
00:09:13,840 --> 00:09:17,250
全球最顶尖的两家模型公司

221
00:09:17,250 --> 00:09:20,309
第五层智能代理带来的持续收费

222
00:09:20,309 --> 00:09:24,169
这种收割逻辑最终将随着aging tik时代的到来

223
00:09:24,169 --> 00:09:25,940
被推向更高的维度

224
00:09:25,940 --> 00:09:27,299
随着cloud4.5

225
00:09:27,299 --> 00:09:30,859
4.6这类具备长城推理能力的模型出现

226
00:09:30,859 --> 00:09:34,399
AI更像一个长时间在线的数字员工

227
00:09:34,399 --> 00:09:36,179
这对亚马逊意味着什么

228
00:09:36,179 --> 00:09:39,200
意味着AWS和bad rock的收费逻辑

229
00:09:39,200 --> 00:09:41,259
会从过去更像按次收费

230
00:09:41,259 --> 00:09:44,539
走向一种更像按驻留时间收费

231
00:09:44,539 --> 00:09:46,649
按任务价值收费的模式

232
00:09:46,649 --> 00:09:48,570
以前模型回答完一句话

233
00:09:48,570 --> 00:09:49,769
资源就释放了

234
00:09:49,769 --> 00:09:51,809
但到了智能代理时代

235
00:09:51,809 --> 00:09:54,549
一个AI可能连续工作几个小时

236
00:09:54,549 --> 00:09:55,259
几天

237
00:09:55,259 --> 00:09:59,299
持续占用算力存储状态管理和上下文环境

238
00:09:59,299 --> 00:10:01,500
这时候卖的就不只是模型调用

239
00:10:01,500 --> 00:10:04,769
而是整套数字劳动力的运行环境

240
00:10:04,769 --> 00:10:07,828
而且亚马逊已经在提前卡这个位置

241
00:10:07,828 --> 00:10:12,089
他和合作伙伴正在推进状态化运行时环境

242
00:10:12,089 --> 00:10:14,849
让AI代理在bad rock中具备持久

243
00:10:14,849 --> 00:10:17,110
记忆和上下文延续能力

244
00:10:17,110 --> 00:10:20,399
这意味着亚马逊不只是卖模型卖云

245
00:10:20,399 --> 00:10:22,539
他开始卖未来智能代理

246
00:10:22,539 --> 00:10:25,980
真正上班所需要的那套底层基础设施

247
00:10:25,980 --> 00:10:27,620
这套递进的霸权逻辑

248
00:10:27,620 --> 00:10:29,980
已经开始在财务报表里体现

249
00:10:29,980 --> 00:10:31,679
根据off ft的预测

250
00:10:31,679 --> 00:10:33,360
随着ANTHROPIC爆发

251
00:10:33,360 --> 00:10:36,460
以及100万颗TRILIUM芯片的逐步集成

252
00:10:36,460 --> 00:10:40,659
AWS在2026年和2027年的收入增速

253
00:10:40,659 --> 00:10:42,279
有望达到30%

254
00:10:42,279 --> 00:10:45,350
高于市场此前的25%的预期

255
00:10:45,350 --> 00:10:46,950
到2027年底

256
00:10:46,950 --> 00:10:48,429
AWS总收入

257
00:10:48,429 --> 00:10:51,360
预计将突破2165亿美元

258
00:10:51,360 --> 00:10:54,120
其中仅ANTHROPIC相关业务的贡献

259
00:10:54,120 --> 00:10:56,568
就可能占到11%以上

260
00:10:56,568 --> 00:10:58,889
美的测算也印证了这一趋势

261
00:10:58,889 --> 00:11:01,299
仅ANTHROPIC1家客户

262
00:11:01,299 --> 00:11:05,139
在2026年一季度的贡献就高达13亿美元

263
00:11:05,139 --> 00:11:07,659
这一数字甚至超过了市场

264
00:11:07,659 --> 00:11:11,279
对整个aw s当季增长额的预期

265
00:11:11,279 --> 00:11:14,269
所以把整件事总结成一句话

266
00:11:14,269 --> 00:11:16,269
就是ANTHROPI爆发

267
00:11:16,269 --> 00:11:18,549
表面上成就的是一家模型公司

268
00:11:18,549 --> 00:11:20,690
但真正持续兑现的是

269
00:11:20,690 --> 00:11:23,980
亚马逊在AI下半场对云平台

270
00:11:23,980 --> 00:11:27,139
自研芯片和运行环境的全面卡位

271
00:11:27,139 --> 00:11:28,220
如果你看到这里

272
00:11:28,220 --> 00:11:28,820
已经发现

273
00:11:28,820 --> 00:11:31,500
亚马逊根本不是你原来以为的那个亚马逊

274
00:11:31,500 --> 00:11:33,120
那就顺手点赞订阅

275
00:11:33,120 --> 00:11:35,450
如果这期视频点赞能过1000

276
00:11:35,450 --> 00:11:38,549
我会把手上最新调研的那批高弹性

277
00:11:38,549 --> 00:11:41,870
真正有机会走出大行情的公司加速整理出来

278
00:11:41,870 --> 00:11:43,460
第一时间做给你看

279
00:11:43,460 --> 00:11:44,320
说到这里

280
00:11:44,320 --> 00:11:46,299
亚马逊讲的其实是AI

281
00:11:46,299 --> 00:11:49,120
下半场里最先兑现利润的逻辑

282
00:11:49,120 --> 00:11:52,399
那接下来我们要看的就是AI下半场里

283
00:11:52,399 --> 00:11:55,539
最有可能率先被重估的硬件方向

284
00:11:55,539 --> 00:11:58,490
数据中心的需求结构正在发生变化

285
00:11:58,490 --> 00:12:01,029
过去传统AI数据中心里

286
00:12:01,029 --> 00:12:05,600
CPU和GPU的配比大约是一比4~1比八

287
00:12:05,600 --> 00:12:07,820
但到了agent推理时代

288
00:12:07,820 --> 00:12:11,789
这个比例正在向一比一甚至一比二靠近

289
00:12:11,789 --> 00:12:14,929
而且变化还不只是配比重估

290
00:12:14,929 --> 00:12:17,960
CPU的总需求也在同步放大

291
00:12:17,960 --> 00:12:20,019
传统数据中心每1G瓦

292
00:12:20,019 --> 00:12:22,799
大约需要3000万个CPU核心

293
00:12:22,799 --> 00:12:25,080
可一旦进入agent的场景

294
00:12:25,080 --> 00:12:28,279
这个数字可能直接升到1.2亿格

295
00:12:28,279 --> 00:12:30,360
相当于一下放大了四倍

296
00:12:30,360 --> 00:12:31,950
我们前面已经讲过

297
00:12:31,950 --> 00:12:35,149
AI1旦从回答问题走向执行任务市场

298
00:12:35,149 --> 00:12:38,230
真正开始重估的就不再只是单点算力

299
00:12:38,230 --> 00:12:40,849
而是整套系统的运行能力

300
00:12:40,849 --> 00:12:42,729
所以接下来真正重要的问题

301
00:12:42,729 --> 00:12:44,609
不是CPU会不会变重要

302
00:12:44,609 --> 00:12:48,509
而是这条逻辑最先会兑现到谁身上

303
00:12:48,509 --> 00:12:50,009
如果只看这个问题

304
00:12:50,009 --> 00:12:51,710
我会先看AMD

305
00:12:51,710 --> 00:12:53,029
原因呢很简单

306
00:12:53,029 --> 00:12:56,509
这轮变化对整个CPU产业链当然都是利好

307
00:12:56,509 --> 00:12:58,029
英特尔呢会受益

308
00:12:58,029 --> 00:13:00,690
原厂商自研CPU也会受益

309
00:13:00,690 --> 00:13:02,490
但如果只看二级市场

310
00:13:02,490 --> 00:13:04,649
谁最有机会把需求抬升

311
00:13:04,649 --> 00:13:06,690
最快变成业绩弹性

312
00:13:06,690 --> 00:13:09,269
AMD的位置明显更舒服

313
00:13:09,269 --> 00:13:09,889
为什么

314
00:13:09,889 --> 00:13:11,909
因为AMD吃的不是单一逻辑

315
00:13:11,909 --> 00:13:14,710
而是两层甚至三层逻辑叠加

316
00:13:14,710 --> 00:13:18,690
第一层是它本来就在服务器CPU这条线上

317
00:13:18,690 --> 00:13:20,578
占着非常关键的位置

318
00:13:20,578 --> 00:13:23,619
AI下半场一旦真正落地到企业流程

319
00:13:23,619 --> 00:13:26,849
数据调度任务编排状态管理数据库

320
00:13:26,849 --> 00:13:28,168
访问系统调用

321
00:13:28,168 --> 00:13:30,169
这些工作都会变得更重

322
00:13:30,169 --> 00:13:33,730
而这些环节本来就不是GPU1家独大的世界

323
00:13:33,730 --> 00:13:35,809
谁能提供更强更高密度

324
00:13:35,809 --> 00:13:37,850
更高效率的服务器CPU

325
00:13:37,850 --> 00:13:40,710
谁就更容易先把这部需求接住

326
00:13:40,710 --> 00:13:43,029
AMD呢正好站在这里

327
00:13:43,029 --> 00:13:45,220
第二层呢更关键

328
00:13:45,220 --> 00:13:47,659
AMD不是一家纯CPU公司

329
00:13:47,659 --> 00:13:49,840
它是一手抓着服务器CPU

330
00:13:49,840 --> 00:13:52,250
另一手又有自己的AIGPU

331
00:13:52,250 --> 00:13:53,090
这意味着

332
00:13:53,090 --> 00:13:56,590
如果未来AI数据中心真的从单一GPU逻辑

333
00:13:56,590 --> 00:13:59,490
走向GPU加CPU协同逻辑

334
00:13:59,490 --> 00:14:02,009
那AMD的弹性就会比很多

335
00:14:02,009 --> 00:14:04,700
只卡住单一环节的玩家更大

336
00:14:04,700 --> 00:14:09,100
说白了市场过去看AAMD更多是在问一个问题

337
00:14:09,100 --> 00:14:12,379
他能不能在GPU上追上英伟达

338
00:14:12,379 --> 00:14:13,679
但到了现在

339
00:14:13,679 --> 00:14:16,700
这已经不是看MD最核心的角度了

340
00:14:16,700 --> 00:14:19,299
因为AMD真正值得重估的地方

341
00:14:19,299 --> 00:14:21,620
未必只是GPU追到哪里

342
00:14:21,620 --> 00:14:25,899
而是他同时卡住了CPU和GPU两个关键位置

343
00:14:25,899 --> 00:14:29,990
一旦数据中心的定价逻辑开始从谁的显卡最强

344
00:14:29,990 --> 00:14:33,230
转向谁能把整套系统真正跑起来

345
00:14:33,230 --> 00:14:35,289
AMD的估值逻辑就会变

346
00:14:35,289 --> 00:14:39,490
所以过去AAMD更像一个AIGPU替代品的故事

347
00:14:39,490 --> 00:14:40,549
但到了现在

348
00:14:40,549 --> 00:14:44,720
他开始更像一个AI运行时代的系统级受益者

349
00:14:44,720 --> 00:14:48,440
而且这件事最近已经开始有新的验证了

350
00:14:48,440 --> 00:14:51,549
AMD本周四大涨7.8%

351
00:14:51,549 --> 00:14:53,210
背后其实是几条逻辑

352
00:14:53,210 --> 00:14:54,590
开始往一起收

353
00:14:54,590 --> 00:14:55,470
一方面

354
00:14:55,470 --> 00:14:59,350
BUSTIN周四把AMD的目标价从235美元

355
00:14:59,350 --> 00:15:01,220
上调到265美元

356
00:15:01,220 --> 00:15:04,250
继续强化服务器业务走强这条主线

357
00:15:04,250 --> 00:15:05,149
另一方面

358
00:15:05,149 --> 00:15:09,980
渠道侧对CPU和GPU的环比反馈也都还在改善

359
00:15:09,980 --> 00:15:12,379
同时市场也开始重新定价

360
00:15:12,379 --> 00:15:14,360
另一条更有想象力的线索

361
00:15:14,360 --> 00:15:18,139
随着顶级模型厂商对算力的需求继续抬升

362
00:15:18,139 --> 00:15:20,000
AMD正在越来越有机会

363
00:15:20,000 --> 00:15:22,889
进入更核心的算力分配体系

364
00:15:22,889 --> 00:15:25,559
smile analysis最新判断提到

365
00:15:25,559 --> 00:15:28,419
AERROPIC之所以把AMD纳入讨论

366
00:15:28,419 --> 00:15:30,779
核心原因就在于它需要更多算力

367
00:15:30,779 --> 00:15:35,440
而AMD的加入能够直接扩大它可调用的总算力尺

368
00:15:35,440 --> 00:15:37,220
如果这条线继续往下走

369
00:15:37,220 --> 00:15:40,669
那M的意义就不再只是有没有机会追英伟达

370
00:15:40,669 --> 00:15:43,110
而是开始进入顶级模型公司

371
00:15:43,110 --> 00:15:45,200
真实的算力调度体系

372
00:15:45,200 --> 00:15:46,700
这一点呢非常重要

373
00:15:46,700 --> 00:15:50,279
因为它说明未来真正支撑AMDGPU业务的

374
00:15:50,279 --> 00:15:51,419
不再只是单一客户

375
00:15:51,419 --> 00:15:52,259
难一订单

376
00:15:52,259 --> 00:15:56,409
而是顶级模型厂对第二供应体系的集体需求

377
00:15:56,409 --> 00:15:58,110
从更大的格局看

378
00:15:58,110 --> 00:15:59,509
Open ai meta

379
00:15:59,509 --> 00:16:02,220
再加上现在被市场讨论的ANTHROPIC

380
00:16:02,220 --> 00:16:02,940
正在一起

381
00:16:02,940 --> 00:16:07,659
把mi450在2027年的需求底盘做后

382
00:16:07,659 --> 00:16:10,840
这对AMD来说是非常关键的信号

383
00:16:10,840 --> 00:16:14,578
第三层也是市场很容易忽略的一点

384
00:16:14,578 --> 00:16:18,198
AMD这条线最吸引人的不只是他能受益

385
00:16:18,198 --> 00:16:21,519
而是它的受益方式更容易被市场快速定价

386
00:16:21,519 --> 00:16:22,759
什么意思呢

387
00:16:22,759 --> 00:16:26,019
因为亚马逊那条线市场看的是平台税

388
00:16:26,019 --> 00:16:28,320
生态绑定长期基础设施

389
00:16:28,320 --> 00:16:29,759
收益逻辑更强

390
00:16:29,759 --> 00:16:32,320
但兑现节奏偏平台型

391
00:16:32,320 --> 00:16:35,240
而AMD这条线一旦市场真的确认

392
00:16:35,240 --> 00:16:37,809
CPU在AI下半场不只是配角

393
00:16:37,809 --> 00:16:39,750
它的业绩弹性和估值弹性

394
00:16:39,750 --> 00:16:43,350
会更容易在股价上被快速反映出来

395
00:16:43,350 --> 00:16:44,269
这也是为什么

396
00:16:44,269 --> 00:16:46,720
我会把亚马逊和AMD看成

397
00:16:46,720 --> 00:16:48,799
两种完全不同的赚钱逻辑

398
00:16:48,799 --> 00:16:50,419
亚马逊赚的是最稳

399
00:16:50,419 --> 00:16:51,559
最先兑现的钱

400
00:16:51,559 --> 00:16:53,360
AMD赚的是结构变化

401
00:16:53,360 --> 00:16:54,539
一旦被确认之后

402
00:16:54,539 --> 00:16:56,220
最容易被重估的钱

403
00:16:56,220 --> 00:16:57,490
那英特尔呢

404
00:16:57,490 --> 00:16:59,450
英特尔当然不是没机会

405
00:16:59,450 --> 00:17:01,409
只要CPU需求真的抬升

406
00:17:01,409 --> 00:17:03,590
作为传统服务器CPU龙头

407
00:17:03,590 --> 00:17:05,328
它理论上一定受益

408
00:17:05,328 --> 00:17:08,019
只不过两条逻辑还是不一样

409
00:17:08,019 --> 00:17:11,940
AMD这条线更偏向需求抬升之后的直接受益

410
00:17:11,940 --> 00:17:13,240
而英特尔这条线

411
00:17:13,240 --> 00:17:16,098
市场交易的更多的还是修复预期

412
00:17:16,098 --> 00:17:17,578
包括制程推进

413
00:17:17,578 --> 00:17:19,549
产品兑现节奏落地

414
00:17:19,549 --> 00:17:20,529
换句话说

415
00:17:20,529 --> 00:17:22,750
这轮变化里两家都会受益

416
00:17:22,750 --> 00:17:24,769
但AAMD更偏兑现

417
00:17:24,769 --> 00:17:26,549
英特尔更偏修复

418
00:17:26,549 --> 00:17:27,589
这也是为什么

419
00:17:27,589 --> 00:17:30,430
如果你问我在AI下半场这条线里

420
00:17:30,430 --> 00:17:32,970
谁是最值得重点讲的那个名字

421
00:17:32,970 --> 00:17:35,470
我还是会先看AMD视频

422
00:17:35,470 --> 00:17:36,670
最后我们总结一下

423
00:17:36,670 --> 00:17:40,659
亚马逊和AMD代表的是两种完全不同的赚钱逻辑

424
00:17:40,659 --> 00:17:43,479
一个赚的是AI生态里最稳最先兑现的钱

425
00:17:43,479 --> 00:17:46,499
一个赌的是AI下半场硬件结构变化

426
00:17:46,499 --> 00:17:48,059
带来的重估空间

427
00:17:48,059 --> 00:17:50,769
所以最后我也把这个问题留给你

428
00:17:50,769 --> 00:17:52,609
如果现在二选一

429
00:17:52,609 --> 00:17:56,549
你更愿意压住亚马逊这种先收钱的底层平台

430
00:17:56,549 --> 00:17:59,730
还是AMD这种结构变化的弹性受益者

431
00:17:59,730 --> 00:18:02,069
为什么评论区告诉我你的答案

432
00:18:02,069 --> 00:18:05,250
我想看看到底是稳着收钱更有吸引力

433
00:18:05,250 --> 00:18:07,779
还是弹性重估更有杀伤力

434
00:18:07,779 --> 00:18:08,640
感谢观看

435
00:18:08,640 --> 00:18:09,519
下期再见

评论采集失败

采集完成，该视频无弹幕