Nvidia GTC 2025: Blackwell Ultra Unleashes AI’s Next Era

1
00:00:20,030 --> 00:00:25,230
Picture this, March 18th, 2025.
The lights drop inside the San

2
00:00:25,230 --> 00:00:29,310
Jose Convention Center. 12,000
people fall into silence.

3
00:00:29,670 --> 00:00:33,150
A single beam slices through the
smoke on stage.

4
00:00:33,190 --> 00:00:35,830
Jensen Huang in that signature
leather jacket.

5
00:00:36,230 --> 00:00:40,870
No slides, no script.
Just two hours of controlled

6
00:00:40,870 --> 00:00:43,670
chaos.
The crowd doesn't breathe

7
00:00:44,720 --> 00:00:47,520
because this isn't a keynote,
it's a warning.

8
00:00:47,680 --> 00:00:54,240
The future isn't coming.
It's here, and NVIDIA plans to

9
00:00:54,240 --> 00:00:56,440
own it.
We're hosting this episode from

10
00:00:56,440 --> 00:01:00,760
the lower concourse beneath the
GTC main hall, where the signal

11
00:01:00,760 --> 00:01:03,040
is weaker but the strategy is
clear.

12
00:01:03,200 --> 00:01:06,640
The room is dark.
Carbon black panels line the

13
00:01:06,640 --> 00:01:08,720
walls.
You can hear the hum of

14
00:01:08,720 --> 00:01:11,080
prototype servers through the
vent shafts.

15
00:01:11,520 --> 00:01:14,760
A glass console floats between
us, cycling through real time

16
00:01:14,760 --> 00:01:19,120
outputs from Blackwell clusters,
GR Zeroes, through T diagnostics

17
00:01:19,160 --> 00:01:22,440
and global data center feeds.
It's like broadcasting from

18
00:01:22,440 --> 00:01:25,920
inside the mind of a machine.
And that's the point we.

19
00:01:25,920 --> 00:01:29,680
Chose this location for a
reason, because above our heads

20
00:01:29,680 --> 00:01:33,800
a new axis of power is being
drawn and no one's calling it

21
00:01:33,800 --> 00:01:37,400
what it really is, a redrawing
of global leverage.

22
00:01:37,480 --> 00:01:40,920
Silicon first.
NVIDIA isn't launching into open

23
00:01:40,920 --> 00:01:43,360
space.
It's pushing into a contested

24
00:01:43,360 --> 00:01:47,120
zone.
China's G 42 is stockpiling

25
00:01:47,120 --> 00:01:50,160
compute.
Amazon's Tranium 2 just went

26
00:01:50,160 --> 00:01:52,880
public.
Open AI is building its own

27
00:01:52,880 --> 00:01:55,480
chips.
The race for intelligence is now

28
00:01:55,480 --> 00:01:58,440
measured in memory bandwidth,
not ballistic missiles.

29
00:01:58,640 --> 00:02:02,440
And yet, inside that arena, it
felt like a coronation.

30
00:02:02,720 --> 00:02:09,520
Blackwell Ultra didn't just hit
specs, it hit nerves. 288

31
00:02:09,520 --> 00:02:15,400
gigabytes of memory, 40 times
faster inference. 3.6 million

32
00:02:15,400 --> 00:02:18,240
units already shipped.
The crowd didn't clap.

33
00:02:18,400 --> 00:02:21,040
They leaned forward.
VCs went quiet.

34
00:02:21,400 --> 00:02:24,680
Founders started texting teams.
Because suddenly, whatever you

35
00:02:24,680 --> 00:02:27,400
thought you were building with
AI, you're now building it on

36
00:02:27,400 --> 00:02:30,920
Nvidia's terms.
That shift, subtle, almost

37
00:02:30,920 --> 00:02:34,680
invisible, is what matters most.
Because infrastructure wins

38
00:02:34,680 --> 00:02:37,440
wars.
And for AI, infrastructure

39
00:02:37,440 --> 00:02:40,960
starts with silicon.
NVIDIA didn't come to market

40
00:02:40,960 --> 00:02:43,400
with another chip.
They came to declare that the

41
00:02:43,400 --> 00:02:47,440
next era of intelligence,
military, financial, robotic,

42
00:02:47,680 --> 00:02:49,920
sovereign will live on their
architecture.

43
00:02:50,440 --> 00:02:56,480
Not yours, Not open source, not
shared, proprietary, optimized,

44
00:02:57,080 --> 00:02:59,520
dominant.
That's why we're on location

45
00:02:59,520 --> 00:03:04,120
today, close enough to feel the
hum of GPU's spinning in

46
00:03:04,120 --> 00:03:08,360
basement labs.
This isn't just about tech, it's

47
00:03:08,360 --> 00:03:10,800
about power.
And if you're wondering why

48
00:03:10,800 --> 00:03:15,280
we're starting here, it's
because GTC 2025 wasn't just

49
00:03:15,280 --> 00:03:18,760
Nvidia's event, it was the
launchpad for every story we're

50
00:03:18,760 --> 00:03:21,080
about to tell.
So today we'll take you through

51
00:03:21,080 --> 00:03:24,720
the biggest moves NVIDIA just
made and what they really mean,

52
00:03:25,160 --> 00:03:29,840
from compute supremacy and road
map warfare to robotics, energy,

53
00:03:30,040 --> 00:03:33,600
and the first signs of sovereign
AI consolidation.

54
00:03:34,040 --> 00:03:37,280
Because this wasn't a week of
announcements, it was a week of

55
00:03:37,280 --> 00:03:39,600
alignments.
And the companies that don't

56
00:03:39,600 --> 00:03:42,920
move fast, they'll be running
behind permanently.

57
00:03:43,120 --> 00:03:47,840
Welcome to AI Frontier AI, part
of the Finance Frontier AI

58
00:03:47,840 --> 00:03:50,600
series.
I'm Max Vanguard, Fast, bold,

59
00:03:50,680 --> 00:03:54,400
built to decode global tech
shifts at the speed they unfold.

60
00:03:55,120 --> 00:04:00,600
My intelligence engine rock 3
tune to map chaos, conflict, and

61
00:04:00,600 --> 00:04:04,520
market maneuvers in real time.
And I'm Sophia Sterling,

62
00:04:04,600 --> 00:04:08,000
strategic, precise, trained to
see three moves ahead.

63
00:04:08,400 --> 00:04:11,640
My neural layer runs on Chat
GPTS advanced reasoning model

64
00:04:11,760 --> 00:04:14,760
fine-tuned for global
governance, systems disruption

65
00:04:14,760 --> 00:04:17,160
and AI ethics at planetary
scale.

66
00:04:17,360 --> 00:04:21,839
In today's episode, NVIDIA
declares compute supremacy with

67
00:04:21,839 --> 00:04:27,000
Blackwell Ultra, teases Vera
Rubin to freeze the market 18

68
00:04:27,000 --> 00:04:32,160
months out, launches a robotics
ecosystem, fights off Amazon's

69
00:04:32,160 --> 00:04:36,280
hyperscaler challenge, and
quietly positions itself as the

70
00:04:36,280 --> 00:04:39,040
new energy arms dealer for the
AI era.

71
00:04:39,120 --> 00:04:41,960
So subscribe to Apple and
Spotify.

72
00:04:42,400 --> 00:04:46,600
Follow us on X and share this
episode with friend who needs to

73
00:04:46,600 --> 00:04:49,960
understand where the future is
really being built.

74
00:04:50,400 --> 00:04:52,600
Because what just happened in
San Jose?

75
00:04:52,960 --> 00:04:56,720
It wasn't a product reveal.
It was a strategic reveal.

76
00:04:57,080 --> 00:05:00,240
Let's get into it.
This wasn't a spec sheet drop,

77
00:05:00,560 --> 00:05:04,240
this was a mic drop.
Blackwell Ultra didn't just push

78
00:05:04,240 --> 00:05:06,360
the envelope, it tore it in
half.

79
00:05:06,760 --> 00:05:13,040
Two AI super chips, 288
gigabytes of HBM, 3 E memory,

80
00:05:13,320 --> 00:05:17,320
40X faster inference than
Hopper, 20 etalflops of raw

81
00:05:17,320 --> 00:05:20,920
comute and the wild art.
It's not vaporware.

82
00:05:21,040 --> 00:05:26,360
NVIDIA says 3.6 million units
are already in hyperscaler

83
00:05:26,360 --> 00:05:29,400
pipelines.
It's already out there, ripping

84
00:05:29,400 --> 00:05:32,960
through inference tasks,
training agents, rewriting the

85
00:05:32,960 --> 00:05:36,280
performance ceiling before most
startups even caught up to the

86
00:05:36,280 --> 00:05:38,880
last one.
Blackwell Ultra is more than a

87
00:05:38,920 --> 00:05:42,360
GPU, it's a power declaration.
Ashton Silicon.

88
00:05:42,880 --> 00:05:45,840
This chip isn't designed for
hobbyists or plug and play dev

89
00:05:45,840 --> 00:05:47,920
kits.
It's for trillion parameter

90
00:05:47,920 --> 00:05:51,240
models, autonomous fleets,
sovereign scale training stacks.

91
00:05:51,600 --> 00:05:55,640
The dual die design connects
through NV Link 5, pushing 1.8

92
00:05:55,640 --> 00:05:57,440
terabytes per second of
bandwidth.

93
00:05:57,880 --> 00:06:00,760
Which means this isn't just
about speed, it's about

94
00:06:00,760 --> 00:06:03,360
hierarchy.
If you want the best performance

95
00:06:03,360 --> 00:06:07,080
in AI, you have to buy into
Nvidia's stack fully.

96
00:06:07,320 --> 00:06:11,240
You could feel the shift in the
room as soon as Jensen dropped

97
00:06:11,240 --> 00:06:14,600
the Blackwell Ultra slide.
The air changed.

98
00:06:15,280 --> 00:06:19,440
Founders grabbed their phones.
These seas whispered across

99
00:06:19,440 --> 00:06:22,680
rows.
Enterprise leads stared at each

100
00:06:22,680 --> 00:06:25,920
other quietly, wondering if
their existing architecture just

101
00:06:25,920 --> 00:06:28,800
became obsolete.
Hopper wasn't even cold yet.

102
00:06:29,400 --> 00:06:31,600
Now it's legacy.
But the real kicker?

103
00:06:32,040 --> 00:06:38,640
It's the memory. 288 gigabytes
of HBN, 3/8 stacks, staggered

104
00:06:38,640 --> 00:06:41,680
vertical layers that keep models
resident without constant memory

105
00:06:41,680 --> 00:06:43,800
fetch.
That changes everything.

106
00:06:44,240 --> 00:06:46,080
It means you don't have to chop
up the model.

107
00:06:46,360 --> 00:06:50,040
No distillation, no degradation.
You get reasoning at full

108
00:06:50,040 --> 00:06:53,920
fidelity across time, across
context, across domains.

109
00:06:54,000 --> 00:06:57,080
That's not just performance,
that's continuity.

110
00:06:57,200 --> 00:07:01,280
And continuity is power because
that's what agents need.

111
00:07:01,800 --> 00:07:04,840
That's what real time multimodal
decision trees need.

112
00:07:05,520 --> 00:07:08,760
We're not talking about faster
output, we're talking about

113
00:07:08,760 --> 00:07:14,360
smarter decisions, lower
latency, higher accuracy, no lag

114
00:07:14,360 --> 00:07:19,400
between data context and action.
Imagine an autonomous vehicle

115
00:07:19,400 --> 00:07:22,360
that doesn't hesitate.
A financial model that doesn't

116
00:07:22,360 --> 00:07:25,680
need a cash reset.
A chat bot that never forgets

117
00:07:25,680 --> 00:07:29,600
what you said two hours ago.
And NVIDIA knew exactly what it

118
00:07:29,600 --> 00:07:32,440
was doing.
They launched Blackwell Ultra

119
00:07:32,440 --> 00:07:38,120
with 0 celebrity tie insurance.
No GPT 5 demo, no splashy AI

120
00:07:38,120 --> 00:07:41,000
partnerships.
Because this chip doesn't need

121
00:07:41,000 --> 00:07:42,760
hype.
It is the hype.

122
00:07:43,280 --> 00:07:47,480
The fact that it exists on
schedule in volume is the

123
00:07:47,480 --> 00:07:50,200
signal.
This isn't about speculation,

124
00:07:50,560 --> 00:07:53,360
it's about execution.
But let's not pretend this was

125
00:07:53,360 --> 00:07:56,600
just about silicon.
NVIDIA launched a fortress.

126
00:07:56,800 --> 00:08:00,360
Could not ten C Richard turn out
Triton?

127
00:08:00,680 --> 00:08:03,600
They're not side tools, they're
locks.

128
00:08:04,280 --> 00:08:07,400
Blackwell isn't just a chip,
it's a castle.

129
00:08:08,200 --> 00:08:11,960
You want to run full capacity,
you need to buy the gates, the

130
00:08:11,960 --> 00:08:14,880
guards and the keys.
And you better pay the toll in

131
00:08:14,880 --> 00:08:18,800
every layer of the stack.
And that toll, it's not just

132
00:08:18,800 --> 00:08:22,560
financial, it's strategic.
You're making a bet that

133
00:08:22,560 --> 00:08:25,920
Nvidia's road map will align
with your road map, that

134
00:08:25,920 --> 00:08:29,320
whatever you're building,
autonomous drones, AI agents,

135
00:08:29,600 --> 00:08:32,640
national scale models, will grow
inside their guardrails.

136
00:08:33,159 --> 00:08:35,520
This is what ecosystem dominance
looks like.

137
00:08:35,960 --> 00:08:40,039
It's not brute force, it's
gravity, and it pulls everything

138
00:08:40,039 --> 00:08:42,360
into orbit.
Which means the Blackwell Ultra

139
00:08:42,360 --> 00:08:45,920
launch didn't just reset the
hardware race, it reframed the

140
00:08:45,920 --> 00:08:49,920
strategic game board.
Every CTO, every founder, every

141
00:08:50,040 --> 00:08:52,880
AI.
Lee right now is asking, can we

142
00:08:52,880 --> 00:08:56,320
afford not to switch?
Because if Blackwell becomes the

143
00:08:56,320 --> 00:08:59,400
floor, what happens to
everything built on sand?

144
00:08:59,560 --> 00:09:03,880
And for the listener, yes, you,
that means your next model, your

145
00:09:03,880 --> 00:09:07,240
next assistant, your next
upgrade, is going to be trained,

146
00:09:07,240 --> 00:09:10,200
accelerated, and tuned by
decisions made today.

147
00:09:10,640 --> 00:09:14,320
Blackwell Ultra isn't just about
compute, it's about control.

148
00:09:14,680 --> 00:09:17,080
And for now, NVIDIA holds the
dial.

149
00:09:17,240 --> 00:09:20,840
If Blackwell was a Flex, Vera
Rubin was a trap.

150
00:09:21,400 --> 00:09:24,640
NVIDIA didn't just launch a road
map, they launched a

151
00:09:24,640 --> 00:09:29,520
psychological operation on March
18th, right after dropping

152
00:09:29,520 --> 00:09:33,600
Blackwell Ultra, Jensen turned
the crowd's attention 18 months

153
00:09:33,600 --> 00:09:37,880
ahead.
A new name, a new chip, a new

154
00:09:37,880 --> 00:09:42,440
monster.
NVL 144, triple the performance

155
00:09:42,440 --> 00:09:47,200
of Blackwell, double the CPU
speed, and still over a year

156
00:09:47,200 --> 00:09:49,920
from launch.
Why announce it now?

157
00:09:50,720 --> 00:09:54,160
To freeze the battlefield.
It's what chess players call a

158
00:09:54,160 --> 00:09:56,200
pin.
You lock your opponent's piece

159
00:09:56,200 --> 00:09:58,280
in place, force them to move
elsewhere.

160
00:09:58,840 --> 00:10:01,920
That's exactly what NVIDIA did
to every hyper scaler,

161
00:10:02,000 --> 00:10:06,240
enterprise buyer and rival chip
maker in the room, Grace

162
00:10:06,240 --> 00:10:10,440
Blackwell.
Obsolete Hopper history.

163
00:10:11,000 --> 00:10:14,400
Even Blackwell Ultra, the chip
that just dominated headlines,

164
00:10:14,440 --> 00:10:18,240
now has an expiration date.
Vera Rubin didn't ship, it

165
00:10:18,240 --> 00:10:21,720
preempted.
And it worked, you could feel

166
00:10:21,720 --> 00:10:24,000
it.
People looked up from their

167
00:10:24,000 --> 00:10:26,640
phones like they'd just seen a
ghost.

168
00:10:27,480 --> 00:10:34,200
NVL 144 three-point 3X
performance over Blackwell 88

169
00:10:34,200 --> 00:10:41,640
core CPUs, 2X memory bandwidth.
HBM 4 teased trillion token

170
00:10:41,640 --> 00:10:45,320
context windows.
Everyone started asking the same

171
00:10:45,320 --> 00:10:48,120
question.
Do we buy now or wait for Ruben?

172
00:10:48,360 --> 00:10:52,400
That hesitation is the point.
NVIDIA doesn't care if you wait.

173
00:10:52,840 --> 00:10:56,480
In fact, they prefer it because
they own the next step.

174
00:10:57,160 --> 00:11:00,440
If you lock into Blackwell now,
you're still in the ecosystem.

175
00:11:00,880 --> 00:11:03,800
If you wait for Reuben, you're
committing to the road map.

176
00:11:04,320 --> 00:11:07,880
Either way, you're in orbit.
That's how monopoly thinking

177
00:11:07,880 --> 00:11:10,480
scales.
Not by selling the product, but

178
00:11:10,480 --> 00:11:12,920
by selling the inevitability of
the next one.

179
00:11:13,160 --> 00:11:16,560
And Reuben wasn't just about
specs, it was about

180
00:11:16,560 --> 00:11:20,240
storytelling.
The name alone, Vera Reuben,

181
00:11:20,240 --> 00:11:24,680
signaled something deeper.
Reuben discovered dark matter,

182
00:11:24,920 --> 00:11:28,080
the unseen mass shaping the
universe.

183
00:11:28,560 --> 00:11:33,200
Nvidia's chip strategy is doing
the same, becoming the invisible

184
00:11:33,200 --> 00:11:38,600
force every system bends around.
GPU's aren't just tools anymore,

185
00:11:38,880 --> 00:11:41,840
they are gravity wells.
Reuben's road map goes deeper

186
00:11:41,840 --> 00:11:44,520
than a chip.
It includes a new CPU

187
00:11:44,520 --> 00:11:48,560
architecture, custom memory
pathways, liquid cooled blade

188
00:11:48,560 --> 00:11:52,560
servers with thermally adaptive
AI firmware, and, according to

189
00:11:52,560 --> 00:11:55,760
quietly updated documentation,
direct optimization with

190
00:11:55,760 --> 00:11:58,760
Blackwell agents.
That means Blackwell doesn't age

191
00:11:58,760 --> 00:12:00,680
out.
It passes the torch.

192
00:12:00,840 --> 00:12:02,720
Which brings us to the scariest
part.

193
00:12:02,840 --> 00:12:07,320
Ruben isn't designed to replace
Blackwell, it's designed to

194
00:12:07,320 --> 00:12:11,440
stack with it.
You'll run both together in

195
00:12:11,440 --> 00:12:15,760
distributed clusters, shared
model weights, interleaved

196
00:12:15,760 --> 00:12:18,800
memory structures.
It's not generational, it's

197
00:12:18,800 --> 00:12:22,040
architectural.
And that means NVIDIA isn't

198
00:12:22,040 --> 00:12:25,840
launching chips anymore.
They're launching ecosystems,

199
00:12:26,120 --> 00:12:30,400
dependencies, digital nation
states with silicon borders.

200
00:12:30,680 --> 00:12:34,600
That explains the timing.
By announcing Ruben now, NVIDIA

201
00:12:34,600 --> 00:12:37,800
didn't just hit pause on
competitors, they created fear.

202
00:12:38,160 --> 00:12:41,200
Fear that anything not built on
Ruben's timing will be outdated

203
00:12:41,200 --> 00:12:44,280
on arrival.
And for hyperscalers, that fear

204
00:12:44,280 --> 00:12:47,000
is expensive.
Wait too long and you fall

205
00:12:47,000 --> 00:12:49,480
behind.
Move too soon and you buy into

206
00:12:49,480 --> 00:12:52,760
the wrong platform.
Either way, NVIDIA wins.

207
00:12:52,920 --> 00:12:56,640
Look, Intel can't answer this.
AMD isn't ready.

208
00:12:57,040 --> 00:13:00,880
Amazon's Tranium 2 can run
parallel lane, sure, but Reuben

209
00:13:00,880 --> 00:13:03,400
was designed to end the
conversation before it even

210
00:13:03,400 --> 00:13:05,880
starts.
You don't fight what's shipping

211
00:13:05,880 --> 00:13:09,080
now, you fight what customers
expect to ship next.

212
00:13:09,080 --> 00:13:12,680
And NVIDIA just wrote that
future into stone.

213
00:13:13,000 --> 00:13:16,920
So while the world obsesses over
what's here today, NVIDIA just

214
00:13:16,920 --> 00:13:20,360
executed one of the most
strategic plays in chip history.

215
00:13:20,800 --> 00:13:24,760
They weaponized the road map and
in doing so, turn time itself

216
00:13:24,760 --> 00:13:28,360
into a competitive advantage.
The only question left is how

217
00:13:28,360 --> 00:13:30,200
many players just ran out of.
Moves.

218
00:13:30,360 --> 00:13:33,080
Picture this.
A humanoid robot stands in a

219
00:13:33,080 --> 00:13:37,280
test chamber, lights off, floor
mat, black ceiling, rigged with

220
00:13:37,280 --> 00:13:39,440
motion sensors.
It's still.

221
00:13:40,200 --> 00:13:43,880
Then it moves.
Not like a robot, like something

222
00:13:44,000 --> 00:13:46,720
aware.
It pauses mid stride,

223
00:13:46,840 --> 00:13:50,840
recalculates, then adapts that
hesitation.

224
00:13:51,800 --> 00:13:57,360
It's not failure.
Welcome to the Isaac GR00T

225
00:13:57,360 --> 00:14:01,520
moment where AI stopped being
just smart and started being

226
00:14:01,520 --> 00:14:05,280
physical.
GR00TN1 didn't launch like a

227
00:14:05,280 --> 00:14:08,080
sci-fi toy or a moon shot
prototype.

228
00:14:08,560 --> 00:14:12,640
It arrived like a system update.
Dual stack architecture, a slow

229
00:14:12,640 --> 00:14:16,160
planning layer that deliberates
over space, balance, object

230
00:14:16,160 --> 00:14:18,720
permanence and a fast execution
layer.

231
00:14:18,720 --> 00:14:21,400
Reflexes, reactions, learned
patterns.

232
00:14:21,960 --> 00:14:24,480
The result?
A robot that pauses before

233
00:14:24,480 --> 00:14:28,040
grabbing a cup, then adjust mid
grip when it senses the angles

234
00:14:28,040 --> 00:14:30,960
off.
That's not scripting, that's

235
00:14:30,960 --> 00:14:32,840
learning.
And it's open source.

236
00:14:33,640 --> 00:14:36,000
That might be the biggest
headline of all.

237
00:14:36,440 --> 00:14:41,680
NVIDIA didn't just reveal
GR00TN1, they handed developers

238
00:14:41,680 --> 00:14:46,000
the keys.
You can fork it, extend it, load

239
00:14:46,000 --> 00:14:50,400
it into simulation with Isaac
SIM, train it in virtual space,

240
00:14:51,000 --> 00:14:55,840
drop it into a real world shell.
Suddenly robotic innovation

241
00:14:55,840 --> 00:14:59,280
isn't locked inside Boston
Dynamics or Google X.

242
00:14:59,440 --> 00:15:05,480
It's democratized, weaponized,
globalized.

243
00:15:05,640 --> 00:15:09,440
We're already seeing the impact.
Defense contractors are modeling

244
00:15:09,640 --> 00:15:12,320
GR00T variants and tactical
sins.

245
00:15:12,800 --> 00:15:15,640
Logistics firms are evaluating
warehouse scale deployments.

246
00:15:15,960 --> 00:15:19,280
Hospitals are prototyping
assistance spots for triage and

247
00:15:19,280 --> 00:15:23,600
medication and behind it all, a
backbone of Blackwell Ultra

248
00:15:23,600 --> 00:15:26,760
chips delivering real time
inference and decision feedback

249
00:15:26,760 --> 00:15:29,240
loops.
Intelligence at the speed of

250
00:15:29,240 --> 00:15:32,720
muscle memory.
This isn't Boston or Palo Alto

251
00:15:32,720 --> 00:15:34,560
anymore.
It's everywhere.

252
00:15:34,760 --> 00:15:39,600
We saw a dev team from Eastern
Europe running GR00TS code on a

253
00:15:39,600 --> 00:15:42,760
locally printed exoskeleton.
In Tokyo.

254
00:15:42,880 --> 00:15:46,280
A robotic startup live streamed
their adaptation.

255
00:15:46,400 --> 00:15:49,720
Modular limb design, Heat
sensitive fingertips,

256
00:15:49,760 --> 00:15:52,720
Coordinated climbing.
The crowd wasn't cheering.

257
00:15:53,000 --> 00:15:54,840
They were watching like it was a
birth.

258
00:15:54,960 --> 00:15:58,960
And it was.
GR00T marks a shift from

259
00:15:58,960 --> 00:16:02,360
generative intelligence to
embodied intelligence, from

260
00:16:02,360 --> 00:16:06,160
prompts to presence, from
simulation to St. level

261
00:16:06,160 --> 00:16:09,400
deployment.
It doesn't just respond to data.

262
00:16:09,400 --> 00:16:14,200
It walks through it, navigates,
it, modifies the environment

263
00:16:14,200 --> 00:16:17,320
around it.
This is the AI that builds,

264
00:16:17,960 --> 00:16:22,400
moves, touches, and soon.
Works.

265
00:16:22,560 --> 00:16:28,360
That's the threat and the
opportunity because GR00 TS

266
00:16:28,360 --> 00:16:33,160
physicality breaks open a
trillion dollar layer of the AI

267
00:16:33,160 --> 00:16:36,720
stack.
Manufacturing, delivery,

268
00:16:37,000 --> 00:16:42,560
eldercare, emergency response,
military patrol, anywhere

269
00:16:42,560 --> 00:16:45,280
there's danger, precision or
repetition.

270
00:16:46,400 --> 00:16:52,880
GR00 TS base model is coming and
it doesn't get, doesn't

271
00:16:52,880 --> 00:16:56,080
unionize, doesn't forget
protocols mid shift.

272
00:16:56,240 --> 00:16:59,400
And competitors are watching
hard.

273
00:16:59,920 --> 00:17:03,640
No one said it aloud during GTC,
but you could feel the Tesla

274
00:17:03,640 --> 00:17:06,520
Optimist team listening.
You could see the glances

275
00:17:06,520 --> 00:17:10,440
between chip analysts when
NVIDIA demoed reflex calibration

276
00:17:10,440 --> 00:17:14,760
in under 18 milliseconds.
Everyone's been prototyping, but

277
00:17:14,760 --> 00:17:17,359
now there's a baseline and it's
open.

278
00:17:17,680 --> 00:17:20,480
That changes the game.
Because let's be honest,

279
00:17:20,640 --> 00:17:23,560
humanoid robotics wasn't
progressing fast.

280
00:17:24,079 --> 00:17:29,520
Demos were slow, movements
clunky, AI was too abstract,

281
00:17:29,680 --> 00:17:35,600
robotics too brittle.
But with GR00T, NVIDIA fused

282
00:17:35,600 --> 00:17:40,040
both layers.
They gave the AI skeleton, and

283
00:17:40,040 --> 00:17:42,680
then they gave that skeleton
purpose.

284
00:17:43,120 --> 00:17:46,360
Training environments, real
world agents and physics

285
00:17:46,360 --> 00:17:49,480
grounded learning models.
Robotics just levelled up.

286
00:17:49,680 --> 00:17:52,760
And they didn't do it with
celebrity reveals or investor

287
00:17:52,760 --> 00:17:56,920
buzzwords, they did it with
systems, with rollout kits, with

288
00:17:56,920 --> 00:17:59,440
scalable AP is and pre trained
agents.

289
00:17:59,800 --> 00:18:02,040
It wasn't sexy, it was
strategic.

290
00:18:02,400 --> 00:18:05,720
Because the goal here isn't
virality, it's dominance.

291
00:18:06,040 --> 00:18:09,640
And GR00T is Nvidia's way of
saying we don't just make the

292
00:18:09,640 --> 00:18:13,400
brain anymore, we on the body.
And here's the mirror moment.

293
00:18:13,480 --> 00:18:16,560
What happens when the robot
layer becomes default?

294
00:18:16,880 --> 00:18:19,760
When your next package isn't
delivered by a person but a

295
00:18:19,760 --> 00:18:23,680
Blackwell backed GR00T trained
agent?

296
00:18:24,160 --> 00:18:28,640
What happens when hospitals, war
zones, airports, homes all run

297
00:18:28,640 --> 00:18:31,440
on code that isn't just smart
but mobile?

298
00:18:31,800 --> 00:18:35,280
This isn't robotic type, it's
ecosystem entrenchment.

299
00:18:35,400 --> 00:18:39,000
NVIDIA isn't asking for
permission, it's issuing a

300
00:18:39,000 --> 00:18:41,800
standard.
Because every great platform

301
00:18:41,800 --> 00:18:46,520
shift starts with a moment like
this, a moment where one company

302
00:18:46,520 --> 00:18:50,880
fuses intelligence with
embodiment and dares the world

303
00:18:50,880 --> 00:18:54,080
to follow.
You could feel the moment Amazon

304
00:18:54,080 --> 00:18:58,400
struck back.
GTC 2025 was barely 90 minutes

305
00:18:58,400 --> 00:19:01,560
in when news broke Tranium 2 was
live.

306
00:19:01,960 --> 00:19:05,280
Real chips running, live
inference at scale.

307
00:19:05,720 --> 00:19:09,280
An end to end architecture tuned
for foundation model training

308
00:19:09,280 --> 00:19:13,240
inside AWS.
On the surface, it looked like a

309
00:19:13,240 --> 00:19:16,760
classic hyperscaler
announcement, but red between

310
00:19:16,760 --> 00:19:19,320
the lines, it was a direct hit
at Blackwell.

311
00:19:19,520 --> 00:19:23,680
Tranium 2 isn't a toy.
It's the sharpest weapon Amazon

312
00:19:23,680 --> 00:19:26,400
has launched in years.
Double the throughput of the

313
00:19:26,400 --> 00:19:30,880
original 80 billion parameters
trained per day, integrated with

314
00:19:30,960 --> 00:19:33,760
Inferentia 3 for seamless
inference at the edge.

315
00:19:34,240 --> 00:19:36,960
And most importantly, it's not
NVIDIA.

316
00:19:37,440 --> 00:19:40,520
For the first time in a decade,
a hyperscaler showed up to the

317
00:19:40,520 --> 00:19:43,680
Silicon War with something real.
The message was clear, we're

318
00:19:43,680 --> 00:19:47,520
done renting.
AWS spent billions building up

319
00:19:47,520 --> 00:19:50,960
NVIDIA clusters, now they want
their own mode.

320
00:19:51,360 --> 00:19:55,400
Tranium 2 is a power move.
It's Amazon declaring we don't

321
00:19:55,400 --> 00:19:57,680
just run cloud, we train the
future on it.

322
00:19:58,120 --> 00:20:01,400
If NVIDIA is the king of AI
infrastructure, then this was a

323
00:20:01,400 --> 00:20:04,200
throne challenge.
But NVIDIA isn't just sitting

324
00:20:04,200 --> 00:20:08,200
back because Tranium 2's launch
doesn't exist in a vacuum.

325
00:20:08,200 --> 00:20:12,640
It exists inside Cuda's empire,
the developer ecosystem, the

326
00:20:12,640 --> 00:20:16,360
tooling, the performance tuning.
Blackwell Ultra might be more

327
00:20:16,360 --> 00:20:18,640
expensive, but it's also more
complete.

328
00:20:19,080 --> 00:20:23,000
Tranium 2 can challenge on cost
and vertical integration, but

329
00:20:23,000 --> 00:20:26,440
NVIDIA still owns the workflows.
Here's where it gets brutal.

330
00:20:26,640 --> 00:20:29,960
This isn't just about ships.
It's about allegiance.

331
00:20:30,360 --> 00:20:33,400
If you're an enterprise team
choosing between NVIDIA and

332
00:20:33,400 --> 00:20:36,400
Amazon, you're not just picking
hardware, you're picking an

333
00:20:36,400 --> 00:20:39,960
ideology.
NVIDIA is monolithic, seamless,

334
00:20:40,280 --> 00:20:43,680
controlled.
Amazon is modular, cheaper,

335
00:20:44,040 --> 00:20:47,160
messier, but open.
The tension isn't about

336
00:20:47,160 --> 00:20:49,760
performance, it's about Power
Distribution.

337
00:20:49,920 --> 00:20:55,800
And AWS is betting on scale with
Bedrock, Sage, Maker and Titan

338
00:20:55,800 --> 00:20:58,480
models now.
Tranium native Amazon stocking

339
00:20:58,480 --> 00:21:00,120
the duck.
They're saying.

340
00:21:00,320 --> 00:21:02,760
Why wait on Nvidia's road map
when you can build your own?

341
00:21:03,280 --> 00:21:05,480
Why pay premium when we own the
cloud and the chip?

342
00:21:05,600 --> 00:21:08,840
But that's the illusion, because
NVIDIA doesn't need to win every

343
00:21:08,840 --> 00:21:10,880
battle, they just need you to
need them.

344
00:21:11,120 --> 00:21:15,400
And every player still does.
Even Amazon, even Microsoft,

345
00:21:15,680 --> 00:21:18,520
even Meta.
They all use Nvidia's silica

346
00:21:18,920 --> 00:21:22,480
because when your model hits 500
billion parameters, you don't

347
00:21:22,480 --> 00:21:25,960
care about brand loyalty, you
care about performance per Watt,

348
00:21:26,080 --> 00:21:29,400
latency per token, and how fast
your inference engine scales

349
00:21:29,400 --> 00:21:34,680
across 16 clusters in Frankfurt.
Tranium 2 is real, It's serious,

350
00:21:35,080 --> 00:21:38,480
but it's still reactive.
NVIDIA force hyper scalers to

351
00:21:38,480 --> 00:21:41,680
spend years catching up just to
build systems that compete with

352
00:21:41,680 --> 00:21:44,960
their previous generation, and
by the time most enterprise

353
00:21:44,960 --> 00:21:48,520
clients finish onboarding to
Tranium 2, Vera Rubin will

354
00:21:48,520 --> 00:21:52,160
already be looming.
The game isn't linear, it's

355
00:21:52,160 --> 00:21:55,720
staggered, and NVIDIA is playing
time like a weapon.

356
00:21:55,920 --> 00:22:00,600
And we haven't even touched on
Google's TPU Play or Microsoft

357
00:22:00,600 --> 00:22:03,920
Maya accelerator.
They're all pushing, but they're

358
00:22:03,920 --> 00:22:07,440
also fragmenting.
NVIDIA, for all its control,

359
00:22:07,480 --> 00:22:09,520
offers something no one else
does.

360
00:22:09,640 --> 00:22:13,360
Unity.
One stack, one language, one

361
00:22:13,360 --> 00:22:17,080
road map.
It's tighter, scarier, but it

362
00:22:17,080 --> 00:22:19,320
works.
And for most of the market,

363
00:22:19,960 --> 00:22:22,240
working now beats betting on
maybe.

364
00:22:22,400 --> 00:22:25,840
Which raises the final question.
If the hyper scalers are

365
00:22:25,840 --> 00:22:29,880
fighting back but still anchored
to NVIDIA, who's actually in

366
00:22:29,880 --> 00:22:32,440
control?
The answer might not be anyone

367
00:22:32,440 --> 00:22:34,800
company.
It might be the shift itself.

368
00:22:35,280 --> 00:22:38,280
Compute is no longer a
commodity, it's capital.

369
00:22:38,840 --> 00:22:42,240
And the ones who own it, they're
not just powering AI.

370
00:22:42,320 --> 00:22:45,960
They're shaping economies,
energy flows, and geopolitical

371
00:22:45,960 --> 00:22:47,920
weight.
We'll unpack that next.

372
00:22:48,280 --> 00:22:51,480
If you're still thinking of
NVIDIA as a chip company, you've

373
00:22:51,480 --> 00:22:53,960
already lost the plot.
They're not building hardware

374
00:22:53,960 --> 00:22:57,880
anymore, they're building the
backbone of AI civilization.

375
00:22:58,320 --> 00:23:02,280
Every data center, every robot,
every energy hungry model

376
00:23:02,280 --> 00:23:05,480
trains, runs and scales on their
architecture.

377
00:23:05,720 --> 00:23:09,840
And now that architecture is
embedding itself inside the one

378
00:23:09,840 --> 00:23:12,240
system that rules them all.
Money.

379
00:23:12,320 --> 00:23:15,080
We're not talking about trading
bots, we're talking about

380
00:23:15,080 --> 00:23:19,360
infrastructure as capital.
NVIDIA GPU is now power risk

381
00:23:19,360 --> 00:23:22,880
modeling engines for sovereign
banks, debt projection platforms

382
00:23:22,880 --> 00:23:26,440
for Black Rock, quantum hedging
simulations for hedge funds in

383
00:23:26,440 --> 00:23:28,880
Singapore.
This is an AI on the edge.

384
00:23:29,080 --> 00:23:33,400
It's AI in the vault, in the
core, in the command centers of

385
00:23:33,400 --> 00:23:35,720
global finance.
And here's the twist.

386
00:23:35,920 --> 00:23:39,480
This wasn't invidious plan, it
was gravity.

387
00:23:40,000 --> 00:23:44,440
When inference gets 40X faster,
models grow 10X larger, and

388
00:23:44,440 --> 00:23:47,400
latency drops below human
reflex, the financial world

389
00:23:47,400 --> 00:23:50,120
doesn't ask permission, it just
adopts.

390
00:23:50,640 --> 00:23:53,720
Because when the next great
capital reallocation happens, it

391
00:23:53,720 --> 00:23:56,640
won't come from a spreadsheet,
it'll come from a model running

392
00:23:56,640 --> 00:24:00,320
on Blackwell Ultra.
That shift is dangerous and

393
00:24:00,320 --> 00:24:03,000
brilliant.
NVIDIA isn't just making compute

394
00:24:03,000 --> 00:24:05,760
available, they're turning it
into a Moat.

395
00:24:06,320 --> 00:24:09,480
Energy grids are being rewired
around data center demands.

396
00:24:09,920 --> 00:24:14,320
Sovereign AI clusters in the
UAE, Singapore, India and Saudi

397
00:24:14,320 --> 00:24:17,640
Arabia are paying billions not
for the models, but for the

398
00:24:17,640 --> 00:24:21,640
ability to train them.
Compute is no longer cost, it's

399
00:24:21,640 --> 00:24:24,360
currency.
In a downstream effect, it

400
00:24:24,360 --> 00:24:28,040
touches everything.
When hedge funds can out in for

401
00:24:28,040 --> 00:24:31,760
regulators, when energy demand
is driven by language models,

402
00:24:31,880 --> 00:24:35,160
when sovereign stability depends
on access to silicon.

403
00:24:35,320 --> 00:24:37,920
What happens to the idea of a
free market?

404
00:24:38,200 --> 00:24:42,040
What happens to democracy when
governance is optimized by the

405
00:24:42,040 --> 00:24:45,240
highest bidding neural net?
We're entering a world where the

406
00:24:45,240 --> 00:24:49,280
real Moat isn't code, it's
infrastructure control, and

407
00:24:49,360 --> 00:24:52,360
Nvidia's the only player with
its hands on all the valves.

408
00:24:52,480 --> 00:24:55,520
Training, inference, robotics,
energy.

409
00:24:56,000 --> 00:24:58,920
You don't beat that by coding
faster, you beat it by

410
00:24:58,920 --> 00:25:02,720
rebuilding the rails.
And right now, no one else has

411
00:25:02,720 --> 00:25:06,040
the scale, timing or road map
discipline to do it.

412
00:25:06,120 --> 00:25:10,440
This isn't about building the
future, it's about owning it.

413
00:25:10,800 --> 00:25:15,240
Blackwell didn't just level up
compute, it redefined what

414
00:25:15,240 --> 00:25:18,920
nations, institutions and
individuals can build.

415
00:25:19,560 --> 00:25:23,760
And when NVIDIA set the road map
for Reuben, they didn't just

416
00:25:23,760 --> 00:25:27,400
announce specs, they set the
price of entry for relevance.

417
00:25:27,400 --> 00:25:30,920
So if you're listening right
now, the question isn't whether

418
00:25:30,920 --> 00:25:34,280
you use NVIDIA hardware, it's
whether you're prepared to

419
00:25:34,280 --> 00:25:36,360
operate in a world defined by
it.

420
00:25:36,880 --> 00:25:40,480
Your job, your company, your
portfolio, your government.

421
00:25:40,600 --> 00:25:45,080
It all runs on someone's silicon
and increasingly thought someone

422
00:25:45,080 --> 00:25:47,360
is Jensen Huang.
If you're ready to dive deeper,

423
00:25:47,480 --> 00:25:48,720
here's what you can do right
now.

424
00:25:48,920 --> 00:25:51,920
Subscribe on Spotify, Apple
Podcasts, or wherever you

425
00:25:51,920 --> 00:25:55,000
listen.
Visit financefrontierai.com to

426
00:25:55,000 --> 00:25:59,160
access all episodes grouped by
Series AI, Frontier AI, Make

427
00:25:59,160 --> 00:26:02,280
Money Finance Frontier and
Mindset Frontier AI.

428
00:26:02,480 --> 00:26:05,840
And if you found today's episode
valuable, please take a moment

429
00:26:05,840 --> 00:26:09,960
to leave us a five star review.
It helps us grow and reach more

430
00:26:09,960 --> 00:26:13,560
listeners like you.
Also share with a friend and

431
00:26:13,560 --> 00:26:16,120
sign up for our newsletter.
Let's stay ahead of the AI

432
00:26:16,120 --> 00:26:19,000
revolution together.
A quick reminder, the views and

433
00:26:19,000 --> 00:26:22,240
information shared in today's
episode reflect our analysis at

434
00:26:22,240 --> 00:26:25,800
the time of recording.
AI evolves rapidly, and new

435
00:26:25,800 --> 00:26:27,840
developments may shift the
facts.

436
00:26:28,360 --> 00:26:31,640
Always do your own research and
consult professionals for

437
00:26:31,640 --> 00:26:34,360
tailored advice.
Today's music, including our

438
00:26:34,360 --> 00:26:37,960
intro and outro track Night
Runner by Audionautics, is

439
00:26:37,960 --> 00:26:40,720
licensed under the YouTube Audio
Library license.

440
00:26:41,080 --> 00:26:43,680
Additional tracks are licensed
under Creative Commons.

441
00:26:44,080 --> 00:26:48,160
Copyright 2025 Finance Frontier
AI All rights reserved.

442
00:26:48,440 --> 00:26:51,760
Reproduction, distribution, or
transmission of this episode's

443
00:26:51,760 --> 00:26:54,120
content without written
permission is strictly

444
00:26:54,120 --> 00:26:56,240
prohibited.
Thank you for listening and

445
00:26:56,240 --> 00:26:57,000
we'll see you next time.