1
00:00:00,200 --> 00:00:00,680
Speaker 1: What if.

2
00:00:01,000 --> 00:00:06,280
Speaker 2: What if the most analyzed, the most debated tech magnate

3
00:00:06,280 --> 00:00:09,400
in the world isn't actually competing in the industries we

4
00:00:09,519 --> 00:00:11,679
think he is, right, I mean, think about it, while

5
00:00:11,720 --> 00:00:14,880
everyone else, you know, Microsoft, Google, Amazon, they are all

6
00:00:15,000 --> 00:00:20,320
aggressively playing chess on this incredibly crowded board, fighting over

7
00:00:20,399 --> 00:00:23,719
the exact same scarce resources.

8
00:00:23,199 --> 00:00:25,039
Speaker 3: The same chips, the same power grid.

9
00:00:25,000 --> 00:00:28,079
Speaker 1: Exactly, And meanwhile Elon Musk is just off in the

10
00:00:28,079 --> 00:00:30,280
corner building an entirely new board from scratch.

11
00:00:30,519 --> 00:00:33,719
Speaker 3: It really fundamentally changes how you look at the whole

12
00:00:33,719 --> 00:00:36,359
tech landscape. I mean, he isn't trying to out maneuver

13
00:00:36,439 --> 00:00:40,479
the Titans by you know, writing slightly more efficient code.

14
00:00:40,359 --> 00:00:43,039
Speaker 1: Right, or getting a fractionally better enterprise contract.

15
00:00:43,200 --> 00:00:46,960
Speaker 3: Yeah, exactly. He is creating a completely separate gain where

16
00:00:46,960 --> 00:00:50,200
the existing rules of scarcity, software integration, all of that,

17
00:00:50,560 --> 00:00:51,840
it just simply does not apply.

18
00:00:52,000 --> 00:00:55,359
Speaker 1: Well, welcome to thrilling Threads. And if you are joining

19
00:00:55,399 --> 00:00:57,280
us today, we know exactly who you are.

20
00:00:57,399 --> 00:00:57,679
Speaker 3: We do.

21
00:00:57,840 --> 00:00:59,640
Speaker 1: You're the kind of listener who doesn't just want the

22
00:00:59,679 --> 00:01:02,399
morning headlines, right, You want to understand the actual underlying

23
00:01:02,479 --> 00:01:04,239
architecture of where the future is heading.

24
00:01:05,400 --> 00:01:07,879
Speaker 3: So whether you are catching up on the absolute latest

25
00:01:07,879 --> 00:01:11,959
tech trends for a massive strategy meeting, or you're just

26
00:01:12,079 --> 00:01:16,239
insanely curious about how all these seemingly disconnected pieces of

27
00:01:16,280 --> 00:01:19,920
the modern world actually fit together. Consider this your shortcut

28
00:01:20,000 --> 00:01:21,840
to a massive aha moment and.

29
00:01:21,799 --> 00:01:25,599
Speaker 1: The analysis where unpacking today is. Honestly, it's truly sweeping.

30
00:01:25,760 --> 00:01:28,480
It connects dots that most financial analysts and even tech

31
00:01:28,519 --> 00:01:30,680
reporters they don't even realize are on the same piece

32
00:01:30,680 --> 00:01:31,079
of paper.

33
00:01:31,280 --> 00:01:33,400
Speaker 3: Yeah, they're looking at the trees and missing the forest.

34
00:01:34,040 --> 00:01:36,959
So our mission today is to thoroughly unpack this fascinating

35
00:01:37,239 --> 00:01:38,920
comprehensive framework we've.

36
00:01:38,719 --> 00:01:41,680
Speaker 1: Received, and the core thesis is wild, It really is.

37
00:01:42,040 --> 00:01:46,840
Speaker 3: The thesis is that Elon Musk's ventures right Tesla, SpaceX

38
00:01:47,120 --> 00:01:51,480
XAI optimists they are not a portfolio of separate, siloed.

39
00:01:51,040 --> 00:01:53,719
Speaker 1: Companies, which is how everyone treats them, right, but.

40
00:01:53,719 --> 00:01:59,239
Speaker 3: They are actually a single, unified, vertically integrated physical AI system. Yeah,

41
00:01:59,280 --> 00:02:01,920
and we are going to explore every single layer of

42
00:02:01,959 --> 00:02:06,640
this sprawling ecosystem today, from orbiting solar powered data centers

43
00:02:06,640 --> 00:02:10,719
in space to cars parked in your driveway that moonlight

44
00:02:10,759 --> 00:02:13,639
as office workers. It sounds like sci fi but it's happening.

45
00:02:13,759 --> 00:02:16,919
Speaker 1: We want to see exactly how each piece feeds all

46
00:02:16,960 --> 00:02:20,439
the other pieces. So where do we even begin with

47
00:02:20,479 --> 00:02:22,319
an architecture that is this massive?

48
00:02:22,680 --> 00:02:24,759
Speaker 3: Well, I think we have to start with the fundamental

49
00:02:24,759 --> 00:02:28,919
paradigm shift in how artificial intelligence actually interacts with the world. Okay,

50
00:02:29,000 --> 00:02:33,599
the overarching idea here is changing the playing field entirely, right,

51
00:02:33,960 --> 00:02:36,280
And the most vivid illustration of this is something that

52
00:02:36,319 --> 00:02:39,000
was announced fairly recently back on March eleventh.

53
00:02:39,080 --> 00:02:41,199
Speaker 1: Oh right, the joint project.

54
00:02:40,919 --> 00:02:45,280
Speaker 3: Yeah, between Tesla and Xai. Internally, they refer to it

55
00:02:45,280 --> 00:02:49,439
as Macrohard or sometimes Digital Optimists Macrohard.

56
00:02:49,520 --> 00:02:51,280
Speaker 1: I mean, I do love that. It's a very cheeky

57
00:02:51,400 --> 00:02:53,639
literal inversion of Microsoft.

58
00:02:53,240 --> 00:02:53,919
Speaker 3: Very on brand.

59
00:02:54,000 --> 00:02:58,599
Speaker 1: Yeah, totally. So what exactly is this macrohard initiative attempting

60
00:02:58,639 --> 00:02:58,840
to do?

61
00:02:59,199 --> 00:03:01,879
Speaker 3: So It's an AIA agent designed to essentially do the

62
00:03:01,879 --> 00:03:03,759
work of an entire software company.

63
00:03:03,840 --> 00:03:04,240
Speaker 1: Wow.

64
00:03:04,479 --> 00:03:07,039
Speaker 3: But to understand why this is a completely different game,

65
00:03:07,520 --> 00:03:09,719
we have to look at how the rest of the industry,

66
00:03:09,879 --> 00:03:12,879
you know, Microsoft's at Google's anthropic, how they are currently

67
00:03:12,919 --> 00:03:17,599
building their AI agents. When you use say Microsoft Copilot

68
00:03:17,680 --> 00:03:21,560
or Claude to automated task how is that AI actually

69
00:03:21,599 --> 00:03:22,800
talking to the software.

70
00:03:22,960 --> 00:03:26,840
Speaker 1: Well, it's communicating through the back end. It uses APIs

71
00:03:27,000 --> 00:03:31,520
application programming interfaces exactly. It's basically reading the underlying code

72
00:03:31,560 --> 00:03:34,120
of the software to figure out what data is there,

73
00:03:34,159 --> 00:03:37,080
and then it injects its own code to make actions happen.

74
00:03:37,719 --> 00:03:40,400
So the developers of the AI and the developers of

75
00:03:40,400 --> 00:03:43,280
the software have to explicitly build those bridges so the

76
00:03:43,319 --> 00:03:45,080
two systems can even talk to each other.

77
00:03:45,360 --> 00:03:49,039
Speaker 3: And building those bridges is incredibly labor intensive. I mean,

78
00:03:49,080 --> 00:03:52,639
you have to write custom integration code for every single

79
00:03:52,639 --> 00:03:54,639
application you want the AI to interact with.

80
00:03:54,759 --> 00:03:55,520
Speaker 1: It's exhausting.

81
00:03:55,639 --> 00:03:58,319
Speaker 3: Yeah, and if an app updates its UI or you know,

82
00:03:58,479 --> 00:04:02,039
changes its API structure, the bridge breaks. The developers have

83
00:04:02,080 --> 00:04:03,280
to scramble to fix it.

84
00:04:03,280 --> 00:04:05,080
Speaker 1: It's just brittle, very fragile.

85
00:04:05,439 --> 00:04:08,360
Speaker 3: Musk's approach with digital optimists throws all of that out

86
00:04:08,360 --> 00:04:11,159
the window. Instead of building an AI that reads back

87
00:04:11,240 --> 00:04:14,479
end code, he is building an AI that functions exactly

88
00:04:14,520 --> 00:04:15,599
like a human being does.

89
00:04:15,840 --> 00:04:19,600
Speaker 1: Okay, So it's the difference between say, hiring a team

90
00:04:19,639 --> 00:04:23,439
of interpreters to translate every single local dialect in the

91
00:04:23,480 --> 00:04:26,560
world for you, versus just taking a magic pill that

92
00:04:26,680 --> 00:04:29,199
lets you universally understand human intent.

93
00:04:29,560 --> 00:04:30,560
Speaker 3: That's a great way to put it.

94
00:04:30,639 --> 00:04:34,000
Speaker 1: If an API based AI is a train, it's stuck

95
00:04:34,040 --> 00:04:37,360
on predetermined tracks, right yep. It can only go where

96
00:04:37,399 --> 00:04:40,279
the custom integrations and the specific code have already been

97
00:04:40,360 --> 00:04:44,199
laid down fast on its specific route. Sure, but completely

98
00:04:44,279 --> 00:04:46,920
useless if you want to go off course. Exactly so,

99
00:04:47,079 --> 00:04:49,120
how does digital optimists go off road?

100
00:04:49,399 --> 00:04:54,000
Speaker 3: By using pure vision? Digitally speaking, it sees what you see, okay.

101
00:04:54,040 --> 00:04:57,079
It continuously processes the last five seconds of screen video,

102
00:04:57,480 --> 00:05:00,399
taking in the visual layout of a desktop, a web site,

103
00:05:00,519 --> 00:05:03,839
or you know, a proprietary enterprise app, and it executes

104
00:05:03,879 --> 00:05:07,199
actions using a digital mouse and a digital keyboard, exactly

105
00:05:07,240 --> 00:05:11,120
like a human operator would. No custom integrations, no back

106
00:05:11,240 --> 00:05:11,920
end APIs.

107
00:05:12,079 --> 00:05:13,839
Speaker 1: I have to admit I struggle with this a bit

108
00:05:14,079 --> 00:05:16,759
oh so well. I understand the elegance of the idea,

109
00:05:16,839 --> 00:05:20,319
but navigating a chaotic computer desktop purely by looking at

110
00:05:20,399 --> 00:05:22,639
it seems incredibly difficult.

111
00:05:22,199 --> 00:05:23,480
Speaker 3: For a machine it is.

112
00:05:23,560 --> 00:05:28,439
Speaker 1: I know, Ashock Elswami, Tesla's VP of AI, He's quoted

113
00:05:28,480 --> 00:05:31,600
saying it's so obvious you can solve this with cameras.

114
00:05:31,600 --> 00:05:36,000
It's an AI problem. But isn't analyzing live high resolution

115
00:05:36,199 --> 00:05:40,399
video of a computer screen vastly more compute intensive, Oh absolutely,

116
00:05:40,480 --> 00:05:43,360
and frankly more prone to hallucination than just reading a

117
00:05:43,399 --> 00:05:47,879
clean text feed from a structured API. Like, why take

118
00:05:47,920 --> 00:05:48,879
the harder road.

119
00:05:49,079 --> 00:05:53,000
Speaker 3: Because if you solve the harder road, you achieve absolute universality,

120
00:05:53,160 --> 00:05:55,839
meaning what exactly think about it? The digital world, our

121
00:05:55,839 --> 00:05:59,240
computer interfaces, our software websites, it was all built for

122
00:05:59,759 --> 00:06:03,560
qu eyes in human hands. We didn't build graphical user

123
00:06:03,600 --> 00:06:06,360
interfaces for machines to read in the background. We built

124
00:06:06,360 --> 00:06:09,519
them with buttons drop down. Men use visual cues that

125
00:06:09,560 --> 00:06:11,199
are meant for biological processing.

126
00:06:11,319 --> 00:06:11,839
Speaker 1: That makes sense.

127
00:06:11,879 --> 00:06:14,480
Speaker 3: So if you TACAI to look at a screen and understand, ah,

128
00:06:14,560 --> 00:06:17,319
that red rectangles is a submit button and that blinking

129
00:06:17,360 --> 00:06:20,120
line is the text field, then that AI can suddenly

130
00:06:20,199 --> 00:06:23,800
use every single piece of software ever created, past, present

131
00:06:23,839 --> 00:06:24,680
and future.

132
00:06:24,720 --> 00:06:27,839
Speaker 1: Without writing a single line of custom integration code. So

133
00:06:27,959 --> 00:06:30,519
if a company updates their app tomorrow and moves the

134
00:06:30,519 --> 00:06:33,079
button to the left side of the screen, the AI

135
00:06:33,439 --> 00:06:35,759
just sees it moved and clicks it anyway, just like

136
00:06:35,759 --> 00:06:36,319
a human would.

137
00:06:36,399 --> 00:06:41,639
Speaker 3: Precisely you bypass the entire API bottleneck. Now regarding your

138
00:06:41,639 --> 00:06:44,879
point about compute intensity, you are entirely correct. Yeah, videos

139
00:06:44,920 --> 00:06:47,759
have Analyzing live video is immensely heavy, and this is

140
00:06:47,759 --> 00:06:50,720
where the vertical integration really starts to show its teeth.

141
00:06:50,759 --> 00:06:51,560
Speaker 1: Okay, wait on me.

142
00:06:51,879 --> 00:06:55,920
Speaker 3: Digital Optimus isn't running on some generic cloud server. It

143
00:06:56,000 --> 00:06:59,720
runs on Tesla's in house AI four hardware, which they

144
00:06:59,720 --> 00:07:02,519
may facture themselves for only about six hundred and fifty

145
00:07:02,560 --> 00:07:03,439
dollars per unit.

146
00:07:03,480 --> 00:07:04,199
Speaker 1: Wow that cheap.

147
00:07:04,480 --> 00:07:08,120
Speaker 3: Yeah. It processes the visual input locally right there on

148
00:07:08,160 --> 00:07:11,800
the chip and only calls out to Xai's massive GROC

149
00:07:11,839 --> 00:07:15,160
cloud when it needs heavy strategic.

150
00:07:14,639 --> 00:07:17,720
Speaker 1: Reasoning what they call system two thinking exactly, So, the

151
00:07:17,759 --> 00:07:20,920
local chip handles the reflexes like where the mouse is,

152
00:07:21,240 --> 00:07:24,079
what the icon looks like, and GROC handles the logic

153
00:07:24,120 --> 00:07:25,519
of what the task actually is.

154
00:07:25,639 --> 00:07:26,160
Speaker 3: You got it.

155
00:07:26,439 --> 00:07:30,000
Speaker 1: But the hardware aspect leads to an extrapolation that honestly,

156
00:07:30,040 --> 00:07:33,319
it sounds like science fiction. There is a stated vision

157
00:07:33,600 --> 00:07:37,040
that your personal Tesla, while it is just sitting parked,

158
00:07:37,360 --> 00:07:39,399
can actually do this office work for you.

159
00:07:39,560 --> 00:07:42,160
Speaker 3: Right if you follow the logic of the hardware, it

160
00:07:42,199 --> 00:07:46,319
makes perfect sense. How so, every single Tesla equipped with

161
00:07:46,360 --> 00:07:51,279
this AI four computer is essentially a highly capable localized

162
00:07:51,360 --> 00:07:54,920
inference node. Okay, when you aren't actively driving it, that

163
00:07:54,959 --> 00:07:57,519
computer is just sitting idle in your garage or parking

164
00:07:57,600 --> 00:08:00,879
lot doing nothing, doing absolutely nothing. So the vision is

165
00:08:00,920 --> 00:08:04,120
to turn those millions of parked cars into a massive

166
00:08:04,240 --> 00:08:05,759
distributed compute network.

167
00:08:05,879 --> 00:08:08,160
Speaker 1: So wait, while I'm sleeping, the computer inside my car

168
00:08:08,279 --> 00:08:11,639
boots up digital optimists and just starts churning through digital

169
00:08:11,680 --> 00:08:12,399
office tasks.

170
00:08:12,480 --> 00:08:15,920
Speaker 3: Yes, and think about the contrast here. Everyone else in

171
00:08:15,959 --> 00:08:18,560
the tech industry right now is fighting tooth and nail

172
00:08:18,680 --> 00:08:23,399
to build massive centralized server farms for cloud compute.

173
00:08:23,160 --> 00:08:24,279
Speaker 1: Right data centers.

174
00:08:24,360 --> 00:08:28,199
Speaker 3: They are begging local governments for power permits. Meanwhile, Tesla

175
00:08:28,240 --> 00:08:31,319
already has millions of these highly capable compute nodes deployed

176
00:08:31,360 --> 00:08:33,519
all over the world. Yeah, and the best part, they

177
00:08:33,519 --> 00:08:34,720
were paid for by the consumer.

178
00:08:34,799 --> 00:08:36,399
Speaker 1: Oh wow, that is wild.

179
00:08:36,679 --> 00:08:40,279
Speaker 3: And to power them for heavy compute tasks, they are

180
00:08:40,279 --> 00:08:45,360
deploying millions of dedicated units at supercharger stations. Those stations

181
00:08:45,440 --> 00:08:49,720
represent roughly seven gigawatts of available deployed power.

182
00:08:49,919 --> 00:08:51,159
Speaker 1: Seven gigawats.

183
00:08:51,240 --> 00:08:51,519
Speaker 3: Yeah.

184
00:08:51,639 --> 00:08:54,600
Speaker 1: To put that in perspective for you listening, one gigawatt

185
00:08:54,679 --> 00:08:58,399
can power a mid sized city that is a staggering

186
00:08:58,440 --> 00:09:00,840
amount of energy infrastructure just sitting there.

187
00:09:01,000 --> 00:09:01,600
Speaker 3: It really is.

188
00:09:01,919 --> 00:09:05,320
Speaker 1: So instead of building a multi billion dollar data center

189
00:09:05,360 --> 00:09:08,120
from scratch in the middle of a desert somewhere, he

190
00:09:08,279 --> 00:09:11,039
just taps into the computers that people have already bought,

191
00:09:11,159 --> 00:09:13,559
which are plugged into a power network he already owned.

192
00:09:13,679 --> 00:09:15,919
Speaker 3: It's a completely decentralized supercomputer.

193
00:09:16,200 --> 00:09:16,480
Speaker 2: Yeah.

194
00:09:16,519 --> 00:09:18,879
Speaker 3: But we have to ask the critical question here, which is,

195
00:09:19,279 --> 00:09:22,159
how does a car company get so absurdly good at

196
00:09:22,200 --> 00:09:26,000
computer vision that they can confidently apply it to complex

197
00:09:26,080 --> 00:09:27,039
office software?

198
00:09:27,360 --> 00:09:30,720
Speaker 1: Right, Because navigating a cluttered spreadsheet is hard, but navigating

199
00:09:30,720 --> 00:09:34,120
a cluttered intersection in Mumbai is a nightmare exactly. And

200
00:09:34,120 --> 00:09:37,080
that's the secret, isn't it. They've been training this visual

201
00:09:37,080 --> 00:09:40,679
intelligence on the chaotic physical streets for years.

202
00:09:40,799 --> 00:09:43,679
Speaker 3: Which brings us to the foundation of this entire ecosystem.

203
00:09:44,120 --> 00:09:47,200
We have to completely reframe how we view Tesla as

204
00:09:47,200 --> 00:09:47,679
a company.

205
00:09:47,919 --> 00:09:49,600
Speaker 1: Yeah. I think people still get this wrong.

206
00:09:49,960 --> 00:09:54,960
Speaker 3: Most financial analysts still value Tesla primarily as an automotive

207
00:09:54,960 --> 00:09:58,159
company that happens to have some cool AI features tacked

208
00:09:58,159 --> 00:10:01,960
onto the dashboard hardware company, right, But the reality is

209
00:10:02,240 --> 00:10:06,440
the exact inverse. Tesla is an AI company that happens

210
00:10:06,440 --> 00:10:08,840
to sell cars to fund its data collection.

211
00:10:09,320 --> 00:10:11,519
Speaker 1: So if they are an AI company, what is their

212
00:10:11,519 --> 00:10:12,639
proprietary data?

213
00:10:12,759 --> 00:10:13,080
Speaker 3: Video?

214
00:10:13,360 --> 00:10:16,519
Speaker 1: Video? Right, every single tesla on the road is equipped

215
00:10:16,559 --> 00:10:20,720
with eight cameras capturing continuous high definition video.

216
00:10:20,519 --> 00:10:22,600
Speaker 3: Of the real world, every single one, every.

217
00:10:22,480 --> 00:10:25,120
Speaker 1: Time someone drives to the grocery store, those cameras are

218
00:10:25,120 --> 00:10:29,080
feeding the largest real world AI training data set on Earth, and.

219
00:10:29,039 --> 00:10:32,639
Speaker 3: The scale of this data set is frankly difficult to comprehend.

220
00:10:33,240 --> 00:10:36,600
The full self driving or FSD fleet has logged over

221
00:10:36,679 --> 00:10:39,200
eight point four billion cumulative miles.

222
00:10:39,360 --> 00:10:42,240
Speaker 1: Eight point four billion miles, but that's sa gift. I

223
00:10:42,279 --> 00:10:45,360
mean the Earth is roughly twenty four thousand miles in circumference.

224
00:10:45,440 --> 00:10:47,919
We are talking about driving around the entire planet hundreds

225
00:10:47,960 --> 00:10:50,600
of thousands of times. It's not just highway driving in

226
00:10:50,720 --> 00:10:55,000
sunny California either. It's snowstorms in Minnesota, deer jumping out

227
00:10:55,039 --> 00:10:58,759
in rural Texas, chaotic pedestrian traffic in Manhattan.

228
00:10:59,120 --> 00:11:02,639
Speaker 3: And the growth rate is purely exponential. They went from

229
00:11:02,720 --> 00:11:05,720
six million miles in twenty twenty one to four point

230
00:11:05,799 --> 00:11:08,600
twenty five billion by the start of twenty twenty five. Okay,

231
00:11:09,240 --> 00:11:11,559
then in just the first fifty days of twenty twenty

232
00:11:11,600 --> 00:11:14,480
six alone, they added another one billion miles.

233
00:11:14,720 --> 00:11:18,080
Speaker 1: A billion miles in fifty days. Yeah, that means they

234
00:11:18,080 --> 00:11:21,639
are capturing every conceivable edge case out there. Think about

235
00:11:21,639 --> 00:11:23,159
the weirdest things you've seen while driving.

236
00:11:23,200 --> 00:11:24,639
Speaker 3: Oh, there's always something like.

237
00:11:24,559 --> 00:11:27,080
Speaker 1: A person walking a dog while riding a unicycle in

238
00:11:27,120 --> 00:11:30,440
a chicken suit. A Tesla camera has almost certainly seen it,

239
00:11:30,679 --> 00:11:33,200
categorized it, and trained the neural network on how to

240
00:11:33,240 --> 00:11:33,840
react to it.

241
00:11:34,039 --> 00:11:36,879
Speaker 3: That volume of edge case data is the ultimate mote.

242
00:11:37,399 --> 00:11:41,720
Achieving safe unsupervised self driving at scale requires an estimated

243
00:11:41,799 --> 00:11:45,559
ten billion mile benchmark just to iron out those statistical anomalies.

244
00:11:45,919 --> 00:11:48,480
At their current pace, they will likely cross that threshold

245
00:11:48,480 --> 00:11:49,159
this year now.

246
00:11:49,279 --> 00:11:52,639
Speaker 1: Waimo cruz zekes these companies are also trying to build

247
00:11:52,639 --> 00:11:55,120
self driving cars. They are, but they are playing a

248
00:11:55,159 --> 00:11:57,600
fundamentally different game, right. They aren't trying to solve the

249
00:11:57,679 --> 00:11:59,799
vision problem the way Tesla is not at all.

250
00:12:00,120 --> 00:12:02,960
Speaker 3: They are relying on a completely different architecture. Yeah, the

251
00:12:03,000 --> 00:12:06,879
competitors use expensive light OAR sensors which shoot thousands of

252
00:12:06,919 --> 00:12:10,200
lasers out to create exact three D maps of the environment.

253
00:12:10,679 --> 00:12:13,559
Speaker 1: So they use lasers, not just cameras, right, and.

254
00:12:13,559 --> 00:12:17,039
Speaker 3: They combine those lasers with heavily pre mapped environments and

255
00:12:17,200 --> 00:12:21,519
rigid geofence zones. They are solving autonomous driving city by city,

256
00:12:21,759 --> 00:12:23,120
map by map, which.

257
00:12:22,960 --> 00:12:25,679
Speaker 1: Goes back to the train analogy from earlier exactly a

258
00:12:25,759 --> 00:12:29,159
WAYML car is a train on highly specific tracks. As

259
00:12:29,200 --> 00:12:32,080
long as the car stays within the perfectly mapped, perfectly

260
00:12:32,159 --> 00:12:36,399
laser scan zone of downtown Phoenix or San Francisco, it operates.

261
00:12:36,039 --> 00:12:38,399
Speaker 3: Beautifully, flawlessly usually, But.

262
00:12:38,360 --> 00:12:41,840
Speaker 1: If a construction crew alters an intersection overnight, or if

263
00:12:41,879 --> 00:12:43,440
you ask it to drive on a dirt road in

264
00:12:43,480 --> 00:12:46,759
Montana that hasn't been pre mapped, the system encounters the scenario,

265
00:12:46,879 --> 00:12:48,799
it doesn't understand, and it just stops.

266
00:12:48,919 --> 00:12:53,120
Speaker 3: It's a brittle system. Tesla conversely, relies on pure vision,

267
00:12:53,480 --> 00:12:57,240
no expensive laser sensors, no pre map geofences, just cameras

268
00:12:57,240 --> 00:12:58,159
and a neural network.

269
00:12:58,360 --> 00:13:02,080
Speaker 1: So WEAIMO is trying to solve the specific task of driving, yes,

270
00:13:02,120 --> 00:13:05,200
but Tesla is trying to solve the universal challenge of

271
00:13:05,320 --> 00:13:07,360
vision and general spatial intelligence.

272
00:13:07,480 --> 00:13:11,480
Speaker 3: Exactly because if you saw real world navigation using only cameras,

273
00:13:11,960 --> 00:13:14,360
you haven't just built a chauffeur. What have you built?

274
00:13:14,519 --> 00:13:18,879
You've built an artificial intelligence that inherently understands physics, depth,

275
00:13:19,120 --> 00:13:22,960
object permanence, and human behavior. Wow, it knows that a

276
00:13:23,000 --> 00:13:25,759
ball rolling into the street means a child might be

277
00:13:25,840 --> 00:13:26,360
chasing it.

278
00:13:26,440 --> 00:13:30,159
Speaker 1: But the industry debate around this is fierce, very fierce.

279
00:13:30,279 --> 00:13:32,840
And I'll play devil's advocate here because the criticism of

280
00:13:32,879 --> 00:13:34,159
pure vision is heavy.

281
00:13:34,399 --> 00:13:34,840
Speaker 3: Go for it.

282
00:13:35,120 --> 00:13:40,200
Speaker 1: Ladar provides exact mathematical distance. It shoots a laser, it

283
00:13:40,240 --> 00:13:43,200
bounces back, The computer calculates the time of flight, and

284
00:13:43,240 --> 00:13:46,159
it knows exactly to the millimeter how far away that

285
00:13:46,200 --> 00:13:47,039
pedestrian is.

286
00:13:47,120 --> 00:13:47,679
Speaker 3: That's true.

287
00:13:47,799 --> 00:13:52,000
Speaker 1: Pure vision, however, relies on inference. The computer has to

288
00:13:52,039 --> 00:13:55,480
guess the distance based on flat two D pixels. It

289
00:13:55,519 --> 00:13:58,480
has to use parallax and temporal memory to infer three

290
00:13:58,559 --> 00:14:01,600
D depth. When human lives are at stake at seventy

291
00:14:01,639 --> 00:14:06,279
miles per hour, isn't relying on inference incredibly risky compared

292
00:14:06,279 --> 00:14:07,639
to the certainty of a laser.

293
00:14:08,000 --> 00:14:11,840
Speaker 3: It's the central tension of the autonomous vehicle race. The

294
00:14:11,879 --> 00:14:17,320
defense of pure vision anchors back to biological equivalents. Biological equivalent, Yes,

295
00:14:17,559 --> 00:14:20,879
the physical world was built for humans, and humans navigate

296
00:14:20,919 --> 00:14:23,159
it without lasers shooting out of our foreheads.

297
00:14:23,279 --> 00:14:24,159
Speaker 1: That's a good point.

298
00:14:24,320 --> 00:14:27,759
Speaker 3: We drive using two optical sensors, our eyes and a

299
00:14:27,799 --> 00:14:32,159
biological neural network behind them, our brain. We infer depth, speed,

300
00:14:32,159 --> 00:14:35,039
and trajectory purely from visual data, and we do it

301
00:14:35,080 --> 00:14:35,960
remarkably well.

302
00:14:36,000 --> 00:14:38,159
Speaker 1: We do most of the time anyway, right.

303
00:14:38,200 --> 00:14:40,879
Speaker 3: So the argument is that if a biological neural network

304
00:14:40,919 --> 00:14:43,840
can do it safely, a digital neural network trained on

305
00:14:43,919 --> 00:14:46,919
ten billion miles of diverse global data can do it

306
00:14:46,960 --> 00:14:48,559
even safer and without.

307
00:14:48,240 --> 00:14:49,840
Speaker 1: The crazy hardware.

308
00:14:49,399 --> 00:14:53,240
Speaker 3: Costs, exactly without the crushing hardware costs and geofencing limitations

309
00:14:53,240 --> 00:14:53,720
of light ar.

310
00:14:54,240 --> 00:14:57,240
Speaker 1: It's the ultimate bet on software over hardware. And if

311
00:14:57,279 --> 00:15:00,360
that bet on purevision pays off, the economic implicate cations

312
00:15:00,360 --> 00:15:03,759
are just wild. They're staggering, because if you have millions

313
00:15:03,799 --> 00:15:06,720
of data collecting AI robots on wheels rolling around the

314
00:15:06,759 --> 00:15:08,639
world and suddenly you prove they don't need a human

315
00:15:08,679 --> 00:15:12,720
driver at all, the entire economic model of car ownership

316
00:15:12,840 --> 00:15:14,879
flips completely upside down.

317
00:15:14,960 --> 00:15:18,559
Speaker 3: Which brings us to the concept of transportation as a service. Yeah,

318
00:15:18,600 --> 00:15:21,399
and what is widely misunderstood about the robotaxi model.

319
00:15:21,480 --> 00:15:24,759
Speaker 1: Yeah, let's talk about robotaxi. Whenever the media covers it,

320
00:15:24,799 --> 00:15:29,240
the framing is almost always Tesla versus Uber. Always it's

321
00:15:29,279 --> 00:15:32,200
portrayed as a battle of ride hailing apps competing for

322
00:15:32,279 --> 00:15:36,279
market share. But that framing completely misses the forest for

323
00:15:36,399 --> 00:15:37,480
the trees, doesn't it.

324
00:15:37,480 --> 00:15:41,039
Speaker 3: It really does. It ignores the fundamental inefficiency of the

325
00:15:41,039 --> 00:15:45,159
modern automobile. Cars as they exist today are arguably the

326
00:15:45,159 --> 00:15:47,159
most inefficient assets in our society.

327
00:15:47,200 --> 00:15:49,360
Speaker 1: Inefficient how well the average.

328
00:15:48,960 --> 00:15:51,679
Speaker 3: Car sits idle for roughly twenty two hours a day.

329
00:15:51,799 --> 00:15:52,559
Speaker 1: Twenty two hours.

330
00:15:52,639 --> 00:15:56,279
Speaker 3: Yeah, it is a rapidly depreciating piece of heavy machinery

331
00:15:56,399 --> 00:15:58,879
that just sits in a driveway or a parking garage,

332
00:15:59,120 --> 00:16:01,720
bleeding money in the form of insurance, maintenance, and loan

333
00:16:01,759 --> 00:16:03,200
payments every single month.

334
00:16:03,279 --> 00:16:05,879
Speaker 1: It's a very expensive two ton metal brick taking a

335
00:16:05,960 --> 00:16:08,200
valuable real estate ninety percent of the time.

336
00:16:08,399 --> 00:16:12,279
Speaker 3: Precisely, the Robotaxi plan aims to take the seven million

337
00:16:12,279 --> 00:16:15,679
tesla's already on the road and transform them from depreciating

338
00:16:15,759 --> 00:16:20,080
liabilities into revenue generating assets. Okay, how when you aren't

339
00:16:20,120 --> 00:16:22,279
using your vehicle, you tap a button on your phone

340
00:16:22,320 --> 00:16:25,480
and the car wakes up, unplugs itself, and drives off

341
00:16:25,519 --> 00:16:27,559
to provide rise to other people on the network.

342
00:16:28,120 --> 00:16:31,559
Speaker 1: The proposed economics of this are fascinating. They are the

343
00:16:31,600 --> 00:16:33,679
owner of the car keeps seventy five percent of the

344
00:16:33,720 --> 00:16:37,720
revenue generated, while the network Tesla takes twenty five percent

345
00:16:37,759 --> 00:16:41,480
to cover the routing, software, network management, and insurance. The

346
00:16:41,639 --> 00:16:44,360
estimates suggests an owner could earn anywhere from ten thousand

347
00:16:44,399 --> 00:16:46,679
dollars to fifty thousand dollars a year, depending on their

348
00:16:46,720 --> 00:16:48,600
lotation and how often they let the car work.

349
00:16:49,200 --> 00:16:52,679
Speaker 3: Consider the financial relief that offers the average consumer. If

350
00:16:52,679 --> 00:16:55,200
your car is generating even ten thousand dollars a year

351
00:16:55,240 --> 00:16:57,399
while you are asleep or sitting at your office desk,

352
00:16:57,759 --> 00:17:00,519
the vehicle is literally paying for its own fire, financing,

353
00:17:00,559 --> 00:17:01,279
and insurance.

354
00:17:01,720 --> 00:17:04,920
Speaker 1: It transitions from an expense to an income stream exactly.

355
00:17:05,119 --> 00:17:08,400
Speaker 3: And from a corporate perspective, the scale is an absolute

356
00:17:08,519 --> 00:17:11,400
cheat code, a cheat code. Think about it. Even if

357
00:17:11,440 --> 00:17:15,400
only ten percent of current Tesla owners opt into this network,

358
00:17:15,680 --> 00:17:18,960
just a ten percent adoption rate that immediately yields a

359
00:17:18,960 --> 00:17:22,880
fleet of seven hundred thousand autonomous vehicles operating globally.

360
00:17:23,039 --> 00:17:25,799
Speaker 1: And what is the capital expenditure to acquire those seven

361
00:17:25,880 --> 00:17:29,880
hundred thousand vehicles zero zero. The consumer already bought the

362
00:17:29,920 --> 00:17:33,240
hardware Uber and weimo have to go out raise capital,

363
00:17:33,359 --> 00:17:37,960
purchase expensive specialized vehicles, outfit them with tens of thousands

364
00:17:38,000 --> 00:17:40,599
of dollars of light our equipment, and deploy them city

365
00:17:40,640 --> 00:17:44,400
by city in a highly capital intensive rollout. It's painfully slow,

366
00:17:44,559 --> 00:17:47,839
but Tesla simply pushes an over the air software update

367
00:17:48,160 --> 00:17:51,079
to cars that are already sitting in people's driveways. It

368
00:17:51,200 --> 00:17:53,880
is an instantaneous, distributed deployment.

369
00:17:54,160 --> 00:17:56,599
Speaker 3: The Airbnb model taken to its absolute extreme.

370
00:17:56,720 --> 00:17:59,200
Speaker 1: Oh totally, Airbnb let you monetize your spare bedroom, but

371
00:17:59,240 --> 00:18:00,680
you still had to watch the sheets and you know,

372
00:18:00,720 --> 00:18:04,559
interact with the guests. With this, your spare bedroom unplugs itself,

373
00:18:04,680 --> 00:18:07,240
leaves your house, generates cash flow all night, and parks

374
00:18:07,240 --> 00:18:09,480
itself back in your driveway before you even wake up.

375
00:18:09,559 --> 00:18:13,480
Speaker 3: This completely alters the trajectory of personal transportation. It's not

376
00:18:13,559 --> 00:18:16,960
just about undercutting ubers pricing, though. The forthcoming Cybercab is

377
00:18:17,000 --> 00:18:18,119
designed to do exactly that.

378
00:18:18,480 --> 00:18:19,720
Speaker 1: Oh the Cybercab.

379
00:18:19,799 --> 00:18:23,279
Speaker 3: Yeah, the Cybercab, which is a purpose built robotaxi entering

380
00:18:23,319 --> 00:18:26,799
production next month it will cost under thirty thousand dollars

381
00:18:26,880 --> 00:18:30,640
to manufacture. It won't even have a steering wheel or pedals.

382
00:18:30,359 --> 00:18:32,640
Speaker 1: Wait, no pedals at all, none, and.

383
00:18:32,680 --> 00:18:35,440
Speaker 3: Operating costs are projected it under twenty cents per mile.

384
00:18:35,319 --> 00:18:37,279
Speaker 1: Under twenty cents a mile, That is up to ninety

385
00:18:37,279 --> 00:18:40,960
percent cheaper than a current uber or lyft ride. If

386
00:18:41,039 --> 00:18:45,559
point to point transportation becomes that radically cheap and instantly accessible,

387
00:18:46,279 --> 00:18:48,559
why would anyone living in a suburb or a city

388
00:18:48,640 --> 00:18:49,880
ever buy a personal car?

389
00:18:49,960 --> 00:18:53,960
Speaker 3: Again, that is the ultimate disruption. Transportation as a service

390
00:18:54,279 --> 00:18:59,400
makes personal car ownership optional, even financially irrational for many people. Absolutely,

391
00:18:59,519 --> 00:19:03,000
you'll limit the need for personal auto insurance, routine maintenance,

392
00:19:03,200 --> 00:19:06,480
and home parking spaces. Yeah, the ripple effects through urban

393
00:19:06,519 --> 00:19:10,599
planning alone, like turning massive concrete parking lots into housing

394
00:19:10,680 --> 00:19:12,640
or green spaces, are staggering.

395
00:19:12,759 --> 00:19:15,440
Speaker 1: It's a total re architecting of modern society. It is,

396
00:19:15,720 --> 00:19:19,960
and yet incredibly solving the autonomous car is only the

397
00:19:20,000 --> 00:19:21,240
beginning of the vision.

398
00:19:21,240 --> 00:19:21,920
Speaker 3: Just the beginning.

399
00:19:22,160 --> 00:19:25,440
Speaker 1: Because if you have successfully developed an AI brain that

400
00:19:25,559 --> 00:19:29,240
is sophisticated enough to navigate the unpredictable chaos of a

401
00:19:29,279 --> 00:19:33,200
busy city street and versatile enough to manage a digital

402
00:19:33,240 --> 00:19:36,920
computer desktop. What happens when you take that exact same

403
00:19:37,000 --> 00:19:39,279
generalized intelligence and give it arms and legs.

404
00:19:39,279 --> 00:19:43,319
Speaker 3: You get Optimists and the true manifestation of a unified

405
00:19:43,400 --> 00:19:45,039
physical AI ecosystem.

406
00:19:45,160 --> 00:19:47,799
Speaker 1: When most people think of humanoid robots right their minds

407
00:19:47,799 --> 00:19:50,519
immediately go to Boston Dynamics. Sure, we've all seen the

408
00:19:50,599 --> 00:19:54,000
viral videos of the Atlas robot doing backflips, parkour and

409
00:19:54,119 --> 00:19:57,920
dancing to pop songs. It's brilliant engineering, very impressive hardware.

410
00:19:58,160 --> 00:20:01,559
But comparing Optimists to Boston Dynamics is fundamentally flawed.

411
00:20:01,240 --> 00:20:05,839
Speaker 3: Isn't it entirely Yeah, Boston Dynamics build standalone, highly specialized

412
00:20:05,920 --> 00:20:10,240
robotic products. They are incredible pieces of megatronics, but they

413
00:20:10,279 --> 00:20:15,359
are bespoke machines built for specific physical tasks. Optimism is

414
00:20:15,400 --> 00:20:18,039
not a separate product, and the traditional sense, it is

415
00:20:18,119 --> 00:20:21,279
quite literally the physical manifestation of the car's AI.

416
00:20:21,720 --> 00:20:24,440
Speaker 1: So the car and the robot are sharing a brain.

417
00:20:24,759 --> 00:20:27,880
Speaker 3: They were running the exact same end to end neural network.

418
00:20:28,160 --> 00:20:30,960
They use a single neural world simulator for training.

419
00:20:31,119 --> 00:20:31,640
Speaker 1: Oh wow.

420
00:20:31,799 --> 00:20:34,880
Speaker 3: The underlying architecture that takes visual input from a camera

421
00:20:35,039 --> 00:20:37,960
on a car's bumper and translates it into steering and

422
00:20:37,960 --> 00:20:41,480
breaking commands. Is simply adapted to take visual input from

423
00:20:41,519 --> 00:20:45,920
the cameras in the robot's head and translate it into arm, hand,

424
00:20:46,000 --> 00:20:46,960
and leg movements.

425
00:20:47,240 --> 00:20:50,519
Speaker 1: That is a staggering concept. That means the eight point

426
00:20:50,559 --> 00:20:53,839
four billion miles of driving data we discussed earlier that

427
00:20:54,039 --> 00:20:56,880
massive mode of data isn't just making the cars better

428
00:20:56,920 --> 00:21:00,640
at driving. It's actively making the humanoid robots it's smarter.

429
00:21:01,039 --> 00:21:04,440
Speaker 3: Yes, because the fundamental challenge of both driving and robotics

430
00:21:04,640 --> 00:21:08,200
isn't the specific task itself. The challenge is understanding the

431
00:21:08,200 --> 00:21:11,759
physical world. Right by navigating billions of miles of roads,

432
00:21:12,079 --> 00:21:15,279
the AI has learned the universal rules of physics. It

433
00:21:15,400 --> 00:21:20,720
understands gravity, momentum, depth of perception, how solid objects interact,

434
00:21:20,920 --> 00:21:23,119
and how humans move unpredictably through space.

435
00:21:23,359 --> 00:21:25,880
Speaker 1: I understand the theory, but help me bridge the practical

436
00:21:25,880 --> 00:21:28,799
gap here. If the AI learns how to successfully identify

437
00:21:28,839 --> 00:21:31,960
and stop for a pedestrian at a busy crosswalk in Chicago,

438
00:21:32,559 --> 00:21:36,759
how does that specific automotive knowledge translate to helping an

439
00:21:36,759 --> 00:21:40,519
optimist robot fold laundry in my living room or assemble

440
00:21:40,559 --> 00:21:42,480
a battery pack in a factory in Texas.

441
00:21:42,640 --> 00:21:46,759
Speaker 3: It comes down to generalize spatial awareness. When the car

442
00:21:46,839 --> 00:21:49,519
learns to stop for a pedestrian, it isn't just learning

443
00:21:49,559 --> 00:21:52,240
a rigid rule that says stop at white lines, okay.

444
00:21:52,680 --> 00:21:55,839
It is learning complex spatial reasoning. It is learning to

445
00:21:55,920 --> 00:21:59,119
identify a three D object moving through a three D environment,

446
00:21:59,480 --> 00:22:03,240
predict its physical trajectory, and manipulate its own physical form

447
00:22:03,599 --> 00:22:06,640
the car to interact safely with that object. All Right

448
00:22:06,880 --> 00:22:09,519
said that underlying spatial reasoning is identical to the reasoning

449
00:22:09,599 --> 00:22:12,480
required for a robot hand to reach out graph the

450
00:22:12,519 --> 00:22:16,200
delicate piece of fabric, understand the fabric's physical properties, and

451
00:22:16,279 --> 00:22:20,279
fold it without tearing it. It's all just manipulating physical

452
00:22:20,279 --> 00:22:23,599
matter in three dimensional space based on visual input.

453
00:22:23,799 --> 00:22:26,640
Speaker 1: And because they are sharing this unified neural brain, it

454
00:22:26,720 --> 00:22:28,960
creates an unstoppable flywheel effect.

455
00:22:29,119 --> 00:22:32,160
Speaker 3: The flywheel is the core driver of evaluation model. Here,

456
00:22:32,200 --> 00:22:32,920
let's trace the loop.

457
00:22:33,039 --> 00:22:33,480
Speaker 1: Let's do it.

458
00:22:33,559 --> 00:22:36,400
Speaker 3: You have millions of cars on the road collecting edge

459
00:22:36,440 --> 00:22:40,640
case data. More cars equals more very data. More data

460
00:22:40,759 --> 00:22:44,759
trains a vastly smarter generalized AI. Because they share a brain,

461
00:22:45,279 --> 00:22:50,119
that smarter AI instantly creates smarter more capable optimist robots.

462
00:22:51,160 --> 00:22:54,240
Those smarter robots are then deployed directly into tessel zone

463
00:22:54,240 --> 00:22:58,200
factories to manufacture things cheaper, safer, and faster.

464
00:22:58,400 --> 00:23:01,000
Speaker 1: And what are those robots manufacts uring. They are building

465
00:23:01,000 --> 00:23:03,920
more cars and more robots exactly, which then deploy into

466
00:23:03,960 --> 00:23:06,680
the world to collect even more data, spinning the flywheel

467
00:23:06,759 --> 00:23:09,960
faster and faster, driving costs down at every revolution.

468
00:23:10,319 --> 00:23:13,599
Speaker 3: The scale of this robotic deployment is already underway. Over

469
00:23:13,680 --> 00:23:17,400
one thousand Gen three optimist robots have already been deployed internally.

470
00:23:17,519 --> 00:23:18,000
Speaker 1: Wow.

471
00:23:18,240 --> 00:23:21,480
Speaker 3: Furthermore, they are currently converting the Fremont factories Model S

472
00:23:21,519 --> 00:23:24,720
and X lines to produce one million robots per year

473
00:23:25,000 --> 00:23:28,279
one million, and with the massive gigatexis facility, they are

474
00:23:28,279 --> 00:23:31,480
targeting a staggering ten million robots per year with the

475
00:23:31,559 --> 00:23:32,759
upcoming Gen four.

476
00:23:32,880 --> 00:23:36,279
Speaker 1: Ten million humanoid robots a year. It's almost impossible to

477
00:23:36,359 --> 00:23:39,240
visualize that level of manufacturing. It is, but there's a

478
00:23:39,279 --> 00:23:42,680
hard physics limit to this robotic flywheel, isn't there. You

479
00:23:42,720 --> 00:23:46,160
can build ten million humanoids, you can deploy millions of

480
00:23:46,200 --> 00:23:50,119
autonomous cars, and you can run digital optimists on millions

481
00:23:50,119 --> 00:23:54,640
of parked computers, but all of that artificial intelligence requires

482
00:23:54,640 --> 00:24:00,319
a terrifying, almost incomprehensible amount of computational power. Yes, and

483
00:24:00,359 --> 00:24:02,440
you can't just plug all of that into the wall.

484
00:24:02,519 --> 00:24:05,839
The electrical grid is already tapped out. This brings us

485
00:24:05,839 --> 00:24:09,960
to a massive bottleneck that threatens this entire vision the

486
00:24:10,000 --> 00:24:11,319
global energy crisis.

487
00:24:11,599 --> 00:24:14,119
Speaker 3: If you look at the macrotns and technology right now,

488
00:24:14,559 --> 00:24:18,279
energy is the only true currency. Everyone is building colossal

489
00:24:18,359 --> 00:24:20,960
AI data centers to train their models, and they are

490
00:24:21,039 --> 00:24:22,880
utterly draining regional power grids.

491
00:24:22,960 --> 00:24:23,599
Speaker 1: How bad is it?

492
00:24:23,759 --> 00:24:27,000
Speaker 3: US data centers are currently drawing forty one gigawants of power.

493
00:24:27,160 --> 00:24:29,759
That represents one hundred and fifty percent increase in just

494
00:24:29,799 --> 00:24:30,480
the last five.

495
00:24:30,400 --> 00:24:33,400
Speaker 1: Years to put forty one gigawatts. In perspective, that rivals

496
00:24:33,440 --> 00:24:35,359
the output of all the nuclear power plants in the

497
00:24:35,480 --> 00:24:36,559
United States combined.

498
00:24:36,720 --> 00:24:37,640
Speaker 3: Just for data centers.

499
00:24:37,720 --> 00:24:43,359
Speaker 1: It's for data centers, and the trajectory is completely unsustainable. PJM,

500
00:24:43,480 --> 00:24:45,839
which is the largest grid operator in the country serving

501
00:24:45,839 --> 00:24:50,160
over sixty five million people across multiple states, projects they

502
00:24:50,200 --> 00:24:54,319
will be six gigawats short of basic reliability requirements by

503
00:24:54,359 --> 00:24:57,240
twenty twenty seven, that is next year. That is next year.

504
00:24:57,279 --> 00:25:01,519
Speaker 3: We are facing the very real potential for blackouts simply

505
00:25:01,519 --> 00:25:04,319
because server farms are consuming too much electricity.

506
00:25:04,440 --> 00:25:07,319
Speaker 1: What's fascinating about the grid crisis is how it completely

507
00:25:07,359 --> 00:25:09,400
crosses traditional political divides.

508
00:25:09,480 --> 00:25:10,240
Speaker 3: It really does.

509
00:25:10,400 --> 00:25:13,200
Speaker 1: You have politicians from completely different sides of the aisle,

510
00:25:13,279 --> 00:25:16,720
like Senator Bernie Sanders and Governor Ron DeSantis both pointing

511
00:25:16,759 --> 00:25:19,960
out the exact same issue. And we are just looking

512
00:25:20,000 --> 00:25:22,440
at the math here right, not endorsing any political stance.

513
00:25:22,759 --> 00:25:25,680
But when both sides are publicly raising alarms about the

514
00:25:25,680 --> 00:25:28,720
sheer strain these data centers are putting on local power grids,

515
00:25:28,759 --> 00:25:29,720
you know it's serious.

516
00:25:29,759 --> 00:25:30,279
Speaker 3: Absolutely.

517
00:25:30,319 --> 00:25:33,359
Speaker 1: When massive tech companies siphon off gigawatts of power, it

518
00:25:33,440 --> 00:25:37,519
restricts supply, which drives up utility bills for everyday consumers.

519
00:25:37,960 --> 00:25:41,359
The math doesn't care about politics. A gigawatch shortfall is

520
00:25:41,359 --> 00:25:44,680
a gigawatch shortfall, and constituents are getting angry.

521
00:25:44,720 --> 00:25:48,920
Speaker 3: With grid power becoming this incredibly scarce, politically volatile resource.

522
00:25:49,400 --> 00:25:50,759
The major tech titans.

523
00:25:50,440 --> 00:25:51,720
Speaker 1: Are panicking, they really are.

524
00:25:52,000 --> 00:25:55,720
Speaker 3: Google, Microsoft and Amazon are desperately fighting each other for

525
00:25:55,799 --> 00:25:58,880
whatever grid power is left. They're locking up long term

526
00:25:59,000 --> 00:26:03,200
nuclear contracts. Microsoft is famously involved in a massive effort

527
00:26:03,319 --> 00:26:08,160
to restart the decommissioned three Mile Island nuclear facility. They're

528
00:26:08,200 --> 00:26:13,119
completely reliant on and effectively begging traditional utility companies for

529
00:26:13,200 --> 00:26:14,279
permission to grow.

530
00:26:14,319 --> 00:26:16,359
Speaker 1: But Musk is playing a different game here too. He

531
00:26:16,400 --> 00:26:19,720
has a three prong strategy to completely bypass this earthly

532
00:26:19,759 --> 00:26:20,599
grid bottleneck.

533
00:26:20,680 --> 00:26:22,680
Speaker 3: We of number one Tesla Energy.

534
00:26:22,839 --> 00:26:23,640
Speaker 1: Let's talk about it.

535
00:26:23,720 --> 00:26:26,759
Speaker 3: While Microsoft and Google are begging for power, Tesla is

536
00:26:26,839 --> 00:26:31,160
actually selling the power infrastructure. Last year alone, they deployed

537
00:26:31,240 --> 00:26:35,200
forty seven gigawatt hours of commercial energy storage, generating over

538
00:26:35,279 --> 00:26:38,640
twelve billion dollars in revenue. Wow, and they rely on

539
00:26:38,680 --> 00:26:43,839
their own hardware. Xai's massive Memphis supercomputer named Colossus isn't

540
00:26:43,880 --> 00:26:46,160
just subject to the whims of the local Tennessee grid.

541
00:26:46,200 --> 00:26:46,839
Speaker 1: What do they use?

542
00:26:47,119 --> 00:26:50,079
Speaker 3: It is powered and stabilized by one hundred and sixty

543
00:26:50,079 --> 00:26:53,759
eight Tesla megapas, representing hundreds of millions of dollars of

544
00:26:53,759 --> 00:26:57,440
proprietary energy storage, acting as a massive buffer.

545
00:26:57,359 --> 00:27:00,920
Speaker 1: So they insulate themselves from grid instability. But storing power

546
00:27:01,000 --> 00:27:03,880
still requires generating power, which brings us to move number

547
00:27:03,880 --> 00:27:06,119
two solar generation right.

548
00:27:06,519 --> 00:27:09,440
Speaker 3: The stated target is to build out one hundred gigawatts

549
00:27:09,480 --> 00:27:12,799
per year of US solar manufacturing.

550
00:27:12,119 --> 00:27:15,160
Speaker 1: Capacity one hundred gigawatts a year. That is more than

551
00:27:15,279 --> 00:27:19,200
double what the entire US data center industry is currently consuming.

552
00:27:19,400 --> 00:27:22,960
Speaker 3: The goal is complete energy autonomy. By pairing that massive

553
00:27:23,000 --> 00:27:26,279
solar generation with the megapex storage, they can create self

554
00:27:26,319 --> 00:27:29,400
sufficient mini power plants for their data centers and factories,

555
00:27:29,599 --> 00:27:32,160
cutting energy costs by up to forty percent compared to

556
00:27:32,200 --> 00:27:33,799
traditional fossil fuel reliance.

557
00:27:33,920 --> 00:27:36,039
Speaker 1: It's like everyone in the tech industry is crammed into

558
00:27:36,119 --> 00:27:39,559
a tiny kitchen viciously fighting over the very last slice

559
00:27:39,559 --> 00:27:42,160
of pie, and Elon just walks outside and builds a

560
00:27:42,160 --> 00:27:45,119
massive solar powered bakery in the backyard. Great analogy, but

561
00:27:45,200 --> 00:27:48,920
eventually even the backyard isn't big enough because move number

562
00:27:48,920 --> 00:27:53,839
three is where the architecture leaves the atmosphere entirely. Let's

563
00:27:53,880 --> 00:27:54,880
talk about the space play.

564
00:27:55,200 --> 00:27:58,039
Speaker 3: This is perhaps the most audacious paradigm shift in the

565
00:27:58,200 --> 00:28:02,559
entire thesis has filed with the FCC for a constellation

566
00:28:02,640 --> 00:28:07,160
of up to one million low Earth orbit satellites. A million,

567
00:28:07,279 --> 00:28:10,440
but these are not for providing starlink internet to remote areas.

568
00:28:10,759 --> 00:28:15,359
They're specifically designed to be solar powered orbital AI data centers.

569
00:28:15,480 --> 00:28:17,599
Speaker 1: Orbital data centers. They want to put the heavy compute

570
00:28:17,599 --> 00:28:20,359
in space. I hear that, and my initial reaction is

571
00:28:20,480 --> 00:28:23,759
why go through the immense difficulty of launching servers into

572
00:28:23,799 --> 00:28:24,759
the vacuum of space?

573
00:28:24,920 --> 00:28:27,160
Speaker 3: Because in lower th orbit there are no grid constraints.

574
00:28:27,480 --> 00:28:29,640
There are no local zoning boards to fight, no land

575
00:28:29,720 --> 00:28:33,960
use battles with municipalities, and no politicians complaining about consumer

576
00:28:34,000 --> 00:28:35,000
electricity prices.

577
00:28:35,079 --> 00:28:35,599
Speaker 1: That's true.

578
00:28:35,640 --> 00:28:39,599
Speaker 3: Most importantly, if you position them in specific sun syncredous orbits,

579
00:28:39,759 --> 00:28:43,759
the sun never sets. You have continuous, unfiltered solar energy

580
00:28:44,119 --> 00:28:46,640
bombarding the solar panels twenty four hours a day, seven

581
00:28:46,720 --> 00:28:48,920
days a week. You never need a battery backup.

582
00:28:49,160 --> 00:28:53,359
Speaker 1: The proposed math is staggering. SpaceX estimates that launching one

583
00:28:53,359 --> 00:28:57,079
million tons of these satellites annually, with each ton generating

584
00:28:57,160 --> 00:29:01,319
one hundred kilowatts of compute power, could add one hundred

585
00:29:01,319 --> 00:29:05,519
gigawatts of AI compute capacity every single year, every single year,

586
00:29:05,680 --> 00:29:09,359
free energy beamed directly from the Sun into silicon processors

587
00:29:09,400 --> 00:29:11,880
floating in zero gravity. But I'm sorry I have to

588
00:29:11,880 --> 00:29:14,440
push back on this. Sure it sounds like a billionaire's

589
00:29:14,559 --> 00:29:17,880
sci fi vanity project. The sheer cost of rocketry aside.

590
00:29:18,279 --> 00:29:20,880
What about the laws of physics? What about latency?

591
00:29:21,160 --> 00:29:22,039
Speaker 3: H latency?

592
00:29:22,200 --> 00:29:24,839
Speaker 1: Yeah, If I am sitting in my parked car and

593
00:29:24,920 --> 00:29:26,920
digital optimist is trying to click a button on a

594
00:29:26,920 --> 00:29:29,759
spreadsheet and it has to beam that request up to

595
00:29:29,799 --> 00:29:33,359
a satellite, process it and beam it back down, the

596
00:29:33,480 --> 00:29:36,960
lag would make the system completely unusable for real time applications.

597
00:29:37,000 --> 00:29:39,680
Speaker 3: If they were relying on orbit for real time inference,

598
00:29:39,720 --> 00:29:42,400
you'd be absolutely right. The round trip latency would be

599
00:29:42,400 --> 00:29:45,319
a deal breaker for reflexes. But that's exactly why the

600
00:29:45,400 --> 00:29:48,799
vertical integration saves. The concept has the local AI four

601
00:29:48,839 --> 00:29:51,920
hardware inside the cars, and the robots handles the split

602
00:29:52,000 --> 00:29:56,240
second real time decisions the fast reflexes. The orbital compute

603
00:29:56,279 --> 00:29:59,359
layer would be strictly reserved for the heavy asynchronous training

604
00:29:59,400 --> 00:30:00,839
of the massive foundation models.

605
00:30:00,960 --> 00:30:02,200
Speaker 1: Oh, I get it right.

606
00:30:02,240 --> 00:30:05,319
Speaker 3: When you are training a neural network on billions of

607
00:30:05,400 --> 00:30:08,559
miles of video data over weeks or months, it doesn't

608
00:30:08,599 --> 00:30:10,559
matter if it takes a few extra milliseconds for the

609
00:30:10,640 --> 00:30:11,720
data packet to retorbit.

610
00:30:11,880 --> 00:30:13,799
Speaker 1: Ah. So the heavy lifting is done in the sky

611
00:30:13,880 --> 00:30:15,960
and the quick reflexes happen on the ground.

612
00:30:16,319 --> 00:30:20,119
Speaker 3: Exactly and regarding your point on the sheer cost of rocketry.

613
00:30:20,920 --> 00:30:23,839
That is exactly why SpaceX has spent the last decade

614
00:30:24,240 --> 00:30:29,319
relentlessly developing fully reusable, massive payload rockets like Starship. Right,

615
00:30:29,799 --> 00:30:32,480
the entire economic purpose of Starship is to drive the

616
00:30:32,559 --> 00:30:35,759
cost orbit down to a level where launching a million

617
00:30:35,839 --> 00:30:38,400
tons of server racks becomes financially viable.

618
00:30:38,559 --> 00:30:41,759
Speaker 1: It all connects, It's a closed loop. SpaceX exists to

619
00:30:41,880 --> 00:30:45,000
launch the data centers which provide the infinite compute to

620
00:30:45,079 --> 00:30:48,200
train the AI which runs the robots which build the

621
00:30:48,200 --> 00:30:50,759
cars and the rocket. It is dizzying, it really is.

622
00:30:51,680 --> 00:30:55,000
But to pull off orbital compute, millions of robots and

623
00:30:55,039 --> 00:30:58,160
autonomous cars at this global scale, you can't just rely

624
00:30:58,240 --> 00:31:01,839
on outside suppliers for your basic building blocks. You have

625
00:31:01,920 --> 00:31:05,720
to build the machine that builds the machine. Yes, which

626
00:31:05,759 --> 00:31:08,480
brings us to the crucial concept of owning the stack.

627
00:31:08,799 --> 00:31:12,799
Speaker 3: Vertical integration is the invisible connective tissue holding this entire

628
00:31:12,839 --> 00:31:16,920
ecosystem together. Yeah, Apple, Microsoft, Google, they are all tech titans,

629
00:31:16,920 --> 00:31:18,559
but none of them own the entire stack.

630
00:31:18,720 --> 00:31:19,200
Speaker 1: That's true.

631
00:31:19,359 --> 00:31:22,680
Speaker 3: Apple doesn't own its own power grid, Microsoft doesn't manufacture

632
00:31:22,720 --> 00:31:25,839
its own vehicles, Google doesn't fabricate its own silicon or

633
00:31:25,920 --> 00:31:29,680
launch its own space data centers. Musk's architecture seeks to

634
00:31:29,680 --> 00:31:33,720
control the entire vertical, the energy, the chips, the factories,

635
00:31:33,759 --> 00:31:35,920
the hardware, and the software models.

636
00:31:36,119 --> 00:31:38,599
Speaker 1: Let's start with chips, because right now silicon is the

637
00:31:38,599 --> 00:31:41,880
single biggest bottleneck in AI outside of energy. It is

638
00:31:42,000 --> 00:31:43,559
if you are a tech company and you want to

639
00:31:43,599 --> 00:31:46,000
train in AI, you are desperately waiting in line at

640
00:31:46,039 --> 00:31:50,000
one specific company TSMC and Taiwan. You weigh in line,

641
00:31:50,000 --> 00:31:52,640
you pay whatever in Nvidia charges for their GPUs, and

642
00:31:52,680 --> 00:31:55,599
you hope geopolitical tensions don't disrupt the supply chain.

643
00:31:55,839 --> 00:31:59,480
Speaker 3: But Musk is bypassing that line entirely. They are building

644
00:31:59,599 --> 00:32:03,519
terrafact out, a massive twenty to twenty five billion dollar

645
00:32:03,640 --> 00:32:07,519
chip fabrication facility in Texas, targeting a formal launch around

646
00:32:07,559 --> 00:32:10,759
March twenty first. They are aiming for a cutting edge

647
00:32:10,880 --> 00:32:15,640
two nanometer process technology, projecting a manufacturing capacity of one

648
00:32:15,720 --> 00:32:18,359
million wafers per month by twenty thirty.

649
00:32:18,720 --> 00:32:21,440
Speaker 1: Let's break down what a two minimeter process actually means,

650
00:32:21,480 --> 00:32:23,720
because it sounds like jargon to a lot of people. Sure,

651
00:32:23,839 --> 00:32:27,480
when we talk about nanometers in ships, imagine the transistor

652
00:32:27,599 --> 00:32:31,599
as a microscopic drawbridge for electricity. The smaller you make

653
00:32:31,640 --> 00:32:34,319
those drawbridges down to just a few atoms wide, the

654
00:32:34,359 --> 00:32:36,119
more of them you can pack onto a single piece

655
00:32:36,119 --> 00:32:36,720
of silicon.

656
00:32:36,880 --> 00:32:41,680
Speaker 3: Exactly, more transistors mean exponentially more computing power. But crucially,

657
00:32:41,920 --> 00:32:45,160
because the electrical current has less distance to travel, smaller

658
00:32:45,200 --> 00:32:48,880
transistors require far less electricity and generate significantly less heat,

659
00:32:49,039 --> 00:32:51,759
which is key absolutely when you are trying to solve

660
00:32:51,799 --> 00:32:55,599
a global energy bottleneck and put computers inside humanoid robots.

661
00:32:55,960 --> 00:32:59,440
Highly efficient microscopic chips are just as vital as massive

662
00:32:59,440 --> 00:33:01,119
solar farms, and the.

663
00:33:01,079 --> 00:33:03,799
Speaker 1: Hardware they are baking in Texas is custom designed for

664
00:33:03,839 --> 00:33:08,400
their exact neural network architecture. The upcoming AI five chip

665
00:33:08,519 --> 00:33:11,559
is projected to have forty to fifty times more compute

666
00:33:11,559 --> 00:33:14,640
power and nine times more memory than the current AI

667
00:33:14,640 --> 00:33:19,599
four generation. Usually, by designing and fabricating their own custom silicon,

668
00:33:19,799 --> 00:33:24,319
they completely eliminate their dependency on nvidious pricing and TSMC's

669
00:33:24,359 --> 00:33:25,920
geographic vulnerabilities.

670
00:33:26,000 --> 00:33:29,880
Speaker 3: It guarantees total hardware sovereignty. And speaking of hardware, we

671
00:33:29,960 --> 00:33:32,720
have to look at how they actually assemble these physical products.

672
00:33:33,400 --> 00:33:36,839
The manufacturing process itself is a proprietary.

673
00:33:36,240 --> 00:33:38,519
Speaker 1: Advantage right unboxed manufacturing.

674
00:33:38,599 --> 00:33:42,519
Speaker 3: With the upcoming cybercab, they're introducing a manufacturing revolution. They

675
00:33:42,519 --> 00:33:44,319
call it the unboxed process.

676
00:33:44,519 --> 00:33:47,640
Speaker 1: To really appreciate unboxed manufacturing, you have to look at

677
00:33:47,640 --> 00:33:50,559
the physics of a traditional auto assembly line. For over

678
00:33:50,599 --> 00:33:53,960
a century since Henry Ford, automakers have built cars sequentially.

679
00:33:54,440 --> 00:33:56,480
You start with a metal frame and it moves slowly

680
00:33:56,519 --> 00:33:59,640
down a massive mile long conveyor belt. A worker puts

681
00:33:59,680 --> 00:34:02,240
on a door or the line moves. Another worker drops

682
00:34:02,240 --> 00:34:05,079
in an engine, the line moves. It's entirely linear, like.

683
00:34:05,000 --> 00:34:08,199
Speaker 3: Stringing pearls on a necklace, one piece at a time

684
00:34:08,360 --> 00:34:11,960
in a straight line. It's slow, it requires massive amounts

685
00:34:11,960 --> 00:34:14,920
of factory floor space, and if one station breaks down,

686
00:34:15,039 --> 00:34:16,360
the entire line stops.

687
00:34:16,400 --> 00:34:17,159
Speaker 1: It's a bottleneck.

688
00:34:17,320 --> 00:34:21,119
Speaker 3: The unboxed process shatters that linear constraint. It is modular

689
00:34:21,400 --> 00:34:22,440
parallel assembly.

690
00:34:22,719 --> 00:34:25,360
Speaker 1: Think of it more like a group of people building

691
00:34:25,400 --> 00:34:29,360
a complex lego set. Instead of passing one single block around,

692
00:34:29,519 --> 00:34:33,480
different teams are building different sections simultaneously exactly. One group

693
00:34:33,519 --> 00:34:36,320
builds the front end, another builds the rear casting, another

694
00:34:36,360 --> 00:34:40,880
builds the interior seats. All these complex subassemblies are completed independently,

695
00:34:41,199 --> 00:34:43,039
and then they are brought together at the very end

696
00:34:43,119 --> 00:34:46,000
and snapped together in one fluid automated motion.

697
00:34:46,280 --> 00:34:50,880
Speaker 3: The efficiency gains are incredible. This parallel modular approach cuts

698
00:34:50,880 --> 00:34:54,159
the necessary factory footprint by forty percent, it reduces labor

699
00:34:54,199 --> 00:34:57,079
costs by thirty percent, and it cuts total assembly time

700
00:34:57,119 --> 00:34:58,000
by fifty percent.

701
00:34:58,159 --> 00:34:58,679
Speaker 2: Wow.

702
00:34:58,920 --> 00:35:01,920
Speaker 3: Their target cycle time T is under ten seconds per vehicle.

703
00:35:02,239 --> 00:35:05,400
That is in traditional automotive manufacturing anymore. That is the

704
00:35:05,440 --> 00:35:08,320
speed and precision of consumer electronics. That is how you

705
00:35:08,360 --> 00:35:12,360
assemble iPhones, applied to two ton autonomous vehicles, and.

706
00:35:12,360 --> 00:35:16,519
Speaker 1: This manufacturing prowess is applied universally across the ecosystem. The

707
00:35:16,599 --> 00:35:20,320
factories themselves are just nodes in the network. The Fremont

708
00:35:20,320 --> 00:35:23,880
factory converts to building robots. Texas bills the custom chips

709
00:35:23,880 --> 00:35:28,199
and the cybercabs. Houston builds the Megapax to power the supercomputers.

710
00:35:28,400 --> 00:35:30,119
Speaker 3: They own the factories that build.

711
00:35:29,840 --> 00:35:33,559
Speaker 1: The machines, and finally, governing all of this physical hardware

712
00:35:33,719 --> 00:35:37,039
is the top layer of the stack, the software mind.

713
00:35:37,199 --> 00:35:40,480
We touched on Xai's GROC earlier, but it is much

714
00:35:40,559 --> 00:35:43,800
more than just a consumer chatbot generating text right.

715
00:35:43,840 --> 00:35:46,599
Speaker 3: The consumer side of GROC is projected to generate roughly

716
00:35:46,639 --> 00:35:50,480
one point two billion dollars in subscription revenue, which is nice,

717
00:35:50,760 --> 00:35:53,199
But the real strategic value isn't a few dollars a

718
00:35:53,199 --> 00:35:56,000
month from Internet users now, the real value is having

719
00:35:56,000 --> 00:35:59,039
an in house foundation model. GROC is the system to

720
00:35:59,239 --> 00:35:59,920
reasoning layer.

721
00:36:00,119 --> 00:36:01,039
Speaker 1: Okay, unpack that.

722
00:36:01,320 --> 00:36:03,239
Speaker 3: While a local AI on the car or a robot

723
00:36:03,320 --> 00:36:07,039
handles the immediate real time spatial execution, rock ACTS is

724
00:36:07,079 --> 00:36:11,360
the deep strategic thinker. It provides the contextual logical understanding

725
00:36:11,400 --> 00:36:15,440
required for digital optimists to navigate a complex enterprise software suite,

726
00:36:15,679 --> 00:36:17,800
or for a humanoid robot to plan out a multi

727
00:36:17,920 --> 00:36:19,760
step physical manufacturing task.

728
00:36:20,000 --> 00:36:22,440
Speaker 1: So it's the brains of the operation exactly.

729
00:36:22,920 --> 00:36:26,559
Speaker 3: Having an in house competitor to open AI's GPT or

730
00:36:26,599 --> 00:36:31,400
Google's Gemini that is natively seamlessly integrated into your own

731
00:36:31,480 --> 00:36:35,519
proprietary silicon in your own physical robots is the ultimate

732
00:36:35,719 --> 00:36:37,599
master stroke of vertical integration.

733
00:36:37,880 --> 00:36:40,559
Speaker 1: Every single layer, from the solar panel to the silicon

734
00:36:40,599 --> 00:36:44,320
ship to the software model makes all the other layers cheaper, faster,

735
00:36:44,440 --> 00:36:44,920
and smarter.

736
00:36:45,639 --> 00:36:46,639
Speaker 3: That's the ecosystem.

737
00:36:46,719 --> 00:36:49,440
Speaker 1: Okay, we have just painted a picture of an absolute

738
00:36:49,480 --> 00:36:52,960
technological juggernaut. It sounds invincible. It sounds like a closed

739
00:36:52,960 --> 00:36:56,480
loop empire that cannot possibly fail. But what happens when

740
00:36:56,480 --> 00:37:00,280
this grand vision collides with reality? We need to ground this.

741
00:37:00,639 --> 00:37:03,320
Let's look at the immense risks and what financial analysts

742
00:37:03,400 --> 00:37:05,039
call the unpriced optionality.

743
00:37:05,320 --> 00:37:08,679
Speaker 3: We absolutely must address the execution risks because a plan

744
00:37:08,719 --> 00:37:12,440
of this magnitude carries existential threats. First and foremost are

745
00:37:12,480 --> 00:37:16,480
the timelines. Oh yeah, Elon Musk is notoriously, perhaps pathologically

746
00:37:16,559 --> 00:37:19,840
optimistic about when these bleeding edge technologies will actually be finished.

747
00:37:19,880 --> 00:37:22,599
Speaker 1: To put it mildly, the Robotaxi network was supposed to

748
00:37:22,639 --> 00:37:25,440
be driving US around years ago. Optimists was supposed to

749
00:37:25,440 --> 00:37:28,679
be a household helper by now right. Terafab is a

750
00:37:28,719 --> 00:37:32,199
twenty five billion dollar construction project that will take years

751
00:37:32,239 --> 00:37:35,239
to fully ramp up, assuming there are no supply chain disasters,

752
00:37:35,840 --> 00:37:37,199
and the orbital data.

753
00:37:36,960 --> 00:37:38,320
Speaker 3: Center still theoretical.

754
00:37:38,519 --> 00:37:41,440
Speaker 1: Yeah, the technology to process heavy AI compute in the

755
00:37:41,519 --> 00:37:44,679
radiation of space barely even exists yet, and launching a

756
00:37:44,719 --> 00:37:49,079
million tons of hardware certainly doesn't have regulatory or FCC approval.

757
00:37:49,320 --> 00:37:52,079
Speaker 3: And then there's the massive capital expenditure required to keep

758
00:37:52,119 --> 00:37:57,719
the flywheel spinning while these technologies mature. Tesla's projected CAPEX

759
00:37:57,760 --> 00:38:02,119
in twenty twenty six alone needs twenty billion dollars.

760
00:38:02,400 --> 00:38:03,400
Speaker 1: That's a lot of money.

761
00:38:03,559 --> 00:38:06,719
Speaker 3: That is an astronomical amount of cash to burn. If

762
00:38:06,760 --> 00:38:10,880
these long term bets hit unexpected roadblocks, If the Rootaxi

763
00:38:10,920 --> 00:38:14,400
network gets bogged down in endless municipal regulatory hearings, or

764
00:38:14,440 --> 00:38:16,920
if the Optimist robot proves too clumsy to work safely

765
00:38:16,960 --> 00:38:20,320
outside of a highly controlled factory environment, tens of billions

766
00:38:20,320 --> 00:38:22,400
of dollars in capital could be completely scranded.

767
00:38:22,719 --> 00:38:25,320
Speaker 1: Not to mention, the existing tech titans aren't just taking

768
00:38:25,320 --> 00:38:29,280
a nap. Weimo is actively expanding their operational footprint right now,

769
00:38:29,320 --> 00:38:33,039
mapping new cities. Google and Anthropic are releasing vastly smarter

770
00:38:33,119 --> 00:38:37,639
AI models every few months. Nvidia is entrenching its GPU

771
00:38:37,679 --> 00:38:41,480
ecosystem deeper into every enterprise on Earth. Just because a

772
00:38:41,519 --> 00:38:44,559
new custom stack is being built doesn't automatically mean the

773
00:38:44,599 --> 00:38:47,320
existing massive industry stack just vanishes.

774
00:38:47,440 --> 00:38:51,199
Speaker 3: Those are very real existential risks, but this provides a

775
00:38:51,280 --> 00:38:54,960
vital framework for you, the listener. As you evaluate this ecosystem,

776
00:38:55,400 --> 00:38:57,519
you have to step back and look at the underlying

777
00:38:57,559 --> 00:38:58,679
macroeconomic trends.

778
00:38:58,760 --> 00:39:00,000
Speaker 1: Okay, let's walk through them.

779
00:39:00,320 --> 00:39:03,440
Speaker 3: Ask yourself, over the next decade, are we going to

780
00:39:03,559 --> 00:39:07,000
need more AI compute globally or less?

781
00:39:07,119 --> 00:39:07,880
Speaker 1: Definitely more?

782
00:39:08,079 --> 00:39:11,039
Speaker 3: Is energy generation and grid capacity going to be the

783
00:39:11,119 --> 00:39:12,880
ultimate bottleneck for that compute?

784
00:39:12,920 --> 00:39:14,679
Speaker 1: The data strongly suggests yes.

785
00:39:14,840 --> 00:39:19,000
Speaker 3: Is pure vertical integration fundamentally more defensible and agile than

786
00:39:19,039 --> 00:39:22,880
being totally dependent on third party international suppliers for your

787
00:39:22,960 --> 00:39:23,840
chips and software?

788
00:39:23,880 --> 00:39:24,880
Speaker 1: Almost always yes?

789
00:39:24,920 --> 00:39:29,679
Speaker 3: And is real world, messy, chaotic visual data more valuable

790
00:39:29,760 --> 00:39:34,360
for training physical AI than clean, simulated laboratory data. So

791
00:39:34,400 --> 00:39:37,639
the final synthesis of this framework is this. The risk

792
00:39:37,840 --> 00:39:41,000
is that this totally different game fails because the timelines

793
00:39:41,039 --> 00:39:44,119
are simply too long, or the capital requirements are too heavy,

794
00:39:44,320 --> 00:39:46,440
or the physics are just too hard to overcome. Right,

795
00:39:46,920 --> 00:39:49,400
But the opportunity is that no one else is even

796
00:39:49,400 --> 00:39:53,360
playing this specific game. If the timeline for the Robotaxi

797
00:39:53,480 --> 00:39:56,840
or optimists slips by two years or even five years,

798
00:39:57,159 --> 00:40:02,800
the underlying vertically integrated architecture being built remains entirely without peer.

799
00:40:03,039 --> 00:40:05,800
Speaker 1: It's like judging an Olympic decaflee only by their score.

800
00:40:05,800 --> 00:40:08,000
In the pole vault. If you look at Tesla and say, well,

801
00:40:08,000 --> 00:40:11,079
their automotive profit margins dipped by two percent this quarter,

802
00:40:11,840 --> 00:40:14,880
you are entirely missing the big picture of what they're

803
00:40:14,920 --> 00:40:18,039
actually training for, completely missing it, and that brings us

804
00:40:18,039 --> 00:40:20,480
to the concept of unpriced optionality.

805
00:40:20,840 --> 00:40:23,480
Speaker 3: This is crucial for anyone trying to understand the financial

806
00:40:23,480 --> 00:40:27,880
reality of this system. Currently, the broader market still largely

807
00:40:27,920 --> 00:40:30,559
prices Tesla as strictly as a traditional car company.

808
00:40:30,719 --> 00:40:33,440
Speaker 1: Yeah, they treat it like Ford or GM exactly.

809
00:40:33,800 --> 00:40:37,400
Speaker 3: Analysts look at traditional automotive multiples, like how many metal

810
00:40:37,400 --> 00:40:40,119
boxes they sold, what the margin on the metal was,

811
00:40:40,559 --> 00:40:43,880
and they price the stock accordingly, perhaps sprinkling in a

812
00:40:43,880 --> 00:40:45,360
slight premium for the AI hype.

813
00:40:45,719 --> 00:40:49,519
Speaker 1: But if this vertically integrated thesis holds true, that pricing

814
00:40:49,559 --> 00:40:52,920
model is completely disconnected from reality.

815
00:40:52,639 --> 00:40:57,280
Speaker 3: Entirely disconnected. Because a scaled ROBOTAXI network wiping out personal

816
00:40:57,280 --> 00:41:02,079
car ownership that's not priced in, enterprise grade digital optimists

817
00:41:02,199 --> 00:41:05,440
functionally doing the back office work of entire software companies

818
00:41:05,719 --> 00:41:09,400
not priced in a twenty five billion dollar terrafab churning

819
00:41:09,440 --> 00:41:13,559
out proprietary hyper efficient AI chips not priced in, and

820
00:41:13,840 --> 00:41:18,000
orbital compute data centers floating in space harvesting infinite solar energy.

821
00:41:18,119 --> 00:41:19,880
That's not even on a market's radar yet.

822
00:41:20,000 --> 00:41:22,039
Speaker 1: So for the learners out there who want to watch

823
00:41:22,079 --> 00:41:24,320
this play out in real time over the next few years,

824
00:41:24,599 --> 00:41:26,559
what are the key metrics we should track to see

825
00:41:26,559 --> 00:41:28,920
if this grand vision is actually coming together or if

826
00:41:28,960 --> 00:41:29,760
it's falling apart.

827
00:41:29,960 --> 00:41:34,079
Speaker 3: There are three main indicators to watch closely. First, watch

828
00:41:34,119 --> 00:41:37,639
the cybercab production ramps starting next month. If they can

829
00:41:37,679 --> 00:41:42,320
truly utilize that modular, unboxed manufacturing process at scale to

830
00:41:42,400 --> 00:41:45,519
turn out cheap vehicles, it validates their entire hardware and

831
00:41:45,599 --> 00:41:49,599
factory thesis. Okay, that's one second. Closely, watch the commercial

832
00:41:49,840 --> 00:41:54,280
energy storage deployment numbers. Every megapac they sell or install

833
00:41:54,599 --> 00:41:58,719
is another brick in their independent decentralized power infrastructure. Got it?

834
00:41:58,960 --> 00:41:59,719
Speaker 1: And the third.

835
00:42:00,119 --> 00:42:04,679
Speaker 3: Third, we have to watch that ten billion mile FSD benchmark. Yeah,

836
00:42:04,800 --> 00:42:07,280
when they cross that massive data line, which is imminent,

837
00:42:07,760 --> 00:42:10,679
we need to see if it actually unlocks the generalized

838
00:42:10,800 --> 00:42:14,760
unsupervised autonomy they've been promising, proving that pure vision can

839
00:42:14,760 --> 00:42:16,000
conquer the physical world.

840
00:42:16,320 --> 00:42:19,320
Speaker 1: If those three metrics hold strong, the flywheel is spinning

841
00:42:19,480 --> 00:42:23,039
exactly as design exactly. This has been an absolutely staggering

842
00:42:23,119 --> 00:42:26,280
journey through a truly mind bending framework. But before we

843
00:42:26,320 --> 00:42:28,360
wrap up today, I want to leave you with one final,

844
00:42:28,519 --> 00:42:31,840
deeply provocative thought, Tom all Over, something that builds on

845
00:42:31,880 --> 00:42:33,960
all this but takes it one step further into the future.

846
00:42:34,039 --> 00:42:34,800
Speaker 3: Okay, I'm ready.

847
00:42:34,920 --> 00:42:38,840
Speaker 1: We've talked extensively today about leveraging vertical integration to drive

848
00:42:38,880 --> 00:42:42,800
the cost of compute, transportation and manufacturing into the ground.

849
00:42:43,440 --> 00:42:47,320
But if you extrapolate this ecosystem to its logical conclusion, right,

850
00:42:47,400 --> 00:42:51,760
if energy eventually trends towards zero cost because of infinite

851
00:42:51,880 --> 00:42:56,039
solar arrays in orbit, and if physical labor trends towards

852
00:42:56,079 --> 00:43:00,039
zero cost because of millions of tireless humanoid robots, So

853
00:43:00,119 --> 00:43:03,519
what actually happens to the fundamental nature of human economics?

854
00:43:04,280 --> 00:43:06,440
Speaker 3: That is the million dollar question, or I guess the

855
00:43:06,480 --> 00:43:10,599
trillion dollar question. Our entire global economic system, capitalism itself,

856
00:43:10,920 --> 00:43:13,840
is based entirely on the management of scarcity.

857
00:43:13,480 --> 00:43:17,519
Speaker 1: Exactly, the scarcity of human labor, the scarcity of physical resources,

858
00:43:17,559 --> 00:43:18,800
the scarcity of energy.

859
00:43:18,960 --> 00:43:23,440
Speaker 3: If this unified physical AI system successfully replaces scarcity with

860
00:43:23,519 --> 00:43:26,599
absolute abundance, the rules of the game. Don't just change

861
00:43:26,599 --> 00:43:29,719
for tech companies, they change for humanity. If energy and

862
00:43:29,800 --> 00:43:33,920
labor are essentially free, what becomes the new most valuable

863
00:43:33,960 --> 00:43:37,880
resource in our society? Is it pure human creativity? Is

864
00:43:37,880 --> 00:43:42,119
it authentic emotional connection? We might be watching the infrastructure

865
00:43:42,400 --> 00:43:44,840
of a post scarcity world being built right in front

866
00:43:44,840 --> 00:43:47,760
of us, cleverly disguised as a car company.

867
00:43:48,159 --> 00:43:49,599
Speaker 1: That is a thought that is going to keep me

868
00:43:49,679 --> 00:43:53,480
up tonight. So we've laid out the pieces of this massive,

869
00:43:53,559 --> 00:43:56,480
complex puzzle. Now we want to know where you stand.

870
00:43:57,000 --> 00:44:00,800
Do you think this grand vertically integrated everything engine will

871
00:44:00,880 --> 00:44:05,360
actually succeed and fundamentally reshape our economic reality? Or do

872
00:44:05,400 --> 00:44:07,760
you think it is just a multi billion dollar house

873
00:44:07,760 --> 00:44:10,800
of cards waiting for a stiff breeze of regulatory or

874
00:44:10,840 --> 00:44:14,280
engineering reality to knock it down. Drop a comment, share

875
00:44:14,320 --> 00:44:16,239
your thoughts, and let's keep this conversation going.

876
00:44:16,639 --> 00:44:18,400
Speaker 3: Thank you so much for joining us as we unpack

877
00:44:18,440 --> 00:44:21,239
this incredible vision. It's always a pleasure to explore these

878
00:44:21,239 --> 00:44:22,239
paradigm shifts with you.

879
00:44:22,360 --> 00:44:25,119
Speaker 1: Absolutely until next time, keep questioning the board you're playing on,

880
00:44:25,280 --> 00:44:28,280
and thanks for joining us on this edition of Thrilling Threads.