1
00:00:03,160 --> 00:00:03,560
Speaker 1: Media.

2
00:00:04,720 --> 00:00:06,960
Speaker 2: Hello and welcome to Better Offline. I'm your host ed

3
00:00:07,000 --> 00:00:21,239
Speaker 2: Sir Tron. Now, before we go any further, I hate

4
00:00:21,280 --> 00:00:23,080
Speaker 2: to ask you to do this, but I need your help.

5
00:00:23,440 --> 00:00:25,520
Speaker 2: I'm up for this year's Webbe's for the Best Business

6
00:00:25,520 --> 00:00:27,760
Speaker 2: Podcast Award. I know it's a pain in the ars

7
00:00:27,760 --> 00:00:29,600
Speaker 2: to register, but can you sign up and vote for

8
00:00:29,640 --> 00:00:31,920
Speaker 2: Better Offline? Never won an award in my life, and

9
00:00:31,920 --> 00:00:33,480
Speaker 2: I have enough listeners that I think we can tip

10
00:00:33,479 --> 00:00:37,000
Speaker 2: the scales. But on with the show. The following two

11
00:00:37,000 --> 00:00:39,760
Speaker 2: part episode is the culmination of months of research, presenting

12
00:00:39,760 --> 00:00:41,920
Speaker 2: a case I'd been making in parts since July of

13
00:00:42,000 --> 00:00:45,559
Speaker 2: last year. Open AYE is a financial abomination, a thing

14
00:00:45,600 --> 00:00:48,120
Speaker 2: that should not be, an aberration, a symbol of rot

15
00:00:48,280 --> 00:00:51,239
Speaker 2: at the heart of Silicon Valley, a company that unrepentantly

16
00:00:51,280 --> 00:00:53,880
Speaker 2: and needlessly burns billions of dollars with no end in sight,

17
00:00:54,240 --> 00:00:57,480
Speaker 2: helmed by a creticist, dis ingenuous billionaire who continually lies

18
00:00:57,720 --> 00:00:59,760
Speaker 2: about what it is that it will do because he,

19
00:01:00,280 --> 00:01:02,920
Speaker 2: like I've been saying since last year, knows that Generative

20
00:01:02,920 --> 00:01:05,640
Speaker 2: AI can't do much more than it does today. What

21
00:01:05,720 --> 00:01:07,320
Speaker 2: I'm going to lay out for you is my case

22
00:01:07,360 --> 00:01:11,040
Speaker 2: the Open AI can't survive, that it's boadline impossible under

23
00:01:11,240 --> 00:01:14,120
Speaker 2: any of the current terms for this company to continue,

24
00:01:14,160 --> 00:01:16,080
Speaker 2: and that its demise will bring about the death of

25
00:01:16,120 --> 00:01:19,400
Speaker 2: a or significant harm of several other firms. A cottage

26
00:01:19,440 --> 00:01:23,240
Speaker 2: industry of despicable, billion dollar burn rate capitalistic monstrosities has

27
00:01:23,240 --> 00:01:26,280
Speaker 2: sprung up around this stupid fucking company in the hopes

28
00:01:26,280 --> 00:01:28,639
Speaker 2: of further inflating a bubble set to burst at any moment.

29
00:01:29,360 --> 00:01:31,880
Speaker 2: My discuss for the parties involved is unrelenting. This is

30
00:01:32,120 --> 00:01:34,200
Speaker 2: my discuss for a media industry that failed to even

31
00:01:34,200 --> 00:01:36,959
Speaker 2: attempt to tackle the subject matter I'm going to detail.

32
00:01:37,280 --> 00:01:39,080
Speaker 2: I believe at the end of these episodes you'll see

33
00:01:39,080 --> 00:01:41,600
Speaker 2: my point and at the very least agree that open

34
00:01:41,600 --> 00:01:45,440
Speaker 2: AI's current situation is totally untenable. If I'm right, open

35
00:01:45,480 --> 00:01:47,640
Speaker 2: ai will go down in history as an abdication of

36
00:01:47,720 --> 00:01:51,920
Speaker 2: due diligence, fiscal responsibility, in frankly common sense, both in

37
00:01:51,960 --> 00:01:54,280
Speaker 2: the venture capitalists and entities that propped it up and

38
00:01:54,320 --> 00:01:57,160
Speaker 2: a tech media that was more concerned with taking detailed

39
00:01:57,200 --> 00:01:59,760
Speaker 2: notes on my comings and goings than knowing their ass

40
00:01:59,760 --> 00:02:02,400
Speaker 2: from their ear hole. Those who have failed to hold

41
00:02:02,400 --> 00:02:05,200
Speaker 2: men like Sam Altman and darra Ama Day Accountable have

42
00:02:05,320 --> 00:02:09,639
Speaker 2: tacitly approved a financial and environmentally destructive movement that will

43
00:02:09,720 --> 00:02:12,840
Speaker 2: lead to the very little actually happening or changing in

44
00:02:12,840 --> 00:02:15,440
Speaker 2: the world other than damaging our power grid and the

45
00:02:15,440 --> 00:02:19,040
Speaker 2: theft of art from millions of people. These episodes will

46
00:02:19,080 --> 00:02:21,920
Speaker 2: be full of numbers and statements and very few declarations

47
00:02:22,000 --> 00:02:24,360
Speaker 2: or personal opinions, though probably a few swear words. If

48
00:02:24,400 --> 00:02:26,959
Speaker 2: I'm honest, I don't need to get that personal here.

49
00:02:27,200 --> 00:02:30,880
Speaker 2: The numbers in question, they're damaging, they're staggering, they're worrying,

50
00:02:30,880 --> 00:02:34,119
Speaker 2: and ultimately spell collapse. The truth is, we don't even

51
00:02:34,120 --> 00:02:36,960
Speaker 2: need to talk about tarifs for things to go sideways

52
00:02:36,960 --> 00:02:39,519
Speaker 2: for this industry. The price of a GPU could rise

53
00:02:39,520 --> 00:02:42,160
Speaker 2: one hundred percent, or it could have it really wouldn't

54
00:02:42,160 --> 00:02:44,799
Speaker 2: make much of a difference to open AI's chances of survival.

55
00:02:45,200 --> 00:02:48,400
Speaker 2: That's how bad the fundamentals are. And to illustrate that point,

56
00:02:48,400 --> 00:02:51,120
Speaker 2: I'm going to ask a number of relatively simple questions

57
00:02:51,120 --> 00:02:53,680
Speaker 2: over the next couple of episodes and make an attempt

58
00:02:53,720 --> 00:02:57,560
Speaker 2: to answer them. First, let's start with something simple, how

59
00:02:57,639 --> 00:03:02,080
Speaker 2: much cash does open ai have. At the start of April,

60
00:03:02,120 --> 00:03:05,280
Speaker 2: open Ai closed what was called the largest private tech

61
00:03:05,320 --> 00:03:09,720
Speaker 2: funding round in history, where it raised and astonishing forty

62
00:03:09,760 --> 00:03:12,200
Speaker 2: billion dollars. And the reason I'm saying this with the

63
00:03:12,240 --> 00:03:15,000
Speaker 2: sarcastic inflection is that open ai has only actually raised

64
00:03:15,040 --> 00:03:16,960
Speaker 2: ten billion dollars of the forty billion dollars, with the

65
00:03:16,960 --> 00:03:19,920
Speaker 2: rest arriving by the end of the year, and even

66
00:03:19,960 --> 00:03:22,000
Speaker 2: as I record this, I don't know if the money's

67
00:03:22,000 --> 00:03:25,880
Speaker 2: actually gone yet. I'll get it into that in a minute. Now.

68
00:03:25,919 --> 00:03:27,799
Speaker 2: A lot can happen in a year, and the remaining

69
00:03:28,200 --> 00:03:31,120
Speaker 2: thirty billion dollars twenty billion dollars of which will allegedly

70
00:03:31,160 --> 00:03:33,760
Speaker 2: be provided by soft Bank, is partially contingent on an

71
00:03:33,760 --> 00:03:36,640
Speaker 2: open AI's conversion from a nonprofit to a for profit

72
00:03:36,720 --> 00:03:38,920
Speaker 2: by the end of twenty twenty five, and if it fails,

73
00:03:38,960 --> 00:03:41,200
Speaker 2: soft Bank will only give open Ai a further twenty

74
00:03:41,200 --> 00:03:44,600
Speaker 2: billion dollars. I'll get into how fucking stupid this gets later.

75
00:03:45,480 --> 00:03:49,400
Speaker 2: The round also valued open Ai and astonishing three hundred

76
00:03:49,440 --> 00:03:51,960
Speaker 2: billion dollars. To put that in context, open Ai had

77
00:03:52,000 --> 00:03:54,560
Speaker 2: revenues of four billion dollars in twenty twenty four. This

78
00:03:54,640 --> 00:03:57,760
Speaker 2: deal values open Ai at seventy five times its revenue.

79
00:03:57,960 --> 00:04:00,560
Speaker 2: That's more than Tesla even at its most looted Chris Peak.

80
00:04:01,480 --> 00:04:04,120
Speaker 2: I also want to add that, as of writing this sentence,

81
00:04:04,160 --> 00:04:07,040
Speaker 2: this money is yet to arrive. Maybe it will arrive

82
00:04:07,080 --> 00:04:09,440
Speaker 2: by the time this is out. Maybe what really make

83
00:04:09,480 --> 00:04:12,560
Speaker 2: me look stupid. But based on soft Bank's filings, which

84
00:04:12,560 --> 00:04:15,040
Speaker 2: I'll link to in the spreadsheet for this episode, say

85
00:04:15,040 --> 00:04:17,679
Speaker 2: that the money will arrive mid April. It's April fourteenth

86
00:04:17,720 --> 00:04:20,000
Speaker 2: as I'm recording this, and that soft Bank would be

87
00:04:20,080 --> 00:04:23,920
Speaker 2: borrowing as much as ten billion dollars of the financing

88
00:04:23,920 --> 00:04:26,200
Speaker 2: for the round, with the option to syndicate, meaning you

89
00:04:26,200 --> 00:04:29,039
Speaker 2: bring in other investors the rest of them. For the

90
00:04:29,040 --> 00:04:31,359
Speaker 2: sake of argument, I'm going to assume the money actually arrives,

91
00:04:31,360 --> 00:04:36,080
Speaker 2: though filings also suggest that, and I'm quoting in certain circumstances,

92
00:04:36,080 --> 00:04:38,600
Speaker 2: the second thirty billion dollar tranch could arrive in early

93
00:04:38,640 --> 00:04:43,200
Speaker 2: twenty twenty six. This isn't great. It also seems that

94
00:04:43,240 --> 00:04:45,839
Speaker 2: soft banks ten billion dollar commitment is contingent on getting

95
00:04:45,839 --> 00:04:49,719
Speaker 2: a loan, as it says it's financed through borrowings from

96
00:04:50,160 --> 00:04:54,920
Speaker 2: Mizuho Bank Limited, among other financial institutions. Soft Banks had

97
00:04:54,920 --> 00:04:56,600
Speaker 2: plenty of loans in the past, so I think they're

98
00:04:56,640 --> 00:04:58,600
Speaker 2: going to get it. But I think this is one

99
00:04:58,640 --> 00:05:01,320
Speaker 2: of their biggest open a I also revealed it now

100
00:05:01,320 --> 00:05:04,799
Speaker 2: has twenty million paying subscribers and over five hundred million

101
00:05:04,800 --> 00:05:07,599
Speaker 2: weekly active users. If you're wondering why it doesn't talking

102
00:05:07,640 --> 00:05:10,440
Speaker 2: about monthly users, it's because they're likely much higher than

103
00:05:10,440 --> 00:05:13,600
Speaker 2: five hundred million, which would reveal exactly how poorly open

104
00:05:13,600 --> 00:05:17,120
Speaker 2: ai converts free chat GPT users to paying ones. The

105
00:05:17,200 --> 00:05:20,120
Speaker 2: Information reported back in January that open ai was generating

106
00:05:20,120 --> 00:05:22,560
Speaker 2: twenty five million dollars in revenue a month from its

107
00:05:22,560 --> 00:05:25,359
Speaker 2: two hundred dollars a month pro subscribers, and just so

108
00:05:25,400 --> 00:05:27,880
Speaker 2: we're clear, they lose money on every one of those two,

109
00:05:28,040 --> 00:05:30,640
Speaker 2: suggesting that they have around one hundred and twenty five

110
00:05:30,680 --> 00:05:34,560
Speaker 2: thousand Chat GPT pro subscribers, each losing the money somehow.

111
00:05:34,960 --> 00:05:38,080
Speaker 2: Assuming the other nineteen million, eight hundred and seventy five

112
00:05:38,160 --> 00:05:41,000
Speaker 2: thousand users are paying twenty bucks a month, that puts

113
00:05:41,040 --> 00:05:43,280
Speaker 2: revenue about four hundred and twenty three million dollars a month,

114
00:05:43,400 --> 00:05:45,080
Speaker 2: or at least five billion dollars a year from chat

115
00:05:45,120 --> 00:05:48,640
Speaker 2: GPT subscriptions. This is what reporters mean when they say

116
00:05:48,760 --> 00:05:51,640
Speaker 2: annualized revenue. By the way, it's literally the month. The

117
00:05:51,680 --> 00:05:53,680
Speaker 2: monthly revenue that the money they're making in one month

118
00:05:53,720 --> 00:05:55,960
Speaker 2: multiplied by twelve, and you'll be surprised to hear that

119
00:05:56,000 --> 00:05:59,240
Speaker 2: people play silly buggers with that all the time. In March,

120
00:05:59,240 --> 00:06:02,120
Speaker 2: Bloomberg reports the open ai expects its revenue to triple

121
00:06:02,160 --> 00:06:04,719
Speaker 2: to twelve point seven billion dollars in twenty twenty five,

122
00:06:04,920 --> 00:06:07,440
Speaker 2: assuming a similar split of revenue to twenty twenty four.

123
00:06:07,560 --> 00:06:10,320
Speaker 2: This would require open ai to nearly double its annualized

124
00:06:10,320 --> 00:06:13,360
Speaker 2: subscription revenue from Q one twenty twenty five from five

125
00:06:13,360 --> 00:06:15,640
Speaker 2: billion dollars right now to around nine point twenty seven

126
00:06:15,680 --> 00:06:19,600
Speaker 2: billion dollars, and nearly kuldruple API revenue from twenty twenty

127
00:06:19,640 --> 00:06:22,559
Speaker 2: four's revenue of one billion, which includes Microsoft's twenty percent

128
00:06:22,600 --> 00:06:25,520
Speaker 2: payment for access to open AI's models, and that would

129
00:06:25,560 --> 00:06:27,400
Speaker 2: get them about three point four to three billion. We

130
00:06:27,440 --> 00:06:29,800
Speaker 2: all need to super worry about these numbers. And I

131
00:06:29,880 --> 00:06:33,359
Speaker 2: realized these are messy numbers. It's just unclear how open

132
00:06:33,400 --> 00:06:35,240
Speaker 2: ai intends to pull any of this off. It's an

133
00:06:35,240 --> 00:06:39,320
Speaker 2: incredible leap, and open AI's own plans don't exactly inspire confidence.

134
00:06:39,680 --> 00:06:43,120
Speaker 2: They're really good at getting free subscribers, they don't seem

135
00:06:43,160 --> 00:06:45,359
Speaker 2: to be able to get paying ones in quite the

136
00:06:45,400 --> 00:06:48,680
Speaker 2: same number, and also they even lose them money. Every

137
00:06:48,760 --> 00:06:50,560
Speaker 2: time I think about this company, I start feeling a

138
00:06:50,600 --> 00:06:53,440
Speaker 2: little crazy, and if I'm on it anyway, The Information

139
00:06:53,520 --> 00:06:56,039
Speaker 2: reported in February that if the open ai plan to

140
00:06:56,040 --> 00:06:58,360
Speaker 2: grow its revenue by making three billion dollars a year

141
00:06:58,400 --> 00:07:01,520
Speaker 2: selling agents with chat GBT subscriptions seven point nine billion

142
00:07:01,520 --> 00:07:03,720
Speaker 2: dollars and API cous one point eight billion dollars making

143
00:07:03,800 --> 00:07:07,600
Speaker 2: up the rest. This is, of course, what's the technical term, Oh,

144
00:07:07,640 --> 00:07:11,360
Speaker 2: it's bollocks. It's complete fucking bollocks. I'm sorry. Agents, by

145
00:07:11,400 --> 00:07:14,000
Speaker 2: the way, are AI chatbots that can do something like

146
00:07:14,080 --> 00:07:17,160
Speaker 2: interact with another program on the user's behalf. Open aiy's

147
00:07:17,200 --> 00:07:20,120
Speaker 2: agents can't even do the simplest tasks. And three billion

148
00:07:20,160 --> 00:07:22,440
Speaker 2: dollars of twelve point seven billion dollar figure appears to

149
00:07:22,480 --> 00:07:24,960
Speaker 2: be a commitment made by soft Bank to purchase three

150
00:07:25,000 --> 00:07:28,160
Speaker 2: billion dollars a year of Open Aiyes, tech, now, let's

151
00:07:28,480 --> 00:07:32,080
Speaker 2: pass out these numbers precisely. So incoming monthly revenue roughly

152
00:07:32,080 --> 00:07:34,280
Speaker 2: four hundred and twenty five million dollars give or take

153
00:07:34,800 --> 00:07:37,760
Speaker 2: theoretical revenue from soft Banks completely made up thing two

154
00:07:37,800 --> 00:07:40,440
Speaker 2: hundred and fifty million dollars a month. However, I really

155
00:07:40,440 --> 00:07:43,000
Speaker 2: can find no proof that soft Bank has begun to

156
00:07:43,000 --> 00:07:45,520
Speaker 2: make these payments, or indeed that it intends to make

157
00:07:45,560 --> 00:07:49,200
Speaker 2: them even how it intends to. Now, let's talk liquidity.

158
00:07:49,920 --> 00:07:52,320
Speaker 2: Open Ai has ten billion dollars that they yet to

159
00:07:52,360 --> 00:07:54,720
Speaker 2: receive a recording this, but let's assume they get it,

160
00:07:54,880 --> 00:07:56,840
Speaker 2: and it will be ten billion dollars seven point five

161
00:07:56,880 --> 00:07:59,760
Speaker 2: billion from soft Bank and a syndicate of investors including

162
00:07:59,760 --> 00:08:02,760
Speaker 2: my mic Krossoft and Cutchiu and a few others. Potentially.

163
00:08:03,520 --> 00:08:06,000
Speaker 2: There's also an indeterminate amount of remaining capital on the

164
00:08:06,000 --> 00:08:08,840
Speaker 2: four billion dollar credit facility provided by multiple banks back

165
00:08:08,880 --> 00:08:11,240
Speaker 2: in October twenty twenty four, and that was raised alongside

166
00:08:11,240 --> 00:08:13,000
Speaker 2: of funding ground that valued the company one hundred and

167
00:08:13,040 --> 00:08:16,480
Speaker 2: fifty seven billion dollars now. As a note, the October

168
00:08:16,480 --> 00:08:19,800
Speaker 2: announcement stated that open ai had access to over ten

169
00:08:19,840 --> 00:08:22,360
Speaker 2: billion dollars in liquidity, giving us a sense of how

170
00:08:22,360 --> 00:08:24,320
Speaker 2: fast it's burning through the cash it has on hand,

171
00:08:24,320 --> 00:08:26,560
Speaker 2: as that was pretty much all the money they'd raised

172
00:08:26,560 --> 00:08:29,400
Speaker 2: at the time. Based on reports, open Ai will not

173
00:08:29,440 --> 00:08:31,600
Speaker 2: have access to the rest of the forty billion dollars

174
00:08:32,000 --> 00:08:34,120
Speaker 2: that soft Bank is funding them with. Until the end

175
00:08:34,120 --> 00:08:35,880
Speaker 2: of the year. And it's unclear what part of the

176
00:08:35,960 --> 00:08:38,760
Speaker 2: end of the year that is, but SoftBank's filing says December.

177
00:08:39,080 --> 00:08:41,520
Speaker 2: Will it be December first? Will it be Christmas time?

178
00:08:41,800 --> 00:08:43,960
Speaker 2: Will Sam Mortman look on to the tree and Messiyoshi

179
00:08:43,960 --> 00:08:46,560
Speaker 2: san will be there? We'll find out, won't we. But

180
00:08:46,640 --> 00:08:49,120
Speaker 2: we can assume in this case that open ai likely has,

181
00:08:49,120 --> 00:08:52,240
Speaker 2: in the best case scenario, access to about roughly sixteen

182
00:08:52,240 --> 00:08:55,319
Speaker 2: billion dollars in liquidity at any given time. It's reasonable

183
00:08:55,400 --> 00:08:57,680
Speaker 2: to believe that open ai will raise more debt this

184
00:08:57,800 --> 00:09:01,040
Speaker 2: year despite this massive raise, and it does so to

185
00:09:01,080 --> 00:09:03,800
Speaker 2: the tune of around five or six billion dollars. Without it,

186
00:09:04,000 --> 00:09:06,880
Speaker 2: I'm genuinely not sure what they're going to do. And

187
00:09:07,040 --> 00:09:10,040
Speaker 2: as a reminder, kids, open ai loses money on every

188
00:09:10,080 --> 00:09:24,400
Speaker 2: single user, free or paying. Now here's my next question,

189
00:09:24,880 --> 00:09:28,600
Speaker 2: what are open AI's obligations? Now? When I put out

190
00:09:28,600 --> 00:09:30,839
Speaker 2: how does open ai survive and open aiy is a

191
00:09:30,880 --> 00:09:33,480
Speaker 2: bad business? I used reported information to explain how this

192
00:09:33,559 --> 00:09:36,960
Speaker 2: company was at its court unsustainable. I'll link to the

193
00:09:37,000 --> 00:09:39,240
Speaker 2: newsletters and the spreadsheet in the podcast too. But let's

194
00:09:39,240 --> 00:09:43,240
Speaker 2: refresh our memory, shall we. Okay, compute costs at least

195
00:09:43,240 --> 00:09:46,280
Speaker 2: thirteen billion dollars of which are going to Microsoft a

196
00:09:46,280 --> 00:09:48,400
Speaker 2: loan in twenty twenty five, and as much as five

197
00:09:48,559 --> 00:09:51,320
Speaker 2: hundred and ninety four million dollars to my favorite company,

198
00:09:51,400 --> 00:09:55,360
Speaker 2: core Wave. Core Weave's back. Every time I think I

199
00:09:55,440 --> 00:09:58,040
Speaker 2: got away from core Weave, they pop out like a

200
00:09:58,120 --> 00:10:01,080
Speaker 2: sour turd coming back in the toilet. Anyway, it seems

201
00:10:01,080 --> 00:10:03,559
Speaker 2: from even a cursory glance that open AI's costs are

202
00:10:03,559 --> 00:10:07,040
Speaker 2: increasing dramatically. The information reported earlier in the year the

203
00:10:07,040 --> 00:10:10,280
Speaker 2: open ai projects to spend thirteen billion dollars on compute

204
00:10:10,320 --> 00:10:13,199
Speaker 2: with Microsoft the loan in twenty twenty five, nearly tripling

205
00:10:13,280 --> 00:10:15,400
Speaker 2: what it spent in total compute in twenty twenty four,

206
00:10:15,400 --> 00:10:17,960
Speaker 2: which was five billion dollars, with three billion for training

207
00:10:18,000 --> 00:10:20,719
Speaker 2: and two billion for running their models. This suggests that

208
00:10:20,760 --> 00:10:24,160
Speaker 2: open aiyes costs as skyrocketing, and that was before the

209
00:10:24,240 --> 00:10:26,880
Speaker 2: launch of their new image generator, which led to multiple

210
00:10:26,920 --> 00:10:29,520
Speaker 2: complaints from Sam Mormon about a lack of available GPUs,

211
00:10:29,800 --> 00:10:33,360
Speaker 2: leading to open Aiy's Terrible Little Man CEO saying that

212
00:10:33,400 --> 00:10:37,720
Speaker 2: he expected stuff to break and delays in new products. Nevertheless,

213
00:10:37,760 --> 00:10:39,480
Speaker 2: even if we assume open ai in factored in the

214
00:10:39,480 --> 00:10:42,400
Speaker 2: compute increases into its projections, it still expects to pay

215
00:10:42,440 --> 00:10:46,280
Speaker 2: that thirteen billion dollars to Microsoft a loan. This number, however,

216
00:10:46,280 --> 00:10:49,360
Speaker 2: doesn't include the twelve point nine billion dollar five year

217
00:10:49,400 --> 00:10:51,760
Speaker 2: long compute deal signed with core Weave, a deal that

218
00:10:51,800 --> 00:10:54,360
Speaker 2: was a result of Microsoft declining to pick up the

219
00:10:54,400 --> 00:10:57,760
Speaker 2: option to buy the compute themselves for open Ai. As

220
00:10:57,800 --> 00:11:00,840
Speaker 2: an aside from what I am here and this is

221
00:11:01,080 --> 00:11:04,840
Speaker 2: sources telling me most of the Microsoft compute was open Ai.

222
00:11:04,880 --> 00:11:08,240
Speaker 2: It's basically they put an open Ai sticker over Microsoft.

223
00:11:08,320 --> 00:11:11,120
Speaker 2: It's basically the same thing. Now, payments for this deal

224
00:11:11,480 --> 00:11:14,320
Speaker 2: with with open Ai and core Weave they start in

225
00:11:14,320 --> 00:11:18,600
Speaker 2: October twenty twenty five according to the information, and assuming

226
00:11:18,640 --> 00:11:22,000
Speaker 2: that it's even paid, or like evenly paid or even

227
00:11:22,040 --> 00:11:24,680
Speaker 2: like anyone gets the cash. But these contracts are weird,

228
00:11:24,720 --> 00:11:27,760
Speaker 2: you never know exactly how they're paid, this would still

229
00:11:27,880 --> 00:11:31,600
Speaker 2: amount to roughly two point three eight billion dollars a year. However,

230
00:11:31,600 --> 00:11:33,960
Speaker 2: for the sake of argument, let's consider the payments are

231
00:11:33,960 --> 00:11:36,199
Speaker 2: around one hundred and ninety eight million dollars a month.

232
00:11:36,559 --> 00:11:39,800
Speaker 2: Though there are scenarios such as say core weaves build

233
00:11:39,840 --> 00:11:42,200
Speaker 2: out partner not being able to build the data centers

234
00:11:42,240 --> 00:11:43,800
Speaker 2: or core weave not having the money to pay to

235
00:11:43,800 --> 00:11:46,400
Speaker 2: build them, where open ai might end up paying less.

236
00:11:46,440 --> 00:11:49,800
Speaker 2: And don't worry, I'll get to that later. Let's talk

237
00:11:49,800 --> 00:11:54,240
Speaker 2: about Stargate. You heard this, You've seen this. Stargate is

238
00:11:54,280 --> 00:11:57,040
Speaker 2: this data center project the open ai is going into.

239
00:11:57,120 --> 00:11:59,720
Speaker 2: And they've dedicated somewhere in the region of nineteen billion

240
00:11:59,720 --> 00:12:02,880
Speaker 2: dollars to this project, along with another nineteen billion dollars

241
00:12:02,920 --> 00:12:07,800
Speaker 2: provided by Software Bank at some indeterminate time. And there

242
00:12:07,800 --> 00:12:10,400
Speaker 2: are other partners too, like oor Coins. Not even obvious

243
00:12:10,440 --> 00:12:12,920
Speaker 2: what they're putting in either. One thing I can tell you, though,

244
00:12:13,000 --> 00:12:16,000
Speaker 2: is the Trump administration is not actually spending any money

245
00:12:16,000 --> 00:12:18,600
Speaker 2: on this anyone. I've had tons of people emailing me

246
00:12:18,679 --> 00:12:21,480
Speaker 2: being like, hey, yeah, Donald Trump was there. Yeah, Donald

247
00:12:21,520 --> 00:12:25,000
Speaker 2: Trump was there. Doesn't mean shit anyway. Based on reporting

248
00:12:25,000 --> 00:12:28,240
Speaker 2: from Bloomberg, open Ai plans to have sixty four thousand

249
00:12:28,240 --> 00:12:30,760
Speaker 2: Blackwell GPUs running by the end of twenty twenty six,

250
00:12:31,200 --> 00:12:33,720
Speaker 2: or roughly about three point eighty four billion dollars worth

251
00:12:33,760 --> 00:12:36,760
Speaker 2: of them. I should also note that Bloomberg said that

252
00:12:36,800 --> 00:12:39,360
Speaker 2: sixteen thousand of these chips would be operational by summer

253
00:12:39,360 --> 00:12:42,120
Speaker 2: twenty twenty five, though it's unclear if that will actually happen,

254
00:12:42,160 --> 00:12:44,559
Speaker 2: and as I'll get to later, I got some more

255
00:12:44,640 --> 00:12:48,400
Speaker 2: questions about what exactly is happening at Stargate or told.

256
00:12:48,600 --> 00:12:51,680
Speaker 2: Though it's unclear who actually pays what parts of Stargate,

257
00:12:51,679 --> 00:12:53,560
Speaker 2: it's safe to assume that open Ai will have to

258
00:12:53,600 --> 00:12:55,760
Speaker 2: at least put a billion dollars into a project that

259
00:12:55,840 --> 00:12:57,240
Speaker 2: is meant to be up and running by the end

260
00:12:57,280 --> 00:13:01,640
Speaker 2: of twenty twenty six, if not more. Now Stargate really

261
00:13:01,679 --> 00:13:04,640
Speaker 2: on has one like data center project under development is

262
00:13:04,679 --> 00:13:08,160
Speaker 2: in Abilene, Texas, and as I've mentioned, it's not really

263
00:13:08,160 --> 00:13:10,080
Speaker 2: clear how it's going though. Are recent piece from the

264
00:13:10,080 --> 00:13:13,720
Speaker 2: Information reported that it was currently empty and incomplete, and

265
00:13:13,720 --> 00:13:16,160
Speaker 2: if it stays that way, and I quote again, open

266
00:13:16,200 --> 00:13:18,080
Speaker 2: Ai could walk away from the deal, which would cost

267
00:13:18,120 --> 00:13:21,600
Speaker 2: Oracle billions of dollars. Though the article takes great pains

268
00:13:21,600 --> 00:13:23,840
Speaker 2: to assure the reader that that won't be likely. Even

269
00:13:23,880 --> 00:13:26,920
Speaker 2: an inkling of such a possibility is a bad fucking sign.

270
00:13:27,440 --> 00:13:30,120
Speaker 2: Business insiders reported on the site in Abilene calls it

271
00:13:30,160 --> 00:13:33,280
Speaker 2: a three point four billion dollar data center development, as

272
00:13:33,320 --> 00:13:35,520
Speaker 2: did the press release from the developer Crusoe, who will

273
00:13:35,520 --> 00:13:38,720
Speaker 2: get too later. Though these numbers don't include GPUs hardware

274
00:13:38,800 --> 00:13:41,880
Speaker 2: or the labour necessary to run them. Right now, Cruso is,

275
00:13:41,880 --> 00:13:44,719
Speaker 2: according to Business Insider, building six new data centers, each

276
00:13:44,760 --> 00:13:47,520
Speaker 2: with a minimum square footage, which will join the two

277
00:13:47,600 --> 00:13:50,720
Speaker 2: it is already constructing for Oracle. Oracle is signed, according

278
00:13:50,760 --> 00:13:53,000
Speaker 2: to the information, a fifteen year long lease with Cruiser

279
00:13:53,000 --> 00:13:55,440
Speaker 2: for its data centers, all of which will be rented

280
00:13:55,480 --> 00:13:58,880
Speaker 2: to open Ai. In any case, open AI's exposure could

281
00:13:58,920 --> 00:14:02,040
Speaker 2: be much, much, much higher than billion dollars. And I'll

282
00:14:02,080 --> 00:14:04,120
Speaker 2: explain in greater depth how I've reached the figure at

283
00:14:04,120 --> 00:14:06,640
Speaker 2: the women I get there. But nevertheless, it could be

284
00:14:06,720 --> 00:14:09,520
Speaker 2: much higher. If OpenAI has to contribute significantly to the

285
00:14:09,600 --> 00:14:13,760
Speaker 2: costs of associating with Stargate in general, it could cost

286
00:14:13,800 --> 00:14:15,839
Speaker 2: them a lot of money. Data centers aren't something you

287
00:14:15,880 --> 00:14:19,720
Speaker 2: can do. Funny money math with Data Center Dynamics reports

288
00:14:19,720 --> 00:14:21,560
Speaker 2: that the Abilene site is meant to have two hundred

289
00:14:21,600 --> 00:14:24,120
Speaker 2: megawatts of compute capacity in the first half of twenty

290
00:14:24,160 --> 00:14:25,920
Speaker 2: twenty five, and then as much as one point two

291
00:14:25,960 --> 00:14:28,680
Speaker 2: gigawats by mid twenty twenty six. To give you a

292
00:14:28,720 --> 00:14:31,120
Speaker 2: sense of the total cost of the project, Former Microsoft

293
00:14:31,240 --> 00:14:33,920
Speaker 2: VP of Energy Brian Janis said in January that it

294
00:14:33,960 --> 00:14:36,840
Speaker 2: costs twenty five million dollars a megawat or about twenty

295
00:14:36,840 --> 00:14:39,640
Speaker 2: five billion dollars a gigawe, meaning that the initial capital

296
00:14:39,680 --> 00:14:42,000
Speaker 2: expensures for Stargate to spin up its first two hundred

297
00:14:42,040 --> 00:14:45,760
Speaker 2: megawat data center will be around five billion dollars, spiraling

298
00:14:45,760 --> 00:14:48,280
Speaker 2: to thirty billion dollars or more for the entire project.

299
00:14:48,960 --> 00:14:51,120
Speaker 2: The Information has reported that the site, which could be

300
00:14:51,120 --> 00:14:53,320
Speaker 2: and I quote potentially one of the biggest AAY identify

301
00:14:53,560 --> 00:14:56,040
Speaker 2: data centers, could cost fifty billion to one hundred billion

302
00:14:56,040 --> 00:15:00,280
Speaker 2: dollars in the coming years. Where's that fucking money coming from? Me?

303
00:15:00,320 --> 00:15:02,520
Speaker 2: Stick with the lower end of these cost estimates, it's

304
00:15:02,560 --> 00:15:04,520
Speaker 2: likely the open AI is on the hook for over

305
00:15:04,560 --> 00:15:07,200
Speaker 2: five billion dollars for the Abilene site. Based on the

306
00:15:07,280 --> 00:15:10,400
Speaker 2: nineteen billion dollars it's committed to the overall Stargate Data

307
00:15:10,400 --> 00:15:13,800
Speaker 2: center project. This expenditure won't come all at once and

308
00:15:13,840 --> 00:15:17,000
Speaker 2: will be spread across several years. Still, assuming even the

309
00:15:17,160 --> 00:15:19,640
Speaker 2: rosiest numbers, it's hard to see how open ai doesn't

310
00:15:19,640 --> 00:15:21,680
Speaker 2: have to pony up at least a billi in twenty

311
00:15:21,720 --> 00:15:24,200
Speaker 2: twenty five, and that's likely because the development at this

312
00:15:24,320 --> 00:15:26,480
Speaker 2: site is going to be heavily delayed by both tariffs,

313
00:15:26,720 --> 00:15:30,760
Speaker 2: labor shortages, and oracles as reported by Information. Well, they're

314
00:15:30,800 --> 00:15:33,000
Speaker 2: trusting and I this is the kind of quote you

315
00:15:33,000 --> 00:15:36,640
Speaker 2: really want to hear. They're trusting scrappy but unproven startups

316
00:15:36,640 --> 00:15:39,280
Speaker 2: to develop the project. Is that good? Is that who

317
00:15:39,320 --> 00:15:41,680
Speaker 2: you want doing this? I know when I get a

318
00:15:41,720 --> 00:15:44,720
Speaker 2: contractor in to fix something, my first thought is, is

319
00:15:44,760 --> 00:15:48,280
Speaker 2: this guy's scrappy? Anyway, Let's talk about the other costs.

320
00:15:48,440 --> 00:15:51,120
Speaker 2: They're at least three point five billion based on reporting

321
00:15:51,160 --> 00:15:53,480
Speaker 2: from the information from last year, open ai will spend

322
00:15:53,480 --> 00:15:56,520
Speaker 2: at least two point five billion dollars across salaries, data

323
00:15:56,600 --> 00:15:59,400
Speaker 2: referring to buying data from other companies, hosting another cost

324
00:15:59,440 --> 00:16:01,920
Speaker 2: of sales in sales and marketing, and then another billion

325
00:16:01,960 --> 00:16:05,080
Speaker 2: on what infrastructure open ai owns. I expect the latter

326
00:16:05,120 --> 00:16:07,960
Speaker 2: cost to balloon with open AI's investment in physical infrastructure

327
00:16:08,000 --> 00:16:11,640
Speaker 2: for stargame. Here's another bloody question. How does open ai

328
00:16:11,760 --> 00:16:15,480
Speaker 2: meet its obligations? Based on previous estimates, open ai spends

329
00:16:15,520 --> 00:16:17,560
Speaker 2: about two dollars and twenty five cents to make a buck.

330
00:16:17,920 --> 00:16:20,200
Speaker 2: At that rate, it's likely the open AI's cost in

331
00:16:20,280 --> 00:16:23,440
Speaker 2: its most rosy reed revenue projections of twelve point seven

332
00:16:23,440 --> 00:16:26,240
Speaker 2: billion dollars or at least twenty eight billion, meaning that

333
00:16:26,280 --> 00:16:28,960
Speaker 2: it's on course to burn at least fourteen billion dollars

334
00:16:29,000 --> 00:16:31,920
Speaker 2: in twenty twenty five a loan, assuming that open ai

335
00:16:32,120 --> 00:16:35,560
Speaker 2: has literally all of the money they had from last year.

336
00:16:35,640 --> 00:16:37,920
Speaker 2: They don't, but for the sake of argument, let's pretend

337
00:16:37,920 --> 00:16:40,440
Speaker 2: they have ten billion dollars, as well as ten billion

338
00:16:40,480 --> 00:16:43,440
Speaker 2: from soft Bank. It's still unclear how they pay for

339
00:16:43,520 --> 00:16:48,120
Speaker 2: everything now. While open ai likely has preferential payment structures

340
00:16:48,160 --> 00:16:51,160
Speaker 2: with all of their vendors, such as discounted rates with Microsoft,

341
00:16:51,240 --> 00:16:54,880
Speaker 2: Razi or cloud Services, it will eventually have to pay someone,

342
00:16:55,000 --> 00:16:57,720
Speaker 2: especially in the case of costs related to stargate, many

343
00:16:57,760 --> 00:17:00,880
Speaker 2: of which are upfront and involve physical things happening. In

344
00:17:00,920 --> 00:17:03,600
Speaker 2: the event that its costs a severe as reporting suggests,

345
00:17:03,640 --> 00:17:06,320
Speaker 2: open AI's revenue comes at a terrible cost and will

346
00:17:06,400 --> 00:17:09,280
Speaker 2: likely be immediately funneled directly into funding the obscene cost

347
00:17:09,320 --> 00:17:12,680
Speaker 2: behind inference and training models like GPT four point five,

348
00:17:12,880 --> 00:17:17,280
Speaker 2: which Sam Wuonton called a giant expensive model, and yeah,

349
00:17:17,320 --> 00:17:21,000
Speaker 2: nevertheless he pushed that to every single user. Worse. Still,

350
00:17:21,080 --> 00:17:24,560
Speaker 2: open ai has while delaying its next model, GPT five

351
00:17:24,800 --> 00:17:27,720
Speaker 2: promised to launch after all that its O three reasoning

352
00:17:27,760 --> 00:17:30,440
Speaker 2: model after saying it wouldn't do so, which is strange

353
00:17:30,480 --> 00:17:33,200
Speaker 2: because it turns up three is actually way more expensive

354
00:17:33,200 --> 00:17:35,960
Speaker 2: to run that people thought. With arc Price Foundation, a

355
00:17:36,000 --> 00:17:39,520
Speaker 2: nonprofit that makes the RKGI test for benchmarking models, because

356
00:17:39,560 --> 00:17:41,480
Speaker 2: real tests don't work on them, they need to make

357
00:17:41,520 --> 00:17:45,000
Speaker 2: something out. Well, they're estimating they will cost about thirty

358
00:17:45,040 --> 00:17:48,159
Speaker 2: grand a task, which is that's a lot of money.

359
00:17:49,280 --> 00:17:52,440
Speaker 2: And yes, they again it. Open Ai has a new

360
00:17:52,480 --> 00:17:54,760
Speaker 2: sugar Daddy in the form of SoftBank. But SoftBank has

361
00:17:54,800 --> 00:17:58,959
Speaker 2: to borrow money to meet its obligations for Stargate and

362
00:17:59,200 --> 00:18:03,439
Speaker 2: also open Ai, and this is leading to its financial

363
00:18:03,480 --> 00:18:07,240
Speaker 2: condition likely deteriorating. And that's S and P Global that

364
00:18:07,440 --> 00:18:10,320
Speaker 2: you know, you know, the S and P five those people,

365
00:18:10,920 --> 00:18:13,240
Speaker 2: those people said that that's not what you want to hear.

366
00:18:13,720 --> 00:18:15,680
Speaker 2: Let's look soft Bank. As of right now, soft Bank

367
00:18:15,720 --> 00:18:18,359
Speaker 2: has committed to the following at least thirty billion dollars

368
00:18:18,359 --> 00:18:20,720
Speaker 2: in funding, is part of open AI's recent forty billion

369
00:18:20,760 --> 00:18:24,080
Speaker 2: dollar funding round. Now. Soft Bank's filing surrounding open AI's

370
00:18:24,119 --> 00:18:26,720
Speaker 2: funding also suggests that they are ultimately on the hook

371
00:18:26,760 --> 00:18:30,280
Speaker 2: for the entire forty billion, but they can syndicate it.

372
00:18:30,359 --> 00:18:33,879
Speaker 2: Like I mentioned earlier, reporting suggests that the syndication will happen,

373
00:18:34,040 --> 00:18:36,280
Speaker 2: and it will happen with people at COUTE, Microsoft and

374
00:18:36,320 --> 00:18:40,000
Speaker 2: other investors. Now here's the funny part. If open ai

375
00:18:40,080 --> 00:18:42,320
Speaker 2: fails to convert to a for profit, that forty billion

376
00:18:42,320 --> 00:18:44,960
Speaker 2: dollars slashed down to a paltry thirty billion dollars, although

377
00:18:44,960 --> 00:18:48,000
Speaker 2: again soft Banks shares contingent upon whether it finds other

378
00:18:48,080 --> 00:18:51,080
Speaker 2: investors to join the deal. Now there's another three billion

379
00:18:51,119 --> 00:18:54,000
Speaker 2: dollars that soft Bank's promised to spend on open AI's tech,

380
00:18:54,560 --> 00:18:57,680
Speaker 2: and then nineteen billion dollars for the Stargate Data Center project,

381
00:18:57,680 --> 00:19:01,119
Speaker 2: which soft Bank is taking full financial responsibility for. The

382
00:19:01,119 --> 00:19:04,920
Speaker 2: information reports, and the total is either fifty two billion

383
00:19:05,000 --> 00:19:07,040
Speaker 2: dollars or sixty two billion dollars, with at least ten

384
00:19:07,080 --> 00:19:09,240
Speaker 2: billion dollars due by the end of twenty twenty five,

385
00:19:09,640 --> 00:19:12,679
Speaker 2: but more like twenty Like there's so much money. To

386
00:19:12,760 --> 00:19:16,679
Speaker 2: be clear, SoftBank had to borrow all of the ten

387
00:19:16,760 --> 00:19:19,120
Speaker 2: billion dollars. How the fuck are they meant to get

388
00:19:19,119 --> 00:19:22,399
Speaker 2: the thirty billion dollars? Now I kind of mentioned s

389
00:19:22,440 --> 00:19:24,240
Speaker 2: and P. But I want to really return to that now,

390
00:19:24,280 --> 00:19:27,320
Speaker 2: SoftBank's exposure to open Ai is materially harming the company,

391
00:19:27,359 --> 00:19:30,600
Speaker 2: and I'm quoting the Wall Street Journal here. Ratings agency

392
00:19:30,640 --> 00:19:33,560
Speaker 2: S and P Global said Tuesday that SoftBank's financial condition

393
00:19:33,600 --> 00:19:36,640
Speaker 2: will likely deteriorate as a result of the open Ai investment,

394
00:19:36,880 --> 00:19:39,280
Speaker 2: and that its plans to add debt could lead the

395
00:19:39,280 --> 00:19:43,560
Speaker 2: agency to consider downgrading SoftBank's ratings. While one might argue

396
00:19:43,560 --> 00:19:45,440
Speaker 2: that SoftBank has a good amount of cash, the journal

397
00:19:45,480 --> 00:19:48,000
Speaker 2: also adds that they're somewhat hamstrung in its use as

398
00:19:48,040 --> 00:19:51,800
Speaker 2: a result of CEO Massaoshi's son's dick shit reckless gambles.

399
00:19:51,920 --> 00:19:54,600
Speaker 2: Again quoting from the Wall Street Journal, SoftBank had a

400
00:19:54,600 --> 00:19:57,320
Speaker 2: decent buffer of thirty one billion dollars of cash as

401
00:19:57,359 --> 00:19:59,600
Speaker 2: of December thirty first, but the company is also pledged

402
00:19:59,600 --> 00:20:01,440
Speaker 2: to hold mone much of that in reserve to quell

403
00:20:01,480 --> 00:20:04,479
Speaker 2: worried investors. Soft Bank is committed to not borrow more

404
00:20:04,520 --> 00:20:06,359
Speaker 2: than twenty five percent of the value of its holdings,

405
00:20:06,359 --> 00:20:08,560
Speaker 2: which means it will likely need to sell some of

406
00:20:08,600 --> 00:20:10,520
Speaker 2: the other parts of its empire to pay the rest

407
00:20:10,520 --> 00:20:14,320
Speaker 2: of the open Ai deal. Worse still, it seems that

408
00:20:14,400 --> 00:20:17,760
Speaker 2: as mentioned before that SoftBank will be financing the entirety

409
00:20:17,840 --> 00:20:20,280
Speaker 2: of the first ten billion seven point five billion if

410
00:20:20,280 --> 00:20:23,360
Speaker 2: they're able to syndicate it. As a result, soft Bank

411
00:20:23,440 --> 00:20:25,000
Speaker 2: is likely going to have to sell off parts of

412
00:20:25,040 --> 00:20:28,240
Speaker 2: their actually valuable holdings in companies like Alababa or arm

413
00:20:28,359 --> 00:20:30,840
Speaker 2: or worse still, parts of their ailing investments from a

414
00:20:30,960 --> 00:20:34,400
Speaker 2: vision fund, resulting in a material loss and its underwater deals.

415
00:20:35,200 --> 00:20:39,560
Speaker 2: This is an untenable strategy, and I'd like to explain why. First.

416
00:20:39,680 --> 00:20:42,440
Speaker 2: Open Ai needs at least forty billion dollars a year

417
00:20:42,480 --> 00:20:45,679
Speaker 2: to survive and its costs are only increasing. While we

418
00:20:45,760 --> 00:20:48,600
Speaker 2: don't have much transparency into open AI's actual day to

419
00:20:48,680 --> 00:20:51,720
Speaker 2: day finances, we can make educated guesses that its costs

420
00:20:51,720 --> 00:20:54,240
Speaker 2: are increasing based on the amount of capital it's raising.

421
00:20:54,560 --> 00:20:57,200
Speaker 2: If open AI's costs were flat or only mildly increasing,

422
00:20:57,280 --> 00:20:59,720
Speaker 2: we'd expect to see raises roughly the same size as

423
00:20:59,720 --> 00:21:01,680
Speaker 2: the pre vis one six point six billions something in

424
00:21:01,760 --> 00:21:04,680
Speaker 2: that range. A forty billion dollar raise is nearly six

425
00:21:04,800 --> 00:21:08,640
Speaker 2: times the previous funding round. Admittedly, multiples like that aren't

426
00:21:08,680 --> 00:21:11,679
Speaker 2: particularly unusual. If a company raises three hundred grand in

427
00:21:11,680 --> 00:21:13,960
Speaker 2: the pre seed round and a three million dollar series

428
00:21:14,000 --> 00:21:17,600
Speaker 2: a round in funding. That's a tenfold increase. But we're

429
00:21:17,600 --> 00:21:20,040
Speaker 2: not talking about hundreds of thousands of dollars or even

430
00:21:20,119 --> 00:21:23,360
Speaker 2: millions of dollars. We're talking about billions of dollars. If

431
00:21:23,400 --> 00:21:26,040
Speaker 2: open AI's funding round with soft Bank goes as planned,

432
00:21:26,200 --> 00:21:29,280
Speaker 2: it will raise the equivalent of the entire GDP of Estonia,

433
00:21:29,520 --> 00:21:32,200
Speaker 2: a fairly wealthy country itself that's also a member of

434
00:21:32,280 --> 00:21:34,879
Speaker 2: NATO and the European Union. That alone should give you

435
00:21:34,920 --> 00:21:38,680
Speaker 2: a sense of how fucking stupid this all is. Stupid, sure,

436
00:21:38,760 --> 00:21:42,639
Speaker 2: but undoubtedly necessary. Per the information, open ai plans to

437
00:21:42,720 --> 00:21:45,040
Speaker 2: spend as much as twenty eight billion dollars in compute

438
00:21:45,040 --> 00:21:47,840
Speaker 2: on Microsoft dosy or cloud in twenty twenty eight. Over

439
00:21:47,880 --> 00:21:50,560
Speaker 2: a third of open AI's revenue, per the same article,

440
00:21:50,760 --> 00:21:55,000
Speaker 2: will come from SoftBank's alleged spend. It's reasonable to believe

441
00:21:55,040 --> 00:21:57,200
Speaker 2: that open ai will, as a result, need to raise

442
00:21:57,240 --> 00:21:59,880
Speaker 2: in excess of forty billion dollars a year, though it's

443
00:21:59,880 --> 00:22:01,840
Speaker 2: reasonable to believe that it will need to raise more

444
00:22:01,880 --> 00:22:05,320
Speaker 2: like fifty billion a year until they reach profitability you

445
00:22:05,359 --> 00:22:09,240
Speaker 2: know never And now this actually has a reason. It's

446
00:22:09,320 --> 00:22:11,199
Speaker 2: due to the growing cost of doing business as well

447
00:22:11,240 --> 00:22:13,879
Speaker 2: as the various infrastructure commitments they've made, both in the

448
00:22:13,920 --> 00:22:16,680
Speaker 2: terms of Stargate as well as deals with third party

449
00:22:16,680 --> 00:22:20,960
Speaker 2: suppliers like core Weavan, indeed, Microsoft. Open Ai CEO Sam

450
00:22:20,960 --> 00:22:23,960
Speaker 2: Mortman's statement around costs also suggests that they're going up

451
00:22:24,080 --> 00:22:27,120
Speaker 2: quite fiercely. In late February, he claimed that open ai

452
00:22:27,359 --> 00:22:29,960
Speaker 2: was out of GPUs. While this suggests that there is

453
00:22:30,280 --> 00:22:33,360
Speaker 2: demand for some products like it's horrifying him as generating

454
00:22:33,520 --> 00:22:37,359
Speaker 2: tech that has made abominations that insult Miyazaki and I

455
00:22:37,480 --> 00:22:40,080
Speaker 2: will hate them for it forever and thus made them

456
00:22:40,119 --> 00:22:42,280
Speaker 2: go viral in much it also means that to meet

457
00:22:42,320 --> 00:22:46,080
Speaker 2: the demand of the horrible abominations, open Ai needs to

458
00:22:46,080 --> 00:22:48,480
Speaker 2: spend more, and at the risk of repeating myself, that

459
00:22:48,560 --> 00:22:53,760
Speaker 2: demand doesn't necessarily translate into revenue or profitability. Now. Second,

460
00:22:53,920 --> 00:22:56,960
Speaker 2: soft Bank cannot fund open ai long term. I must

461
00:22:57,040 --> 00:22:59,480
Speaker 2: be clear. Open AI's costs are projected to be three

462
00:22:59,600 --> 00:23:04,119
Speaker 2: hundred twenty billion dollars over the next five years. SoftBank

463
00:23:04,160 --> 00:23:07,320
Speaker 2: has to overcome significant challenges to fund both open ai

464
00:23:07,480 --> 00:23:09,720
Speaker 2: and it's part of Stargate, and when I say fund,

465
00:23:09,760 --> 00:23:13,160
Speaker 2: I mean fund the current state of both projects, assuming

466
00:23:13,240 --> 00:23:17,720
Speaker 2: no further obligations or complications. The Information reports the open

467
00:23:17,760 --> 00:23:20,160
Speaker 2: Ai forecast that will spend as I mentioned, twenty eight

468
00:23:20,200 --> 00:23:22,360
Speaker 2: billion dollars on compute with Microsoft a loan in twenty

469
00:23:22,400 --> 00:23:25,359
Speaker 2: twenty eight, and I apologize for repeating myself, these numbers

470
00:23:25,400 --> 00:23:28,400
Speaker 2: fucking matter. The same arcle also reports that open ai

471
00:23:28,480 --> 00:23:30,400
Speaker 2: would turn profitable by the end of the decade after

472
00:23:30,440 --> 00:23:34,240
Speaker 2: building out Stargate, suggesting that open AI's operating expend ditches

473
00:23:34,280 --> 00:23:37,880
Speaker 2: will grow exponentially year over year. According to the Information,

474
00:23:37,920 --> 00:23:41,280
Speaker 2: it expects its costs to surpass three hundred and twenty

475
00:23:41,320 --> 00:23:43,639
Speaker 2: billion dollars between twenty twenty five and twenty thirty, with

476
00:23:43,720 --> 00:23:46,160
Speaker 2: half of that going towards funding model training and development.

477
00:23:47,400 --> 00:23:50,159
Speaker 2: How the fuck does building stargate make them profitable? I

478
00:23:50,200 --> 00:23:53,919
Speaker 2: really can't anyway, it won't. It's the same shit. But

479
00:23:54,000 --> 00:23:56,800
Speaker 2: SoftBank has had to and will continue having to go

480
00:23:56,880 --> 00:24:00,280
Speaker 2: to remarkable lengths to fund open AI's current forty billion

481
00:24:00,320 --> 00:24:02,919
Speaker 2: dollar round, lengths so significant that it may lead to

482
00:24:02,960 --> 00:24:06,080
Speaker 2: their credit rating being further downgraded. As I mentioned, even

483
00:24:06,119 --> 00:24:09,879
Speaker 2: if we assume the best case scenario, open ai successfully

484
00:24:09,920 --> 00:24:11,760
Speaker 2: converts to a for profit entity by the end of

485
00:24:11,760 --> 00:24:14,560
Speaker 2: twenty twenty five and gets that full thirty billion dollars.

486
00:24:15,200 --> 00:24:18,199
Speaker 2: It seems unlikely, if not impossible, for it to continue

487
00:24:18,280 --> 00:24:21,600
Speaker 2: raising the amount of capital they need to to continue operations.

488
00:24:22,040 --> 00:24:24,359
Speaker 2: As I've argued in multiple news letters and podcasts, there

489
00:24:24,359 --> 00:24:26,200
Speaker 2: are only a few entities that can provide the kind

490
00:24:26,200 --> 00:24:28,840
Speaker 2: of funding that open AI needs. These include big tech

491
00:24:28,880 --> 00:24:31,960
Speaker 2: focused investment firms like soft Bank, sovereign wealth funds like

492
00:24:31,960 --> 00:24:34,359
Speaker 2: those of Saudi Arabia and the UAE, and perhaps the

493
00:24:34,440 --> 00:24:37,560
Speaker 2: largest tech companies. I also want to be clear that

494
00:24:37,920 --> 00:24:40,600
Speaker 2: I keep getting messages being like the gatherment could do

495
00:24:40,720 --> 00:24:42,679
Speaker 2: it daily, couldn't it get me word? If the government

496
00:24:42,680 --> 00:24:44,840
Speaker 2: did get it's not going to happen. Stop fucking asking me.

497
00:24:45,040 --> 00:24:47,720
Speaker 2: It's just not going to happen. The government's not sending

498
00:24:47,800 --> 00:24:50,480
Speaker 2: forty billion dollars to Sam Altman. They're not going to

499
00:24:50,520 --> 00:24:53,400
Speaker 2: do it. I will apologize personally to each and every

500
00:24:53,480 --> 00:24:55,280
Speaker 2: one of you if I'm wrong, but they will need

501
00:24:55,320 --> 00:24:58,359
Speaker 2: to send forty billion dollars cash to open ai. But

502
00:24:58,440 --> 00:25:01,160
Speaker 2: these entities can meet open ai needs for some time,

503
00:25:01,200 --> 00:25:05,119
Speaker 2: but not all the time. It's not realistic to expect

504
00:25:05,119 --> 00:25:07,639
Speaker 2: soft Bank or Microsoft or the Soald's Oracle or whoever

505
00:25:07,680 --> 00:25:10,640
Speaker 2: to provide forty billion dollars every year for the foreseeable future.

506
00:25:11,000 --> 00:25:15,040
Speaker 2: And just to be clear, this is what open Ai needs. Eventually,

507
00:25:15,040 --> 00:25:17,800
Speaker 2: even the Souds have to have a break. And I

508
00:25:17,800 --> 00:25:19,560
Speaker 2: don't know if you remember from a previous episode, but

509
00:25:19,560 --> 00:25:21,840
Speaker 2: Masayoshi's son gets a lot of his money from them,

510
00:25:22,040 --> 00:25:24,680
Speaker 2: and they're not happy with him. In fact, Masaoshi's son

511
00:25:24,720 --> 00:25:27,080
Speaker 2: has said a few years ago that he owed Muhammad

512
00:25:27,080 --> 00:25:31,080
Speaker 2: bill and Salmon. Not a great guy to owe anyway.

513
00:25:31,119 --> 00:25:33,160
Speaker 2: This really is especially true for soft Bank by the way.

514
00:25:33,200 --> 00:25:36,000
Speaker 2: They're bruised. They're battered after several rough years, including a

515
00:25:36,000 --> 00:25:39,800
Speaker 2: failed multi billion dollar investment in we Work. Based on

516
00:25:39,840 --> 00:25:41,639
Speaker 2: its current promise to not borrow more than twenty five

517
00:25:41,680 --> 00:25:44,320
Speaker 2: percent of its holdings, it's near impossible that SoftBank will

518
00:25:44,359 --> 00:25:46,920
Speaker 2: be able to continue to fund open Ai at this rate,

519
00:25:47,400 --> 00:25:50,399
Speaker 2: and forty billion dollars a year may not actually be enough.

520
00:25:51,520 --> 00:25:54,640
Speaker 2: Based on soft Bank's last reported equity value of its holdings,

521
00:25:54,720 --> 00:25:57,400
Speaker 2: they have about two hundred and twenty nine billion dollars

522
00:25:57,440 --> 00:25:59,920
Speaker 2: of stuff, meaning that they can borrow just over fifty

523
00:26:00,080 --> 00:26:03,600
Speaker 2: seven billion while remaining compliant with these guidelines. In any case,

524
00:26:03,640 --> 00:26:07,000
Speaker 2: it's unclear how soft Bank can fund open Ai, but

525
00:26:07,040 --> 00:26:24,040
Speaker 2: it's far clearer that nobody else is willing to. Now

526
00:26:24,440 --> 00:26:26,920
Speaker 2: we're going to move off the questions for a minute

527
00:26:26,960 --> 00:26:29,680
Speaker 2: because I just want to get into some problems. I've

528
00:26:29,720 --> 00:26:32,280
Speaker 2: popped because it turns out that open ai they got

529
00:26:32,320 --> 00:26:36,399
Speaker 2: doodoo ass. That's a professional finance term. I did just

530
00:26:36,400 --> 00:26:38,520
Speaker 2: say financi and we're just going to keep that. Now.

531
00:26:38,600 --> 00:26:41,119
Speaker 2: Open Ai has started running into capacity issues, and this

532
00:26:41,240 --> 00:26:44,119
Speaker 2: is a real problem, and it suggests material instability in

533
00:26:44,160 --> 00:26:47,600
Speaker 2: their business or infrastructure, and it's really not clear how

534
00:26:47,640 --> 00:26:51,479
Speaker 2: open ai expands further. Let me explain it's important to

535
00:26:51,480 --> 00:26:53,720
Speaker 2: note that open ai does not really have any of

536
00:26:53,720 --> 00:26:56,840
Speaker 2: its own compute infrastructure. The majority of its compute is

537
00:26:56,880 --> 00:27:00,679
Speaker 2: provided by Microsoft, though as mentioned previously, open now has

538
00:27:00,680 --> 00:27:02,880
Speaker 2: a deal with core Weave to take over the capacity

539
00:27:02,880 --> 00:27:05,200
Speaker 2: that Microsoft was going to have twelve billion dollars or

540
00:27:05,240 --> 00:27:08,400
Speaker 2: so of capacity in the future. Anyway, in the last

541
00:27:08,520 --> 00:27:10,560
Speaker 2: ninety days, Sam Altman has complained about the lack of

542
00:27:10,640 --> 00:27:13,919
Speaker 2: GPUs and pressure on open ai servers many times. In

543
00:27:13,960 --> 00:27:16,640
Speaker 2: my newsletter published a few days ago, I was did

544
00:27:16,720 --> 00:27:20,320
Speaker 2: six such examples. Should you be curious, these statements in

545
00:27:20,359 --> 00:27:23,000
Speaker 2: a bubble seem either harmless or like open AI's growth

546
00:27:23,080 --> 00:27:25,600
Speaker 2: is skyrocketing, the latter of which might indeed be true,

547
00:27:25,600 --> 00:27:27,760
Speaker 2: but botzil for a company that burns money with every

548
00:27:27,760 --> 00:27:31,240
Speaker 2: single user, any mention of rate limits or performance issues

549
00:27:31,240 --> 00:27:34,399
Speaker 2: suggests that open ai is having significant capacity issues, and

550
00:27:34,480 --> 00:27:36,679
Speaker 2: at this point it's unclear what further capacity it can

551
00:27:36,720 --> 00:27:40,439
Speaker 2: actually expand to outside of that currently available. Sam Wilton's

552
00:27:40,440 --> 00:27:44,600
Speaker 2: complaining about melting GPUs, You've got the lead from Sourer saying, yeah,

553
00:27:44,640 --> 00:27:47,160
Speaker 2: we're gonna have some problems showing you stuff for a minute,

554
00:27:47,359 --> 00:27:49,520
Speaker 2: just because like, yeah, we're melting GPUs. There was a

555
00:27:49,560 --> 00:27:52,720
Speaker 2: whole thing about how you should expect delays on product

556
00:27:52,800 --> 00:27:56,720
Speaker 2: launches and service problems. None of it's really good, And

557
00:27:56,920 --> 00:27:59,560
Speaker 2: like I said, it isn't really obvious how open ai

558
00:27:59,680 --> 00:28:03,000
Speaker 2: is going to expand much further. Remember Microsoft has now

559
00:28:03,040 --> 00:28:06,160
Speaker 2: pulled as much as two gigawatts of data center projects,

560
00:28:06,280 --> 00:28:09,240
Speaker 2: walked away from a billion dollar data center development in

561
00:28:09,280 --> 00:28:11,840
Speaker 2: Ohio and declined the option the one I just mentioned

562
00:28:11,920 --> 00:28:14,360
Speaker 2: on twelve billion dollars a compute from Corewave that open

563
00:28:14,400 --> 00:28:17,439
Speaker 2: ai had to pick up, Meaning that open ai may

564
00:28:17,480 --> 00:28:20,200
Speaker 2: be pushing up against the limits of what is physically available.

565
00:28:20,720 --> 00:28:24,120
Speaker 2: While The total available capacity of GPUs at many providers

566
00:28:24,160 --> 00:28:26,400
Speaker 2: like Lambda and Crusoe is unknown, and indeed, I don't

567
00:28:26,400 --> 00:28:28,919
Speaker 2: know if Cruso has a single data center at this point.

568
00:28:29,119 --> 00:28:31,879
Speaker 2: We know that Corewave has upon approximately three hundred and

569
00:28:31,960 --> 00:28:35,720
Speaker 2: sixty megawats available compared to microsoft six point five to

570
00:28:35,760 --> 00:28:38,960
Speaker 2: seven point five gigawads, a large chunk of which I

571
00:28:39,000 --> 00:28:42,360
Speaker 2: think powers open Ai. If open ai is running into

572
00:28:42,360 --> 00:28:44,880
Speaker 2: capacity issues, it could be one of the following. They

573
00:28:44,880 --> 00:28:46,800
Speaker 2: could be running up against the limit of what Microsoft

574
00:28:46,800 --> 00:28:49,000
Speaker 2: has available or is willing to offer the company. The

575
00:28:49,000 --> 00:28:51,680
Speaker 2: Information reported in October twenty twenty four the open Ai

576
00:28:51,880 --> 00:28:55,040
Speaker 2: was frustrated with Microsoft and said it wasn't moving fast

577
00:28:55,120 --> 00:28:58,200
Speaker 2: enough to supply open ai with servers and now. It

578
00:28:58,200 --> 00:29:00,920
Speaker 2: could also be that while open ai is cony is sufficient,

579
00:29:00,920 --> 00:29:03,719
Speaker 2: it does not have the resources available to easily handle

580
00:29:03,760 --> 00:29:06,320
Speaker 2: bursts in user growth in the stable manner. Per The

581
00:29:06,360 --> 00:29:09,880
Speaker 2: Information's reporting, Microsoft promised open Ai three hundred thousand in

582
00:29:09,960 --> 00:29:12,560
Speaker 2: Vidio GB two hundred Blackwell chips. By the end of

583
00:29:12,560 --> 00:29:15,600
Speaker 2: twenty twenty five, are roughly eighteen billion dollars worth of them.

584
00:29:15,800 --> 00:29:18,720
Speaker 2: It's unclear if this has changed since Microsoft allowed open

585
00:29:18,760 --> 00:29:21,160
Speaker 2: Ai to seek other compute from other companies in late

586
00:29:21,240 --> 00:29:24,200
Speaker 2: January twenty twenty five. I also don't believe that open

587
00:29:24,200 --> 00:29:27,640
Speaker 2: ai has any other viable options for existing compute infrastructure

588
00:29:27,640 --> 00:29:31,320
Speaker 2: outside of Microsoft. Corweep's current data centers mostly feature in

589
00:29:31,400 --> 00:29:34,840
Speaker 2: Vidia's aging Hopper GPUs, and while it could and likely

590
00:29:35,000 --> 00:29:38,400
Speaker 2: is retrofitting its current infrastructure with Blackwell chips, doing so

591
00:29:38,520 --> 00:29:41,920
Speaker 2: isn't easy or cheap. Blackwell chips require far more powerful

592
00:29:41,960 --> 00:29:44,600
Speaker 2: cooling and server infrastructure to make them run smoothly, a

593
00:29:44,600 --> 00:29:46,400
Speaker 2: problem which led to a delay in their delivery to

594
00:29:46,440 --> 00:29:49,240
Speaker 2: most customers according to the information. And even if Corey

595
00:29:49,480 --> 00:29:52,400
Speaker 2: was able to replace every last Hopper GPU with a Blackwell,

596
00:29:52,440 --> 00:29:55,280
Speaker 2: and they won't, it still wouldn't match what open ai

597
00:29:55,440 --> 00:29:58,440
Speaker 2: needs to expand one might argue that it simply needs

598
00:29:58,440 --> 00:30:00,800
Speaker 2: to wait for the construction of the Stargus data center,

599
00:30:00,960 --> 00:30:02,560
Speaker 2: or for core we have to finish the gigawa or

600
00:30:02,600 --> 00:30:05,800
Speaker 2: so of construction it's working on. I want to be

601
00:30:05,880 --> 00:30:08,480
Speaker 2: clear how impossible that is. I need to be clear.

602
00:30:08,560 --> 00:30:10,920
Speaker 2: One of my least favorite responses to my work as

603
00:30:10,920 --> 00:30:13,440
Speaker 2: people saying they'll just build more data centers. They just

604
00:30:13,480 --> 00:30:14,920
Speaker 2: go and build them inn They're just going to build

605
00:30:14,960 --> 00:30:17,640
Speaker 2: them right now. They won't. You can't just data centers

606
00:30:17,680 --> 00:30:20,680
Speaker 2: don't just grow from the fucking ground. As I've argued

607
00:30:20,680 --> 00:30:23,120
Speaker 2: in the past, I have serious concerns over the viability

608
00:30:23,120 --> 00:30:26,160
Speaker 2: of core We've ever completing. It's alleged contracted one point

609
00:30:26,200 --> 00:30:29,440
Speaker 2: three gigawatts of capacity. Based on calculations, it'll have to

610
00:30:29,440 --> 00:30:32,080
Speaker 2: spend in excess of thirty nine billion dollars to build it.

611
00:30:32,080 --> 00:30:34,960
Speaker 2: It's unclear how that will happen, and it doesn't have

612
00:30:35,040 --> 00:30:38,360
Speaker 2: the money to do so, like it actually does not

613
00:30:38,400 --> 00:30:41,160
Speaker 2: have the cash. I'll get into that later, but they

614
00:30:41,200 --> 00:30:44,680
Speaker 2: don't have the money. However, even if I were to

615
00:30:44,800 --> 00:30:47,959
Speaker 2: humor this idea, it's impossible that any of this project

616
00:30:48,040 --> 00:30:49,640
Speaker 2: is done by the end of twenty twenty five, and

617
00:30:49,680 --> 00:30:52,560
Speaker 2: I'd argue even in twenty twenty six. I can find

618
00:30:52,560 --> 00:30:54,800
Speaker 2: no commitments to any time scale, other than the fact

619
00:30:54,800 --> 00:30:57,040
Speaker 2: that open Ai will allegedly start paying core Weave in

620
00:30:57,080 --> 00:31:00,080
Speaker 2: October per the information, which could very well be using

621
00:31:00,120 --> 00:31:03,960
Speaker 2: their current capacity. I also can't find any evidence that Cruso,

622
00:31:04,040 --> 00:31:06,760
Speaker 2: the company building the Stargate data center in Texas, has

623
00:31:06,840 --> 00:31:10,520
Speaker 2: any compute anywhere else. Lander, a GPU compute company that

624
00:31:10,640 --> 00:31:12,840
Speaker 2: raised three hundred and twenty million dollars earlier in the year,

625
00:31:13,160 --> 00:31:15,840
Speaker 2: and according to Data Center Dynamics, operates out of co

626
00:31:15,960 --> 00:31:19,200
Speaker 2: location data centers in San Francisco, California, and Allen, Texas.

627
00:31:19,320 --> 00:31:21,120
Speaker 2: It's backed by Moorner eight hundred and twenty million in

628
00:31:21,200 --> 00:31:24,640
Speaker 2: funds raised this year. All of that just kind of

629
00:31:25,000 --> 00:31:27,040
Speaker 2: doesn't say that they have a data center at all.

630
00:31:27,040 --> 00:31:30,480
Speaker 2: A colocation. That means you're in someone else's buddy. You

631
00:31:30,520 --> 00:31:33,640
Speaker 2: don't own a house, you're ranting one. And just to

632
00:31:33,640 --> 00:31:36,640
Speaker 2: be clear, open AI's ability to scale is entirely contingent

633
00:31:36,640 --> 00:31:39,239
Speaker 2: on the availability of whatever data center providers it has

634
00:31:39,280 --> 00:31:42,640
Speaker 2: relationships with, and all of their growths coming from these

635
00:31:42,640 --> 00:31:46,600
Speaker 2: to these two companies. I'll get there, don't worry. But

636
00:31:47,040 --> 00:31:50,480
Speaker 2: every time I kind of say this stuff out loud,

637
00:31:50,520 --> 00:31:54,200
Speaker 2: I feel my soul slightly stripped from me. I feel

638
00:31:54,200 --> 00:31:57,360
Speaker 2: like I feel like I'm in hell because this is

639
00:31:57,400 --> 00:32:00,480
Speaker 2: an insane thing. If you or I win to someone

640
00:32:00,720 --> 00:32:03,160
Speaker 2: went like, hey, look, I'm going to lose five billion dollars,

641
00:32:03,200 --> 00:32:05,240
Speaker 2: but I promise you God will come out of the computer.

642
00:32:05,960 --> 00:32:09,560
Speaker 2: They institutionalize you or me, they definitely would instalize me.

643
00:32:11,000 --> 00:32:13,760
Speaker 2: It's just fucking stupid, But in any case, this means

644
00:32:13,760 --> 00:32:16,160
Speaker 2: that open AI's only real choice for GPUs is either

645
00:32:16,240 --> 00:32:19,800
Speaker 2: Core Weave or Microsoft. And while it's hard to calculate precisely,

646
00:32:19,880 --> 00:32:23,960
Speaker 2: open AI's best case scenario is that sixteen thousand GPUs

647
00:32:24,000 --> 00:32:25,880
Speaker 2: come online in the summer of twenty twenty five as

648
00:32:25,880 --> 00:32:28,200
Speaker 2: part of the Stargate Data Center project. And that's a

649
00:32:28,280 --> 00:32:30,440
Speaker 2: drop in the bucket compared to the three hundred thousand

650
00:32:30,480 --> 00:32:34,120
Speaker 2: of the fucking things that Microsoft have previously promised. Now,

651
00:32:34,160 --> 00:32:39,240
Speaker 2: one last thing, any capacity or expansion issues that happen

652
00:32:39,320 --> 00:32:42,760
Speaker 2: with open Ai will needcap this company, open ai is,

653
00:32:42,800 --> 00:32:45,680
Speaker 2: regardless of how you or I may feel about generative Ai,

654
00:32:46,240 --> 00:32:49,080
Speaker 2: one of the fastest growing companies of all time. It

655
00:32:49,120 --> 00:32:52,440
Speaker 2: currently has, according to its own statements, over five hundred

656
00:32:52,520 --> 00:32:56,400
Speaker 2: million weekly active users. Putting aside that each user is unprofitable.

657
00:32:56,680 --> 00:32:59,640
Speaker 2: Such remarkable growth, especially as it's partially as a result

658
00:32:59,680 --> 00:33:03,400
Speaker 2: of its extremely resourced intensive image generator, is a massive,

659
00:33:03,840 --> 00:33:07,640
Speaker 2: horrifying strain on their infrastructure. The vast majority of open

660
00:33:07,680 --> 00:33:11,080
Speaker 2: aiyes users are free customers using chat GPT with only,

661
00:33:11,160 --> 00:33:14,080
Speaker 2: like I mentioned earlier, twenty million paying subscribers, and the

662
00:33:14,160 --> 00:33:16,720
Speaker 2: vast majority of them on the cheapest twenty bucks plan.

663
00:33:17,280 --> 00:33:20,680
Speaker 2: Open Aiyes, services, even in the case of image generation,

664
00:33:21,240 --> 00:33:24,480
Speaker 2: are relatively commoditized, meaning that users can, if they really care,

665
00:33:24,520 --> 00:33:27,480
Speaker 2: go and use any number of other large language model services.

666
00:33:27,800 --> 00:33:30,840
Speaker 2: They could use bing Oh, they could use stable Diffusion,

667
00:33:30,840 --> 00:33:33,400
Speaker 2: they could even use GROC if they really I don't

668
00:33:33,480 --> 00:33:37,240
Speaker 2: like saying it, but free users they're a burden on

669
00:33:37,280 --> 00:33:40,560
Speaker 2: the company, especially with such piss poor conversion rates, losing

670
00:33:40,840 --> 00:33:43,280
Speaker 2: money with each prompt, which is by the way, also

671
00:33:43,880 --> 00:33:46,760
Speaker 2: case with paying subscribers, and the remarkable popularity of its

672
00:33:46,760 --> 00:33:51,840
Speaker 2: horrible image generator, it only threatens to bring more burdens

673
00:33:51,840 --> 00:33:54,360
Speaker 2: than one off customers that will generate a few abominable

674
00:33:54,400 --> 00:33:58,000
Speaker 2: studio ghibli pictures of Garfield with giant knockers and then

675
00:33:58,040 --> 00:34:02,200
Speaker 2: never return. If AI's growth continues at this rate, it

676
00:34:02,280 --> 00:34:04,800
Speaker 2: will run into capacity issues, and it does not have

677
00:34:04,880 --> 00:34:07,240
Speaker 2: much room to expand. Well, we don't know how much

678
00:34:07,280 --> 00:34:10,040
Speaker 2: capacity they're taking up with Microsoft, or indeed whether Microsoft

679
00:34:10,040 --> 00:34:13,040
Speaker 2: is approaching capacity or otherwise limiting them. We do know

680
00:34:13,080 --> 00:34:15,480
Speaker 2: that open ai has seen reason to beg for access

681
00:34:15,480 --> 00:34:18,640
Speaker 2: for more GPUs. In simpler terms, even if open ai

682
00:34:18,800 --> 00:34:21,319
Speaker 2: wasn't running out of money, even If open ai wasn't

683
00:34:21,440 --> 00:34:25,240
Speaker 2: un horrifyingly unprofitable, it also may not have enough GPUs

684
00:34:25,280 --> 00:34:28,560
Speaker 2: to continue providing its services in a reliable manner. If

685
00:34:28,560 --> 00:34:30,799
Speaker 2: that's the case, there really isn't that much that can

686
00:34:30,840 --> 00:34:33,719
Speaker 2: be done other than significantly limiting free users activity on

687
00:34:33,760 --> 00:34:36,720
Speaker 2: the platform, which is open AI's primary mechanism for revenue

688
00:34:36,760 --> 00:34:40,280
Speaker 2: growth and customer acquisition. Limiting activity or changing the economics

689
00:34:40,280 --> 00:34:43,480
Speaker 2: behind its paid product. And this is quoting sam Orman.

690
00:34:44,200 --> 00:34:46,640
Speaker 2: They could potentially find some way to let people pay

691
00:34:47,040 --> 00:34:49,919
Speaker 2: for compute they want to use more dynamically, that's not good.

692
00:34:50,400 --> 00:34:52,800
Speaker 2: Almans come up with some other ideas, like an idea

693
00:34:52,840 --> 00:34:55,840
Speaker 2: for paid plans on March fourth, where twenty bucks a

694
00:34:55,880 --> 00:34:58,680
Speaker 2: month goes to credits which you can use across features

695
00:34:58,719 --> 00:35:01,319
Speaker 2: like deep research, oh one at four point five sore

696
00:35:01,360 --> 00:35:03,680
Speaker 2: and so on, with no fixed limits per feature, and

697
00:35:03,840 --> 00:35:05,760
Speaker 2: you choose what you want. If you run out of credits,

698
00:35:05,760 --> 00:35:07,600
Speaker 2: you can buy more. I just want to be clear

699
00:35:07,640 --> 00:35:10,000
Speaker 2: that that is a terrible fucking deal. We have no

700
00:35:10,120 --> 00:35:12,240
Speaker 2: idea what the credits would be, and it would definitely

701
00:35:12,239 --> 00:35:14,359
Speaker 2: be rigged so that you would have to buy more.

702
00:35:14,920 --> 00:35:18,360
Speaker 2: He's also brought up things like mentioning losing two hundred

703
00:35:18,400 --> 00:35:20,799
Speaker 2: bucks a month on the pro subscription, but he's a

704
00:35:20,840 --> 00:35:23,799
Speaker 2: funny one. Buried in an article from the Information from

705
00:35:23,800 --> 00:35:26,319
Speaker 2: March fifth is a comment that suggests that open ai

706
00:35:26,440 --> 00:35:29,640
Speaker 2: is considering measures like changing its pricing model entirely, with

707
00:35:29,719 --> 00:35:32,960
Speaker 2: Altman reportedly telling developers in London in February twenty twenty

708
00:35:33,000 --> 00:35:35,279
Speaker 2: five the open aiy is prime to judge twenty or

709
00:35:35,320 --> 00:35:38,279
Speaker 2: thirty percent of pro customers a higher price because of

710
00:35:38,320 --> 00:35:41,520
Speaker 2: how many research queries they're doing. But he suggested in

711
00:35:41,680 --> 00:35:43,960
Speaker 2: a la carte or pays you go approach when it

712
00:35:43,960 --> 00:35:46,279
Speaker 2: comes to agents, though we have to judge more than

713
00:35:46,280 --> 00:35:49,319
Speaker 2: two hundred dollars a month. Fucking just this fucking guy.

714
00:35:50,239 --> 00:35:52,480
Speaker 2: The problem is that with all of these measures, even

715
00:35:52,520 --> 00:35:54,760
Speaker 2: if they succeeded in generating more money for the company,

716
00:35:54,880 --> 00:35:58,319
Speaker 2: also need to reduce the burden on open aiyes available infrastructure.

717
00:35:58,719 --> 00:36:03,040
Speaker 2: Remember remember data centers can take three to six years

718
00:36:03,080 --> 00:36:05,759
Speaker 2: to build, and even with stargates accelerated, and I'd argue

719
00:36:05,800 --> 00:36:09,400
Speaker 2: on realistic timelines, open ai isn't unlocking a tenth of

720
00:36:09,440 --> 00:36:12,880
Speaker 2: the promised compute that Microsoft gave them. Three hundred thousand

721
00:36:12,920 --> 00:36:16,400
Speaker 2: GPUs is a lot sixteen thousand. Really not, so what

722
00:36:16,560 --> 00:36:19,799
Speaker 2: might these capacities look like? What are the consequences? You

723
00:36:19,840 --> 00:36:23,440
Speaker 2: love payal horses. I love payal horses. Let's get riding.

724
00:36:24,080 --> 00:36:27,160
Speaker 2: Though downtime might be an obvious choice. Capacity issues that

725
00:36:27,200 --> 00:36:29,799
Speaker 2: open ai will likely manifest in hard limits on what

726
00:36:29,880 --> 00:36:34,000
Speaker 2: free users can do, some of which I've documented previously. Nevertheless,

727
00:36:34,040 --> 00:36:36,920
Speaker 2: I believe the real pay all horses of capacity issues

728
00:36:36,920 --> 00:36:40,000
Speaker 2: come from arbitrary limits on any given user group, meaning

729
00:36:40,040 --> 00:36:42,880
Speaker 2: both free users and paid users. Some limits on what

730
00:36:42,920 --> 00:36:44,719
Speaker 2: a user can do, a reduction in the number of

731
00:36:44,760 --> 00:36:48,440
Speaker 2: generations of images for paid users, any introduction of peak hours,

732
00:36:48,680 --> 00:36:50,960
Speaker 2: or any increases in prices are assigned that open ai

733
00:36:51,080 --> 00:36:53,920
Speaker 2: is running out of GPUs, which it's already said as

734
00:36:54,000 --> 00:36:57,960
Speaker 2: happening publicly. However, the really obvious thing, the real obvious

735
00:36:58,040 --> 00:37:01,400
Speaker 2: pale horse, would be service degradation delays in generations of

736
00:37:01,440 --> 00:37:04,520
Speaker 2: any kind, five hundred status code errors, or chat GBT

737
00:37:05,000 --> 00:37:08,080
Speaker 2: just failing to produce an answer. Open ai has up

738
00:37:08,160 --> 00:37:11,719
Speaker 2: until this point had fairly impressive uptime. Still, it's if

739
00:37:11,719 --> 00:37:13,839
Speaker 2: it's running up against the wall this streak will end.

740
00:37:15,239 --> 00:37:17,920
Speaker 2: The consequences depend on how often these issues occur. And

741
00:37:17,960 --> 00:37:20,840
Speaker 2: to whom they occur. If free users face service degradation,

742
00:37:20,920 --> 00:37:23,239
Speaker 2: they'll bounce off the product, as their US is likely

743
00:37:23,280 --> 00:37:25,719
Speaker 2: far more fleeting than the paid user, which will begin

744
00:37:25,800 --> 00:37:30,000
Speaker 2: to erode open AI's growth. Ironically, rapid and especially unprecedented

745
00:37:30,040 --> 00:37:32,800
Speaker 2: growth in one of open AI's competitors like Higher or

746
00:37:32,840 --> 00:37:35,560
Speaker 2: Anthropic could also represent a pale horse for open Ai,

747
00:37:35,680 --> 00:37:38,880
Speaker 2: though based on the monthly active users I've seen from Anthropic,

748
00:37:38,920 --> 00:37:41,160
Speaker 2: I don't think that's going to be a problem now.

749
00:37:41,200 --> 00:37:45,120
Speaker 2: If paid users face service degradation, it will likely cause

750
00:37:45,120 --> 00:37:47,120
Speaker 2: the most harm to the company is while paid users

751
00:37:47,200 --> 00:37:49,200
Speaker 2: still lose open Ai money, in the end, at least

752
00:37:49,200 --> 00:37:52,840
Speaker 2: they receive some money. Open Ai is effectively one choice

753
00:37:52,840 --> 00:37:56,400
Speaker 2: here get more GPUs from Microsoft, and really its future

754
00:37:56,440 --> 00:37:59,560
Speaker 2: depends heavily on both Microsoft's generosity and there being enough

755
00:37:59,600 --> 00:38:01,719
Speaker 2: of them time when Microsoft is pulled back from over

756
00:38:01,760 --> 00:38:05,520
Speaker 2: two gigabats of data centers, specifically according to td COD

757
00:38:05,760 --> 00:38:10,560
Speaker 2: because of it moving away from providing compute to open Ai. Now, admittedly,

758
00:38:10,600 --> 00:38:13,600
Speaker 2: open Ai has previously spent more on training models than inference,

759
00:38:13,640 --> 00:38:15,880
Speaker 2: and that's the actual running of them, and the company

760
00:38:15,920 --> 00:38:18,360
Speaker 2: might be able to smooth downtime issues by shifting capacity.

761
00:38:18,400 --> 00:38:20,000
Speaker 2: This would of course have a knock on effect on

762
00:38:20,040 --> 00:38:22,160
Speaker 2: its ability to develop new models, and the company is

763
00:38:22,200 --> 00:38:25,160
Speaker 2: already losing ground, particularly when it comes to Chinese rivals

764
00:38:25,200 --> 00:38:28,000
Speaker 2: like Deepseek. Now, I know this has been a long

765
00:38:28,040 --> 00:38:30,320
Speaker 2: episode and the fact is I'm not even close to finish.

766
00:38:30,480 --> 00:38:33,440
Speaker 2: I have some more tough, tough questions and tough problems

767
00:38:33,440 --> 00:38:36,080
Speaker 2: for open Ai. Tune in to the next episode to

768
00:38:36,120 --> 00:38:38,440
Speaker 2: hear them. I realized this has been a lot, and

769
00:38:38,480 --> 00:38:40,400
Speaker 2: I know you're very patient with me. You let me

770
00:38:40,440 --> 00:38:43,799
Speaker 2: read the long scripts. But this stuff is important. It's

771
00:38:43,840 --> 00:38:46,080
Speaker 2: important to me, and I think you're going to find

772
00:38:46,120 --> 00:38:50,120
Speaker 2: it important too. And I'll have a nice sexy conclusion

773
00:38:50,120 --> 00:38:51,960
Speaker 2: to this at the end of the next one. But

774
00:38:52,239 --> 00:38:54,720
Speaker 2: before I go, please vote for me in the fucking Webbies.

775
00:38:54,760 --> 00:38:56,799
Speaker 2: I want to win. It's my pin post on Blue

776
00:38:56,800 --> 00:38:59,040
Speaker 2: Sky and to wear in my handles eadsit Tran dot

777
00:38:59,080 --> 00:39:00,600
Speaker 2: Com and Blue Sky. Please he vote for me and

778
00:39:00,640 --> 00:39:11,759
Speaker 2: the Webbies. Anyway, catch you next episode. Thank you for

779
00:39:11,800 --> 00:39:14,480
Speaker 2: listening to Better Offline. The editor and composer of the

780
00:39:14,480 --> 00:39:17,600
Speaker 2: Better Offline theme song is Matasowski. You can check out

781
00:39:17,600 --> 00:39:21,279
Speaker 2: more of his music and audio projects at Matasowski dot com,

782
00:39:21,440 --> 00:39:24,760
Speaker 2: M A T T O S O W s ki

783
00:39:25,160 --> 00:39:28,040
Speaker 2: dot com. You can email me at easy at better

784
00:39:28,080 --> 00:39:30,480
Speaker 2: offline dot com or visit better offline dot com to

785
00:39:30,560 --> 00:39:33,400
Speaker 2: find more podcast links and of course my newsletter. I

786
00:39:33,440 --> 00:39:36,160
Speaker 2: also really recommend you go to chat dot Where's Youreed

787
00:39:36,239 --> 00:39:38,480
Speaker 2: dot at to visit the discord, and go to our

788
00:39:38,560 --> 00:39:41,880
Speaker 2: slash Better Offline to check out our reddit. Thank you

789
00:39:41,920 --> 00:39:42,880
Speaker 2: so much for listening.

790
00:39:43,719 --> 00:39:46,400
Speaker 1: Better Offline is a production of cool Zone Media. For

791
00:39:46,520 --> 00:39:49,719
Speaker 1: more from cool Zone Media, visit our website cool Zonemedia

792
00:39:49,760 --> 00:39:52,600
Speaker 1: dot com, or check us out on the iHeartRadio app,

793
00:39:52,640 --> 00:40:15,680
Speaker 1: Apple Podcasts, or wherever you get your podcasts.