NVIDIA: From 1993 to the AI Age | A SleepWise Story

1
00:00:00,120 --> 00:00:06,400
Tonight we drift back to the
early 1990s, when computers

2
00:00:06,400 --> 00:00:13,200
still felt like quiet furniture,
beige towers under desks, soft

3
00:00:13,200 --> 00:00:19,240
fans turning, a monitor glowing
with a kind of patient effort.

4
00:00:19,720 --> 00:00:24,560
The world on those screens was
mostly flat, made of windows and

5
00:00:24,640 --> 00:00:30,280
icons, and yet you could already
sense a hunger for depth, for

6
00:00:30,280 --> 00:00:34,480
movement, for something that
looked more like the world

7
00:00:34,480 --> 00:00:39,440
outside the room.
In cinemas, new computer imagery

8
00:00:39,440 --> 00:00:45,320
was beginning to shimmer and
convince dinosaurs breathing in

9
00:00:45,320 --> 00:00:51,400
moonlit rain, impossible scenes
made believable by patient

10
00:00:51,400 --> 00:00:55,080
calculation.
In arcades and living rooms,

11
00:00:55,560 --> 00:01:01,040
games were leaning toward
illusion corridors that seemed

12
00:01:01,040 --> 00:01:06,520
to stretch away creatures that
lurked in shadow, textures that

13
00:01:06,520 --> 00:01:10,640
hinted at stone and metal even
if they were made of clever

14
00:01:10,640 --> 00:01:14,440
pixels.
On APC, though, the central

15
00:01:14,440 --> 00:01:20,960
processor did almost everything,
and it was tired, asked to

16
00:01:20,960 --> 00:01:28,480
calculate logic, sound input,
and then, on top of it all, the

17
00:01:28,480 --> 00:01:32,720
heavy math of drawing A3
dimensional scene.

18
00:01:33,200 --> 00:01:38,200
The result was often a kind of
compromise, A stutter between

19
00:01:38,200 --> 00:01:44,240
frames, a world that wanted to
be smooth but could only manage

20
00:01:44,480 --> 00:01:49,680
a rough sketch of itself.
And in that moment, in Silicon

21
00:01:49,680 --> 00:01:54,560
Valley air that smelled faintly
of solder and coffee, three

22
00:01:54,560 --> 00:01:58,600
engineers began circling a
simple idea.

23
00:01:59,320 --> 00:02:04,800
What if the work of seeing could
be separated from the work of

24
00:02:04,800 --> 00:02:08,280
thinking?
What if there were a dedicated

25
00:02:08,280 --> 00:02:13,680
chip, built for repetition and
rhythm, doing many small visual

26
00:02:13,680 --> 00:02:19,280
calculations at once, so the
main processor could breathe?

27
00:02:19,800 --> 00:02:27,680
And the images could flow.
In 1993, Jensen Huang, Chris

28
00:02:27,680 --> 00:02:33,360
Malakowski and Curtis Prim
started a new company with that

29
00:02:33,360 --> 00:02:37,520
question tucked inside it.
And they gave it a name that

30
00:02:37,520 --> 00:02:41,920
sounded like a bright place
beyond the horizon.

31
00:02:42,680 --> 00:02:48,520
NVIDIA the story says their
first serious conversations

32
00:02:48,520 --> 00:02:53,920
happened in an ordinary diner
over breakfast, the kind of

33
00:02:53,920 --> 00:02:58,760
place where the coffee keeps
coming and big ideas can feel

34
00:02:58,760 --> 00:03:02,840
strangely practical.
They were not alone in chasing

35
00:03:02,840 --> 00:03:04,640
the future.
Not at all.

36
00:03:05,240 --> 00:03:09,880
Dozens of companies were
building graphics chips, each

37
00:03:09,880 --> 00:03:15,320
trying to become the one that
mattered, each racing to survive

38
00:03:15,320 --> 00:03:20,240
the next product cycle.
So from the beginning, there was

39
00:03:20,240 --> 00:03:24,880
pressure in the background, like
a steady clock.

40
00:03:25,600 --> 00:03:31,040
Deadlines, prototypes, decisions
that could not be undone once

41
00:03:31,040 --> 00:03:35,120
the silicon was made.
But there is a particular kind

42
00:03:35,120 --> 00:03:39,560
of calm that comes from being
early, from committing to

43
00:03:39,560 --> 00:03:45,520
something before it is obvious,
before it has a market with a

44
00:03:45,520 --> 00:03:49,800
neat label and a confident
number attached to it.

45
00:03:50,160 --> 00:03:52,960
To night.
As the world slows.

46
00:03:53,440 --> 00:03:57,560
We'll follow that quiet
commitment from the first

47
00:03:57,560 --> 00:04:03,560
uncertain designs to the moment
a new kind of processor began to

48
00:04:03,560 --> 00:04:08,400
lift the weight of images from
the shoulders of the CPU.

49
00:04:08,680 --> 00:04:12,720
We won't rush, and you don't
have to remember a thing.

50
00:04:13,120 --> 00:04:18,360
Just listen for the gentle hum
of machines and the way a few

51
00:04:18,360 --> 00:04:24,960
patient decisions can, overtime,
reshape what the world is able

52
00:04:24,960 --> 00:04:28,960
to imagine.
In the beginning, the future did

53
00:04:28,960 --> 00:04:34,640
not arrive as a clean, confident
breakthrough, but as a messy

54
00:04:34,640 --> 00:04:40,880
pile of possibilities.
Sound, video, controller ports

55
00:04:41,080 --> 00:04:45,720
and graphics all tugging at the
same small chip.

56
00:04:46,080 --> 00:04:51,440
Nvidia's first attempt in the
mid 1990s tried to be everything

57
00:04:51,440 --> 00:04:58,280
at once, like a multi tool
carried into a storm, and for a

58
00:04:58,280 --> 00:05:01,920
moment it felt clever, even
elegant on paper.

59
00:05:02,280 --> 00:05:07,000
But the world outside the lab
had its own preferences, and

60
00:05:07,000 --> 00:05:10,840
software developers were
gathering around new standards,

61
00:05:11,480 --> 00:05:17,200
choosing the simplest shapes to
build their 3D worlds triangles,

62
00:05:17,640 --> 00:05:22,520
predictable and fast.
When the chip arrived, the

63
00:05:22,520 --> 00:05:29,240
market did what markets often do
It shrugged, it moved on, and it

64
00:05:29,240 --> 00:05:33,320
returned boxes that had once
seemed like a promise.

65
00:05:33,600 --> 00:05:38,520
Inside the company, the room
grew quieter, not with despair

66
00:05:39,040 --> 00:05:45,440
but with focus, because a young
team learns quickly that silicon

67
00:05:45,440 --> 00:05:52,480
is unforgiving, and there is no
undo once the design is baked.

68
00:05:52,960 --> 00:05:56,560
They were surrounded by
competitors, too, a whole

69
00:05:56,560 --> 00:06:01,360
crowded field of graphics
companies, each racing to make

70
00:06:01,360 --> 00:06:06,400
the next card, each hoping to be
the name printed on the box.

71
00:06:06,640 --> 00:06:11,560
It was the kind of era where
magazines compared frame rates

72
00:06:11,560 --> 00:06:16,840
the way people once compared
engines, and gamers could feel

73
00:06:16,840 --> 00:06:21,280
the difference between smooth
motion and a world that

74
00:06:21,280 --> 00:06:25,680
stumbled.
So NVIDIA narrowed its gaze,

75
00:06:25,920 --> 00:06:31,040
letting go of the octopus dream
and choosing A simpler goal.

76
00:06:31,760 --> 00:06:37,760
Make one thing work beautifully
and make it arrive on time.

77
00:06:38,120 --> 00:06:44,240
In those months, time became a
texture, like sand in a pocket.

78
00:06:44,840 --> 00:06:49,360
Because there is a special
pressure when you know you have

79
00:06:49,360 --> 00:06:53,720
only one more product cycle to
earn your right to continue.

80
00:06:54,120 --> 00:06:59,880
And yet there was also craft
engineers running long

81
00:06:59,880 --> 00:07:05,600
simulations, testing a design
before it ever touched a

82
00:07:05,600 --> 00:07:10,280
factory, listening for errors
the way you might listen for a

83
00:07:10,280 --> 00:07:15,360
Creek in an old house at night.
When the new chip finally

84
00:07:15,360 --> 00:07:20,920
landed, it worked, and it sold,
and it did what it was meant to

85
00:07:20,920 --> 00:07:23,960
do.
It helped images move more

86
00:07:23,960 --> 00:07:28,160
smoothly, and it gave the
company another morning.

87
00:07:28,560 --> 00:07:34,360
Tonight, we can hold that moment
gently, because it's not triumph

88
00:07:34,360 --> 00:07:41,520
or failure that defines a long
story, but endurance, the steady

89
00:07:41,520 --> 00:07:46,200
choice to try again with clearer
eyes.

90
00:07:46,520 --> 00:07:50,960
It was around the time when
games like Quake were teaching

91
00:07:50,960 --> 00:07:56,640
the eye to expect depth, and
when each new graphics card felt

92
00:07:56,640 --> 00:08:02,400
like a new pair of glasses, the
same world, but sharper.

93
00:08:02,880 --> 00:08:07,400
The chip's name mattered less
than its timing, but people

94
00:08:07,400 --> 00:08:12,520
remember it as one of the
Rivacards, an unromantic word

95
00:08:12,520 --> 00:08:16,240
that none the less became a
bridge back to stability.

96
00:08:16,560 --> 00:08:20,520
And once the bridge was built,
the pace changed.

97
00:08:20,920 --> 00:08:26,840
Because survival has a way of
sharpening decisions, the team

98
00:08:27,080 --> 00:08:32,760
could now iterate faster, learn
faster, and keep the fan hum

99
00:08:32,760 --> 00:08:37,679
steady through the night.
With stability came a quiet kind

100
00:08:37,679 --> 00:08:42,240
of confidence, and with
confidence came permission to

101
00:08:42,240 --> 00:08:46,120
think bigger.
Not louder, just deeper.

102
00:08:46,400 --> 00:08:54,120
In 1999, NVIDIA introduced a new
card under a new name, G Force,

103
00:08:54,680 --> 00:08:58,960
and with it they offered a
phrase that would stick the

104
00:08:58,960 --> 00:09:03,440
graphics processing unit, or
GPU.

105
00:09:03,720 --> 00:09:09,080
It was, in a sense, a small act
of renaming the world.

106
00:09:09,640 --> 00:09:15,520
Because until then, graphics
were treated as an accessory, a

107
00:09:15,520 --> 00:09:20,560
helper, something the CPU could
manage if it had enough time.

108
00:09:20,880 --> 00:09:26,720
But time was exactly what games
did not have, because a moving

109
00:09:26,720 --> 00:09:31,080
world cannot pause while the
processor catches its breath,

110
00:09:31,800 --> 00:09:34,960
and the eye will always notice
the hesitation.

111
00:09:35,360 --> 00:09:40,360
So the GPU was designed to take
on the repeated work of

112
00:09:40,360 --> 00:09:45,880
transforming 3D shapes,
calculating light, preparing

113
00:09:45,880 --> 00:09:52,520
triangles and pushing pixels, A
specialized rhythm done in

114
00:09:52,520 --> 00:09:57,160
parallel, like many hands
folding the same kind of paper.

115
00:09:57,600 --> 00:10:02,120
This was not only about beauty,
though beauty mattered, but

116
00:10:02,120 --> 00:10:06,000
about freeing the CPU to do what
it did best.

117
00:10:06,520 --> 00:10:12,720
Logic decisions, the subtle
choreography of a game's rules.

118
00:10:13,000 --> 00:10:17,240
If you played in those years,
you may remember the feeling of

119
00:10:17,240 --> 00:10:22,400
a new graphics card, like a
secret upgrade to reality

120
00:10:23,240 --> 00:10:29,200
corridors in Quake 3 that
suddenly felt smoother, arenas

121
00:10:29,200 --> 00:10:34,640
that shimmered without tearing.
The GPU also carried something

122
00:10:34,640 --> 00:10:39,560
else, almost hidden at first.
It was becoming more

123
00:10:39,560 --> 00:10:45,080
programmable, less like a fixed
tool and more like a small

124
00:10:45,080 --> 00:10:47,800
machine that could learn new
tricks.

125
00:10:48,160 --> 00:10:53,080
That shift would take years to
fully unfold, but you can

126
00:10:53,080 --> 00:10:58,080
already sense it here, like a
door cracked open in a quiet

127
00:10:58,080 --> 00:11:02,320
hallway.
Around this same time, Nvidia's

128
00:11:02,320 --> 00:11:07,360
chips began to appear not only
in PCs but in the broader

129
00:11:07,360 --> 00:11:12,480
culture of gaming hardware, as
console makers looked for

130
00:11:12,480 --> 00:11:17,920
graphics power that felt both
affordable and advanced.

131
00:11:18,320 --> 00:11:25,360
A console is, in its own way, a
promise of consistency, the same

132
00:11:25,360 --> 00:11:30,840
experience for everyone, the
same frames, the same light, the

133
00:11:30,840 --> 00:11:36,320
same reliable hum.
And NVIDIA wanted to be part of

134
00:11:36,320 --> 00:11:40,360
that steadiness.
So the company kept moving,

135
00:11:40,480 --> 00:11:45,880
generation by generation, each
one a little faster, a little

136
00:11:45,880 --> 00:11:51,320
more capable, each one teaching
developers to expect more from

137
00:11:51,320 --> 00:11:56,640
what a screen could hold.
And beneath it all was a calm

138
00:11:56,640 --> 00:12:02,520
idea that would later become
enormous, that some work is

139
00:12:02,520 --> 00:12:07,480
better done by many simple
processors acting together,

140
00:12:08,000 --> 00:12:12,120
rather than one processor doing
everything alone.

141
00:12:12,440 --> 00:12:18,040
When NVIDIA called it a GPU, it
was partly a statement of

142
00:12:18,040 --> 00:12:22,520
ambition.
Not just a faster graphics chip,

143
00:12:23,160 --> 00:12:28,080
but a new class of processor
built to do a certain kind of

144
00:12:28,080 --> 00:12:33,000
math with steady repetition.
It was the beginning of a long

145
00:12:33,000 --> 00:12:39,560
relationship between silicon and
imagination, where artists and

146
00:12:39,760 --> 00:12:46,280
engineers met in the middle, one
drawing what they wished to see,

147
00:12:46,800 --> 00:12:51,280
the other finding a way to
compute it in time.

148
00:12:51,720 --> 00:12:57,080
In the early 2000's, the world
of games grew more cinematic,

149
00:12:57,640 --> 00:13:02,920
not because stories became
louder, but because light and

150
00:13:02,920 --> 00:13:08,400
texture became more convincing.
A wall could look like stone

151
00:13:08,400 --> 00:13:12,640
instead of a pattern, a sky
could fade gently instead of

152
00:13:12,640 --> 00:13:18,280
switching colors, and shadows
could suggest distance the way

153
00:13:18,280 --> 00:13:22,240
evening does.
This happened through a quiet

154
00:13:22,240 --> 00:13:27,320
evolution called
programmability, small pieces of

155
00:13:27,320 --> 00:13:33,040
code called shaders that told
the GPU how to color a pixel or

156
00:13:33,040 --> 00:13:37,200
bend a surface like instructions
whispered to a brush.

157
00:13:37,600 --> 00:13:43,520
For developers, this was a new
palette, and for players it was

158
00:13:43,520 --> 00:13:49,280
a gradual shift in expectation,
as if the mind learned to

159
00:13:49,280 --> 00:13:52,640
believe the screen a little more
each year.

160
00:13:53,040 --> 00:13:58,200
NVIDIA and its rivals pushed
each other forward, and

161
00:13:58,200 --> 00:14:02,360
generation by generation, the
numbers climbed.

162
00:14:03,000 --> 00:14:08,560
More transistors, more memory
bandwidth, more parallel work

163
00:14:08,760 --> 00:14:12,480
per heartbeat.
The cultural world around it

164
00:14:12,480 --> 00:14:17,400
changed, too, because gaming
moved from basements to

165
00:14:17,400 --> 00:14:21,880
bedrooms, from a private hobby
to a shared language.

166
00:14:22,280 --> 00:14:27,840
You might remember land parties,
the glow of monitors in a dim

167
00:14:27,840 --> 00:14:33,360
room, the soft clatter of
keyboards and the quick laughter

168
00:14:33,360 --> 00:14:38,320
between rounds.
Later came broadband, and then

169
00:14:38,320 --> 00:14:43,040
the strange comfort of watching
someone else play, as if the act

170
00:14:43,040 --> 00:14:47,800
of play could become a story
you'd listen to while doing

171
00:14:47,800 --> 00:14:51,400
something else.
Platforms like YouTube and

172
00:14:51,400 --> 00:14:56,040
Twitch would eventually make
that watching feel effortless,

173
00:14:56,640 --> 00:15:00,840
but even before they were
household names, the need was

174
00:15:00,840 --> 00:15:06,720
already there, the desire to
share a moment in real time.

175
00:15:07,080 --> 00:15:13,520
A GPU helped with that as well,
because the same parallel power

176
00:15:13,520 --> 00:15:17,920
that drew a game could also help
compress it into a stream,

177
00:15:18,480 --> 00:15:22,720
turning private pixels into a
public broadcast.

178
00:15:23,080 --> 00:15:29,000
And slowly, without any single
turning point, games became

179
00:15:29,000 --> 00:15:36,000
arenas and arenas became events,
and events became a kind of

180
00:15:36,000 --> 00:15:41,280
modern gathering, lit by screens
instead of stadium lights.

181
00:15:41,720 --> 00:15:46,440
Through it all, NVIDIA kept
refining the craft of fast

182
00:15:46,440 --> 00:15:50,640
drawing, building not only chips
but a relationship with

183
00:15:50,640 --> 00:15:57,640
developers, tools, drivers and
the gentle promise that next

184
00:15:57,640 --> 00:16:01,680
year's hardware would make new
ideas possible.

185
00:16:01,960 --> 00:16:07,600
In that era, names like DirectX
and Open GL floated through

186
00:16:07,600 --> 00:16:12,880
conversations not as dry
acronyms, but as common roads

187
00:16:12,880 --> 00:16:17,960
that everyone agreed to travel
so games could run on many

188
00:16:17,960 --> 00:16:21,720
machines.
Nvidia's chips found their way

189
00:16:21,720 --> 00:16:27,520
into consoles as well, and if
you ever held an original Xbox

190
00:16:27,520 --> 00:16:32,760
controller, heavy and sure in
the hands, there was a bit of

191
00:16:32,760 --> 00:16:36,760
Nvidia's work helping that world
feel smooth.

192
00:16:37,240 --> 00:16:42,000
Even when you didn't know the
chip's name, you could sense its

193
00:16:42,000 --> 00:16:48,240
effect, The steadier frame rate,
the richer scenes, the feeling

194
00:16:48,240 --> 00:16:51,600
that the machine was keeping up
with your intention.

195
00:16:51,960 --> 00:16:57,160
All of this was still graphics
officially, yet it carried a

196
00:16:57,160 --> 00:17:02,960
hidden lesson that parallel
computation, done well, could

197
00:17:02,960 --> 00:17:07,000
make complex experiences feel
simple.

198
00:17:07,280 --> 00:17:12,680
At some point, usually in the
middle of a long technology

199
00:17:12,680 --> 00:17:18,280
story, people begin to use a
tool for something it was not

200
00:17:18,280 --> 00:17:22,319
built to do.
They do it quietly at first, in

201
00:17:22,319 --> 00:17:27,640
labs and late nights, not to
rebel, but to see what happens

202
00:17:27,640 --> 00:17:32,840
when you turn a familiar object
slightly and look at it from

203
00:17:32,840 --> 00:17:36,880
another angle.
A GPU, after all, was already a

204
00:17:36,880 --> 00:17:39,720
machine for repeating simple
math very quickly.

205
00:17:40,160 --> 00:17:44,440
And many scientific problems are
also made of repeated math,

206
00:17:44,720 --> 00:17:49,680
stacked like waves.
So researchers began to wonder,

207
00:17:50,160 --> 00:17:55,800
if a GPU can apply light to
millions of pixels, could it

208
00:17:55,800 --> 00:18:00,680
also apply equations to millions
of data points?

209
00:18:01,040 --> 00:18:05,960
In the early days, this was
awkward, like writing poetry

210
00:18:05,960 --> 00:18:11,760
with a screwdriver, because
programming a GPU meant

211
00:18:11,760 --> 00:18:15,960
pretending your problem was a
graphics problem.

212
00:18:16,320 --> 00:18:20,720
You would encode numbers as
colors, calculations as

213
00:18:20,720 --> 00:18:25,120
textures, and you would work
through the constraints with

214
00:18:25,120 --> 00:18:29,680
patient creativity.
But when it worked, it felt like

215
00:18:29,680 --> 00:18:35,000
a small miracle.
A simulation that once took days

216
00:18:35,000 --> 00:18:39,520
could finish in hours, and the
computer seemed to breathe in a

217
00:18:39,520 --> 00:18:43,040
new way.
NVIDIA noticed these

218
00:18:43,040 --> 00:18:48,600
experiments, and instead of
dismissing them as edge cases,

219
00:18:49,120 --> 00:18:54,120
they listened.
Because listening is often how a

220
00:18:54,120 --> 00:19:00,280
company discovers its next life,
the question became sharper,

221
00:19:01,080 --> 00:19:06,920
What if we made it normal to
program the GPU directly, not as

222
00:19:06,920 --> 00:19:11,000
a graphics trick, but as a
computing platform?

223
00:19:11,400 --> 00:19:17,480
That idea would require software
tools, a new language of sorts,

224
00:19:18,040 --> 00:19:23,440
and a willingness to invite
people into the chips inner

225
00:19:23,440 --> 00:19:26,480
rhythm.
It also required humility,

226
00:19:27,040 --> 00:19:30,560
because it meant admitting that
the future might not be only

227
00:19:30,560 --> 00:19:35,320
about games, even if games had
built the company.

228
00:19:35,560 --> 00:19:41,200
So as the world kept playing,
NVIDIA began preparing another

229
00:19:41,200 --> 00:19:45,960
doorway, one that would lead
from rendering scenes to

230
00:19:45,960 --> 00:19:51,680
accelerating science.
Tonight we can feel that doorway

231
00:19:52,040 --> 00:19:57,080
as a soft draft in a hallway,
the hint of a larger room

232
00:19:57,080 --> 00:20:02,440
beyond, filled with equations,
data, and the steady hum of

233
00:20:02,440 --> 00:20:08,200
parallel machines.
ACPU is like a careful solo

234
00:20:08,200 --> 00:20:15,480
thinker, stepping through tasks
one by one, while a GPU is more

235
00:20:15,480 --> 00:20:20,920
like a choir.
Many small voices each singing a

236
00:20:20,920 --> 00:20:25,760
simple note at the same time.
When the notes line up, the

237
00:20:25,760 --> 00:20:31,360
result is not just faster but
different, because problems that

238
00:20:31,360 --> 00:20:37,120
were too large to attempt can
suddenly be held in memory and

239
00:20:37,120 --> 00:20:41,440
worked on patiently.
Think of medical images are

240
00:20:41,440 --> 00:20:46,280
reconstructed from many slices,
or seismic readings turned into

241
00:20:46,280 --> 00:20:51,760
maps of hidden rock, or
molecules simulated as they fold

242
00:20:51,760 --> 00:20:58,280
and bump in liquid, all of it
made of repeating math, all of

243
00:20:58,280 --> 00:21:03,440
it hungry for parallel effort.
In those years, the people doing

244
00:21:03,440 --> 00:21:08,840
this work often built their own
clusters from ordinary graphics

245
00:21:08,840 --> 00:21:15,560
cards stacked in metal frames
with extra fans, turning gaming

246
00:21:15,560 --> 00:21:19,160
hardware into homemade
supercomputers.

247
00:21:19,480 --> 00:21:24,240
It was an improvised kind of
progress, but it carried a clear

248
00:21:24,240 --> 00:21:29,560
message the GPU was ready to
become something more.

249
00:21:29,920 --> 00:21:37,600
In 2006, NVIDIA opened that
doorway with a name that sounded

250
00:21:37,840 --> 00:21:44,080
like a quiet spark, CUDA.
CUDA was a way of speaking to

251
00:21:44,080 --> 00:21:49,080
the GPU more directly, letting
programmers write ordinary

252
00:21:49,080 --> 00:21:54,560
looking code and letting the GPU
translate it into thousands of

253
00:21:54,560 --> 00:21:59,080
tiny parallel steps.
It did not change the laws of

254
00:21:59,080 --> 00:22:04,960
physics, but it changed who
could use them, because you no

255
00:22:04,960 --> 00:22:11,120
longer had to disguise your work
as graphics to access the GPU's

256
00:22:11,120 --> 00:22:14,920
power.
Now a researcher could say in

257
00:22:14,920 --> 00:22:21,400
plain terms, take this matrix,
apply this operation, repeat it

258
00:22:21,400 --> 00:22:25,680
across the data and do it all at
once.

259
00:22:26,040 --> 00:22:32,360
The GPU in response would not
think harder, but it would work

260
00:22:32,360 --> 00:22:38,040
wider, spreading the task across
many cores, each one doing a

261
00:22:38,040 --> 00:22:41,200
small piece with steady
patience.

262
00:22:41,440 --> 00:22:46,440
NVIDIA began to build products
specifically for this world.

263
00:22:47,000 --> 00:22:52,880
GPU's intended for data centers
and labs, often sold not as

264
00:22:52,880 --> 00:22:59,640
shiny gaming cards, but as
accelerators, quiet engines

265
00:23:00,040 --> 00:23:04,040
bolted into racks.
This was the beginning of a

266
00:23:04,040 --> 00:23:07,520
pattern that would repeat again
and again.

267
00:23:08,120 --> 00:23:12,400
Hardware paired with an
ecosystem, Software paired with

268
00:23:12,400 --> 00:23:16,680
community.
The chip and the tools growing

269
00:23:16,680 --> 00:23:20,840
together.
Scientists used CUDA to speed up

270
00:23:20,840 --> 00:23:26,560
simulations, engineers used it
to test designs, and

271
00:23:26,560 --> 00:23:31,880
universities began teaching the
new approach, training students

272
00:23:31,880 --> 00:23:35,800
to think in parallel.
If you stood in one of those

273
00:23:35,800 --> 00:23:40,960
labs, you might hear the same
familiar sound as in a gaming

274
00:23:40,960 --> 00:23:45,880
PC.
Fans, airflow, a constant hum.

275
00:23:46,600 --> 00:23:49,040
But the work inside was
different.

276
00:23:49,360 --> 00:23:55,240
Instead of drawing Dragons or
stadiums, the GPU's were tracing

277
00:23:55,240 --> 00:24:03,640
protein folds, reconstructing CT
scans, simulating fluids or

278
00:24:03,640 --> 00:24:06,280
searching patterns in noisy
signals.

279
00:24:06,600 --> 00:24:12,800
And still, nothing about it felt
flashy, because the best

280
00:24:12,800 --> 00:24:18,240
infrastructure rarely does.
It simply makes the impossible

281
00:24:18,240 --> 00:24:23,000
feel routine, and the routine
feel gentle.

282
00:24:23,240 --> 00:24:28,600
What began as an experiment at
the edges was becoming a second

283
00:24:28,600 --> 00:24:33,640
foundation beneath the company,
and NVIDIA was learning how to

284
00:24:33,640 --> 00:24:40,360
stand on both the playful world
of games and the patient world

285
00:24:40,360 --> 00:24:44,960
of science.
In some places, professors built

286
00:24:44,960 --> 00:24:50,000
clusters from off the shelf
cards, lining them up like books

287
00:24:50,000 --> 00:24:56,000
on a shelf, cooling them with
box fans, and calling it a

288
00:24:56,000 --> 00:25:00,880
personal supercomputer.
And as these experiments spread,

289
00:25:01,360 --> 00:25:06,360
large supercomputing centers
began to take notice.

290
00:25:06,960 --> 00:25:13,680
Because the math was persuasive.
For certain workloads, one GPU

291
00:25:13,960 --> 00:25:19,320
could replace many CPUs.
Before long, the fastest

292
00:25:19,320 --> 00:25:24,080
machines in the world were
mixing processes the way a good

293
00:25:24,080 --> 00:25:33,240
kitchen mixes tools.
CPO's for general work, GPU's

294
00:25:33,240 --> 00:25:38,400
for the heavy parallel stirring,
each doing what it did best.

295
00:25:38,800 --> 00:25:44,000
The idea of accelerated
computing began to sound less

296
00:25:44,000 --> 00:25:50,640
like a slogan and more like a
new normal, a calm shift in how

297
00:25:50,640 --> 00:25:53,840
the world's hardest problems
would be approached.

298
00:25:54,200 --> 00:25:59,040
By the early 2000 tens, another
quiet community was gathering

299
00:25:59,040 --> 00:26:03,400
around GPU's, not because of
graphics but because of

300
00:26:03,880 --> 00:26:07,320
learning.
Machine learning had existed for

301
00:26:07,320 --> 00:26:11,840
a long time like a library of
ideas waiting for the right kind

302
00:26:11,840 --> 00:26:17,040
of electricity, but it often
felt slow and limited by the

303
00:26:17,040 --> 00:26:20,880
available compute.
Deep neural networks in

304
00:26:20,880 --> 00:26:27,320
particular needed two things to
truly flourish large data sets

305
00:26:27,800 --> 00:26:31,160
and an immense amount of
repeated math.

306
00:26:31,520 --> 00:26:37,080
That repeated math was mostly
linear algebra, the kind of

307
00:26:37,080 --> 00:26:42,360
multiplication and addition that
can be broken into many small

308
00:26:42,360 --> 00:26:48,040
pieces and done in parallel,
like raindrops on a window.

309
00:26:48,360 --> 00:26:53,720
In 2012, a group of researchers
trained a deep network to

310
00:26:53,720 --> 00:26:59,880
recognize images in a major
competition, and they used GPUs

311
00:26:59,880 --> 00:27:06,800
to do it, because without GPUs
it would have taken far too

312
00:27:06,800 --> 00:27:09,960
long.
The results were not a small

313
00:27:09,960 --> 00:27:14,400
improvement.
They were a clear leap, a sudden

314
00:27:14,400 --> 00:27:19,560
drop in error that made the
field look up and pay attention.

315
00:27:19,920 --> 00:27:26,000
It was one of those moments that
feels later, like a hinge, a

316
00:27:26,000 --> 00:27:31,560
door swinging open, revealing a
new room in the House of

317
00:27:31,560 --> 00:27:35,400
technology.
After that, more people tried

318
00:27:35,400 --> 00:27:39,200
deep learning because the path
was now visible.

319
00:27:39,920 --> 00:27:45,960
If you can train the network
faster, you can iterate, and

320
00:27:45,960 --> 00:27:51,000
iteration is how intuition
becomes progress.

321
00:27:51,360 --> 00:27:57,360
NVIDIA, watching this,
understood something important.

322
00:27:58,040 --> 00:28:03,760
The Jeep AU had become a general
engine for modern pattern

323
00:28:03,760 --> 00:28:07,520
recognition.
A network learns by adjusting

324
00:28:07,520 --> 00:28:12,600
countless tiny weights, nudging
them again and again, and the

325
00:28:12,600 --> 00:28:19,080
GPU's parallel structure made
that nudging feasible at scale.

326
00:28:19,480 --> 00:28:23,520
The company began to serve this
emerging world with more

327
00:28:23,520 --> 00:28:29,960
intention, better software
libraries, specialized hardware

328
00:28:29,960 --> 00:28:35,800
features, whole systems designed
to train models efficiently.

329
00:28:36,160 --> 00:28:41,600
And while the public still
thought of GPU's as gaming parts

330
00:28:42,320 --> 00:28:48,160
in quiet server rooms, the story
was shifting because the same

331
00:28:48,160 --> 00:28:53,760
silicon that drew game worlds
was now helping machines see.

332
00:28:54,120 --> 00:28:57,800
You could almost feel the
cultural change approaching.

333
00:28:58,480 --> 00:29:02,960
Cameras everywhere, data
everywhere, and the growing

334
00:29:02,960 --> 00:29:08,640
sense that computers might learn
from experience rather than only

335
00:29:08,640 --> 00:29:13,560
follow explicit rules.
In the sleep wise calm of this

336
00:29:13,560 --> 00:29:19,360
night, we can hold that change
softly, not as a sudden

337
00:29:19,360 --> 00:29:26,280
explosion, but as a slow dawn,
the kind that brightens the sky

338
00:29:26,280 --> 00:29:30,160
before anyone notices.
The benchmark was called

339
00:29:30,400 --> 00:29:35,480
Imagenet, and its millions of
labeled pictures were like a

340
00:29:35,480 --> 00:29:40,200
training ground for digital
vision, a way to measure whether

341
00:29:40,200 --> 00:29:45,960
a machine could recognize a cat,
a street sign, a face.

342
00:29:46,280 --> 00:29:51,960
The network that one often
remembered as Alex Net was

343
00:29:51,960 --> 00:29:57,520
trained on consumer GPU's, the
kind someone could buy for a

344
00:29:57,520 --> 00:30:02,280
gaming PC.
And that detail mattered.

345
00:30:03,000 --> 00:30:07,040
It meant the new era did not
require a secret government

346
00:30:07,040 --> 00:30:10,160
machine, only the right
approach.

347
00:30:10,480 --> 00:30:15,680
Soon the internet's endless
photos and videos became more

348
00:30:15,680 --> 00:30:20,880
than content.
They became data, and the GPU

349
00:30:20,880 --> 00:30:26,080
became the quiet mill that could
grind that data into learned

350
00:30:26,080 --> 00:30:29,680
ability.
Once a new door opens, the world

351
00:30:29,680 --> 00:30:35,200
tends to rush through it.
But NVIDIA moved in its own way,

352
00:30:35,840 --> 00:30:41,720
deliberate building not a single
product but an entire path.

353
00:30:42,080 --> 00:30:45,680
The company began to speak more
often about accelerated

354
00:30:45,680 --> 00:30:51,360
computing, about GPU's and data
centers, about a future where

355
00:30:51,360 --> 00:30:56,200
training a model would be as
common as hosting a website.

356
00:30:56,600 --> 00:31:01,560
To make that future easier, they
built complete machines designed

357
00:31:01,560 --> 00:31:08,000
for AI boxes of GPU's tuned to
work together like a small

358
00:31:08,000 --> 00:31:13,080
orchestra in a single chassis.
These systems had names like

359
00:31:13,080 --> 00:31:18,800
DGX, but the names were less
important than the feeling they

360
00:31:18,800 --> 00:31:24,600
offered that AI training could
be packaged, repeatable, and

361
00:31:24,600 --> 00:31:27,640
reliable.
They also refined the chips

362
00:31:27,640 --> 00:31:31,920
themselves, adding features
meant for the new workload

363
00:31:32,600 --> 00:31:37,240
circuits that speed up the
matrix math of neural networks,

364
00:31:37,640 --> 00:31:41,480
doing the same kind of
multiplication again and again

365
00:31:41,760 --> 00:31:45,520
with less energy.
People sometimes called those

366
00:31:45,520 --> 00:31:50,680
features tenser cores, but you
can think of them as a gentle

367
00:31:50,680 --> 00:31:55,280
shortcut, a way of making the
most common steps in learning

368
00:31:55,280 --> 00:31:59,160
faster and smoother at the same
time.

369
00:31:59,840 --> 00:32:05,440
Gaming did not disappear.
It continued vivid and bright.

370
00:32:05,840 --> 00:32:11,080
And NVIDIA kept improving the
visual magic too, even

371
00:32:11,320 --> 00:32:16,920
introducing ray tracing, A
technique that imitates how

372
00:32:16,920 --> 00:32:21,960
light bounces.
It was an unusual balance, 1

373
00:32:21,960 --> 00:32:26,920
foot in play, 1 foot in
research, and both supported by

374
00:32:26,920 --> 00:32:32,440
the same foundational idea of
parallel work in server rooms.

375
00:32:32,800 --> 00:32:38,840
The GPU's began to cluster in
large groups, linked by fast

376
00:32:38,840 --> 00:32:44,520
connections, because training
bigger models meant coordinating

377
00:32:44,520 --> 00:32:47,720
many chips as if they were one
machine.

378
00:32:48,040 --> 00:32:52,680
This is where the story begins
to feel less like a consumer

379
00:32:52,680 --> 00:32:57,640
tale and more like
infrastructure, the kind of

380
00:32:57,640 --> 00:33:03,160
technology you rarely see but
you depend on all the same.

381
00:33:03,480 --> 00:33:09,640
And yet even infrastructure has
a mood, and the mood here was

382
00:33:09,640 --> 00:33:15,200
steady ambition, the patient
confidence that if you build the

383
00:33:15,200 --> 00:33:19,440
right platform, others will
build wonders on top of it.

384
00:33:19,800 --> 00:33:25,560
So NVIDIA kept doing what it had
learned in its earliest years.

385
00:33:26,120 --> 00:33:29,920
Iterate.
Refine and listen.

386
00:33:30,480 --> 00:33:35,240
While the hum of computation
grew a little louder in the

387
00:33:35,240 --> 00:33:39,760
background of the modern world,
cloud companies began offering

388
00:33:39,760 --> 00:33:44,360
GPU's by the hour, as if
parallel compute were a kind of

389
00:33:44,640 --> 00:33:49,240
calm utility.
And researchers rented time the

390
00:33:49,240 --> 00:33:52,520
way earlier generations rented
film cameras.

391
00:33:52,840 --> 00:33:57,840
Software frameworks like
Tensorflow and π Torch made it

392
00:33:57,840 --> 00:34:02,800
easier to describe learning
models, and beneath them,

393
00:34:03,160 --> 00:34:09,440
Nvidia's libraries helped the
GPU do the heavy lifting with

394
00:34:09,440 --> 00:34:15,719
quiet efficiency.
In this way, the GPU became less

395
00:34:15,719 --> 00:34:21,520
like a component and more like a
place, an environment where

396
00:34:21,520 --> 00:34:27,360
ideas could run, repeat, and
improve while humans slept and

397
00:34:27,360 --> 00:34:31,760
the machines kept training.
The world was entering an era

398
00:34:31,760 --> 00:34:37,040
where progress often meant more
data, more compute, and more

399
00:34:37,040 --> 00:34:43,159
patience, and NVIDIA had
positioned itself at the center

400
00:34:43,159 --> 00:34:47,000
of that triangle.
As AI models grew larger,

401
00:34:47,320 --> 00:34:49,960
something surprising became
obvious.

402
00:34:50,760 --> 00:34:53,280
It wasn't enough to have fast
chips.

403
00:34:53,679 --> 00:34:57,520
You also needed fast
conversation between chips.

404
00:34:57,880 --> 00:35:02,400
Training a large model is like
teaching a choir a new song.

405
00:35:03,040 --> 00:35:07,960
Each section can practice its
part, but they must hear each

406
00:35:07,960 --> 00:35:12,200
other to stay in time.
So the network inside a data

407
00:35:12,200 --> 00:35:17,440
center became as important as
the processors themselves, and

408
00:35:17,600 --> 00:35:23,120
NVIDIA began to think about the
whole system, not just the GPU.

409
00:35:23,400 --> 00:35:28,800
That thinking LED them to buy
Mellanox, a company known for

410
00:35:28,800 --> 00:35:34,200
high speed networking, the kind
used in supercomputers where

411
00:35:34,200 --> 00:35:39,840
every microsecond matters.
With that, NVIDIA could offer

412
00:35:39,920 --> 00:35:44,520
not only the engines that
compute, but the pathways that

413
00:35:44,520 --> 00:35:49,760
move data connecting racks and
nodes into one coordinated

414
00:35:49,760 --> 00:35:54,200
Organism.
Jensen Huang often described a

415
00:35:54,200 --> 00:36:01,360
data center as a single giant
computer spread across many

416
00:36:01,360 --> 00:36:07,600
machines, and it's a calming
image, like a city seen from

417
00:36:07,600 --> 00:36:11,760
above, lit by many small
synchronized lights.

418
00:36:12,200 --> 00:36:16,760
This was also the period when
NVIDIA explored becoming more

419
00:36:16,760 --> 00:36:22,360
than a GPU company, imagining a
future where it could shape both

420
00:36:22,360 --> 00:36:26,920
the graphics side and the
broader processor landscape.

421
00:36:27,360 --> 00:36:33,480
In 2000, Twenties NVIDIA
announced a plan to acquire ARM,

422
00:36:34,040 --> 00:36:39,080
whose CPU designs live inside
countless phones and devices,

423
00:36:39,680 --> 00:36:43,320
and whose architecture was
spreading towards servers.

424
00:36:43,720 --> 00:36:49,000
The attempt did not succeed,
held back by regulatory concerns

425
00:36:49,200 --> 00:36:53,680
and the delicate politics of
foundational technology.

426
00:36:54,360 --> 00:36:57,680
But even the attempt revealed
the direction of the company's

427
00:36:57,680 --> 00:37:00,400
mind.
Whether through ownership or

428
00:37:00,400 --> 00:37:07,040
partnership, NVIDIA wanted a
world where GPU's and CPU's work

429
00:37:07,040 --> 00:37:12,600
together more closely, each
optimized for its role, each

430
00:37:12,600 --> 00:37:17,160
helping the other.
So they kept building new kinds

431
00:37:17,160 --> 00:37:21,960
of server CPU, used new
interconnects, new ways to

432
00:37:21,960 --> 00:37:25,000
stitch computing into a seamless
fabric.

433
00:37:25,360 --> 00:37:31,680
It's a story that feels, at this
point, less like a product line

434
00:37:32,000 --> 00:37:38,520
and more like an ecosystem, a
quiet landscape of tools

435
00:37:38,920 --> 00:37:43,240
designed to make complex work
feel natural.

436
00:37:43,520 --> 00:37:48,080
And somewhere in that landscape,
the original dream is still

437
00:37:48,080 --> 00:37:53,040
present, not just to make images
beautiful, but to make

438
00:37:53,040 --> 00:37:58,640
computation itself smoother,
wider, and more alive.

439
00:37:59,080 --> 00:38:04,760
They even began speaking about
new categories like the DPU, a

440
00:38:04,760 --> 00:38:09,040
data processing unit meant to
handle the busy work of

441
00:38:09,040 --> 00:38:16,000
networking and security so CPU's
and GPU's could focus on their

442
00:38:16,000 --> 00:38:20,920
main tasks.
In a large cloud, that division

443
00:38:20,920 --> 00:38:28,120
of Labor matters, because every
saved cycle becomes space for

444
00:38:28,120 --> 00:38:33,480
another training run, another
simulation, another quiet

445
00:38:33,480 --> 00:38:36,920
experiment.
So the architecture of modern

446
00:38:36,920 --> 00:38:41,520
computing began to look like a
well run port at night.

447
00:38:42,200 --> 00:38:48,000
Different cranes, different
lanes, each moving its share so

448
00:38:48,000 --> 00:38:51,160
the whole system flows without
strain.

449
00:38:51,640 --> 00:38:56,720
By the time the 20 twenties
arrived, the GPU had become

450
00:38:56,720 --> 00:39:01,320
something you rarely saw
directly, like electricity

451
00:39:01,320 --> 00:39:07,840
behind a wall, present,
necessary and mostly unnoticed

452
00:39:08,480 --> 00:39:11,240
until something remarkable
happens.

453
00:39:11,600 --> 00:39:16,520
That remarkable thing for many
people was the sudden feeling

454
00:39:16,520 --> 00:39:21,680
that a computer could speak not
in stiff commands, but in

455
00:39:21,680 --> 00:39:26,080
flowing sentences, answering
questions, writing drafts,

456
00:39:26,520 --> 00:39:32,720
composing small pieces of code.
Generative AI is the name often

457
00:39:32,720 --> 00:39:36,960
used for this, but the
experience is simpler than the

458
00:39:36,960 --> 00:39:40,760
label.
You type a few words and the

459
00:39:40,760 --> 00:39:44,800
machine responds with something
that seems to understand the

460
00:39:44,800 --> 00:39:49,280
shape of language.
Behind that softness is an

461
00:39:49,280 --> 00:39:54,640
immense amount of training, a
long process where a model reads

462
00:39:54,640 --> 00:40:00,800
and predicts and adjusts again
and again, learning patterns the

463
00:40:00,800 --> 00:40:07,200
way a musician learns scales.
This training is repetitive math

464
00:40:07,320 --> 00:40:13,880
on a vast scale, and it thrives
on parallel effort, which is why

465
00:40:13,880 --> 00:40:20,440
GPU's became the engines of the
moment, packed into data centers

466
00:40:20,720 --> 00:40:25,520
like rows of quiet lanterns in
2016.

467
00:40:25,880 --> 00:40:31,480
Jensen Huang even carried one of
Nvidia's early AI systems to

468
00:40:31,800 --> 00:40:38,520
open AI by hand, a symbolic
gesture that felt almost old

469
00:40:38,520 --> 00:40:45,640
fashioned, a founder delivering
a machine to a small team as if

470
00:40:45,880 --> 00:40:48,440
dropping off tools for a
workshop.

471
00:40:48,840 --> 00:40:55,080
Years later, those tools had
become entire fleets, and the

472
00:40:55,080 --> 00:40:59,800
fleets had become part of a new
global infrastructure, built so

473
00:40:59,800 --> 00:41:04,360
models could learn faster and be
updated more often.

474
00:41:04,760 --> 00:41:08,960
If you think of the Internet as
a library, then these training

475
00:41:08,960 --> 00:41:14,600
clusters are the late night
readers, turning pages at

476
00:41:14,600 --> 00:41:20,400
extraordinary speed, not for
pleasure but for pattern.

477
00:41:20,840 --> 00:41:23,640
The cultural ripple was
immediate.

478
00:41:24,320 --> 00:41:29,560
Images generated from prompts,
voices that could be

479
00:41:29,560 --> 00:41:35,120
synthesized, assistance that
could help with daily writing,

480
00:41:35,840 --> 00:41:40,920
and a renewed debate about what
creativity means.

481
00:41:41,280 --> 00:41:46,520
Yet the hardware story remains
steady, fans spinning, heat

482
00:41:46,520 --> 00:41:52,760
rising, racks glowing softly,
and the GPU doing what it has

483
00:41:52,760 --> 00:41:59,160
always done best, many simple
steps repeated in parallel until

484
00:41:59,160 --> 00:42:04,320
complexity feels smooth.
And as we move forward, we can

485
00:42:04,320 --> 00:42:09,160
keep one gentle thought in mind
that the most surprising

486
00:42:09,160 --> 00:42:14,560
revolutions are often powered by
the least glamorous things, by

487
00:42:14,600 --> 00:42:19,680
infrastructure that simply keeps
working all night long.

488
00:42:20,040 --> 00:42:28,240
For many, the first encounter
had a name like ChatGPT, a small

489
00:42:28,240 --> 00:42:33,560
window on a screen that made the
future feel oddly familiar, like

490
00:42:33,560 --> 00:42:36,480
sending a message to a
thoughtful stranger.

491
00:42:36,840 --> 00:42:43,560
And almost overnight, more
people learned the word GPU not

492
00:42:43,560 --> 00:42:49,040
from gaming forums but from
headlines and conversations at

493
00:42:49,040 --> 00:42:53,800
work, as if the hidden engine
had finally become part of the

494
00:42:53,800 --> 00:42:59,840
public vocabulary.
Still, the GPU itself did not

495
00:42:59,840 --> 00:43:03,800
change its personality.
It remained a patient

496
00:43:03,800 --> 00:43:09,120
calculator, turning data into
learned behavior with the same

497
00:43:09,120 --> 00:43:14,400
calm repetition it once used to
turn geometry into light.

498
00:43:14,760 --> 00:43:19,840
When a new technology becomes
widely desired, the world begins

499
00:43:19,840 --> 00:43:24,600
to measure it in physical terms.
How many chips?

500
00:43:24,840 --> 00:43:29,080
How many servers?
How many megawatts of power?

501
00:43:29,440 --> 00:43:32,480
How much cooling?
How much space?

502
00:43:33,040 --> 00:43:39,360
AI training is not just
software, it is a landscape of

503
00:43:39,360 --> 00:43:45,400
machines, and that landscape has
temperature, weight, wiring and

504
00:43:45,400 --> 00:43:49,640
timing.
Nvidia's newer data center GPU's

505
00:43:49,640 --> 00:43:53,720
arrived with names that sounded
like quiet tributes to

506
00:43:53,720 --> 00:43:59,880
scientists, and each generation
brought more parallel capacity,

507
00:44:00,400 --> 00:44:05,440
more efficiency, more ability to
move data through the chip

508
00:44:05,720 --> 00:44:10,240
without bottlenecks.
But the real change was not only

509
00:44:10,240 --> 00:44:15,440
inside the silicon, it was in
how the systems were built.

510
00:44:16,280 --> 00:44:22,680
GPU's linked together with fast
connections, treated as one

511
00:44:22,680 --> 00:44:26,960
shared pool of compute.
This is why networking mattered

512
00:44:26,960 --> 00:44:30,880
so much, and why NVIDIA cared
about the whole stack.

513
00:44:31,520 --> 00:44:37,560
The GPU, the interconnect, the
software libraries, the tools

514
00:44:37,560 --> 00:44:42,320
that schedule work, the way
memory is managed.

515
00:44:42,680 --> 00:44:48,920
It is also why cloud companies
invested heavily, turning AI

516
00:44:48,920 --> 00:44:55,080
compute into a service, renting
out acceleration the way a city

517
00:44:55,080 --> 00:44:59,640
rents out office space.
If you imagine a data center at

518
00:44:59,640 --> 00:45:03,080
night, it can feel surprisingly
peaceful.

519
00:45:03,720 --> 00:45:09,120
A warehouse of quiet order,
blinking lights, controlled air,

520
00:45:09,360 --> 00:45:14,800
steady redundancy inside.
Training jobs run for days,

521
00:45:15,120 --> 00:45:19,760
sometimes weeks, and the
progress is incremental.

522
00:45:20,440 --> 00:45:25,240
A little better accuracy, a
little more fluency, a little

523
00:45:25,240 --> 00:45:30,120
more stability.
This kind of progress rewards

524
00:45:30,120 --> 00:45:34,440
patience, and patience has
always been part of Nvidia's

525
00:45:34,440 --> 00:45:39,520
character, shaped early by the
knowledge that one failed chip

526
00:45:39,800 --> 00:45:44,040
can end a company.
So the firm kept refining the

527
00:45:44,040 --> 00:45:49,600
same fundamental lesson.
Parallel computing is a way to

528
00:45:49,600 --> 00:45:54,800
make the heavy feel light, as
long as you respect the details.

529
00:45:55,120 --> 00:45:59,480
And as the demand rose, the
world began to notice another

530
00:45:59,480 --> 00:46:05,360
truth that computation is a
physical resource, and that

531
00:46:05,360 --> 00:46:10,560
every new capability must
eventually be paid for in power,

532
00:46:10,760 --> 00:46:16,360
heat, and careful engineering.
Engineers began talking about

533
00:46:16,360 --> 00:46:21,320
packaging and memory as much as
raw compute, because feeding the

534
00:46:21,320 --> 00:46:29,240
GPU is as important as the GPU
itself, like making sure a fast

535
00:46:29,240 --> 00:46:34,800
train has enough track.
High bandwidth memory stacked

536
00:46:34,800 --> 00:46:39,960
close to the processor helped
keep the data flowing, and new

537
00:46:39,960 --> 00:46:45,120
systems used liquid cooling or
carefully designed airflow to

538
00:46:45,120 --> 00:46:50,920
carry heat away quietly.
In some places, the constraint

539
00:46:50,920 --> 00:46:55,680
became electricity, and
companies planned new

540
00:46:55,680 --> 00:47:02,120
substations and new power
contracts, as if AI were not

541
00:47:02,120 --> 00:47:07,080
only a software wave but a new
kind of industry.

542
00:47:07,400 --> 00:47:12,120
And yet the mood inside those
rooms remained calm.

543
00:47:12,800 --> 00:47:18,120
The hum of fans, the soft blink
of status lights, and the sense

544
00:47:18,120 --> 00:47:22,960
that thinking in our modern
world has become something we

545
00:47:23,160 --> 00:47:27,920
build.
It's easy now to look at NVIDIA

546
00:47:28,120 --> 00:47:33,040
and see inevitability, to
imagine a straight line from the

547
00:47:33,040 --> 00:47:36,360
diner breakfast to the modern
data center.

548
00:47:36,720 --> 00:47:41,760
But the line was never straight,
and the calm truth is that the

549
00:47:41,760 --> 00:47:46,960
company survived by repeatedly
choosing the uncomfortable

550
00:47:46,960 --> 00:47:49,760
future over the comfortable
present.

551
00:47:50,200 --> 00:47:56,480
Jensen Huang liked to talk about
investing in zero billion dollar

552
00:47:56,480 --> 00:48:02,200
markets, meaning places where
the opportunity is real but not

553
00:48:02,200 --> 00:48:06,720
yet visible, like a shoreline
before sunrise.

554
00:48:07,120 --> 00:48:13,720
In the 1990s, 3D gaming felt
like that, Exciting but small

555
00:48:14,200 --> 00:48:20,000
and easy to dismiss.
In the mid 2000s, GPU computing

556
00:48:20,000 --> 00:48:25,240
for science felt like that,
useful but niche and hard to

557
00:48:25,240 --> 00:48:30,160
monetize quickly.
In the early 2000, Tens, deep

558
00:48:30,160 --> 00:48:36,120
learning felt like that, a
research curiosity that suddenly

559
00:48:36,120 --> 00:48:42,880
became practical only when data
and GPUs made it practical.

560
00:48:43,280 --> 00:48:47,600
To keep making those bets,
NVIDIA had to build a culture

561
00:48:47,600 --> 00:48:50,120
that could live with long
timelines.

562
00:48:50,600 --> 00:48:53,120
And that is not a glamorous
skill.

563
00:48:53,440 --> 00:48:59,080
It means spending years on
software ecosystems that do not

564
00:48:59,080 --> 00:49:04,560
show up in product photos,
compilers, libraries,

565
00:49:05,040 --> 00:49:11,640
documentation, developer
support, the slow work of making

566
00:49:11,640 --> 00:49:16,080
tools trustworthy.
It means accepting that a

567
00:49:16,080 --> 00:49:21,400
platform is not a single
invention, but a relationship,

568
00:49:21,960 --> 00:49:26,880
one that deepens as more people
build on it and as the company

569
00:49:26,880 --> 00:49:29,800
responds.
If you have ever learned a

570
00:49:29,800 --> 00:49:35,120
craft, you know the feeling.
Progress doesn't happen as a

571
00:49:35,120 --> 00:49:39,520
single leap.
It happens as 1000 small

572
00:49:39,520 --> 00:49:43,200
corrections, each one barely
visible.

573
00:49:43,600 --> 00:49:49,920
Nvidia's rise is built on those
corrections, the way a graphics

574
00:49:49,920 --> 00:49:53,680
company learned to serve
scientists, then learned to

575
00:49:53,680 --> 00:49:57,720
serve machine learning, then
learn to serve the world's

576
00:49:57,720 --> 00:50:01,960
largest clouds.
And through that, the company's

577
00:50:01,960 --> 00:50:04,960
original question remained
intact.

578
00:50:05,600 --> 00:50:10,760
How do we make computation feel
fluid so creativity and

579
00:50:10,760 --> 00:50:14,280
discovery can move without
friction?

580
00:50:14,720 --> 00:50:19,240
Tonight, we can let that
question settle like a warm

581
00:50:19,240 --> 00:50:26,400
stone in the hand.
Simple, steady and powerful in

582
00:50:26,400 --> 00:50:29,840
its persistence.
You can hear the lesson most

583
00:50:29,840 --> 00:50:36,840
clearly in the crisis years,
when early products failed, cash

584
00:50:36,840 --> 00:50:43,760
grew tight and the team shrank
not from lack of talent but from

585
00:50:43,760 --> 00:50:48,880
the reality that hardware is
expensive to get wrong.

586
00:50:49,200 --> 00:50:54,480
In that period, they learned to
test a chip before it existed,

587
00:50:54,960 --> 00:51:00,280
running long emulations and
listening closely for mistakes

588
00:51:00,840 --> 00:51:05,040
because they could not afford a
second miss.

589
00:51:05,440 --> 00:51:10,200
When the recovery chip worked
and sold, it did more than save

590
00:51:10,200 --> 00:51:14,000
the company.
It taught them that speed is

591
00:51:14,000 --> 00:51:18,960
useful only when paired with
rigor, and that survival is

592
00:51:18,960 --> 00:51:25,040
often a matter of details done
calmly, correctly, and on time.

593
00:51:25,440 --> 00:51:29,120
That memory stayed with them,
and it shaped how they

594
00:51:29,120 --> 00:51:34,640
approached every later pivot,
with patience, with discipline,

595
00:51:35,120 --> 00:51:39,400
and with a willingness to
rebuild their future from the

596
00:51:39,400 --> 00:51:44,440
inside out.
Sometimes it helps to remember

597
00:51:45,160 --> 00:51:48,920
that all of this began with
light.

598
00:51:49,360 --> 00:51:53,720
A game world is mostly light and
timing.

599
00:51:54,600 --> 00:52:00,040
The illusion that a surface is
curved, the suggestion that a

600
00:52:00,040 --> 00:52:05,880
hallway recedes, the gentle
persistence of motion from frame

601
00:52:05,880 --> 00:52:10,120
to frame.
In the Doom and Quake era, those

602
00:52:10,120 --> 00:52:14,760
worlds were assembled with
clever shortcuts and you could

603
00:52:14,760 --> 00:52:19,400
almost see the seams.
But the excitement was real

604
00:52:19,400 --> 00:52:24,800
because your imagination filled
in what the hardware could not.

605
00:52:25,280 --> 00:52:32,240
Then GPUs arrived and the seams
softened, textures became

606
00:52:32,240 --> 00:52:38,640
richer, geometry more complex,
and the mind stopped noticing

607
00:52:38,640 --> 00:52:42,000
the tricks.
Later, with programmable

608
00:52:42,000 --> 00:52:48,160
shaders, developers could paint
with equations telling the GPU

609
00:52:48,160 --> 00:52:52,520
how water should shimmer, how
smoke should fade, how skin

610
00:52:52,520 --> 00:52:58,160
should scatter light softly.
Eventually, ray tracing brought

611
00:52:58,160 --> 00:53:03,880
the old dream closer, simulating
the path of light itself,

612
00:53:04,520 --> 00:53:09,840
letting reflections and shadows
behave more like they do in real

613
00:53:09,840 --> 00:53:12,840
rooms.
It is easy to think of this as

614
00:53:13,080 --> 00:53:19,200
graphics, yet it is also a
lesson about simulation, about

615
00:53:19,200 --> 00:53:23,280
using computation to imitate the
physical world.

616
00:53:23,600 --> 00:53:30,200
And simulation in a calm way is
a cousin of learning, because

617
00:53:30,200 --> 00:53:35,440
both require a great many small
calculations to produce

618
00:53:35,440 --> 00:53:41,080
something that feels coherent.
When AI models learn patterns,

619
00:53:41,600 --> 00:53:46,880
they are doing their own kind of
rendering, not of light, but of

620
00:53:46,880 --> 00:53:52,160
language, of images, of the
statistical shape of the world.

621
00:53:52,560 --> 00:53:56,840
The GPU's gift is the same in
both cases.

622
00:53:57,400 --> 00:54:00,640
It handles repetition without
complaint.

623
00:54:01,240 --> 00:54:04,560
It spreads work across many
cores.

624
00:54:05,080 --> 00:54:11,160
It makes the heavy field smooth.
So the cultural story and the

625
00:54:11,160 --> 00:54:17,240
technical story keep touching A
teenager playing Unreal

626
00:54:17,240 --> 00:54:23,240
Tournament in a dim bedroom and
a researcher training an image

627
00:54:23,240 --> 00:54:28,000
model in a bright lab, both
relying on the same kind of

628
00:54:28,000 --> 00:54:31,880
parallel engine.
Tonight, you don't need to

629
00:54:31,880 --> 00:54:36,880
separate these threads.
You can let them weave together

630
00:54:37,400 --> 00:54:43,360
and feel how a technology can
grow by serving play first, then

631
00:54:43,360 --> 00:54:48,120
serving discovery, and
eventually serving almost

632
00:54:48,120 --> 00:54:51,360
everything.
Even the sounds are related.

633
00:54:51,840 --> 00:54:58,160
The click of a mouse, the low
whirl of a fan, the faint heat

634
00:54:58,400 --> 00:55:02,000
rising from a PC case after a
long session.

635
00:55:02,400 --> 00:55:07,720
Years later, the same kind of
hum fills a data center, only

636
00:55:07,720 --> 00:55:13,400
steadier, more distant, as if
the world's computing has moved

637
00:55:13,400 --> 00:55:20,160
from bedrooms into warehouses.
And yet the feeling is familiar,

638
00:55:20,720 --> 00:55:26,040
a machine working beside you,
not demanding attention, just

639
00:55:26,320 --> 00:55:30,760
providing the smoothness that
lets your mind stay in the

640
00:55:30,760 --> 00:55:35,040
story.
This is why Nvidia's rise can be

641
00:55:35,040 --> 00:55:43,040
told as a bedtime story at all,
because it is, at heart about

642
00:55:43,040 --> 00:55:49,200
making experiences feel calm and
continuous, even when the work

643
00:55:49,200 --> 00:55:56,480
underneath is vast, The world
slows, and the frames keep

644
00:55:56,480 --> 00:56:00,480
coming.
A chip is a piece of silicon,

645
00:56:01,040 --> 00:56:06,760
but a platform is a habit.
CUDA became that habit slowly,

646
00:56:06,880 --> 00:56:11,520
as more developers learned it,
as more libraries were built on

647
00:56:11,520 --> 00:56:16,760
top of it, and as more research
results depended on its rhythms.

648
00:56:17,120 --> 00:56:23,080
Overtime, Nvidia's advantage was
not only that its GPU's were

649
00:56:23,080 --> 00:56:27,960
fast, but that the path to use
them was well lit.

650
00:56:28,560 --> 00:56:34,760
Documentation tools optimized
code for common tasks.

651
00:56:35,080 --> 00:56:40,480
In machine learning, people
began to rely on libraries that

652
00:56:40,480 --> 00:56:45,040
handled the hardest parts.
The deep inner loops of

653
00:56:45,040 --> 00:56:51,040
training, the careful math for
neural networks, the patterns

654
00:56:51,040 --> 00:56:54,400
that appear in almost every
model.

655
00:56:54,680 --> 00:56:59,280
When those libraries are tuned
for a specific platform, the

656
00:56:59,280 --> 00:57:05,240
platform becomes sticky not
through coercion, but through

657
00:57:05,240 --> 00:57:10,560
convenience and trust.
Trust matters in computing

658
00:57:11,040 --> 00:57:15,920
because a researcher does not
want to spend months debugging

659
00:57:15,920 --> 00:57:21,240
hardware quirks when the real
goal is discovery.

660
00:57:21,640 --> 00:57:26,640
A company building an AI product
does not want to reinvent the

661
00:57:26,640 --> 00:57:32,520
foundations, it wants to stand
on something stable and keep

662
00:57:32,520 --> 00:57:36,720
moving.
So NVIDIA kept doing what it had

663
00:57:36,720 --> 00:57:42,720
learned from gaming Cultivate
developers, support the people

664
00:57:42,720 --> 00:57:48,400
building content, because
content is what makes hardware

665
00:57:48,400 --> 00:57:52,920
meaningful.
In games, the content is worlds

666
00:57:52,920 --> 00:57:59,960
and stories and competition.
In AI, the content is models and

667
00:58:00,000 --> 00:58:04,720
data and applications.
Speech recognition, vision

668
00:58:04,720 --> 00:58:11,120
systems, recommendations,
scientific tools, creative

669
00:58:11,120 --> 00:58:14,800
assistance.
The platform connects the mall

670
00:58:15,080 --> 00:58:20,520
like a slow river that carries
many boats, and the river's

671
00:58:20,520 --> 00:58:25,520
value increases as more people
choose to float on it.

672
00:58:25,920 --> 00:58:29,880
Of course, the river is not
alone.

673
00:58:30,760 --> 00:58:36,320
Competitors exist, alternative
approaches exist, and the

674
00:58:36,320 --> 00:58:40,480
industry is always searching for
efficiency.

675
00:58:40,920 --> 00:58:45,760
But Nvidia's strategy was steady
Build the ecosystem, build the

676
00:58:45,760 --> 00:58:51,240
hardware, and keep them aligned
so the experience stays smooth.

677
00:58:51,640 --> 00:58:56,480
If you listen closely, you can
hear how this echoes the

678
00:58:56,520 --> 00:59:01,120
earliest days.
The same desire to reduce

679
00:59:01,120 --> 00:59:07,000
friction, to make complex work
feel simple, to keep the images

680
00:59:07,160 --> 00:59:13,440
or the learning flowing.
In supercomputing centers, GPU's

681
00:59:13,440 --> 00:59:18,840
became standard equipment, and
in cloud data centers they

682
00:59:18,840 --> 00:59:24,400
became rentable resources.
So a student with a small budget

683
00:59:24,800 --> 00:59:30,240
could access the same kind of
acceleration as a large lab.

684
00:59:30,760 --> 00:59:34,200
That accessibility changed the
pace of innovation.

685
00:59:34,800 --> 00:59:38,840
More experiments, more
iterations, more chances for

686
00:59:38,840 --> 00:59:43,080
unexpected breakthroughs.
And each time someone learned

687
00:59:43,080 --> 00:59:48,880
CUDA, taught it, used it, or
built on it, the ecosystem

688
00:59:48,880 --> 00:59:52,360
deepened.
The way a city becomes more

689
00:59:52,360 --> 00:59:58,600
livable as more roads, signs and
shared customs appear.

690
00:59:58,880 --> 01:00:01,560
It's a gentle kind of
compounding.

691
01:00:01,560 --> 01:00:05,400
Less dramatic than a headline,
but more durable.

692
01:00:05,920 --> 01:00:11,080
The accumulation of tools,
habits, and trust all circling a

693
01:00:11,200 --> 01:00:14,960
parallel engine.
So when the world suddenly

694
01:00:14,960 --> 01:00:18,840
wanted more AI, NVIDIA was
already there.

695
01:00:19,040 --> 01:00:23,880
Not just with chips, but with a
well worn path that made those

696
01:00:23,880 --> 01:00:28,080
chips usable as NVIDIA grew into
AI.

697
01:00:28,480 --> 01:00:34,320
It also began to imagine where
AI would live, not only in data

698
01:00:34,320 --> 01:00:39,520
centers, but in machines that
move through the physical world.

699
01:00:39,960 --> 01:00:45,360
A self driving car is, in a
quiet sense, a perception

700
01:00:45,360 --> 01:00:49,000
machine.
Cameras and sensors streaming

701
01:00:49,000 --> 01:00:52,560
data, software deciding what
matters.

702
01:00:53,160 --> 01:00:56,520
Computation, turning patterns
into action.

703
01:00:56,920 --> 01:01:01,880
This requires fast inference,
the moment to moment use of a

704
01:01:01,880 --> 01:01:08,080
trained model, and it requires
reliability because the road

705
01:01:08,360 --> 01:01:12,000
does not pause while the
computer thinks so.

706
01:01:12,000 --> 01:01:17,560
NVIDIA built platforms for
automotive computing, pairing

707
01:01:17,560 --> 01:01:21,720
specialized hardware with
software that could process

708
01:01:21,720 --> 01:01:26,120
sensor data and run neural
networks in real time.

709
01:01:26,520 --> 01:01:31,440
The goal was not to sell a
single chip, but to offer a

710
01:01:31,440 --> 01:01:35,880
foundation that car makers and
robotics teams could build on,

711
01:01:36,480 --> 01:01:41,440
adapting it to their needs.
At the smaller end, NVIDIA

712
01:01:41,440 --> 01:01:48,080
created compact modules for Edge
AI, tiny computers with GPU

713
01:01:48,080 --> 01:01:54,760
power used in drones, cameras,
robots, and industrial devices.

714
01:01:55,080 --> 01:02:00,720
These systems are like the data
center's distant cousins, still

715
01:02:00,720 --> 01:02:06,760
parallel, still repetitive, but
living closer to the world of

716
01:02:06,760 --> 01:02:12,680
motion and dust and vibration.
And beyond physical machines,

717
01:02:12,920 --> 01:02:18,120
NVIDIA leaned into simulation,
the idea that you can build a

718
01:02:18,120 --> 01:02:23,040
digital environment, a kind of
twin, and test systems there

719
01:02:23,040 --> 01:02:26,960
safely.
Simulation has always been part

720
01:02:26,960 --> 01:02:32,280
of Nvidia's heritage, because
graphics is a form of

721
01:02:32,280 --> 01:02:37,640
simulation, and the GPU is a
natural tool for building

722
01:02:37,640 --> 01:02:41,800
believable virtual spaces in
robotics.

723
01:02:42,280 --> 01:02:49,560
Simulation is a gentle teacher,
a place where a robot can fail

724
01:02:49,560 --> 01:02:54,800
without harm, learn without
fear, and repeat an action

725
01:02:54,800 --> 01:02:59,080
thousands of times without
getting tired.

726
01:02:59,400 --> 01:03:04,840
In industry, simulation helps
engineers model factories,

727
01:03:04,960 --> 01:03:11,000
logistics, traffic flow, and
energy systems, letting them

728
01:03:11,000 --> 01:03:16,240
explore possibilities before
making expensive changes.

729
01:03:16,560 --> 01:03:22,680
So the company's story broadened
again, from drawing pixels to

730
01:03:22,680 --> 01:03:27,680
accelerating science, to
supporting learning, and now to

731
01:03:27,680 --> 01:03:33,160
helping machines perceive and
act in the physical world.

732
01:03:33,480 --> 01:03:40,600
It is all 1 theme really, giving
computation a body and giving

733
01:03:40,600 --> 01:03:46,560
complex systems a calm place to
practice before they meet

734
01:03:46,560 --> 01:03:49,920
reality.
NVIDIA called some of these

735
01:03:50,080 --> 01:03:56,880
simulation efforts Omniverse, a
shared 3D workspace where many

736
01:03:56,880 --> 01:04:03,200
tools can meet and where scenes
can be rendered, changed and

737
01:04:03,200 --> 01:04:08,240
replayed with quiet precision.
For robotics, they built

738
01:04:08,240 --> 01:04:13,880
software stacks that helped
teams train and deploy systems,

739
01:04:14,400 --> 01:04:19,680
often blending computer vision,
planning and control into one

740
01:04:19,680 --> 01:04:23,680
coherent flow.
If you picture it, it feels like

741
01:04:23,680 --> 01:04:28,680
a calm rehearsal room.
Virtual light, virtual gravity,

742
01:04:29,040 --> 01:04:35,000
virtual friction, all computed
patiently so the real machine

743
01:04:35,240 --> 01:04:41,800
can move with confidence later.
And so even here, the GPU's role

744
01:04:41,800 --> 01:04:46,800
remains consistent.
It makes rich simulation and

745
01:04:46,800 --> 01:04:51,960
fast learning possible, turning
repetition into readiness.

746
01:04:52,280 --> 01:04:56,880
While AI drew headlines,
another, quieter thread

747
01:04:56,880 --> 01:05:01,440
continued, one that NVIDIA had
always served the work of

748
01:05:01,440 --> 01:05:06,200
creators who build images,
spaces and stories.

749
01:05:06,600 --> 01:05:10,320
Film studios render scenes that
never existed.

750
01:05:10,600 --> 01:05:13,800
Architects preview buildings
before they are built.

751
01:05:14,360 --> 01:05:19,760
Designers rotate objects in 3D
until the shape feels right.

752
01:05:20,120 --> 01:05:24,960
These tasks are heavy with
geometry and light, and they

753
01:05:24,960 --> 01:05:30,640
benefit from the same parallel
computation that powers games.

754
01:05:31,000 --> 01:05:36,160
So Nvidia's professional
visualization tools kept

755
01:05:36,160 --> 01:05:41,280
evolving, supporting artists who
care about subtle reflections,

756
01:05:41,680 --> 01:05:46,520
soft shadows, the way a surface
looks under morning light.

757
01:05:46,920 --> 01:05:52,440
In some ways, this creative
thread is a gentle bridge

758
01:05:52,440 --> 01:05:59,760
between eras, because it reminds
us that the GPU's earliest

759
01:05:59,840 --> 01:06:04,600
purpose was to help imagination
appear on a screen.

760
01:06:05,040 --> 01:06:10,680
What changed is scale and
audience, and the way images now

761
01:06:10,680 --> 01:06:15,400
travel.
A render created in one city can

762
01:06:15,400 --> 01:06:21,880
be shared instantly, edited
collaboratively, and experienced

763
01:06:21,880 --> 01:06:27,120
on a phone across the world.
The boundary between real and

764
01:06:27,120 --> 01:06:31,240
rendered grew softer, and not in
a frightening way, but in a

765
01:06:31,240 --> 01:06:38,200
practical 1 because simulation
helps us plan, rehearse, and

766
01:06:38,200 --> 01:06:41,520
communicate.
Even a simple social media

767
01:06:41,520 --> 01:06:46,720
filter, smoothing a face or
adjusting a background is part

768
01:06:46,720 --> 01:06:50,960
of this lineage.
A tiny piece of accelerated

769
01:06:50,960 --> 01:06:56,320
computation applied to everyday
life, and when AI began

770
01:06:56,320 --> 01:07:01,640
generating images from text
prompts, it did so using learned

771
01:07:01,640 --> 01:07:08,400
patterns that still rely on GPU
accelerated math, echoing the

772
01:07:08,400 --> 01:07:12,640
old relationship between
computation and art.

773
01:07:13,080 --> 01:07:18,680
You can see the continuity from
a gamer watching a new lighting

774
01:07:18,680 --> 01:07:25,200
effect in a 2004 title to an
artist using a neural tool to

775
01:07:25,200 --> 01:07:32,320
explore variations in 2024, both
seeking the same thing, not

776
01:07:32,320 --> 01:07:38,560
novelty for its own sake, but a
smoother path from idea to

777
01:07:38,600 --> 01:07:42,040
expression.
In the sleep wise pace of this

778
01:07:42,040 --> 01:07:47,440
night, it's comforting to notice
how often technology's goal is

779
01:07:47,440 --> 01:07:52,800
simply to remove friction so the
mind can stay with the creative

780
01:07:52,800 --> 01:07:57,400
act.
The GPU, steady and patient, has

781
01:07:57,400 --> 01:08:03,160
been doing that for decades,
turning mathematics into images

782
01:08:03,640 --> 01:08:08,320
and now turning images and
language into new kinds of

783
01:08:08,320 --> 01:08:13,560
shared possibility.
Real time rendering also changed

784
01:08:13,560 --> 01:08:18,600
how film and media are made,
with directors walking through

785
01:08:18,600 --> 01:08:23,000
virtual sets before cameras
roll, seeing lighting and

786
01:08:23,000 --> 01:08:27,479
composition in the moment
instead of waiting days for

787
01:08:27,479 --> 01:08:31,399
final renders.
That shift, too, is a kind of

788
01:08:31,640 --> 01:08:38,120
calm fewer surprises, more
iteration, more room for gentle

789
01:08:38,120 --> 01:08:41,800
adjustment.
And as the tools became faster,

790
01:08:42,080 --> 01:08:46,279
the creative process became more
like conversation.

791
01:08:46,960 --> 01:08:51,680
Try aversion.
See it instantly refine, repeat

792
01:08:52,279 --> 01:08:54,840
until the result feels
inevitable.

793
01:08:55,240 --> 01:09:00,479
In that way, Nvidia's rise is
not only about machines

794
01:09:00,479 --> 01:09:07,240
learning, but about humans
creating with less delay, as if

795
01:09:07,240 --> 01:09:12,600
the distance between thought and
screen has quietly shortened.

796
01:09:12,960 --> 01:09:18,160
Every large story has its
limits, and for modern computing

797
01:09:18,479 --> 01:09:25,560
the limits are often physical
energy, materials, manufacturing

798
01:09:25,560 --> 01:09:31,439
capacity, and the quiet
complexity of global supply

799
01:09:31,439 --> 01:09:35,680
chains.
A high end processor is not just

800
01:09:35,680 --> 01:09:42,080
design, it is fabrication,
packaging, shipping, testing,

801
01:09:42,520 --> 01:09:47,640
and the careful assembly of
components that arrive from many

802
01:09:47,640 --> 01:09:51,600
places.
In the AI boom, those chains

803
01:09:51,600 --> 01:09:57,200
became visible because demand
rose quickly and the world

804
01:09:57,200 --> 01:10:02,520
learned that not all chips can
be produced at will.

805
01:10:02,960 --> 01:10:07,120
There were also rules and
restrictions shifting over time,

806
01:10:07,520 --> 01:10:11,920
influencing where certain
advanced technologies could be

807
01:10:11,920 --> 01:10:18,240
sold, reminding everyone that
computing lives inside a wider

808
01:10:18,240 --> 01:10:21,960
world of policies and
priorities.

809
01:10:22,320 --> 01:10:28,680
NVIDIA navigated this by doing
what it has always done,

810
01:10:29,360 --> 01:10:34,800
focusing on engineering, on
efficiency, on making each

811
01:10:34,800 --> 01:10:39,960
generation of hardware do more
work per unit of energy.

812
01:10:40,080 --> 01:10:45,680
Efficiency is a soothing concept
when you think about it.

813
01:10:46,480 --> 01:10:51,520
The idea that you can accomplish
more with less strain, that

814
01:10:51,520 --> 01:10:54,760
progress does not have to mean
waste.

815
01:10:55,080 --> 01:11:00,480
In data centers, that efficiency
shows up as better cooling,

816
01:11:00,800 --> 01:11:05,160
smarter scheduling, and
architectures that avoid

817
01:11:05,200 --> 01:11:12,000
unnecessary movement of data.
Because moving data can be as

818
01:11:12,000 --> 01:11:17,120
costly as computing.
In chips, it shows up as

819
01:11:17,120 --> 01:11:23,960
specialized units for common
operations, and in software, it

820
01:11:23,960 --> 01:11:28,800
shows up as libraries that
squeeze more usefulness out of

821
01:11:28,800 --> 01:11:32,480
each computation.
All of this is a kind of

822
01:11:32,680 --> 01:11:37,600
invisible stewardship, the quiet
effort to make the system

823
01:11:37,600 --> 01:11:42,320
sustainable.
As it scales and as AI becomes

824
01:11:42,320 --> 01:11:47,360
more common, the world will keep
asking the same calm question in

825
01:11:47,360 --> 01:11:51,640
new forms.
How do we build intelligence

826
01:11:51,880 --> 01:11:54,080
without building too much
friction?

827
01:11:54,440 --> 01:12:00,960
Nvidia's answer again and again
has been acceleration paired

828
01:12:00,960 --> 01:12:07,040
with integration, making the
processor faster, yes, but also

829
01:12:07,040 --> 01:12:09,720
making the whole pathway
smoother.

830
01:12:10,200 --> 01:12:15,760
Tonight you don't need to hold
the details, only the feeling

831
01:12:15,760 --> 01:12:20,240
that behind the scenes, an
immense amount of careful

832
01:12:20,240 --> 01:12:25,080
engineering is trying to keep
the future gentle.

833
01:12:25,360 --> 01:12:30,960
If you picture a silicon wafer,
it's like a dark mirror, round

834
01:12:30,960 --> 01:12:35,520
and delicate, carrying many
future machines in a single

835
01:12:35,520 --> 01:12:38,920
sheet.
Lithography prints patterns too

836
01:12:38,920 --> 01:12:45,400
small to see layer by layer, and
each layer must align as if the

837
01:12:45,400 --> 01:12:50,480
chip is being written in a
microscopic calligraphy.

838
01:12:50,840 --> 01:12:55,240
Then comes packaging, where
memory and processor are brought

839
01:12:55,240 --> 01:12:59,120
close together, and where the
physical distance between

840
01:12:59,120 --> 01:13:03,200
components becomes part of
performance.

841
01:13:03,560 --> 01:13:09,120
When people say compute is the
new oil, they often forget how

842
01:13:09,120 --> 01:13:14,760
tangible it is.
Metal, water, power lines and

843
01:13:14,760 --> 01:13:18,480
the calm logistics of keeping
everything running.

844
01:13:18,800 --> 01:13:23,520
So the rise of AI has quietly
renewed an old respect for

845
01:13:23,520 --> 01:13:29,560
infrastructure, and NVIDIA, a
company born in graphics, has

846
01:13:29,560 --> 01:13:33,440
become one of the architects of
that infrastructure.

847
01:13:33,760 --> 01:13:38,800
No company rises alone, and
Nvidia's story has always

848
01:13:38,800 --> 01:13:44,040
included rivals, alternatives
and the steady pressure that

849
01:13:44,040 --> 01:13:47,960
keeps a craft honest.
In gaming, there were

850
01:13:47,960 --> 01:13:53,400
competitors making their own
graphics cards, pushing price

851
01:13:53,400 --> 01:13:58,640
and performance in a cycle that
benefited players in data

852
01:13:58,640 --> 01:14:02,880
centers.
Cloud companies explored custom

853
01:14:02,880 --> 01:14:07,840
chips, building specialized
processors for their own needs,

854
01:14:08,160 --> 01:14:13,320
and researchers experimented
with many kinds of accelerators.

855
01:14:13,640 --> 01:14:19,960
This is natural because when
something becomes foundational,

856
01:14:20,680 --> 01:14:26,440
everyone wants to shape it.
Yet Nvidia's endurance has often

857
01:14:26,440 --> 01:14:32,320
come from its platform approach,
not only shipping hardware, but

858
01:14:32,320 --> 01:14:36,600
supporting it with software that
makes the hardware practical.

859
01:14:37,000 --> 01:14:42,600
A custom chip can be powerful,
but power without a mature

860
01:14:42,680 --> 01:14:49,000
ecosystem can feel lonely, like
a road without signs.

861
01:14:49,320 --> 01:14:55,480
So NVIDIA continued to invest in
the tools that let people build

862
01:14:55,480 --> 01:15:01,600
quickly Compilers, libraries,
drivers, and a wide set of

863
01:15:01,600 --> 01:15:06,120
partnerships across academia and
industry.

864
01:15:06,520 --> 01:15:11,120
Competition also pushed NVIDIA
to stay efficient, to keep

865
01:15:11,120 --> 01:15:16,520
improving performance per Watt,
to design systems that can scale

866
01:15:16,520 --> 01:15:21,920
without losing reliability.
And in the background, a quieter

867
01:15:21,920 --> 01:15:26,720
competition existed as well.
The competition against time,

868
01:15:27,120 --> 01:15:32,320
against complexity, against the
tendency of systems to become

869
01:15:32,360 --> 01:15:36,600
fragile as they grow.
It's surprisingly hard to keep

870
01:15:36,600 --> 01:15:43,160
an ecosystem stable while it
expands, and stability is part

871
01:15:43,160 --> 01:15:47,640
of what people pay for, even if
they don't always name it.

872
01:15:48,040 --> 01:15:54,320
So the rise of NVIDIA is not a
story of a single victory, but

873
01:15:54,320 --> 01:16:00,800
of continuous tuning, like
keeping an instrument in tune as

874
01:16:00,800 --> 01:16:06,960
the temperature changes tonight,
it's calming to remember that

875
01:16:06,960 --> 01:16:13,440
progress in technology is
usually not a conquest, but a

876
01:16:13,440 --> 01:16:20,280
conversation.
Many ideas, many approaches, and

877
01:16:20,280 --> 01:16:25,400
a slow selection of what proves
dependable.

878
01:16:25,720 --> 01:16:30,680
NVIDIA has been one of the
dependable answers for a long

879
01:16:30,680 --> 01:16:36,600
time, partly because it learned
early that survival requires

880
01:16:36,600 --> 01:16:42,000
both speed and care.
When engineers choose tools,

881
01:16:42,320 --> 01:16:48,200
they often choose the one that
reduces uncertainty, the one

882
01:16:48,200 --> 01:16:53,440
that works across operating
systems, across clusters, across

883
01:16:53,800 --> 01:16:58,000
generations of hardware.
They choose the one with a

884
01:16:58,000 --> 01:17:03,480
community, with examples, with
people who have already solved

885
01:17:03,480 --> 01:17:07,280
the sharp corners.
That is why, even as

886
01:17:07,280 --> 01:17:13,320
alternatives grow stronger, many
teams still return to Nvidia's

887
01:17:13,320 --> 01:17:17,560
stack.
Because it feels like a well

888
01:17:17,560 --> 01:17:20,600
maintained path through dense
forest.

889
01:17:20,960 --> 01:17:24,880
And as those alternatives
improve, the whole industry

890
01:17:24,880 --> 01:17:28,800
improves.
Better standards, better ideas,

891
01:17:28,800 --> 01:17:34,840
better efficiency, all feeding
back into a healthier ecosystem.

892
01:17:35,200 --> 01:17:41,640
In that way, competition becomes
another form of calm, a steady

893
01:17:41,640 --> 01:17:45,520
wind that keeps the air moving,
preventing stagnation,

894
01:17:45,760 --> 01:17:49,600
encouraging everyone to keep
refining their craft.

895
01:17:49,960 --> 01:17:56,000
If you step back, Nvidia's rise
can feel like a shift from the

896
01:17:56,000 --> 01:18:01,120
visible to the invisible.
At first, you could see the

897
01:18:01,120 --> 01:18:06,400
change on your own screen.
Smoother motion, Richer colors.

898
01:18:06,960 --> 01:18:11,720
A game world that suddenly felt
more alive.

899
01:18:12,160 --> 01:18:16,320
Then the change moved into
places you rarely visit.

900
01:18:16,800 --> 01:18:21,840
Server rooms, research labs,
data centers at the edge of

901
01:18:21,840 --> 01:18:25,960
towns, humming behind locked
doors.

902
01:18:26,280 --> 01:18:30,520
And now the results return to
you through everyday

903
01:18:30,520 --> 01:18:34,120
experiences.
Better translations.

904
01:18:34,440 --> 01:18:41,640
Smarter search recommendations,
Image tools, voice assistance,

905
01:18:42,280 --> 01:18:47,040
and the quiet sense that
software is becoming more

906
01:18:47,040 --> 01:18:50,840
conversational.
It is the loop of modern

907
01:18:50,840 --> 01:18:55,000
infrastructure.
Build it far away and it comes

908
01:18:55,000 --> 01:19:01,080
back as something intimate.
There is also a human loop, a

909
01:19:01,080 --> 01:19:05,840
loop of patience, founders
taking risks, engineers

910
01:19:05,840 --> 01:19:10,520
iterating, developers adopting,
researchers discovering

911
01:19:10,880 --> 01:19:16,120
companies building on top.
The story rewards the same

912
01:19:16,120 --> 01:19:22,440
qualities again and again,
endurance, clarity and the

913
01:19:22,440 --> 01:19:27,480
willingness to invest in tools
that pay off slowly.

914
01:19:27,840 --> 01:19:30,960
It is tempting to focus on the
famous moments.

915
01:19:31,440 --> 01:19:38,280
The first time the word GPU was
said, The release of CUDA, the

916
01:19:38,280 --> 01:19:43,320
Imagenet breakthrough, the
arrival of generative AI.

917
01:19:43,680 --> 01:19:46,720
But the sleep wise truth is
softer.

918
01:19:47,320 --> 01:19:52,720
Those moments mattered because
thousands of quieter decisions

919
01:19:53,360 --> 01:19:57,520
made them possible.
A driver update that improves

920
01:19:57,520 --> 01:20:01,000
stability.
A library that speeds up a

921
01:20:01,000 --> 01:20:05,040
common operation.
An engineering practice that

922
01:20:05,040 --> 01:20:10,240
prevents A costly error.
A support team that answers a

923
01:20:10,240 --> 01:20:15,800
question at the right time.
This is the texture of real

924
01:20:15,800 --> 01:20:21,000
progress, Not fireworks, but
steady refinement.

925
01:20:21,400 --> 01:20:27,040
And perhaps that is why NVIDIA
fits so well into a bedtime

926
01:20:27,040 --> 01:20:30,920
story.
Because it is not a tale of

927
01:20:30,920 --> 01:20:34,920
sudden magic, but of compounding
craft.

928
01:20:35,360 --> 01:20:41,200
A company that began by making
pixels move smoothly spent

929
01:20:41,200 --> 01:20:46,000
decades learning how to make
computation itself move

930
01:20:46,000 --> 01:20:51,640
smoothly, and in doing so, it
helped the world build new kinds

931
01:20:51,640 --> 01:20:55,760
of thinking.
Tonight we can let that idea

932
01:20:56,200 --> 01:21:01,240
settle gently, that behind the
modern glow of screens and

933
01:21:01,240 --> 01:21:06,640
services there are patient
engines and patient people

934
01:21:07,320 --> 01:21:12,720
repeating small steps until the
future feels natural.

935
01:21:13,080 --> 01:21:17,000
Some of those engines sit in
cold climates, where air is

936
01:21:17,000 --> 01:21:22,800
cheap and cooling is easy, and
some sit in warm places where

937
01:21:22,800 --> 01:21:26,800
cooling systems work quietly all
day.

938
01:21:27,200 --> 01:21:32,720
They sit near fiber lines and
power substations, near highways

939
01:21:32,720 --> 01:21:38,000
that deliver replacement parts,
near the unnoticed logistics

940
01:21:38,000 --> 01:21:43,360
that keep everything dependable.
And while we sleep, they run

941
01:21:44,000 --> 01:21:49,320
training, serving, calculating,
making the modern world feel

942
01:21:49,320 --> 01:21:55,160
responsive, as if intelligence
is always on call.

943
01:21:55,520 --> 01:22:00,160
It is a strange comfort, when
you think about it, that so much

944
01:22:00,160 --> 01:22:04,280
activity can exist without
demanding our attention.

945
01:22:05,120 --> 01:22:10,600
A hum we don't have to hear, A
light we don't have to watch.

946
01:22:11,000 --> 01:22:17,400
So tonight, as the room darkens,
we can carry Nvidia's story like

947
01:22:17,440 --> 01:22:22,160
a small Lantern.
Steady, not urgent.

948
01:22:22,600 --> 01:22:26,920
A company begins with a
question, survives its early

949
01:22:26,920 --> 01:22:33,240
missteps, learns to build with
rigor, and then discovers that

950
01:22:33,240 --> 01:22:36,400
its tool can serve more than the
first dream.

951
01:22:36,800 --> 01:22:42,520
First it helps the world see
smoother games, richer images,

952
01:22:42,960 --> 01:22:48,160
light rendered with patients.
Then it helps the world measure

953
01:22:48,160 --> 01:22:54,360
and simulate molecules, weather,
medical images, the slow

954
01:22:54,360 --> 01:22:57,040
problems that require
repetition.

955
01:22:57,440 --> 01:23:02,800
Then it helps the world learn
networks, adjusting their

956
01:23:02,800 --> 01:23:09,560
weights, patterns emerging from
data machines, becoming better

957
01:23:09,560 --> 01:23:14,840
at recognizing what we already
know how to recognize.

958
01:23:15,200 --> 01:23:22,000
And then, quietly, it helps the
world speak language models that

959
01:23:22,000 --> 01:23:28,200
respond, image models that
imagine systems that feel like

960
01:23:28,200 --> 01:23:31,320
assistance rather than
calculators.

961
01:23:31,720 --> 01:23:37,840
Underneath every stage is the
same rhythm, parallel work, many

962
01:23:37,840 --> 01:23:42,400
small calculations done
together, so complexity can

963
01:23:42,400 --> 01:23:47,200
arrive without strain.
You can picture it now as a

964
01:23:47,200 --> 01:23:53,640
gentle current, electrons moving
through etched pathways, heat

965
01:23:53,760 --> 01:23:58,920
carried away by airflow, status
lights blinking softly in a

966
01:23:58,920 --> 01:24:02,000
distant room.
And you can picture the older

967
01:24:02,000 --> 01:24:07,600
version, too.
A bedroom PC in 1999, a bulky

968
01:24:07,600 --> 01:24:13,040
monitor, a new graphics card.
A game loading, a world

969
01:24:13,040 --> 01:24:17,480
appearing frame by frame.
If there is a lesson worth

970
01:24:17,480 --> 01:24:22,920
keeping, it is that technology
rarely arrives in a single leap.

971
01:24:23,680 --> 01:24:29,320
It arrives through patience,
iteration and years of careful

972
01:24:29,320 --> 01:24:32,160
building.
It moves through people willing

973
01:24:32,160 --> 01:24:37,720
to work on foundations, to make
tools reliable, to choose the

974
01:24:37,720 --> 01:24:41,560
long path when the short 1 looks
tempting.

975
01:24:41,960 --> 01:24:46,560
Now the world slows, and you do
not need to hold any of it

976
01:24:46,560 --> 01:24:50,200
tightly.
Let the images fade, let the

977
01:24:50,200 --> 01:24:56,360
servers hum somewhere far away,
and let parallel work become a

978
01:24:56,360 --> 01:25:03,240
lullaby, small efforts working
together so the whole system can

979
01:25:03,240 --> 01:25:07,240
rest somewhere.
A founder's early breakfast

980
01:25:07,240 --> 01:25:12,880
conversation becomes a decade of
prototypes, then a platform,

981
01:25:13,440 --> 01:25:19,320
then an ecosystem that thousands
of engineers quietly depend on.

982
01:25:19,640 --> 01:25:25,200
Somewhere a creator renders a
scene in real time, adjusting

983
01:25:25,200 --> 01:25:31,320
light until it feels like dusk,
and the GPU turns math into

984
01:25:31,320 --> 01:25:37,360
atmosphere without complaint.
And somewhere a curious person

985
01:25:37,360 --> 01:25:42,640
opens a small window on a
screen, asks a question, and

986
01:25:42,640 --> 01:25:47,720
receives an answer that feels
calm, as if the machine has

987
01:25:47,720 --> 01:25:50,560
learned the contours of
language.

988
01:25:50,960 --> 01:25:55,600
All of it is connected not by
spectacle, but by steady

989
01:25:55,600 --> 01:26:00,880
repetition, the kind that makes
hard things seem simple and

990
01:26:00,880 --> 01:26:07,160
makes simple things feel smooth.
So as you settle deeper into the

991
01:26:07,160 --> 01:26:13,000
night, you can let the story
drift away from names and dates

992
01:26:13,640 --> 01:26:16,120
and toward a softer
understanding.

993
01:26:17,000 --> 01:26:22,160
Progress is often a quiet
compounding of small choices.

994
01:26:22,560 --> 01:26:27,120
The hum continues in the
distance like wind through

995
01:26:27,120 --> 01:26:33,520
trees, like water through pipes,
like a city breathing while its

996
01:26:33,520 --> 01:26:38,160
streets sleep.
And here in your own room, you

997
01:26:38,160 --> 01:26:43,960
can do the same.
Breathe, soften, and let the day

998
01:26:43,960 --> 01:26:46,040
dissolve.
Good night.