美股 英伟达最强竞争对手要上市，20倍超额认购，值不值买？ #CBRS

1
00:00:00,040 --> 00:00:02,240
美股AI量化分析工具

2
00:00:02,240 --> 00:00:02,899
大家好

3
00:00:02,899 --> 00:00:04,360
我们是美股投资网

4
00:00:04,360 --> 00:00:07,108
用大数据驱动你的交易决策

5
00:00:07,108 --> 00:00:10,069
英伟达的护城河可能正在被撕开

6
00:00:10,069 --> 00:00:11,519
一道真正的裂口

7
00:00:11,519 --> 00:00:12,380
本周四

8
00:00:12,380 --> 00:00:15,160
一家AI算力公司three birds

9
00:00:15,160 --> 00:00:17,960
代号CBRS将在美股上市

10
00:00:17,960 --> 00:00:20,600
他做的芯片连英伟达自己都不敢想

11
00:00:20,600 --> 00:00:24,079
英伟达呢是把成千上万颗GPU连在一起

12
00:00:24,079 --> 00:00:25,350
拼算力集群

13
00:00:25,350 --> 00:00:26,870
three ripers更激进

14
00:00:26,870 --> 00:00:30,289
直接把一整片晶圆做成一颗超级芯片

15
00:00:30,289 --> 00:00:31,350
打个比方

16
00:00:31,350 --> 00:00:33,829
英伟达在拼乐高sweepers

17
00:00:33,829 --> 00:00:36,240
想直接造出一整面算力墙

18
00:00:36,240 --> 00:00:39,719
现在呢IPO路演已经把市场情绪点着了

19
00:00:39,719 --> 00:00:43,250
这周一因为订单需求直接超出发行量20倍

20
00:00:43,250 --> 00:00:44,049
Three rivers

21
00:00:44,049 --> 00:00:47,310
自己把发行价从115~125美元

22
00:00:47,310 --> 00:00:50,018
提高到了150~160美元

23
00:00:50,018 --> 00:00:52,298
发行量也加了200万股

24
00:00:52,298 --> 00:00:53,399
按上限算

25
00:00:53,399 --> 00:00:55,820
估值冲到了350亿美元

26
00:00:55,820 --> 00:00:57,079
在这个行情下

27
00:00:57,079 --> 00:00:58,380
还敢主动提价

28
00:00:58,380 --> 00:00:59,799
提了还被抢光

29
00:00:59,799 --> 00:01:02,219
你就知道机构有多看好他

30
00:01:02,219 --> 00:01:04,540
他手里还握着两个最强背书

31
00:01:04,540 --> 00:01:06,000
OpenAI和亚马逊

32
00:01:06,000 --> 00:01:09,840
three birds跟OpenAI签了超过200亿美元的照付

33
00:01:09,840 --> 00:01:10,789
不易合同

34
00:01:10,789 --> 00:01:12,728
1000就到2028年

35
00:01:12,728 --> 00:01:13,448
另一边

36
00:01:13,448 --> 00:01:17,609
AWSS宣布把three birds芯片搬进自家数据中心

37
00:01:17,609 --> 00:01:20,120
跟自研的trillium芯片一起干活

38
00:01:20,120 --> 00:01:21,599
一家没上市的公司

39
00:01:21,599 --> 00:01:24,180
同时拿下AI巨头和云巨头的订单

40
00:01:24,180 --> 00:01:26,959
这个信号不需要我再多解释了

41
00:01:26,959 --> 00:01:29,920
今天这期视频我会讲清楚四件事

42
00:01:29,920 --> 00:01:31,540
strippers到底是什么

43
00:01:31,540 --> 00:01:33,599
他凭什么敢挑战英伟达

44
00:01:33,599 --> 00:01:38,579
第二它的WSSE芯片和英伟达GPU本质区别在哪

45
00:01:38,579 --> 00:01:41,920
第三他在商业模式靠什么赚钱

46
00:01:41,920 --> 00:01:43,739
财务数据增长速度

47
00:01:43,739 --> 00:01:45,640
估值逻辑又该怎么看

48
00:01:45,640 --> 00:01:50,629
第四散户怎么看待这次IPO上市首日能不能追

49
00:01:50,629 --> 00:01:53,569
背后最大的机会和风险分别是什么

50
00:01:53,569 --> 00:01:57,909
这可能是今年AI硬件赛道最受关注的一支新股

51
00:01:57,909 --> 00:02:00,950
但越火的股票越不能只看故事视频

52
00:02:00,950 --> 00:02:02,909
最后我会把机会风险

53
00:02:02,909 --> 00:02:05,510
估值和参与策略一次性讲透

54
00:02:05,510 --> 00:02:06,750
千万别走开

55
00:02:06,750 --> 00:02:08,550
three birds到底是干嘛的

56
00:02:08,550 --> 00:02:10,079
一句话就可以讲清楚

57
00:02:10,079 --> 00:02:12,998
别人是把晶圆切成很多小芯片

58
00:02:12,998 --> 00:02:17,038
three birth是直接把整片晶圆当成一颗超级芯片

59
00:02:17,038 --> 00:02:18,909
采用什么是晶圆

60
00:02:18,909 --> 00:02:21,610
就是那块圆圆的亮晶晶的硅片

61
00:02:21,610 --> 00:02:23,270
直径呢300mm

62
00:02:23,270 --> 00:02:25,349
正常的芯片厂是怎么做呢

63
00:02:25,349 --> 00:02:28,050
先在晶圆上做出几百颗芯片

64
00:02:28,050 --> 00:02:29,610
然后像切披萨一样

65
00:02:29,610 --> 00:02:31,770
咔咔咔切成一小块一小块

66
00:02:31,770 --> 00:02:33,949
每一小块呢就是一颗芯片

67
00:02:33,949 --> 00:02:35,889
我们平时用的手机芯片

68
00:02:35,889 --> 00:02:36,990
电脑的CPU

69
00:02:36,990 --> 00:02:39,909
英伟达的GPU大体呢都是这个逻辑

70
00:02:39,909 --> 00:02:42,169
那英伟达的AI算力为什么这么强

71
00:02:42,169 --> 00:02:43,710
因为他不止卖芯片

72
00:02:43,710 --> 00:02:45,349
他卖的是整套系统

73
00:02:45,349 --> 00:02:46,789
GPU显存

74
00:02:46,789 --> 00:02:48,469
N类link交换机

75
00:02:48,469 --> 00:02:49,969
光模块服务器

76
00:02:49,969 --> 00:02:53,770
再加上库的软件生态几乎全给你打包好了

77
00:02:53,770 --> 00:02:55,569
客户买到的不是一颗芯片

78
00:02:55,569 --> 00:02:58,479
而是一座可以直接开工的AI工厂

79
00:02:58,479 --> 00:03:01,500
但是这个模式有一个逃不掉的代价

80
00:03:01,500 --> 00:03:02,659
GPU越多

81
00:03:02,659 --> 00:03:05,960
数据就越需要在芯片之间来回传输

82
00:03:05,960 --> 00:03:07,919
你可以想象一个超级工厂

83
00:03:07,919 --> 00:03:10,050
每个车间呢都非常先进

84
00:03:10,050 --> 00:03:13,610
但原材料要在几百个车间来回运

85
00:03:13,610 --> 00:03:15,090
最后卡住你的是什么

86
00:03:15,090 --> 00:03:16,469
不是机器不够快

87
00:03:16,469 --> 00:03:17,789
而是物流太慢

88
00:03:17,789 --> 00:03:18,960
AI也一样

89
00:03:18,960 --> 00:03:20,219
数据搬来搬去

90
00:03:20,219 --> 00:03:21,439
延迟就上来了

91
00:03:21,439 --> 00:03:22,900
耗电也上来了

92
00:03:22,900 --> 00:03:24,840
带宽压力也跟着爆了

93
00:03:24,840 --> 00:03:27,379
ZABUDS的创始人叫ANDREWFELDMAN

94
00:03:27,379 --> 00:03:29,900
这人早年做过一家服务器公司

95
00:03:29,900 --> 00:03:31,340
后来卖给了AMD

96
00:03:31,340 --> 00:03:33,289
他太清楚问题的本质了

97
00:03:33,289 --> 00:03:36,629
算力系统瓶颈从来不在单颗芯片跑多快

98
00:03:36,629 --> 00:03:38,229
而在数据怎么流动

99
00:03:38,229 --> 00:03:38,930
于是呢

100
00:03:38,930 --> 00:03:40,750
他想了一个特别极端的办法

101
00:03:40,750 --> 00:03:43,250
既然跨芯片传数据那么麻烦

102
00:03:43,250 --> 00:03:45,870
那我干脆不切了一整片晶圆

103
00:03:45,870 --> 00:03:47,349
直接做成一颗芯片

104
00:03:47,349 --> 00:03:50,300
这就是WSE晶圆级引擎

105
00:03:50,300 --> 00:03:52,639
第三代叫WS1杠三

106
00:03:52,639 --> 00:03:53,080
装

107
00:03:53,080 --> 00:03:54,579
在cs3系统里

108
00:03:54,579 --> 00:03:55,878
核心思路一句话

109
00:03:55,878 --> 00:03:58,959
把所有东西尽量塞在一块硅片上

110
00:03:58,959 --> 00:04:00,618
让数据少往外跑

111
00:04:00,618 --> 00:04:02,258
菲奥特曼自己说过一句话

112
00:04:02,258 --> 00:04:03,419
我印象特别深

113
00:04:03,419 --> 00:04:05,780
我们发明了过去75年

114
00:04:05,780 --> 00:04:09,120
计算机行业里被认为最不可能实现的技术

115
00:04:09,120 --> 00:04:11,500
把处理器做到餐盘那么大

116
00:04:11,500 --> 00:04:13,800
而传统处理器只有邮票大小

117
00:04:13,800 --> 00:04:14,860
从数据上看

118
00:04:14,860 --> 00:04:16,629
这条路线还不错

119
00:04:16,629 --> 00:04:19,670
cs3的峰值算力是125

120
00:04:19,670 --> 00:04:20,629
Petterflops

121
00:04:20,629 --> 00:04:24,379
单科英伟达B2版是4.5PETTERFLOPS

122
00:04:24,379 --> 00:04:27,639
所以cs31颗顶28颗B200

123
00:04:27,639 --> 00:04:30,040
但英伟达不是只卖单颗芯片

124
00:04:30,040 --> 00:04:31,959
它卖的是整柜系统

125
00:04:31,959 --> 00:04:36,540
如果你拿cs3和英伟达的整柜GB200NVL

126
00:04:36,540 --> 00:04:40,329
72比后者峰值算力是360

127
00:04:40,329 --> 00:04:42,829
PETERFPS比cs3还要高

128
00:04:42,829 --> 00:04:46,509
所以你不能简单说three birth算力比英伟达强

129
00:04:46,509 --> 00:04:49,769
真正让three birth与众不同的是另外两个东西

130
00:04:49,769 --> 00:04:52,350
内存带宽和架构设计

131
00:04:52,350 --> 00:04:54,970
我们先了解推理是怎么工作的

132
00:04:54,970 --> 00:04:58,970
所有模型权重都需要从内存传输到计算单元

133
00:04:58,970 --> 00:05:01,829
整个过程是逐词顺序进行的

134
00:05:01,829 --> 00:05:04,410
一个700亿参数的模型权重

135
00:05:04,410 --> 00:05:08,509
数据量相当于大约100部高清电影生成

136
00:05:08,509 --> 00:05:09,129
每一个词

137
00:05:09,129 --> 00:05:13,389
都需要把这么多的数据从内存搬到计算单元

138
00:05:13,389 --> 00:05:15,089
一个1000字的回答

139
00:05:15,089 --> 00:05:18,509
相当于传输了10万部高清电影的数据量

140
00:05:18,509 --> 00:05:20,610
在传统GPU上

141
00:05:20,610 --> 00:05:24,069
从HBM传输数据到计算单元的速度

142
00:05:24,069 --> 00:05:25,800
大约是8tb每秒

143
00:05:25,800 --> 00:05:27,680
而在three pers的晶圆上

144
00:05:27,680 --> 00:05:30,819
这个速度超过了二点10000tb每秒

145
00:05:30,819 --> 00:05:32,788
快了2600多倍

146
00:05:32,788 --> 00:05:34,149
为什么能做到

147
00:05:34,149 --> 00:05:36,528
因为CEREBRAS用的是s RM

148
00:05:36,528 --> 00:05:39,759
不是h b m s ram本身速度就快

149
00:05:39,759 --> 00:05:41,800
但传统上它的问题是

150
00:05:41,800 --> 00:05:44,939
每平方毫米能做的容量有限

151
00:05:44,939 --> 00:05:47,459
zebras的解法很简单粗暴

152
00:05:47,459 --> 00:05:49,079
我芯片做的足够大

153
00:05:49,079 --> 00:05:51,180
就能放足够多的s ram

154
00:05:51,180 --> 00:05:54,610
这就是整片晶圆不切割的核心价值所在

155
00:05:54,610 --> 00:05:55,930
再看另一个数据

156
00:05:55,930 --> 00:06:01,069
cs3的内存带宽是26750tb每秒

157
00:06:01,069 --> 00:06:05,589
而GB2版NVL72只有130tb每秒

158
00:06:05,589 --> 00:06:07,029
差了200多倍

159
00:06:07,029 --> 00:06:09,418
three birds自己总结的很直接

160
00:06:09,418 --> 00:06:12,499
我们的处理器比最大的GPU大58倍

161
00:06:12,499 --> 00:06:15,338
内存带宽高出2500倍以上

162
00:06:15,338 --> 00:06:18,259
这正是高速推理的基础好

163
00:06:18,259 --> 00:06:19,839
那实际跑任务呢

164
00:06:19,839 --> 00:06:22,160
three bas官方公布过一个数据

165
00:06:22,160 --> 00:06:25,220
cs3跑拉玛3.1710B推理

166
00:06:25,220 --> 00:06:28,180
比英伟达D7XB200快21倍

167
00:06:28,180 --> 00:06:30,000
成本低32%

168
00:06:30,000 --> 00:06:31,560
功耗也低1/3

169
00:06:31,560 --> 00:06:35,589
第三方测试也有PAMA3.18B推理

170
00:06:35,589 --> 00:06:37,350
每秒1800个token

171
00:06:37,350 --> 00:06:39,589
是常规GPU方案的20倍

172
00:06:39,589 --> 00:06:42,329
还有一个更夸张的碳捕获模拟

173
00:06:42,329 --> 00:06:44,529
比H100快了200多倍

174
00:06:44,529 --> 00:06:48,139
但是注意这些数字都是在特定任务下跑出来的

175
00:06:48,139 --> 00:06:50,259
您要是换成复杂的混合运算

176
00:06:50,259 --> 00:06:52,379
或者重度依赖库代生态的活儿

177
00:06:52,379 --> 00:06:54,399
英伟达依然稳如泰山

178
00:06:54,399 --> 00:06:56,060
这时候你可能会问

179
00:06:56,060 --> 00:06:58,329
那GPU为什么不也跑快一点

180
00:06:58,329 --> 00:07:01,069
three birth揭示了一个很值得注意的现象

181
00:07:01,069 --> 00:07:02,990
叫做GPU的速度陷阱

182
00:07:02,990 --> 00:07:05,529
GPU在低速运行时效率极高

183
00:07:05,529 --> 00:07:08,089
比如每秒35个token的慢速

184
00:07:08,089 --> 00:07:11,129
下一台GB200NNL72

185
00:07:11,129 --> 00:07:13,959
可以同时支持数万名用户

186
00:07:13,959 --> 00:07:15,129
成本极低

187
00:07:15,129 --> 00:07:17,189
但当用户要求提速

188
00:07:17,189 --> 00:07:19,610
比如每秒270个token

189
00:07:19,610 --> 00:07:21,949
这台机器只能服务一名用户

190
00:07:21,949 --> 00:07:24,610
成本瞬间飙升到天上去

191
00:07:24,610 --> 00:07:26,370
three rivers自己的说法是

192
00:07:26,370 --> 00:07:29,250
从每秒100~150个token开始

193
00:07:29,250 --> 00:07:32,569
GPU变得极其昂贵且功耗效率低下

194
00:07:32,569 --> 00:07:35,959
而我们的成本与功耗仅为其极小一部分

195
00:07:35,959 --> 00:07:38,939
没有任何数量的GPU能达到我们的速度

196
00:07:38,939 --> 00:07:41,319
GPU根本不可能实现这种性能

197
00:07:41,319 --> 00:07:42,459
在这种速度下

198
00:07:42,459 --> 00:07:44,120
我们毫无竞争对手

199
00:07:44,120 --> 00:07:45,120
那问题来了

200
00:07:45,120 --> 00:07:47,759
为什么一定要追求这种极端速度

201
00:07:47,759 --> 00:07:49,560
对普通用户来说

202
00:07:49,560 --> 00:07:52,259
虽然每秒35个token的阅读速度

203
00:07:52,259 --> 00:07:53,649
其实已经够了

204
00:07:53,649 --> 00:07:55,209
但现实很残酷

205
00:07:55,209 --> 00:07:56,649
哪怕只卖一点点

206
00:07:56,649 --> 00:07:57,850
用户就会流失

207
00:07:57,850 --> 00:08:00,670
就像YC创始人program说的

208
00:08:00,670 --> 00:08:02,959
如果ChatGPT反应变慢

209
00:08:02,959 --> 00:08:06,100
他回过头去用谷歌的AI就会多出一倍

210
00:08:06,100 --> 00:08:08,819
这揭示了一个残酷的商业真相

211
00:08:08,819 --> 00:08:11,439
快就是最硬的护城河

212
00:08:11,439 --> 00:08:13,660
some奥特曼追求的极致速度

213
00:08:13,660 --> 00:08:18,139
本质上是在消除用户切换平台的任何微小念头

214
00:08:18,139 --> 00:08:19,860
但这只是硬币的一面

215
00:08:19,860 --> 00:08:21,139
另一面更重要

216
00:08:21,139 --> 00:08:24,959
整个AI产业的算力命脉几乎都攥在英伟达手里

217
00:08:24,959 --> 00:08:27,959
这不是任何一个大模型公司愿意看到的

218
00:08:27,959 --> 00:08:31,699
所以OpenAI花这个钱是两个原因叠加的结果

219
00:08:31,699 --> 00:08:35,860
第一three birds确实快在极致速度这条赛道上

220
00:08:35,860 --> 00:08:38,298
GPU根本跑不过three birds

221
00:08:38,298 --> 00:08:40,658
第二必须降低对英伟达的依赖

222
00:08:40,658 --> 00:08:43,658
把鸡蛋从英伟达一个篮子里分一部分

223
00:08:43,658 --> 00:08:45,120
到three birds这个篮子里

224
00:08:45,120 --> 00:08:45,980
两个原因

225
00:08:45,980 --> 00:08:46,960
一个追求性能

226
00:08:46,960 --> 00:08:47,840
一个追求安全

227
00:08:47,840 --> 00:08:51,159
这就是200多亿美元合同签下来的底层逻辑

228
00:08:51,159 --> 00:08:53,250
第二个大客户呢是亚马逊

229
00:08:53,250 --> 00:08:54,830
2026年3月

230
00:08:54,830 --> 00:08:59,110
AWS宣布将three birds cs3部署到自家数据中心

231
00:08:59,110 --> 00:09:02,289
通过亚马逊bad rock提供服务

232
00:09:02,289 --> 00:09:07,169
方案呢是用自研的TRALIUM芯片处理预填充阶段

233
00:09:07,169 --> 00:09:09,509
用three birds处理解码阶段

234
00:09:09,509 --> 00:09:11,328
各干最擅长的事

235
00:09:11,328 --> 00:09:14,389
AWS计算副总裁david brown的原话

236
00:09:14,389 --> 00:09:17,440
这套组合的推理速度将比现有方案快

237
00:09:17,440 --> 00:09:18,379
一个数量级

238
00:09:18,379 --> 00:09:20,120
一家没上市的公司

239
00:09:20,120 --> 00:09:23,789
同时拿下OpenAI和AWS的订单

240
00:09:23,789 --> 00:09:26,250
背后还有甲骨文metal等客户的关注

241
00:09:26,250 --> 00:09:28,399
可见还是非常有实力的

242
00:09:28,399 --> 00:09:30,779
如果你觉得本视频对你有帮助

243
00:09:30,779 --> 00:09:31,519
老规矩

244
00:09:31,519 --> 00:09:32,980
先点赞再收藏

245
00:09:32,980 --> 00:09:34,389
关键时刻能帮忙

246
00:09:34,389 --> 00:09:37,210
美国热线6263783637

247
00:09:37,210 --> 00:09:39,690
接下来我们看three birds的财务

248
00:09:39,690 --> 00:09:43,320
这家公司最大的亮点是增长速度非常快

249
00:09:43,320 --> 00:09:45,580
它的营收呢从2022年的

250
00:09:45,580 --> 00:09:47,179
2460万美元

251
00:09:47,179 --> 00:09:50,919
增长到2023年的7870万美元

252
00:09:50,919 --> 00:09:53,299
2024年达到2.9亿美元

253
00:09:53,299 --> 00:09:56,269
2025年进一步增长到5.1亿美元

254
00:09:56,269 --> 00:09:58,870
3年时间营收增长了接近20倍

255
00:09:58,870 --> 00:10:01,929
这个速度呢在半导体里公司非常少见

256
00:10:01,929 --> 00:10:06,019
但问题是增长快不代表已经真正赚钱

257
00:10:06,019 --> 00:10:07,100
表面上看

258
00:10:07,100 --> 00:10:07,820
Cl bras

259
00:10:07,820 --> 00:10:11,820
在2025年实现了2.378亿美元的gap

260
00:10:11,820 --> 00:10:12,740
净利润

261
00:10:12,740 --> 00:10:14,799
看起来已经扭亏为盈

262
00:10:14,799 --> 00:10:18,360
但这笔利润里有很大一部分来自会计调整

263
00:10:18,360 --> 00:10:19,559
不是卖芯片

264
00:10:19,559 --> 00:10:21,519
卖算力赚出来的钱

265
00:10:21,519 --> 00:10:22,919
公开资料显示

266
00:10:22,919 --> 00:10:23,840
Siri brass

267
00:10:23,840 --> 00:10:26,279
2020年有一笔约3.6

268
00:10:26,279 --> 00:10:28,639
3亿美元的远期合同负债

269
00:10:28,639 --> 00:10:30,549
公允价值变动收益

270
00:10:30,549 --> 00:10:32,750
如果剔除这些因素

271
00:10:32,750 --> 00:10:36,379
公司2025年仍然录得1.459亿美元

272
00:10:36,379 --> 00:10:37,220
经营亏损

273
00:10:37,220 --> 00:10:40,779
non gap净亏损约7570万美元

274
00:10:40,779 --> 00:10:43,320
所以这里不能简单说它已经盈利

275
00:10:43,320 --> 00:10:44,779
更准确的说法是

276
00:10:44,779 --> 00:10:46,340
财报上看到了盈利

277
00:10:46,340 --> 00:10:48,899
但核心经营还在亏钱

278
00:10:48,899 --> 00:10:50,320
再看毛利率

279
00:10:50,320 --> 00:10:53,659
three birds的毛利率从2022年的12%

280
00:10:53,659 --> 00:10:56,230
提升到了2025年的39%

281
00:10:56,230 --> 00:10:59,350
进步呢很明显说明随着收入扩大

282
00:10:59,350 --> 00:11:00,909
产品交付增加

283
00:11:00,909 --> 00:11:03,480
它的成本结构确实在改善

284
00:11:03,480 --> 00:11:06,960
但39%的毛利率和英伟达相比呢

285
00:11:06,960 --> 00:11:08,120
还有很大的差距

286
00:11:08,120 --> 00:11:09,759
原因呢也不难理解

287
00:11:09,759 --> 00:11:12,039
three birds做的是晶圆级芯片

288
00:11:12,039 --> 00:11:12,980
制造难度高

289
00:11:12,980 --> 00:11:14,500
系统复杂度高

290
00:11:14,500 --> 00:11:17,200
而且现在规模还没有完全跑起来

291
00:11:17,200 --> 00:11:19,080
它要达到更高毛利率

292
00:11:19,080 --> 00:11:21,159
必须靠后续交付规模扩大

293
00:11:21,159 --> 00:11:24,599
制造效率提升以及云服务模式跑通

294
00:11:24,599 --> 00:11:27,479
真正让市场兴奋的是他手里的订单

295
00:11:27,479 --> 00:11:29,119
截至2025年底

296
00:11:29,119 --> 00:11:31,619
three birth披露的待确认订单

297
00:11:31,619 --> 00:11:33,730
达到246亿美元

298
00:11:33,730 --> 00:11:37,730
其中包括与OpenAI相关的约200亿美元合作

299
00:11:37,730 --> 00:11:39,980
也包括亚马逊等客户订单

300
00:11:39,980 --> 00:11:41,379
这个数字呢非常大

301
00:11:41,379 --> 00:11:44,059
说明市场对它的算力确实有需求

302
00:11:44,059 --> 00:11:46,679
但投资者一定要冷静看两个问题

303
00:11:46,679 --> 00:11:49,950
第一呢这些订单不是马上变成收入公司

304
00:11:49,950 --> 00:11:51,509
预计呢到2027年底

305
00:11:51,509 --> 00:11:54,250
大约只能确认其中15%的收入

306
00:11:54,250 --> 00:11:56,389
之后25~48个月

307
00:11:56,389 --> 00:11:58,100
再确认约43%

308
00:11:58,100 --> 00:11:59,580
剩下的还要更久

309
00:11:59,580 --> 00:12:01,639
也就是说订单很大

310
00:12:01,639 --> 00:12:04,809
但钱要分几年慢慢落袋

311
00:12:04,809 --> 00:12:06,809
第二呢交付这些订单之前

312
00:12:06,809 --> 00:12:08,539
公司还要花很多的钱

313
00:12:08,539 --> 00:12:10,200
他要提前采购晶圆

314
00:12:10,200 --> 00:12:11,080
建设系统

315
00:12:11,080 --> 00:12:12,220
部署数据中心

316
00:12:12,220 --> 00:12:15,080
还要解决电力供应链和客服交付的问题

317
00:12:15,080 --> 00:12:16,639
EBS自己也提示

318
00:12:16,639 --> 00:12:18,320
订单能不能变成收入

319
00:12:18,320 --> 00:12:19,480
取决于制造产能

320
00:12:19,480 --> 00:12:21,389
基础设施部署和电力供应

321
00:12:21,389 --> 00:12:25,070
那上市之后我们的核心操盘思路是什么呢

322
00:12:25,070 --> 00:12:26,539
先说我的判断

323
00:12:26,539 --> 00:12:28,899
第一阶段大概是炒情绪

324
00:12:28,899 --> 00:12:31,700
炒它是今年最稀缺的AI硬科技

325
00:12:31,700 --> 00:12:32,360
Ipo

326
00:12:32,360 --> 00:12:36,860
炒OpenAI那200多亿美元的照付不易合同炒

327
00:12:36,860 --> 00:12:39,440
跟AWS的合作炒一个故事

328
00:12:39,440 --> 00:12:42,240
AI算力不一定只有英伟达这一条路

329
00:12:42,240 --> 00:12:44,639
这个阶段情绪主导资金涌入

330
00:12:44,639 --> 00:12:47,120
首日涨100%都有可能

331
00:12:47,120 --> 00:12:48,519
但情绪炒完之后

332
00:12:48,519 --> 00:12:50,480
第二阶段会非常的现实

333
00:12:50,480 --> 00:12:52,120
上面讲的那些风险点

334
00:12:52,120 --> 00:12:54,599
市场会一条一条拿出来对账

335
00:12:54,599 --> 00:12:57,918
订单能不能变成收入750兆瓦的部署

336
00:12:57,918 --> 00:12:59,259
能不能按时交付

337
00:12:59,259 --> 00:13:02,649
毛利率能不能从50%爬到60%

338
00:13:02,649 --> 00:13:04,409
客户集中度会不会出问题

339
00:13:04,409 --> 00:13:06,490
你到底是在补充英伟达

340
00:13:06,490 --> 00:13:08,649
还是有本事抢走他的份额

341
00:13:08,649 --> 00:13:10,330
这些问题上市后

342
00:13:10,330 --> 00:13:12,970
第一个财报季就会被摆到台面上

343
00:13:12,970 --> 00:13:15,049
因为现在市场给three birds的

344
00:13:15,049 --> 00:13:17,559
已经不是普通半导体公司的估值了

345
00:13:17,559 --> 00:13:20,200
而是AI基础设施新路线的估值

346
00:13:20,200 --> 00:13:22,600
这个估值的前提是你代表一条路

347
00:13:22,600 --> 00:13:24,299
而不只是卖了一个产品

348
00:13:24,299 --> 00:13:28,360
如果后面收入兑现跟不上或者交付能力出问题

349
00:13:28,360 --> 00:13:30,559
股价回撤呢会非常快

350
00:13:30,559 --> 00:13:34,419
最后我们来说风险第一呢客户高度集中

351
00:13:34,419 --> 00:13:35,549
这不是秘密

352
00:13:35,549 --> 00:13:37,309
2024年上半年

353
00:13:37,309 --> 00:13:40,509
three birth87%的收入都来自一家公司

354
00:13:40,509 --> 00:13:42,429
阿联酋的G42

355
00:13:42,429 --> 00:13:43,570
到了2025年

356
00:13:43,570 --> 00:13:45,990
G42占比虽然降到了24%

357
00:13:45,990 --> 00:13:47,490
但别急着松口气

358
00:13:47,490 --> 00:13:49,450
M b z u a i

359
00:13:52,889 --> 00:13:54,690
跟G42关系极深

360
00:13:54,690 --> 00:13:57,350
他一家就占了62%的收入

361
00:13:57,350 --> 00:13:58,480
算下来

362
00:13:58,480 --> 00:13:59,399
Three birds

363
00:13:59,399 --> 00:14:01,600
对中东这几个大客户的依赖

364
00:14:01,600 --> 00:14:03,720
仍然接近90%

365
00:14:03,720 --> 00:14:07,539
而且OpenAI那张200亿美元的大单也不是白给的

366
00:14:07,539 --> 00:14:08,200
合同

367
00:14:08,200 --> 00:14:10,429
对交付进度要求非常严格

368
00:14:10,429 --> 00:14:12,230
首批250兆瓦算力

369
00:14:12,230 --> 00:14:13,429
如果交付延迟

370
00:14:13,429 --> 00:14:15,990
three birds面临的就不是少赚点钱的问题

371
00:14:15,990 --> 00:14:18,849
而是违约和信誉双重打击

372
00:14:18,849 --> 00:14:21,188
第二呢供应链压力不小

373
00:14:21,188 --> 00:14:24,048
three birds的晶圆级芯片制造难度非常高

374
00:14:24,048 --> 00:14:26,570
关键制造环节高度依赖台积电

375
00:14:26,570 --> 00:14:28,029
他用的是5NM

376
00:14:28,029 --> 00:14:30,850
不需要最紧张的HBM和covers

377
00:14:30,850 --> 00:14:32,480
这确实是一个优势

378
00:14:32,480 --> 00:14:37,000
但晶圆级芯片本身良率成本交付节奏都很敏感

379
00:14:37,000 --> 00:14:39,139
只要制造或者产能出现波动

380
00:14:39,139 --> 00:14:41,828
都会直接影响毛利率和订单交付

381
00:14:41,828 --> 00:14:42,808
第三呢

382
00:14:42,808 --> 00:14:45,269
软件生态还远远不如英伟达

383
00:14:45,269 --> 00:14:47,549
three birds可以对接PATROCH

384
00:14:47,549 --> 00:14:49,330
TENSORFLOW等主流框架

385
00:14:49,330 --> 00:14:51,330
但要把性能完全榨出来

386
00:14:51,330 --> 00:14:53,950
还是需要适配它自己的软件站

387
00:14:53,950 --> 00:14:57,330
相比英伟达库的几十年积累下来的开发者生态

388
00:14:57,330 --> 00:14:59,090
工具链和客户习惯

389
00:14:59,090 --> 00:15:01,568
three birds短期内很难追上

390
00:15:01,568 --> 00:15:04,828
第四呢是成本和可靠性还需要验证

391
00:15:04,828 --> 00:15:06,168
晶圆级芯片很强

392
00:15:06,168 --> 00:15:07,928
但也意味着供电散热

393
00:15:07,928 --> 00:15:10,879
封装长期稳定性要求都非常高

394
00:15:10,879 --> 00:15:12,899
大规模部署到数据中心之后

395
00:15:12,899 --> 00:15:14,759
能不能长期稳定运行

396
00:15:14,759 --> 00:15:16,279
能不能把成本压下来

397
00:15:16,279 --> 00:15:17,828
还需要时间证明

398
00:15:17,828 --> 00:15:19,808
第五竞争不会停下来

399
00:15:19,808 --> 00:15:23,779
英伟达不会坐着等人挑战blackwell robin

400
00:15:23,779 --> 00:15:25,620
后面一代接一代升级

401
00:15:25,620 --> 00:15:28,730
GPU集群的性能和能效还会继续提高

402
00:15:28,730 --> 00:15:29,830
更重要的是

403
00:15:29,830 --> 00:15:32,068
OpenAI自己也在做芯片

404
00:15:32,068 --> 00:15:34,349
如果未来自研芯片推进顺利

405
00:15:34,349 --> 00:15:37,629
今天的大客户未来也可能变成潜在竞争者

406
00:15:37,629 --> 00:15:40,299
所以你看three pers不是没有风险

407
00:15:40,299 --> 00:15:41,899
客户集中交付压力

408
00:15:41,899 --> 00:15:44,799
供应链软件生态散热成本竞争升级

409
00:15:44,799 --> 00:15:46,370
这些问题一个都不小

410
00:15:46,370 --> 00:15:48,809
它绝对不是一上市就能取代英伟达

411
00:15:48,809 --> 00:15:50,759
也不是没有短板的完美公司

412
00:15:50,759 --> 00:15:53,700
但重点在于市场现在看的不是three birds

413
00:15:53,700 --> 00:15:55,580
明天能不能打败英伟达

414
00:15:55,580 --> 00:15:57,159
而是一个更大的信号

415
00:15:57,159 --> 00:15:59,460
AI巨头们正在认真寻找

416
00:15:59,460 --> 00:16:01,700
英伟达之外的第二套算力方案

417
00:16:01,700 --> 00:16:04,740
这才是英伟达接下来财报最敏感的地方

418
00:16:04,740 --> 00:16:07,539
因为英伟达真正的风险不是这个季度

419
00:16:07,539 --> 00:16:08,919
GPU卖不卖得动

420
00:16:08,919 --> 00:16:09,899
短期来看

421
00:16:09,899 --> 00:16:11,879
AI芯片需求依然很强

422
00:16:11,879 --> 00:16:13,539
市场真正担心的是

423
00:16:13,539 --> 00:16:17,039
英伟达最大的客户正在一边买他的GPU

424
00:16:17,039 --> 00:16:18,740
一边加速自研芯片

425
00:16:18,740 --> 00:16:21,409
甚至扶持新的替代路线

426
00:16:21,409 --> 00:16:22,570
过去两年

427
00:16:22,570 --> 00:16:23,730
微软谷歌

428
00:16:23,730 --> 00:16:27,549
亚马逊metal这些云巨头疯狂采购英伟达GPU

429
00:16:27,549 --> 00:16:30,309
把英伟达推上了AI时代的王座

430
00:16:30,309 --> 00:16:32,909
但现在谷歌在推TPU

431
00:16:32,909 --> 00:16:35,870
亚马逊在推吹联微软和OpenAI

432
00:16:35,870 --> 00:16:37,779
也在布局自己的AI芯片

433
00:16:37,779 --> 00:16:38,720
与此同时

434
00:16:38,720 --> 00:16:41,259
Three birds grock testrent

435
00:16:41,259 --> 00:16:45,568
这些新架构公司也在从推理市场寻找突破口

436
00:16:45,568 --> 00:16:46,749
换句话说

437
00:16:46,749 --> 00:16:49,249
因为答案最大的客户正在慢慢变成

438
00:16:49,249 --> 00:16:50,859
他潜在的竞争对手

439
00:16:50,859 --> 00:16:52,759
中国市场也是一个变量

440
00:16:52,759 --> 00:16:53,979
受出口限制影响

441
00:16:53,979 --> 00:16:57,720
英伟达高端AI芯片在中国的销售空间被压缩

442
00:16:57,720 --> 00:17:01,610
华为ascend正在承接更多国产替代需求

443
00:17:01,610 --> 00:17:04,710
如果中国大厂更多转向国产芯片

444
00:17:04,710 --> 00:17:07,859
英伟达在中国市场的增长空间也会受到影响

445
00:17:07,859 --> 00:17:11,140
所以英伟达财报真正要看的不只是营收和利润

446
00:17:11,140 --> 00:17:11,940
有没有超预期

447
00:17:11,940 --> 00:17:13,078
而是三个问题

448
00:17:13,078 --> 00:17:15,659
大客户的采购节奏有没有变化

449
00:17:15,659 --> 00:17:19,019
自研芯片和替代方案有没有开始影响市场预期

450
00:17:19,019 --> 00:17:22,380
英伟达还能不能继续维持现在的这种高增长

451
00:17:22,380 --> 00:17:23,960
高毛利和高估值

452
00:17:23,960 --> 00:17:26,828
这也是为什么three birds上市值得关注

453
00:17:26,828 --> 00:17:28,808
他表面上是一只AI新股

454
00:17:28,808 --> 00:17:30,548
背后却代表着整个行业

455
00:17:30,548 --> 00:17:33,490
正在寻找后英伟达时代的可能性

456
00:17:33,490 --> 00:17:34,150
好了

457
00:17:34,150 --> 00:17:36,170
今天的视频到这里就结束了

458
00:17:36,170 --> 00:17:37,309
three birds上市

459
00:17:37,309 --> 00:17:39,109
如果一开盘180美元

460
00:17:39,109 --> 00:17:40,750
你还会不会冲进去

461
00:17:40,750 --> 00:17:41,789
是快进快出

462
00:17:41,789 --> 00:17:42,630
赚一点就跑

463
00:17:42,630 --> 00:17:44,069
还是等回调再进

464
00:17:44,069 --> 00:17:45,969
还是隔山看好戏

465
00:17:45,969 --> 00:17:47,769
你又对英伟达财报怎么看

466
00:17:47,769 --> 00:17:49,249
评论区聊聊你的看法

467
00:17:49,249 --> 00:17:50,339
我们一起讨论

评论采集失败

采集完成，该视频无弹幕