1
00:00:03,200 --> 00:00:07,840
Speaker 1: Media. Hello, and welcome to this week's Better Offline Monologue.

2
00:00:07,920 --> 00:00:19,280
Speaker 1: I'm your host ed ze Trunk. Now. I'm turning forty

3
00:00:19,400 --> 00:00:21,800
Speaker 1: in a month or so, and at four years young,

4
00:00:21,920 --> 00:00:24,840
Speaker 1: I'm old enough to remember as far back as December eleventh,

5
00:00:24,840 --> 00:00:27,400
Speaker 1: twenty twenty five, when Disney and open Ai reached an

6
00:00:27,440 --> 00:00:31,040
Speaker 1: agreement to bring beloved characters from across Disney's brands to

7
00:00:31,160 --> 00:00:34,760
Speaker 1: sourer open AI's video platform. As part of the deal,

8
00:00:34,800 --> 00:00:37,479
Speaker 1: Disney would and i quote, become a major customer of

9
00:00:37,520 --> 00:00:40,560
Speaker 1: open Ai, use its API to build new products, tools

10
00:00:40,560 --> 00:00:43,560
Speaker 1: and experiences, as well as showing Sora videos in Disney

11
00:00:43,560 --> 00:00:46,680
Speaker 1: Plus and deploy chat GPT for its employees, as well

12
00:00:46,680 --> 00:00:50,040
Speaker 1: as making a one billion dollar equity investment in open Ai.

13
00:00:50,720 --> 00:00:53,720
Speaker 1: Except now, Sora is dead shot in the head by Clammy,

14
00:00:53,760 --> 00:00:56,680
Speaker 1: Sam Altman and open Ai because it cost fifteen million

15
00:00:56,720 --> 00:00:59,240
Speaker 1: dollars or more to run and had everybody making videos

16
00:00:59,240 --> 00:01:02,800
Speaker 1: of Goofy saying gorsh hold fuck anything that moves like

17
00:01:02,880 --> 00:01:06,040
Speaker 1: Frank from Blue Velvet. To be clear, open aye is

18
00:01:06,040 --> 00:01:08,760
Speaker 1: shutting down both Sora the app and access to the

19
00:01:08,840 --> 00:01:12,720
Speaker 1: SAA video model through its API, framing the changes, focusing

20
00:01:12,760 --> 00:01:15,400
Speaker 1: the company on business and enterprise customers, which is a

21
00:01:15,440 --> 00:01:18,480
Speaker 1: euphemism for a product that never had any real business

22
00:01:18,480 --> 00:01:21,560
Speaker 1: model or product market fit other than burning millions of

23
00:01:21,560 --> 00:01:26,200
Speaker 1: dollars of computerday and violating copyright. I think it's also

24
00:01:26,319 --> 00:01:29,600
Speaker 1: completely insane that nobody bothered to check whether Disney had

25
00:01:29,640 --> 00:01:33,160
Speaker 1: ever invested or even talked about investing that billion dollars.

26
00:01:33,400 --> 00:01:34,800
Speaker 1: I went and had a look, and there wasn't a

27
00:01:34,880 --> 00:01:37,640
Speaker 1: damn thing in any of the Disney's most recent quarterly

28
00:01:37,720 --> 00:01:40,720
Speaker 1: or annual reports. I should have checked this myself, but

29
00:01:40,760 --> 00:01:43,280
Speaker 1: it's wild that nobody else did either. There are business

30
00:01:43,319 --> 00:01:45,840
Speaker 1: and tech reporters, there are people that do this full

31
00:01:46,240 --> 00:01:49,680
Speaker 1: I but I also want to take this moment to

32
00:01:49,680 --> 00:01:54,640
Speaker 1: say something for the bottom my heart. I was right.

33
00:01:54,800 --> 00:01:56,880
Speaker 1: I was fucking right. I said it back in October.

34
00:01:57,000 --> 00:01:59,000
Speaker 1: I've said it. I said it back in twenty twenty four.

35
00:01:59,200 --> 00:02:03,720
Speaker 1: SAA was never business. SAURA was never replacing anything. SAURA

36
00:02:03,920 --> 00:02:08,400
Speaker 1: was always unreliable because the hallucination prone, large language model

37
00:02:08,400 --> 00:02:11,919
Speaker 1: bullshit it's built upon. It was never gonna work. Everyone

38
00:02:11,919 --> 00:02:14,960
Speaker 1: who said otherwise is a goddamn rube, and so many

39
00:02:15,000 --> 00:02:18,480
Speaker 1: people were wrong. The app that CNBC said was challenging

40
00:02:18,480 --> 00:02:21,280
Speaker 1: Hollywood and freaking out the movie industry, and The Hollywood

41
00:02:21,280 --> 00:02:24,560
Speaker 1: Reporter would suggest could somehow challenge Pixar and with sam

42
00:02:24,800 --> 00:02:28,280
Speaker 1: when successfully playing Hollywood, and the fucking anklo who should

43
00:02:28,280 --> 00:02:31,200
Speaker 1: be ashamed of themselves said, was Open Ai going to

44
00:02:31,240 --> 00:02:34,360
Speaker 1: war with Hollywood as it shook the industry, and Deadline

45
00:02:34,600 --> 00:02:37,600
Speaker 1: said made Hollywood sow and Boardrooms said was in a

46
00:02:37,639 --> 00:02:40,720
Speaker 1: standoff with Hollywood, and The La Times said was deepening

47
00:02:40,760 --> 00:02:43,400
Speaker 1: a battle between Hollywood and Open AI and igniting a

48
00:02:43,440 --> 00:02:47,480
Speaker 1: firestorm in Hollywood, and Puck said had Hollywood panicking, and

49
00:02:47,600 --> 00:02:50,200
Speaker 1: tech No Lama said was the end of copyright as

50
00:02:50,240 --> 00:02:52,560
Speaker 1: we know it, and Slat said was a case of

51
00:02:52,600 --> 00:02:56,320
Speaker 1: AI crushing Hollywood as it as we've known it, and

52
00:02:56,320 --> 00:02:57,880
Speaker 1: that I swear to God I wrote that right in

53
00:02:57,919 --> 00:03:00,600
Speaker 1: the script. But I'm keeping going anyway. This thing's completely

54
00:03:00,639 --> 00:03:03,080
Speaker 1: dead a little more than five months after everybody claimed

55
00:03:03,080 --> 00:03:06,680
Speaker 1: it was changing everything. You all were wrong, every single

56
00:03:06,720 --> 00:03:10,240
Speaker 1: goddamn wrong, wrong, wrong, wrong, wrong, wrong, on the wrong,

57
00:03:10,280 --> 00:03:12,639
Speaker 1: a sandwich of wrong that you're shoving up your asshole.

58
00:03:12,680 --> 00:03:16,840
Speaker 1: I'm so fucking tired of this. I cannot express how

59
00:03:16,880 --> 00:03:21,120
Speaker 1: angry I am, because it's almost as if everybody making

60
00:03:21,160 --> 00:03:25,200
Speaker 1: these proclamations, was instinctively printing whatever marketing copy had been

61
00:03:25,240 --> 00:03:29,040
Speaker 1: imagined by the AI labs to promote compute intensive vaporware.

62
00:03:29,200 --> 00:03:31,960
Speaker 1: And absolutely nobody is going to apologize to the people

63
00:03:32,000 --> 00:03:34,560
Speaker 1: working in the entertainment industry for scaring the fuck out

64
00:03:34,600 --> 00:03:37,680
Speaker 1: of them with ghost stories for no apparent reason. Every

65
00:03:37,920 --> 00:03:41,000
Speaker 1: single person who blindly repeated that SAURA existed and was

66
00:03:41,080 --> 00:03:44,400
Speaker 1: changing everything should be forced to apologize to their readers.

67
00:03:45,360 --> 00:03:48,280
Speaker 1: I cannot express the sheer amount of panic that spread

68
00:03:48,280 --> 00:03:51,040
Speaker 1: through every single part of the entertainment industry as a

69
00:03:51,040 --> 00:03:54,600
Speaker 1: result of these specious, poorly founded mythologies spread by people

70
00:03:54,600 --> 00:03:56,840
Speaker 1: that didn't give enough of a shit to understand what

71
00:03:57,040 --> 00:04:00,480
Speaker 1: was actually going on. Saura too, was always an act

72
00:04:00,480 --> 00:04:03,160
Speaker 1: of desperation, an attempt to create a marketing cycle to

73
00:04:03,200 --> 00:04:04,760
Speaker 1: prop up a tool that was burning as much as

74
00:04:04,800 --> 00:04:07,360
Speaker 1: fifty million dollars a day. The most of the mainstream

75
00:04:07,400 --> 00:04:10,720
Speaker 1: media bought into because they believe everything Open AI says

76
00:04:10,880 --> 00:04:13,320
Speaker 1: and are willing to extrapolate the destruction of an entire

77
00:04:13,360 --> 00:04:16,279
Speaker 1: industry from a fucking facade driven by a clammy man

78
00:04:16,320 --> 00:04:19,520
Speaker 1: that lies for a living. I am tired of everybody

79
00:04:19,560 --> 00:04:22,760
Speaker 1: scarfing down the slot from these companies. I'm tired of

80
00:04:22,800 --> 00:04:25,039
Speaker 1: the majority of reporters not doing the hard work to

81
00:04:25,040 --> 00:04:28,800
Speaker 1: actually understand what's going on. And good lord, do I

82
00:04:28,839 --> 00:04:31,960
Speaker 1: not have another example? Well? I do, Eh, you get

83
00:04:31,960 --> 00:04:35,000
Speaker 1: what I mean. I spent the last week or so

84
00:04:35,080 --> 00:04:37,800
Speaker 1: trying to work out how many actual data centers got

85
00:04:37,839 --> 00:04:40,919
Speaker 1: built last year and found that despite somewhere between one

86
00:04:41,000 --> 00:04:44,280
Speaker 1: hundred and ninety two hundred and forty gigawats supposed capacity

87
00:04:44,360 --> 00:04:47,480
Speaker 1: being announced, based on my analysis, only three gigawats of

88
00:04:47,520 --> 00:04:50,120
Speaker 1: IT load, which is the GPUs and the associated hardware

89
00:04:50,120 --> 00:04:53,000
Speaker 1: that's being turned on, got built in America in twenty

90
00:04:53,040 --> 00:04:56,680
Speaker 1: twenty five. To explain how bad this number, three gigawatts

91
00:04:56,720 --> 00:04:59,640
Speaker 1: of critical IT hardware is about ninety billion dollars per

92
00:05:00,240 --> 00:05:02,599
Speaker 1: Darling of TD Cohen and in Video sold about one

93
00:05:02,680 --> 00:05:04,800
Speaker 1: hundred and thirty five billion dollars worth of GPUs and

94
00:05:04,920 --> 00:05:08,960
Speaker 1: hardware to America in its last fiscal year. Based on

95
00:05:09,000 --> 00:05:11,240
Speaker 1: the last quarter of sales, which is in the sixty

96
00:05:11,279 --> 00:05:13,960
Speaker 1: two billion dollar range, sixty nine percent of that nice

97
00:05:14,080 --> 00:05:17,160
Speaker 1: is America. It now takes about six months to install

98
00:05:17,160 --> 00:05:21,520
Speaker 1: a single quarters sales worth of GPUs and associated hardware,

99
00:05:22,400 --> 00:05:24,400
Speaker 1: even if you think AI is the biggest and most

100
00:05:24,440 --> 00:05:27,160
Speaker 1: hugest and most special. Boy, what's the fucking point of

101
00:05:27,200 --> 00:05:30,679
Speaker 1: buying these GPUs two to four years in advance, Because

102
00:05:30,720 --> 00:05:33,599
Speaker 1: even if it takes six it takes six months to

103
00:05:33,640 --> 00:05:36,719
Speaker 1: install a single quarters worth of sales. From what I found,

104
00:05:36,760 --> 00:05:39,680
Speaker 1: it's actually taking a lot longer. In some cases. Jensen

105
00:05:39,760 --> 00:05:42,279
Speaker 1: Hwang is announcing a new GPU every year. By the

106
00:05:42,320 --> 00:05:45,040
Speaker 1: time you install your Blackwells, they'll already have Vera Ruben.

107
00:05:45,240 --> 00:05:47,640
Speaker 1: By the time they have Vera Ruben, they'll have something else.

108
00:05:47,680 --> 00:05:49,840
Speaker 1: I don't know, Daphnie from Scooby Doo. They can have

109
00:05:50,200 --> 00:05:52,680
Speaker 1: no These things are the only cool thing that Nvidia

110
00:05:52,720 --> 00:05:55,880
Speaker 1: does is choosing the names of people anyway. And with

111
00:05:55,920 --> 00:05:59,359
Speaker 1: every new generation of these GPUs comes new power requirements

112
00:05:59,360 --> 00:06:02,320
Speaker 1: and rack side While the first generation of Vera Ruben,

113
00:06:02,360 --> 00:06:04,480
Speaker 1: which is the next gen coming out very soon, will

114
00:06:04,520 --> 00:06:07,559
Speaker 1: fit into the current Oberon in video racks and video

115
00:06:07,640 --> 00:06:11,480
Speaker 1: is already teasing it's Kiber racks, which will require even

116
00:06:11,480 --> 00:06:14,560
Speaker 1: more power, even more liquid cooling, which will mean those

117
00:06:14,880 --> 00:06:19,520
Speaker 1: ooh ew ew, those old, stinky, dirty blackwells, You're gonna

118
00:06:19,520 --> 00:06:21,680
Speaker 1: have to throw them right in the trash. No more

119
00:06:21,720 --> 00:06:25,640
Speaker 1: black well for me. Thanks, they've got Verra Rubin in town. Ew.

120
00:06:25,839 --> 00:06:30,600
Speaker 1: You're gonna make me use blackwell? How gross? Ew Jensen

121
00:06:30,640 --> 00:06:34,320
Speaker 1: put those stinky blackwells in the toilet. Ew, you couldn't

122
00:06:34,320 --> 00:06:37,160
Speaker 1: fit in the far too large I've tried yet. Things

123
00:06:37,240 --> 00:06:39,680
Speaker 1: only get worse when you look at the actual construction

124
00:06:39,800 --> 00:06:43,200
Speaker 1: in process per Siteline Climate. Despite over one hundred and

125
00:06:43,240 --> 00:06:46,839
Speaker 1: ninety gigawatts of global capacity that's supposedly in process, with

126
00:06:46,960 --> 00:06:50,480
Speaker 1: sixteen gigwatts that's meant to come online this year, only

127
00:06:50,640 --> 00:06:54,280
Speaker 1: five gigawatts off, it is actually under construction, and under

128
00:06:54,279 --> 00:06:57,520
Speaker 1: constructure can mean everything from a single steel beam to

129
00:06:57,560 --> 00:07:01,000
Speaker 1: a nearly finished buildout. Walk around London some time. You

130
00:07:01,040 --> 00:07:03,640
Speaker 1: should just just look at building projects in London or

131
00:07:03,680 --> 00:07:06,240
Speaker 1: New York. Look at how quickly or not quickly those

132
00:07:06,279 --> 00:07:10,720
Speaker 1: get built. Oh, I don't know. Go look around Stargate

133
00:07:10,760 --> 00:07:13,400
Speaker 1: and all the data centers there. One of the infrastructure

134
00:07:13,400 --> 00:07:17,000
Speaker 1: people from open AI was boasting of the Wisconsin data

135
00:07:17,000 --> 00:07:19,360
Speaker 1: center the other day and he just had a single

136
00:07:19,400 --> 00:07:21,240
Speaker 1: steel beam. I'm not even fucking with you. It's just

137
00:07:21,280 --> 00:07:23,320
Speaker 1: a one beam. And they were like four guys. There's

138
00:07:23,360 --> 00:07:25,520
Speaker 1: three guys working on the beam and another guy walking

139
00:07:25,600 --> 00:07:29,360
Speaker 1: up to it. I love builders, but anyway, in my premium,

140
00:07:29,520 --> 00:07:32,240
Speaker 1: shit in my premium this week, just hit my microphone,

141
00:07:32,320 --> 00:07:35,480
Speaker 1: just going to keep going. I also dug into most

142
00:07:35,520 --> 00:07:37,720
Speaker 1: of the major data center projects that have been announced,

143
00:07:37,760 --> 00:07:40,040
Speaker 1: and then basically every case found that they had either

144
00:07:40,120 --> 00:07:43,240
Speaker 1: barely broken ground, hadn't actually picked the land yet, or

145
00:07:43,240 --> 00:07:46,160
Speaker 1: it stalled entirely, as was the case with fermi supposed

146
00:07:46,240 --> 00:07:49,200
Speaker 1: eleven giga what data center project in Amarillo, Texas, which

147
00:07:49,280 --> 00:07:53,200
Speaker 1: promptly dismissed workers in February after moving forward without permitting

148
00:07:53,280 --> 00:07:57,400
Speaker 1: or funding necessary to actually build the fucking thing. That's

149
00:07:57,400 --> 00:08:01,480
Speaker 1: a public company by the way, for Merick's gone, Oh god,

150
00:08:02,360 --> 00:08:07,280
Speaker 1: it slightly Avengers of assholes. Despite the many, many statements

151
00:08:07,520 --> 00:08:10,240
Speaker 1: the AI is inevitable or unstoppable, the reality is that

152
00:08:10,280 --> 00:08:12,800
Speaker 1: the things aren't getting built, billions of dollars aren't getting sent,

153
00:08:13,000 --> 00:08:16,360
Speaker 1: industries are left undisrupted, and the only people actually benefiting

154
00:08:16,360 --> 00:08:19,560
Speaker 1: are grifters, venture capitalists and banks providing debt to non

155
00:08:19,560 --> 00:08:23,120
Speaker 1: existent data center projects. Look, I believe by the end

156
00:08:23,160 --> 00:08:25,680
Speaker 1: of this era, to quote just Dario one him and

157
00:08:25,760 --> 00:08:28,680
Speaker 1: Kakashi the lone Gunman, these guys are two of the

158
00:08:28,720 --> 00:08:32,240
Speaker 1: best out there and their pseudonymous accounts. This era will

159
00:08:32,280 --> 00:08:34,920
Speaker 1: be remembered as the largest waste of capital in history.

160
00:08:35,400 --> 00:08:37,800
Speaker 1: Most data centers aren't getting built, and those that are

161
00:08:37,840 --> 00:08:40,320
Speaker 1: won't be ready before twenty to twenty seven at best.

162
00:08:40,920 --> 00:08:43,720
Speaker 1: Most AI startups make piddley amounts of revenue and burn

163
00:08:44,000 --> 00:08:46,480
Speaker 1: hundreds of millions of dollars to make single digit millions

164
00:08:46,480 --> 00:08:49,040
Speaker 1: of dollars of revenue, and constantly piss off their users

165
00:08:49,080 --> 00:08:52,000
Speaker 1: by changing rate limits because every single one is subsidized,

166
00:08:52,200 --> 00:08:54,559
Speaker 1: every single one is spending a couple of dollars or

167
00:08:54,600 --> 00:08:57,000
Speaker 1: more per dollar of revenue, and nothing is changing and

168
00:08:57,040 --> 00:09:00,160
Speaker 1: nothing is improving. Fuck off with your whole inferance this

169
00:09:00,240 --> 00:09:03,360
Speaker 1: profitable thing, by the fucking way, to paraphrase my good

170
00:09:03,400 --> 00:09:05,920
Speaker 1: friend Casey Kagawa, who made this point to me over

171
00:09:06,160 --> 00:09:11,200
Speaker 1: Signal fairly recently. Either inference isn't profitable or API cars

172
00:09:11,240 --> 00:09:14,560
Speaker 1: are just a big they're the biggest ripoff ever that

173
00:09:14,640 --> 00:09:17,600
Speaker 1: they could be charging even less than that. Indeed, if

174
00:09:17,600 --> 00:09:20,880
Speaker 1: inference was so profitable, why are they charging a couple

175
00:09:20,920 --> 00:09:24,400
Speaker 1: of bucks for a million dollars worth of input tokens?

176
00:09:24,400 --> 00:09:27,640
Speaker 1: They should be charging pennies. Why are they charging so

177
00:09:27,720 --> 00:09:31,320
Speaker 1: much of it's so profitable take that eighth of dits anyway,

178
00:09:31,360 --> 00:09:33,080
Speaker 1: And for the most part, the media has failed to

179
00:09:33,120 --> 00:09:35,840
Speaker 1: notice all this stuff. I really must be clear how

180
00:09:36,040 --> 00:09:39,680
Speaker 1: shocking this is. I really have read all over the shop.

181
00:09:39,720 --> 00:09:43,040
Speaker 1: I still see people making the Amazon Web Services lost

182
00:09:43,040 --> 00:09:45,440
Speaker 1: a lot of money. Argument, that's about thirty eight billion

183
00:09:45,520 --> 00:09:48,200
Speaker 1: dollars in over a decade that was spent on Amazon

184
00:09:48,200 --> 00:09:50,880
Speaker 1: Web Services. So walk that one bag. Well, yeah, I

185
00:09:50,880 --> 00:09:53,640
Speaker 1: don't know. I still see the same talking points pedaled

186
00:09:53,720 --> 00:09:56,640
Speaker 1: and it's just ridiculous. Big numbers get quoted without a

187
00:09:56,679 --> 00:10:00,800
Speaker 1: second thought. Unrealistic or fantastical timelines are public strad verbatim.

188
00:10:01,000 --> 00:10:04,800
Speaker 1: I'm talking about that. What ten twenty billion dollar nebuis

189
00:10:04,920 --> 00:10:08,320
Speaker 1: deal with meta that just data center doesn't exist, haven't

190
00:10:08,360 --> 00:10:11,560
Speaker 1: even raised the money. And scrutiny appears, by the way,

191
00:10:12,040 --> 00:10:15,080
Speaker 1: to only be reserved for haters or skeptics, of those

192
00:10:15,080 --> 00:10:18,720
Speaker 1: who say fuck too much on a podcast. Well, the

193
00:10:18,760 --> 00:10:22,360
Speaker 1: scrutiny machine over here is just getting started, and you know,

194
00:10:22,840 --> 00:10:27,719
Speaker 1: you just know that I will be absolutely insufferable when

195
00:10:27,760 --> 00:10:31,920
Speaker 1: I'm finally proven rain. Thank you all for listening. Should

196
00:10:31,960 --> 00:10:34,400
Speaker 1: have some fun stuff next week. I'm still working on

197
00:10:35,160 --> 00:10:42,800
Speaker 1: I love you all,