1
00:00:00,040 --> 00:00:01,500
伴随AI需求的爆发
2
00:00:01,500 --> 00:00:02,080
这几年来
3
00:00:02,080 --> 00:00:04,309
科技巨头纷纷拼命加码资本支出
4
00:00:04,309 --> 00:00:05,349
甚至连微软
5
00:00:05,349 --> 00:00:05,849
谷歌
6
00:00:05,849 --> 00:00:08,210
亚马逊这样的印钞机级别的超级巨头
7
00:00:08,210 --> 00:00:09,750
都因为庞大的资本支出
8
00:00:09,750 --> 00:00:11,830
导致自由现金流所剩无几
9
00:00:11,830 --> 00:00:14,470
这些资本支出一度让市场感觉到恐慌
10
00:00:14,470 --> 00:00:15,990
觉得AI泡沫来了
11
00:00:15,990 --> 00:00:19,269
但是今天的主题我们会从租金回报率的角度看一看
12
00:00:19,269 --> 00:00:21,350
这些大厂资本支出到底合不合理
13
00:00:21,350 --> 00:00:24,920
以及英伟达GPU在AI时代到底有多值钱
14
00:00:24,920 --> 00:00:27,839
一基础计算
15
00:00:28,980 --> 00:00:33,320
目前单块B200云上租金大概是6.1美元每小时左右
16
00:00:33,320 --> 00:00:36,689
这让我们可以站在大厂投资角度算一笔账
17
00:00:36,689 --> 00:00:40,030
对于有72个B200的NVL72机架来说
18
00:00:40,030 --> 00:00:42,570
它的采购价格是260万美元
19
00:00:42,570 --> 00:00:44,450
我们按照六美元每小时计算
20
00:00:44,450 --> 00:00:48,049
72张卡一年租金也有378万美元
21
00:00:48,049 --> 00:00:49,070
考虑运维成本
22
00:00:49,070 --> 00:00:50,259
参考下图估算
23
00:00:50,259 --> 00:00:52,820
运维成本大概是机价总价值的15%
24
00:00:52,820 --> 00:00:56,320
大厂买一套NVL72机架回来几乎是一年就回本
25
00:00:56,320 --> 00:00:58,768
大厂还有比投数据中心更好的投资吗
26
00:00:58,768 --> 00:01:02,090
这比回购自家股票搞战略投资来得好多了
27
00:01:02,090 --> 00:01:05,510
所以千万不要低估这些科技巨头的效率和判断力
28
00:01:05,510 --> 00:01:07,769
GPU在AI需求的爆发式增长下
29
00:01:07,769 --> 00:01:09,289
几乎就是印钞机
30
00:01:09,289 --> 00:01:12,180
另外目前也看不到折旧和空置的风险
31
00:01:12,180 --> 00:01:14,180
因为token消耗量几何式增长
32
00:01:14,180 --> 00:01:16,420
6年前的A100和4年前的H100
33
00:01:16,420 --> 00:01:17,420
不仅是被租满
34
00:01:17,420 --> 00:01:19,439
作为老卡租金居然也在涨
35
00:01:19,439 --> 00:01:21,040
其中H100和A100
36
00:01:21,040 --> 00:01:22,939
近期云上租金都创出新高
37
00:01:22,939 --> 00:01:23,799
更重要的是
38
00:01:23,799 --> 00:01:27,430
GPU不是想要多少就有多少的受到台积电产能
39
00:01:27,430 --> 00:01:28,909
内存供应等诸多限制
40
00:01:28,909 --> 00:01:31,200
这使得供应很难短期拉上来
41
00:01:31,200 --> 00:01:36,150
所以明年B200这种高端GPU算力的租金大概率比现在还要高
42
00:01:36,150 --> 00:01:38,209
参考new cloud代表厂家NBAS
43
00:01:38,209 --> 00:01:39,269
在今年6月1日起
44
00:01:39,269 --> 00:01:41,709
将大幅提高GPU云上租金
45
00:01:41,709 --> 00:01:44,010
供不应求的趋势非常明显
46
00:01:44,010 --> 00:01:47,569
2space x招股书透露的秘密
47
00:01:48,890 --> 00:01:51,290
在实际建设和部署数据中心的过程中
48
00:01:51,290 --> 00:01:54,329
成本显然远不是我们第一部分估计的这样简单
49
00:01:54,329 --> 00:01:56,010
还要考虑到土地审批
50
00:01:56,010 --> 00:01:56,810
资金成本
51
00:01:56,810 --> 00:01:58,500
税务等等各方面的开支
52
00:01:58,500 --> 00:02:02,340
以及并非所有GPU都会七天24小时满额出租
53
00:02:02,340 --> 00:02:05,900
所以一年回本只能作为理想状态下的理论值
54
00:02:05,900 --> 00:02:08,379
实际上的云上收租回报率到底有多高
55
00:02:08,379 --> 00:02:10,979
我们可以参考SPACEX最近被曝光的招股书
56
00:02:10,979 --> 00:02:13,069
其中透露一些非常有营养的数据
57
00:02:13,069 --> 00:02:14,389
根据招股书的内容
58
00:02:14,389 --> 00:02:18,009
我们知道ANTHROPIC在锁定class1集群算力后
59
00:02:18,009 --> 00:02:21,219
又继续获得class s2中更强的BLACKWELL算力
60
00:02:21,219 --> 00:02:22,159
为了这些算力
61
00:02:22,159 --> 00:02:24,219
从现在到2029年5月
62
00:02:24,219 --> 00:02:28,120
ANTHROPIC每个月将向SPACEX支付12.5亿美元的算力租金
63
00:02:28,120 --> 00:02:29,949
一年就是150亿美元
64
00:02:29,949 --> 00:02:33,939
单单这一项收入就差不多是space x所有其他收入之和
65
00:02:33,939 --> 00:02:35,180
CLOS1数据中心
66
00:02:35,180 --> 00:02:36,240
这个合作是双赢
67
00:02:36,240 --> 00:02:38,259
对于即将IPO的SPACEX来说
68
00:02:38,259 --> 00:02:39,930
大幅改善了公司现金流
69
00:02:39,930 --> 00:02:41,250
对于ANTHROPIC来说
70
00:02:41,250 --> 00:02:43,490
则是获得了严重短缺的宝贵算力
71
00:02:43,490 --> 00:02:44,810
我们现在知道了租金
72
00:02:44,810 --> 00:02:47,409
剩下的就是算两个集群资产的成本了
73
00:02:47,409 --> 00:02:48,729
先看克劳斯一
74
00:02:48,729 --> 00:02:52,310
已知这是一个拥有22万张英伟达GPU的算力集群
75
00:02:52,310 --> 00:02:55,810
包括H100H200GB230种GPU
76
00:02:55,810 --> 00:02:57,409
H100和H200
77
00:02:57,409 --> 00:02:59,949
单卡成本大概在3.5到4万美元
78
00:02:59,949 --> 00:03:02,469
GB200单卡成本是3.6万美元
79
00:03:02,469 --> 00:03:04,949
考虑到CLOS1的建设和部署时间
80
00:03:04,949 --> 00:03:05,629
可以推测
81
00:03:05,629 --> 00:03:08,150
其中H100和H200比例比较大
82
00:03:08,150 --> 00:03:11,289
我们以4万美元为整个集群所有卡的平均成本
83
00:03:11,289 --> 00:03:15,969
则CLAUSE集群用来买英伟达GPU的成本大概是88亿美元
84
00:03:15,969 --> 00:03:16,930
除了GPU外
85
00:03:16,930 --> 00:03:19,050
数据中心基础设施建设也要成本
86
00:03:19,050 --> 00:03:20,829
这是SPACEX引以为豪的事
87
00:03:20,829 --> 00:03:22,088
他们在招股书中说
88
00:03:22,088 --> 00:03:23,848
其通过极致的供应链控制
89
00:03:23,848 --> 00:03:28,509
将数据中心的建设成本压低至惊人的每兆瓦270万美元
90
00:03:28,509 --> 00:03:31,949
远低于业界基准的每兆瓦1230万美元
91
00:03:31,949 --> 00:03:34,409
而class1的容量是300兆瓦
92
00:03:34,409 --> 00:03:37,319
所以基础设施建设是8.1亿美元
93
00:03:37,319 --> 00:03:41,900
因此CLAUS1总成本等于88亿美元的GPU采购费
94
00:03:41,900 --> 00:03:46,199
加8.1亿美元的基础设施建设费等于96.1亿美元
95
00:03:46,199 --> 00:03:47,539
再看class2
96
00:03:47,539 --> 00:03:49,620
这个集群比class1更大
97
00:03:49,620 --> 00:03:50,899
更强也更贵
98
00:03:50,899 --> 00:03:52,789
这是一个纯black well集群
99
00:03:52,789 --> 00:03:54,210
里面不是GB200
100
00:03:54,210 --> 00:03:55,389
就是GB300
101
00:03:55,389 --> 00:03:57,310
当前已上线430兆瓦
102
00:03:57,310 --> 00:04:00,269
对应22万张GB200或GB300
103
00:04:00,269 --> 00:04:03,329
下一阶段后容量将增长至830兆瓦
104
00:04:03,329 --> 00:04:04,609
约44万张
105
00:04:04,609 --> 00:04:06,439
GB200或GB300
106
00:04:06,439 --> 00:04:08,360
都按NVL72机柜计算
107
00:04:08,360 --> 00:04:10,439
GB200单价3.6万美元
108
00:04:10,439 --> 00:04:12,680
GB300单价4.2万美元
109
00:04:12,680 --> 00:04:13,919
据平均数估算
110
00:04:13,919 --> 00:04:17,759
class2中的GPU平均每个成本是3.9万美元
111
00:04:17,759 --> 00:04:20,620
44万个的成本就是171亿美元
112
00:04:20,620 --> 00:04:24,319
同样考虑每兆瓦270万美元的基础设施建设成本
113
00:04:24,319 --> 00:04:28,500
close2对应830兆瓦的成本是22.4亿美元
114
00:04:28,500 --> 00:04:33,658
因此clothes2总成本等于171亿美元的GPU采购费
115
00:04:33,658 --> 00:04:36,538
加22.4亿美元的基础设施建设费
116
00:04:36,538 --> 00:04:38,819
等于193.4亿美元
117
00:04:38,819 --> 00:04:41,860
我们把class1和CLAUS2的成本加起来
118
00:04:41,860 --> 00:04:43,560
即96.1亿美元
119
00:04:43,560 --> 00:04:45,259
加193.4亿美元
120
00:04:45,259 --> 00:04:47,399
等于289.5亿美元
121
00:04:47,399 --> 00:04:51,139
现在我们再来回头看看ANTHROPIC付给space x的租金
122
00:04:51,139 --> 00:04:52,620
一年150亿美元
123
00:04:52,620 --> 00:04:55,199
就明白这个出租算力的生意有多香了
124
00:04:55,199 --> 00:04:57,639
我们上文分享了一个大体的数据中心
125
00:04:57,639 --> 00:05:01,019
年度运维成本大概是GPU机架成本的15%
126
00:05:01,019 --> 00:05:03,480
两个集群大概是35.5亿美元
127
00:05:03,480 --> 00:05:05,639
所以回报率的计算就很简单了
128
00:05:05,639 --> 00:05:08,600
租金减去运维成本除以总资产投入
129
00:05:08,600 --> 00:05:11,079
也就是150-35.5
130
00:05:11,079 --> 00:05:12,439
除以289.5
131
00:05:12,439 --> 00:05:14,180
等于39.6%
132
00:05:14,180 --> 00:05:18,360
这样一个具体的数字足够让我们感受到英伟达GPU到底有多值钱
133
00:05:18,360 --> 00:05:20,310
到底是一个怎样的优质资产
134
00:05:20,310 --> 00:05:21,649
还需要特别注意的是
135
00:05:21,649 --> 00:05:23,029
39.6%
136
00:05:23,029 --> 00:05:25,819
这个回报率大概率是被显著低估的
137
00:05:25,819 --> 00:05:27,339
因为目前公开信息只是说
138
00:05:27,339 --> 00:05:30,620
ANTHROPIC和space x的协议已从独占colossus1
139
00:05:30,620 --> 00:05:32,579
扩展到包含colossus2的算力
140
00:05:32,579 --> 00:05:35,990
没说ANTHROPIC已经把colossus2的全部容量独占
141
00:05:35,990 --> 00:05:42,288
所以每年150亿美元的算力租金对应的只是全部class1加部分class2
142
00:05:42,288 --> 00:05:45,329
我们按两个集群的全量投资计算出来的租金回报率
143
00:05:45,329 --> 00:05:47,918
实际上还是会显著低于实际的回报率
144
00:05:47,918 --> 00:05:51,519
只要CLAUS2里还有一小部分算力尚未租给ANTHOPIC
145
00:05:51,519 --> 00:05:54,119
实际租金回报率可能轻松超越5%
146
00:05:54,119 --> 00:05:54,439
十
147
00:05:54,439 --> 00:05:56,538
两年内回本的出租生意
148
00:05:56,538 --> 00:05:58,499
三真实需求
149
00:05:58,499 --> 00:06:03,108
一各家科技巨头的apex和云算力出租的关系
150
00:06:03,108 --> 00:06:05,809
就像是建房子和出租房子的关系
151
00:06:05,809 --> 00:06:07,908
如果你建了一大堆房子却租不出去
152
00:06:07,908 --> 00:06:09,470
这个时候就要警惕泡沫了
153
00:06:09,470 --> 00:06:11,569
但如果你建一栋就租满一栋
154
00:06:11,569 --> 00:06:13,230
租金高到你两年回本
155
00:06:13,230 --> 00:06:14,910
而且之前几年建的老房子
156
00:06:14,910 --> 00:06:16,550
现在租金还在不断上涨
157
00:06:16,550 --> 00:06:19,629
这显然是一个供不应求的正在爆发的市场
158
00:06:19,629 --> 00:06:21,970
如果说建房或买房可能涉及到融资
159
00:06:21,970 --> 00:06:24,019
那么租房行为则是纯消费
160
00:06:24,019 --> 00:06:25,139
在上文的案例中
161
00:06:25,139 --> 00:06:28,420
ANTHROPIC作为租客不可能闲的没事给对方送钱
162
00:06:28,420 --> 00:06:31,889
他们愿意拿一年150亿美元的租金给到SPACEX
163
00:06:31,889 --> 00:06:36,209
就说明ANTHROPIC能用这个算力换来远超150亿美元的回报
164
00:06:36,209 --> 00:06:39,410
未来的AI算力一定会像现在的电力或石油一样
165
00:06:39,410 --> 00:06:42,019
成为几乎每一个人生活和工作中的刚需
166
00:06:42,019 --> 00:06:44,379
我们每个月也会多一份token账单
167
00:06:44,379 --> 00:06:47,319
本期主题内容来自过去一段时间星球的部分主题
168
00:06:47,319 --> 00:06:49,899
另外最近投资户限制趋于严格
169
00:06:49,899 --> 00:06:52,500
关于如何保护存量账户和未来窗口
170
00:06:52,500 --> 00:06:54,120
除了上上篇星球主题外
171
00:06:54,120 --> 00:06:57,560
我还有更多总结发布在星球这个半私域空间
172
00:06:57,560 --> 00:07:01,750
感兴趣的朋友也欢迎扫描公号文末二维码加入星球
173
00:07:01,750 --> 00:07:04,110
最近还有不少网友问我的小号是什么
174
00:07:04,110 --> 00:07:05,649
这里备注一下备用工号
175
00:07:05,649 --> 00:07:09,829
也谢谢感兴趣的朋友多用一个关注名额关注我的备用号
176
00:07:10,569 --> 00:07:12,110
如果你喜欢本期视频
177
00:07:12,110 --> 00:07:13,529
希望可以点赞关注
178
00:07:13,529 --> 00:07:14,269
支持一波
179
00:07:14,269 --> 00:07:16,560
这将对我有很大的帮助