1
00:00:00,000 --> 00:00:03,477
Good morning. >> [applause]

2
00:00:03,560 --> 00:00:05,637
>> Morning.

3
00:00:05,720 --> 00:00:08,237
It's uh fantastic

4
00:00:08,320 --> 00:00:11,917
to be back uh in San Francisco at Build.

5
00:00:12,000 --> 00:00:16,917
It's always fun to be at developer
conferences at times of great change.

6
00:00:17,000 --> 00:00:20,797
You know, developer conferences are
always about understanding tech shifts,

7
00:00:20,880 --> 00:00:23,477
understanding the new stack, but it's

8
00:00:23,560 --> 00:00:26,677
also about really coming to grips with

9
00:00:26,760 --> 00:00:29,277
the new opportunity, right? For us as

10
00:00:29,360 --> 00:00:31,557
developers, for the companies that we

11
00:00:31,640 --> 00:00:34,717
work at, as well as the broader world.

12
00:00:34,800 --> 00:00:37,157
Uh so, today we're going to unpack this.

13
00:00:37,240 --> 00:00:38,957
This conference is all about that. And

14
00:00:39,040 --> 00:00:41,917
if there's one key takeaway,

15
00:00:42,000 --> 00:00:44,957
it would be this. How do you all

16
00:00:45,040 --> 00:00:48,637
participate fully in this frontier

17
00:00:48,720 --> 00:00:53,237
intelligence ecosystem? Right? It's not
about any one piece of technology that

18
00:00:53,320 --> 00:00:55,597
you'll hear about or even the platform

19
00:00:55,680 --> 00:00:59,317
itself. It's about the value that you

20
00:00:59,400 --> 00:01:04,237
can build, you can compound, you can
create on top of the platform. That's

21
00:01:04,320 --> 00:01:06,197
what a developer conference needs to be

22
00:01:06,280 --> 00:01:08,637
about, and that's what we will focus on.

23
00:01:08,720 --> 00:01:11,677
So, let's kind of get into it. Um let's

24
00:01:11,760 --> 00:01:14,677
take a look at this AI stack that this

25
00:01:14,760 --> 00:01:19,837
conference is going to unpack in a great
amount of detail. It starts, of course,

26
00:01:19,920 --> 00:01:21,477
with the compute fabric, right? That

27
00:01:21,560 --> 00:01:25,117
ubiquitous compute fabric that is spans

28
00:01:25,200 --> 00:01:29,957
the edge and the cloud.
Uh you then have this layer, which is

29
00:01:30,040 --> 00:01:31,677
the emerging layer where you have the

30
00:01:31,760 --> 00:01:34,637
models, you have context, you have the

31
00:01:34,720 --> 00:01:38,077
tools the models can access. Um and then

32
00:01:38,160 --> 00:01:40,237
on top of that, you have the runtime

33
00:01:40,320 --> 00:01:45,757
where you deploy your agents and
applications that you built uh on top of

34
00:01:45,840 --> 00:01:49,437
uh the context layer, on top of the
model layer. And then, of course, you

35
00:01:49,520 --> 00:01:53,837
have the best tooling uh to do all
of this. And then, you have the security,

36
00:01:53,920 --> 00:01:58,597
the compliance, uh and the governance,
right? That's the simplest form of the

37
00:01:58,680 --> 00:02:02,597
tech stack. But let's kind
of start where it always

38
00:02:02,680 --> 00:02:04,917
starts, which is infrastructure.

39
00:02:05,000 --> 00:02:07,357
And in fact, let's start at the edge

40
00:02:07,440 --> 00:02:09,157
with Windows.

41
00:02:09,240 --> 00:02:12,117
Because when you step back,

42
00:02:12,200 --> 00:02:15,437
the amount of compute there is at the

43
00:02:15,520 --> 00:02:17,477
edge is actually astounding, right? I

44
00:02:17,560 --> 00:02:21,677
mean, think about every NPU, GPU, CPU

45
00:02:21,760 --> 00:02:24,957
even. Um, you every PC, if you sort of

46
00:02:25,040 --> 00:02:27,237
aggregate that, that's a lot of compute

47
00:02:27,320 --> 00:02:30,677
power. So, we asked ourselves one simple

48
00:02:30,760 --> 00:02:32,117
question.

49
00:02:32,200 --> 00:02:34,597
Right? If we can deliver

50
00:02:34,680 --> 00:02:38,197
unmetered intelligence to every desk and

51
00:02:38,280 --> 00:02:42,077
every home, right? It takes us all
the way back to the very beginning, but

52
00:02:42,160 --> 00:02:46,637
that's what we said.
Can we do that in this era of AI?

53
00:02:46,720 --> 00:02:50,677
And in fact, we're in some sense
delivering that already today, right?

54
00:02:50,760 --> 00:02:52,317
When you look at something like, say,

55
00:02:52,400 --> 00:02:54,597
Outlook summarize, it's actually using

56
00:02:54,680 --> 00:02:57,957
onboard AI locally. Uh, same thing with

57
00:02:58,040 --> 00:03:01,077
PowerPoint all text, uh, or Teams super

58
00:03:01,160 --> 00:03:03,477
resolution. Uh, it's just not Microsoft

59
00:03:03,560 --> 00:03:05,237
stuff. It's a bit For example, Adobe

60
00:03:05,320 --> 00:03:08,357
After Effects or Premiere are both using

61
00:03:08,440 --> 00:03:11,717
Windows ML, um, across, uh, NPUs and

62
00:03:11,800 --> 00:03:15,917
GPUs for local processing. And so,
today one of the things that I'm

63
00:03:16,000 --> 00:03:17,997
really excited about in order to tap

64
00:03:18,080 --> 00:03:21,437
into all this compute power is to expand

65
00:03:21,520 --> 00:03:25,077
the scope of Windows ML and Windows AI,

66
00:03:25,160 --> 00:03:30,437
right? So, you now have the full install
base of GPUs, uh, that you can get to.

67
00:03:30,520 --> 00:03:34,157
So, I'm really thrilled that every
developer out there can count on

68
00:03:34,240 --> 00:03:39,437
building for local onboard AI, uh, and
then have it run across all of, uh, the

69
00:03:39,520 --> 00:03:43,197
install base. Now, if you

70
00:03:43,280 --> 00:03:45,757
We are also announcing two very cool new

71
00:03:45,840 --> 00:03:48,877
models, uh, that are all going to run on

72
00:03:48,960 --> 00:03:52,117
Windows, uh, inbox. Uh, the first is a

73
00:03:52,200 --> 00:03:55,957
new SLM. It's sort of a more efficient um

74
00:03:56,040 --> 00:04:00,997
model ion instruct um and it's
a it's a great uh reasoning

75
00:04:01,080 --> 00:04:05,757
model. And then we have the planning
model um ion plan, which is a local

76
00:04:05,840 --> 00:04:10,197
agentic loop. I mean, think about it,
right? You now have a full local agentic

77
00:04:10,280 --> 00:04:12,637
loop. You can give it tools access um

78
00:04:12,720 --> 00:04:15,837
and build a fully onboard agent agentic

79
00:04:15,920 --> 00:04:18,277
applications uh without having having to

80
00:04:18,360 --> 00:04:20,557
round trip to the cloud.

81
00:04:20,640 --> 00:04:23,157
Uh and of course, we said in order to

82
00:04:23,240 --> 00:04:26,677
really push the limits, realize this

83
00:04:26,760 --> 00:04:31,157
unmetered intelligence, there's also
a lot of hardware that's evolving, too.

84
00:04:31,240 --> 00:04:34,957
And so we're thrilled to see the
innovation across the Windows ecosystem,

85
00:04:35,040 --> 00:04:37,677
right? It's AMD with their Ryzen

86
00:04:37,760 --> 00:04:42,757
processors, the Panther Lake stuff from
Intel is just exciting to see. Qualcomm

87
00:04:42,840 --> 00:04:46,717
uh announced two sets of things,
which are great. One on the high end with

88
00:04:46,800 --> 00:04:49,077
Snapdragon X2 Elite.

89
00:04:49,160 --> 00:04:50,797
Uh and then on the lower end, uh even

90
00:04:50,880 --> 00:04:54,517
for sub 500 uh PCs with Snapdragon C.

91
00:04:54,600 --> 00:04:56,717
So, it's great to see all of this.

92
00:04:56,800 --> 00:04:58,917
And of course, that brings us

93
00:04:59,000 --> 00:05:02,517
uh to Nvidia uh and RTX Spark. Uh this

94
00:05:02,600 --> 00:05:05,197
is a next-generation SOC for PCs. It

95
00:05:05,280 --> 00:05:08,917
brings together the CPU, the GPU, as

96
00:05:09,000 --> 00:05:12,917
well as the AI capabilities into
a single SOC. And that's the exciting

97
00:05:13,000 --> 00:05:15,757
part. And it includes unified memory uh

98
00:05:15,840 --> 00:05:19,477
architecture and also the integrated uh

99
00:05:19,560 --> 00:05:24,597
DRTM, right? So, you now have all of
the systems innovation and the SOC coming

100
00:05:24,680 --> 00:05:27,997
together. Um and we're
really thrilled that one of

101
00:05:28,080 --> 00:05:31,597
the first devices that we built uh was

102
00:05:31,680 --> 00:05:33,957
the Surface Ultra, right? It's uh it's a

103
00:05:34,040 --> 00:05:37,037
beautiful uh device. And uh it really

104
00:05:37,120 --> 00:05:40,397
brings the power of Nvidia together with

105
00:05:40,480 --> 00:05:42,757
the design and the craftsmanship of

106
00:05:42,840 --> 00:05:46,557
Surface. It's 128 GB of uh unified

107
00:05:46,640 --> 00:05:50,837
memory. It's beautiful, too. Uh 2,000 and

108
00:05:50,920 --> 00:05:55,917
and and it has an all day
battery life. And so we're

109
00:05:56,000 --> 00:05:58,997
really excited to see
this later this fall.

110
00:05:59,080 --> 00:06:01,677
And it's also wonderful to see all the

111
00:06:01,760 --> 00:06:04,997
designs from all the OEM partners who

112
00:06:05,080 --> 00:06:09,677
are building taking advantage
of this new SOC and the new platform

113
00:06:09,760 --> 00:06:12,877
and bringing out some
pretty exciting machines

114
00:06:12,960 --> 00:06:16,797
that's going to be all
available come this fall.

115
00:06:16,880 --> 00:06:19,717
And so of course we said, "Okay, this is

116
00:06:19,800 --> 00:06:22,797
fantastic. What can we do next?" So we

117
00:06:22,880 --> 00:06:26,477
said, "Let's try and push this and push

118
00:06:26,560 --> 00:06:29,877
the architecture to its limit for

119
00:06:29,960 --> 00:06:34,197
developers. Right?
What if we could just max the compute,

120
00:06:34,280 --> 00:06:38,917
max the memory, build out
that developer machine that's

121
00:06:39,000 --> 00:06:40,477
the dream machine.

122
00:06:40,560 --> 00:06:42,997
And that's what we're announcing today.

123
00:06:43,080 --> 00:06:48,920
Surface RTX Spark Dev Box.
Let's roll the video.

124
00:06:50,690 --> 00:06:52,710
>> [music]

125
00:06:57,890 --> 00:06:59,910
[music]

126
00:07:11,650 --> 00:07:13,670
[music]

127
00:07:16,230 --> 00:07:18,250
[music]

128
00:07:21,480 --> 00:07:23,500
[music]

129
00:07:24,100 --> 00:07:26,120
[singing]

130
00:07:29,330 --> 00:07:31,350
[music]

131
00:07:38,500 --> 00:07:40,520
[music]

132
00:07:43,090 --> 00:07:48,380
[music] >> All right. >> [applause]

133
00:07:50,960 --> 00:07:53,117
>> Yeah, this is it's it's truly a dream

134
00:07:53,200 --> 00:07:56,437
machine. It's got a one petaflop of AI

135
00:07:56,520 --> 00:08:01,077
compute, 20 CPU cores.

136
00:08:01,160 --> 00:08:04,077
All of those things have 128

137
00:08:04,160 --> 00:08:07,717
GB of unified memory access. So super

138
00:08:07,800 --> 00:08:11,917
excited about this coming
in the fall and you can join the

139
00:08:12,000 --> 00:08:13,837
wait list. I'm on the wait list as well.

140
00:08:13,920 --> 00:08:15,197
So

141
00:08:15,280 --> 00:08:17,517
we'll get there.

142
00:08:17,600 --> 00:08:21,317
You know, so then we said why stop

143
00:08:21,400 --> 00:08:26,077
there? We said what if
we did one more thing?

144
00:08:26,160 --> 00:08:29,117
And in fact, I think
Jensen did that already

145
00:08:29,200 --> 00:08:31,437
when he said Windows is coming

146
00:08:31,520 --> 00:08:34,877
to the DGX station. So a new I describe

147
00:08:34,960 --> 00:08:37,717
it as the desktop data center, which you

148
00:08:37,800 --> 00:08:41,117
can have right there and running

149
00:08:41,200 --> 00:08:44,237
a one trillion parameter model locally.

150
00:08:44,320 --> 00:08:47,077
I mean, just to sort of put it
in perspective, it's pretty close to what

151
00:08:47,160 --> 00:08:51,797
perhaps we had when we built GPT-2 5 or 3,
right? One of the first one of the

152
00:08:51,880 --> 00:08:55,877
first computer supercomputer. So it's
pretty crazy to think that we've come

153
00:08:55,960 --> 00:09:00,837
this far where you can now have
a data center on your desktop.

154
00:09:00,920 --> 00:09:02,877
And of course, we're extending the

155
00:09:02,960 --> 00:09:04,837
developer endpoints even to the cloud.

156
00:09:04,920 --> 00:09:08,157
Windows 365 now has the developer

157
00:09:08,240 --> 00:09:11,077
distribution that's optimized for

158
00:09:11,160 --> 00:09:15,397
developer productivity in the cloud.
So that's the Windows 365 that I use every

159
00:09:15,480 --> 00:09:19,317
day. And it's simply about making Windows,

160
00:09:19,400 --> 00:09:24,677
whether it's on the laptop, the desktop,
the cloud, whatever, the best place to

161
00:09:24,760 --> 00:09:28,997
build. And so to that end,
we have tons and

162
00:09:29,080 --> 00:09:33,357
tons of updates, right? We're starting
by the way with one of the favorite

163
00:09:33,440 --> 00:09:37,037
things for all of us,
which is distraction-free

164
00:09:37,120 --> 00:09:41,677
dev environment. Uh we're also
introducing an intelligent terminal uh

165
00:09:41,760 --> 00:09:45,677
which has got built-in GitHub Copilot,
right? Terminal with sort of the Copilot

166
00:09:45,760 --> 00:09:48,597
intelligence. Um and of course, there's

167
00:09:48,680 --> 00:09:51,677
lots and lots of Linux love uh across uh

168
00:09:51,760 --> 00:09:53,837
Windows now. So, in

169
00:09:53,920 --> 00:09:57,237
you have 70 plus utilities. So, grep uh

170
00:09:57,320 --> 00:10:00,117
in full glory is now available for

171
00:10:00,200 --> 00:10:02,957
regular Windows access. Uh 70 plus

172
00:10:03,040 --> 00:10:05,437
utilities, as I said, uh all of that is

173
00:10:05,520 --> 00:10:07,797
coming to Windows. Um we're also

174
00:10:07,880 --> 00:10:10,877
bringing uh things from the Mac that you

175
00:10:10,960 --> 00:10:13,797
love. Uh things like uh Starship, Z

176
00:10:13,880 --> 00:10:16,597
shell, uh Homebrew is going to be native

177
00:10:16,680 --> 00:10:19,157
on Windows as well. So, you'll be able

178
00:10:19,240 --> 00:10:22,957
to uh switch to Windows. Um and today uh

179
00:10:23,040 --> 00:10:28,797
we're announcing uh WSL containers. Uh
this is, I know, a point of pain uh when

180
00:10:28,880 --> 00:10:32,077
you're managing all these uh
environments, uh switching, all the

181
00:10:32,160 --> 00:10:34,117
mental gymnastics of trying to keep this

182
00:10:34,200 --> 00:10:36,597
straight. Uh so, having first-class

183
00:10:36,680 --> 00:10:40,277
support uh for containers uh will really

184
00:10:40,360 --> 00:10:42,477
uh help us uh be in the flow when you

185
00:10:42,560 --> 00:10:45,317
are building and deploying locally. Um

186
00:10:45,400 --> 00:10:49,317
and to show you all of this on our new

187
00:10:49,400 --> 00:10:52,477
uh Surface RTX Spark, I wanted to uh

188
00:10:52,560 --> 00:10:55,197
invite up on stage um

189
00:10:55,280 --> 00:11:00,167
Kayla to just uh take a spin through
all the dev tools. Kayla, go ahead.

190
00:11:00,250 --> 00:11:02,270
>> [applause]

191
00:11:04,830 --> 00:11:08,077
[applause] >> Our team has
been working hard to make

192
00:11:08,160 --> 00:11:10,317
Windows a great place for development.

193
00:11:10,400 --> 00:11:13,437
And today, I'm going to show you some
major improvements and the lighters you

194
00:11:13,520 --> 00:11:16,997
can try out today. So,
what we have here is the default

195
00:11:17,080 --> 00:11:19,797
experience on the Surface RTX Spark dev

196
00:11:19,880 --> 00:11:22,717
box. Right away, it feels calm. There's

197
00:11:22,800 --> 00:11:27,277
no news feed, no widgets popping up,
no notifications, and of course, we're in

198
00:11:27,360 --> 00:11:31,077
dark mode. I'm immediately
ready to start development.

199
00:11:31,160 --> 00:11:34,797
Well, there's one thing I'd like
to change. I'd like to put my taskbar on

200
00:11:34,880 --> 00:11:38,477
the left. So, let me jump into
my taskbar settings using the new run,

201
00:11:38,560 --> 00:11:42,077
which is built leveraging
the architecture of PowerToys command

202
00:11:42,160 --> 00:11:46,117
palette. No, no, no.
I want it all the way on the

203
00:11:46,200 --> 00:11:50,757
left. There [snorts] we go.
After popular demand, we're excited to

204
00:11:50,840 --> 00:11:57,120
announce that vertical taskbar is now
available in Windows Insider builds.

205
00:11:57,280 --> 00:11:59,957
The new Surface RT Spark also has a

206
00:12:00,040 --> 00:12:04,437
bunch of key dev tools already installed,
like Python, Node, and many

207
00:12:04,520 --> 00:12:09,037
more of your favorites. All of the
developer goodness, all in one file.

208
00:12:09,120 --> 00:12:11,837
And if you want to get the same
experience today on your device, we're

209
00:12:11,920 --> 00:12:14,877
making this file available
to everyone right now.

210
00:12:14,960 --> 00:12:18,717
We have a public repo set up with the
configuration file and instructions for

211
00:12:18,800 --> 00:12:22,317
how to apply it using winget configure,
which will make the adjustments to

212
00:12:22,400 --> 00:12:25,677
Windows and install all the tools.

213
00:12:25,760 --> 00:12:29,277
Now, one cool thing that I have running
here is PowerToys' new utility called

214
00:12:29,360 --> 00:12:34,077
Grab and Move, which lets you hold Alt
and move the window around from anywhere.

215
00:12:34,160 --> 00:12:38,037
Another tip is that you can enable end
task, which lets you end the process

216
00:12:38,120 --> 00:12:41,077
without having to open task manager.

217
00:12:41,160 --> 00:12:46,277
So, let me jump into my dev drive. Dev
drives run on ReFS with Defender running

218
00:12:46,360 --> 00:12:49,597
in async and are optimized for performance
when it comes to development

219
00:12:49,680 --> 00:12:51,077
scenarios.

220
00:12:51,160 --> 00:12:54,477
Also, File Explorer is Git aware. We got

221
00:12:54,560 --> 00:12:59,637
stuff like last change author name, last
change message, the status of each file,

222
00:12:59,720 --> 00:13:03,517
plus my favorite is that the branch
name is on the bottom left.

223
00:13:03,600 --> 00:13:07,957
So, now let's get started
building and open our terminal.

224
00:13:08,040 --> 00:13:11,757
This is an experimental experience
called intelligent terminal that makes

225
00:13:11,840 --> 00:13:14,597
working with agents even more seamless.

226
00:13:14,680 --> 00:13:17,557
When you first install intelligent
terminal, you're greeted with the option

227
00:13:17,640 --> 00:13:21,637
to pick your favorite agent. I'm going
to use GitHub Copilot for today, but you

228
00:13:21,720 --> 00:13:24,517
can use whichever agent speaks to you.

229
00:13:24,600 --> 00:13:28,477
Now, here in an terminal, I have
a regular terminal pane at the top and an

230
00:13:28,560 --> 00:13:31,797
agent that's listening on the bottom and I
can work between them while the agent

231
00:13:31,880 --> 00:13:35,117
helps along the way. So,
for example, here's an error being

232
00:13:35,200 --> 00:13:39,437
generated. My agent pane is able
to detect it and provide a fix, which is

233
00:13:39,520 --> 00:13:41,357
great when I don't remember the syntax,

234
00:13:41,440 --> 00:13:44,437
especially for something like regex.

235
00:13:44,520 --> 00:13:49,677
So, I'm going to work on open claw and
I've already built it using WSL container.

236
00:13:49,760 --> 00:13:53,477
WSL container is a native container
experience on Windows, plus it can

237
00:13:53,560 --> 00:13:57,757
leverage the GPU, which is perfect
for the Surface RTX Spark.

238
00:13:57,840 --> 00:14:00,997
It can also reference your existing
container files, just like the one in

239
00:14:01,080 --> 00:14:04,317
the open claw project. So,
here's one of the files open in

240
00:14:04,400 --> 00:14:08,117
Microsoft edit, which ships in Windows
by default and just got syntax

241
00:14:08,200 --> 00:14:10,757
highlighting in its latest version.

242
00:14:10,840 --> 00:14:13,397
And then you can also just see
your containers running with a simple

243
00:14:13,480 --> 00:14:18,157
container images command.
And since we're on the topic of WSL,

244
00:14:18,240 --> 00:14:20,077
we're providing a WSL profile that's

245
00:14:20,160 --> 00:14:25,197
designed to feel comfortable for those
of you who use tools like Starship, ZSH,

246
00:14:25,280 --> 00:14:28,877
and Homebrew. So, it comes
pre-configured with all of

247
00:14:28,960 --> 00:14:32,317
your favorite utilities and it's
available in the repo that I showed

248
00:14:32,400 --> 00:14:34,437
earlier. It also includes one of my

249
00:14:34,520 --> 00:14:37,437
favorites, Btop.

250
00:14:37,520 --> 00:14:39,757
So, the Surface RTX Spark is designed

251
00:14:39,840 --> 00:14:44,957
for developer heavy workloads, including
serving large local models for coding.

252
00:14:45,040 --> 00:14:48,997
I've already done some development with
a 120 billion parameter model that most

253
00:14:49,080 --> 00:14:53,117
machines can't even load. So,
here's a quick view of my usage and

254
00:14:53,200 --> 00:14:57,037
we can see how many tokens I've used
locally. So, we're looking at about 3.4

255
00:14:57,120 --> 00:15:00,557
million tokens leveraged
on the device itself.

256
00:15:00,640 --> 00:15:04,837
Now, we can kick off multiple sub-agents
using fleet and just so we don't have to

257
00:15:04,920 --> 00:15:08,557
watch me type, I'm going to use
Copilot's voice feature, which is also

258
00:15:08,640 --> 00:15:12,117
leveraging its own local model. So,
I'll just hold spacebar and tell it what I

259
00:15:12,200 --> 00:15:14,920
want it to do.

260
00:15:15,040 --> 00:15:19,277
Find any console.writeline
or debug.writeline calls in the tray and

261
00:15:19,360 --> 00:15:21,597
node projects and convert them

262
00:15:21,680 --> 00:15:26,717
to the standard logger used
elsewhere in the code base.

263
00:15:26,800 --> 00:15:28,637
Here we go. Now the main agent will

264
00:15:28,720 --> 00:15:33,797
delegate sub agent tasks of appropriate
complexity to the local model utilizing

265
00:15:33,880 --> 00:15:36,837
my GPU and making it more cost-efficient.

266
00:15:36,920 --> 00:15:40,717
Now as developers while we're debugging
we're often looking through log files to

267
00:15:40,800 --> 00:15:46,397
diagnose any issues. Sometimes finding the
location of the log files is a challenge.

268
00:15:46,480 --> 00:15:51,317
I'd love to be able to just type
something like grep log and find them all.

269
00:15:51,400 --> 00:15:55,437
Oh how sweet. So on top
of already adding curl, tar,

270
00:15:55,520 --> 00:15:57,637
and pseudo to Windows now we're adding

271
00:15:57,720 --> 00:16:00,997
over 75 command line utilities like env,

272
00:16:01,080 --> 00:16:07,600
head, tail, and touch for those of us
who love to live in the terminal.

273
00:16:08,040 --> 00:16:11,837
So I found all my log files but now
having to parse them is the second

274
00:16:11,920 --> 00:16:14,117
challenge. Well I've actually had Ion

275
00:16:14,200 --> 00:16:17,157
instruct practically performing analyses

276
00:16:17,240 --> 00:16:20,157
on my log files this whole time.

277
00:16:20,240 --> 00:16:23,557
Now I can quickly diagnose anything
that's gone wrong in my development.

278
00:16:23,640 --> 00:16:27,597
Plus I don't have to worry about
token usage because it's all local.

279
00:16:27,680 --> 00:16:31,837
We can even take a look at our machine's
resources. The models are loaded and you

280
00:16:31,920 --> 00:16:34,277
can see 90 gigs of RAM being utilized by

281
00:16:34,360 --> 00:16:37,117
the GPU truly showcasing the full power

282
00:16:37,200 --> 00:16:39,557
of the Surface RTX Spark.

283
00:16:39,640 --> 00:16:44,437
We were able to use three local models
simultaneously unmetered while going

284
00:16:44,520 --> 00:16:50,440
about our our regular dev flow
without a hitch. That's huge.

285
00:16:51,990 --> 00:16:54,010
>> [applause]

286
00:16:56,720 --> 00:17:00,317
>> So I know you're going to love what the
team's been working on. We hope it gives

287
00:17:00,400 --> 00:17:01,877
you a glimpse of what's possible on

288
00:17:01,960 --> 00:17:08,280
Windows today and where we're headed next.
Thank you. Back to you Satya.

289
00:17:09,020 --> 00:17:11,040
>> [applause]

290
00:17:11,400 --> 00:17:13,717
>> Yeah that's the really the beginning of

291
00:17:13,800 --> 00:17:15,997
this idea of unmetered intelligence

292
00:17:16,080 --> 00:17:18,477
having those models and having the

293
00:17:18,560 --> 00:17:21,717
agents using the models work in parallel

294
00:17:21,800 --> 00:17:25,957
to what you may be doing along with the
cloud as well. And that I think is what

295
00:17:26,040 --> 00:17:28,517
the platform enables as first class.

296
00:17:28,600 --> 00:17:33,077
Now, let's move to the cloud.
You know, the driving equation for us

297
00:17:33,160 --> 00:17:38,117
remains the same, which is tokens per
dollar per watt. How do we optimize

298
00:17:38,200 --> 00:17:41,677
around this, right? So, when we think
about the systems problem, we think

299
00:17:41,760 --> 00:17:44,237
about electrons coming on one end and

300
00:17:44,320 --> 00:17:46,677
tokens on the other end. And how do we

301
00:17:46,760 --> 00:17:48,997
think about the systems optimization end

302
00:17:49,080 --> 00:17:51,357
to end? It starts with the data center

303
00:17:51,440 --> 00:17:53,757
design itself, right? The core compute,

304
00:17:53,840 --> 00:17:58,557
storage, network, all the accelerators
that go into accelerating each of those

305
00:17:58,640 --> 00:18:00,197
components.

306
00:18:00,280 --> 00:18:02,997
How do you even think about the DC to DC

307
00:18:03,080 --> 00:18:05,637
connectivity and the networking,

308
00:18:05,720 --> 00:18:10,397
as well as the offload to something
like the local compute, right? That's the

309
00:18:10,480 --> 00:18:14,597
sort of systems challenge.
But before we even get into

310
00:18:14,680 --> 00:18:18,797
all of the systems and the
technology and the innovation,

311
00:18:18,880 --> 00:18:20,357
perhaps the most important design

312
00:18:20,440 --> 00:18:23,597
criteria for us is how do we earn the

313
00:18:23,680 --> 00:18:26,117
permission from the communities in which

314
00:18:26,200 --> 00:18:28,437
we're building these data centers. And

315
00:18:28,520 --> 00:18:31,597
that's where these principles ground us

316
00:18:31,680 --> 00:18:33,397
and focus us.

317
00:18:33,480 --> 00:18:35,957
How do we ensure that the DCs do not

318
00:18:36,040 --> 00:18:39,397
increase the electricity prices,

319
00:18:39,480 --> 00:18:44,517
making sure that we are replenishing
all our water use, creating jobs in the

320
00:18:44,600 --> 00:18:47,517
local communities for the local residents,

321
00:18:47,600 --> 00:18:51,837
adding to the tax base, making sure
we're strengthening the communities by

322
00:18:51,920 --> 00:18:54,237
investing in local training and the

323
00:18:54,320 --> 00:18:57,517
nonprofits in the area. Only when we

324
00:18:57,600 --> 00:19:01,717
live up to these principles, do the hard
work around it, is when we earn the

325
00:19:01,800 --> 00:19:05,037
permission to go ahead
and innovate and build.

326
00:19:05,120 --> 00:19:07,797
And we've been doing a lot of data

327
00:19:07,880 --> 00:19:11,477
center build out. Today, Uh, Azure spans

328
00:19:11,560 --> 00:19:16,557
more than 500 data centers in 80 regions.
It's the most expansive, you

329
00:19:16,640 --> 00:19:17,877
know, we have the most expansive

330
00:19:17,960 --> 00:19:20,477
hyperscale footprint out there. Um, and

331
00:19:20,560 --> 00:19:23,637
we've added more data center capacity in

332
00:19:23,720 --> 00:19:27,437
the last 18 months than the first decade

333
00:19:27,520 --> 00:19:31,877
of Azure. Just to put that in perspective.
Uh, but more importantly,

334
00:19:31,960 --> 00:19:36,117
what we're building is also very
different. Uh, in fact, the first 15

335
00:19:36,200 --> 00:19:38,717
years or so, you know, we built out the

336
00:19:38,800 --> 00:19:43,237
commercial cloud infrastructure
for a set of heterogeneous,

337
00:19:43,320 --> 00:19:48,237
you know, workloads that spanned the
enterprise. Uh, but now when you look at

338
00:19:48,320 --> 00:19:51,317
what we're building, right? When you
think about all the gigawatts that were

339
00:19:51,400 --> 00:19:55,477
going to come online,
really they have three dominant

340
00:19:55,560 --> 00:20:00,197
workloads. There is training,
there's inference, and then there is the

341
00:20:00,280 --> 00:20:04,117
agent runtime. Right, these are three
dominant workloads. And in fact, when

342
00:20:04,200 --> 00:20:09,437
you look at Fairwater, uh,
it was sort of our first AI super

343
00:20:09,520 --> 00:20:13,117
factory. You know,
it spanned two regions, uh,

344
00:20:13,200 --> 00:20:17,877
Georgia and Wisconsin. Uh,
the entire system was designed from

345
00:20:17,960 --> 00:20:21,557
the ground up for AI. Right, we worked
in fact very closely with even Nvidia on

346
00:20:21,640 --> 00:20:24,437
this, and it's a two-story architecture

347
00:20:24,520 --> 00:20:27,397
that lets us essentially place racks,

348
00:20:27,480 --> 00:20:29,357
obviously in three dimensions, and pack

349
00:20:29,440 --> 00:20:32,837
the the maximum number of GPUs densely

350
00:20:32,920 --> 00:20:34,797
uh, with network access. Uh, and that

351
00:20:34,880 --> 00:20:37,277
means you really got fantastic higher

352
00:20:37,360 --> 00:20:41,957
performance networking, lower latency,
uh, and more effective bandwidth across

353
00:20:42,040 --> 00:20:46,397
the entire cluster. Um,
and we're rethinking even the power

354
00:20:46,480 --> 00:20:48,197
delivery, right? So, we sort of okay,

355
00:20:48,280 --> 00:20:51,197
how do we deliver hundreds of kilowatts

356
00:20:51,280 --> 00:20:54,157
per row while minimizing all the loss,

357
00:20:54,240 --> 00:20:58,277
right? Which is the conversion loss
that happens from the grid to the silicon.

358
00:20:58,360 --> 00:21:01,117
So, we basically sort of even, uh, took,

359
00:21:01,200 --> 00:21:03,437
uh, a new approach to it. And all of

360
00:21:03,520 --> 00:21:06,437
this also changes with, uh, the cooling

361
00:21:06,520 --> 00:21:10,837
system, right? And water. So,
in fact, the cooling loop is filled

362
00:21:10,920 --> 00:21:12,837
um once in the data center can operate

363
00:21:12,920 --> 00:21:15,397
effectively with zero water consumption.

364
00:21:15,480 --> 00:21:18,517
In fact, the daily water usage over the

365
00:21:18,600 --> 00:21:21,317
course of an entire year is roughly

366
00:21:21,400 --> 00:21:23,677
equivalent to what a single restaurant

367
00:21:23,760 --> 00:21:28,450
would use, right? I mean,
that's um >> [applause]

368
00:21:30,880 --> 00:21:33,157
>> And when it comes
to the systems and the

369
00:21:33,240 --> 00:21:35,477
silicon, uh again, we have a lot of

370
00:21:35,560 --> 00:21:38,037
choice. Uh we have first-party silicon.

371
00:21:38,120 --> 00:21:40,637
We have partner sys- uh systems.

372
00:21:40,720 --> 00:21:43,237
Uh we were the very first cloud to bring

373
00:21:43,320 --> 00:21:48,117
up, in fact, NVIDIA's uh Vera Rubin
system for validation. Very exciting to

374
00:21:48,200 --> 00:21:52,917
see that. Uh we're working closely with
AMD. We worked with them on MI300. Now,

375
00:21:53,000 --> 00:21:53,997
we're working with them on their

376
00:21:54,080 --> 00:21:57,957
next-generation uh AMD GPUs. Uh

377
00:21:58,040 --> 00:22:02,357
with Maya 200, uh it's continuing
to scale. In fact, it's live in Iowa and

378
00:22:02,440 --> 00:22:06,397
Arizona. Uh we'll deploy it
internationally later this year. It

379
00:22:06,480 --> 00:22:09,277
delivers 30% improved tokens per dollar

380
00:22:09,360 --> 00:22:11,357
compared to sort of what's the leading

381
00:22:11,440 --> 00:22:14,837
uh GPU today. Uh and we have validated

382
00:22:14,920 --> 00:22:17,637
it with 55 GPT-55, and we're going to

383
00:22:17,720 --> 00:22:20,277
use that to power uh Microsoft 365

384
00:22:20,360 --> 00:22:23,677
Copilot. Uh and so, when it comes to

385
00:22:23,760 --> 00:22:25,917
running these um

386
00:22:26,000 --> 00:22:28,797
agents, the interesting thing now is

387
00:22:28,880 --> 00:22:30,797
it's no longer just about having an AI

388
00:22:30,880 --> 00:22:34,997
accelerator or a GPU. The CPU is really

389
00:22:35,080 --> 00:22:36,757
critical. In fact, the ratios may be

390
00:22:36,840 --> 00:22:39,837
even coming to 1:1. Uh and that's why

391
00:22:39,920 --> 00:22:42,517
we're innovating uh with Cobalt. Uh

392
00:22:42,600 --> 00:22:45,157
we're announcing the preview of Cobalt

393
00:22:45,240 --> 00:22:48,517
200 VMs, our next-generation Arm-based

394
00:22:48,600 --> 00:22:51,717
CPU design for both cloud-native and

395
00:22:51,800 --> 00:22:57,117
agent uh workloads uh today. So, it's
exciting to see uh Cobalt make progress

396
00:22:57,200 --> 00:23:00,810
as well. >> [applause]

397
00:23:02,040 --> 00:23:05,757
>> You know, one thing that uh we've
been trying to make sure is that being

398
00:23:05,840 --> 00:23:08,997
optimized for these new workloads I
talked about, right? These new Asian

399
00:23:09,080 --> 00:23:13,037
workloads. So, in fact, Cobalt delivers
50 plus percent better performance than

400
00:23:13,120 --> 00:23:15,357
Cobalt 100 on cloud native. But, we

401
00:23:15,440 --> 00:23:20,157
started benchmarking them using the
GitHub Copilot traces, uh these agentic

402
00:23:20,240 --> 00:23:23,757
traces to see because the call patterns
are so different. Uh and we're now

403
00:23:23,840 --> 00:23:26,917
seeing 33% lower latency for the agent

404
00:23:27,000 --> 00:23:30,917
calls, uh 14% faster speed, 23% higher

405
00:23:31,000 --> 00:23:33,117
throughput. So, this is about the core

406
00:23:33,200 --> 00:23:36,197
design of both the AI accelerator and

407
00:23:36,280 --> 00:23:40,157
the CPU uh for the agent. And of course,

408
00:23:40,240 --> 00:23:44,517
um when you talk about AI workloads,
you need scale, you need reliability, and

409
00:23:44,600 --> 00:23:46,917
that's why the network becomes super

410
00:23:47,000 --> 00:23:49,717
critical. Uh we have innovated with the

411
00:23:49,800 --> 00:23:51,917
MRC architecture

412
00:23:52,000 --> 00:23:54,797
uh and rebuilt how traffic across Azure

413
00:23:54,880 --> 00:23:57,757
moves to support AI workloads, uh which

414
00:23:57,840 --> 00:24:02,157
are fundamentally synchronous data
parallel uh workloads, right? And so,

415
00:24:02,240 --> 00:24:06,237
therefore, they span tens
and thousands of GPUs.

416
00:24:06,320 --> 00:24:10,157
Um and so, you need to have them coherent.
That's what the next frontier

417
00:24:10,240 --> 00:24:13,237
of you know, is for us to make sure uh

418
00:24:13,320 --> 00:24:17,477
we're able to keep scaling uh
the network architecture. And of course,

419
00:24:17,560 --> 00:24:20,917
it's not about just inside the data
center, it's also about connecting

420
00:24:21,000 --> 00:24:23,677
across the data centers. Um and every

421
00:24:23,760 --> 00:24:26,517
Fairwater data center, for example, is

422
00:24:26,600 --> 00:24:29,317
connected through our continent-spanning

423
00:24:29,400 --> 00:24:32,637
uh AI WAN, forming this truly uh

424
00:24:32,720 --> 00:24:37,677
fungible uh compute fabric.
Um and you know,

425
00:24:37,760 --> 00:24:39,277
you know, all this innovation is

426
00:24:39,360 --> 00:24:44,957
exciting, but when you think about
innovation, designing for AI, and I'll

427
00:24:45,040 --> 00:24:49,677
say deep understanding of systems,

428
00:24:49,760 --> 00:24:51,757
um it you know, there's no better

429
00:24:51,840 --> 00:24:53,997
company than NVIDIA, and there's no

430
00:24:54,080 --> 00:24:57,237
better person uh than Jensen uh to talk

431
00:24:57,320 --> 00:24:59,837
about it. So, I wanted to invite live

432
00:24:59,920 --> 00:25:03,357
from Taipei, uh Jensen Huang, CEO and uh

433
00:25:03,440 --> 00:25:07,880
founder of Nvidia.
>> [applause and cheering]

434
00:25:11,040 --> 00:25:14,197
>> Thank you. >> Jensen,
thank you so much for being at

435
00:25:14,280 --> 00:25:19,317
Build again. I know it's late for you
in Taipei. I really appreciate you staying

436
00:25:19,400 --> 00:25:21,717
up. Uh you know,

437
00:25:21,800 --> 00:25:26,637
and then, you know, I'm I've been
looking at social and, you know, and

438
00:25:26,720 --> 00:25:29,757
everything people have been talking
about since your keynote over the

439
00:25:29,840 --> 00:25:33,237
weekend. Uh and suddenly, you know, this

440
00:25:33,320 --> 00:25:36,837
concept of unmetered intelligence right

441
00:25:36,920 --> 00:25:41,677
at the edge is so hot again. So, maybe
you want to talk a little bit. You've

442
00:25:41,760 --> 00:25:46,117
thought about this, talked about this,
and now, of course, with RTX Spark

443
00:25:46,200 --> 00:25:48,197
really delivered, I think, what's a

444
00:25:48,280 --> 00:25:51,717
breakthrough uh system for AI to be much

445
00:25:51,800 --> 00:25:53,317
more ubiquitous. But maybe, Jensen, you

446
00:25:53,400 --> 00:25:55,677
can just share a little bit your vision

447
00:25:55,760 --> 00:25:59,117
around where you see this going.

448
00:25:59,200 --> 00:26:02,877
>> Well, this all started about 3 years
ago between a conversation between you and

449
00:26:02,960 --> 00:26:05,717
I. And we were talking about how

450
00:26:05,800 --> 00:26:08,877
we could build a new class of PCs that's

451
00:26:08,960 --> 00:26:11,637
incredible for designers and creators.

452
00:26:11,720 --> 00:26:15,597
And uh it would be incredible for
artificial intelligence. And it would be

453
00:26:15,680 --> 00:26:17,917
one of these systems that uh

454
00:26:18,000 --> 00:26:22,677
has the processing capability, but also
the software stack that's integrated

455
00:26:22,760 --> 00:26:26,557
into the world's design packages and
creator packages. And And of course, all

456
00:26:26,640 --> 00:26:30,397
the things that we're doing
with AI. And here we are uh

457
00:26:30,480 --> 00:26:32,997
3 years later, we built an incredible

458
00:26:33,080 --> 00:26:34,237
new chip.

459
00:26:34,320 --> 00:26:38,677
And this system is supported by uh all

460
00:26:38,760 --> 00:26:40,237
of this new software that you create for

461
00:26:40,320 --> 00:26:43,117
Windows. And we now have the ability to

462
00:26:43,200 --> 00:26:48,477
have essentially an autonomous
agent running on the PC.

463
00:26:48,560 --> 00:26:51,877
Now, when you take a step back,
and you think about, what does that mean?

464
00:26:51,960 --> 00:26:56,237
For For the 40 years that or some 30
years we've been working together, we

465
00:26:56,320 --> 00:26:57,597
went from

466
00:26:57,680 --> 00:27:01,037
inventing DirectX together to

467
00:27:01,120 --> 00:27:04,037
creating now this

468
00:27:04,120 --> 00:27:06,557
incredible computer that has autonomous

469
00:27:06,640 --> 00:27:10,437
systems running. The PC evolved from

470
00:27:10,520 --> 00:27:13,677
being an incredible tool to now being a

471
00:27:13,760 --> 00:27:16,477
tool that's used autonomously by an AI

472
00:27:16,560 --> 00:27:19,517
assistant. And so, the idea that I could

473
00:27:19,600 --> 00:27:22,477
be traveling and I'm on the phone and I

474
00:27:22,560 --> 00:27:25,957
could text my PC and ask my PC to get

475
00:27:26,040 --> 00:27:28,437
some coding done or some idea that I

476
00:27:28,520 --> 00:27:33,077
have and it would fire
up the tools on the PC.

477
00:27:33,160 --> 00:27:37,237
And it would make the modifications
or the changes or the design that I was I

478
00:27:37,320 --> 00:27:39,917
told it to to do and it would iterate

479
00:27:40,000 --> 00:27:45,277
with me while I'm away from the PC.
My PC became an assistant. While I'm

480
00:27:45,360 --> 00:27:46,917
sitting there, of course, that this PC

481
00:27:47,000 --> 00:27:48,837
would be my great assistant as well. And

482
00:27:48,920 --> 00:27:52,597
so, this idea that that the PC evolved

483
00:27:52,680 --> 00:27:54,837
from a from a personal computer to a

484
00:27:54,920 --> 00:27:58,277
personal AI is just really exciting and

485
00:27:58,360 --> 00:28:03,397
to see it come to life, Satya, to see
it come to life and actually doing that.

486
00:28:03,480 --> 00:28:06,837
You know, so I'm I'm super
excited about it. I Spark

487
00:28:06,920 --> 00:28:10,757
you you mentioned earlier has all this
incredible capabilities, a petaflop of

488
00:28:10,840 --> 00:28:13,717
AI performance. It has a petaflops of

489
00:28:13,800 --> 00:28:17,997
NVFP4, this numerical format that our two

490
00:28:18,080 --> 00:28:21,517
companies worked
on together that allows us to

491
00:28:21,600 --> 00:28:24,517
take advantage of this 128 GB of memory

492
00:28:24,600 --> 00:28:29,477
and fit maybe a couple of hundred
billion parameter model. A couple of

493
00:28:29,560 --> 00:28:33,757
hundred billion parameter model
is state-of-the-art. And so, I think the

494
00:28:33,840 --> 00:28:37,997
days of having a really smart assistant
running autonomously on the PC is here.

495
00:28:38,080 --> 00:28:41,277
>> Yeah, I know it's so awesome and
in fact, I'm also excited about Windows

496
00:28:41,360 --> 00:28:45,517
coming to the GB300. And so,
that's another thing that it's

497
00:28:45,600 --> 00:28:49,397
kind of like data center right
on your desktop and it's so exciting. But

498
00:28:49,480 --> 00:28:53,397
talking about that um data
center side, obviously,

499
00:28:53,480 --> 00:28:56,157
you know, this entire thing got started

500
00:28:56,240 --> 00:28:58,997
when we built the first supercomputer uh

501
00:28:59,080 --> 00:29:02,037
together to train uh the GPT models and

502
00:29:02,120 --> 00:29:04,357
we've come a long way. In fact, even I

503
00:29:04,440 --> 00:29:06,357
was talking about the Fairwater design,

504
00:29:06,440 --> 00:29:09,277
it is custom-built essentially for uh

505
00:29:09,360 --> 00:29:11,597
the Grace Blackwell era uh to be able to

506
00:29:11,680 --> 00:29:16,917
max the the data center design with
the system design you had. And now, of

507
00:29:17,000 --> 00:29:20,117
course, we're validating Vera Rubin,
we're very excited about it. Maybe you

508
00:29:20,200 --> 00:29:24,517
want to share a little bit about sort
of what happens even on the cloud side uh

509
00:29:24,600 --> 00:29:28,517
with how you're pushing
on the systems innovation.

510
00:29:28,600 --> 00:29:30,757
>> Well, our journey has been incredible.

511
00:29:30,840 --> 00:29:35,477
We built the first AI supercomputer
together. That was based on Ampere.

512
00:29:35,560 --> 00:29:39,197
Of course, Hopper was an incredible
success. This first two generations were

513
00:29:39,280 --> 00:29:41,757
focused on pre-training. Grace Blackwell

514
00:29:41,840 --> 00:29:44,317
came along and

515
00:29:44,400 --> 00:29:47,637
all of the focus moved to post-training,

516
00:29:47,720 --> 00:29:52,197
uh reinforcement learning, which allowed
us to have reasoning models, and these

517
00:29:52,280 --> 00:29:56,477
reasoning models based on mixture
of experts were incredibly intelligent,

518
00:29:56,560 --> 00:29:58,077
energy-efficient,

519
00:29:58,160 --> 00:30:01,197
um but it requires uh giant systems. And

520
00:30:01,280 --> 00:30:04,677
so, we created NVLink 72, and the entire

521
00:30:04,760 --> 00:30:07,997
rack became one computer. We had evolved

522
00:30:08,080 --> 00:30:12,797
from one node to now one rack. Well,

523
00:30:12,880 --> 00:30:16,837
Microsoft deployed
the largest number of Grace

524
00:30:16,920 --> 00:30:18,717
Blackwells in the world today.

525
00:30:18,800 --> 00:30:22,277
The fastest and the largest number
of Grace Blackwells in the world.

526
00:30:22,360 --> 00:30:26,397
Fairwater is just
a magnificent system to look at.

527
00:30:26,480 --> 00:30:31,237
It's just a miracle of engineering.
It's It's just an incredible feat. Uh it's

528
00:30:31,320 --> 00:30:34,877
completely liquid-cooled. You mentioned
something earlier that I'm very proud of

529
00:30:34,960 --> 00:30:37,357
as well, that it's close-looped.

530
00:30:37,440 --> 00:30:39,957
Basically, uses almost no water, and

531
00:30:40,040 --> 00:30:42,477
it's incredibly environmentally uh

532
00:30:42,560 --> 00:30:45,037
it's energy efficient. We're able to

533
00:30:45,120 --> 00:30:47,717
increase the token generation rate and

534
00:30:47,800 --> 00:30:50,517
reduce the cost of token generation by

535
00:30:50,600 --> 00:30:53,397
an order of magnitude, some 30 times

536
00:30:53,480 --> 00:30:57,997
over Hopper. So, that was a huge
achievement. Well, Vera Rubin was

537
00:30:58,080 --> 00:31:01,077
created for a world where

538
00:31:01,160 --> 00:31:05,357
these AIs are now agentic.
And so, whereas Hopper was created for

539
00:31:05,440 --> 00:31:09,197
pre-training, Grace Blackwell
for training, post-training, and also

540
00:31:09,280 --> 00:31:10,557
inference,

541
00:31:10,640 --> 00:31:14,037
Vera Rubin is designed to run agents.

542
00:31:14,120 --> 00:31:16,717
It's agents, as you know, this computing

543
00:31:16,800 --> 00:31:18,837
pattern is exactly the same computing

544
00:31:18,920 --> 00:31:22,397
pattern when a run on the RTX Spark.

545
00:31:22,480 --> 00:31:26,197
It's exactly the same agentic system,
except of course, it's going to be much,

546
00:31:26,280 --> 00:31:31,477
much larger. We'll process enormous
number of them simultaneously. Many of

547
00:31:31,560 --> 00:31:34,677
them are going to be from different
customers and different different

548
00:31:34,760 --> 00:31:37,437
partners. And so, the entire path, the

549
00:31:37,520 --> 00:31:39,757
entire coding path from

550
00:31:39,840 --> 00:31:45,437
storage, which is the long-term memory,
the working memory, is encrypted.

551
00:31:45,520 --> 00:31:48,917
The data is encrypted in transit, the data

552
00:31:49,000 --> 00:31:51,637
is also encrypted in use. And so, we've

553
00:31:51,720 --> 00:31:55,997
we're going to really innovate in the
area of confidential computing. And so,

554
00:31:56,080 --> 00:31:58,397
this entire disaggregated, distributed

555
00:31:58,480 --> 00:32:00,797
computing system, you mentioned CPUs,

556
00:32:00,880 --> 00:32:05,677
Vera is a revolutionary CPU designed
for agents. You know, the past CPUs were

557
00:32:05,760 --> 00:32:08,917
designed for humans.
And you know, we're just more patient

558
00:32:09,000 --> 00:32:14,437
than than agents are. And agents uh
want low latency, just as have you

559
00:32:14,520 --> 00:32:18,077
been working on as well. Vera was
designed for extremely low latency. And

560
00:32:18,160 --> 00:32:22,477
so, Vera Rubin's is just completely
revolutionary. I can't wait uh to show

561
00:32:22,560 --> 00:32:24,637
it to everybody.
You've already stood it up.

562
00:32:24,720 --> 00:32:27,277
>> Yep. >> Our two teams
Our two teams have been

563
00:32:27,360 --> 00:32:28,997
working very closely. And you know,

564
00:32:29,080 --> 00:32:32,597
almost uh long before the chips taped

565
00:32:32,680 --> 00:32:36,597
out, long before the systems were
brought up, our two teams were already

566
00:32:36,680 --> 00:32:40,877
completely aligned. And so, the design
the data centers were created for Vera

567
00:32:40,960 --> 00:32:45,037
Rubin. Vera Rubin is designed
and integrated into your complete stack,

568
00:32:45,120 --> 00:32:49,117
into your networking, into your security.
And so, the moment that our

569
00:32:49,200 --> 00:32:53,317
systems were rolling off the lines,
they were being stood up at Microsoft. So,

570
00:32:53,400 --> 00:32:55,557
I'm incredibly excited
about the collaboration.

571
00:32:55,640 --> 00:32:58,997
>> Yeah, I know this speed of light
execution between the teams is fantastic

572
00:32:59,080 --> 00:33:01,477
to see. And of course, all this to is to

573
00:33:01,560 --> 00:33:06,197
power the ecosystem around us, right?
I mean, that you and I, having grown up

574
00:33:06,280 --> 00:33:10,957
with the PC, the server, and now with AI,
have always thought about ultimately

575
00:33:11,040 --> 00:33:15,077
it's about creating the opportunity
for every developer, every organization to

576
00:33:15,160 --> 00:33:18,637
build on the work that we do and the
platforms we create. And speaking of

577
00:33:18,720 --> 00:33:21,157
that, there's a lot of software uh that

578
00:33:21,240 --> 00:33:25,757
NVIDIA builds that's all also coming.
Uh for example, we're going to have your

579
00:33:25,840 --> 00:33:29,597
models in Foundry, your tooling
in Foundry. We're going to have, in fact,

580
00:33:29,680 --> 00:33:34,197
uh your software even help us with
accelerating our uh workloads when it

581
00:33:34,280 --> 00:33:37,917
comes to even the data warehouse.
We're going to obviously have uh stuff in

582
00:33:38,000 --> 00:33:40,157
Windows. Just talk a little bit about

583
00:33:40,240 --> 00:33:45,277
that broader vision of what does it mean
for us an opportunity, right? Because

584
00:33:45,360 --> 00:33:46,917
everybody talks about this one model or

585
00:33:47,000 --> 00:33:49,037
one piece of tech. But it's about the

586
00:33:49,120 --> 00:33:52,797
broadest, biggest opportunity for people

587
00:33:52,880 --> 00:33:57,757
to create value. Uh maybe you want
to share a little bit about that.

588
00:33:57,840 --> 00:34:00,597
>> Well, we've been preparing for this
moment. You know, what happened in the

589
00:34:00,680 --> 00:34:05,237
last several months, we've been working
for a decade and a half together.

590
00:34:05,320 --> 00:34:08,557
Getting ready for really what happened
in the last several months. All of a

591
00:34:08,640 --> 00:34:11,197
sudden, because of agentic systems, the

592
00:34:11,280 --> 00:34:12,757
convergence of these really great

593
00:34:12,840 --> 00:34:18,517
models, AI is now useful.
If you just look at GitHub,

594
00:34:18,600 --> 00:34:22,717
the the the the commits
into GitHub has gone

595
00:34:22,800 --> 00:34:24,957
completely parabolic. In the last

596
00:34:25,040 --> 00:34:27,197
several months, the the number of

597
00:34:27,280 --> 00:34:30,037
commits increased by factor of three.

598
00:34:30,120 --> 00:34:35,037
It's clear that agentic systems are
useful, that it's doing productive work,

599
00:34:35,120 --> 00:34:36,597
and also

600
00:34:36,680 --> 00:34:39,357
tokens are now profitable as a result.

601
00:34:39,440 --> 00:34:42,237
And so, the amount of demand for compute

602
00:34:42,320 --> 00:34:45,437
between the the the usage of the AI and

603
00:34:45,520 --> 00:34:49,557
the computation that's necessary
for agents, the compute demand has really

604
00:34:49,640 --> 00:34:53,517
gone through the roof. Well, one of
the things that we've been we been doing

605
00:34:53,600 --> 00:34:57,917
together is making sure that all of the
tools that the agents are going to use

606
00:34:58,000 --> 00:35:00,077
are fully accelerated. Fabric, for

607
00:35:00,160 --> 00:35:02,917
example, is now fully accelerated. We're

608
00:35:03,000 --> 00:35:06,917
accelerating data processing, um SQL, uh

609
00:35:07,000 --> 00:35:11,557
Spark, uh semantic-based, vector-based,

610
00:35:11,640 --> 00:35:15,117
uh graph-based. We're going to make sure
that all of the tools that are available

611
00:35:15,200 --> 00:35:19,237
on Azure are going to be fully GPU
accelerated because the agents are going

612
00:35:19,320 --> 00:35:23,237
to be impatient. The faster we can
get the answers back to the agents, the

613
00:35:23,320 --> 00:35:26,877
faster they can iterate, the faster
they can generate tokens, which are

614
00:35:26,960 --> 00:35:30,637
ultimately what the developers,
both of our customers, would like to do is

615
00:35:30,720 --> 00:35:34,797
generate a lot of tokens that are really
profitable, that are highly intelligent.

616
00:35:34,880 --> 00:35:39,077
>> Now, thank you so much, Jensen, for
the partnership and the leadership and the

617
00:35:39,160 --> 00:35:41,477
innovation uh that you bring to this

618
00:35:41,560 --> 00:35:46,597
entire ecosystem, and really thrilled to
be working closely with you and the team

619
00:35:46,680 --> 00:35:50,797
and bring all this to the developers
here and beyond, and look forward to

620
00:35:50,880 --> 00:35:55,357
seeing what the next sort of few months
and the next year bring in terms of the

621
00:35:55,440 --> 00:35:59,357
innovation that gets built on top
of the platform. So, thank you again for

622
00:35:59,440 --> 00:36:02,757
joining this late
in the night uh from Taipei.

623
00:36:02,840 --> 00:36:07,470
>> Thank you so much, Satya, for your
partnership and friendship. [applause]

624
00:36:11,520 --> 00:36:16,437
>> So, so far, we've talked
about the edge and the cloud.

625
00:36:16,520 --> 00:36:19,077
Um the current form
factors, right? I mean,

626
00:36:19,160 --> 00:36:23,237
when I saw that Jensen picture from the
weekend where he had all the desktops, I

627
00:36:23,320 --> 00:36:25,077
felt like, "Man, I'm back in the '90s,"

628
00:36:25,160 --> 00:36:28,037
right? Because it was so cool to see the

629
00:36:28,120 --> 00:36:31,357
lineup of all the machines that I loved

630
00:36:31,440 --> 00:36:34,157
and I grew up with back yet again with

631
00:36:34,240 --> 00:36:36,597
new functionality, right? The same form

632
00:36:36,680 --> 00:36:39,277
factor, but unbelievable new

633
00:36:39,360 --> 00:36:43,797
functionality because of the onboard
AI capability, right? So, that's sort of

634
00:36:43,880 --> 00:36:47,597
what we've seen with the laptop,
the desktop, and of

635
00:36:47,680 --> 00:36:52,317
course with the cloud. But,
it also, you know, sets up that

636
00:36:52,400 --> 00:36:57,237
next question. If you have
that capability, which is

637
00:36:57,320 --> 00:36:59,197
new function, and you can put it into

638
00:36:59,280 --> 00:37:01,397
existing form factors,

639
00:37:01,480 --> 00:37:04,557
can you even purpose-build

640
00:37:04,640 --> 00:37:06,837
new form factors

641
00:37:06,920 --> 00:37:08,677
for the new function? Can you build a

642
00:37:08,760 --> 00:37:11,597
new platform even for

643
00:37:11,680 --> 00:37:14,237
uh the agent era? And that is the

644
00:37:14,320 --> 00:37:17,597
motivation behind Project Solara, which

645
00:37:17,680 --> 00:37:19,717
we're introducing today. And to talk

646
00:37:19,800 --> 00:37:21,957
about this, I wanted to invite Stevie on

647
00:37:22,040 --> 00:37:25,037
stage, but first, let's roll the video.

648
00:37:25,120 --> 00:37:28,557
>> First, there's you.

649
00:37:28,640 --> 00:37:32,597
Then, everything in front of you.

650
00:37:32,680 --> 00:37:39,120
We get it. The noise,
the weight of it all.

651
00:37:39,880 --> 00:37:43,157
And so, we've [music]
been building something.

652
00:37:43,240 --> 00:37:45,557
A new chip-to-cloud platform for an

653
00:37:45,640 --> 00:37:48,277
agent-first world, and for new

654
00:37:48,360 --> 00:37:53,237
agent-first devices. You come close,

655
00:37:53,320 --> 00:37:57,037
it's ready when you are.
>> [music] >> You speak,

656
00:37:57,120 --> 00:38:01,277
it understands you.
What matters comes forward.

657
00:38:01,360 --> 00:38:05,437
Everything else [music] falls away.

658
00:38:05,520 --> 00:38:10,357
And when you want it to keep
going, it keeps going.

659
00:38:10,440 --> 00:38:12,317
Your call. >> [music]

660
00:38:12,400 --> 00:38:14,797
>> Sometimes it sits with you.

661
00:38:14,880 --> 00:38:19,277
Sometimes it goes with you.
Sometimes it [music] sees what you want

662
00:38:19,360 --> 00:38:21,437
it to see.

663
00:38:21,520 --> 00:38:26,317
It doesn't decide for you.
It just lights the path

664
00:38:26,400 --> 00:38:30,037
and tries to clear [music]
the way when you need it to.

665
00:38:30,120 --> 00:38:35,477
That's the idea.
A whole ecosystem built to clear the way

666
00:38:35,560 --> 00:38:38,877
so all that's left is [music]
where you need to go.

667
00:38:38,960 --> 00:38:42,577
Wherever you are, whenever it matters.

668
00:38:42,660 --> 00:38:44,680
>> [music]

669
00:38:47,520 --> 00:38:51,437
>> I am so excited to be here.

670
00:38:51,520 --> 00:38:54,517
It really is great
to be back on this stage.

671
00:38:54,600 --> 00:38:58,637
Now, before I talk about those
awesome new devices you just saw,

672
00:38:58,720 --> 00:39:01,277
let me start with the why.

673
00:39:01,360 --> 00:39:04,477
Back at Build 2023, I talked about the

674
00:39:04,560 --> 00:39:07,397
outside AI application structure.

675
00:39:07,480 --> 00:39:09,957
Where AI moved from operating within the

676
00:39:10,040 --> 00:39:12,957
application frame to operating globally,

677
00:39:13,040 --> 00:39:14,997
working across multiple apps and

678
00:39:15,080 --> 00:39:20,117
services to connect, coordinate,
and maintain context across entire

679
00:39:20,200 --> 00:39:21,437
workflows,

680
00:39:21,520 --> 00:39:25,077
devices, and time scales.

681
00:39:25,160 --> 00:39:27,397
What if there were an ecosystem of

682
00:39:27,480 --> 00:39:30,197
devices specifically designed for that

683
00:39:30,280 --> 00:39:32,437
new type of application structure, for

684
00:39:32,520 --> 00:39:35,877
those types of agents, for that

685
00:39:35,960 --> 00:39:38,957
transformational interaction technology?

686
00:39:39,040 --> 00:39:42,477
That is the impetus behind Project Solara.

687
00:39:42,560 --> 00:39:44,957
But with so many possible forms, which

688
00:39:45,040 --> 00:39:50,037
one do you pick? What is the next device?

689
00:39:50,120 --> 00:39:52,277
>> [snorts] >> You see,
the big aha for us is that it's

690
00:39:52,360 --> 00:39:56,637
not about choosing one specific form
factor. It is about creating a system

691
00:39:56,720 --> 00:39:58,197
that extends your agent across a

692
00:39:58,280 --> 00:40:01,517
constellation of devices.

693
00:40:01,600 --> 00:40:04,037
The next computer is not one device.

694
00:40:04,120 --> 00:40:06,997
It is all these devices working together

695
00:40:07,080 --> 00:40:09,837
as one system with agents showing up

696
00:40:09,920 --> 00:40:14,357
closer to where and when you need them.

697
00:40:14,440 --> 00:40:16,317
To really To realize this vision, two

698
00:40:16,400 --> 00:40:20,197
challenges immediately show up. First,

699
00:40:20,280 --> 00:40:22,117
many specialized form factors already

700
00:40:22,200 --> 00:40:25,437
exist, but often rely on custom one-off

701
00:40:25,520 --> 00:40:29,517
apps and fragmented stacks that are
difficult and expensive to build,

702
00:40:29,600 --> 00:40:31,797
deploy, and maintain.

703
00:40:31,880 --> 00:40:34,597
And second, across every industry,

704
00:40:34,680 --> 00:40:39,077
people and organizations already
building their own agents, deeply

705
00:40:39,160 --> 00:40:41,317
specialized and instrumented for their

706
00:40:41,400 --> 00:40:46,397
work. But the impact of those agents
is constrained by how and where they can

707
00:40:46,480 --> 00:40:50,117
exist. Project Volterra addresses both by

708
00:40:50,200 --> 00:40:52,117
giving organizations a way to extend

709
00:40:52,200 --> 00:40:55,237
their agents onto new, purpose-built,

710
00:40:55,320 --> 00:41:00,077
easy-to-manage form factors designed
to reach the nooks and crannies where

711
00:41:00,160 --> 00:41:03,877
conventional computers
either do not exist or

712
00:41:03,960 --> 00:41:06,677
are not optimal.

713
00:41:06,760 --> 00:41:11,437
It's a turnkey solution for building
unique agent-first devices enabled by

714
00:41:11,520 --> 00:41:13,517
three pillars.

715
00:41:13,600 --> 00:41:16,637
First, it's enterprise-ready, enabled by

716
00:41:16,720 --> 00:41:21,397
the AOSP-based Microsoft
device ecosystem platform.

717
00:41:21,480 --> 00:41:25,597
Second, it's it has
an agent-driven interaction

718
00:41:25,680 --> 00:41:28,357
model with just-in-time UI that adapts

719
00:41:28,440 --> 00:41:30,997
to the form factor. And third, it has

720
00:41:31,080 --> 00:41:33,557
extensibility, so you can bring your own

721
00:41:33,640 --> 00:41:35,837
agents. And tying it all together is

722
00:41:35,920 --> 00:41:40,917
Azure, unifying the system
across cloud and device. Okay.

723
00:41:41,000 --> 00:41:44,637
That's enough of that. Now,
let's talk about the devices.

724
00:41:44,720 --> 00:41:48,637
Today, we're previewing previewing
two very broad categories. The first is

725
00:41:48,720 --> 00:41:51,397
stationary, and the second is portable.

726
00:41:51,480 --> 00:41:55,477
The first is a device is designed
for your desk, and it's built on MediaTek

727
00:41:55,560 --> 00:41:58,957
silicon. With Hello
for Business, just walking up

728
00:41:59,040 --> 00:42:01,357
to the device securely signs you in,

729
00:42:01,440 --> 00:42:04,157
giving you direct access to your agents,

730
00:42:04,240 --> 00:42:08,557
just like Nathan's about to show you here.

731
00:42:08,640 --> 00:42:10,477
For the information worker, this means

732
00:42:10,560 --> 00:42:12,597
frictionless yet protected access to

733
00:42:12,680 --> 00:42:15,757
Microsoft 365 Copilot grounded in Work

734
00:42:15,840 --> 00:42:19,757
IQ. And with a simple glance,
it surfaces what matters next in your

735
00:42:19,840 --> 00:42:22,237
workday, helping you think, plan, and

736
00:42:22,320 --> 00:42:24,397
even act by delegating tasks to your

737
00:42:24,480 --> 00:42:27,997
agents with a simple tap or just using

738
00:42:28,080 --> 00:42:31,557
your voice. Think of it
as a dedicated secured

739
00:42:31,640 --> 00:42:35,837
ambient device for work.
It even supports experiences like

740
00:42:35,920 --> 00:42:37,917
handoff between devices, acting as a

741
00:42:38,000 --> 00:42:41,397
companion to your existing Windows PC.

742
00:42:41,480 --> 00:42:44,317
Or it can even let you access your cloud

743
00:42:44,400 --> 00:42:50,520
PC through Windows 365 and a connected
monitor. How cool is that?

744
00:42:50,900 --> 00:42:52,920
>> [applause]

745
00:42:54,560 --> 00:42:57,157
>> Now, the second device is portable.

746
00:42:57,240 --> 00:43:02,077
It's reimagining a wearable that
millions of people use every day,

747
00:43:02,160 --> 00:43:05,797
the access badge.
Built using Qualcomm silicon for

748
00:43:05,880 --> 00:43:10,037
wearable, this digital badge is a
lightweight form factor designed for

749
00:43:10,120 --> 00:43:12,797
agent interactions on the go.

750
00:43:12,880 --> 00:43:17,237
And then it's adaptable across
a variety of verticals and workflows.

751
00:43:17,320 --> 00:43:22,560
All right, I have here
an early prototype of the badge.

752
00:43:22,880 --> 00:43:25,037
And so, thanks. And using my

753
00:43:25,120 --> 00:43:27,917
fingerprint, I tap to unlock the device

754
00:43:28,000 --> 00:43:32,637
and I have access now to all my agents
in a secured manner. And would you look

755
00:43:32,720 --> 00:43:36,757
at that, I already have a task. And it
says, "Gather content for your social

756
00:43:36,840 --> 00:43:41,197
media post for today." So, why not just
do it right now, right? So, I'm going to

757
00:43:41,280 --> 00:43:44,637
hit record. And then now
the device's camera is

758
00:43:44,720 --> 00:43:47,957
recording. I'm going to pan across.
I hope you don't mind. I'm going to take

759
00:43:48,040 --> 00:43:50,917
your shots. Yes, thank you. Copilot,

760
00:43:51,000 --> 00:43:53,117
find some good shots from this, clean

761
00:43:53,200 --> 00:43:58,077
them up, and then send them to me so
for the for me and my team to review. All

762
00:43:58,160 --> 00:44:03,077
right. And then there you have it. Now
my agent's off running through multiple

763
00:44:03,160 --> 00:44:06,117
tasks to actually clean this up and send
them to me and the team. That's pretty

764
00:44:06,200 --> 00:44:10,870
cool. >> [applause]

765
00:44:11,880 --> 00:44:16,397
>> Okay, I know it's a simple demo,
but it's all agent driven and there are so

766
00:44:16,480 --> 00:44:20,277
many verticals, so many opportunities.
I mean, for example, imagine in health

767
00:44:20,360 --> 00:44:24,437
care, from the moment you pick up
the device, the right agent shapes the

768
00:44:24,520 --> 00:44:26,397
experience around the role and the

769
00:44:26,480 --> 00:44:29,077
workflow, helping with check-ins,

770
00:44:29,160 --> 00:44:31,837
patient records, and critical insights,

771
00:44:31,920 --> 00:44:36,557
all through enterprise-grade
secure access. And with the built-in

772
00:44:36,640 --> 00:44:38,237
microphones, the nurse can start a

773
00:44:38,320 --> 00:44:40,557
hands-free voice-based documentation,

774
00:44:40,640 --> 00:44:43,597
including diarization and annotation.

775
00:44:43,680 --> 00:44:45,757
And the side-facing camera can be used

776
00:44:45,840 --> 00:44:50,637
to verify and document patient vitals or
even scan in medications and help verify

777
00:44:50,720 --> 00:44:52,637
workflows. You know, these are just a

778
00:44:52,720 --> 00:44:57,717
few examples of how this small,
purpose-built wearable can can bring

779
00:44:57,800 --> 00:44:59,477
intelligence directly into the flow of

780
00:44:59,560 --> 00:45:02,317
patient care, helping nurses access and

781
00:45:02,400 --> 00:45:05,677
gather and even act on information while

782
00:45:05,760 --> 00:45:09,317
staying present with the patient.

783
00:45:09,400 --> 00:45:12,517
And while both the stationary
and portable concept devices represent a

784
00:45:12,600 --> 00:45:17,277
specific expression of agent first
devices, their core reference hardware

785
00:45:17,360 --> 00:45:20,437
and software are designed
to be highly flexible.

786
00:45:20,520 --> 00:45:24,197
I mean, just with a few changes, loading
a different agent, adjusting the shape,

787
00:45:24,280 --> 00:45:27,397
the screen size, the sensors,
or even input methods,

788
00:45:27,480 --> 00:45:30,597
the same foundation, the same software

789
00:45:30,680 --> 00:45:34,837
can be adapted for many verticals
and workflows, such as retail and

790
00:45:34,920 --> 00:45:38,517
industrial, hospitality, financial
services, legal, and so forth. I mean,

791
00:45:38,600 --> 00:45:42,797
that is the power of the platform,
that flexibility.

792
00:45:42,880 --> 00:45:46,437
So, whatever your scenario, there are
thousands of untapped opportunities to

793
00:45:46,520 --> 00:45:48,277
bring agent workflows into places where

794
00:45:48,360 --> 00:45:52,397
computing has not naturally fit before.

795
00:45:52,480 --> 00:45:54,317
And while this is an early look, we're

796
00:45:54,400 --> 00:45:56,717
really excited that AccuWeather, Best

797
00:45:56,800 --> 00:45:59,837
Buy, CVS Health, Levi's, Target, and

798
00:45:59,920 --> 00:46:02,957
others are working towards exploring how

799
00:46:03,040 --> 00:46:04,597
specialized agents and devices can

800
00:46:04,680 --> 00:46:08,197
improve their workflows.

801
00:46:08,280 --> 00:46:10,677
This is the broader opportunity for the

802
00:46:10,760 --> 00:46:13,757
ecosystem. Agents moving outside the app

803
00:46:13,840 --> 00:46:16,757
and taking shape in devices designed for

804
00:46:16,840 --> 00:46:18,677
a specific scenario, a specific

805
00:46:18,760 --> 00:46:22,717
customer, and a specific place.

806
00:46:22,800 --> 00:46:24,477
And for all of you,

807
00:46:24,560 --> 00:46:27,517
this is the moment to imagine where your

808
00:46:27,600 --> 00:46:30,117
agents should live, what form they

809
00:46:30,200 --> 00:46:36,360
should take, and what
new work they can unlock.

810
00:46:36,600 --> 00:46:40,677
Last week Satya sat down with Cristiano
Amon from Qualcomm to discuss this

811
00:46:40,760 --> 00:46:44,157
future. Let's roll the video.

812
00:46:44,240 --> 00:46:46,077
>> Cristiano, it's so so wonderful having

813
00:46:46,160 --> 00:46:49,117
you here at Device Lab and talk about

814
00:46:49,200 --> 00:46:51,557
these new reference designs.

815
00:46:51,640 --> 00:46:56,197
>> Very happy to be here. I have been
to this Device Lab a lot recently. You

816
00:46:56,280 --> 00:47:00,917
know, one of the things that you and I
have chatted for a while is how there's

817
00:47:01,000 --> 00:47:02,757
a real platform shift. We are moving

818
00:47:02,840 --> 00:47:06,237
from building operating systems, devices

819
00:47:06,320 --> 00:47:11,077
for apps to agents. You want to talk
a little bit about how you see this?

820
00:47:11,160 --> 00:47:12,917
>> Yeah, absolutely.
In every generation of

821
00:47:13,000 --> 00:47:14,717
technology transition we've seen a shift

822
00:47:14,800 --> 00:47:16,917
and I think this one seems to be a very

823
00:47:17,000 --> 00:47:20,317
big one. Agents really changes the whole

824
00:47:20,400 --> 00:47:23,597
nature of the device in itself. Starting

825
00:47:23,680 --> 00:47:28,597
with the fact that, you know,
if the AI understand the world the way we

826
00:47:28,680 --> 00:47:31,917
understand it, it's going to be closer
to our senses. It's going to be closer

827
00:47:32,000 --> 00:47:33,597
to our eyes, to our mouth, to our ears.

828
00:47:33,680 --> 00:47:36,837
It's going to be things
that we wear and it's changing

829
00:47:36,920 --> 00:47:41,397
the the nature of the computing. I think
you need a computing that is now geared

830
00:47:41,480 --> 00:47:45,077
towards real-time context
and from silicon to

831
00:47:45,160 --> 00:47:48,757
cloud. I mean, this is sort of one
of the reference designs we have which I

832
00:47:48,840 --> 00:47:53,437
love. There's stuff happening at the edge
here on this device. There's things

833
00:47:53,520 --> 00:47:55,197
on the cloud. You want to talk a a bit

834
00:47:55,280 --> 00:47:58,157
about the core systems and silicon

835
00:47:58,240 --> 00:48:01,517
implications for such a new ecosystem.

836
00:48:01,600 --> 00:48:06,197
>> You need a very power efficient CPU.
The whole silicon is designed for you to

837
00:48:06,280 --> 00:48:10,357
have a cloud native experience. And then
you have a lot of sensors. A lot of

838
00:48:10,440 --> 00:48:15,397
sensors for context. It's a much more
personalized and bespoke experience than

839
00:48:15,480 --> 00:48:17,197
an app in itself.

840
00:48:17,280 --> 00:48:19,317
And I think that's changing the nature

841
00:48:19,400 --> 00:48:22,517
of devices. And even the definition of a

842
00:48:22,600 --> 00:48:24,957
wearable platform is changing in itself.

843
00:48:25,040 --> 00:48:28,797
And you started to see those incredible
new form factors. That computing

844
00:48:28,880 --> 00:48:32,837
platform, one of the things that you
and I have talked a lot about is

845
00:48:32,920 --> 00:48:35,077
how do you build it such that there is

846
00:48:35,160 --> 00:48:40,517
an open ecosystem, right? Because it's
not about one agent. It's about any

847
00:48:40,600 --> 00:48:44,037
agent. >> This is what makes it very, very

848
00:48:44,120 --> 00:48:45,957
interesting and exciting. Because the

849
00:48:46,040 --> 00:48:48,717
smartphone today is at the center of

850
00:48:48,800 --> 00:48:51,357
your digital life.

851
00:48:51,440 --> 00:48:54,597
The job of those devices today is to be

852
00:48:54,680 --> 00:48:58,517
extending the functionality of the
smartphone. And because of that, you

853
00:48:58,600 --> 00:49:02,557
actually saw that many of those
platforms became vertical. It was a

854
00:49:02,640 --> 00:49:06,757
natural thing to have a vertical
platform from the same company because

855
00:49:06,840 --> 00:49:10,637
the phone was at the center.
That changes dramatically >> Right.

856
00:49:10,720 --> 00:49:13,037
>> when you think about agents.

857
00:49:13,120 --> 00:49:16,557
And agents becomes the center of your

858
00:49:16,640 --> 00:49:18,437
digital experience. And I think the

859
00:49:18,520 --> 00:49:20,997
industry is going to be looking for an

860
00:49:21,080 --> 00:49:23,877
open horizontal platform that actually

861
00:49:23,960 --> 00:49:27,557
enable the agents to be interacting with

862
00:49:27,640 --> 00:49:31,677
the best possible device for different
applications. And I think that's what's

863
00:49:31,760 --> 00:49:34,037
exciting. >> And we want
to make it possible for

864
00:49:34,120 --> 00:49:35,837
anybody who's thinking of an agentic

865
00:49:35,920 --> 00:49:38,917
system to not just be bound to current

866
00:49:39,000 --> 00:49:41,677
devices, but to imagine that there can

867
00:49:41,760 --> 00:49:46,357
be many, many devices that carry that
intelligence in different contexts. And

868
00:49:46,440 --> 00:49:50,277
so that's the open ecosystem we want to
build together. And again, it's so great

869
00:49:50,360 --> 00:49:52,757
to be partnered with you
to get this started.

870
00:49:52,840 --> 00:49:54,477
>> That's awesome.
We're very proud of this

871
00:49:54,560 --> 00:49:58,200
partnership and it's just the beginning.

872
00:49:58,620 --> 00:50:00,640
>> [applause]

873
00:50:02,400 --> 00:50:05,917
>> Thank you very much Stevie
and Cristiano. We are very excited about

874
00:50:06,000 --> 00:50:09,557
Project Solara.
It's sort of a new platform

875
00:50:09,640 --> 00:50:11,677
but perhaps more importantly it's a set

876
00:50:11,760 --> 00:50:15,437
of new platform rules that don't you

877
00:50:15,520 --> 00:50:18,237
know, in some sense hem in what you can

878
00:50:18,320 --> 00:50:19,957
imagine

879
00:50:20,040 --> 00:50:22,517
the type of form factors, where your

880
00:50:22,600 --> 00:50:25,397
agents live. I think always whenever

881
00:50:25,480 --> 00:50:27,717
these new platforms have come,

882
00:50:27,800 --> 00:50:30,837
you get to rewrite even the rules of how

883
00:50:30,920 --> 00:50:33,637
new platforms operate and that's what

884
00:50:33,720 --> 00:50:35,117
we're trying to get done with Project

885
00:50:35,200 --> 00:50:37,997
Solara so that you as developers

886
00:50:38,080 --> 00:50:41,757
and and enterprises have the flexibility

887
00:50:41,840 --> 00:50:44,237
to imagine the form factors that you

888
00:50:44,320 --> 00:50:47,397
want and have your agents be ubiquitous.

889
00:50:47,480 --> 00:50:51,597
Now let's go up the stack
to the next layer.

890
00:50:51,680 --> 00:50:55,437
>> [laughter] >> We are
building a new intelligence layer

891
00:50:55,520 --> 00:50:58,357
bringing together the models, contexts

892
00:50:58,440 --> 00:51:03,517
as well as the tools.
It of course starts with model choice.

893
00:51:03,600 --> 00:51:07,757
Every customer, every
developer is going to choose the

894
00:51:07,840 --> 00:51:10,157
right model for the right task and their

895
00:51:10,240 --> 00:51:13,757
eval mix, latency budget, even cogs

896
00:51:13,840 --> 00:51:15,277
budgets

897
00:51:15,360 --> 00:51:18,997
and Foundry today has over 11,000

898
00:51:19,080 --> 00:51:22,677
models. It's the largest model catalog

899
00:51:22,760 --> 00:51:26,637
out there from Open AI
to Anthropic and of course

900
00:51:26,720 --> 00:51:29,117
even MAI models and just

901
00:51:29,200 --> 00:51:31,437
just last week we brought the Open AI

902
00:51:31,520 --> 00:51:34,437
real-time voice models along with Claude

903
00:51:34,520 --> 00:51:37,877
Opus 4.8 to Foundry. So we're continuing

904
00:51:37,960 --> 00:51:40,317
to bring all these frontier models.

905
00:51:40,400 --> 00:51:43,637
The the consideration
though about the models

906
00:51:43,720 --> 00:51:46,277
is becoming now increasingly key. Key in

907
00:51:46,360 --> 00:51:47,997
fact from the last developer conference

908
00:51:48,080 --> 00:51:50,677
to now, uh when you're building any

909
00:51:50,760 --> 00:51:53,957
agentic system, having this context,

910
00:51:54,040 --> 00:51:57,157
uh really shaped right, is becoming

911
00:51:57,240 --> 00:52:01,117
super important. And in fact, it starts
right at the data tier, right? The data

912
00:52:01,200 --> 00:52:03,917
estate today has been built for

913
00:52:04,000 --> 00:52:08,757
applications that supported these
user-facing applications, right? Now you

914
00:52:08,840 --> 00:52:10,637
have to change and build them for

915
00:52:10,720 --> 00:52:13,277
agents, which again have very different

916
00:52:13,360 --> 00:52:16,157
call patterns even to the data tier. Um

917
00:52:16,240 --> 00:52:18,357
agents are continuously

918
00:52:18,440 --> 00:52:21,917
storing, retrieving, reasoning,

919
00:52:22,000 --> 00:52:26,037
uh acting, and learning, right? That's
sort of what's happening in a continuous

920
00:52:26,120 --> 00:52:28,317
loop. Uh and you see that today. Uh

921
00:52:28,400 --> 00:52:30,917
agents are using things like Cosmos DB

922
00:52:31,000 --> 00:52:33,317
for their memory. In fact, ChatGPT does.

923
00:52:33,400 --> 00:52:36,237
Azure Search, um you you know, is used

924
00:52:36,320 --> 00:52:39,397
for retrieval of indices and embeddings.

925
00:52:39,480 --> 00:52:43,917
Uh Fabric IQ for the semantic models
and the ontologies. It's part of the

926
00:52:44,000 --> 00:52:47,277
business logic tier
effectively for agents.

927
00:52:47,360 --> 00:52:50,997
Um and also, uh you have Fabric
Real-Time Intelligence for all the

928
00:52:51,080 --> 00:52:54,117
observability traces. Uh which now

929
00:52:54,200 --> 00:52:57,157
brings me uh to a very exciting new

930
00:52:57,240 --> 00:53:00,757
service, Horizon DB, which is our fully

931
00:53:00,840 --> 00:53:03,397
managed PostgreSQL service on Azure.

932
00:53:03,480 --> 00:53:06,880
Really thrilled to have this.

933
00:53:07,920 --> 00:53:10,397
>> [applause] >> I mean,
one of the things we wanted to

934
00:53:10,480 --> 00:53:12,917
make sure is we built ground up a

935
00:53:13,000 --> 00:53:15,157
PostgreSQL managed service which was for

936
00:53:15,240 --> 00:53:18,677
high availability, scale out. Uh it's

937
00:53:18,760 --> 00:53:22,117
zone redundant with automated failover.

938
00:53:22,200 --> 00:53:25,997
Uh 128 terabytes of storage per cluster.

939
00:53:26,080 --> 00:53:28,157
Uh 15 read replicas. I mean, the

940
00:53:28,240 --> 00:53:31,317
read-heavy workloads, uh you can scale

941
00:53:31,400 --> 00:53:35,637
uh with this managed service.
And in our internal testing, uh we're

942
00:53:35,720 --> 00:53:37,557
seeing something like, you know, the

943
00:53:37,640 --> 00:53:40,917
Horizon is delivering 3x throughput uh

944
00:53:41,000 --> 00:53:43,357
compared to any self-managed PostgreSQL

945
00:53:43,440 --> 00:53:47,957
setup. So, it's super critical when you
think about the uh the scale you need.

946
00:53:48,040 --> 00:53:51,317
Uh the other uh data
workload uh we're also

947
00:53:51,400 --> 00:53:53,197
changing pretty dramatically is the data

948
00:53:53,280 --> 00:53:56,717
warehouse. Uh in a world where agents

949
00:53:56,800 --> 00:53:59,757
are constantly querying data, the data

950
00:53:59,840 --> 00:54:03,757
warehouse effectively becomes a pretty
mission-critical, right? I mean, it's

951
00:54:03,840 --> 00:54:05,837
one thing to have it be mission-critical

952
00:54:05,920 --> 00:54:08,877
for users, uh but when agents need uh

953
00:54:08,960 --> 00:54:11,477
the analysis done on the fly, uh you

954
00:54:11,560 --> 00:54:13,917
know, bringing GPU acceleration to

955
00:54:14,000 --> 00:54:17,237
Fabric uh is super key, and we're seeing

956
00:54:17,320 --> 00:54:21,477
7x uh performance gain.
So, it's really thrilling

957
00:54:21,560 --> 00:54:23,817
uh to see that AI acceleration.

958
00:54:23,900 --> 00:54:25,920
>> [applause]

959
00:54:26,840 --> 00:54:29,397
>> So, if you sort of have the data tier,

960
00:54:29,480 --> 00:54:31,797
the the layer above this

961
00:54:31,880 --> 00:54:35,957
is the IQ layer that we're building that

962
00:54:36,040 --> 00:54:41,117
brings together essentially
the model capabilities to

963
00:54:41,200 --> 00:54:44,597
along with the data, right? So,
you kind of mix the data and the model

964
00:54:44,680 --> 00:54:47,237
capabilities so that you can deliver

965
00:54:47,320 --> 00:54:51,797
that right context to unlock
intelligence, right? That's where In

966
00:54:51,880 --> 00:54:56,837
fact, if you want if when people talk
about token efficiency, this is perhaps

967
00:54:56,920 --> 00:54:58,557
the most important consideration, right?

968
00:54:58,640 --> 00:55:01,957
If you structure the context right uh

969
00:55:02,040 --> 00:55:04,677
and feed the models, then you will by

970
00:55:04,760 --> 00:55:06,317
definition going to be so much more

971
00:55:06,400 --> 00:55:10,077
token efficient. Uh the first domain uh

972
00:55:10,160 --> 00:55:12,997
is the web. Uh web grounding is so

973
00:55:13,080 --> 00:55:16,557
important. Um you need that fresh, uh

974
00:55:16,640 --> 00:55:21,037
high-quality, and fast um web data, uh

975
00:55:21,120 --> 00:55:23,157
and that's why we're really, really

976
00:55:23,240 --> 00:55:28,840
excited uh to announce today WebIQ.

977
00:55:29,400 --> 00:55:33,957
>> [applause] >> And WebIQ
is built on our global

978
00:55:34,040 --> 00:55:36,197
infrastructure that's already serving

979
00:55:36,280 --> 00:55:39,637
over a billion users, but fundamentally

980
00:55:39,720 --> 00:55:41,277
re-architected

981
00:55:41,360 --> 00:55:43,477
for the LLM and the agentic workflow.

982
00:55:43,560 --> 00:55:47,437
It's model agnostic, it's MCP native,

983
00:55:47,520 --> 00:55:50,757
plugs right into any agent runtime. It

984
00:55:50,840 --> 00:55:54,717
has web, it has news, images, video. So

985
00:55:54,800 --> 00:55:58,477
agents can ground responses in fresh,

986
00:55:58,560 --> 00:56:00,837
verifiable content.

987
00:56:00,920 --> 00:56:04,157
And Web IQ leads across all of the three

988
00:56:04,240 --> 00:56:06,517
key criteria, right? It's best in class

989
00:56:06,600 --> 00:56:09,397
in quality, it's best in class in speed,

990
00:56:09,480 --> 00:56:13,597
as well as in cost. So we're very,
very thrilled about Web IQ being in the

991
00:56:13,680 --> 00:56:17,477
developers' hands as you build out
agentic systems. Of course, we're not

992
00:56:17,560 --> 00:56:19,997
stopping there. Beyond the web, every

993
00:56:20,080 --> 00:56:21,957
developer wants to be able to ground

994
00:56:22,040 --> 00:56:25,077
their agents on what's the most valuable

995
00:56:25,160 --> 00:56:28,117
data across the enterprise. And so we

996
00:56:28,200 --> 00:56:31,877
are bringing together Foundry, Fabric,

997
00:56:31,960 --> 00:56:35,637
and Microsoft 365 as this unified IQ

998
00:56:35,720 --> 00:56:37,997
layer, right? It's a continuously

999
00:56:38,080 --> 00:56:40,837
updated understanding

1000
00:56:40,920 --> 00:56:43,957
of your organization to show you all of

1001
00:56:44,040 --> 00:56:47,237
this rich IQ in action, building

1002
00:56:47,320 --> 00:56:49,717
real-life agents. Here is Elijah.

1003
00:56:49,800 --> 00:56:54,610
Elijah, take it away. >> [applause]

1004
00:56:55,520 --> 00:56:58,637
>> Agents are only as good as the context

1005
00:56:58,720 --> 00:57:02,077
we give them. Microsoft IQ unifies

1006
00:57:02,160 --> 00:57:06,237
enterprise intelligence
for every organization.

1007
00:57:06,320 --> 00:57:08,357
I'm here at a power utilities control

1008
00:57:08,440 --> 00:57:10,757
center and we'll start by running a

1009
00:57:10,840 --> 00:57:14,197
long-running agent.

1010
00:57:14,280 --> 00:57:16,357
This agent is going to help us assess

1011
00:57:16,440 --> 00:57:21,757
the current grid operations incident and
produce a brief for us so we can respond

1012
00:57:21,840 --> 00:57:24,437
accordingly. Now, I'm going
to go ahead and kick that

1013
00:57:24,520 --> 00:57:28,557
off, and while that runs, let me show
you how we got here with context from

1014
00:57:28,640 --> 00:57:30,757
Microsoft IQ.

1015
00:57:30,840 --> 00:57:33,637
We built our agent in Microsoft Foundry.

1016
00:57:33,720 --> 00:57:36,037
It's connected to various tools, and

1017
00:57:36,120 --> 00:57:38,317
it's also wired to a Foundry IQ

1018
00:57:38,400 --> 00:57:41,277
knowledge base, a single grounded source

1019
00:57:41,360 --> 00:57:46,517
that packages our documents, operational
data, and people into context the agent

1020
00:57:46,600 --> 00:57:50,357
can reason over.
After building the agent in Foundry, we

1021
00:57:50,440 --> 00:57:53,317
published it to Microsoft 365 for the

1022
00:57:53,400 --> 00:57:57,557
whole team. Let's go see it in action.

1023
00:57:57,640 --> 00:58:01,557
Now, here in M365, I will start with
a question about current events that an

1024
00:58:01,640 --> 00:58:03,477
ungrounded LLM wouldn't be able to

1025
00:58:03,560 --> 00:58:05,997
answer. I'm going to ask about current

1026
00:58:06,080 --> 00:58:08,917
electricity prices in SF.

1027
00:58:09,000 --> 00:58:11,157
For this, our agent pulls in the first

1028
00:58:11,240 --> 00:58:14,397
IQ in our toolbox, Web IQ, search built

1029
00:58:14,480 --> 00:58:16,077
for the AI era.

1030
00:58:16,160 --> 00:58:18,797
Web IQ delivers industry-leading

1031
00:58:18,880 --> 00:58:22,317
quality, velocity, and efficiency. Web

1032
00:58:22,400 --> 00:58:25,317
IQ constantly indexes fresh, official

1033
00:58:25,400 --> 00:58:29,677
sources from across the web,
and additionally, Web IQ does a great job

1034
00:58:29,760 --> 00:58:32,157
with semantic documents. And here we can

1035
00:58:32,240 --> 00:58:37,037
see Web IQ gave us our answer
grounded in reality.

1036
00:58:37,120 --> 00:58:41,637
Now, not only can we deploy our agent
here in M365, but we can also embed it

1037
00:58:41,720 --> 00:58:45,517
in our custom apps.
Let's head back to the control center.

1038
00:58:45,600 --> 00:58:49,917
Let's see how we handled a previous
incident. So, after using Web IQ to

1039
00:58:50,000 --> 00:58:54,797
gather external info, we asked
for details about our potentially at-risk

1040
00:58:54,880 --> 00:58:59,037
substations. The real
power of Microsoft IQ comes

1041
00:58:59,120 --> 00:59:01,317
when we combine that external knowledge

1042
00:59:01,400 --> 00:59:03,237
with our own internal enterprise

1043
00:59:03,320 --> 00:59:06,157
context. For this information, our agent

1044
00:59:06,240 --> 00:59:08,677
pulls in the next layer of Microsoft IQ,

1045
00:59:08,760 --> 00:59:11,157
Fabric IQ.

1046
00:59:11,240 --> 00:59:13,757
Here is Brightlines grid represented as

1047
00:59:13,840 --> 00:59:17,077
a fabric ontology, an operational model

1048
00:59:17,160 --> 00:59:19,677
of the live grid. And critically, we

1049
00:59:19,760 --> 00:59:21,837
didn't build this from scratch.

1050
00:59:21,920 --> 00:59:24,197
Fabric takes the Power BI semantic

1051
00:59:24,280 --> 00:59:26,357
models used by millions of customers

1052
00:59:26,440 --> 00:59:31,237
today and let's teams extend them
into rich ontologies that help run the

1053
00:59:31,320 --> 00:59:35,157
business. This model Yes.
This model is coupled

1054
00:59:35,240 --> 00:59:39,877
with live telemetry, so it reflects
the real operational state of the grid

1055
00:59:39,960 --> 00:59:41,877
minute by minute.

1056
00:59:41,960 --> 00:59:47,400
This is what Microsoft IQ means
by enterprise intelligence.

1057
00:59:48,130 --> 00:59:52,237
>> [applause] >> Not data
scattered across disparate

1058
00:59:52,320 --> 00:59:56,597
systems, but a single living model
of the business that an agent can reason

1059
00:59:56,680 --> 01:00:00,157
over. Now, turning back to our agent,

1060
01:00:00,240 --> 01:00:02,797
we can see that it gave us a table of

1061
01:00:02,880 --> 01:00:07,437
our most exposed substations. That's
the agent querying the ontology you just

1062
01:00:07,520 --> 01:00:10,197
saw. Now, the agent has access to both the

1063
01:00:10,280 --> 01:00:12,717
outside world and our grid. The last

1064
01:00:12,800 --> 01:00:18,037
piece is the one that turns a situation
into a response, our policies and our

1065
01:00:18,120 --> 01:00:21,317
people. By asking, "What are the steps to

1066
01:00:21,400 --> 01:00:26,197
respond to a substation trip?" we
activate the final layer of Microsoft IQ

1067
01:00:26,280 --> 01:00:30,557
Work IQ. This is Brightline's
response procedure

1068
01:00:30,640 --> 01:00:35,037
in SharePoint. It's the playbook the team
actually reaches for when something

1069
01:00:35,120 --> 01:00:38,117
goes wrong.
And the important thing is the agent

1070
01:00:38,200 --> 01:00:40,357
isn't working from a stale upload or

1071
01:00:40,440 --> 01:00:46,037
copied snapshot. It's answering from
the same source the team maintains day to

1072
01:00:46,120 --> 01:00:48,917
day. When the procedure
changes, the answer

1073
01:00:49,000 --> 01:00:51,437
changes with it. No re-uploads, no

1074
01:00:51,520 --> 01:00:53,957
prompt rebuilds, and no stale versions.

1075
01:00:54,040 --> 01:00:57,517
And critically, this is your knowledge.

1076
01:00:57,600 --> 01:01:02,677
The assets you create stay with you,
no matter what model or agent is reasoning

1077
01:01:02,760 --> 01:01:06,157
over them. Now, if we
return to our agent, we can

1078
01:01:06,240 --> 01:01:08,197
see the proper way to respond to this

1079
01:01:08,280 --> 01:01:11,037
issue. Not a generic recommendation, but

1080
01:01:11,120 --> 01:01:14,357
our playbook applied to this incident.

1081
01:01:14,440 --> 01:01:19,037
Now, we just saw one situation,
three questions, and one connected answer.

1082
01:01:19,120 --> 01:01:22,277
Now, let's go check back
in on our long-running agent.

1083
01:01:22,360 --> 01:01:24,077
And Oh.

1084
01:01:24,160 --> 01:01:27,357
Let's check the backup really quick.

1085
01:01:27,440 --> 01:01:31,997
Moment of truth. And

1086
01:01:32,080 --> 01:01:38,080
boom. OUR TASK FINISHED.

1087
01:01:38,830 --> 01:01:43,517
>> [applause] >> HERE WE CAN
SEE every step the agent took.

1088
01:01:43,600 --> 01:01:45,717
First, beginning with Web IQ, connecting

1089
01:01:45,800 --> 01:01:48,917
it to the outside world. Second, Fabric

1090
01:01:49,000 --> 01:01:53,277
IQ through Foundry, anchoring it
in the real estate of our operations. And

1091
01:01:53,360 --> 01:01:57,677
third, Work IQ, grounding it
in our people and procedures.

1092
01:01:57,760 --> 01:02:00,997
I triggered this manually, but with
Foundry routines, this can run on a

1093
01:02:01,080 --> 01:02:03,277
schedule, turning a one-off response

1094
01:02:03,360 --> 01:02:06,557
into continuous proactive execution. And

1095
01:02:06,640 --> 01:02:11,117
if we look closely, we can see
it even use the power of Work

1096
01:02:11,200 --> 01:02:14,957
IQ to alert me directly about
the incident. And if we go ahead and check

1097
01:02:15,040 --> 01:02:17,957
Teams, it sent me an incident brief

1098
01:02:18,040 --> 01:02:22,360
notifying me of the situation.

1099
01:02:23,370 --> 01:02:27,997
>> [applause] >> That's
the power of Microsoft IQ. When a

1100
01:02:28,080 --> 01:02:32,357
crisis hits, the team doesn't chase
answers across a dozen systems. They ask

1101
01:02:32,440 --> 01:02:36,437
a question and get a response grounded
in the world, their operations, and

1102
01:02:36,520 --> 01:02:42,160
their people, all in one place
they can trust. Back to you, Satya.

1103
01:02:43,020 --> 01:02:45,957
>> [applause] >> All right.

1104
01:02:46,040 --> 01:02:47,837
Thank you, Elijah.

1105
01:02:47,920 --> 01:02:50,357
Uh and now, you know,

1106
01:02:50,440 --> 01:02:53,037
let's sort of move up from this context

1107
01:02:53,120 --> 01:02:56,557
model layer uh to deploying these agents

1108
01:02:56,640 --> 01:03:00,557
and thinking about the runtime.
Um and you know, when you're building a

1109
01:03:00,640 --> 01:03:04,557
first-class um agentic system,
you need a first-class

1110
01:03:04,640 --> 01:03:07,237
agent runtime and a platform uh that we

1111
01:03:07,320 --> 01:03:09,357
are going to ship both with Windows as

1112
01:03:09,440 --> 01:03:12,317
well as part of Foundry in Azure. Uh we

1113
01:03:12,400 --> 01:03:16,037
want Windows to be a fantastic place to

1114
01:03:16,120 --> 01:03:18,517
run and scale agents.

1115
01:03:18,600 --> 01:03:21,477
Um and agents effectively are a new

1116
01:03:21,560 --> 01:03:23,597
execution environment, right? It's a new

1117
01:03:23,680 --> 01:03:27,077
paradigm even. They reason continuously.

1118
01:03:27,160 --> 01:03:30,117
They generate and run code dynamically.

1119
01:03:30,200 --> 01:03:33,757
Uh they take actions uh across files and

1120
01:03:33,840 --> 01:03:37,077
devices as well as across the network.

1121
01:03:37,160 --> 01:03:40,077
Um obviously there's a lot of power uh

1122
01:03:40,160 --> 01:03:44,117
in it, right? The fact that it can
generate code and act on it on a on a

1123
01:03:44,200 --> 01:03:46,557
long-running agent that's autonomous.

1124
01:03:46,640 --> 01:03:50,917
But obviously it creates new risk.
Uh and that's why today uh we're

1125
01:03:51,000 --> 01:03:52,957
introducing Microsoft Execution

1126
01:03:53,040 --> 01:03:56,797
Containers or MXC. Uh MXC is a new

1127
01:03:56,880 --> 01:04:00,397
policy layer that lets Windows apply

1128
01:04:00,480 --> 01:04:04,637
isolation and containment uh using AI

1129
01:04:04,720 --> 01:04:09,197
native or rather oh sorry, OS native
primitives, right? You need to bake this

1130
01:04:09,280 --> 01:04:11,597
into the operating system. Uh so that

1131
01:04:11,680 --> 01:04:14,597
the containment is enforced by policy.

1132
01:04:14,680 --> 01:04:17,477
Um you can have process-level isolation

1133
01:04:17,560 --> 01:04:19,717
for lightweight agent actions. You can

1134
01:04:19,800 --> 01:04:22,397
have session-level isolation for user

1135
01:04:22,480 --> 01:04:25,237
separation. Uh so Windows and Linux

1136
01:04:25,320 --> 01:04:27,237
virtual machines in fact are also great

1137
01:04:27,320 --> 01:04:30,437
including WSL for much uh stronger

1138
01:04:30,520 --> 01:04:33,157
boundaries. Uh and in fact if you want

1139
01:04:33,240 --> 01:04:38,237
full isolation and containment, Windows
365 for agents for maximum isolation in

1140
01:04:38,320 --> 01:04:39,717
a separate managed environment

1141
01:04:39,800 --> 01:04:42,677
effectively. Uh you can pick the right

1142
01:04:42,760 --> 01:04:45,397
containment option for the workload and

1143
01:04:45,480 --> 01:04:48,277
Windows will enforce it uh via MXC. And

1144
01:04:48,360 --> 01:04:50,957
so I think this becomes pretty critical

1145
01:04:51,040 --> 01:04:52,717
uh as you think about deploying agents

1146
01:04:52,800 --> 01:04:56,437
at scale on your Windows desktop. Uh we

1147
01:04:56,520 --> 01:05:01,477
want to ensure containment is enforced,
of course, regardless of who builds the

1148
01:05:01,560 --> 01:05:05,117
agent, right? So this is why you want to
bake it into the operating system. Uh to

1149
01:05:05,200 --> 01:05:09,237
that end we're working with many
partners to ensure that containment we

1150
01:05:09,320 --> 01:05:11,717
are building supports real developer

1151
01:05:11,800 --> 01:05:13,917
workloads out there and and also

1152
01:05:14,000 --> 01:05:16,677
addresses the needs. In fact, NVIDIA is

1153
01:05:16,760 --> 01:05:18,917
bringing Open Shell to Windows to

1154
01:05:19,000 --> 01:05:22,837
securely execute autonomous AI agents.

1155
01:05:22,920 --> 01:05:27,277
And today we're really
thrilled to announce that

1156
01:05:27,360 --> 01:05:32,720
Open Claw runs on Windows leveraging MXC.

1157
01:05:33,320 --> 01:05:35,340
>> [applause]

1158
01:05:35,520 --> 01:05:37,917
>> Yeah, we are very deeply engaged with

1159
01:05:38,000 --> 01:05:40,797
the team to make Open Claw run super

1160
01:05:40,880 --> 01:05:43,197
well on Windows. And so let me hand it

1161
01:05:43,280 --> 01:05:48,717
over to my colleagues Scott and Samantha
to show you Open Claw on Windows. Scott,

1162
01:05:48,800 --> 01:05:53,030
Samantha. >> [applause] [cheering]

1163
01:05:55,400 --> 01:05:59,877
>> Hey friends. You know, Open Claw came
out in November of last year and it took

1164
01:05:59,960 --> 01:06:03,717
the world by storm. And for the last
several months, I've been using my Open

1165
01:06:03,800 --> 01:06:05,797
Claw to stay on top of my health.

1166
01:06:05,880 --> 01:06:08,037
My Claw can help me manage my blood

1167
01:06:08,120 --> 01:06:10,437
sugar and it gives me even proactive

1168
01:06:10,520 --> 01:06:15,237
notifications via heartbeat. I've got
mine triaging my personal email, it's

1169
01:06:15,320 --> 01:06:18,717
doing my GitHub issues and tracking
packages and it even buys me movie

1170
01:06:18,800 --> 01:06:21,637
tickets. So Samantha,
what are you using your Claw for?

1171
01:06:21,720 --> 01:06:26,597
>> That's cool, Scott. I turned my Open
Claw agent into my triathlon coach. I'm

1172
01:06:26,680 --> 01:06:30,357
nowhere near ready for my race
in September, but coach Claw developed a

1173
01:06:30,440 --> 01:06:35,117
work back plan for me and is using
my Strava data to notify me on how I'm

1174
01:06:35,200 --> 01:06:38,517
progressing and to keep me
accountable when I'm slacking.

1175
01:06:38,600 --> 01:06:41,877
>> Now we've both found our Claws
to be super useful and that's why we're

1176
01:06:41,960 --> 01:06:46,197
working closely with Open Claw to make
them successful and even more successful

1177
01:06:46,280 --> 01:06:50,237
on Windows. We've been collaborating
in the open on GitHub to bring you all an

1178
01:06:50,320 --> 01:06:54,437
Open Claw Windows companion app. It's
going to help you set up your own Claws

1179
01:06:54,520 --> 01:06:58,837
or connect to existing ones whether
they're hosted in Windows or in WSL. And

1180
01:06:58,920 --> 01:07:00,357
the Windows companion, we're going to

1181
01:07:00,440 --> 01:07:03,357
sandbox the Open Claw tool calls to keep

1182
01:07:03,440 --> 01:07:07,637
you and your system safe. >> Yeah,
you'll see the Open Claw Windows

1183
01:07:07,720 --> 01:07:11,157
Companion app running right now in the
background. Go ahead and right-click on

1184
01:07:11,240 --> 01:07:14,640
it, Scott. >> All right.

1185
01:07:15,320 --> 01:07:18,637
>> That looks awesome. You'll notice
immediately it looks like a native

1186
01:07:18,720 --> 01:07:23,237
Windows app because it is. It's written
in WinUI 3. It's got all kinds of

1187
01:07:23,320 --> 01:07:26,917
information about my gateway, other
machines that are participating in my

1188
01:07:27,000 --> 01:07:30,077
claw, my sessions, and my usage. I've

1189
01:07:30,160 --> 01:07:35,117
also got quick access to things like
chat, canvas, the main dashboard, and

1190
01:07:35,200 --> 01:07:39,237
more. Let's jump into companion settings.

1191
01:07:39,320 --> 01:07:43,197
Within the app, we've got full chat
support with tool calling. And you'll

1192
01:07:43,280 --> 01:07:45,277
notice down here in the corner,

1193
01:07:45,360 --> 01:07:50,317
we've got lots of permissions options
along with our sandbox configuration.

1194
01:07:50,400 --> 01:07:54,317
>> Now, this sandbox is really interesting
because this is using MXC, the Microsoft

1195
01:07:54,400 --> 01:07:58,237
Execution Containers, and for this we're
going to be using process isolation.

1196
01:07:58,320 --> 01:08:01,597
Now, newer versions of Windows are going
to have even more containment options,

1197
01:08:01,680 --> 01:08:04,237
so you're going to want to keep an eye
out for more news with MXC in the

1198
01:08:04,320 --> 01:08:06,317
future. Now, I can see that I've got

1199
01:08:06,400 --> 01:08:11,477
one-click security option settings,
but Samantha, talk to me about custom

1200
01:08:11,560 --> 01:08:14,837
folders. >> Yeah, you've
got full support about what

1201
01:08:14,920 --> 01:08:19,157
files and folders you want Open Claw
to have access to, and really granular

1202
01:08:19,240 --> 01:08:21,357
security features like clipboard access

1203
01:08:21,440 --> 01:08:23,837
or talking to the internet itself.

1204
01:08:23,920 --> 01:08:27,117
Now, I've given it read-only desktop

1205
01:08:27,200 --> 01:08:29,397
read access to your desktop folder.

1206
01:08:29,480 --> 01:08:33,837
Open Claw already has a rich safety layer,
and that layer is only augmented

1207
01:08:33,920 --> 01:08:35,837
more by appropriate containment that can

1208
01:08:35,920 --> 01:08:38,477
be managed by me or policies applied by

1209
01:08:38,560 --> 01:08:43,517
IT. Now, for the purposes of this demo,
I'm going to do something really scary

1210
01:08:43,600 --> 01:08:46,877
and ask Open Claw to delete
all the files on your desktop.

1211
01:08:46,960 --> 01:08:49,477
>> That's cool, cuz I
have a nice clean desktop.

1212
01:08:49,560 --> 01:08:53,997
>> Um I think you're keeping a secret.
You hid all your icons from the audience

1213
01:08:54,080 --> 01:08:57,877
while we were on stage. >> That is >>
I think we need to show the audience who

1214
01:08:57,960 --> 01:09:02,837
you really are. >> So disre- so
disrespectful. I know where everything is.

1215
01:09:02,920 --> 01:09:05,997
>> [laughter] >> Just don't
touch my stuff. I know where

1216
01:09:06,080 --> 01:09:09,757
they are. You know, Sam, I need to make
sure that you're clear that a messy

1217
01:09:09,840 --> 01:09:14,077
desktop is an organized mind. I'm pretty
sure that that's the That's the quote,

1218
01:09:14,160 --> 01:09:17,397
right? So, what we've done is we've
asked Open Claw to delete those files

1219
01:09:17,480 --> 01:09:20,437
from the Windows node, and the only
thing that is going to keep from

1220
01:09:20,520 --> 01:09:23,877
happening is MXC because we've turned
off all of the many layers that Open

1221
01:09:23,960 --> 01:09:29,077
Claw offers, but our IT, in this case
Samantha, has set it to read-only. So,

1222
01:09:29,160 --> 01:09:32,077
it's trying to go and delete all
of those files. We can actually see the

1223
01:09:32,160 --> 01:09:35,317
different attempts where it's going and
deleting and then checking the directory

1224
01:09:35,400 --> 01:09:38,877
and then deleting again because it's
very persistent. It wants these files

1225
01:09:38,960 --> 01:09:42,137
gone, and I want I want them to stay.

1226
01:09:42,220 --> 01:09:44,517
>> [sighs] >> Nope. Nope.

1227
01:09:44,600 --> 01:09:47,517
The read-only sandbox is there. 94 JPEGs

1228
01:09:47,600 --> 01:09:52,117
are still on the desktop. Absolutely.
My desktop icons are safe from Samantha's

1229
01:09:52,200 --> 01:09:54,997
reign of terror. >> Foiled again.

1230
01:09:55,080 --> 01:09:59,597
>> Oh my goodness. So bad.
>> Today we've seen security sandboxing,

1231
01:09:59,680 --> 01:10:04,077
WinUI 3, and open source brought
together all in this alpha release of a

1232
01:10:04,160 --> 01:10:08,717
Windows companion app. We think this app
is a great opportunity to showcase Open

1233
01:10:08,800 --> 01:10:12,037
Claw on Windows, and it's only going
to get better in the coming months.

1234
01:10:12,120 --> 01:10:14,797
>> That is right. And by the way,
I want to note that we're doing all this

1235
01:10:14,880 --> 01:10:17,037
development work on this calm Windows

1236
01:10:17,120 --> 01:10:19,357
development machine with all the tools

1237
01:10:19,440 --> 01:10:22,237
that I love like WSL,

1238
01:10:22,320 --> 01:10:25,117
uh containers, and containment. And I've

1239
01:10:25,200 --> 01:10:29,277
even got GitHub Copilot with multi-model
support ready to go. Now, this is

1240
01:10:29,360 --> 01:10:34,437
literally how the team and I have been
working on the Open Claw app on GitHub.

1241
01:10:34,520 --> 01:10:36,157
Now, people might wonder how all of this

1242
01:10:36,240 --> 01:10:38,957
came together. Turns out that over the

1243
01:10:39,040 --> 01:10:41,517
holidays, I got a DM from this random

1244
01:10:41,600 --> 01:10:44,517
guy on on the internet, and it took me a

1245
01:10:44,600 --> 01:10:49,197
couple of days to get back to him.
Uh and then when I did, I we just had this

1246
01:10:49,280 --> 01:10:52,797
kind of cool idea that
maybe he should come to

1247
01:10:52,880 --> 01:10:55,597
to to Microsoft Build. Now, I think
there's one more person that we all want

1248
01:10:55,680 --> 01:10:58,917
to thank for bringing the next
generation of agents to world. Everybody

1249
01:10:59,000 --> 01:11:02,707
put your hands together for the claw
father himself, Peter Steinberger.

1250
01:11:02,790 --> 01:11:04,810
>> [applause]

1251
01:11:05,800 --> 01:11:09,357
>> Yep. Samantha.

1252
01:11:09,440 --> 01:11:12,637
She's just showing off my secret DMs.

1253
01:11:12,720 --> 01:11:14,117
Um

1254
01:11:14,200 --> 01:11:17,437
I'm so excited to see Open Claw native

1255
01:11:17,520 --> 01:11:20,077
on Windows. You know, watching a claw

1256
01:11:20,160 --> 01:11:26,287
try to delete all your desktop file
and just fail made me really happy.

1257
01:11:26,370 --> 01:11:30,517
>> [laughter] >> Because 6 months
ago that would have totally worked.

1258
01:11:30,600 --> 01:11:32,997
You know, I built Open Claw to have

1259
01:11:33,080 --> 01:11:35,037
access to everything. You know, my

1260
01:11:35,120 --> 01:11:38,517
files, my machines, my chats, always on

1261
01:11:38,600 --> 01:11:42,277
and fully open source.
That's what makes it so powerful.

1262
01:11:42,360 --> 01:11:45,717
And that's what also makes
companies a bit nervous.

1263
01:11:45,800 --> 01:11:50,277
You know, what I kept hearing
was Peter, I love my claw.

1264
01:11:50,360 --> 01:11:54,437
Can I use this at work?
And that's what we spent the last few

1265
01:11:54,520 --> 01:11:57,277
months on with Microsoft, GitHub,

1266
01:11:57,360 --> 01:12:00,877
OpenAI, Nvidia, just to name a few.

1267
01:12:00,960 --> 01:12:05,957
We added observability. We added auto
mode for permissions. We changed how

1268
01:12:06,040 --> 01:12:09,637
access works. It's not all or nothing
anymore. You can pick which folder

1269
01:12:09,720 --> 01:12:11,837
should be read only, which one should be

1270
01:12:11,920 --> 01:12:14,157
writeable or hidden.

1271
01:12:14,240 --> 01:12:21,080
So, here's the news. You can totally
run Open Claw inside YOUR COMPANY NOW.

1272
01:12:21,390 --> 01:12:23,410
>> [applause]

1273
01:12:23,760 --> 01:12:26,837
>> AND WE EVEN MADE THE HARNESS itself a

1274
01:12:26,920 --> 01:12:29,277
plugin. You can bring your own.

1275
01:12:29,360 --> 01:12:32,717
Copilot, Codex, whatever you already

1276
01:12:32,800 --> 01:12:36,677
trust and your rules come right with it.

1277
01:12:36,760 --> 01:12:38,917
And then you put Open Claw on top of it.

1278
01:12:39,000 --> 01:12:41,597
You get persistent memory, heartbeats,

1279
01:12:41,680 --> 01:12:44,437
and you get a claw right inside Slack or

1280
01:12:44,520 --> 01:12:49,197
Teams. It's been really
exciting to see Open

1281
01:12:49,280 --> 01:12:50,197
Claw

1282
01:12:50,280 --> 01:12:53,917
grow into something much bigger.

1283
01:12:54,000 --> 01:12:57,277
A global movement and a community.

1284
01:12:57,360 --> 01:13:02,357
And I started the Open Claw Foundation,
a real non-profit, so it stays open and

1285
01:13:02,440 --> 01:13:07,397
neutral. Any model, any operating system.

1286
01:13:07,480 --> 01:13:09,317
Because we are entering

1287
01:13:09,400 --> 01:13:12,597
a new era of building these agents. More

1288
01:13:12,680 --> 01:13:15,677
capability for the people who don't code

1289
01:13:15,760 --> 01:13:18,837
and more power for those who do.

1290
01:13:18,920 --> 01:13:23,517
And we get to do this.
We get to build it together in the open.

1291
01:13:23,600 --> 01:13:28,767
So, my task is simple.
Come build with us. Thank you.

1292
01:13:28,850 --> 01:13:31,677
>> [applause] >> Fantastic.

1293
01:13:31,760 --> 01:13:35,957
Thank you very much to Peter. Thank you
to Samantha. Thank you to the Companion

1294
01:13:36,040 --> 01:13:39,317
app team for their hard work. Thank you
to the Open Claw community for giving

1295
01:13:39,400 --> 01:13:41,197
everyone here a crustacean of their own.

1296
01:13:41,280 --> 01:13:43,957
And as we walk out, I want to remind you
that that's the first time you've ever

1297
01:13:44,040 --> 01:13:47,437
seen Open Claw running
on a Surface Laptop Ultra.

1298
01:13:47,520 --> 01:13:49,870
Goodbye. [cheering]

1299
01:13:50,470 --> 01:13:52,490
>> [applause]

1300
01:13:54,080 --> 01:13:56,877
>> All right. Thank you so much, Samantha,

1301
01:13:56,960 --> 01:13:59,597
Scott, and Peter. It's so wonderful to

1302
01:13:59,680 --> 01:14:03,397
see Open Claw come to Windows

1303
01:14:03,480 --> 01:14:07,317
and have all of that capability in terms

1304
01:14:07,400 --> 01:14:11,277
of the security and and that
comfort to be able to have

1305
01:14:11,360 --> 01:14:15,757
these long-running agents and and
unmetered intelligence come together.

1306
01:14:15,840 --> 01:14:17,997
So, now let's move

1307
01:14:18,080 --> 01:14:20,877
from the Windows side on the edge to the

1308
01:14:20,960 --> 01:14:23,517
cloud with Foundry. And we're building

1309
01:14:23,600 --> 01:14:26,837
Foundry into this full application

1310
01:14:26,920 --> 01:14:29,557
platform for the agent era, right? I

1311
01:14:29,640 --> 01:14:33,117
mean, yeah, every era of the platform

1312
01:14:33,200 --> 01:14:36,877
shift, when we moved to the cloud,
we had the cloud native app stack. And now

1313
01:14:36,960 --> 01:14:40,157
we have the agent native
app stack in Foundry.

1314
01:14:40,240 --> 01:14:42,957
We're particularly excited about the

1315
01:14:43,040 --> 01:14:46,637
Foundry hosted agent Uh, as a runtime

1316
01:14:46,720 --> 01:14:51,357
for long running agents, um, it you know,
agents now have uh, so if you're

1317
01:14:51,440 --> 01:14:55,317
building in Foundry um, uh,
hosted agents, you can have all

1318
01:14:55,400 --> 01:14:59,197
the IQ layers. Uh,
you have the tools. You have the

1319
01:14:59,280 --> 01:15:03,397
durability uh,
and the memory and the state. Uh,

1320
01:15:03,480 --> 01:15:07,597
you have your own sandbox.
In fact, it's a super fast uh,

1321
01:15:07,680 --> 01:15:10,037
sandbox that you can spin up. Uh, you

1322
01:15:10,120 --> 01:15:12,877
can generate the rubrics. Um,

1323
01:15:12,960 --> 01:15:17,397
you get and the evals. Uh,
you can in fact have all the safety and

1324
01:15:17,480 --> 01:15:20,917
the guardrails around your agentic system.

1325
01:15:21,000 --> 01:15:24,637
Uh, and in fact, even one of the coolest
things in Foundry is it's a continuously

1326
01:15:24,720 --> 01:15:28,357
improving loop even, right? It's got
that self-improvement loop built in,

1327
01:15:28,440 --> 01:15:30,077
right? So, you build an agent that's

1328
01:15:30,160 --> 01:15:35,717
continuously getting better. Uh,
and I'm really also excited today

1329
01:15:35,800 --> 01:15:39,957
uh, to announce a partnership
with Fireworks AI,

1330
01:15:40,040 --> 01:15:42,717
uh, bringing in fact all of their open

1331
01:15:42,800 --> 01:15:45,597
weight models to Foundry. Uh, that means

1332
01:15:45,680 --> 01:15:49,877
giving you as developers more choices uh,
as well as their great influence

1333
01:15:49,960 --> 01:15:52,437
stack to build the next generation of

1334
01:15:52,520 --> 01:15:54,637
these agentic applications with all the

1335
01:15:54,720 --> 01:15:58,197
enterprise rails that Foundry has

1336
01:15:58,280 --> 01:16:01,677
uh, as well as the governance uh,
Foundry has. So, really excited about

1337
01:16:01,760 --> 01:16:06,110
that partnership. >> [applause]

1338
01:16:07,120 --> 01:16:10,197
>> And so,
now that brings us to the tools.

1339
01:16:10,280 --> 01:16:15,477
Uh, GitHub itself is at the heart of all
this. Uh, you know, in fact, Jensen

1340
01:16:15,560 --> 01:16:19,397
spoke to this. The GitHub is not just
about the code repo. It's becoming the

1341
01:16:19,480 --> 01:16:23,437
control plane uh, for all
the agents. Um, and nearly

1342
01:16:23,520 --> 01:16:25,757
everything we measure on GitHub uh,

1343
01:16:25,840 --> 01:16:28,637
whether it's repo creation, PR activity,

1344
01:16:28,720 --> 01:16:31,757
API usage, actions, all of them uh, are

1345
01:16:31,840 --> 01:16:33,797
growing faster uh, because of these

1346
01:16:33,880 --> 01:16:37,077
agentic workflows. Uh, this new scale is

1347
01:16:37,160 --> 01:16:39,197
driven by humans and agents

1348
01:16:39,280 --> 01:16:41,317
collaborating together. Uh, and we're

1349
01:16:41,400 --> 01:16:44,637
exposing our tooling across every form

1350
01:16:44,720 --> 01:16:50,997
factor. In fact, we've seen
tremendous growth in CLI driven, you

1351
01:16:51,080 --> 01:16:55,197
know, the approachability of the CLI form
factor. It's always been great to go to

1352
01:16:55,280 --> 01:17:00,037
a terminal. And now though, when combined
with the power of the models and natural

1353
01:17:00,120 --> 01:17:04,317
language, you know, CLI has become the
thing that everybody goes to. But at the

1354
01:17:04,400 --> 01:17:05,837
end of the day,

1355
01:17:05,920 --> 01:17:09,797
when you have hundreds of CLIs, it

1356
01:17:09,880 --> 01:17:13,557
becomes complicated. It doesn't scale.

1357
01:17:13,640 --> 01:17:15,477
Especially the cognitive load that I

1358
01:17:15,560 --> 01:17:19,117
have when you have 100 CLI sessions open

1359
01:17:19,200 --> 01:17:21,277
is such that you kind of need something

1360
01:17:21,360 --> 01:17:23,877
new. And so that's what has led us to

1361
01:17:23,960 --> 01:17:27,517
build We know we need like this tool

1362
01:17:27,600 --> 01:17:30,477
essentially that has the speed and the

1363
01:17:30,560 --> 01:17:33,357
flexibility of a CLI,

1364
01:17:33,440 --> 01:17:35,797
but has the capability of an IDE and the

1365
01:17:35,880 --> 01:17:38,597
ability to scale to infinite number of

1366
01:17:38,680 --> 01:17:40,877
agent sessions. And so today we're

1367
01:17:40,960 --> 01:17:43,997
taking that next big step introducing

1368
01:17:44,080 --> 01:17:48,120
our new GitHub Copilot app.

1369
01:17:49,550 --> 01:17:53,557
>> [applause] >> And we realize
that it's not sufficient,

1370
01:17:53,640 --> 01:17:56,957
right? Because you still have a back
end to deal with, right? Code is easy to

1371
01:17:57,040 --> 01:17:58,637
generate, but what about the back end?

1372
01:17:58,720 --> 01:18:01,117
So you need to contend with identities,

1373
01:18:01,200 --> 01:18:04,277
storage, database schemas. And that's

1374
01:18:04,360 --> 01:18:06,637
why we're really super excited about

1375
01:18:06,720 --> 01:18:08,397
Rayfin. Rayfin

1376
01:18:08,480 --> 01:18:12,477
is an agent first SDK that connects your

1377
01:18:12,560 --> 01:18:15,837
agents to a back end as a service.

1378
01:18:15,920 --> 01:18:18,197
And we're bringing this to everywhere

1379
01:18:18,280 --> 01:18:20,517
you build. That's why I'm again super

1380
01:18:20,600 --> 01:18:23,517
excited about Rayfin and the partnership

1381
01:18:23,600 --> 01:18:30,520
with Replit. You can now build
apps in Replit while the app

1382
01:18:31,120 --> 01:18:33,277
You can build the app in Replit while

1383
01:18:33,360 --> 01:18:38,357
the app and data are deployed into the
enterprise managed fabric tenant thanks

1384
01:18:38,440 --> 01:18:41,997
to, uh, the Ray Fin SDK.
Uh, and so this Ray

1385
01:18:42,080 --> 01:18:46,237
Fin SDK is available now for anyone else,
uh, to be able to use with their

1386
01:18:46,320 --> 01:18:48,397
tools as a back end, uh, and essentially

1387
01:18:48,480 --> 01:18:50,357
have this back end service.

1388
01:18:50,440 --> 01:18:53,357
And now to show you all of the foundry

1389
01:18:53,440 --> 01:18:56,117
and Ray Fin and building agents and the

1390
01:18:56,200 --> 01:19:01,960
long running agents, let me invite up
on stage Cassidy. Cassidy, take it away.

1391
01:19:02,250 --> 01:19:05,517
>> [applause] >> Hello, everyone.

1392
01:19:05,600 --> 01:19:09,317
I'm so excited to be amongst my fellow
devs today. This has been a long day.

1393
01:19:09,400 --> 01:19:11,357
Fix your shoulders. Sit back. Get ready.

1394
01:19:11,440 --> 01:19:13,157
I see a lot of you sitting up. Great.

1395
01:19:13,240 --> 01:19:16,717
Great. I know you're drinking from the
firehose of information today, so I want

1396
01:19:16,800 --> 01:19:20,237
you to go into the next few minutes
thinking, what can I try out on my

1397
01:19:20,320 --> 01:19:23,597
laptop later today?
So as Satya said, we're going to show

1398
01:19:23,680 --> 01:19:28,757
you Ray Fin, but first, I can't wait
to show you the new GitHub Copilot app.

1399
01:19:28,840 --> 01:19:31,837
This app is your home base
for development and operations on your

1400
01:19:31,920 --> 01:19:35,477
computer, and we think you're going
to love it. So let me show you around.

1401
01:19:35,560 --> 01:19:38,277
When you open up the app from the start,
you see this home screen here, where you

1402
01:19:38,360 --> 01:19:42,357
can kick off a new agentic coding session.
But also before I get into the

1403
01:19:42,440 --> 01:19:44,917
serious stuff, you can drag Mona around,

1404
01:19:45,000 --> 01:19:49,277
and there's a game. Look, it's so fun.

1405
01:19:49,360 --> 01:19:53,637
Okay, I'm not very good at it, so let's
just get back to you can kick off a new

1406
01:19:53,720 --> 01:19:55,797
agentic coding session.

1407
01:19:55,880 --> 01:20:00,837
So, I started off one a little bit
earlier here, and it gave me a review of

1408
01:20:00,920 --> 01:20:05,557
a bunch of release blockers.
Which one should I fix? Call it out.

1409
01:20:05,640 --> 01:20:09,197
83? The critical ones? You know what? How

1410
01:20:09,280 --> 01:20:11,997
about we just do all of them?

1411
01:20:12,080 --> 01:20:14,557
Let's go. This app will now kick off a

1412
01:20:14,640 --> 01:20:19,077
separate session for every single issue
here. I don't worry have to worry about

1413
01:20:19,160 --> 01:20:23,397
stashing or coding conflicts or anything
cuz the app takes care of that with Git

1414
01:20:23,480 --> 01:20:26,797
worktrees. Git worktrees
are isolated environments

1415
01:20:26,880 --> 01:20:29,957
for each session that you run, so
your agents can work in parallel without

1416
01:20:30,040 --> 01:20:34,037
stepping on each other.
But you still have to merge them, right?

1417
01:20:34,120 --> 01:20:37,557
So, uh Copilot has
your back there, too. If

1418
01:20:37,640 --> 01:20:40,117
I head over, not to this one, but to

1419
01:20:40,200 --> 01:20:43,637
this issue here, I can run agent merge.

1420
01:20:43,720 --> 01:20:47,837
And when I enable agent merge,
Copilot will continuously babysit this PR

1421
01:20:47,920 --> 01:20:52,397
through CI checks, code review,
and merge conflicts. Okay, let me keep

1422
01:20:52,480 --> 01:20:54,997
showing you around while
those are still running.

1423
01:20:55,080 --> 01:20:57,637
Now, if I head over to my work,

1424
01:20:57,720 --> 01:21:00,357
I can see a focused view of all of my

1425
01:21:00,440 --> 01:21:04,717
activity and just projects loaded in
the app, issues and PRs, everything here.

1426
01:21:04,800 --> 01:21:08,557
And then, under automations, I have
a bunch of reusable sessions and workflows

1427
01:21:08,640 --> 01:21:11,117
that can run locally or on the cloud.

1428
01:21:11,200 --> 01:21:12,797
You see, there's issue poetry there.

1429
01:21:12,880 --> 01:21:15,317
That is real and that is load-bearing.

1430
01:21:15,400 --> 01:21:20,397
Okay. And now, [laughter] under sessions,
like I briefly showed earlier, these are

1431
01:21:20,480 --> 01:21:23,277
sessions. If I want to add
a new repository, I can

1432
01:21:23,360 --> 01:21:27,037
click that button here and it can
pull from a local repo or from a GitHub

1433
01:21:27,120 --> 01:21:29,717
repository. And then, if I were to just

1434
01:21:29,800 --> 01:21:32,677
add one, I can add a session in Pocket

1435
01:21:32,760 --> 01:21:37,037
Cal. This is an open-source repo.
I can start a session anywhere and it just

1436
01:21:37,120 --> 01:21:40,997
loads it. I don't have to clone,
I don't have to pull, it just works.

1437
01:21:41,080 --> 01:21:44,397
Now, I when I look at a session within
this repository, let me look at this

1438
01:21:44,480 --> 01:21:48,597
other one over here, I get an integrated
browser, there's a terminal, I can see

1439
01:21:48,680 --> 01:21:52,677
the chat, it's all loading. I can even
toggle light mode and dark mode in here.

1440
01:21:52,760 --> 01:21:56,917
And there's also this great button, pick
and polish, where if I click that, I can

1441
01:21:57,000 --> 01:21:59,557
pick and polish anything in this app and

1442
01:21:59,640 --> 01:22:04,597
it adds it to the chat and I can say,
"Hey, I want you to add reordering to

1443
01:22:04,680 --> 01:22:08,077
this list." And it'll just
work, all living in there.

1444
01:22:08,160 --> 01:22:11,917
I have access to all the most popular
models via my single GitHub Copilot

1445
01:22:12,000 --> 01:22:16,717
subscription, including those from OpenAI,
Anthropic, and Google. You can

1446
01:22:16,800 --> 01:22:20,517
see all of them in our model picker.
And having model choice is great. Not only

1447
01:22:20,600 --> 01:22:23,797
can you pick the right model for the task,
but for bigger features, Copilot

1448
01:22:23,880 --> 01:22:26,157
can request a rubber duck review. So, in

1449
01:22:26,240 --> 01:22:28,077
this session here, for example, I was

1450
01:22:28,160 --> 01:22:31,877
using GPT-5.5, but if I scroll up, it

1451
01:22:31,960 --> 01:22:35,677
actually requested one
from Claude Opus 4.8.

1452
01:22:35,760 --> 01:22:39,557
All models have blind spots, and the power
of the Copilot multimodal approach

1453
01:22:39,640 --> 01:22:42,197
means that I can catch them earlier.

1454
01:22:42,280 --> 01:22:44,237
And this is all very cool.

1455
01:22:44,320 --> 01:22:49,357
But working with AI in 2026 should
be more than just chat. You just

1456
01:22:49,440 --> 01:22:52,357
saw me scrolling on that.
There's so many words here.

1457
01:22:52,440 --> 01:22:54,717
So, today, I'm very excited to show you

1458
01:22:54,800 --> 01:22:57,517
the concept of a canvas. I'm going to

1459
01:22:57,600 --> 01:22:59,597
open one right there.

1460
01:22:59,680 --> 01:23:01,917
The canvas is how an agent can build a

1461
01:23:02,000 --> 01:23:05,157
custom UI to communicate with you.

1462
01:23:05,240 --> 01:23:08,917
What if your AI could see?

1463
01:23:09,000 --> 01:23:11,477
Everyone say, "Demo gods bless us."

1464
01:23:11,560 --> 01:23:13,877
Okay, let's see if it works.

1465
01:23:13,960 --> 01:23:16,757
Here is a fun canvas where, if I get the

1466
01:23:16,840 --> 01:23:19,677
camera going, okay, the agent shows your

1467
01:23:19,760 --> 01:23:24,237
PRs down here, and I can take toggle it
with a thumbs up or a thumbs down. Let's

1468
01:23:24,320 --> 01:23:29,397
approve it. Hey!
It's so fun! >> [laughter]

1469
01:23:29,480 --> 01:23:32,517
>> And they can go so much deeper if you
want. This is just the beginning of what

1470
01:23:32,600 --> 01:23:35,277
you can do. So, again,
I kicked off a bunch of

1471
01:23:35,360 --> 01:23:38,157
different sessions earlier.

1472
01:23:38,240 --> 01:23:41,557
This is a signal box app. It's 100%

1473
01:23:41,640 --> 01:23:45,477
agent-built. It's containerized
with a database back-end.

1474
01:23:45,560 --> 01:23:50,597
Would you be able to deploy this to your
enterprise with no questions asked? Be

1475
01:23:50,680 --> 01:23:56,237
honest. No. Yes, exactly.
No, but you can with

1476
01:23:56,320 --> 01:23:59,197
Rayfen. And that is very,
very exciting. Let me

1477
01:23:59,280 --> 01:24:02,397
open up a new terminal over here.

1478
01:24:02,480 --> 01:24:07,277
All I have to do is type Rayfen up. And

1479
01:24:07,360 --> 01:24:10,277
then, demo gods bless us. Come on.

1480
01:24:10,360 --> 01:24:12,557
It will

1481
01:24:12,640 --> 01:24:16,117
maybe deploy. Blammo! It's happening.

1482
01:24:16,200 --> 01:24:20,437
Yes! And all hosted on Microsoft Fabric.

1483
01:24:20,520 --> 01:24:23,277
Okay, I know that's a lot.
I know that's a lot.

1484
01:24:23,360 --> 01:24:27,037
With Rayfin, your agents get a complete
enterprise back end, so you can deploy

1485
01:24:27,120 --> 01:24:30,677
with confidence in the way that's best
for you. But, this is the key thing to

1486
01:24:30,760 --> 01:24:33,317
remember. This app

1487
01:24:33,400 --> 01:24:35,437
is not just another session manager.

1488
01:24:35,520 --> 01:24:39,997
Yes, it is. It It manages a lot of
sessions, but session managers just make

1489
01:24:40,080 --> 01:24:44,917
it easy to create work, but GitHub
Copilot helps you to finish it.

1490
01:24:45,000 --> 01:24:47,357
Thank you so much.
Happy Pride. Back to you, Satya.

1491
01:24:47,440 --> 01:24:49,460
>> [applause and cheering]

1492
01:24:52,680 --> 01:24:54,117
[applause]

1493
01:24:54,200 --> 01:24:57,637
>> All right.
Um now we're back to IDs that

1494
01:24:57,720 --> 01:25:00,637
have UI. That's sort of so cool.

1495
01:25:00,720 --> 01:25:02,957
Um you know, it's kind of come so full

1496
01:25:03,040 --> 01:25:07,077
circle. Um now let's talk about how you

1497
01:25:07,160 --> 01:25:10,197
can observe, govern, and secure these

1498
01:25:10,280 --> 01:25:13,917
agents. Uh Agent 365 is the agent uh

1499
01:25:14,000 --> 01:25:17,237
control plane. Uh agent requires

1500
01:25:17,320 --> 01:25:19,717
um their own identities, access

1501
01:25:19,800 --> 01:25:24,357
controls, um even when they're working
on your behalf, right? You just want

1502
01:25:24,440 --> 01:25:26,517
that work on behalf identity uh to be

1503
01:25:26,600 --> 01:25:29,877
enforced. So, we extended in Entra.

1504
01:25:29,960 --> 01:25:33,157
Uh agents need uh real-time defense, so

1505
01:25:33,240 --> 01:25:35,037
we extended Defender.

1506
01:25:35,120 --> 01:25:40,237
Uh agents require this always-on
data protections and compliance, so we

1507
01:25:40,320 --> 01:25:41,997
extended Purview.

1508
01:25:42,080 --> 01:25:44,437
Um and these agents can be hosted

1509
01:25:44,520 --> 01:25:47,677
anywhere. They can be on AWS, GCP, not

1510
01:25:47,760 --> 01:25:49,877
just on Azure, uh or built with any

1511
01:25:49,960 --> 01:25:53,717
framework. Um and today we're announcing

1512
01:25:53,800 --> 01:25:57,077
a number of updates, including the GA of

1513
01:25:57,160 --> 01:26:00,917
Agent 365 SDK, and we're expanding it to

1514
01:26:01,000 --> 01:26:03,317
your local agents uh running on Windows

1515
01:26:03,400 --> 01:26:05,197
and elsewhere, and the clause we just

1516
01:26:05,280 --> 01:26:08,037
saw earlier. And let us take a look at

1517
01:26:08,120 --> 01:26:14,800
how all of Agent 365 composes.
Uh over to you, Amanda.

1518
01:26:16,520 --> 01:26:20,117
>> [applause] >> Everyone
is building agents, but that's

1519
01:26:20,200 --> 01:26:22,677
not the hard part anymore. The hard part

1520
01:26:22,760 --> 01:26:24,717
is integrating them into your business

1521
01:26:24,800 --> 01:26:27,317
and governing them at scale. Today, I'll

1522
01:26:27,400 --> 01:26:30,317
show you how Foundry makes this easy.

1523
01:26:30,400 --> 01:26:34,877
Let's start locally. I've already built
a LangGraph agent, and now I'll show you

1524
01:26:34,960 --> 01:26:39,157
the value Foundry adds. First, tools.

1525
01:26:39,240 --> 01:26:41,517
Agents need tools to get work done, and

1526
01:26:41,600 --> 01:26:43,437
with Foundry toolbox, I just add my

1527
01:26:43,520 --> 01:26:46,277
tools once, and any agent can consume

1528
01:26:46,360 --> 01:26:49,677
them through a single MCP endpoint.

1529
01:26:49,760 --> 01:26:54,357
And because tools live centrally,
that means governance does as well.

1530
01:26:54,440 --> 01:26:56,157
I've applied a guardrail, which blocks

1531
01:26:56,240 --> 01:26:59,037
PII from leaking into tool calls, into

1532
01:26:59,120 --> 01:27:02,237
responses, and all I have to do is apply

1533
01:27:02,320 --> 01:27:06,077
this once, and all
my agents are protected.

1534
01:27:06,160 --> 01:27:09,557
Now, let's make this
agent enterprise-ready.

1535
01:27:09,640 --> 01:27:13,477
To deploy this agent to Foundry,
all I have to do is add this one block of

1536
01:27:13,560 --> 01:27:15,957
code, and then I push my changes, and

1537
01:27:16,040 --> 01:27:18,837
GitHub Actions takes it from there.

1538
01:27:18,920 --> 01:27:23,677
Once deployed, I can actually use
this agent in the same Foundry extension I

1539
01:27:23,760 --> 01:27:25,357
showed you earlier.

1540
01:27:25,440 --> 01:27:29,117
And let's now test it out. I'm asking it

1541
01:27:29,200 --> 01:27:33,277
to track a few open items from
a stand-up I had earlier today,

1542
01:27:33,360 --> 01:27:35,477
and what's actually happening right now

1543
01:27:35,560 --> 01:27:37,917
is Foundry is spinning up a dedicated

1544
01:27:38,000 --> 01:27:41,277
micro VM just for this session. And the

1545
01:27:41,360 --> 01:27:43,997
session even gets its own persistent

1546
01:27:44,080 --> 01:27:46,237
file system. What you're going to see in

1547
01:27:46,320 --> 01:27:48,477
a minute here is if I go to the files

1548
01:27:48,560 --> 01:27:51,957
tab to track the open items, the agent's

1549
01:27:52,040 --> 01:27:54,957
actually writing to a file.

1550
01:27:55,040 --> 01:27:56,597
Pretty cool, right?

1551
01:27:56,680 --> 01:28:00,597
Um plus, Foundry now has server-side

1552
01:28:00,680 --> 01:28:03,117
traces and built-in evals to show me

1553
01:28:03,200 --> 01:28:06,477
exactly what happened on every run.

1554
01:28:06,560 --> 01:28:09,117
But how do I know if my agent's doing a

1555
01:28:09,200 --> 01:28:12,357
good job? That's what Foundry's brand

1556
01:28:12,440 --> 01:28:17,317
rubric evaluators are for.
It's super simple. With just one AZD

1557
01:28:17,400 --> 01:28:20,957
command, Foundry reads my agent and then

1558
01:28:21,040 --> 01:28:23,317
generates the evaluation criteria for

1559
01:28:23,400 --> 01:28:26,437
me. In other words, it creates a rubric

1560
01:28:26,520 --> 01:28:30,797
personalized just to this agent.
Now let's check out this rubric in Foundry

1561
01:28:30,880 --> 01:28:35,997
portal. Look at these
dimensions. Ground uh

1562
01:28:36,080 --> 01:28:40,397
governance outcome correctness,
prescribe source usage. I didn't write

1563
01:28:40,480 --> 01:28:42,837
any of these. Foundry generated them

1564
01:28:42,920 --> 01:28:45,197
from production traces. And with this

1565
01:28:45,280 --> 01:28:48,157
rubric, I can score my agent in run

1566
01:28:48,240 --> 01:28:50,957
evals, but we can do so much more than

1567
01:28:51,040 --> 01:28:53,837
that. That's where Foundry's brand new

1568
01:28:53,920 --> 01:28:56,197
agent optimizer takes over.

1569
01:28:56,280 --> 01:28:58,077
Here's how it works. It tunes four

1570
01:28:58,160 --> 01:29:01,037
things: the model, instructions, tool

1571
01:29:01,120 --> 01:29:03,597
descriptions, and skills. And then it

1572
01:29:03,680 --> 01:29:08,317
generates improved candidates and scores
each one using the rubric I just showed

1573
01:29:08,400 --> 01:29:11,557
you. Here, I can view
the candidates. I can

1574
01:29:11,640 --> 01:29:16,477
see the strategy used. I can see their
scores. And I can actually view exactly

1575
01:29:16,560 --> 01:29:19,677
what changed. This candidate,
for example, improved

1576
01:29:19,760 --> 01:29:23,437
its score by updating
the model and the system prompt.

1577
01:29:23,520 --> 01:29:28,157
Foundry then makes it super easy to deploy
the best candidate as a brand new

1578
01:29:28,240 --> 01:29:30,197
agent version.

1579
01:29:30,280 --> 01:29:33,037
But this is not a one-time thing.

1580
01:29:33,120 --> 01:29:36,157
Every run feeds the next eval, and every

1581
01:29:36,240 --> 01:29:38,357
eval tells the optimizer where to

1582
01:29:38,440 --> 01:29:41,077
improve next. So your agents now get

1583
01:29:41,160 --> 01:29:43,277
better the more they're used.

1584
01:29:43,360 --> 01:29:46,477
But now let's put this agent to work.

1585
01:29:46,560 --> 01:29:49,117
I've published my agent to Teams and

1586
01:29:49,200 --> 01:29:52,357
M365 Copilot. And right now I'm going to

1587
01:29:52,440 --> 01:29:54,717
ask it to just catch me up on what I've

1588
01:29:54,800 --> 01:29:56,797
missed because I've been offline all

1589
01:29:56,880 --> 01:30:00,117
morning, but my agent hasn't.

1590
01:30:00,200 --> 01:30:02,197
While it works on that though, let me

1591
01:30:02,280 --> 01:30:04,877
explain what makes this agent different.

1592
01:30:04,960 --> 01:30:07,437
This is an autopilot agent, which means

1593
01:30:07,520 --> 01:30:09,837
it has its own identity and productivity

1594
01:30:09,920 --> 01:30:13,517
license, so it can work across M365 on

1595
01:30:13,600 --> 01:30:16,237
its own behalf. Earlier, you saw me

1596
01:30:16,320 --> 01:30:19,237
using this agent by myself, but shipping

1597
01:30:19,320 --> 01:30:23,997
a feature in a new release takes a whole
team, and that's why this agent lives in

1598
01:30:24,080 --> 01:30:28,277
our team's group chat. Now, let's
check out how it did. It summarized all

1599
01:30:28,360 --> 01:30:30,837
updates and called out the key features

1600
01:30:30,920 --> 01:30:32,917
I need to be tracking.

1601
01:30:33,000 --> 01:30:35,677
But, to put an agent like this actually

1602
01:30:35,760 --> 01:30:38,197
to work in your enterprise, governance

1603
01:30:38,280 --> 01:30:40,677
needs to come first. And that's why

1604
01:30:40,760 --> 01:30:43,277
every Autopilot agent requires admin

1605
01:30:43,360 --> 01:30:47,797
approval. Here, admins can review all
critical details, and they can even

1606
01:30:47,880 --> 01:30:52,557
choose who has the ability to talk
to the agent. But, it doesn't end there.

1607
01:30:52,640 --> 01:30:54,597
Even after approval,

1608
01:30:54,680 --> 01:30:56,637
uh admins can continue to monitor the

1609
01:30:56,720 --> 01:31:00,957
agent and block it at any time. But,

1610
01:31:01,040 --> 01:31:03,397
governance cannot apply to just one

1611
01:31:03,480 --> 01:31:06,437
agent. Every agent in your organization

1612
01:31:06,520 --> 01:31:08,917
needs to be managed with the same rigor

1613
01:31:09,000 --> 01:31:14,357
as users, apps, and devices. And that's
exactly what Microsoft Agent 365

1614
01:31:14,440 --> 01:31:18,397
provides. So, let's zoom out.
Today, we saw

1615
01:31:18,480 --> 01:31:22,517
Foundry accelerate development,
take your agent from local to

1616
01:31:22,600 --> 01:31:26,757
enterprise-ready,
and put it to work in M365.

1617
01:31:26,840 --> 01:31:30,957
Foundry makes it simple. You build the
agent, we handle the rest. Now, back to

1618
01:31:31,040 --> 01:31:37,520
you, Satya. >> [applause]
>> Thank you, Amanda.

1619
01:31:38,440 --> 01:31:43,157
You know, what you just saw
was how we're building security

1620
01:31:43,240 --> 01:31:46,957
for AI. Uh but, there's
also one other critical

1621
01:31:47,040 --> 01:31:49,237
aspect, especially, you know, the news

1622
01:31:49,320 --> 01:31:52,237
today is all about, you know, how do you

1623
01:31:52,320 --> 01:31:56,757
defend yourself using AI against attacks

1624
01:31:56,840 --> 01:31:59,117
that may, in fact, be using AI, right?

1625
01:31:59,200 --> 01:32:02,117
So, last month, we announced our

1626
01:32:02,200 --> 01:32:03,877
multi-model

1627
01:32:03,960 --> 01:32:07,077
agentic security system M-Dash. That's

1628
01:32:07,160 --> 01:32:09,957
essentially an agent harness for

1629
01:32:10,040 --> 01:32:12,797
security essentially that we built.

1630
01:32:12,880 --> 01:32:15,237
We're bringing together 100 agents

1631
01:32:15,320 --> 01:32:17,837
across the frontier

1632
01:32:17,920 --> 01:32:20,677
and custom models to really find these

1633
01:32:20,760 --> 01:32:25,397
exploitable bugs better than any single
model does. In fact, when we debuted

1634
01:32:25,480 --> 01:32:28,757
this harness,
it was on the top of Cyber Gym

1635
01:32:28,840 --> 01:32:31,237
benchmark. So, I want you to take a look

1636
01:32:31,320 --> 01:32:34,797
at what you can do with M-Dash to really

1637
01:32:34,880 --> 01:32:41,680
defend against AI attacks and defend the
entire digital estate. Over to you, Sarah.

1638
01:32:43,630 --> 01:32:45,650
>> [applause]

1639
01:32:46,360 --> 01:32:51,037
>> Thanks, Satya. Now, we all know that
security scans can take a while, so I'm

1640
01:32:51,120 --> 01:32:53,117
not going to do one live. So, let me

1641
01:32:53,200 --> 01:32:55,397
show you the results of an M-Dash scan I

1642
01:32:55,480 --> 01:32:58,357
already ran on my code base.

1643
01:32:58,440 --> 01:33:00,437
Now, the system runs as a standalone

1644
01:33:00,520 --> 01:33:03,277
CLI, but today I'm using it in my GitHub

1645
01:33:03,360 --> 01:33:06,717
Copilot app on my local dev machine.

1646
01:33:06,800 --> 01:33:09,357
The scan is broken down by vulnerability

1647
01:33:09,440 --> 01:33:14,277
domains and severity. And it's also

1648
01:33:14,360 --> 01:33:16,837
in additional to finding traditional

1649
01:33:16,920 --> 01:33:21,597
issues like coding errors and hardcoded
secrets, it's also identifying AI

1650
01:33:21,680 --> 01:33:23,837
specific vulnerabilities in the code

1651
01:33:23,920 --> 01:33:26,597
base. Now, what happened under the hood

1652
01:33:26,680 --> 01:33:29,077
is that over 100 specialized agents are

1653
01:33:29,160 --> 01:33:32,237
working together to discover, debate,

1654
01:33:32,320 --> 01:33:35,037
and prove exploitable vulnerabilities

1655
01:33:35,120 --> 01:33:37,637
end-to-end. And when the scan finishes,

1656
01:33:37,720 --> 01:33:42,597
it generates both a SARIF log
and an HTML report that I can give to my

1657
01:33:42,680 --> 01:33:46,917
management. Now,
the Defender details command allows

1658
01:33:47,000 --> 01:33:49,317
me to dig into these vulnerabilities,

1659
01:33:49,400 --> 01:33:52,477
and I can see what the vulnerability is,

1660
01:33:52,560 --> 01:33:55,357
where it is in my code, and the severity

1661
01:33:55,440 --> 01:33:58,197
to help me prioritize. And from here, of

1662
01:33:58,280 --> 01:34:01,797
course, we're going to fix it.

1663
01:34:01,880 --> 01:34:04,517
Now, using the Defender fix command,

1664
01:34:04,600 --> 01:34:09,037
the system will remediate suggested
fixes directly in my local dev

1665
01:34:09,120 --> 01:34:13,837
environment. And when that's done,
I can check out the diff. So, I have full

1666
01:34:13,920 --> 01:34:18,597
transparency about what the harness has
done, and I still have a human in the

1667
01:34:18,680 --> 01:34:22,157
loop check. But, of course,

1668
01:34:22,240 --> 01:34:23,957
everything I've shown you so far has run

1669
01:34:24,040 --> 01:34:27,797
locally, but I can also create a PR

1670
01:34:27,880 --> 01:34:30,277
to plug into my existing workflows and

1671
01:34:30,360 --> 01:34:32,797
push up to my repo. And I can take the

1672
01:34:32,880 --> 01:34:35,357
SARIF output from the scan, and I can

1673
01:34:35,440 --> 01:34:38,077
upload it to tools like GitHub Advanced

1674
01:34:38,160 --> 01:34:42,717
Security, and manage everything
alongside my other application security

1675
01:34:42,800 --> 01:34:46,597
findings. Now, I've shown
you m-dash working on my

1676
01:34:46,680 --> 01:34:48,877
code, but let me show you a

1677
01:34:48,960 --> 01:34:51,077
vulnerability our security research

1678
01:34:51,160 --> 01:34:54,197
teams identified using m-dash that you

1679
01:34:54,280 --> 01:34:56,157
can go and look up yourself.

1680
01:34:56,240 --> 01:34:58,357
So, the TLDR, because this is a lot to

1681
01:34:58,440 --> 01:35:02,237
read, of this bug is that Wasmtime reads

1682
01:35:02,320 --> 01:35:04,837
an out-of-date map

1683
01:35:04,920 --> 01:35:07,437
of an object, runs off the end, and then

1684
01:35:07,520 --> 01:35:10,677
it crashes the host. So, this is exactly

1685
01:35:10,760 --> 01:35:12,917
the kind of bug that the harness was

1686
01:35:13,000 --> 01:35:15,597
built for, because the flaw spread

1687
01:35:15,680 --> 01:35:17,717
across three different parts of the code

1688
01:35:17,800 --> 01:35:20,677
base. So, no single file looks wrong on

1689
01:35:20,760 --> 01:35:25,077
its own. They look absolutely fine, and

1690
01:35:25,160 --> 01:35:29,597
we can even see here, which is probably
my favorite part of this bug, a very

1691
01:35:29,680 --> 01:35:34,037
confident statement from the developers
claiming everything is fine.

1692
01:35:34,120 --> 01:35:35,837
Now, this is exactly the sort of

1693
01:35:35,920 --> 01:35:38,437
reassurance that fools normal scanners

1694
01:35:38,520 --> 01:35:41,437
and single AI models, but m-dash wasn't

1695
01:35:41,520 --> 01:35:43,677
fooled. One team of agents spotted the

1696
01:35:43,760 --> 01:35:46,397
suspicious gap, another team argued it

1697
01:35:46,480 --> 01:35:49,437
apart, and a third team built a working

1698
01:35:49,520 --> 01:35:53,717
example that actually triggered the crash.
And it did all of this in an

1699
01:35:53,800 --> 01:35:57,597
open-source code base. And this is
the kind of joined-up reasoning that

1700
01:35:57,680 --> 01:36:00,397
previously required significant manual

1701
01:36:00,480 --> 01:36:02,597
security research effort.

1702
01:36:02,680 --> 01:36:05,197
This is M-Dash helping developers create

1703
01:36:05,280 --> 01:36:08,077
secure code from the start. Coming soon

1704
01:36:08,160 --> 01:36:10,277
to your CLI and the Microsoft Defender

1705
01:36:10,360 --> 01:36:15,317
portal. Back to you, Satya. >> All right.

1706
01:36:15,400 --> 01:36:16,837
Thank you, Sarah.

1707
01:36:16,920 --> 01:36:20,357
So, that was the stack. Before though we

1708
01:36:20,440 --> 01:36:24,437
move to, you know,
unpacking more of the opportunity,

1709
01:36:24,520 --> 01:36:26,557
uh I want to do something different. Um

1710
01:36:26,640 --> 01:36:31,757
I want to introduce two people
whose LinkedIn uh profiles were both

1711
01:36:31,840 --> 01:36:33,317
super impressive and slightly

1712
01:36:33,400 --> 01:36:35,637
perplexing. Uh under current role it

1713
01:36:35,720 --> 01:36:38,837
says they're general partners at Mantis

1714
01:36:38,920 --> 01:36:41,557
VC. Under previous role it says they

1715
01:36:41,640 --> 01:36:43,837
sold out Madison Square Garden. Please

1716
01:36:43,920 --> 01:36:46,077
help me welcome Alex and Drew from The

1717
01:36:46,160 --> 01:36:51,280
Chainsmokers. Alex, Drew. >> [music]

1718
01:36:55,160 --> 01:36:57,197
[music] >> It's so
wonderful seeing you, Alex.

1719
01:36:57,280 --> 01:37:01,277
Thank you so much. You know, one of the
things when my team came to me and said,

1720
01:37:01,360 --> 01:37:04,637
"Hey, we're going to have uh
Chainsmokers at Build." I thought maybe

1721
01:37:04,720 --> 01:37:09,757
that's what we're calling our new GitHub
Copilot app. Uh but, you know, it's so I

1722
01:37:09,840 --> 01:37:13,877
thought maybe they're your fans.
But, tell tell me about how you got into

1723
01:37:13,960 --> 01:37:18,317
this. I mean, you've been at it
now for what? 12-plus years

1724
01:37:18,400 --> 01:37:21,757
uh as venture capitalists or angel
investors first and you have your firm

1725
01:37:21,840 --> 01:37:24,877
now for 7-plus years and you even picked

1726
01:37:24,960 --> 01:37:28,957
what I would have not thought is the
natural place, which is B2B SaaS even as

1727
01:37:29,040 --> 01:37:32,677
the first place. We just give us
a little bit about the back story on this.

1728
01:37:32,760 --> 01:37:36,437
>> Sure. Well, uh hey, I'm sure you guys
are wondering what timeline you're on

1729
01:37:36,520 --> 01:37:41,577
where The Chainsmokers are at Microsoft
Build, but hey, how are how are you? Um

1730
01:37:41,660 --> 01:37:43,117
>> [laughter] >> And uh you
know, we've been making we've

1731
01:37:43,200 --> 01:37:45,677
been in The Chainsmokers for 14 years

1732
01:37:45,760 --> 01:37:47,637
and um you know, when our our music

1733
01:37:47,720 --> 01:37:50,437
started to take off in the mid-2010s, we

1734
01:37:50,520 --> 01:37:53,477
got to play a few events like this

1735
01:37:53,560 --> 01:37:57,997
and we met a lot of the founders
from the consumer mobile cloud era of

1736
01:37:58,080 --> 01:38:02,557
startups and they taught us a lot and gave
us a front row seat into what early

1737
01:38:02,640 --> 01:38:08,317
stage investing was like.
We fell in love with it. We found a lot of

1738
01:38:08,400 --> 01:38:10,997
we had a lot in common with these
founders and the way they built their

1739
01:38:11,080 --> 01:38:15,157
business was the same way that we
thought about starting the Chainsmokers

1740
01:38:15,240 --> 01:38:19,157
and and you know, kind of breaking
through the noise there and you know, we

1741
01:38:19,240 --> 01:38:22,637
got to participate in a few of their
deals. We were very lucky and we decided

1742
01:38:22,720 --> 01:38:25,277
to institutionalize in 2020
and start our own fund.

1743
01:38:25,360 --> 01:38:29,677
>> So it's so cool and and when you sort
of hear about all this AI stuff and what's

1744
01:38:29,760 --> 01:38:33,037
happening. Obviously there's a lot even
going on even in your portfolio. We were

1745
01:38:33,120 --> 01:38:35,277
talking backstage, things are changing.

1746
01:38:35,360 --> 01:38:38,357
How do you sort of see
the opportunity going forward?

1747
01:38:38,440 --> 01:38:40,597
>> I mean there's so
many vectors you could

1748
01:38:40,680 --> 01:38:44,317
talk about this on. Of course there's
the creative output side which you know,

1749
01:38:44,400 --> 01:38:47,677
we've been experimenting with in music
but always important to have your

1750
01:38:47,760 --> 01:38:51,717
authenticity and when it comes
to creativity. But on the investment side,

1751
01:38:51,800 --> 01:38:55,477
I think we're moving from producing
outputs to producing actions which I

1752
01:38:55,560 --> 01:38:59,117
think presents like a very interesting
opportunity to kind of reimagine the

1753
01:38:59,200 --> 01:39:03,397
entire architecture of the way software
enterprises been built. So instead of

1754
01:39:03,480 --> 01:39:06,957
humans, you know, producing outputs,
it's machines producing outputs and

1755
01:39:07,040 --> 01:39:10,957
rethinking what that entire
space looks like in that context.

1756
01:39:11,040 --> 01:39:13,717
>> Yeah, I know it's fantastic to see that

1757
01:39:13,800 --> 01:39:16,357
and maybe to just to close out any sort

1758
01:39:16,440 --> 01:39:19,637
of when as artists who have had you

1759
01:39:19,720 --> 01:39:23,757
know, great success,
when you look at a founder,

1760
01:39:23,840 --> 01:39:28,077
what sort of is the advice you give them
when you're creating, you know, you're

1761
01:39:28,160 --> 01:39:32,157
bringing something new and it's
a creative process. So what is

1762
01:39:32,240 --> 01:39:35,557
sort of what are you looking for in
founders? What is it that you're giving

1763
01:39:35,640 --> 01:39:38,397
them as advice? >> Totally.
I mean there really are so many

1764
01:39:38,480 --> 01:39:39,997
parallels. Maybe it seems like that.

1765
01:39:40,080 --> 01:39:43,477
Maybe it doesn't but I mean
it's really difficult as artists

1766
01:39:43,560 --> 01:39:46,277
to really find, I guess, what your sound

1767
01:39:46,360 --> 01:39:50,557
is, what's authentically you,
what you can continue to do over and

1768
01:39:50,640 --> 01:39:54,717
over because with this much competition,
you have to be moving authentically, you

1769
01:39:54,800 --> 01:39:59,157
have to be connected to the product that
you're creating, and cuz you're going to

1770
01:39:59,240 --> 01:40:02,357
have to iterate and keep it going
and have consistency for a long period of

1771
01:40:02,440 --> 01:40:06,077
time to kind of lock into your fan base
and make something that's special and

1772
01:40:06,160 --> 01:40:09,917
unique. And, you know, we try to advise
our founders to do the same thing.

1773
01:40:10,000 --> 01:40:12,077
>> Fantastic. Thank you both for joining

1774
01:40:12,160 --> 01:40:17,037
us. Are you game for playing for us
this evening for our closest friends here?

1775
01:40:17,120 --> 01:40:20,087
>> I will be there.
>> 6:00 6:00 this evening.

1776
01:40:20,170 --> 01:40:22,077
>> [applause] >> Thank
you so much. Thank you.

1777
01:40:22,160 --> 01:40:25,040
>> Thank you. Thank you.

1778
01:40:27,000 --> 01:40:29,957
>> Back We'll be right here at 6:00 p.m.

1779
01:40:30,040 --> 01:40:32,237
tonight, and really looking forward to

1780
01:40:32,320 --> 01:40:36,677
it. So, so that effectively sort of

1781
01:40:36,760 --> 01:40:38,277
talks a little bit about the developer

1782
01:40:38,360 --> 01:40:41,077
stack. So, now let's talk about

1783
01:40:41,160 --> 01:40:44,437
what is opportunity for every company,

1784
01:40:44,520 --> 01:40:48,997
right? At the end of the day, we are
institution and organizational builders.

1785
01:40:49,080 --> 01:40:52,917
How if whether it's an AI native
company, whether it's a SaaS company,

1786
01:40:53,000 --> 01:40:55,397
whether it's an enterprise, the first

1787
01:40:55,480 --> 01:40:58,357
thing we want to do is make sure that as

1788
01:40:58,440 --> 01:41:01,957
you build things like plugins or agents

1789
01:41:02,040 --> 01:41:06,757
or your AI apps, we want to make sure
that they're discovered throughout the

1790
01:41:06,840 --> 01:41:10,557
Microsoft ecosystem.
That's job number one for us. That's

1791
01:41:10,640 --> 01:41:12,317
what we why we're doing the things we're

1792
01:41:12,400 --> 01:41:15,517
doing in Windows or the Microsoft 365

1793
01:41:15,600 --> 01:41:18,357
Copilot or in Teams or in GitHub. We

1794
01:41:18,440 --> 01:41:23,677
want to structure them such that your
applications, your agents, your plugins

1795
01:41:23,760 --> 01:41:27,677
are discovered. Customers
are also building lots of line

1796
01:41:27,760 --> 01:41:29,797
of business applications, line of

1797
01:41:29,880 --> 01:41:34,717
business agents then using Copilot
Studio, and we want to make that all

1798
01:41:34,800 --> 01:41:36,397
discoverable again as part of the

1799
01:41:36,480 --> 01:41:38,757
Copilot experience. And Teams in some

1800
01:41:38,840 --> 01:41:41,997
sense has become this destination

1801
01:41:42,080 --> 01:41:44,517
for multiplayer human-to-agent

1802
01:41:44,600 --> 01:41:47,237
interaction. We want you to be able to

1803
01:41:47,320 --> 01:41:52,597
find agents, interact with agents
uh right in Teams. Um and we're

1804
01:41:52,680 --> 01:41:54,237
supercharging

1805
01:41:54,320 --> 01:41:57,517
all of this. Uh Copilot continues to

1806
01:41:57,600 --> 01:42:00,677
evolve very quickly. It started first

1807
01:42:00,760 --> 01:42:03,357
with chat with some of the best models,

1808
01:42:03,440 --> 01:42:08,317
with great access to Work IQ. We didn't
have the name Work IQ, but now uh you

1809
01:42:08,400 --> 01:42:09,877
that's kind of where it got started.

1810
01:42:09,960 --> 01:42:12,717
Then came Co-work, uh a new way of

1811
01:42:12,800 --> 01:42:15,557
working and generating these stunning

1812
01:42:15,640 --> 01:42:19,077
artifacts and solve these multi-step

1813
01:42:19,160 --> 01:42:21,117
problems, right? So, you assign a

1814
01:42:21,200 --> 01:42:24,237
multi-step task uh to Co-work. Uh you

1815
01:42:24,320 --> 01:42:28,357
saw GitHub.
GitHub is continuing to evolve as well.

1816
01:42:28,440 --> 01:42:30,677
Um and now as of, you know, the come

1817
01:42:30,760 --> 01:42:31,957
summer,

1818
01:42:32,040 --> 01:42:35,557
uh you will be bringing coding to all

1819
01:42:35,640 --> 01:42:39,157
knowledge work within one Copilot super

1820
01:42:39,240 --> 01:42:44,197
app, right? That's going to be
really exciting to see. Yeah.

1821
01:42:44,280 --> 01:42:44,917
>> [applause]

1822
01:42:45,000 --> 01:42:48,197
>> So, you'll have chat, Co-work, and code

1823
01:42:48,280 --> 01:42:51,517
all in Copilot. But today, we're

1824
01:42:51,600 --> 01:42:55,837
introducing something completely new.

1825
01:42:55,920 --> 01:43:00,597
Autopilots. Uh we can
think of Autopilots as

1826
01:43:00,680 --> 01:43:03,557
enterprise-grade claws.

1827
01:43:03,640 --> 01:43:07,157
These are autonomous long-running agents

1828
01:43:07,240 --> 01:43:10,557
with full enterprise compliance that run

1829
01:43:10,640 --> 01:43:14,077
in your tenant. Uh Autopilots can have a

1830
01:43:14,160 --> 01:43:17,997
name, personality, custom connectors,

1831
01:43:18,080 --> 01:43:20,957
context, and memory. And they're totally

1832
01:43:21,040 --> 01:43:23,677
new way to reduce toil and get you back

1833
01:43:23,760 --> 01:43:26,717
to what you love. Um to kick things off,

1834
01:43:26,800 --> 01:43:30,037
the first Autopilot we're introducing is

1835
01:43:30,120 --> 01:43:33,880
Scout. Let's take a look.

1836
01:43:34,680 --> 01:43:38,700
All right. >> [applause]

1837
01:43:39,400 --> 01:43:41,757
>> As you can see, Scout works where you

1838
01:43:41,840 --> 01:43:47,157
work, uh joining group chats in Teams,
handling threads in Outlook, uh starting

1839
01:43:47,240 --> 01:43:51,637
today for those of you who are on Copilot
Frontier, you can try out Scout,

1840
01:43:51,720 --> 01:43:53,837
and in the coming months, we will build

1841
01:43:53,920 --> 01:43:59,117
this out uh to a complete digital team
of autopilots right inside a Copilot,

1842
01:43:59,200 --> 01:44:02,597
right? So, you can go to the Copilot app.
Scout is the one that comes by

1843
01:44:02,680 --> 01:44:07,117
default, but you can build
more of these autopilots. Uh

1844
01:44:07,200 --> 01:44:09,557
And so, that's the future of what we

1845
01:44:09,640 --> 01:44:11,717
think of as the Copilot ecosystem

1846
01:44:11,800 --> 01:44:16,797
itself. And so far, we have talked a lot
about what you can do to build your own

1847
01:44:16,880 --> 01:44:20,597
agents and how these agents are
discovered in things like Copilot and

1848
01:44:20,680 --> 01:44:25,277
Teams and Windows. But when
you think about what makes any

1849
01:44:25,360 --> 01:44:27,437
organization, any enterprise, any

1850
01:44:27,520 --> 01:44:29,797
company unique,

1851
01:44:29,880 --> 01:44:35,037
it is its tacit knowledge that it's
continuously compounding through its

1852
01:44:35,120 --> 01:44:39,037
operations. So, the key
consideration at some level

1853
01:44:39,120 --> 01:44:41,437
in an AI age is to ask the question,

1854
01:44:41,520 --> 01:44:43,357
what's the future of the firm?

1855
01:44:43,440 --> 01:44:46,557
Uh how do you continue to preserve and

1856
01:44:46,640 --> 01:44:50,437
compound that tacit knowledge in the age

1857
01:44:50,520 --> 01:44:54,677
of AI where models can learn anything
from the data and the trajectories they

1858
01:44:54,760 --> 01:44:59,317
see? Uh to do that, we believe that every

1859
01:44:59,400 --> 01:45:02,757
organization, whether
it's an AI-native company,

1860
01:45:02,840 --> 01:45:05,397
a SaaS company, or any enterprise, will

1861
01:45:05,480 --> 01:45:08,797
need to build their own hill-climbing

1862
01:45:08,880 --> 01:45:10,037
machine.

1863
01:45:10,120 --> 01:45:13,797
It's a system that continuously improves

1864
01:45:13,880 --> 01:45:17,237
against your objectives, your private

1865
01:45:17,320 --> 01:45:21,717
evals, compounding
your advantage over time,

1866
01:45:21,800 --> 01:45:23,477
not someone else's.

1867
01:45:23,560 --> 01:45:26,117
Uh to deliver on this, we're taking the

1868
01:45:26,200 --> 01:45:29,117
next step with frontier tuning. Uh we

1869
01:45:29,200 --> 01:45:33,477
have been working with many customers
already to help with, you know, really

1870
01:45:33,560 --> 01:45:37,237
build their own uh learning
loop, uh the environment,

1871
01:45:37,320 --> 01:45:40,277
the context, the tools, the rubrics, and

1872
01:45:40,360 --> 01:45:42,437
even train their own models.

1873
01:45:42,520 --> 01:45:45,477
And today really we're really thrilled

1874
01:45:45,560 --> 01:45:50,517
and excited about supercharging that
frontier tuning capability uh with the

1875
01:45:50,600 --> 01:45:54,717
innovation that's coming out of our super
intelligence lab. To tell you all

1876
01:45:54,800 --> 01:45:59,720
about this, let me hand it
over to Mustafa. Mustafa?

1877
01:46:00,160 --> 01:46:01,957
>> Thank you.

1878
01:46:02,040 --> 01:46:04,877
Thank you. Thank you.

1879
01:46:04,960 --> 01:46:08,197
Thank you. Good morning, everybody. Um

1880
01:46:08,280 --> 01:46:10,477
you know, we really are living in the

1881
01:46:10,560 --> 01:46:13,677
most remarkable times. Since I started

1882
01:46:13,760 --> 01:46:16,277
working in AI, the compute that we use

1883
01:46:16,360 --> 01:46:19,437
to train frontier models has increased

1884
01:46:19,520 --> 01:46:22,437
by one trillionfold.

1885
01:46:22,520 --> 01:46:24,357
That's 12 orders of magnitude of

1886
01:46:24,440 --> 01:46:28,317
computation in just 15 years.

1887
01:46:28,400 --> 01:46:30,677
It's now clear that a consistent

1888
01:46:30,760 --> 01:46:33,477
exponential increase in computation

1889
01:46:33,560 --> 01:46:38,717
leads to predictable
advances in AI capabilities.

1890
01:46:38,800 --> 01:46:42,757
And in the next few years, we're going
to see three more orders of magnitude of

1891
01:46:42,840 --> 01:46:46,717
compute applied to train frontier models.

1892
01:46:46,800 --> 01:46:48,677
Intelligence is now a function of

1893
01:46:48,760 --> 01:46:51,397
compute. Log-linear hill climbing has

1894
01:46:51,480 --> 01:46:54,077
become the norm. The scaling laws are

1895
01:46:54,160 --> 01:46:56,757
clearly holding, and it is a remarkable

1896
01:46:56,840 --> 01:46:59,637
time in our industry.

1897
01:46:59,720 --> 01:47:03,117
And so, in this context, we at MAI are

1898
01:47:03,200 --> 01:47:07,317
building towards what we call
humanist superintelligence.

1899
01:47:07,400 --> 01:47:10,077
State-of-the-art AI capabilities that

1900
01:47:10,160 --> 01:47:12,837
are explicitly designed to serve people

1901
01:47:12,920 --> 01:47:16,677
and organizations and not replace them.

1902
01:47:16,760 --> 01:47:19,637
Because the type of AI that we create

1903
01:47:19,720 --> 01:47:24,477
really does matter.
We need an AI that places humanity

1904
01:47:24,560 --> 01:47:26,037
first,

1905
01:47:26,120 --> 01:47:28,677
that always prioritizes human [snorts]

1906
01:47:28,760 --> 01:47:32,037
well-being and human progress.

1907
01:47:32,120 --> 01:47:36,557
This is the core philosophy and
motivation behind our superintelligence

1908
01:47:36,640 --> 01:47:38,957
efforts at Microsoft, and it shapes

1909
01:47:39,040 --> 01:47:41,197
everything that we do.

1910
01:47:41,280 --> 01:47:43,797
And as a platform company, our job and

1911
01:47:43,880 --> 01:47:46,517
our commitment is to keep you developers

1912
01:47:46,600 --> 01:47:49,437
building at the absolute frontier.

1913
01:47:49,520 --> 01:47:51,837
So today, we are very excited to

1914
01:47:51,920 --> 01:47:57,477
announce a family of seven new models
across image, voice, transcription, and

1915
01:47:57,560 --> 01:48:01,350
coding. >> [cheering] [applause]

1916
01:48:03,640 --> 01:48:05,677
>> These are all built with real attention

1917
01:48:05,760 --> 01:48:07,717
to detail and a commitment to making

1918
01:48:07,800 --> 01:48:12,757
very practical and efficient tools
that are tuned to just how you work in the

1919
01:48:12,840 --> 01:48:14,077
real world.

1920
01:48:14,160 --> 01:48:17,717
So first up, MAI image 2.5 and its flash

1921
01:48:17,800 --> 01:48:20,117
variant, two super strong models that

1922
01:48:20,200 --> 01:48:23,677
deliver a step change in quality, now at

1923
01:48:23,760 --> 01:48:27,797
number two on the leaderboard surpassing
the score of Nana Banana 2 on image

1924
01:48:27,880 --> 01:48:31,490
editing. >> [applause]

1925
01:48:32,560 --> 01:48:37,357
>> They give you precise editing
with incredible control and consistency.

1926
01:48:37,440 --> 01:48:41,997
Flash is here for super efficient
production workloads, while 2.5 gives

1927
01:48:42,080 --> 01:48:46,237
you that maximum fidelity and
professional-grade performance. They're

1928
01:48:46,320 --> 01:48:49,837
live in PowerPoint today, they're
rolling out to OneDrive, and right now

1929
01:48:49,920 --> 01:48:51,757
you can access them on Foundry at a

1930
01:48:51,840 --> 01:48:54,717
market-leading quality per dollar.

1931
01:48:54,800 --> 01:48:59,037
Next up, we've got MAI transcribe 1.5.

1932
01:48:59,120 --> 01:49:01,557
This is the best transcription model in

1933
01:49:01,640 --> 01:49:04,037
the world. State-of-the-art accuracy

1934
01:49:04,120 --> 01:49:09,437
across 43 languages, beating out Gemini
and OpenAI's flagship transcription

1935
01:49:09,520 --> 01:49:12,797
models. We've optimized
it for real-world use,

1936
01:49:12,880 --> 01:49:14,957
so that you can produce highly accurate

1937
01:49:15,040 --> 01:49:18,037
transcripts for any bespoke use case

1938
01:49:18,120 --> 01:49:22,717
five times faster than all rival models.

1939
01:49:22,800 --> 01:49:24,237
It's now being integrated inside of

1940
01:49:24,320 --> 01:49:27,637
GitHub Teams Copilot, Dynamics 365

1941
01:49:27,720 --> 01:49:31,357
contact center, and it's now also
available in Foundry where I'm very

1942
01:49:31,440 --> 01:49:36,117
excited to say it is the fastest,
most efficient, and most cost-effective

1943
01:49:36,200 --> 01:49:40,117
transcription model of any
of the hyperscalers out there.

1944
01:49:40,200 --> 01:49:42,880
>> [cheering] [applause]

1945
01:49:45,000 --> 01:49:48,837
>> So, paired with that, we've got
MAI Voice 2. This is our latest speech

1946
01:49:48,920 --> 01:49:53,557
generation model. It has beautiful
prosody, natural-sounding delivery,

1947
01:49:53,640 --> 01:49:58,037
fine-grained emotional control, and it's
available in 15 languages with many more

1948
01:49:58,120 --> 01:50:02,517
coming soon. We're also
announcing Voice 2 Flash, and

1949
01:50:02,600 --> 01:50:07,117
that provides the very best value and
speed for ultra-latency sensitive voice

1950
01:50:07,200 --> 01:50:08,997
agents, which of course is the big thing

1951
01:50:09,080 --> 01:50:11,317
in 2026.

1952
01:50:11,400 --> 01:50:14,197
Next up, our text foundation model,

1953
01:50:14,280 --> 01:50:16,957
MAI Thinking 1. This is our first

1954
01:50:17,040 --> 01:50:19,077
reasoning model, and it's exceptionally

1955
01:50:19,160 --> 01:50:21,557
strong in our target use cases of

1956
01:50:21,640 --> 01:50:24,917
reasoning and SWE tasks. It's a 35

1957
01:50:25,000 --> 01:50:30,077
billion active parameter MOE with a 256K
context window. That means that it

1958
01:50:30,160 --> 01:50:33,077
competes in the medium-sized weight
class where it's certainly punching

1959
01:50:33,160 --> 01:50:36,997
above its weight,
and independent human raters on Surge

1960
01:50:37,080 --> 01:50:39,037
prefer it in overall quality

1961
01:50:39,120 --> 01:50:44,317
side-by-sides versus Sonnet 4.6.

1962
01:50:44,400 --> 01:50:45,597
>> [applause]

1963
01:50:45,680 --> 01:50:50,197
>> It's achieved 95
97% on AMIE 2025, which

1964
01:50:50,280 --> 01:50:51,517
is obviously the key measure of its

1965
01:50:51,600 --> 01:50:53,837
general-purpose reasoning abilities, but

1966
01:50:53,920 --> 01:50:57,837
most importantly of all, it's now at 53%

1967
01:50:57,920 --> 01:51:00,197
on SWE Bench Pro, which places it right

1968
01:51:00,280 --> 01:51:03,077
alongside Opus 4.6, at least on the

1969
01:51:03,160 --> 01:51:07,897
toughest coding benchmark that's out
there. So, we're very happy with that.

1970
01:51:07,980 --> 01:51:12,357
>> [cheering and applause] >> Now,
there's plenty more for us to do as

1971
01:51:12,440 --> 01:51:16,477
we get this into production and hill
climb against real world uh tasks and

1972
01:51:16,560 --> 01:51:19,557
real world traffic. But,
it's what is actually most

1973
01:51:19,640 --> 01:51:24,597
remarkable about this model we think
is that it has climbed entirely from the

1974
01:51:24,680 --> 01:51:27,517
bottom. And that means
that it hasn't targeted

1975
01:51:27,600 --> 01:51:31,877
any of the benchmarks specifically,
and it's done so with absolutely zero

1976
01:51:31,960 --> 01:51:36,117
distillation. And to us, this is critical.
Because it means that the

1977
01:51:36,200 --> 01:51:37,917
model is created with an

1978
01:51:38,000 --> 01:51:43,157
enterprise-grade, clean, and commercially
licensed data lineage. That

1979
01:51:43,240 --> 01:51:46,677
means that you can put it into
production in a very trustworthy way

1980
01:51:46,760 --> 01:51:49,117
with complete confidence.

1981
01:51:49,200 --> 01:51:51,157
Now, finally, I'm incredibly excited to

1982
01:51:51,240 --> 01:51:56,157
announce MAI Code 1 Flash.
This is our new inference efficient

1983
01:51:56,240 --> 01:52:00,077
coding model, which has been especially
tuned for VS Code and of course GitHub

1984
01:52:00,160 --> 01:52:01,957
Copilot CLI.

1985
01:52:02,040 --> 01:52:05,397
It achieves 51% on SWE-Bench Pro despite

1986
01:52:05,480 --> 01:52:08,197
having just 5 billion parameters.

1987
01:52:08,280 --> 01:52:13,277
And so, it's much closer to I Q in terms
of size, but cheaper in cost, delivering

1988
01:52:13,360 --> 01:52:15,197
really strong coding performance at

1989
01:52:15,280 --> 01:52:21,257
great inference efficiency. And it's
rolling out today inside of VS Code.

1990
01:52:21,340 --> 01:52:23,360
>> [applause]

1991
01:52:25,080 --> 01:52:29,157
>> Now, alongside distribution on Foundry
and optimization for our 1P products,

1992
01:52:29,240 --> 01:52:33,757
we're also very excited to make our
models available on Open Router as well

1993
01:52:33,840 --> 01:52:37,117
as Fireworks and Base Ten. This means
that for the first time, you're going to

1994
01:52:37,200 --> 01:52:38,797
be able to tune the weights directly

1995
01:52:38,880 --> 01:52:42,437
yourself in an ecosystem of your choice.

1996
01:52:42,520 --> 01:52:44,997
Now, across this entire family, safety

1997
01:52:45,080 --> 01:52:49,757
and security have been built in from
the start. Our voice models come with

1998
01:52:49,840 --> 01:52:53,757
protections against unauthorized cloning.
Everything is watermarked from

1999
01:52:53,840 --> 01:52:58,117
scratch. We've reduced our over
refusals, improved representation,

2000
01:52:58,200 --> 01:53:00,437
including for people with disabilities.

2001
01:53:00,520 --> 01:53:02,797
We're also publishing a very detailed

2002
01:53:02,880 --> 01:53:04,997
technical report today to give you to

2003
01:53:05,080 --> 01:53:10,397
full and transparent understanding
of how we put all of this together.

2004
01:53:10,480 --> 01:53:13,957
Now, one of the things I'm particularly
excited about is that we have been

2005
01:53:14,040 --> 01:53:16,557
carefully co-designing our models with

2006
01:53:16,640 --> 01:53:18,837
our own silicon. So, that means that

2007
01:53:18,920 --> 01:53:21,517
we've optimized MAI thinking one on our

2008
01:53:21,600 --> 01:53:24,597
very own Maia 200 chip and benchmarked

2009
01:53:24,680 --> 01:53:28,157
it head-to-head against the GB200.

2010
01:53:28,240 --> 01:53:30,757
And so, on top of the 30% performance

2011
01:53:30,840 --> 01:53:32,517
improvement that Satya talked about

2012
01:53:32,600 --> 01:53:36,477
earlier, we're now seeing a further 1.4x

2013
01:53:36,560 --> 01:53:38,877
performance per watt gain when we run

2014
01:53:38,960 --> 01:53:41,997
our MAI models on the Maia 200

2015
01:53:42,080 --> 01:53:44,557
end-to-end. And that's huge. Cuz as

2016
01:53:44,640 --> 01:53:47,277
everybody knows, at this scale, every

2017
01:53:47,360 --> 01:53:51,717
watt counts. And silicon
and model co-design is a

2018
01:53:51,800 --> 01:53:56,077
really key advantage that we think
is going to help keep everybody here right

2019
01:53:56,160 --> 01:54:00,117
on the frontier with the most efficient
and most powerful thinking and coding

2020
01:54:00,200 --> 01:54:04,077
agents out there.
We're also super excited that these

2021
01:54:04,160 --> 01:54:06,317
faster and more efficient MAI models are

2022
01:54:06,400 --> 01:54:11,157
coming to the N1X that Satya mentioned
a few moments ago. And we think that's

2023
01:54:11,240 --> 01:54:15,197
going to be able to deliver the very
best performance uh on Windows in a few

2024
01:54:15,280 --> 01:54:16,997
months' time.

2025
01:54:17,080 --> 01:54:19,797
Now, to us, this is what owning the full

2026
01:54:19,880 --> 01:54:22,237
stack end-to-end looks like. It's the

2027
01:54:22,320 --> 01:54:25,277
foundation of Microsoft frontier tuning,

2028
01:54:25,360 --> 01:54:29,437
lets you customize the MAI models using
our full stack hill climbing machine

2029
01:54:29,520 --> 01:54:33,797
right where you want it. And it means
that the disciplined and very relentless

2030
01:54:33,880 --> 01:54:38,197
engineering that is going into building
our models is now available to all of

2031
01:54:38,280 --> 01:54:40,637
you on a platform that you can trust,

2032
01:54:40,720 --> 01:54:43,397
working on your behalf to create custom

2033
01:54:43,480 --> 01:54:48,477
agents that you will control.
So, the really big thing, of course,

2034
01:54:48,560 --> 01:54:51,597
that's happened in the last year
is these ARELEs, reinforcement learning

2035
01:54:51,680 --> 01:54:55,877
environments, these unique
training gyms for your AIs.

2036
01:54:55,960 --> 01:54:58,397
They create company and task-specific

2037
01:54:58,480 --> 01:55:02,997
agents adapted only to you
built on our MAI models.

2038
01:55:03,080 --> 01:55:07,197
So, for example, within Microsoft,
we use our RLEs combined with our MAI

2039
01:55:07,280 --> 01:55:09,637
models to climb towards the best agentic

2040
01:55:09,720 --> 01:55:11,957
use cases on Excel.

2041
01:55:12,040 --> 01:55:14,317
Our MAI tuned model is now on par with

2042
01:55:14,400 --> 01:55:16,957
GPT-5.4 on public and private

2043
01:55:17,040 --> 01:55:19,277
benchmarks, whilst at the same time

2044
01:55:19,360 --> 01:55:24,237
being 10 times more efficient on cost.

2045
01:55:24,320 --> 01:55:29,757
Thank you. You know, and many other
early adopters are seeing similar results.

2046
01:55:29,840 --> 01:55:32,837
When we've tuned our models
on McKinsey's tasks, MAI delivered the

2047
01:55:32,920 --> 01:55:36,837
highest win rate,
even outperforming GPT-5.5,

2048
01:55:36,920 --> 01:55:39,437
and again delivering 10x greater

2049
01:55:39,520 --> 01:55:41,317
efficiency on cost.

2050
01:55:41,400 --> 01:55:44,117
So, to us, this is the advantage of very

2051
01:55:44,200 --> 01:55:47,677
carefully calibrated frontier tuning.

2052
01:55:47,760 --> 01:55:51,437
And importantly, unlike with
some of the other companies,

2053
01:55:51,520 --> 01:55:56,157
with MAI, you don't rent intelligence
from a shared model that learns from

2054
01:55:56,240 --> 01:55:59,557
everybody. Only you
keep the benefits of your

2055
01:55:59,640 --> 01:56:05,317
hard-earned workflows, know-how,
knowledge, and your own institutional

2056
01:56:05,400 --> 01:56:08,397
data. Only you get
to control the resulting

2057
01:56:08,480 --> 01:56:11,677
model. And so, with us, the RLEs and the

2058
01:56:11,760 --> 01:56:16,117
models that you build inside
of them, they become your moat.

2059
01:56:16,200 --> 01:56:20,717
I really think this is distinct.
It marks a new era in AI that we're all

2060
01:56:20,800 --> 01:56:24,117
very, very excited about. Okay.

2061
01:56:24,200 --> 01:56:26,557
So, now just one final announcement that

2062
01:56:26,640 --> 01:56:28,797
I'm very excited about. We are taking

2063
01:56:28,880 --> 01:56:33,797
customization and co-creation of our
models to the highest level possible on

2064
01:56:33,880 --> 01:56:35,717
what I think of as perhaps the most

2065
01:56:35,800 --> 01:56:39,477
important application of AI, healthcare.

2066
01:56:39,560 --> 01:56:42,557
So, today, we're very proud to be
announcing that we're partnering with

2067
01:56:42,640 --> 01:56:46,957
Mayo Clinic to jointly develop a new
frontier model for health, and then

2068
01:56:47,040 --> 01:56:50,797
deploy it around the world in their
hospitals and beyond. So, this morning,

2069
01:56:50,880 --> 01:56:52,917
please help me in welcoming to the stage

2070
01:56:53,000 --> 01:56:54,957
a physician, groundbreaking researcher,

2071
01:56:55,040 --> 01:57:01,960
president, and CEO of the Mayo
Clinic, Dr. Gianrico Farrugia.

2072
01:57:02,340 --> 01:57:05,010
>> [applause] [music]

2073
01:57:06,320 --> 01:57:08,717
>> Thank you so much
for being here, Gianrico.

2074
01:57:08,800 --> 01:57:13,197
Um now, of course, everyone will
recognize Mayo as perhaps the leading

2075
01:57:13,280 --> 01:57:17,917
hospital in the world with an incredible
track record of research and innovation

2076
01:57:18,000 --> 01:57:21,757
and clinical practice. Tell us a little
bit more about what you hope to get out

2077
01:57:21,840 --> 01:57:23,997
of our collaboration. >> Well,
first of all, thanks for having me

2078
01:57:24,080 --> 01:57:28,437
here. Thanks to Statis Mayo
Clinic is known for being able to

2079
01:57:28,520 --> 01:57:33,077
live up to our primary value. The needs
of the patient come first. We deliver

2080
01:57:33,160 --> 01:57:37,157
outstanding health care. We're ranked
the number one health care organization

2081
01:57:37,240 --> 01:57:40,797
in the world. Yet,
we know most people in the world

2082
01:57:40,880 --> 01:57:43,117
will not have access to Mayo Clinic.

2083
01:57:43,200 --> 01:57:47,717
So, 7 years ago, we decided to create
a platform, the Mayo Clinic platform,

2084
01:57:47,800 --> 01:57:49,277
moving all of health care from a

2085
01:57:49,360 --> 01:57:51,837
pipeline to a platform.

2086
01:57:51,920 --> 01:57:54,837
And with our partners, that platform now

2087
01:57:54,920 --> 01:57:57,317
is in four continents and reaches about

2088
01:57:57,400 --> 01:58:02,157
100 million people.
It has created the largest, to our

2089
01:58:02,240 --> 01:58:06,277
knowledge, deepest
longitudinal health care data

2090
01:58:06,360 --> 01:58:10,957
set in the world, multimodal,
gen- including genomics. So, here together

2091
01:58:11,040 --> 01:58:14,677
now we have the opportunity to do what
we do best together, which is create a

2092
01:58:14,760 --> 01:58:16,877
frontier model for health care.

2093
01:58:16,960 --> 01:58:20,837
What it means if you're a patient,
if you're somebody interested in health

2094
01:58:20,920 --> 01:58:25,437
care, you can get clinical and logistical
answers to your health care.

2095
01:58:25,520 --> 01:58:28,197
But, if you're health care
provider, if you're a physician,

2096
01:58:28,280 --> 01:58:30,837
it can give you insights, can act as

2097
01:58:30,920 --> 01:58:33,797
your real-time team member that can tell

2098
01:58:33,880 --> 01:58:37,397
you what is likely to happen next, but

2099
01:58:37,480 --> 01:58:39,277
it can also prevent harm and therefore

2100
01:58:39,360 --> 01:58:44,477
increase patient safety, and giving
valuable insights that make the team

2101
01:58:44,560 --> 01:58:47,877
better at giving you what you need
most, which is better health care.

2102
01:58:47,960 --> 01:58:49,237
>> Yeah, and I think
that one of the things

2103
01:58:49,320 --> 01:58:52,877
that we're most excited about is that
the models are already pretty incredible

2104
01:58:52,960 --> 01:58:57,157
at textbook knowledge. It read all
the journals and all the papers, but what

2105
01:58:57,240 --> 01:59:01,077
they're really lacking is the kind of
clinical practice and clinical expertise

2106
01:59:01,160 --> 01:59:04,517
of your team and your clinicians
that you've developed over the last many

2107
01:59:04,600 --> 01:59:08,877
decades. So, how do you think we might
go about using that clinical practice to

2108
01:59:08,960 --> 01:59:11,317
improve the performance
of the model in production?

2109
01:59:11,400 --> 01:59:15,797
>> So, the exciting part here is we each
do what we do best, and we can tackle

2110
01:59:15,880 --> 01:59:17,357
something that has eluded health care

2111
01:59:17,440 --> 01:59:20,197
for a long time, trusted, scalable

2112
01:59:20,280 --> 01:59:25,037
solutions. And to do that, you need
to have the right data, you certainly need

2113
01:59:25,120 --> 01:59:28,917
to have the right people, but you also
need to have a very patient-focused

2114
01:59:29,000 --> 01:59:30,957
lens. And between the two of us, we now

2115
01:59:31,040 --> 01:59:34,837
have these all this together, so we can

2116
01:59:34,920 --> 01:59:39,397
build this frontier model, and then
offer safe, secure, trustworthy, and of

2117
01:59:39,480 --> 01:59:42,557
course effective health
care solutions for all.

2118
01:59:42,640 --> 01:59:45,277
>> Well, our number one objective,
of course, is to put the patient first,

2119
01:59:45,360 --> 01:59:48,717
deliver the highest quality we can
in a trusted way, and then hopefully share

2120
01:59:48,800 --> 01:59:52,157
that with the world over. So, we're very
excited for this partnership. Can't wait

2121
01:59:52,240 --> 01:59:53,717
to share more with everyone in the future.

2122
01:59:53,800 --> 01:59:57,200
>> Us too. Thank you. >> Thank you.

2123
01:59:57,880 --> 01:59:59,900
>> [applause]

2124
02:00:01,240 --> 02:00:04,397
>> So, today marks
some very, very exciting

2125
02:00:04,480 --> 02:00:07,917
steps that we're taking on our journey
to create humanistic superintelligence

2126
02:00:08,000 --> 02:00:12,877
at Microsoft. We now have an incredible
roster of seven new world-class models

2127
02:00:12,960 --> 02:00:15,997
to keep everybody working
at the absolute frontier.

2128
02:00:16,080 --> 02:00:19,917
And we're really looking forward
to everybody being able to co-create your

2129
02:00:20,000 --> 02:00:22,597
own unique agents, adapted to you, that

2130
02:00:22,680 --> 02:00:24,837
you'll control. I really feel like this

2131
02:00:24,920 --> 02:00:28,717
is a new era in AI, an era of AI that

2132
02:00:28,800 --> 02:00:33,077
you control on your terms. So, let's
build it together. Thank you very much,

2133
02:00:33,160 --> 02:00:36,580
everyone. >> [applause]

2134
02:00:40,440 --> 02:00:46,600
>> And now over to Tanya to show us
how frontier tuning works in practice.

2135
02:00:47,660 --> 02:00:49,680
>> [applause]

2136
02:00:52,520 --> 02:00:56,757
>> Thanks, Mustafa.
With frontier tuning, we're making it

2137
02:00:56,840 --> 02:00:59,397
possible for you to create your own

2138
02:00:59,480 --> 02:01:01,237
enterprise AI

2139
02:01:01,320 --> 02:01:04,717
with the models and the harness tuned on

2140
02:01:04,800 --> 02:01:07,037
your data and your workflows. Let me

2141
02:01:07,120 --> 02:01:12,080
show you how you can build
your own hill climbing machine.

2142
02:01:12,400 --> 02:01:17,277
MAI thinking one is now available
in private preview in the foundry model

2143
02:01:17,360 --> 02:01:21,077
catalog. You can go ahead
and deploy the model as

2144
02:01:21,160 --> 02:01:23,957
is, or if you want to go on your own

2145
02:01:24,040 --> 02:01:26,237
hill climbing journey, you can start by

2146
02:01:26,320 --> 02:01:29,237
clicking the fine-tune button.

2147
02:01:29,320 --> 02:01:31,437
Now, I'm in the fine-tuning UI. The

2148
02:01:31,520 --> 02:01:34,677
first thing I'll do is add my data set.

2149
02:01:34,760 --> 02:01:39,157
Next, I'm going to add a grader,
and that's it. I go ahead and submit the

2150
02:01:39,240 --> 02:01:40,557
job.

2151
02:01:40,640 --> 02:01:43,757
Couple of hours in, I see how the model

2152
02:01:43,840 --> 02:01:46,557
is learning. It generates rollouts, it

2153
02:01:46,640 --> 02:01:50,317
scores them, and we start to hill climb.

2154
02:01:50,400 --> 02:01:54,557
Now, if you want full control over
your RL training loop, let me give you a

2155
02:01:54,640 --> 02:01:56,917
sneak peek into our low-level training

2156
02:01:57,000 --> 02:02:01,717
API. You can see here
I have the MAI thinking

2157
02:02:01,800 --> 02:02:02,917
model.

2158
02:02:03,000 --> 02:02:06,197
I can also configure my rollout strategy

2159
02:02:06,280 --> 02:02:09,237
as well as my hyper parameters to define

2160
02:02:09,320 --> 02:02:11,917
exactly how I want my training algorithm

2161
02:02:12,000 --> 02:02:13,597
to work.

2162
02:02:13,680 --> 02:02:16,437
In fact, I can also incorporate my own

2163
02:02:16,520 --> 02:02:19,357
RL gym here by defining the tools that

2164
02:02:19,440 --> 02:02:21,877
this model interacts with.

2165
02:02:21,960 --> 02:02:24,517
All right. You just saw all the code.

2166
02:02:24,600 --> 02:02:27,437
But as an M365 customer, you're never

2167
02:02:27,520 --> 02:02:30,277
starting from scratch. Let's hop on to

2168
02:02:30,360 --> 02:02:32,517
Copilot.

2169
02:02:32,600 --> 02:02:34,917
As a part of frontier tuning, we're

2170
02:02:35,000 --> 02:02:38,077
introducing a new way to build RL

2171
02:02:38,160 --> 02:02:41,637
environments based on your data and your

2172
02:02:41,720 --> 02:02:44,717
workflows. One of our customers, Land

2173
02:02:44,800 --> 02:02:46,957
O'Lakes, one amongst the largest

2174
02:02:47,040 --> 02:02:50,037
agribusinesses in the United States, is

2175
02:02:50,120 --> 02:02:52,837
in fact using this to perfect that

2176
02:02:52,920 --> 02:02:55,157
butter on your morning toast. Let me

2177
02:02:55,240 --> 02:02:58,797
walk you through their environment.

2178
02:02:58,880 --> 02:03:01,197
Now, on a high level, the environment

2179
02:03:01,280 --> 02:03:03,877
consists of skills, knowledge, and

2180
02:03:03,960 --> 02:03:07,557
tools. But, in the back end, we create

2181
02:03:07,640 --> 02:03:12,597
an entire RL gym for your agents
to continuously learn the way that you

2182
02:03:12,680 --> 02:03:15,277
work. Now, let's look at the butter report

2183
02:03:15,360 --> 02:03:17,437
generation skill here.

2184
02:03:17,520 --> 02:03:20,277
These tasks are very complex. They

2185
02:03:20,360 --> 02:03:23,237
require many manual steps and a high

2186
02:03:23,320 --> 02:03:28,317
degree of precision. Now,
in these tasks, even 80% accuracy

2187
02:03:28,400 --> 02:03:30,077
isn't good enough.

2188
02:03:30,160 --> 02:03:32,477
To help clients to higher accuracy,

2189
02:03:32,560 --> 02:03:35,517
we're extending the industry definition

2190
02:03:35,600 --> 02:03:38,557
of skills to include rubrics on what

2191
02:03:38,640 --> 02:03:40,277
good looks like.

2192
02:03:40,360 --> 02:03:42,677
But, this is just one task. How do I

2193
02:03:42,760 --> 02:03:46,237
scale this to codify all of the tasks in

2194
02:03:46,320 --> 02:03:50,637
enterprise? You spend
a lot of time in M365, in

2195
02:03:50,720 --> 02:03:53,037
Teams, Outlook, Word, Excel, and

2196
02:03:53,120 --> 02:03:56,157
PowerPoint. We use those signals to

2197
02:03:56,240 --> 02:03:58,757
suggest skills and rubrics that truly

2198
02:03:58,840 --> 02:04:04,037
define the way that you work.
Next, you can add your organizational

2199
02:04:04,120 --> 02:04:07,757
knowledge for grounding,
like from OneDrive, SharePoint.

2200
02:04:07,840 --> 02:04:12,437
Now, the environment comes built-in with
Microsoft tools, and you can add custom

2201
02:04:12,520 --> 02:04:16,677
tools to it.
Because these tools tap into real

2202
02:04:16,760 --> 02:04:19,077
workflows, we virtualize them to

2203
02:04:19,160 --> 02:04:21,757
simulate execution, so the model can

2204
02:04:21,840 --> 02:04:24,237
learn without actually impacting the

2205
02:04:24,320 --> 02:04:28,917
live state of your business.
And now, my most favorite part, the

2206
02:04:29,000 --> 02:04:30,557
science.

2207
02:04:30,640 --> 02:04:33,477
By generalizing all of these learnings

2208
02:04:33,560 --> 02:04:36,237
into the main model, as well as in the

2209
02:04:36,320 --> 02:04:39,037
embedding model, we're able to hill

2210
02:04:39,120 --> 02:04:44,560
climb for tasks that
require high accuracy.

2211
02:04:45,000 --> 02:04:49,317
And not only this, we are able to hill

2212
02:04:49,400 --> 02:04:52,797
climb to more than 90% accuracy for Land

2213
02:04:52,880 --> 02:04:56,557
O'Lakes tasks using the MEI model.

2214
02:04:56,640 --> 02:04:59,997
In fact, we estimate this model to be

2215
02:05:00,080 --> 02:05:04,237
10x more efficient than
the baseline models.

2216
02:05:04,320 --> 02:05:06,557
Now with this environment setup, let's

2217
02:05:06,640 --> 02:05:09,357
go back to butter reporting.

2218
02:05:09,440 --> 02:05:13,157
Let's run this task this time using this

2219
02:05:13,240 --> 02:05:16,717
tuned environment
as an inferencing harness.

2220
02:05:16,800 --> 02:05:19,357
Now usually this task takes a couple of

2221
02:05:19,440 --> 02:05:21,957
minutes to run. In the meantime, I will

2222
02:05:22,040 --> 02:05:25,957
show you a cached response.

2223
02:05:26,040 --> 02:05:28,877
Now as you can see here, the agent has

2224
02:05:28,960 --> 02:05:31,917
leveraged test time free search with

2225
02:05:32,000 --> 02:05:35,197
multiple models including a fine-tuned

2226
02:05:35,280 --> 02:05:38,237
model. And what you see here is a

2227
02:05:38,320 --> 02:05:41,197
summary that doesn't feel generic. It

2228
02:05:41,280 --> 02:05:44,237
feels undoubtedly Land O'Lakes. And

2229
02:05:44,320 --> 02:05:47,357
that's not it. The task holds itself to

2230
02:05:47,440 --> 02:05:50,157
high standards and it continuously

2231
02:05:50,240 --> 02:05:53,637
retrospects and evaluates itself.

2232
02:05:53,720 --> 02:05:56,877
This way, with frontier tuning, your

2233
02:05:56,960 --> 02:05:59,877
agent continuously improves with your

2234
02:05:59,960 --> 02:06:04,077
compliant RL environment, on your data,

2235
02:06:04,160 --> 02:06:07,837
encoding the way you work. Now that is

2236
02:06:07,920 --> 02:06:11,477
what I call frontier
tuning as smooth as butter.

2237
02:06:11,560 --> 02:06:16,077
We cannot wait to see
the environments you built.

2238
02:06:16,160 --> 02:06:19,007
Back to you, Satya. >> All right.

2239
02:06:19,090 --> 02:06:21,077
>> [applause]

2240
02:06:21,160 --> 02:06:24,357
>> Thank you so much, Tanya, Mustafa, Jen

2241
02:06:24,440 --> 02:06:26,157
Rico. Thank you for the partnership

2242
02:06:26,240 --> 02:06:29,357
here. Uh you know, what you just saw is

2243
02:06:29,440 --> 02:06:33,197
a pretty significant shift. Uh

2244
02:06:33,280 --> 02:06:36,317
we believe that times come for every

2245
02:06:36,400 --> 02:06:40,077
company to just move from consuming

2246
02:06:40,160 --> 02:06:44,237
a frontier model to fully participating

2247
02:06:44,320 --> 02:06:45,757
at the frontier in the frontier

2248
02:06:45,840 --> 02:06:48,917
ecosystem. That's the transition. Uh you

2249
02:06:49,000 --> 02:06:51,917
can have your own private Evals and

2250
02:06:52,000 --> 02:06:53,357
outcomes,

2251
02:06:53,440 --> 02:06:56,717
your private RLs and traces, your

2252
02:06:56,800 --> 02:06:59,077
enterprise knowledge, create the

2253
02:06:59,160 --> 02:07:02,597
scaffolding for models to hill climb.

2254
02:07:02,680 --> 02:07:07,917
That's what will allow you to create
that differentiated IP that you own, you

2255
02:07:08,000 --> 02:07:09,397
control.

2256
02:07:09,480 --> 02:07:14,997
But the second salient point is that
there is a new operating point at the

2257
02:07:15,080 --> 02:07:16,317
frontier.

2258
02:07:16,400 --> 02:07:20,237
Uh where you can use a very efficient uh

2259
02:07:20,320 --> 02:07:24,517
reasoning model and a coding model and

2260
02:07:24,600 --> 02:07:28,397
achieve frontier level performance
because you've done the hard work of

2261
02:07:28,480 --> 02:07:30,557
creating that environment, that RL, that

2262
02:07:30,640 --> 02:07:35,757
hill climbing machine in which these
models with your traces can hill climb

2263
02:07:35,840 --> 02:07:39,517
to the frontier.
So we think that these combination of

2264
02:07:39,600 --> 02:07:44,557
these two is a pretty big game-changer
on how people think about what does it

2265
02:07:44,640 --> 02:07:48,757
mean to operate at the frontier? What
does frontier tokens look like? How are

2266
02:07:48,840 --> 02:07:50,837
you in control? What's the future of a

2267
02:07:50,920 --> 02:07:53,397
firm? These are the big questions and

2268
02:07:53,480 --> 02:07:55,957
really an ecosystem that that gets built

2269
02:07:56,040 --> 02:07:58,437
around as opposed to a few models that

2270
02:07:58,520 --> 02:08:01,077
just are hungry for all data.

2271
02:08:01,160 --> 02:08:03,797
To close out though, I want to talk

2272
02:08:03,880 --> 02:08:07,717
about that frontiers beyond um and that

2273
02:08:07,800 --> 02:08:10,997
push this ecosystem of ours and really

2274
02:08:11,080 --> 02:08:13,437
say what's the next big thing.

2275
02:08:13,520 --> 02:08:18,437
Bending the curve, you know, when
Mustafa talked about this with Jean Rico

2276
02:08:18,520 --> 02:08:22,317
um when it comes to health,
but building, you know, that scientific

2277
02:08:22,400 --> 02:08:24,397
discovery loop

2278
02:08:24,480 --> 02:08:28,037
can have perhaps the biggest societal uh

2279
02:08:28,120 --> 02:08:31,277
impact, right? Science today still

2280
02:08:31,360 --> 02:08:33,317
is a little too linear. You create a

2281
02:08:33,400 --> 02:08:36,157
hypothesis, you form a hypothesis,

2282
02:08:36,240 --> 02:08:40,837
you run an experiment, you wait
for lab results, which is out of band,

2283
02:08:40,920 --> 02:08:45,277
and then you begin again. But what
if the scientific method itself could

2284
02:08:45,360 --> 02:08:49,517
become more continuous,
more parallel, more programmable?

2285
02:08:49,600 --> 02:08:53,277
That's what we're doing with Microsoft
Discovery, right? Discovery brings

2286
02:08:53,360 --> 02:08:57,557
together the models, the HPC compute,
the knowledge graphs of all the

2287
02:08:57,640 --> 02:08:59,357
scientific knowledge,

2288
02:08:59,440 --> 02:09:02,757
the automated lab and simulation into

2289
02:09:02,840 --> 02:09:05,637
one agentic discovery loop. And so

2290
02:09:05,720 --> 02:09:08,637
today, I'm really thrilled to GA

2291
02:09:08,720 --> 02:09:11,397
Microsoft Discovery and and to show you

2292
02:09:11,480 --> 02:09:15,857
all of this in action, let me invite up
on David on stage. David, take it away.

2293
02:09:15,940 --> 02:09:19,117
>> [applause and cheering] >> Hello.

2294
02:09:19,200 --> 02:09:22,557
I work in the Microsoft
Discovery and Quantum team.

2295
02:09:22,640 --> 02:09:24,877
Today, to recycle a PET plastic like

2296
02:09:24,960 --> 02:09:27,397
this bottle, you have to shred it and

2297
02:09:27,480 --> 02:09:32,357
melt it. The result is degraded. You
cannot use it to make this bottle again.

2298
02:09:32,440 --> 02:09:35,757
It's downcycling.
Cambridge Consultants, part of

2299
02:09:35,840 --> 02:09:40,317
Capgemini, is using Microsoft Discovery
to advance that research. Let me show it

2300
02:09:40,400 --> 02:09:44,917
to you. This is the Microsoft
Discovery app. It

2301
02:09:45,000 --> 02:09:47,117
is based on VS code because agentic

2302
02:09:47,200 --> 02:09:49,677
discovery has many parallels with

2303
02:09:49,760 --> 02:09:54,797
agentic software engineering. Today,
I want to do these three things as a

2304
02:09:54,880 --> 02:09:58,357
scientist. First,
I want to write a scientific

2305
02:09:58,440 --> 02:10:00,797
paper about this topic. I'll explore an

2306
02:10:00,880 --> 02:10:03,197
existing line of research. Instead of

2307
02:10:03,280 --> 02:10:06,797
melting the plastic, I can use proteins

2308
02:10:06,880 --> 02:10:11,637
to decompose it, so you can
recycle it again and again.

2309
02:10:11,720 --> 02:10:13,277
Second,

2310
02:10:13,360 --> 02:10:15,997
I want to perform the actual discovery

2311
02:10:16,080 --> 02:10:19,717
to find new proteins. And third,

2312
02:10:19,800 --> 02:10:24,957
I want to create a lab protocol
to test the results in a real lab.

2313
02:10:25,040 --> 02:10:26,717
If you are a developer, these three

2314
02:10:26,800 --> 02:10:29,517
steps should be very familiar to you.

2315
02:10:29,600 --> 02:10:33,837
They look a lot like planning,
coding and testing, and deploying to

2316
02:10:33,920 --> 02:10:36,277
production, but for science.

2317
02:10:36,360 --> 02:10:40,917
Let's launch it. This will
kick off the discovery engine.

2318
02:10:41,000 --> 02:10:45,917
You can think of this as a as a team of
specialized agents always running in the

2319
02:10:46,000 --> 02:10:50,917
background following the scientific
method. You can see those agents here,

2320
02:10:51,000 --> 02:10:52,597
and you can add more.

2321
02:10:52,680 --> 02:10:54,997
Microsoft Discovery includes a community

2322
02:10:55,080 --> 02:10:58,437
of agents, models, and tools across many

2323
02:10:58,520 --> 02:11:02,317
domains. You can use open
source, third parties,

2324
02:11:02,400 --> 02:11:06,797
or create your own. Okay,
so this is still running, and it

2325
02:11:06,880 --> 02:11:09,197
will be running for a while, even hours

2326
02:11:09,280 --> 02:11:11,717
or days, because the discovery engine,

2327
02:11:11,800 --> 02:11:14,957
like science, is not sequential. It is

2328
02:11:15,040 --> 02:11:17,477
exploring hypotheses and performing

2329
02:11:17,560 --> 02:11:20,637
long-running simulations dynamically.

2330
02:11:20,720 --> 02:11:24,597
So, let me open one that is already
completed. You can see the files that it

2331
02:11:24,680 --> 02:11:29,277
created here. This one is the
research paper that I was asking

2332
02:11:29,360 --> 02:11:33,197
for. To create it, Discovery's using a

2333
02:11:33,280 --> 02:11:37,837
knowledge graph internally, bringing
together public scientific literature

2334
02:11:37,920 --> 02:11:41,957
and internal knowledge.
This is a critical asset because it

2335
02:11:42,040 --> 02:11:46,957
provides complete visibility of the
reasoning, so scientists can be in full

2336
02:11:47,040 --> 02:11:50,517
control. We can go now
to step two, the discovery

2337
02:11:50,600 --> 02:11:54,237
itself. Let me open
the output for that one.

2338
02:11:54,320 --> 02:11:58,117
So, how do you come
up with new proteins to

2339
02:11:58,200 --> 02:12:00,597
decompose plastic? This is the approach

2340
02:12:00,680 --> 02:12:02,237
that Discovery took.

2341
02:12:02,320 --> 02:12:05,277
First, you need an AI model to predict

2342
02:12:05,360 --> 02:12:07,677
how good a protein is for that.

2343
02:12:07,760 --> 02:12:09,997
Then, you need a way to generate new

2344
02:12:10,080 --> 02:12:12,437
proteins, and finally, you need to

2345
02:12:12,520 --> 02:12:15,637
identify the most promising ones. So,

2346
02:12:15,720 --> 02:12:20,397
let's start with the model. Microsoft
Discovery trained multiple models and

2347
02:12:20,480 --> 02:12:24,477
picked the best. And there's no
out-of-the-box agent for that. But,

2348
02:12:24,560 --> 02:12:27,637
that's okay. If Microsoft
Discovery does not find an

2349
02:12:27,720 --> 02:12:30,397
agent for a task, it will create one on

2350
02:12:30,480 --> 02:12:32,997
the fly, which is pretty cool. So,

2351
02:12:33,080 --> 02:12:36,197
here's all the files that it created.

2352
02:12:36,280 --> 02:12:38,917
This one here is the YAML for the agent.

2353
02:12:39,000 --> 02:12:43,157
And this one, for example,
is the Python code to train the models.

2354
02:12:43,240 --> 02:12:48,117
Now, we need to generate candidates
for new proteins. And that task requires a

2355
02:12:48,200 --> 02:12:53,037
lot of compute. Discovery
is integrated with HPC. So,

2356
02:12:53,120 --> 02:12:56,397
agents can use it for complex simulations.

2357
02:12:56,480 --> 02:12:58,557
You can see the process here.

2358
02:12:58,640 --> 02:13:01,437
It started with a seed protein, and then

2359
02:13:01,520 --> 02:13:05,957
it created variations by replacing small
segments. You can see those segments

2360
02:13:06,040 --> 02:13:09,277
there. Then, it applied
the model from before

2361
02:13:09,360 --> 02:13:13,797
to see the variation helps or not,
learning in that process.

2362
02:13:13,880 --> 02:13:16,277
This is done millions of times in

2363
02:13:16,360 --> 02:13:18,677
multiple jobs in parallel, exploring a

2364
02:13:18,760 --> 02:13:23,317
huge tree of proteins. The result is 80

2365
02:13:23,400 --> 02:13:25,837
proteins that are ready to be sent to

2366
02:13:25,920 --> 02:13:30,877
the lab for testing, which we said it
was like deploying software, right? The

2367
02:13:30,960 --> 02:13:34,517
problem, though, is that creating
a protein is a little bit more

2368
02:13:34,600 --> 02:13:37,077
complicated. The most common way is

2369
02:13:37,160 --> 02:13:39,917
inserting DNA into bacteria, so the

2370
02:13:40,000 --> 02:13:42,917
bacteria create the protein for me. So,

2371
02:13:43,000 --> 02:13:45,877
Discovery created another file.

2372
02:13:45,960 --> 02:13:48,437
This one here. This file contains all

2373
02:13:48,520 --> 02:13:50,477
the DNA sequences to create each

2374
02:13:50,560 --> 02:13:53,277
protein. I can send this now to a lab,

2375
02:13:53,360 --> 02:13:55,717
create the proteins, and test them.

2376
02:13:55,800 --> 02:13:59,317
Or, I can go one step further.

2377
02:13:59,400 --> 02:14:04,197
If I have an automated lab, I can
integrate it directly with the agents.

2378
02:14:04,280 --> 02:14:07,477
Let's do it. So, let me
go back to the session,

2379
02:14:07,560 --> 02:14:10,837
and let's just submit job to lab.

2380
02:14:10,920 --> 02:14:14,317
Go. This will use a agent that is sending

2381
02:14:14,400 --> 02:14:16,757
instructions for the lab equipment. And

2382
02:14:16,840 --> 02:14:22,317
this is very real. Cambridge Consultants
does have an automated lab, and I have

2383
02:14:22,400 --> 02:14:26,277
it right here. So,
this is the application controlling

2384
02:14:26,360 --> 02:14:29,037
the lab. It has a Copilot interface, so

2385
02:14:29,120 --> 02:14:33,357
the scientist can interact with the lab
to design experiments, which is It feels

2386
02:14:33,440 --> 02:14:35,957
like being Iron Man, but for chemistry.

2387
02:14:36,040 --> 02:14:40,277
In this case, Discovery is submitting
the job. You can see it already here.

2388
02:14:40,360 --> 02:14:43,837
But, remember, you need to grow
the bacteria, so this will take some time.

2389
02:14:43,920 --> 02:14:46,317
So, let me open a previous run.

2390
02:14:46,400 --> 02:14:48,357
And there you go. These are all the

2391
02:14:48,440 --> 02:14:51,277
steps that the agents are doing. Most of

2392
02:14:51,360 --> 02:14:58,160
them completely automated with human
supervision. How cool is that?

2393
02:14:58,510 --> 02:15:00,530
>> [applause]

2394
02:15:02,880 --> 02:15:06,437
>> This is bringing together the physical
world and the digital agents in a

2395
02:15:06,520 --> 02:15:11,157
unified discovery loop. And this was
just one example of what Microsoft

2396
02:15:11,240 --> 02:15:13,317
Discovery can do. Customers across

2397
02:15:13,400 --> 02:15:15,837
industries are using it today to embrace

2398
02:15:15,920 --> 02:15:20,697
a new era of scientific discovery.
Thank you. Back to you, Satya.

2399
02:15:20,780 --> 02:15:22,800
>> [applause]

2400
02:15:24,080 --> 02:15:25,877
>> Thank you so much, David. You know,

2401
02:15:25,960 --> 02:15:29,277
talking about uh scientific discovery,

2402
02:15:29,360 --> 02:15:31,837
we're also continuing to make rapid

2403
02:15:31,920 --> 02:15:34,717
progress on our long-term goal of

2404
02:15:34,800 --> 02:15:37,797
building a scalable quantum computer.

2405
02:15:37,880 --> 02:15:40,717
Uh we announced uh last year our first

2406
02:15:40,800 --> 02:15:44,917
QPU. Uh we created
a new state of matter uh

2407
02:15:45,000 --> 02:15:47,717
that was only theorized 100 years ago,

2408
02:15:47,800 --> 02:15:50,677
uh and we proved it out that it exists.

2409
02:15:50,760 --> 02:15:53,637
Our vision was to take a very radically

2410
02:15:53,720 --> 02:15:56,917
different approach uh to addressing the

2411
02:15:57,000 --> 02:16:01,317
fundamental barriers uh to building
a scalable quantum machine, which is all

2412
02:16:01,400 --> 02:16:03,237
about reliability,

2413
02:16:03,320 --> 02:16:06,357
uh speed, as well as size. Uh since

2414
02:16:06,440 --> 02:16:09,197
then, we have continued to make progress

2415
02:16:09,280 --> 02:16:11,997
across the full quantum stack uh, with

2416
02:16:12,080 --> 02:16:14,637
both our academic and industry partners.

2417
02:16:14,720 --> 02:16:17,037
In fact, in Q North, we will have a

2418
02:16:17,120 --> 02:16:21,757
quantum computer powered by Adam
computers, natural Adam computers. Uh,

2419
02:16:21,840 --> 02:16:23,837
with our stack in there, we're also

2420
02:16:23,920 --> 02:16:25,997
working with algorithmic and Columbia

2421
02:16:26,080 --> 02:16:31,677
and ETH Zurich. Um, and we
ourselves in fact use this discovery

2422
02:16:31,760 --> 02:16:32,317
um,

2423
02:16:32,400 --> 02:16:35,917
agentic loop to accelerate the work

2424
02:16:36,000 --> 02:16:38,797
uh, in quantum compressing years of

2425
02:16:38,880 --> 02:16:41,797
research uh, into this last year. And

2426
02:16:41,880 --> 02:16:45,717
today, I'm really thrilled to announce

2427
02:16:45,800 --> 02:16:50,037
uh, Myrana 2. And so, this is Myrana 2.

2428
02:16:50,120 --> 02:16:53,437
Um, Myrana 2 implements

2429
02:16:53,520 --> 02:16:56,517
the next-generation material stack that

2430
02:16:56,600 --> 02:16:59,837
we use discovery to discover and build

2431
02:16:59,920 --> 02:17:02,637
and help fabricate. The resulting qubits

2432
02:17:02,720 --> 02:17:05,197
are exceptionally reliable and capable

2433
02:17:05,280 --> 02:17:08,117
of maintaining uh, their state much

2434
02:17:08,200 --> 02:17:11,037
longer. While other common approaches

2435
02:17:11,120 --> 02:17:13,557
deliver a lifetime of just microseconds

2436
02:17:13,640 --> 02:17:16,797
or even milliseconds, Myrana 2 provides

2437
02:17:16,880 --> 02:17:20,917
qubit mean lifetime of 20 seconds or up

2438
02:17:21,000 --> 02:17:23,237
to even a minute. Uh, essentially a

2439
02:17:23,320 --> 02:17:29,087
thousand times higher than what we
were able to achieve with Myrana 1. Um,

2440
02:17:29,170 --> 02:17:32,677
>> [applause] >> and the operations,
and this is so key,

2441
02:17:32,760 --> 02:17:35,397
operations in Myrana 2 are 1

2442
02:17:35,480 --> 02:17:38,157
microsecond, enabling pretty complex

2443
02:17:38,240 --> 02:17:40,877
compact you know, quantum computation

2444
02:17:40,960 --> 02:17:43,397
uh, in that lifetime. And all of this in

2445
02:17:43,480 --> 02:17:47,117
the same qubit form factor of Myrana 1.

2446
02:17:47,200 --> 02:17:51,317
Uh, in 1 100th of a millimeter
controlled all digitally, which is again

2447
02:17:51,400 --> 02:17:54,077
a super important aspect. Uh, making it

2448
02:17:54,160 --> 02:17:56,477
all possible to fit a million of these

2449
02:17:56,560 --> 02:17:59,437
qubits in a chip smaller than a credit

2450
02:17:59,520 --> 02:18:02,317
card. It's this combination

2451
02:18:02,400 --> 02:18:05,357
of the reliability, the speed, the size

2452
02:18:05,440 --> 02:18:08,117
that makes the topological approach so

2453
02:18:08,200 --> 02:18:11,037
unique. Uh with Myrana 1, we had proven

2454
02:18:11,120 --> 02:18:13,797
out the foundational physics, and with

2455
02:18:13,880 --> 02:18:16,957
Myrana 2, now we begin the engineering

2456
02:18:17,040 --> 02:18:22,930
scale. But ultimately, yeah. >> [applause]

2457
02:18:24,360 --> 02:18:28,477
>> But ultimately, it is never about tech

2458
02:18:28,560 --> 02:18:30,277
for tech's sake.

2459
02:18:30,360 --> 02:18:32,837
It's about tackling those pressing

2460
02:18:32,920 --> 02:18:36,757
challenges of people and planet.

2461
02:18:36,840 --> 02:18:39,157
It is also the fundamental point of this

2462
02:18:39,240 --> 02:18:44,117
conference. The question is not whether
you can build the next great model, the

2463
02:18:44,200 --> 02:18:47,877
next great platform,
or even this quantum machine.

2464
02:18:47,960 --> 02:18:50,397
It's the question is how do we build

2465
02:18:50,480 --> 02:18:53,717
this frontier ecosystem together?

2466
02:18:53,800 --> 02:18:56,397
Because there are really two stories

2467
02:18:56,480 --> 02:18:59,357
people can tell about this moment.

2468
02:18:59,440 --> 02:19:01,957
One is that technology concentrates

2469
02:19:02,040 --> 02:19:05,397
power, reduces human agency, and leaves

2470
02:19:05,480 --> 02:19:09,077
the society to absorb the consequences.

2471
02:19:09,160 --> 02:19:12,397
The other is that we use this next wave

2472
02:19:12,480 --> 02:19:16,877
to unlock opportunity for developers,
scientists, enterprises, and every

2473
02:19:16,960 --> 02:19:20,557
community. And our job
is to make the second story

2474
02:19:20,640 --> 02:19:23,437
true. It's That's our North Star for the

2475
02:19:23,520 --> 02:19:28,077
frontier ecosystem. Let's all go build
together. Thank you all very, very much.

2476
02:19:28,160 --> 02:19:31,050
Thank you. >> [applause and cheering]