1
00:00:01,430 --> 00:00:04,757
Thank you so much
for being at Build again.

2
00:00:04,840 --> 00:00:07,157
I know it's late for you in Taipei.

3
00:00:07,240 --> 00:00:09,510
I really appreciate you staying up.

4
00:00:10,440 --> 00:00:17,167
You know, at the end, I've been looking
at social and you know, and everything

5
00:00:17,250 --> 00:00:20,567
people have been talking about
since your keynote over the weekend.

6
00:00:20,650 --> 00:00:26,107
And suddenly, you know,
this concept of unmetered intelligence

7
00:00:26,190 --> 00:00:29,447
right at the edge is so hot again.

8
00:00:29,530 --> 00:00:31,217
So maybe you want to talk a little bit.

9
00:00:31,300 --> 00:00:33,217
You've thought about this,
talked about this.

10
00:00:33,300 --> 00:00:37,287
And now of course with RTX
Spark really delivered,

11
00:00:37,370 --> 00:00:42,257
I think what's a breakthrough system
for AI to be much more ubiquitous.

12
00:00:42,340 --> 00:00:46,097
But maybe Jensen, you can just share
a little bit your vision around where

13
00:00:46,180 --> 00:00:47,310
you see this going.

14
00:00:48,520 --> 00:00:50,837
Well, this all started
about three years ago

15
00:00:50,920 --> 00:00:55,107
between a conversation between you
and I, and we were talking about how

16
00:00:55,190 --> 00:00:58,207
we could build a new class of PCs.

17
00:00:58,290 --> 00:01:02,717
That's incredible for designers
and creators, and it would be incredible

18
00:01:02,800 --> 00:01:06,087
for artificial intelligence.
And it would be one of these systems

19
00:01:06,170 --> 00:01:11,117
that has the processing capability,
but also the software stack

20
00:01:11,200 --> 00:01:15,127
that's integrated into the world's
design packages and creator packages

21
00:01:15,210 --> 00:01:17,627
and of course, all the things
that we're doing with AI.

22
00:01:17,710 --> 00:01:23,167
And here we are three years later,
we built an incredible new chip.

23
00:01:23,250 --> 00:01:29,107
And this system is supported
by all of this new software

24
00:01:29,190 --> 00:01:33,677
that you created for Windows. And we
now have the ability to have essentially

25
00:01:33,760 --> 00:01:37,530
an autonomous agent running on the PC.

26
00:01:37,800 --> 00:01:41,347
Now, when you take a step back and you
think about what does that mean?

27
00:01:41,430 --> 00:01:45,557
For the 40 years, that or some 30
years we've been working together,

28
00:01:45,640 --> 00:01:52,557
we went from inventing DirectX
together to creating now

29
00:01:52,640 --> 00:01:56,437
this incredible computer that has
autonomous systems running.

30
00:01:56,520 --> 00:02:01,667
The PC evolved from being
an incredible tool

31
00:02:01,750 --> 00:02:06,677
to now being a tool that's used
autonomously by an AI assistant.

32
00:02:06,760 --> 00:02:11,747
And so the idea that I could be
traveling and I'm on the phone

33
00:02:11,830 --> 00:02:16,757
and I could text my PC and ask
my PC to get some coding done

34
00:02:16,840 --> 00:02:21,970
or some idea that I have, and it
would fire up the tools on the PC,

35
00:02:22,380 --> 00:02:26,597
and it would make the modifications
or the changes or the design that I was,

36
00:02:26,680 --> 00:02:32,197
I told it to to do, and it would iterate
with me while I'm away from the PC.

37
00:02:32,280 --> 00:02:35,807
My PC became an assistant.
While I'm sitting there, of course,

38
00:02:35,890 --> 00:02:38,137
this PC would be my great
assistant as well.

39
00:02:38,220 --> 00:02:44,117
And so this idea that that the PC
evolved from a personal computer

40
00:02:44,200 --> 00:02:47,547
to a personal AI,
it's just really exciting.

41
00:02:47,630 --> 00:02:52,887
And to see it come to life, Satya, to see
it come to life and actually doing that,

42
00:02:52,970 --> 00:02:54,727
you know, so I'm super excited about it.

43
00:02:54,810 --> 00:02:59,597
Spark, as you mentioned earlier,
has all this incredible capabilities,

44
00:02:59,680 --> 00:03:05,367
a petaflop of AI performance.
It has a petaflop of NVFP4,

45
00:03:05,450 --> 00:03:09,007
this numerical format that our two
companies worked on together.

46
00:03:09,090 --> 00:03:15,047
That allows us to take advantage
of this 128GB of memory and fit

47
00:03:15,130 --> 00:03:18,577
maybe a couple of hundred
billion parameter model.

48
00:03:18,660 --> 00:03:21,917
A couple of hundred billion
parameter model is state of the art.

49
00:03:22,000 --> 00:03:25,957
And so I think the days of having
a really smart assistant running

50
00:03:26,040 --> 00:03:27,727
autonomously on the PC is here.

51
00:03:27,810 --> 00:03:31,097
Yeah, I know it's so awesome. And in fact,
I'm also excited about Windows

52
00:03:31,180 --> 00:03:36,567
coming to the GB300. And so that's another
thing that it's kind of like data center

53
00:03:36,650 --> 00:03:38,797
right on your desktop.
And it's so exciting.

54
00:03:38,880 --> 00:03:45,037
But talking about that data center side,
obviously, you know, this entire thing

55
00:03:45,120 --> 00:03:48,447
got started when we built
the first supercomputer

56
00:03:48,530 --> 00:03:51,377
together to train the GPT models.

57
00:03:51,460 --> 00:03:56,117
And we have come a long way. In fact, even
I was talking about the Fairwater design.

58
00:03:56,200 --> 00:04:02,527
It was custom built essentially for the
Grace Blackwell era to be able to max the

59
00:04:02,610 --> 00:04:05,997
the data center design
with the system design you had.

60
00:04:06,080 --> 00:04:08,267
And now, of course,
we're validating Vera Rubin.

61
00:04:08,350 --> 00:04:11,497
We're very excited about it. Maybe you
want to share a little bit about sort of

62
00:04:11,580 --> 00:04:15,337
what happens, even on the cloud
side, with how you're pushing

63
00:04:15,420 --> 00:04:20,147
on the systems innovation.
Well, our journey is incredible.

64
00:04:20,230 --> 00:04:24,917
We built the first AI supercomputer
together that was based on Ampere.

65
00:04:25,000 --> 00:04:27,317
Of course, Hopper
was an incredible success.

66
00:04:27,400 --> 00:04:30,487
The first two generations
were focused on pre-training.

67
00:04:30,570 --> 00:04:36,997
Grace Blackwell came along and all
of the focus moved to post training,

68
00:04:37,080 --> 00:04:41,167
reinforcement learning, which
allowed us to have reasoning models.

69
00:04:41,250 --> 00:04:44,297
And these reasoning models,
based on mixture of experts,

70
00:04:44,380 --> 00:04:50,577
were incredibly intelligent, energy
efficient, but it requires giant systems.

71
00:04:50,660 --> 00:04:56,447
And so we created NVLink72 and the
entire rack became one computer.

72
00:04:56,530 --> 00:04:59,887
We had evolved from one
node to now one rack.

73
00:04:59,970 --> 00:05:00,803
Well,

74
00:05:02,240 --> 00:05:07,287
Microsoft deployed the largest number
of Grace Blackwells in the world

75
00:05:07,370 --> 00:05:11,657
today, the fastest and the largest
number of Grace Blackwells in the world.

76
00:05:11,740 --> 00:05:15,927
Fairwater is just
a magnificent system to look at.

77
00:05:16,010 --> 00:05:18,197
It's just a miracle of engineering.

78
00:05:18,280 --> 00:05:22,167
It's just an incredible feat.
It's completely liquid cooled.

79
00:05:22,250 --> 00:05:25,037
You mentioned something earlier
that I'm very proud of as well,

80
00:05:25,120 --> 00:05:28,917
that it's closed looped,
basically uses almost no water,

81
00:05:29,000 --> 00:05:33,487
and it's incredibly environmentally
friendly. It's energy efficient.

82
00:05:33,570 --> 00:05:37,717
We're able to increase the token
generation rate and reduce

83
00:05:37,800 --> 00:05:43,697
the cost of token generation by an order
of magnitude, some 30 times over Hopper.

84
00:05:43,780 --> 00:05:45,597
So that was a huge achievement.

85
00:05:45,680 --> 00:05:52,467
Well, Vera Rubin was created for a world
where these AIs are now agentic.

86
00:05:52,550 --> 00:05:55,907
And so where as Hopper
was created for pretraining, Grace

87
00:05:55,990 --> 00:06:00,047
Blackwell for training,
post training and also inference.

88
00:06:00,130 --> 00:06:03,417
Vera Rubin is designed to run agents.

89
00:06:03,500 --> 00:06:06,917
It's agents. As you know,
this computing pattern

90
00:06:07,000 --> 00:06:11,687
is exactly the same computing pattern
we're going to run on the RTX Spark.

91
00:06:11,770 --> 00:06:15,727
It's exactly the same agentic system,
except of course, it's going to be much,

92
00:06:15,810 --> 00:06:20,427
much larger when process enormous
number of them simultaneously.

93
00:06:20,510 --> 00:06:23,167
Many of them are going to be
from different customers

94
00:06:23,250 --> 00:06:24,697
and different different partners.

95
00:06:24,780 --> 00:06:31,107
And so the entire path, the entire coding
path from storage, which is the long

96
00:06:31,190 --> 00:06:34,877
term memory, the working
memory is encrypted,

97
00:06:34,960 --> 00:06:38,047
the data is encrypted in transit.

98
00:06:38,130 --> 00:06:40,147
The data is also encrypted in use.

99
00:06:40,230 --> 00:06:42,717
And so we've we're
going to really innovate

100
00:06:42,800 --> 00:06:44,717
in the area of confidential computing.

101
00:06:44,800 --> 00:06:49,097
And so this entire disaggregated
distributed computing system

102
00:06:49,180 --> 00:06:53,997
you mentioned CPUs. Vera is a
revolutionary CPU designed for agents.

103
00:06:54,080 --> 00:06:56,437
You know, the past CPUs
were designed for humans.

104
00:06:56,520 --> 00:06:59,707
And, you know, we're just more
patient than than agents are.

105
00:06:59,790 --> 00:07:04,977
And agents want low latency,
just as if you've been working on as well.

106
00:07:05,060 --> 00:07:07,447
And Vera is designed
for extremely low latency.

107
00:07:07,530 --> 00:07:10,447
And so Vera Rubin is just
completely revolutionary.

108
00:07:10,530 --> 00:07:12,877
I can't wait to show it to everybody.

109
00:07:12,960 --> 00:07:14,687
You've already stood it up. Yep.

110
00:07:14,770 --> 00:07:18,017
Our two teams, our two teams
have been working very closely.

111
00:07:18,100 --> 00:07:23,757
And, you know, almost long before
the chips taped out, long before

112
00:07:23,840 --> 00:07:27,227
the systems were brought up, our two
teams were already completely aligned.

113
00:07:27,310 --> 00:07:31,067
And so the design, the data
centers were created for Vera Rubin.

114
00:07:31,150 --> 00:07:34,507
The Vera Rubin is designed
and integrated into your complete stack,

115
00:07:34,590 --> 00:07:36,577
into your networking into your security.

116
00:07:36,660 --> 00:07:40,507
And so the moment that our systems
were rolling off the lines,

117
00:07:40,590 --> 00:07:42,607
they were being stood up at Microsoft.

118
00:07:42,690 --> 00:07:45,247
So I'm incredibly excited
about the collaboration.

119
00:07:45,330 --> 00:07:49,317
Yeah, I know this speed of light execution
between the teams is fantastic to see.

120
00:07:49,400 --> 00:07:53,957
And of course, all this is to power
the ecosystem around us, right?

121
00:07:54,040 --> 00:07:58,697
I mean, you and I, having grown up
with the PC, the server and now with AI,

122
00:07:58,780 --> 00:08:02,627
have always thought about ultimately
it's about creating the opportunity

123
00:08:02,710 --> 00:08:06,467
for every developer, every organization
to build on the work that we do

124
00:08:06,550 --> 00:08:10,737
and the platforms we create. In speaking
of that, there's a lot of software

125
00:08:10,820 --> 00:08:14,147
that Nvidia builds.
That's all also coming.

126
00:08:14,230 --> 00:08:17,917
For example, we're going to have your
models in Foundry your tooling in Foundry.

127
00:08:18,000 --> 00:08:20,917
We're going to have, in fact,
your software even help us

128
00:08:21,000 --> 00:08:25,657
with accelerating our workloads when
it comes to even the data warehouse.

129
00:08:25,740 --> 00:08:28,627
We're going to obviously
have stuff in Windows.

130
00:08:28,710 --> 00:08:33,627
Just talk a little bit about that broader
vision of what does it mean for us,

131
00:08:33,710 --> 00:08:36,097
an opportunity, right.
Because everybody talks about this

132
00:08:36,180 --> 00:08:42,007
one model or one piece of tech, but it's
about the broadest, biggest opportunity

133
00:08:42,090 --> 00:08:47,107
for people to create value. Maybe you
want to share a little bit about that.

134
00:08:47,190 --> 00:08:49,207
Well, we've been
preparing for this moment.

135
00:08:49,290 --> 00:08:51,577
You know what happened
in the last several months.

136
00:08:51,660 --> 00:08:56,287
We've been working for a decade and a
half together, getting ready for really

137
00:08:56,370 --> 00:08:57,787
what happened in the last several months.

138
00:08:57,870 --> 00:09:01,827
All of a sudden, because of agentic
systems, the convergence of these

139
00:09:01,910 --> 00:09:05,257
really great models, AI is now useful.

140
00:09:05,340 --> 00:09:12,267
If you just look at GitHub,
the the commits into GitHub is gone

141
00:09:12,350 --> 00:09:17,237
completely parabolic. In the last
several months, the number of commits

142
00:09:17,320 --> 00:09:19,437
increased by a factor of three.

143
00:09:19,520 --> 00:09:22,617
It's clear that agentic
systems are useful,

144
00:09:22,700 --> 00:09:28,647
that it's doing productive work, and also
tokens are now profitable as a result.

145
00:09:28,730 --> 00:09:31,617
And so the amount of demand for compute

146
00:09:31,700 --> 00:09:37,627
between the usage of the AI and the
computation that's necessary for agents,

147
00:09:37,710 --> 00:09:40,127
the compute demand has
really gone through the roof.

148
00:09:40,210 --> 00:09:44,497
Well, one of the things that we've been
we've been doing together is making sure

149
00:09:44,580 --> 00:09:48,907
that all of the tools that the agents
are going to use are fully accelerated.

150
00:09:48,990 --> 00:09:51,937
Fabric, for example,
is now fully accelerated.

151
00:09:52,020 --> 00:09:57,477
We're accelerating data
processing, SQL, Spark,

152
00:09:57,560 --> 00:10:02,047
semantic based vector,
vector based, graph based.

153
00:10:02,130 --> 00:10:05,417
We're going to make sure that all
of the tools that are available on Azure

154
00:10:05,500 --> 00:10:08,587
are going to be fully GPU
accelerated, because the agents

155
00:10:08,670 --> 00:10:11,427
are going to be impatient.
The faster we can get the answers

156
00:10:11,510 --> 00:10:15,767
back to the agents, the faster they can
iterate, the faster can generate tokens,

157
00:10:15,850 --> 00:10:19,837
which are ultimately what the developers
both of our customers would like

158
00:10:19,920 --> 00:10:22,837
to do is generate a lot of tokens
that are really profitable,

159
00:10:22,920 --> 00:10:25,077
that are highly intelligent. Now.

160
00:10:25,160 --> 00:10:28,647
Thank you so much, Jensen,
for the partnership and the leadership

161
00:10:28,730 --> 00:10:33,717
and the innovation that you bring to
this entire ecosystem, and really thrilled

162
00:10:33,800 --> 00:10:38,187
to be working closely with you and the
team and bring all this to the developers

163
00:10:38,270 --> 00:10:42,927
here and beyond, and look forward
to seeing what the next few months

164
00:10:43,010 --> 00:10:45,627
and the next year bring
in terms of the innovation

165
00:10:45,710 --> 00:10:48,097
that gets built on top of the platform.

166
00:10:48,180 --> 00:10:52,237
So thank you again for joining
this late in the night from Taipei.

167
00:10:52,320 --> 00:10:55,837
Thank you so much Satya for your
partnership and friendship. Thank you.

168
00:10:55,920 --> 00:10:56,753
Thank you.