WEBVTT

00:00:50.330 --> 00:00:54.337
>> Good morning.
It is fantastic to be back in

00:00:54.420 --> 00:00:56.937
San Francisco build.
It is always fun to be a

00:00:57.020 --> 00:00:59.747
developer conference to see

00:00:59.830 --> 00:01:00.747
great change.

00:01:00.830 --> 00:01:02.747
Developer conferences are always

00:01:02.830 --> 00:01:04.277
about understanding tech shifts

00:01:04.360 --> 00:01:07.597
and understanding the new stack.

00:01:07.680 --> 00:01:10.857
It is also about really
coming to grips with the new

00:01:10.940 --> 00:01:12.387
opportunity for us as

00:01:12.470 --> 00:01:13.987
developers, for the companies we

00:01:14.070 --> 00:01:18.927
work at as well as the broader world.

00:01:19.010 --> 00:01:20.627
Today we are going to unpack this.

00:01:20.710 --> 00:01:21.947
This conference is all about

00:01:22.030 --> 00:01:22.347
that.

00:01:22.430 --> 00:01:24.117
If there is one key take away it

00:01:24.200 --> 00:01:27.937
would be this, how do you all

00:01:28.020 --> 00:01:30.477
participate fully in this

00:01:30.560 --> 00:01:33.977
frontier intelligence ecosystem.

00:01:34.060 --> 00:01:36.317
It is not about any one piece of

00:01:36.400 --> 00:01:37.417
technology you hear about or

00:01:37.500 --> 00:01:39.637
even the platform itself, it is

00:01:39.720 --> 00:01:41.497
about the value that you can

00:01:41.580 --> 00:01:43.617
build, you can compound, you can

00:01:43.700 --> 00:01:49.997
create on top of the platform.

00:01:50.080 --> 00:01:54.167
That's what the conference will
be about and that's what we will

00:01:54.250 --> 00:01:54.727
focus on.

00:01:54.810 --> 00:01:56.647
Let's take a look at this AI

00:01:56.730 --> 00:01:57.947
stack this conference is going

00:01:58.030 --> 00:02:03.607
to unpack in a great amount of

00:02:03.690 --> 00:02:06.647
detail. It starts with ubiquitous

00:02:06.730 --> 00:02:10.177
compute fabric that expands
the edge and the cloud.

00:02:10.260 --> 00:02:11.547
You then have this layer which

00:02:11.630 --> 00:02:13.337
is the emerging layer where you

00:02:13.420 --> 00:02:15.387
have the models, you have

00:02:15.470 --> 00:02:17.287
context, you have the tools, the

00:02:17.370 --> 00:02:22.857
models can access, and then you

00:02:22.940 --> 00:02:25.697
have the run time where you

00:02:25.780 --> 00:02:28.887
deploy the agents
and applications that you build on

00:02:28.970 --> 00:02:30.447
top of the content player and

00:02:30.530 --> 00:02:31.617
model layer.

00:02:31.700 --> 00:02:33.967
Of course you have the best

00:02:34.050 --> 00:02:36.437
tooling to do all of this and

00:02:36.520 --> 00:02:37.707
you have the security and

00:02:37.790 --> 00:02:43.077
compliance and governance.

00:02:43.160 --> 00:02:46.297
That's the simplest form.

00:02:46.380 --> 00:02:48.467
Let's start where it all starts

00:02:48.550 --> 00:02:50.717
which is infrastructure.

00:02:50.800 --> 00:02:52.287
In fact let's start at the edge

00:02:52.370 --> 00:02:55.127
with windows.

00:02:55.210 --> 00:02:57.897
The amount of compute there is,

00:02:57.980 --> 00:03:00.427
at the edge is actually

00:03:00.510 --> 00:03:00.967
astounding.

00:03:01.050 --> 00:03:05.867
Think about every NPU, GPU, CPU

00:03:05.950 --> 00:03:11.220
even, every PC

00:03:11.460 --> 00:03:18.460
we ask ourselves one simple

00:03:25.600 --> 00:03:30.957
question. AI.

00:03:31.040 --> 00:03:33.197
In some sense we are
delivering that already.

00:03:33.280 --> 00:03:37.717
When you look at something like
outlook summarized it is using

00:03:37.800 --> 00:03:41.607
on board AI locally.
Same thing with PowerPoint all text.

00:03:41.690 --> 00:03:44.477
Or teams super resolution.

00:03:44.560 --> 00:03:47.747
It is just not Microsoft.

00:03:47.830 --> 00:03:51.747
It is Adobe effects or premier.

00:03:51.830 --> 00:03:58.830
Using windows ML in order to

00:04:00.010 --> 00:04:01.357
happen to all of this compute

00:04:01.440 --> 00:04:05.297
power is to expand the scope of

00:04:05.380 --> 00:04:09.387
windows ML and windows AI.

00:04:09.470 --> 00:04:13.367
You now are full install
base of GPUs you can get to.

00:04:13.450 --> 00:04:16.977
I am thrilled every developer
out there can count on building

00:04:17.060 --> 00:04:18.907
for local on board AI and then

00:04:18.990 --> 00:04:22.647
have it run across all of the

00:04:22.730 --> 00:04:27.097
install base.
Now if you are -- we are also

00:04:27.180 --> 00:04:29.217
announcing two cool models that

00:04:29.300 --> 00:04:36.300
are all going to run on windows

00:04:38.110 --> 00:04:43.587
in INVOX. Eye on instruct, it is a great

00:04:43.670 --> 00:04:43.987
reasoning model.

00:04:44.070 --> 00:04:49.437
We have the planning model eye

00:04:49.520 --> 00:04:53.777
-- Ion plan. You have a full local agentic

00:04:53.860 --> 00:04:55.917
loop you can give it tools

00:04:56.000 --> 00:04:59.097
access and build a fully agentic

00:04:59.180 --> 00:05:02.387
applications without having to

00:05:02.470 --> 00:05:05.587
run trip to the cloud.

00:05:05.670 --> 00:05:07.487
In order to really push the

00:05:07.570 --> 00:05:11.897
limits, realize this unmetered

00:05:11.980 --> 00:05:13.367
intelligence, there is a lot of

00:05:13.450 --> 00:05:15.167
hardware that we are adding too.

00:05:15.250 --> 00:05:21.567
We are thrilled to see
the innovation through the

00:05:21.650 --> 00:05:24.777
ecosystem. The stuff from
Intel is exciting to see.

00:05:24.860 --> 00:05:27.747
Qualcomm announced two sets
of things which are great one on

00:05:27.830 --> 00:05:31.847
the high end snap dragon X on

00:05:31.930 --> 00:05:38.930
the lower end to sub 500PCs.

00:05:39.840 --> 00:05:44.357
Of course that brings us to inn

00:05:44.440 --> 00:05:48.027
NVIDIA. This is next generation SOC for

00:05:48.110 --> 00:05:51.747
PCs. It brings together the CPU the

00:05:51.830 --> 00:05:56.007
GPU as well as AI
capabilities into a single SOC.

00:05:56.090 --> 00:05:58.807
It includes unified memory,

00:05:58.890 --> 00:06:05.890
architecture and also integrated

00:06:06.580 --> 00:06:06.697
DRTM.

00:06:06.780 --> 00:06:12.987
You have the SOC coming

00:06:13.070 --> 00:06:17.227
together. One we build is the Surface

00:06:17.310 --> 00:06:19.467
Ultra it is a beautiful device.

00:06:19.550 --> 00:06:22.347
It brings the power of inn

00:06:22.430 --> 00:06:27.907
NVIDIA together with the design
and Craftsman ship of Surface,

00:06:27.990 --> 00:06:30.377
128 giga bites of unified memory.

00:06:30.460 --> 00:06:33.407
Beautiful 2,000 display, and it

00:06:33.490 --> 00:06:40.490
has an all day battery life.

00:06:41.020 --> 00:06:42.887
Excited to see this later in the fall.

00:06:42.970 --> 00:06:47.627
It is wonderful to see all
of the designs from all of the OEM

00:06:47.710 --> 00:06:51.097
partners building and taking
advantage of the SOC and the new

00:06:51.180 --> 00:06:54.067
platform and bringing out pretty

00:06:54.150 --> 00:06:58.887
exciting machines that are going
to be all available come this fall.

00:06:58.970 --> 00:07:02.277
So, of course, we said,
okay, this is fantastic.

00:07:02.360 --> 00:07:07.547
What can we do next?
We said, let's try and push

00:07:07.630 --> 00:07:10.817
this. Push the architecture to its

00:07:10.900 --> 00:07:12.627
limit for developers.

00:07:12.710 --> 00:07:15.517
What if we could just Max the

00:07:15.600 --> 00:07:18.287
compute, Max the memory, build

00:07:18.370 --> 00:07:23.097
out that developer machine that
is the dream machine, and that's

00:07:23.180 --> 00:07:25.227
what we are announcing today.

00:07:25.310 --> 00:07:31.747
Surface RTX, dev VOX.

00:07:31.830 --> 00:07:38.830
Let's roll the video.

00:08:35.630 --> 00:08:37.917
>> It is truly a dream machine.

00:08:38.000 --> 00:08:42.567
It has 1 Peta flop of AI compute.

00:08:42.650 --> 00:08:48.977
20CPU cores. All of those things
have 128 giga bites of unified memory

00:08:49.060 --> 00:08:49.447
access.

00:08:49.530 --> 00:08:54.537
So super excited.

00:08:54.620 --> 00:08:58.557
I am on the wait list as well,
there's a wait list.

00:08:58.640 --> 00:08:59.587
We will get there.

00:08:59.670 --> 00:09:00.557
[Laughter]

00:09:00.640 --> 00:09:05.697
So then we said why stop there?

00:09:05.780 --> 00:09:08.367
We said what if we did one more

00:09:08.450 --> 00:09:12.867
thing? In fact I think Jensen did that

00:09:12.950 --> 00:09:16.977
already. He said windows is coming to the

00:09:17.060 --> 00:09:17.907
DGX station.

00:09:17.990 --> 00:09:20.247
I describe it as the desktop

00:09:20.330 --> 00:09:24.747
data center which you can have

00:09:24.830 --> 00:09:29.617
running a 1 trillion model locally.

00:09:29.700 --> 00:09:31.217
It is close to what we perhaps

00:09:31.300 --> 00:09:36.827
had when we built GPT-5 or 3 one

00:09:36.910 --> 00:09:39.227
of the first super computers.
It is crazy to think we have

00:09:39.310 --> 00:09:44.647
come this far with where you
can have a data center on your

00:09:44.730 --> 00:09:48.667
desktop. We are extending
the developer end points to the cloud.

00:09:48.750 --> 00:09:53.947
Windows 365 has the developer
distribution that is optimized

00:09:54.030 --> 00:09:59.167
for developer productivity
in the cloud that's windows 365

00:09:59.250 --> 00:10:04.417
that I use every day. It is about making
windows whether it is on the laptop,

00:10:04.500 --> 00:10:07.557
desktop, the cloud wherever
the best place to build.

00:10:07.640 --> 00:10:10.297
And so to that end we have tons

00:10:10.380 --> 00:10:15.197
and tons of updates.
We have starting -- we are

00:10:15.280 --> 00:10:19.767
starting with the favorite
things for all of us, which is

00:10:19.850 --> 00:10:21.267
distraction 3 dev environment.

00:10:21.350 --> 00:10:23.777
We are introducing an

00:10:23.860 --> 00:10:29.577
intelligent terminal that has
built in GitHub Copilot.

00:10:29.660 --> 00:10:30.947
Copilot intelligence.

00:10:31.030 --> 00:10:34.897
Of course lots and lots of Linux

00:10:34.980 --> 00:10:38.917
love for windows now.
You have 70 plus utilities.

00:10:39.000 --> 00:10:45.097
Grab in full glory is now available
for regular windows access.

00:10:45.180 --> 00:10:48.267
70 plus utilities as I said.
All of that is coming to

00:10:48.350 --> 00:10:48.667
windows.

00:10:48.750 --> 00:10:55.750
We are also bringing that you

00:10:58.520 --> 00:10:58.907
love.

00:10:58.990 --> 00:11:05.990
Home brew will be on native as

00:11:16.270 --> 00:11:18.457
well. Switching all of this.

00:11:18.540 --> 00:11:20.727
Having first class support for

00:11:20.810 --> 00:11:26.407
containers while will hp --

00:11:26.490 --> 00:11:28.467
will really help us when

00:11:28.550 --> 00:11:29.937
developing locally.

00:11:30.020 --> 00:11:35.247
Showing you all of this on new

00:11:35.330 --> 00:11:36.047
surface on RTX.

00:11:36.130 --> 00:11:38.377
I wanted to invite up on stage

00:11:38.460 --> 00:11:43.487
Kayla to take a spin through all

00:11:43.570 --> 00:11:47.640
of the Dev tools. Kayla, go ahead.

00:11:47.840 --> 00:11:52.367
>> Our team has been working hard to make
Windows a great place for development.

00:11:52.450 --> 00:11:55.887
Today I am going to show
you major improvements and

00:11:55.970 --> 00:11:57.217
delighters you can try out today.

00:11:57.300 --> 00:11:58.747
What we have here is the default

00:11:58.830 --> 00:12:04.637
experience on the Surface
RTX Spark DevBox.

00:12:04.720 --> 00:12:09.617
It is calm. No news feed no
widgets popping up no notifications.

00:12:09.700 --> 00:12:11.247
We are in dark mode of course.

00:12:11.330 --> 00:12:14.167
I am ready to start development.

00:12:14.250 --> 00:12:15.367
There's one thing I would like to change.

00:12:15.450 --> 00:12:18.487
I like to put my task bar on the left.

00:12:18.570 --> 00:12:22.397
Let me jump into my task bar

00:12:22.480 --> 00:12:27.997
settings using run which is built
with power tools xand pallet.

00:12:28.080 --> 00:12:29.577
No I want it all of the way on the left.

00:12:29.660 --> 00:12:32.287
There we go. After popular demand we are

00:12:32.370 --> 00:12:37.637
excited to announce vertical
task bar is available in Windows

00:12:37.720 --> 00:12:39.277
insider build.

00:12:39.360 --> 00:12:43.917
The new surfacSurface RTX Spark

00:12:44.000 --> 00:12:46.397
bunch of tools already installed

00:12:46.480 --> 00:12:51.787
like python and node.
All developer goodness all in one file.

00:12:51.870 --> 00:12:55.057
If you want the same experience
on your device we are making

00:12:55.140 --> 00:12:57.527
this file available to everyone right now.

00:12:57.610 --> 00:13:02.167
We have a public repo set up
with configuration file and how

00:13:02.250 --> 00:13:08.237
to apply it. It will make adjustments
to windows and install all of the tools.

00:13:08.320 --> 00:13:12.137
One cool thing I have running
here is power toys new utility

00:13:12.220 --> 00:13:16.917
called grab and move. Hold ality
and move the window around from anywhere.

00:13:17.000 --> 00:13:21.637
You can enable end task which
let's you end the process

00:13:21.720 --> 00:13:23.617
without having to open task manager.

00:13:23.700 --> 00:13:26.957
Let me jump into my Dev drive.

00:13:27.040 --> 00:13:31.757
It runs with defender running in

00:13:31.840 --> 00:13:35.817
async and authorize performance
when comes to developer scenarios.

00:13:35.900 --> 00:13:40.237
Get aware. We have stuff like last name

00:13:40.320 --> 00:13:43.777
change message status of each
file and the branch name is on

00:13:43.860 --> 00:13:45.177
the bottom left.

00:13:45.260 --> 00:13:47.607
So now let's get started

00:13:47.690 --> 00:13:54.097
building and open our terminal.

00:13:54.180 --> 00:13:57.157
It helps working with agents
even more seamless.

00:13:57.240 --> 00:14:02.057
When you install intelligent
terminal you can pick your

00:14:02.140 --> 00:14:05.327
favorite agent.
I am using GitHub Copilot today.

00:14:05.410 --> 00:14:08.167
You can use which ever
agent speaks to you.

00:14:08.250 --> 00:14:12.317
In intelligent terminal I have
terminal pane at the top and

00:14:12.400 --> 00:14:14.937
agent working on the bottom.
I, work between them while the

00:14:15.020 --> 00:14:18.837
agent helps along the way.
Here is an air being generated.

00:14:18.920 --> 00:14:20.947
Agent pain is able to detect it

00:14:21.030 --> 00:14:25.997
and provide a fix which is great

00:14:26.080 --> 00:14:28.167
when I don't understand the sin tax.

00:14:28.250 --> 00:14:32.657
I am working on open claw I have
worked it using WSL container.

00:14:32.740 --> 00:14:34.727
It is a native container

00:14:34.810 --> 00:14:37.127
experience on Windows plus it

00:14:37.210 --> 00:14:40.497
can leverage the GPU.

00:14:40.580 --> 00:14:45.767
It can reference your existing
container files just like the

00:14:45.850 --> 00:14:49.667
one in open claw project.
This is open in Microsoft edit

00:14:49.750 --> 00:14:53.207
got sin tax highlighting
and latest version.

00:14:53.290 --> 00:14:56.047
You can see the container

00:14:56.130 --> 00:14:59.077
running with container command.

00:14:59.160 --> 00:15:03.047
Since we are on the topic of WSL

00:15:03.130 --> 00:15:04.687
there is a profile that is

00:15:04.770 --> 00:15:10.957
comfortable for those using
tools like star, CSH and home

00:15:11.040 --> 00:15:14.627
brew. It has all of the favorite
utilities and available in the

00:15:14.710 --> 00:15:19.467
repo I showed earlier
and includes B top one of my

00:15:19.550 --> 00:15:19.767
favorites.

00:15:19.850 --> 00:15:24.577
The Surface RTX Spark serves

00:15:24.660 --> 00:15:27.407
large local models for coding.

00:15:27.490 --> 00:15:31.747
I have done development with 120

00:15:31.830 --> 00:15:35.817
billion parameter model which
most machines can't load.

00:15:35.900 --> 00:15:37.317
We can see how many tokens I

00:15:37.400 --> 00:15:43.087
have used locally.
3.4 million tokens leveraged on

00:15:43.170 --> 00:15:46.457
the device itself.
We can kickoff agents using fleet.

00:15:46.540 --> 00:15:49.517
Just so you don't have to watch
me type I am going to use

00:15:49.600 --> 00:15:53.167
Copilot voice feature which
is leveraging its own local model.

00:15:53.250 --> 00:15:55.007
I will hold face for it and tell

00:15:55.090 --> 00:15:57.137
it what I want it to do.

00:15:57.220 --> 00:16:00.307
Find any console dot right line

00:16:00.390 --> 00:16:03.077
calls in the tray and node

00:16:03.160 --> 00:16:08.367
projects and convert them to the standard
logger used elsewhere in the code base.

00:16:08.450 --> 00:16:12.287
There we go. The main
agent will delegate sub

00:16:12.370 --> 00:16:15.787
agent tasks of appropriate
complexity to the local model

00:16:15.870 --> 00:16:20.267
use liesing GPU and making
it cost efficient.

00:16:20.350 --> 00:16:24.147
As developers while debugging we
are looking through log files to

00:16:24.230 --> 00:16:28.207
diagnose any issues. Finding the location
of the log files is a challenge.

00:16:28.290 --> 00:16:33.947
I would love to be able to type
something like grep log and find them all.

00:16:34.030 --> 00:16:37.547
Sweet. On top of already adding curl,

00:16:37.630 --> 00:16:41.997
car and pseudo to windows now we
are adding over 75 command line

00:16:42.080 --> 00:16:45.657
utilities like end, be head,
tail and touch for those of us

00:16:45.740 --> 00:16:49.397
who love to live in the terminal.

00:16:49.480 --> 00:16:53.727
So I found all of my log files,
but now having to pars them is

00:16:53.810 --> 00:16:57.747
the second challenge.
I have had eye on instruct

00:16:57.830 --> 00:17:01.567
practically performing analyses
on my log files this whole time.

00:17:01.650 --> 00:17:06.137
Now I can quickly diagnose
anything that has gone wrong in

00:17:06.220 --> 00:17:10.747
my development, plus I don't have to worry
about token usage since it is all local.

00:17:10.830 --> 00:17:13.397
We can take a look
at the machinery sources.

00:17:13.480 --> 00:17:15.737
Models are loaded you can see 90

00:17:15.820 --> 00:17:18.257
gigs of ram being utilized by

00:17:18.340 --> 00:17:22.997
the GPU using the full power of

00:17:23.080 --> 00:17:27.767
RTX Spark. We were able to use them

00:17:27.850 --> 00:17:31.117
simultaneously unmetered going
through the dev flow without a

00:17:31.200 --> 00:17:31.587
hitch.

00:17:31.670 --> 00:17:38.390
That's huge.

00:17:38.490 --> 00:17:41.177
So I know you are going to love
what the team has been working on.

00:17:41.260 --> 00:17:45.277
We hope it gives you a glimpse
of what's possible on windows

00:17:45.360 --> 00:17:46.867
today and where we are headed next.

00:17:46.950 --> 00:17:47.397
Thank you.

00:17:47.480 --> 00:17:54.480
Back to you Satya.

00:17:58.010 --> 00:18:02.687
>> That's the beginning
of unmetered intelligence having

00:18:02.770 --> 00:18:07.337
the models and agents using
the models work in parallel to what

00:18:07.420 --> 00:18:12.037
you are doing in the cloud. That's what
the platform enables as first class.

00:18:12.120 --> 00:18:16.947
Let's move to the cloud. The driving
equation for us remains the same, which is

00:18:17.030 --> 00:18:20.997
tokens per dollar per watt.
How do we optimize around this?

00:18:21.080 --> 00:18:24.617
When we think about the systems
problem we think about electrons

00:18:24.700 --> 00:18:28.187
coming on one end and tokens
on the other end and how do we

00:18:28.270 --> 00:18:31.997
think about the systems
optimization end-to-end.

00:18:32.080 --> 00:18:34.927
It starts with the data
center design itself.

00:18:35.010 --> 00:18:37.487
The core compute, storage, network.

00:18:37.570 --> 00:18:41.537
All of the accelerators that are
going to accelerate each of the

00:18:41.620 --> 00:18:45.907
components. How do you
think about the DC to

00:18:45.990 --> 00:18:51.047
DC connectivity and networking as well as
the off load to something like the local

00:18:51.130 --> 00:18:54.297
compute. That's the sort
of systems challenge.

00:18:54.380 --> 00:18:57.857
But before we even get into
all of the systems and the

00:18:57.940 --> 00:19:01.727
technology and the innovation,
perhaps the most important

00:19:01.810 --> 00:19:07.227
design criteria for us is how do
we earn the permission from the

00:19:07.310 --> 00:19:10.917
communities in which we are
building these data centers.

00:19:11.000 --> 00:19:14.637
That's where these principles
ground us and focus us.

00:19:14.720 --> 00:19:17.677
How do we ensure that the DCs do

00:19:17.760 --> 00:19:22.707
not increase the electro city

00:19:22.790 --> 00:19:27.797
prices. Making sure we are using water

00:19:27.880 --> 00:19:30.517
use. Creating jobs
for the local residents.

00:19:30.600 --> 00:19:33.927
Adding to the tax base.
Making sure we are strengthening

00:19:34.010 --> 00:19:36.927
communities by investing
in had local training and the

00:19:37.010 --> 00:19:40.267
nonprofits in the area.

00:19:40.350 --> 00:19:42.027
Only when we live up to the

00:19:42.110 --> 00:19:46.337
principles do the hard work is

00:19:46.420 --> 00:19:49.867
when we innovate and build.
We have been doing a lot of data

00:19:49.950 --> 00:19:52.677
center build up.

00:19:52.760 --> 00:19:57.147
Today Azure spans more than 500

00:19:57.230 --> 00:20:01.337
data centers in the region.
We have the most expansive

00:20:01.420 --> 00:20:02.197
hyperscaler footprint out there.

00:20:02.280 --> 00:20:04.657
We have added more data center

00:20:04.740 --> 00:20:08.457
capacity in the last 18 months

00:20:08.540 --> 00:20:14.767
than the first decade of to put

00:20:14.850 --> 00:20:21.850
that in perspective.

00:20:23.560 --> 00:20:25.837
We built out the cloud

00:20:25.920 --> 00:20:31.077
infrastructure for hetero genius
workloads that span the enterprise.

00:20:31.160 --> 00:20:33.767
When you look at what we are
building when you think of all

00:20:33.850 --> 00:20:39.157
of the gigawatts that are going
to come on-line we have three

00:20:39.240 --> 00:20:42.927
dominant workloads. There is training,
inference and the agent run time.

00:20:43.010 --> 00:20:45.297
These are three dominant workloads.

00:20:45.380 --> 00:20:48.297
In fact when you look at fair

00:20:48.380 --> 00:20:51.847
water it was sort of our first

00:20:51.930 --> 00:20:55.767
AI super factory.
It spanned two regions, Georgia

00:20:55.850 --> 00:20:59.307
and Wisconsin.
The entire system was designed

00:20:59.390 --> 00:21:03.607
from the ground up for AI.
We worked in fact very closely

00:21:03.690 --> 00:21:05.547
with inn NVIDIA on this.

00:21:05.630 --> 00:21:08.167
It is a two-story architecture

00:21:08.250 --> 00:21:13.117
that lets us he goes --

00:21:13.200 --> 00:21:18.027
essentially put the number of GPU
densely with network access.

00:21:18.110 --> 00:21:20.097
Means you really got fantastic

00:21:20.180 --> 00:21:22.697
higher performance networking,

00:21:22.780 --> 00:21:25.647
lower latency and more band

00:21:25.730 --> 00:21:27.507
width across the entire cluster.

00:21:27.590 --> 00:21:29.807
We are rethinking even the power

00:21:29.890 --> 00:21:30.067
delivery.

00:21:30.150 --> 00:21:37.150
Pou do -- how do we deliver 100

00:21:38.500 --> 00:21:41.937
kilowatts her row with the
conversion loss to the grid to

00:21:42.020 --> 00:21:46.787
silicon. We took a new approach to it.

00:21:46.870 --> 00:21:48.587
Also changes with the cooling

00:21:48.670 --> 00:21:51.527
system and water.

00:21:51.610 --> 00:21:54.427
The tooling loop is filled and

00:21:54.510 --> 00:21:59.117
the data center can operate
effectively with zero water

00:21:59.200 --> 00:22:02.307
consumption.
The daily water usage over the

00:22:02.390 --> 00:22:05.137
course of the entire one year is

00:22:05.220 --> 00:22:07.347
equivalent to what the single

00:22:07.430 --> 00:22:13.210
restaurant would use.

00:22:13.320 --> 00:22:18.117
When it comes to the systems
in the silicon, we have a lot of

00:22:18.200 --> 00:22:20.427
choice. We have first party silicon, we

00:22:20.510 --> 00:22:24.527
have partner systems.

00:22:24.610 --> 00:22:27.367
We have the first cloud to bring

00:22:27.450 --> 00:22:32.717
up inn NVIDIA.

00:22:32.800 --> 00:22:34.907
We work close with AMD.

00:22:34.990 --> 00:22:38.377
Working with them on the NX

00:22:38.460 --> 00:22:42.007
generation AMD GPU.

00:22:42.090 --> 00:22:43.907
Meyer 200 is continuing to scale

00:22:43.990 --> 00:22:48.367
it is live in Arizona will

00:22:48.450 --> 00:22:53.857
deploy internationally.
It delivers 30% per tokens per

00:22:53.940 --> 00:22:56.757
dollar verses what's
is the leading GPU today.

00:22:56.840 --> 00:23:02.127
We have validated it with GPD55
and we are going to use that to

00:23:02.210 --> 00:23:03.747
power Microsoft 365 Copilot.

00:23:03.830 --> 00:23:09.037
When it comes to running these

00:23:09.120 --> 00:23:12.007
agents, interesting thing now it

00:23:12.090 --> 00:23:16.747
is no longer about having
AI accelerator or GPU.

00:23:16.830 --> 00:23:21.597
CPU is critical.
Ratios may be coming to one is

00:23:21.680 --> 00:23:24.957
to one. That's why we are innovating

00:23:25.040 --> 00:23:30.127
with cobalt.
We are announcing preview oOf Cob

00:23:30.210 --> 00:23:34.897
Cobalt00VMs. Next generation on generation
CPU designed for cloud native

00:23:34.980 --> 00:23:36.687
and agent workloads today.

00:23:36.770 --> 00:23:39.717
It is exciting to see Cobalt

00:23:39.800 --> 00:23:46.790
make progress as well.

00:23:46.890 --> 00:23:50.877
One thing we are trying to make
sure being optimized for the new

00:23:50.960 --> 00:23:53.297
agent workloads.

00:23:53.380 --> 00:23:57.787
Cobalt delivers 30% better than

00:23:57.870 --> 00:24:01.957
Cobalt on cloud native.
Using GitHub Copilot traces

00:24:02.040 --> 00:24:06.127
these agentic traces to see.
The call patterns are so

00:24:06.210 --> 00:24:06.427
different.

00:24:06.510 --> 00:24:09.367
We are seeing 33% lower latency

00:24:09.450 --> 00:24:14.267
for the agent calls, 14% faster
speed, 23% higher through put.

00:24:14.350 --> 00:24:18.837
This is about the co design
of both the AI accelerator and the

00:24:18.920 --> 00:24:23.007
CPU for the agent.
And of course, when you talk

00:24:23.090 --> 00:24:27.437
about AI workloads you need scale,
you need reliability and

00:24:27.520 --> 00:24:29.397
that's why the network becomes

00:24:29.480 --> 00:24:32.817
super critical.

00:24:32.900 --> 00:24:35.427
We have innovated with the

00:24:35.510 --> 00:24:38.197
architecture and rebuilt how

00:24:38.280 --> 00:24:40.557
traffic across Azure moves to

00:24:40.640 --> 00:24:43.487
support workloads which is

00:24:43.570 --> 00:24:48.837
synchronous data workloads.
They span 10s of thousands of

00:24:48.920 --> 00:24:52.207
GPUs. You need to have them coherent.

00:24:52.290 --> 00:24:53.877
That's what the next frontier is

00:24:53.960 --> 00:24:55.877
to make sure we are able to keep

00:24:55.960 --> 00:25:02.960
scaling the network

00:25:04.500 --> 00:25:06.217
architearchitectur architecture.

00:25:06.300 --> 00:25:08.387
It is not just about outside
it is inside the data centers.

00:25:08.470 --> 00:25:12.457
It is connected through our

00:25:12.540 --> 00:25:17.887
continent spanning AI with the

00:25:17.970 --> 00:25:19.887
truly fungible compute fabric.

00:25:19.970 --> 00:25:22.507
All of this innovation is

00:25:22.590 --> 00:25:25.977
exciting, but when you think

00:25:26.060 --> 00:25:29.907
about innovation designing
for AI and I would say deep

00:25:29.990 --> 00:25:32.747
understanding of systems,

00:25:32.830 --> 00:25:37.657
there's no better company than

00:25:37.740 --> 00:25:41.687
NVIDIA and there's no other
person than Jensen to talk about

00:25:41.770 --> 00:25:43.197
it.

00:25:43.280 --> 00:25:50.280
I want to invite Jensen Wang

00:25:55.250 --> 00:25:59.877
from NVIDIA. I know it is late for you in

00:25:59.960 --> 00:26:00.377
Taipei.

00:26:00.460 --> 00:26:04.717
I appreciate you staying up.

00:26:04.800 --> 00:26:11.800
I have been looking at social

00:26:15.310 --> 00:26:18.427
and people talking about your keynote.

00:26:18.510 --> 00:26:23.097
The concept of unmetered intelligence
right at the edge is so hot again.

00:26:23.180 --> 00:26:25.347
You have thought about this,

00:26:25.430 --> 00:26:26.837
talked about this, now of course

00:26:26.920 --> 00:26:31.277
with RTX Spark really delivered

00:26:31.360 --> 00:26:33.377
what is a break through system

00:26:33.460 --> 00:26:38.047
for AI to be much Moreau --

00:26:38.130 --> 00:26:42.167
more ubiquitous. Maybe you can
share where you see this going?

00:26:42.250 --> 00:26:45.657
>> It all started about three
years ago with a conversation

00:26:45.740 --> 00:26:49.287
between you and I.
We were talking about how we

00:26:49.370 --> 00:26:53.767
could build a new class of PCs
incredible for designers and

00:26:53.850 --> 00:26:57.947
creators and it would be incredible
for artificial intelligence.

00:26:58.030 --> 00:27:01.407
It would be one of these
systems that has the processing

00:27:01.490 --> 00:27:04.037
capability but also the software

00:27:04.120 --> 00:27:09.007
stack that is integrated into
the world's design packages and

00:27:09.090 --> 00:27:12.177
creator packages and all
of the things we are doing with AI.

00:27:12.260 --> 00:27:13.897
Here we are three years later we

00:27:13.980 --> 00:27:16.717
built an incredible new chip,

00:27:16.800 --> 00:27:20.987
and this system is supported by

00:27:21.070 --> 00:27:26.057
all of this new software you

00:27:26.140 --> 00:27:30.747
created for Windows.
We have autonomous agent running

00:27:30.830 --> 00:27:35.837
on the PC. When you step back
and think about what does that mean?

00:27:35.920 --> 00:27:37.007
For 40 years or some 30 years we

00:27:37.090 --> 00:27:40.747
have been working together we

00:27:40.830 --> 00:27:43.147
went from inventing direct X

00:27:43.230 --> 00:27:46.347
together to creating now this

00:27:46.430 --> 00:27:50.787
incredible computer that has
autonomous systems running.

00:27:50.870 --> 00:27:54.927
The PC evolved from being
an incredible tool to now being a

00:27:55.010 --> 00:27:58.527
tool that is used autonomously

00:27:58.610 --> 00:28:04.217
by an AI assistant.

00:28:04.300 --> 00:28:06.247
The idea I could be traveling

00:28:06.330 --> 00:28:11.877
and on the phone and text my PC
and ask my PC to get coding done

00:28:11.960 --> 00:28:15.207
or an idea I have and it would
fire up the tools on the PC and

00:28:15.290 --> 00:28:18.477
it would make the modifications
or the changes or the design

00:28:18.560 --> 00:28:22.667
that I told it to do, and it
would iterate with me while I am

00:28:22.750 --> 00:28:24.587
away from the PC.

00:28:24.670 --> 00:28:30.057
My PC became an assistant.

00:28:30.140 --> 00:28:33.357
While I am sitting there it
would my great system as well.

00:28:33.440 --> 00:28:35.827
The idea that the PC evolved

00:28:35.910 --> 00:28:37.377
from a personal computer to

00:28:37.460 --> 00:28:44.460
personal AI is really exciting.

00:28:44.560 --> 00:28:48.107
To see it come to life
and actually doing that, I am super

00:28:48.190 --> 00:28:51.747
excited about it.
Spark you mentioned earlier has

00:28:51.830 --> 00:28:53.797
all of these capabilities Peta

00:28:53.880 --> 00:28:55.597
flop of AI performance.

00:28:55.680 --> 00:29:01.987
Has a Peta flop of MVFP4

00:29:02.070 --> 00:29:04.127
numerical format our companies
worked on together that allows

00:29:04.210 --> 00:29:07.097
it to take advantage of the 128

00:29:07.180 --> 00:29:11.787
giga bites of memory and fit
maybe a couple hundred billion

00:29:11.870 --> 00:29:16.507
parameter model. A couple hundred billion
parameter model is state of the art.

00:29:16.590 --> 00:29:20.037
The days of having a really
smart assistant running on the

00:29:20.120 --> 00:29:23.207
PC is here.

00:29:23.290 --> 00:29:24.877
>> It is awesome.

00:29:24.960 --> 00:29:27.497
I am excited about Windows

00:29:27.580 --> 00:29:31.287
coming to the GP300.

00:29:31.370 --> 00:29:35.787
It is a data center on your laptop.

00:29:35.870 --> 00:29:37.587
Talking about the data center side.

00:29:37.670 --> 00:29:40.697
This entire thing got started

00:29:40.780 --> 00:29:47.397
when we built the first super
computer to train the GPU

00:29:47.480 --> 00:29:49.037
models. I was talking about the fair

00:29:49.120 --> 00:29:52.677
water design, this custom build

00:29:52.760 --> 00:29:56.477
for the grace black water era
to Max the data center design with

00:29:56.560 --> 00:29:58.247
the system design you had.

00:29:58.330 --> 00:30:04.557
Now of course we are validating bel

00:30:04.640 --> 00:30:06.817
bel belly --.

00:30:06.900 --> 00:30:12.197
Maybe you can share on the clout
side how you are sharing the system.

00:30:12.280 --> 00:30:16.747
>> Our journey is incredible.
We built the first AI super

00:30:16.830 --> 00:30:20.537
computer together.
Hopper was an incredible success.

00:30:20.620 --> 00:30:23.507
The first two generations were

00:30:23.590 --> 00:30:27.847
focused on free training.
Grace Blackwell came along and

00:30:27.930 --> 00:30:31.147
all of the focus moved to post
training, reinforcement learning

00:30:31.230 --> 00:30:34.297
which allowed us to have reasoning models.

00:30:34.380 --> 00:30:38.687
These reasoning models based
on mixture of experts were

00:30:38.770 --> 00:30:42.027
incredibly intelligent, energy efficient.

00:30:42.110 --> 00:30:46.457
It requires giant systems.
We required MV length 72.

00:30:46.540 --> 00:30:49.067
The entire rack became one computer.

00:30:49.150 --> 00:30:51.077
We had evolved from one node to

00:30:51.160 --> 00:30:55.037
now one rack.

00:30:55.120 --> 00:30:56.737
Well, Microsoft deployed the

00:30:56.820 --> 00:30:58.837
largest number of grace

00:30:58.920 --> 00:31:02.647
Blackwells in the world today.

00:31:02.730 --> 00:31:03.947
Largest number of grace

00:31:04.030 --> 00:31:07.627
Blackwells in the world.

00:31:07.710 --> 00:31:09.897
Fair water is a magnificent

00:31:09.980 --> 00:31:13.357
system to look at.

00:31:13.440 --> 00:31:17.357
It is an incredible feet.
It is liquid cooled.

00:31:17.440 --> 00:31:20.897
Something you mentioned it
is close looped, basically uses

00:31:20.980 --> 00:31:25.447
almost no water.
It is incredibly environmentally

00:31:25.530 --> 00:31:29.067
friendly energy efficient.
We are able to increase the

00:31:29.150 --> 00:31:31.407
token generation rate and reduce

00:31:31.490 --> 00:31:36.277
the cost of token generation
by an order of magnitude, some 30

00:31:36.360 --> 00:31:39.377
times over hopper.
That was a huge achievement.

00:31:39.460 --> 00:31:45.157
Vera Ruben was created for a
world where these AIs are now

00:31:45.240 --> 00:31:48.087
agentic. As hopper was created for

00:31:48.170 --> 00:31:49.687
pretraining grace Blackwell for

00:31:49.770 --> 00:31:55.527
training, post training and also

00:31:55.610 --> 00:31:59.247
inference. Vera Rubin is designed to run

00:31:59.330 --> 00:32:03.307
for agents.
It is the same computing pattern

00:32:03.390 --> 00:32:06.237
when run on the RTX Spark.

00:32:06.320 --> 00:32:08.277
It is the same agentic system

00:32:08.360 --> 00:32:13.017
but of course larger win process

00:32:13.100 --> 00:32:19.617
enormous number of them. Many will be
from different customers and different

00:32:19.700 --> 00:32:22.587
partners. The entire path, coding path

00:32:22.670 --> 00:32:26.797
from storage which is the
long-term memory, working memory

00:32:26.880 --> 00:32:27.697
is encrypted.

00:32:27.780 --> 00:32:31.747
The data is encrypted in transit

00:32:31.830 --> 00:32:38.037
and data also incren crepted in

00:32:38.120 --> 00:32:41.277
encrypted in use.
This entire disaggregated

00:32:41.360 --> 00:32:43.747
distributed computing system you

00:32:43.830 --> 00:32:46.417
mentioned CPU Vera is a CPU

00:32:46.500 --> 00:32:50.957
designed for agents.
In the past CPUs were designed

00:32:51.040 --> 00:32:55.157
for humans. We are more
parent than agents are.

00:32:55.240 --> 00:32:57.887
Agents want low latency.

00:32:57.970 --> 00:33:00.057
Vera was designed for extremely

00:33:00.140 --> 00:33:03.747
low latency.

00:33:03.830 --> 00:33:07.567
Vera Rubin I can't wait to show everybody.

00:33:07.650 --> 00:33:13.077
You stood it up.
Our two teams have been working

00:33:13.160 --> 00:33:16.807
closely. Long before the chips taped out

00:33:16.890 --> 00:33:21.297
long before the systems brought
up the two teams were almost

00:33:21.380 --> 00:33:23.517
completely aligned.

00:33:23.600 --> 00:33:28.427
Data centers created for Vera Rubin.

00:33:28.510 --> 00:33:30.827
It is into the networking
and stack and community.

00:33:30.910 --> 00:33:34.357
The moment systems were rolling
off the lines they were building

00:33:34.440 --> 00:33:39.247
stood up at Microsoft. So incredibly
excited about the collaboration.

00:33:39.330 --> 00:33:42.907
>> This speed of light execution
for the teams are fantastic to see.

00:33:42.990 --> 00:33:46.337
All of this is to power
the ecosystem around us.

00:33:46.420 --> 00:33:50.577
You and I having grown-up with
a PC the server and now with AI,

00:33:50.660 --> 00:33:55.537
have always thought about ultimately it is
about creating the opportunity for every

00:33:55.620 --> 00:33:58.487
developer, every organization
to build on the work that we do and

00:33:58.570 --> 00:34:00.487
the platforms we create.

00:34:00.570 --> 00:34:04.197
Speaking of that there's a lot

00:34:04.280 --> 00:34:08.467
of software that vid NVIDIA

00:34:08.550 --> 00:34:13.247
builds we will have the models
in foundry, tooling in foundry.

00:34:13.330 --> 00:34:16.337
Software will help us with
accelerating our workloads when

00:34:16.420 --> 00:34:20.307
it comes to the data warehouse.

00:34:20.390 --> 00:34:23.047
We will have the stuff in windows.

00:34:23.130 --> 00:34:28.367
Talk about the broader vision
of what does it mean as an

00:34:28.450 --> 00:34:31.617
opportunity. Everybody talks about the one

00:34:31.700 --> 00:34:36.757
model and one piece of tech.
It is about the broadest biggest

00:34:36.840 --> 00:34:41.567
opportunity maybe you want
to share a little bit about that.

00:34:41.650 --> 00:34:43.567
>> We have been preparing for this moment.

00:34:43.650 --> 00:34:47.087
What happened the last several
months we have been working a

00:34:47.170 --> 00:34:48.967
decade and a half together
getting ready for really what

00:34:49.050 --> 00:34:50.467
happened in the last several months.

00:34:50.550 --> 00:34:54.807
All of a sudden because
of agentic systems the convergence

00:34:54.890 --> 00:34:57.307
of these great models AI is now

00:34:57.390 --> 00:34:57.547
useful.

00:34:57.630 --> 00:35:03.067
If you look at GitHub, the

00:35:03.150 --> 00:35:07.887
commits into GitHub has
gone completely parabolic.

00:35:07.970 --> 00:35:11.657
In the last several months
the number of commits increase by a

00:35:11.740 --> 00:35:15.367
factor of three.
It is clear agentic systems are

00:35:15.450 --> 00:35:17.677
useful, it is doing productive

00:35:17.760 --> 00:35:23.237
work and also it is profitable

00:35:23.320 --> 00:35:25.637
as a result. The amount of command for

00:35:25.720 --> 00:35:29.177
compute between the usage of AI

00:35:29.260 --> 00:35:35.197
the contribution necessary
for agents it has gone through the

00:35:35.280 --> 00:35:38.357
roof. One thing we have been doing
together is making sure all of

00:35:38.440 --> 00:35:41.987
the tools agents are using are

00:35:42.070 --> 00:35:43.457
fully accelerated.

00:35:43.540 --> 00:35:45.997
Fabric is fully accelerated.

00:35:46.080 --> 00:35:49.597
We are accelerating data

00:35:49.680 --> 00:35:54.467
processing, SQL, Spark, semantic

00:35:54.550 --> 00:35:57.007
based, graph based we are going
to make sure all of the tools

00:35:57.090 --> 00:35:59.477
available on Azure are going to

00:35:59.560 --> 00:36:05.577
be fully GPU accelerated because
the agents are going to be impat

00:36:05.660 --> 00:36:09.837
impatient. The faster it can
iterate and generate tokens which are

00:36:09.920 --> 00:36:13.887
ultimately what the developers
both customers would like to do

00:36:13.970 --> 00:36:19.097
is generate a lot of tokens that are
profitable and highly intelligent.

00:36:19.180 --> 00:36:22.097
>> Thank you so much for the
partnership and leadership and

00:36:22.180 --> 00:36:25.717
innovation that you bring
to this entire ecosystem and really

00:36:25.800 --> 00:36:28.967
thrilled to be working closely
with you and the team and bring

00:36:29.050 --> 00:36:32.537
all of this to the developers
here and beyond and look forward

00:36:32.620 --> 00:36:36.047
to seeing what the next sort
of few months and the next year

00:36:36.130 --> 00:36:41.367
will bring in terns of the innovation
that gets built on top of the platform.

00:36:41.450 --> 00:36:45.917
Thank you again for joining
this late in the night from Taipei.

00:36:46.000 --> 00:36:47.757
>> Thank you so much Satya for

00:36:47.840 --> 00:36:53.310
your partnership and friendship.

00:36:53.610 --> 00:36:57.177
>> So so far we have talked

00:36:57.260 --> 00:37:02.537
about the edge and the cloud.
The form factors, when I saw

00:37:02.620 --> 00:37:07.177
that Jensen picture from the weekend
where he had all of the desktops.

00:37:07.260 --> 00:37:09.577
I felt man I am back in the 90s.

00:37:09.660 --> 00:37:13.277
It is so tool to see the lineup
of all of the machines that I

00:37:13.360 --> 00:37:17.617
loved and grew up with back yet
again with new functionality.

00:37:17.700 --> 00:37:20.217
The same form factor, but

00:37:20.300 --> 00:37:25.127
unbelievable new functionality,
because of the on board AI capability.

00:37:25.210 --> 00:37:30.677
That is sort of what we have seen with the
laptop, the desktop and of course with the

00:37:30.760 --> 00:37:33.887
cloud. But it also, you know, sets up

00:37:33.970 --> 00:37:38.577
that next question.
If you have that capability,

00:37:38.660 --> 00:37:42.207
which is new function and you
can put it into existing form

00:37:42.290 --> 00:37:45.877
fact factors, can you even purpose

00:37:45.960 --> 00:37:49.797
build new form factors
for the new function?

00:37:49.880 --> 00:37:52.857
Can you build a new platform

00:37:52.940 --> 00:37:54.687
even for the agent era?

00:37:54.770 --> 00:37:56.987
And that is the motivation

00:37:57.070 --> 00:38:02.827
behind project Solara.

00:38:02.910 --> 00:38:07.047
I wanted to invite Stevie
on stage but first let's roll the

00:38:07.130 --> 00:38:13.377
video.

00:38:13.460 --> 00:38:15.747
>> First there's you.
Then everything in front of you.

00:38:15.830 --> 00:38:16.947
We get it.

00:38:17.030 --> 00:38:22.667
The noise, the weight of it all.

00:38:22.750 --> 00:38:29.750
We have been building something.

00:38:30.170 --> 00:38:31.527
For an agent first world.

00:38:31.610 --> 00:38:33.027
New agent first devices.

00:38:33.110 --> 00:38:38.317
You come close, it is ready when

00:38:38.400 --> 00:38:40.387
you are. You speak, it understands you.

00:38:40.470 --> 00:38:44.437
What matters comes forward.

00:38:44.520 --> 00:38:46.837
Everything else falls away.

00:38:46.920 --> 00:38:49.777
When you want it to keep going

00:38:49.860 --> 00:38:56.860
it keeps going.

00:38:57.180 --> 00:38:59.987
Sometimes it sits with you.
Sometimes it goes with you.

00:39:00.070 --> 00:39:04.597
Sometimes it sees what you want it to see.

00:39:04.680 --> 00:39:08.927
It doesn't decide for you.
It lights the path.

00:39:09.010 --> 00:39:11.347
Tries to clear the way
when you need it to.

00:39:11.430 --> 00:39:15.567
That's the idea.

00:39:15.650 --> 00:39:19.707
A whole ecosystem built to clear

00:39:19.790 --> 00:39:25.017
the way.

00:39:25.100 --> 00:39:29.767
Wherever you are, whenever it

00:39:29.850 --> 00:39:33.617
matters. >> I am so excited to be here.

00:39:33.700 --> 00:39:36.057
It really is great
to be back on this stage.

00:39:36.140 --> 00:39:37.527
Before I talk about those

00:39:37.610 --> 00:39:39.527
awesome new devices you just

00:39:39.610 --> 00:39:43.367
saw, let me start with the why.

00:39:43.450 --> 00:39:48.037
Back at build 2023 I talked

00:39:48.120 --> 00:39:50.907
about the outside AI

00:39:50.990 --> 00:39:51.577
infrastructure.

00:39:51.660 --> 00:39:55.747
It moves within the application

00:39:55.830 --> 00:40:00.577
frame to operating globally to

00:40:00.660 --> 00:40:02.167
connect maintain context across

00:40:02.250 --> 00:40:04.587
entire workflows, devices and

00:40:04.670 --> 00:40:08.987
time scales.

00:40:09.070 --> 00:40:11.757
What if there were an ecosystem

00:40:11.840 --> 00:40:16.667
of devices specifically designed
for the new type of application

00:40:16.750 --> 00:40:18.747
structure for those types of

00:40:18.830 --> 00:40:22.567
agents for that transaction.

00:40:22.650 --> 00:40:24.837
That's the impetus behind

00:40:24.920 --> 00:40:29.747
project Solara.
With many possible forms which

00:40:29.830 --> 00:40:32.147
one do you pick? What is the next device.

00:40:32.230 --> 00:40:37.367
The big uh-huh for us is that it
is not choosing one form factor

00:40:37.450 --> 00:40:39.187
it is about creating a system

00:40:39.270 --> 00:40:43.757
that extends your agent across a

00:40:43.840 --> 00:40:47.027
cons constellation of devices.

00:40:47.110 --> 00:40:51.347
The next computer is all of the
devices working together as one

00:40:51.430 --> 00:40:55.767
system with agents cloeing up
closer to where and when you

00:40:55.850 --> 00:40:57.137
need them.

00:40:57.220 --> 00:40:59.707
To realize this vision two

00:40:59.790 --> 00:41:03.777
challenges immediately show up.

00:41:03.860 --> 00:41:07.047
First many specialized form
factors already exist but often

00:41:07.130 --> 00:41:11.287
rely on custom one off apps
and fragmented stacks that are

00:41:11.370 --> 00:41:16.087
difficult and expensive to build,
deploy and maintain.

00:41:16.170 --> 00:41:18.857
Second across every industry
people and organizations are

00:41:18.940 --> 00:41:22.997
already building their own
agents deeply specialized and

00:41:23.080 --> 00:41:27.637
instrumented for their work.
The impact of those agents is

00:41:27.720 --> 00:41:28.467
constrained by how and where

00:41:28.550 --> 00:41:31.637
they can exist.

00:41:31.720 --> 00:41:35.547
Project Solara does both by

00:41:35.630 --> 00:41:39.047
extending agents on to new
purpose built easy to manage

00:41:39.130 --> 00:41:41.497
form factors designed to reach

00:41:41.580 --> 00:41:46.787
the nooks and crannies where
project creators do not exist or

00:41:46.870 --> 00:41:50.187
are not optimal.
It is a turnkey solution for

00:41:50.270 --> 00:41:55.727
building unique agent first
devices enabled by three

00:41:55.810 --> 00:41:58.117
pillars. First it is enterprise ready.

00:41:58.200 --> 00:42:01.067
Enabled by the AOSP based

00:42:01.150 --> 00:42:06.237
Microsoft device ecosystem platform.

00:42:06.320 --> 00:42:09.307
Second it has an agent driven
interaction model with just in

00:42:09.390 --> 00:42:13.217
time UI that adapts to the form factor.

00:42:13.300 --> 00:42:17.137
Third it has extensibility so
you can bring your own agents.

00:42:17.220 --> 00:42:22.287
Tying it all together is Azure unifying
the system across cloud and device.

00:42:22.370 --> 00:42:26.427
That's enough of that.
Now let's talk about the

00:42:26.510 --> 00:42:30.127
devices. Today we are previewing
two very broad categories.

00:42:30.210 --> 00:42:33.087
The first is stationary
and the second is port able.

00:42:33.170 --> 00:42:36.067
The first device
is designed for your desk.

00:42:36.150 --> 00:42:39.737
It is built on media tech silicon.

00:42:39.820 --> 00:42:43.207
With hello for business walking
up to the device securely signs

00:42:43.290 --> 00:42:46.517
you in giving you direct
access to your agent.

00:42:46.600 --> 00:42:49.697
Just like Nathan
is about to show you here.

00:42:49.780 --> 00:42:53.017
For the information work this

00:42:53.100 --> 00:42:55.187
means frictionless yet protected

00:42:55.270 --> 00:42:58.927
access to Microsoft 365 Copilot

00:42:59.010 --> 00:43:02.797
grounded in work IQ.
What matters next in your

00:43:02.880 --> 00:43:06.947
workday helping you think,
plan and even act by delegating task

00:43:07.030 --> 00:43:09.267
to your agent with a simple tap

00:43:09.350 --> 00:43:16.247
or just using your voice.

00:43:16.330 --> 00:43:19.777
Dedicated Ambian device for work.

00:43:19.860 --> 00:43:25.787
It acts as campaign onto your
existing windows Copilot Plus

00:43:25.870 --> 00:43:27.187
-- PC.

00:43:27.270 --> 00:43:31.757
Or access windows 365 in a

00:43:31.840 --> 00:43:34.367
connected monitor.

00:43:34.450 --> 00:43:39.730
How cool is that?

00:43:39.830 --> 00:43:43.767
The second is port able. Re-imagining
a wear able millions of people use every

00:43:43.850 --> 00:43:47.307
day. The access badge.

00:43:47.390 --> 00:43:49.977
Built using Qualcomm silicon

00:43:50.060 --> 00:43:54.737
wear able this badge is light
weight form factor designed for

00:43:54.820 --> 00:43:59.257
agents on the go.
Adapt able across variety of

00:43:59.340 --> 00:44:01.117
verti

00:44:01.200 --> 00:44:02.157
verticals and workflows.

00:44:02.240 --> 00:44:08.327
I have an early prototype of the

00:44:08.410 --> 00:44:10.877
badge. Using my finger print I tapped

00:44:10.960 --> 00:44:16.937
to unlock the device now I have
all of my agents in a secure

00:44:17.020 --> 00:44:22.247
manner. Gather content
for social media post for today.

00:44:22.330 --> 00:44:23.547
Why not do it right now.

00:44:23.630 --> 00:44:26.277
I am going to hit record and the

00:44:26.360 --> 00:44:31.217
device's camera is recording.

00:44:31.300 --> 00:44:34.527
Hope you don't mind I
am going to take shots.

00:44:34.610 --> 00:44:38.727
Find good shots from this clean
them up and send them for me and

00:44:38.810 --> 00:44:42.767
my team to review. There you have it.

00:44:42.850 --> 00:44:44.947
Now my agent is off running

00:44:45.030 --> 00:44:47.767
through multiple tasks I clean

00:44:47.850 --> 00:44:52.977
them up and send them to me and the team.

00:44:53.060 --> 00:44:56.777
Pretty cool.
I know it is a simple demo but

00:44:56.860 --> 00:45:01.427
it is all agent driven and there are
so many verticals and opportunity.

00:45:01.510 --> 00:45:05.287
Imagine in healthcare from the
moment you pick up the device

00:45:05.370 --> 00:45:08.787
the right agent shapes
the experience around the roll and

00:45:08.870 --> 00:45:12.727
the workflow helping with check
ins, patient records and

00:45:12.810 --> 00:45:17.267
critical insights. All through
enterprise grade secure access.

00:45:17.350 --> 00:45:21.167
And with the built
in microphones the nurse can start

00:45:21.250 --> 00:45:25.377
a hands free voice based
documentation including

00:45:25.460 --> 00:45:27.677
dirizezation and and notation.

00:45:27.760 --> 00:45:32.547
The side phasing camera can
help with asht vitals or scan

00:45:32.630 --> 00:45:34.897
medications or help verify workflows.

00:45:34.980 --> 00:45:38.787
This is a few examples of how
the small purpose built wear

00:45:38.870 --> 00:45:41.927
able can bring intelligence
directly into the flow of

00:45:42.010 --> 00:45:46.257
patient care helping nurses
access and gather and even act

00:45:46.340 --> 00:45:47.497
on information while staying

00:45:47.580 --> 00:45:54.580
print with the patient.

00:45:55.650 --> 00:46:00.337
Both devices are an expression
of devices the core hardware and

00:46:00.420 --> 00:46:02.777
software are designed
to be highly flexible.

00:46:02.860 --> 00:46:07.647
With a few changes loading
a different agent adjusting the

00:46:07.730 --> 00:46:10.917
shape, screen size,
sensors or input methods the same

00:46:11.000 --> 00:46:14.827
foundation, the same software
can be adapted for many

00:46:14.910 --> 00:46:19.197
verticals and workflows such
as retail, industrial, hospitality,

00:46:19.280 --> 00:46:20.997
financial service, legal and so forth.

00:46:21.080 --> 00:46:22.997
That is the power of the

00:46:23.080 --> 00:46:28.407
platform, that flexibility.

00:46:28.490 --> 00:46:33.377
Whatever scenario there are
places where computing has not

00:46:33.460 --> 00:46:33.907
naturally fit before.

00:46:33.990 --> 00:46:36.107
While this is an early look, we

00:46:36.190 --> 00:46:42.237
are really excited that ak ue

00:46:42.320 --> 00:46:46.687
whether, best buy, target
and others are working towards

00:46:46.770 --> 00:46:49.727
exploring how devices can
improve their workflows.

00:46:49.810 --> 00:46:54.157
This is the broader
opportunity for the ecosystem.

00:46:54.240 --> 00:46:58.547
Agents moving outside of the app
and taking shape in devices

00:46:58.630 --> 00:47:00.147
designed for specific scenario,

00:47:00.230 --> 00:47:02.767
a specific customer and a

00:47:02.850 --> 00:47:05.937
specific place.

00:47:06.020 --> 00:47:10.077
For all of you this is the
moment to imagine where your

00:47:10.160 --> 00:47:12.247
agent should live, what form

00:47:12.330 --> 00:47:14.377
they should take and what new

00:47:14.460 --> 00:47:19.387
work they can unlock.

00:47:19.470 --> 00:47:24.057
Last week Satya sat down with

00:47:24.140 --> 00:47:28.557
Qualcomm, Chrissian yaw to talk

00:47:28.640 --> 00:47:33.487
about the future. >> Incredible to have
you at device lab and talk about the

00:47:33.570 --> 00:47:35.767
reference designs. >> Happy to be here.

00:47:35.850 --> 00:47:38.737
I have been to the device
lab a lot recently.

00:47:38.820 --> 00:47:41.677
>> One of the things that you
and I have chatted for a while

00:47:41.760 --> 00:47:44.617
is how there's a real platform shift.

00:47:44.700 --> 00:47:49.937
We are moving from building
operating systems, devices for

00:47:50.020 --> 00:47:53.217
apps for agents.
You want to talk about how you

00:47:53.300 --> 00:47:57.087
see this?

00:47:57.170 --> 00:48:01.657
>> This is technology this
one seems to be a big one.

00:48:01.740 --> 00:48:03.827
Agents change the whole nature

00:48:03.910 --> 00:48:07.767
of the device in itself.

00:48:07.850 --> 00:48:11.577
Starting with the fact that if
AI understands the world as we

00:48:11.660 --> 00:48:14.737
understand it it will be closer
to our senses, closer to our

00:48:14.820 --> 00:48:19.447
eyes and mouth and ears.

00:48:19.530 --> 00:48:21.497
It will be things we wear.

00:48:21.580 --> 00:48:27.057
You need computing which
is geared towards real time

00:48:27.140 --> 00:48:29.957
context. And from silicon to cloud.

00:48:30.040 --> 00:48:34.097
>> This is one of the references
we have that I love.

00:48:34.180 --> 00:48:37.097
There is stuff happening 07b9 edge here.

00:48:37.180 --> 00:48:39.517
Things on the cloud.
You want to talk about the core

00:48:39.600 --> 00:48:41.837
systems and silicon implications

00:48:41.920 --> 00:48:47.077
for such a new ecosystem.
>> You need a power efficient

00:48:47.160 --> 00:48:51.247
CPU. The silicon is designed
for you to have a cloud native

00:48:51.330 --> 00:48:56.267
experience. You have a lot of sensors for

00:48:56.350 --> 00:48:56.797
context.

00:48:56.880 --> 00:49:01.857
It is personaized.

00:49:01.940 --> 00:49:06.557
It is a thank change
in the nature of devices.

00:49:06.640 --> 00:49:09.397
Wear able platform is changing itself.

00:49:09.480 --> 00:49:12.287
You see incredible factors.

00:49:12.370 --> 00:49:15.767
>> The computer platform, how do

00:49:15.850 --> 00:49:18.507
you build it such that there is

00:49:18.590 --> 00:49:22.847
an open ecosystem.

00:49:22.930 --> 00:49:25.817
It is not about one agent it is

00:49:25.900 --> 00:49:30.167
about any agent.
>> Because the smartphone today

00:49:30.250 --> 00:49:32.587
is at the center of your digital

00:49:32.670 --> 00:49:35.787
life, the job of those devices

00:49:35.870 --> 00:49:40.927
today is to be extending the

00:49:41.010 --> 00:49:42.097
functionality of the smartphone.

00:49:42.180 --> 00:49:45.747
You saw many platforms became vertical.

00:49:45.830 --> 00:49:50.207
It is a natural thing to have a vertical
platform from the same company.

00:49:50.290 --> 00:49:51.407
The phone was at the center.

00:49:51.490 --> 00:49:54.277
That changes dramatically when

00:49:54.360 --> 00:49:57.607
you think about agents.

00:49:57.690 --> 00:49:59.477
Agents becomes the center of

00:49:59.560 --> 00:50:02.767
your digital experience.

00:50:02.850 --> 00:50:07.757
They will be looking for an open
horizontal platform that enable

00:50:07.840 --> 00:50:09.857
the agents to be interacting

00:50:09.940 --> 00:50:12.397
with the best possible device

00:50:12.480 --> 00:50:16.127
for different applications.

00:50:16.210 --> 00:50:20.687
>> We want to make it possible
for anybody working with agentic

00:50:20.770 --> 00:50:25.777
system not just bound to devices
but imagine they can be many

00:50:25.860 --> 00:50:28.607
devices that carry that intelligence
in different context.

00:50:28.690 --> 00:50:32.477
That's the open ecosystem we

00:50:32.560 --> 00:50:36.067
want to build together.
It is great to partner with you

00:50:36.150 --> 00:50:38.847
to get this started.

00:50:38.930 --> 00:50:41.857
>> We are proud of this

00:50:41.940 --> 00:50:46.597
partnership and that's just the beginning.

00:50:46.680 --> 00:50:51.747
>> Thank you very much
Steven and Christiano.

00:50:51.830 --> 00:50:53.517
>> It is a new platform but it

00:50:53.600 --> 00:50:57.337
is a set of new platform rules

00:50:57.420 --> 00:51:01.537
that don't in some sense hem in

00:51:01.620 --> 00:51:05.347
what you can imagine the type of

00:51:05.430 --> 00:51:08.767
form factors where your agents live.

00:51:08.850 --> 00:51:13.457
Where the new platforms come you
get to rewrite even the rules of

00:51:13.540 --> 00:51:17.157
how new platforms operate.
That's what we are trying to get

00:51:17.240 --> 00:51:20.697
done with project Solara so you

00:51:20.780 --> 00:51:24.427
as developers and enterprises
have the flexibility to imagine

00:51:24.510 --> 00:51:27.387
the form factors that you want

00:51:27.470 --> 00:51:32.377
and have agents be ubiquitous.

00:51:32.460 --> 00:51:34.607
Let's go up the stack to the next layer.

00:51:34.690 --> 00:51:38.547
We are building a new
intelligence layer bringing

00:51:38.630 --> 00:51:40.477
together the models, context as

00:51:40.560 --> 00:51:45.587
well as the tools.

00:51:45.670 --> 00:51:49.087
It starts with model choice.
Every customer every developer

00:51:49.170 --> 00:51:51.457
is going to choose the right

00:51:51.540 --> 00:51:55.767
model for the right task and

00:51:55.850 --> 00:51:59.387
eval mix. Agency budget and foundry today

00:51:59.470 --> 00:52:02.407
has over 121 sthou models.

00:52:02.490 --> 00:52:07.307
The largest model catalog out there.

00:52:07.390 --> 00:52:12.177
From OpenAI to an thropic.
Even MAI models.

00:52:12.260 --> 00:52:14.147
Last week we brought the OpenAI

00:52:14.230 --> 00:52:16.237
real time voice models along

00:52:16.320 --> 00:52:19.517
with Claude Opus 4.8 to foundry.

00:52:19.600 --> 00:52:24.627
We are continuing to bring
these frontier models.

00:52:24.710 --> 00:52:29.467
The consideration about
the models is becoming increasingly

00:52:29.550 --> 00:52:33.217
key. The last developer counsel to

00:52:33.300 --> 00:52:36.707
now when you are building
the agentic system having this

00:52:36.790 --> 00:52:40.237
context really shaped right
is becoming super important.

00:52:40.320 --> 00:52:42.547
In fact it starts right at the

00:52:42.630 --> 00:52:43.147
data tier.

00:52:43.230 --> 00:52:47.547
The data to date has been built

00:52:47.630 --> 00:52:51.337
for application that is supported
user facing applications.

00:52:51.420 --> 00:52:53.817
Now you have to change build

00:52:53.900 --> 00:53:00.457
them for agents again different
call patterns even to the data

00:53:00.540 --> 00:53:03.997
tier. Agents are restoring,

00:53:04.080 --> 00:53:05.347
retrieving, reasoning, acting

00:53:05.430 --> 00:53:07.717
and learning.

00:53:07.800 --> 00:53:11.777
That is what is happening in a

00:53:11.860 --> 00:53:12.607
continuous loop.

00:53:12.690 --> 00:53:17.707
They are using Cosmos DB,

00:53:17.790 --> 00:53:21.347
ChatGPT does, Azure search

00:53:21.430 --> 00:53:26.217
retrieval of ind -- indices.

00:53:26.300 --> 00:53:28.957
They started the business logic

00:53:29.040 --> 00:53:31.887
tier effectively for agents.

00:53:31.970 --> 00:53:35.327
You have fabric real time

00:53:35.410 --> 00:53:38.517
intelligence which now brings me

00:53:38.600 --> 00:53:41.887
to an exciting new service

00:53:41.970 --> 00:53:45.407
Verizon DB which is fully

00:53:45.490 --> 00:53:51.847
managed SQL server on Azure.

00:53:51.930 --> 00:53:53.077
One of the things we wanted to

00:53:53.160 --> 00:53:56.767
make sure we built ground up

00:53:56.850 --> 00:53:59.757
Proest gres managed service for

00:53:59.840 --> 00:54:03.727
higher availability scaleout.

00:54:03.810 --> 00:54:08.327
It is automated fill over 128

00:54:08.410 --> 00:54:10.027
terra bites per cluster.

00:54:10.110 --> 00:54:15.937
The read heavy workloads.
You can scale with this managed

00:54:16.020 --> 00:54:16.707
service.

00:54:16.790 --> 00:54:20.237
In our internal testing we are

00:54:20.320 --> 00:54:24.407
seeing Verizon is having 3X

00:54:24.490 --> 00:54:29.397
through put compared to PostgreSQL.

00:54:29.480 --> 00:54:31.667
It is critical thinking
about the scale you need.

00:54:31.750 --> 00:54:34.887
The other data workload we are

00:54:34.970 --> 00:54:35.787
changing dramatically is the

00:54:35.870 --> 00:54:39.927
data warehouse in a world where

00:54:40.010 --> 00:54:44.627
agents are constantly querying

00:54:44.710 --> 00:54:47.487
data, data warehouse is critical.

00:54:47.570 --> 00:54:51.737
One thing to have it mission
critical for users but when

00:54:51.820 --> 00:54:54.277
agents need analysis done on the

00:54:54.360 --> 00:54:56.777
fly bringing GPU acceleration to

00:54:56.860 --> 00:55:03.437
fabric is super key.
We are seeing 7X performance

00:55:03.520 --> 00:55:08.340
gain. It is thrilling
to see that AI acceleration.

00:55:08.570 --> 00:55:11.657
So if you sort of have the data

00:55:11.740 --> 00:55:14.457
tier, the layer about this is

00:55:14.540 --> 00:55:17.867
the IQ layer that we are

00:55:17.950 --> 00:55:23.167
building that brings together essentially
the model capabilities to -- along with

00:55:23.250 --> 00:55:26.777
the data. You kind of mix the data and

00:55:26.860 --> 00:55:31.247
model capability so you can
deliver that right context to

00:55:31.330 --> 00:55:34.017
unlock intelligence.

00:55:34.100 --> 00:55:37.367
In fact when people talk about

00:55:37.450 --> 00:55:42.657
token efficiency this is
perhaps the most important

00:55:42.740 --> 00:55:45.197
consideration.
If you structure the context

00:55:45.280 --> 00:55:50.127
right and feed the models you
are by definition going to be so

00:55:50.210 --> 00:55:53.517
much more token efficient.
The first domain is the web.

00:55:53.600 --> 00:55:55.537
Web grounding is so important.

00:55:55.620 --> 00:55:58.837
You need that fresh, high

00:55:58.920 --> 00:56:03.307
quality and fast web data.

00:56:03.390 --> 00:56:07.577
And that's why we are really,
really excited to announce today

00:56:07.660 --> 00:56:12.100
web IQ.

00:56:12.200 --> 00:56:15.987
And web IQ is built on our

00:56:16.070 --> 00:56:19.697
global infrastructure that is
already serving over a billion

00:56:19.780 --> 00:56:26.117
users. But fundamentally
rearchitected for the LLM and agentic

00:56:26.200 --> 00:56:31.707
workflow. It is model
agnostic, MCP native.

00:56:31.790 --> 00:56:32.877
Plugs into any agent run time.

00:56:32.960 --> 00:56:36.077
It has web and news, images,

00:56:36.160 --> 00:56:38.777
video, so agents can ground

00:56:38.860 --> 00:56:42.027
responses in flesh, verify able

00:56:42.110 --> 00:56:45.557
content, and web IQ leads across

00:56:45.640 --> 00:56:47.917
all of the three key criteria.

00:56:48.000 --> 00:56:51.457
It is best in class in quality,
best in class in speed as well

00:56:51.540 --> 00:56:55.157
as in cost. So we are very, very thrilled

00:56:55.240 --> 00:56:57.727
about web IQ being in the

00:56:57.810 --> 00:57:02.807
developer's hands as you
build out agentic systems.

00:57:02.890 --> 00:57:04.707
We are not stopping beyond the web.

00:57:04.790 --> 00:57:08.237
Every developer wants to ground
their agents on what is the most

00:57:08.320 --> 00:57:10.877
valuable data across the enterprise.

00:57:10.960 --> 00:57:14.147
We are bringing together

00:57:14.230 --> 00:57:16.367
foundry, fabric and Microsoft

00:57:16.450 --> 00:57:21.587
365 as this unified IQ layer.

00:57:21.670 --> 00:57:25.387
Continuously updated understanding of your

00:57:25.470 --> 00:57:26.657
organization to show you all of

00:57:26.740 --> 00:57:30.097
this rich IQ in action building

00:57:30.180 --> 00:57:33.617
real life agents here is Elijah.

00:57:33.700 --> 00:57:36.850
Elijah take it away.

00:57:37.320 --> 00:57:39.707
>> Agents are only as good as

00:57:39.790 --> 00:57:41.677
the context we give them.

00:57:41.760 --> 00:57:44.577
Microsoft IQ unifies enterprise

00:57:44.660 --> 00:57:48.827
intelligence for every organization.

00:57:48.910 --> 00:57:52.987
I am here at a power utilities
control center and will start by

00:57:53.070 --> 00:57:57.487
running a long running agent.
This agent is going to help us

00:57:57.570 --> 00:58:01.327
assess the current grid
operations incident and produce

00:58:01.410 --> 00:58:04.527
a brief for us so we
can respond accordingly.

00:58:04.610 --> 00:58:06.887
Now I am going to go
ahead and kick that off.

00:58:06.970 --> 00:58:08.967
While that runs let me show you

00:58:09.050 --> 00:58:12.907
how we got here.

00:58:12.990 --> 00:58:16.747
We built our agent

00:58:16.830 --> 00:58:21.697
And its also wired to a
Foundry IQ knowledge base.

00:58:21.780 --> 00:58:26.817
A single grounded source that packages our
documents, operational data, and people

00:58:26.900 --> 00:58:29.117
into context the agent can reason over.

00:58:29.200 --> 00:58:30.987
After building the agent in

00:58:31.070 --> 00:58:36.297
Foundry, we published it
to Microsoft 365 for the whole

00:58:36.380 --> 00:58:38.647
team. Let's go see it in action.

00:58:38.730 --> 00:58:41.867
Now, here in M365, I will start
with a question about current

00:58:41.950 --> 00:58:43.907
events that an ungrounded LLM

00:58:43.990 --> 00:58:45.407
wouldn't be able to answer.

00:58:45.490 --> 00:58:49.747
I'm going to ask about current

00:58:49.830 --> 00:58:51.847
electricity prices in SF.

00:58:51.930 --> 00:58:54.977
Our agents pulls in the first IQ

00:58:55.060 --> 00:58:59.687
tool, Web IQ, search built for the AI era.

00:58:59.770 --> 00:59:03.487
It delivers industry-leading
quality, velocity, and

00:59:03.570 --> 00:59:03.787
efficiency.

00:59:03.870 --> 00:59:10.027
Web IQ constantly indexes fresh

00:59:10.110 --> 00:59:14.887
official sources and does
a great job with semantic

00:59:14.970 --> 00:59:18.407
documents. And it give us our answer.

00:59:18.490 --> 00:59:22.007
Grounded in reality.
Not only can we deploy our agent

00:59:22.090 --> 00:59:27.217
here, but we can embed it.
Let's head back to the control

00:59:27.300 --> 00:59:31.437
center. Let's how we handleeds a

00:59:31.520 --> 00:59:35.987
previous incident. Can he asked for --

00:59:36.070 --> 00:59:39.097
Deasked for details
at our at risk substations.

00:59:39.180 --> 00:59:42.627
The power of Microsoft IQ comes
when we combine that external

00:59:42.710 --> 00:59:45.987
knowledge with our own
internal enterprise context.

00:59:46.070 --> 00:59:50.567
For this information our agent
pulls in the next layer, Fabric

00:59:50.650 --> 00:59:52.907
IQ.

00:59:52.990 --> 00:59:55.777
Here is bright line's grid

00:59:55.860 --> 00:59:58.547
represented as a tab Rick on

00:59:58.630 --> 01:00:00.777
theology, an operational of the

01:00:00.860 --> 01:00:04.697
live grid, and we didn't build

01:00:04.780 --> 01:00:08.757
this from scratch.
It takes models used by millions

01:00:08.840 --> 01:00:12.297
of customers and lets teams
extend them into rich ontologies

01:00:12.380 --> 01:00:15.097
that helps run the business.
This model -- yes!

01:00:15.180 --> 01:00:20.117
This model is coupled with live
telemetry, so it reflects the

01:00:20.200 --> 01:00:23.237
real operational state
of the grid minute by minute.

01:00:23.320 --> 01:00:25.537
This is what Microsoft IQ means

01:00:25.620 --> 01:00:31.790
by enterprise intelligence.

01:00:31.890 --> 01:00:35.077
Not data scattered across
disparate systems, but a single

01:00:35.160 --> 01:00:38.537
living model of the business
that an agent can reason over.

01:00:38.620 --> 01:00:41.217
Now, turning back to our agent.

01:00:41.300 --> 01:00:46.097
We can see that it gave us a table
of our most exposed substations.

01:00:46.180 --> 01:00:49.867
That's the agent query
and the ontology you just saw.

01:00:49.950 --> 01:00:53.517
Now the agent has access to the
outside world and our COMBRID.

01:00:53.600 --> 01:00:58.337
The last -- and our grid. The last piece
turns the situation into a response.

01:00:58.420 --> 01:01:00.807
Our policies and our people.

01:01:00.890 --> 01:01:02.847
By asking, "What are the steps

01:01:02.930 --> 01:01:08.147
to respond to a substation
trip," we activate the final

01:01:08.230 --> 01:01:09.897
layer, Work IQ.

01:01:09.980 --> 01:01:15.257
This is Brightline's response
procedure in SharePoint.

01:01:15.340 --> 01:01:18.797
It's the play book the team reaches
for when something goes wrong.

01:01:18.880 --> 01:01:23.167
And the important thing,
the agent isn't working from a stale

01:01:23.250 --> 01:01:26.347
upload or copied snapshot.
It's answering from the same

01:01:26.430 --> 01:01:28.147
source the team maintains day to day.

01:01:28.230 --> 01:01:31.967
When the procedure changes,
the answer changes with it.

01:01:32.050 --> 01:01:34.637
No reuploads. No prompt rebuildings.

01:01:34.720 --> 01:01:38.547
And no stale versions.
And critically, this is your

01:01:38.630 --> 01:01:42.997
knowledge. The assets you create
stay with you, no matter what model or

01:01:43.080 --> 01:01:46.887
agent is reasoning over them.
Now, if we return to our agent,

01:01:46.970 --> 01:01:50.057
we can see the proper way
to respond to this issue.

01:01:50.140 --> 01:01:55.127
Not a generic recommendation,
but our play book applied to

01:01:55.210 --> 01:01:59.177
this incident. Now, we just saw one
situation, three questions, and one

01:01:59.260 --> 01:02:03.937
connected answer. Now let's go check
back in on our long-running agent.

01:02:04.020 --> 01:02:06.907
And... oh! Let's check the backup really

01:02:06.990 --> 01:02:08.137
quick.

01:02:08.220 --> 01:02:11.247
Moment of truth.

01:02:11.330 --> 01:02:15.077
And...

01:02:15.160 --> 01:02:20.800
Boom! Our task finished! Ha-ha!

01:02:20.900 --> 01:02:23.387
Here we can see every step the

01:02:23.470 --> 01:02:24.927
agent took.

01:02:25.010 --> 01:02:30.667
First, beginning with Web IQ,
connecting it on the outside

01:02:30.750 --> 01:02:34.187
world. Second, fabric IQ through
foundry, organicing it in the

01:02:34.270 --> 01:02:35.967
is -- anchoring it
in the state of operations.

01:02:36.050 --> 01:02:39.107
And work IQ, grounding
in people and procedures.

01:02:39.190 --> 01:02:42.177
I triggered this manually,
but with Foundry routines, this can

01:02:42.260 --> 01:02:46.777
run on a schedule, turning a one
off response into continuous,

01:02:46.860 --> 01:02:50.097
proactive execution.
And if we look closely, we can

01:02:50.180 --> 01:02:53.987
see it even use the power
of work IQ to alert me directly

01:02:54.070 --> 01:02:57.957
about the incident.
And if we go ahead and check

01:02:58.040 --> 01:03:00.557
Teams, it sent me a brief

01:03:00.640 --> 01:03:07.330
notifying me of the situation.

01:03:07.430 --> 01:03:10.037
That's the power of Microsoft

01:03:10.120 --> 01:03:16.307
IQ. When the crisis its, teams don't

01:03:16.390 --> 01:03:19.747
chance answer.
They get an answer all in one

01:03:19.830 --> 01:03:25.500
place they can trust. Back to you, Satya.

01:03:25.600 --> 01:03:27.187
>> All right.

01:03:27.270 --> 01:03:29.617
Thank you, Elijah.

01:03:29.700 --> 01:03:33.197
And now, you know, let's move up

01:03:33.280 --> 01:03:37.797
from this context model layer
to deploying these agents and

01:03:37.880 --> 01:03:41.387
thinking about the run time.
And, you know, when you're

01:03:41.470 --> 01:03:46.077
building a first-class agentic
system, you need a first-class

01:03:46.160 --> 01:03:49.577
agent run time and a platform
that we are going to ship both

01:03:49.660 --> 01:03:53.947
with Windows as well
as part of Foundry in Azure.

01:03:54.030 --> 01:03:55.927
We want Windows to be a

01:03:56.010 --> 01:03:59.917
fantastic place to run and scale

01:04:00.000 --> 01:04:01.957
agents, and agents effectively

01:04:02.040 --> 01:04:03.857
are a new execution environment.

01:04:03.940 --> 01:04:07.257
Right? It's a new paradigm, even.

01:04:07.340 --> 01:04:12.917
They reason continuously.
They generate and run code

01:04:13.000 --> 01:04:17.007
dynamically. They take action across
files and devices, as well as across

01:04:17.090 --> 01:04:21.637
the network. Obviously
there's a lot of power in it.

01:04:21.720 --> 01:04:23.877
Right? The fact that it can generate

01:04:23.960 --> 01:04:29.597
code and act on it,
on a long-running agent that's autonomous.

01:04:29.680 --> 01:04:31.097
But obviously it creates new risk.

01:04:31.180 --> 01:04:33.887
And that's why today we're

01:04:33.970 --> 01:04:35.487
introducing Microsoft execution

01:04:35.570 --> 01:04:38.387
containers, or MXC, which is a

01:04:38.470 --> 01:04:41.827
new policy layer that lets

01:04:41.910 --> 01:04:44.927
Windows apply isolation and

01:04:45.010 --> 01:04:49.267
containment, using AI native
-- OS native primitives.

01:04:49.350 --> 01:04:54.077
You need to bake this into
the operating system, so that the

01:04:54.160 --> 01:04:56.577
containment is enforced by policy.

01:04:56.660 --> 01:05:01.117
You can have process level isolation
for lightweight agent actions.

01:05:01.200 --> 01:05:02.797
You can have session-level

01:05:02.880 --> 01:05:05.937
isolation for user separation.

01:05:06.020 --> 01:05:09.887
So Windows and Linux machines,

01:05:09.970 --> 01:05:13.557
including WSL,
for much stronger boundaries.

01:05:13.640 --> 01:05:16.897
And in fact, if you want full
isolation and containment,

01:05:16.980 --> 01:05:20.717
Windows 365 for agents,
for maximum isolation in a separate

01:05:20.800 --> 01:05:23.097
managed environment, effectively.

01:05:23.180 --> 01:05:28.007
You can pick the right
containment option for the work

01:05:28.090 --> 01:05:30.147
load, and windows will enforce it.

01:05:30.230 --> 01:05:33.177
And I think this becomes pretty
critical as you think about

01:05:33.260 --> 01:05:35.767
deploying agents at scale on

01:05:35.850 --> 01:05:40.587
your Windows desktop.
We want to ensure containment is

01:05:40.670 --> 01:05:43.257
enforced, of course,
regardless of who builds the agent.

01:05:43.340 --> 01:05:47.427
This is why you want to bake
it into the operating system.

01:05:47.510 --> 01:05:52.147
On to that end we're working
with many partners to ensure

01:05:52.230 --> 01:05:56.067
that taint, we're supporting
real developer workloads out

01:05:56.150 --> 01:05:58.537
there and addresses the needs.

01:05:58.620 --> 01:06:03.707
And NVIDIA is bring Open Shell

01:06:03.790 --> 01:06:06.577
so Windows, and today, we are

01:06:06.660 --> 01:06:09.567
really thrilled to announce that

01:06:09.650 --> 01:06:13.387
Open Claw redundancy on Windows

01:06:13.470 --> 01:06:16.710
leveraging MXC.

01:06:16.810 --> 01:06:18.727
We are very deeply engaged with

01:06:18.810 --> 01:06:22.557
the team to make Open Claw run

01:06:22.640 --> 01:06:26.617
super well on Windows.
Let me hand it over to my

01:06:26.700 --> 01:06:30.867
colleagues scat and Samantha
toish -- Scott and smoothened to

01:06:30.950 --> 01:06:33.907
show you. Scott and Samantha?

01:06:33.990 --> 01:06:38.877
[Cheers and applause.] >> Hey, friends!

01:06:38.960 --> 01:06:41.817
Open Claw came out in November
of last year, and it took the

01:06:41.900 --> 01:06:45.267
world by on the stomach.
And for the -- by the storm.

01:06:45.350 --> 01:06:48.587
And I've been using it
to stay on top of my health.

01:06:48.670 --> 01:06:51.787
It can help me manage my blood

01:06:51.870 --> 01:06:55.497
sugar and gives me

01:06:55.580 --> 01:06:55.827
notifications.

01:06:55.910 --> 01:07:00.187
I've got line triaging email,

01:07:00.270 --> 01:07:03.307
GitHub issues, and buys movie tickets.

01:07:03.390 --> 01:07:04.767
What are you using your for?

01:07:04.850 --> 01:07:08.437
>> I turned my Open Claw agent

01:07:08.520 --> 01:07:10.977
into my triathlon coach.

01:07:11.060 --> 01:07:12.907
It developed a work back plan

01:07:12.990 --> 01:07:17.367
for me, using my data to notify

01:07:17.450 --> 01:07:19.337
me how I'm progressing and

01:07:19.420 --> 01:07:23.457
keeping me accountable.

01:07:23.540 --> 01:07:29.927
>> We're working closely with Open Claw to
make them successful and more successful

01:07:30.010 --> 01:07:34.387
on Windows. We've collaborated
on GitHub to bring you an Open Claw app.

01:07:34.470 --> 01:07:40.207
It will help you set up your own
on or connect to existing ones.

01:07:40.290 --> 01:07:43.407
And the Windows companion,
we're going to sandbox the Open Claw

01:07:43.490 --> 01:07:47.577
tool calls to keep you
and your system safe.

01:07:47.660 --> 01:07:53.117
>> Yeah! You'll see the Open Claw
windows app rin running in the background.

01:07:53.200 --> 01:07:58.857
Right-click on it, Scott. >> All right.

01:07:58.940 --> 01:08:00.527
>> It looks awesome!

01:08:00.610 --> 01:08:05.527
It looks like a native windows

01:08:05.610 --> 01:08:09.367
app because it is. It's got information
about I gateway, other machines, my

01:08:09.450 --> 01:08:12.967
sessions, and my usage.
I've also got quick access to

01:08:13.050 --> 01:08:17.377
things like chat, Canvas,
the main dashed, and more.

01:08:17.460 --> 01:08:19.947
Let's jump into companion settings.

01:08:20.030 --> 01:08:23.797
Within the app, we've got full
chat support with tool calling,

01:08:23.880 --> 01:08:26.117
and you'll notice in the corner,

01:08:26.200 --> 01:08:31.787
we've got lots of PERMTHSZ options,
along with our sandbox configuration.

01:08:31.870 --> 01:08:33.457
>> Thepermissions options,
along with our sandbox configuration.

01:08:33.540 --> 01:08:36.467
>> The sandbox is interesting
because this is using MXC, and

01:08:36.550 --> 01:08:39.647
for this, we're going to be
using process isolation.

01:08:39.730 --> 01:08:41.587
Now, newer version of Windows

01:08:41.670 --> 01:08:46.837
will have more containment
options, so keep your eyes out

01:08:46.920 --> 01:08:50.847
for news. I've got one-click security
option settings, but Samantha,

01:08:50.930 --> 01:08:53.917
talk to me about custom folders.

01:08:54.000 --> 01:08:58.237
>> Yeah. You've got full support
about what files and folders you want

01:08:58.320 --> 01:09:01.817
Open Claw to have access to,
and really granular security

01:09:01.900 --> 01:09:05.757
features like clipboard access
or talking to the internet itself.

01:09:05.840 --> 01:09:11.367
Now, I've given it read-only
desktop -- read-access to your

01:09:11.450 --> 01:09:12.227
depth folder.

01:09:12.310 --> 01:09:13.947
Open Claw already has a rich

01:09:14.030 --> 01:09:20.207
safety layer, augmented more by

01:09:20.290 --> 01:09:24.247
me or policies by IT.
For the purposes of this demo,

01:09:24.330 --> 01:09:28.847
I'm going to do something scary
and ask Open Claw to delete all

01:09:28.930 --> 01:09:31.667
the fails on your depth
-- files on your depth.

01:09:31.750 --> 01:09:35.687
>> I like a clean desktop.
>> You hid all your icons from

01:09:35.770 --> 01:09:37.857
the audience while we were on stage.

01:09:37.940 --> 01:09:41.957
I think we need to show the audience who
you really are >> That's so disrespectful.

01:09:42.040 --> 01:09:47.347
I know where everything is. [Laughter]
It's just -- don't touch my stuff!

01:09:47.430 --> 01:09:50.837
I know where they are. Sam, I need to make
sure that you're clear that a messy

01:09:50.920 --> 01:09:55.507
desktop is an organized mind. I'm pretty
sure that that's the... that's the quote,

01:09:55.590 --> 01:09:58.807
right? So what we've done, asked
Open Claw to delete those files from

01:09:58.890 --> 01:10:01.647
the Windows node, and the only
thing that is going to keep it

01:10:01.730 --> 01:10:06.037
from happening is MXC because
we've turned off all of the

01:10:06.120 --> 01:10:07.587
layers that Open Claw offers,

01:10:07.670 --> 01:10:09.757
but our IT, in this case,

01:10:09.840 --> 01:10:13.997
Samantha, has stet to

01:10:14.080 --> 01:10:19.847
-- you can see the attempts,
and checking the directory, and

01:10:19.930 --> 01:10:25.537
deleting again. It wants the files
gone, and I want them to stay.

01:10:25.620 --> 01:10:26.177
[Laughter] Ope! No!

01:10:26.260 --> 01:10:32.777
The read-only sandbox is there.

01:10:32.860 --> 01:10:36.097
94 je pegs are still there. They're safe!

01:10:36.180 --> 01:10:38.637
>> Foiled again! >> Oh, my goodness.

01:10:38.720 --> 01:10:43.727
So bad. >> We've seen security

01:10:43.810 --> 01:10:46.357
sandboxing, and brought together
all in this alpha release of a

01:10:46.440 --> 01:10:50.027
Windows companion app. We think this app
is a great opportunity to showcase Open

01:10:50.110 --> 01:10:53.587
Claw on windows, and it's only going
to get better in the coming months

01:10:53.670 --> 01:10:57.067
>> That is right. And, by the way,
I want to note that we're doing a all this

01:10:57.150 --> 01:10:59.937
development work on this calm
Windows development machine with

01:11:00.020 --> 01:11:04.107
all the tools that I love, like

01:11:04.190 --> 01:11:09.117
WSL, containers, a containment, and I've
even got GitHub Copilot with multimodel

01:11:09.200 --> 01:11:11.837
support ready to go.
This is literally how the team

01:11:11.920 --> 01:11:16.287
and I have been working
on the Open Claw app on GitHub.

01:11:16.370 --> 01:11:18.627
Now, people wonder how this came together.

01:11:18.710 --> 01:11:22.827
Turns out that over the holidays,
I got a DM from this

01:11:22.910 --> 01:11:27.447
random guy on the intent, and it
takes me a couple of days to get

01:11:27.530 --> 01:11:31.707
back to him. And then when I did,
we just had this kind of cool idea.

01:11:31.790 --> 01:11:35.177
That maybe he should come
to -- to Microsoft Build.

01:11:35.260 --> 01:11:37.077
I think there's one more
person that we want to thank for

01:11:37.160 --> 01:11:40.517
bringing the next generation
of agents to the world.

01:11:40.600 --> 01:11:43.267
Everybody put your hands

01:11:43.350 --> 01:11:49.857
together for the claw father
himself, Peter Steinberger!

01:11:49.940 --> 01:11:51.957
[Cheers and applause.] >>
Samantha, it's just showing

01:11:52.040 --> 01:11:54.727
off my secret DMs!

01:11:54.810 --> 01:11:59.777
I'm so excited to see Open
Claw native on Windows.

01:11:59.860 --> 01:12:01.787
You know, watching a claw try to

01:12:01.870 --> 01:12:07.737
delete all your desktop files
and just fail makes me really happy.

01:12:07.820 --> 01:12:12.507
[Laughter] Because six months ago,
that totally would've worked.

01:12:12.590 --> 01:12:16.367
[Laughter] I built Open Claw
to have access to everything.

01:12:16.450 --> 01:12:18.767
My files, my machines, my chats.

01:12:18.850 --> 01:12:21.087
Always on, and fully open source.

01:12:21.170 --> 01:12:23.287
That's what makes it so powerful!

01:12:23.370 --> 01:12:24.727
And that's what also makes

01:12:24.810 --> 01:12:27.697
companies a bit nervous.

01:12:27.780 --> 01:12:32.467
You know, what I kept hearing
was, "Peter, I love my claw.

01:12:32.550 --> 01:12:35.447
Can I use this at work?"
And that's what we spent the last

01:12:35.530 --> 01:12:37.267
few months on.

01:12:37.350 --> 01:12:41.477
With Microsoft, git Hick, open

01:12:41.560 --> 01:12:44.177
AI, NVIDIA, just to name a few.

01:12:44.260 --> 01:12:47.677
We had observability,
auto mode for permissions.

01:12:47.760 --> 01:12:50.937
We changed how access works.
It's not all or nothing anymore.

01:12:51.020 --> 01:12:56.287
You can pick which folder should read
only, which ones should be hidden.

01:12:56.370 --> 01:12:59.757
Here's the nutrition: -- Here's the news:.

01:12:59.840 --> 01:13:01.827
You can run it inside your company now.

01:13:01.910 --> 01:13:07.487
[Cheering] And we even made the harness

01:13:07.570 --> 01:13:08.617
itself a plugin.

01:13:08.700 --> 01:13:11.467
You can bring your own.

01:13:11.550 --> 01:13:14.837
Copilot, codex, whatever you

01:13:14.920 --> 01:13:19.347
already trust in and
your rules come with it.

01:13:19.430 --> 01:13:21.817
And you put Open Claw on top of it.

01:13:21.900 --> 01:13:25.437
Persistent memory, heartbeats,
and a claw right inside Slack or

01:13:25.520 --> 01:13:27.287
Teams.

01:13:27.370 --> 01:13:31.087
It's been really exciting to see

01:13:31.170 --> 01:13:35.627
Open Claw grow into something much bigger.

01:13:35.710 --> 01:13:37.327
A global movement.

01:13:37.410 --> 01:13:39.687
And a community.

01:13:39.770 --> 01:13:43.087
And I started the Open Claw

01:13:43.170 --> 01:13:45.467
Foundation, a really nonprofit.

01:13:45.550 --> 01:13:48.407
Any model, any operating system.

01:13:48.490 --> 01:13:51.947
Because we are entering a new

01:13:52.030 --> 01:13:55.847
era of building agents,
more capability for the people who

01:13:55.930 --> 01:13:59.617
don't code, and more
power for those who do.

01:13:59.700 --> 01:14:02.517
And we get to do this, we get to

01:14:02.600 --> 01:14:04.687
build it together in the open.

01:14:04.770 --> 01:14:08.197
So my task is simple: Come build with us.

01:14:08.280 --> 01:14:10.697
Thank you. [Cheers and applause.]

01:14:10.780 --> 01:14:15.647
>> Fantastic.
Thank you very much to Peter, to

01:14:15.730 --> 01:14:18.767
Samantha, to the companion
app team for their hard work.

01:14:18.850 --> 01:14:22.507
Thank you to the Open Claw
community for giving everyone

01:14:22.590 --> 01:14:26.077
here a crucifix crustacean
-- a crustacean of their own.

01:14:26.160 --> 01:14:31.537
And that's the first time you've
even Open Claw running on a

01:14:31.620 --> 01:14:36.857
Surface ultra.
[Cheers and applause.] >> All right.

01:14:36.940 --> 01:14:39.287
Thank you very much,
Samantha, Scott, and Peter.

01:14:39.370 --> 01:14:42.057
It's so wonderful to see Open

01:14:42.140 --> 01:14:44.997
Claw come to Windows.

01:14:45.080 --> 01:14:48.117
And have all of that capability

01:14:48.200 --> 01:14:53.367
in terms of the security and --
and that comfort, to be able to

01:14:53.450 --> 01:14:57.037
have these long running agents
and unmetered intelligence come together.

01:14:57.120 --> 01:15:00.147
So now, let's move from the

01:15:00.230 --> 01:15:02.247
Windows side on the edge to the

01:15:02.330 --> 01:15:04.167
cloud with Foundry briar briar,

01:15:04.250 --> 01:15:08.757
and we're building Foundry into

01:15:08.840 --> 01:15:12.427
a full application
platform for the agent era.

01:15:12.510 --> 01:15:15.397
Right? Every era of platform shift,

01:15:15.480 --> 01:15:17.867
when you move to the cloud, we

01:15:17.950 --> 01:15:24.007
have the cloud stack in Foundry.

01:15:24.090 --> 01:15:28.007
We're excited about the Foundry
hosted agent as a runtime for

01:15:28.090 --> 01:15:29.847
long running agents.

01:15:29.930 --> 01:15:35.217
You know, agents now have --
if you're building in Foundry

01:15:35.300 --> 01:15:38.497
hosted agents, you can
have all the IQ layers.

01:15:38.580 --> 01:15:42.757
You have the tools.
You have the durability and the

01:15:42.840 --> 01:15:45.787
memory and the state.
You have your own sandbox.

01:15:45.870 --> 01:15:49.297
In fact, it's a super-fast

01:15:49.380 --> 01:15:51.297
sandbox that you can spin up.

01:15:51.380 --> 01:15:54.187
You can generate the rubrics.

01:15:54.270 --> 01:15:58.567
You can -- and the evals.
You can in fact have all the

01:15:58.650 --> 01:16:02.437
safety and the guardrails
around your agentic system.

01:16:02.520 --> 01:16:04.847
And in fact, even one of the
coolest things in Foundry is

01:16:04.930 --> 01:16:07.347
it's a continuously improving loop, even.

01:16:07.430 --> 01:16:10.137
It's got that self-improvement
loop built in.

01:16:10.220 --> 01:16:13.617
You build an agent that's
continuously getting better.

01:16:13.700 --> 01:16:16.857
And I'm really also excited

01:16:16.940 --> 01:16:20.757
today to announce a partnership

01:16:20.840 --> 01:16:26.177
with Fireworks AI, bringing
all of their open wave models to

01:16:26.260 --> 01:16:30.507
Foundry. That means giving you
as developmenters more choices, as

01:16:30.590 --> 01:16:32.867
well as that great inference
stack, to build the next

01:16:32.950 --> 01:16:35.837
generation of these agentic

01:16:35.920 --> 01:16:41.317
indications, with all
the enterprise Rayls that Foundry

01:16:41.400 --> 01:16:43.437
has, as well as the governance
that Foundry has.

01:16:43.520 --> 01:16:48.670
Really excited about that partnership.

01:16:48.770 --> 01:16:53.057
And so now that brings us to the tools.

01:16:53.140 --> 01:16:55.097
Data itself is at the heart of all this.

01:16:55.180 --> 01:16:57.897
You know, in fact, Jensen spoke to this.

01:16:57.980 --> 01:17:00.247
GitHub is not just about the code repo.

01:17:00.330 --> 01:17:04.467
It's becoming the control
plane for all the agents.

01:17:04.550 --> 01:17:06.137
And nearly everything we measure

01:17:06.220 --> 01:17:09.037
on GitHub, whether it's repo

01:17:09.120 --> 01:17:11.707
creation, PR activity, API

01:17:11.790 --> 01:17:16.497
usage, actions, all of this are
going faster because of the

01:17:16.580 --> 01:17:22.457
agentic workflows. The new scale is driven
by humans and agents collaborating

01:17:22.540 --> 01:17:22.717
together.

01:17:22.800 --> 01:17:26.897
And they're exposing too long

01:17:26.980 --> 01:17:30.327
-- tooling crass every form factor.

01:17:30.410 --> 01:17:33.017
We've seen tremendous growth
in CLI, driven, you know, the

01:17:33.100 --> 01:17:37.407
approachability of the CLI form
factor, it's allusion -- always

01:17:37.490 --> 01:17:42.707
been great to go to a termal.
And now when combined with

01:17:42.790 --> 01:17:46.047
models and natural language,
CLI is the thing that everything goes to.

01:17:46.130 --> 01:17:48.117
But end of the day, when you

01:17:48.200 --> 01:17:54.687
have hundreds of CLI's,
it becomes pretty complicated.

01:17:54.770 --> 01:17:59.427
It doesn't scale. Especially the cognitive
load that I have when you have 100CLI

01:17:59.510 --> 01:18:03.927
sessions open, is such that you
kind of need something new.

01:18:04.010 --> 01:18:06.227
And so that's what has led us to

01:18:06.310 --> 01:18:11.407
build -- you know, we need like
this tool, essentially, that has

01:18:11.490 --> 01:18:14.407
the speed and the flexibility of a CLI.

01:18:14.490 --> 01:18:18.977
But has the capability of an IE,
and the ability to scale to

01:18:19.060 --> 01:18:21.477
infinite number of agent sessions.

01:18:21.560 --> 01:18:23.727
So today we're taking that next

01:18:23.810 --> 01:18:27.757
big step, introducing our new

01:18:27.840 --> 01:18:33.980
git Hick Copilot app --
Our new GitHub cCopilot app.

01:18:34.070 --> 01:18:35.457
And we realize it's not sufficient.

01:18:35.540 --> 01:18:38.097
Because you still have
a back toned deal with.

01:18:38.180 --> 01:18:40.617
Code is easy to generate,
but what about the back end?

01:18:40.700 --> 01:18:43.117
You need to content with

01:18:43.200 --> 01:18:44.907
identities, storage, database

01:18:44.990 --> 01:18:47.777
schemas, and that's why we're

01:18:47.860 --> 01:18:51.547
really super excited about

01:18:51.630 --> 01:18:56.077
Rayfin, an agent-first SDK that
connects your agents to the back

01:18:56.160 --> 01:19:00.257
end as a service, and we're bringing
this to everywhere you build.

01:19:00.340 --> 01:19:05.087
That's why I'm, again, super
excited about Rayfin and the

01:19:05.170 --> 01:19:12.080
relationship with Replit. You can now
build apps in Replit, while the -- yeah.

01:19:12.180 --> 01:19:16.367
You can build the app in Replit
while the app and data are

01:19:16.450 --> 01:19:18.167
deployed into the

01:19:18.250 --> 01:19:19.707
enterprise-managed fabric

01:19:19.790 --> 01:19:23.577
tenant, thanks to the Rayfin

01:19:23.660 --> 01:19:26.947
SDK, which is available now
for anyone else to be able to use

01:19:27.030 --> 01:19:29.897
with their tools as a back end,
and essentially have the back

01:19:29.980 --> 01:19:34.587
end service.
And now to show you all of the

01:19:34.670 --> 01:19:36.927
Foundry and Rayfin and building

01:19:37.010 --> 01:19:38.587
agents and the long running

01:19:38.670 --> 01:19:43.057
agents, let me invite Cassidy.

01:19:43.140 --> 01:19:46.127
Take with away. >> Hello, everyone!

01:19:46.210 --> 01:19:50.337
I'm so excited to be
amongst my fellow Devs.

01:19:50.420 --> 01:19:52.237
This is a long day. Fix your shoulders.

01:19:52.320 --> 01:19:54.907
Sit back. I see a lot of you sitting up.

01:19:54.990 --> 01:19:57.447
Great. I know you're drinking from
the fire hose of information today,

01:19:57.530 --> 01:20:01.217
so I want you to go into the next
few minutes thinking, "What

01:20:01.300 --> 01:20:03.997
can I drywall try out on -- what

01:20:04.080 --> 01:20:08.617
can I try out later today?"
First, I can't wait to show you

01:20:08.700 --> 01:20:12.557
the new GitHub Copilot app.
This is your home base for

01:20:12.640 --> 01:20:13.927
development and operations
on your computer.

01:20:14.010 --> 01:20:17.097
And we think you're going to love it.
Let me show you around.

01:20:17.180 --> 01:20:20.817
When you open up the app,
you see the home screen where you

01:20:20.900 --> 01:20:26.137
can kick off a new session.
But before night the serious

01:20:26.220 --> 01:20:29.177
stuff, you can drag Mona
around, and there's a game!

01:20:29.260 --> 01:20:32.207
Look! It's so fun!
[Laughter] I'm not very good at it, so

01:20:32.290 --> 01:20:37.327
let's just get back to, you can kick
off a new agentic coding session.

01:20:37.410 --> 01:20:39.987
So I started off one
a little bit earlier here.

01:20:40.070 --> 01:20:43.557
And it gave me a review
of a bunch of release blockers.

01:20:43.640 --> 01:20:48.197
Which one should I fix? Call it out.

01:20:48.280 --> 01:20:50.497
83? The critical ones? You know what?

01:20:50.580 --> 01:20:52.927
How about we just do all of them?

01:20:53.010 --> 01:20:57.197
Let's go! This app will now kick
off a separate session for every

01:20:57.280 --> 01:21:02.537
single issue here. I don't have to worry
about stashing or coding complex or

01:21:02.620 --> 01:21:08.217
anything, because the app
takes care of that with Git work

01:21:08.300 --> 01:21:09.967
trees, that are isolated
environments for each session.

01:21:10.050 --> 01:21:12.667
Your agents can work in parallel
without stepping on each other.

01:21:12.750 --> 01:21:15.187
But you still have to merge them.

01:21:15.270 --> 01:21:18.587
Right? So -- [chuckling] --
Copilot has your back there too.

01:21:18.670 --> 01:21:20.597
If I head over -- not to this one.

01:21:20.680 --> 01:21:23.357
But to this issue here, I can

01:21:23.440 --> 01:21:27.717
run agent merge.
And when I enable agent merge,

01:21:27.800 --> 01:21:33.067
cofight will baby sit this PR through CI
check, code review, and merge collects.

01:21:33.150 --> 01:21:36.437
Okay. Let me keep showing
you around while it's running.

01:21:36.520 --> 01:21:39.177
If I head over to my work, I can

01:21:39.260 --> 01:21:45.337
see a focused view of all of my activity
ask just projects loaded in the app, PRs,

01:21:45.420 --> 01:21:50.357
everything here. And under automation,
I have a bunch of reusable sessions that

01:21:50.440 --> 01:21:52.257
can run locally or on the cloud.

01:21:52.340 --> 01:21:54.397
You see there's issue poetry there.

01:21:54.480 --> 01:21:57.227
That's real. And that is load-bearing.

01:21:57.310 --> 01:22:02.047
Okay! [Chuckling]. And now undersessions,
like I briefly showed earlier, these

01:22:02.130 --> 01:22:03.367
are sessions.

01:22:03.450 --> 01:22:06.567
If I want to add a new

01:22:06.650 --> 01:22:11.647
repository, I can click the button,
and it will pull one,

01:22:11.730 --> 01:22:12.947
and if I were to add one, I can

01:22:13.030 --> 01:22:15.947
add a session, and this is an

01:22:16.030 --> 01:22:22.117
open source -- I can start
a session anywhere and it loads

01:22:22.200 --> 01:22:26.797
it. When I look at a session within
this repository, let me look at

01:22:26.880 --> 01:22:31.597
the other one, I get an integrated
browser, there's a terminal.

01:22:31.680 --> 01:22:35.247
And a chat. I can technology light
and -- toggle light and dark mode.

01:22:35.330 --> 01:22:38.767
And there's a button, pick and polish.

01:22:38.850 --> 01:22:42.237
I can polish anything in this app,
and it adds it to the chat,

01:22:42.320 --> 01:22:46.277
and I can say, hey, I want you
to add reordering to this list.

01:22:46.360 --> 01:22:48.977
And it will just work,
all living in there.

01:22:49.060 --> 01:22:52.797
I have access to all the most
popular models via my single

01:22:52.880 --> 01:22:55.317
GitHub Copilot subscription,

01:22:55.400 --> 01:23:00.627
including from open AI,
Ann throw meic, and Google.

01:23:00.710 --> 01:23:02.757
And having a model change is agreed.

01:23:02.840 --> 01:23:07.077
Not only can you pick the right
one, but Copilot can request a

01:23:07.160 --> 01:23:08.647
rubber duck review.

01:23:08.730 --> 01:23:12.707
In this session, I was use use

01:23:12.790 --> 01:23:15.737
GPT5.5, but it asked one from

01:23:15.820 --> 01:23:20.407
Claude opus 4.8.
All models have blind spots, and

01:23:20.490 --> 01:23:23.197
the power of this approach
means I can catch them earlier.

01:23:23.280 --> 01:23:28.317
This is all very cool.
But working with AI in 2026

01:23:28.400 --> 01:23:29.887
should be more than just chat.

01:23:29.970 --> 01:23:32.957
You just saw me scrolling up.
There's so many words here.

01:23:33.040 --> 01:23:39.327
[Laughter] So today, I'm very
excited to show you the concept of a

01:23:39.410 --> 01:23:43.047
canvass. I'm going to own one right there.

01:23:43.130 --> 01:23:46.467
The canvas is how an agent can
build a custom UI to communicate

01:23:46.550 --> 01:23:47.107
with you.

01:23:47.190 --> 01:23:53.747
What if your AI could see?
Everyone say, demo gods, bless

01:23:53.830 --> 01:23:58.267
us. Let's see if it works.
Here's a fun canvas where, if I

01:23:58.350 --> 01:24:02.657
get the camera going, okay.
The agent shows your PRs down

01:24:02.740 --> 01:24:04.527
here, and I can toggle it with a

01:24:04.610 --> 01:24:10.497
thumbs up or thumbs down.
Let's aapprove it!

01:24:10.580 --> 01:24:12.767
Hey! It's so fun! And goes deeper.

01:24:12.850 --> 01:24:14.747
This is just the beginning
of what you can do.

01:24:14.830 --> 01:24:16.247
So again, I kicked off a bunch

01:24:16.330 --> 01:24:19.867
of different sessions earlier.

01:24:19.950 --> 01:24:24.007
This is a signal box app.
It's 100% agent-build.

01:24:24.090 --> 01:24:27.447
It's containerized
with a database back end.

01:24:27.530 --> 01:24:30.027
Would you be able to deploy
this to your enterprise with no

01:24:30.110 --> 01:24:31.067
questions asked?

01:24:31.150 --> 01:24:37.827
Be honest. No. Yes.

01:24:37.910 --> 01:24:42.757
Exact. No, but you can with Rayfin,
and that is very, very, exciting.

01:24:42.840 --> 01:24:44.927
Let me open up a new terminal.

01:24:45.010 --> 01:24:48.587
All I have to do is type "Rayfin

01:24:48.670 --> 01:24:52.137
up." And then, demo gods bless you!

01:24:52.220 --> 01:24:52.367
Come on!

01:24:52.450 --> 01:24:54.837
It will...

01:24:54.920 --> 01:25:00.407
[Singing] maybe dedeploy! Blammo!

01:25:00.490 --> 01:25:02.317
It happened! Yes!
And hosted on Microsoft fabric.

01:25:02.400 --> 01:25:04.267
I know that's a lot. I know that's a lot.

01:25:04.350 --> 01:25:09.187
With Rayfin, your agents get
a complete enterprise back end, so

01:25:09.270 --> 01:25:12.457
you can deploy with confidence
in the way that's best for you.

01:25:12.540 --> 01:25:15.427
But this is the key thing
to remember: This app is not just

01:25:15.510 --> 01:25:18.567
another session manager. Yes, it is.

01:25:18.650 --> 01:25:20.447
It managing a lot of sessions.

01:25:20.530 --> 01:25:25.167
But session managers just make
it easy to create work, but this

01:25:25.250 --> 01:25:28.077
helps you to finish it.
Thank you very much.

01:25:28.160 --> 01:25:28.737
Happy pride.

01:25:28.820 --> 01:25:35.820
Back to you, Satya.

01:25:35.960 --> 01:25:40.647
>> All right. Now we're
back to IDEs that have UI.

01:25:40.730 --> 01:25:43.587
That sort of -- it's so cool! [Laughter]

01:25:43.670 --> 01:25:45.827
It's come full circle.

01:25:45.910 --> 01:25:48.727
Now, let's talk about how you

01:25:48.810 --> 01:25:51.167
can observe, govern, and secure

01:25:51.250 --> 01:25:53.167
these agents.

01:25:53.250 --> 01:25:57.067
Agent 365 is the agent control

01:25:57.150 --> 01:26:02.577
plane, agent requires their own
identities, access controls.

01:26:02.660 --> 01:26:07.147
Even when they're working on your behalf,
right? You just want that work on behalf

01:26:07.230 --> 01:26:08.417
identity to be enforced.

01:26:08.500 --> 01:26:14.057
So we extended an intraagent

01:26:14.140 --> 01:26:17.957
need real-time defense,
so we extended Defender.

01:26:18.040 --> 01:26:19.957
Agents require this always-on

01:26:20.040 --> 01:26:23.027
data protection and compliance,

01:26:23.110 --> 01:26:27.447
so we extended Purview.
And these agents can be hosted

01:26:27.530 --> 01:26:30.807
everywhere, AWS, GCP, not just on Azure.

01:26:30.890 --> 01:26:36.377
Or build with any framework.
And today we are announcing a

01:26:36.460 --> 01:26:38.277
number of updates, including the

01:26:38.360 --> 01:26:42.817
GA of agent 265SDK, and we're x

01:26:42.900 --> 01:26:45.697
expanding it to your local
agents running on Windows and

01:26:45.780 --> 01:26:47.857
elsewhere, and the claws,
you just saw earlier.

01:26:47.940 --> 01:26:50.517
And let us take a look at how

01:26:50.600 --> 01:26:54.097
all the Agent 365 composes.

01:26:54.180 --> 01:27:00.180
Over to you, Amanda.

01:27:00.270 --> 01:27:03.787
>> Everyone is building agents,
but that's not the hard part anymore.

01:27:03.870 --> 01:27:07.067
The hard part is intrigue
grating them into your business

01:27:07.150 --> 01:27:11.847
and governing them at scale. Today I
will show you how Foundry makes this easy.

01:27:11.930 --> 01:27:16.247
Let's start locally. I've already built a
line graph agent, and now I will show you

01:27:16.330 --> 01:27:21.117
the value Foundry adds. First, tools.

01:27:21.200 --> 01:27:24.187
Agents need tools to get work done,
and with Foundry tool box,

01:27:24.270 --> 01:27:25.927
I just add my tools once, and

01:27:26.010 --> 01:27:27.527
any agent can consume them

01:27:27.610 --> 01:27:31.327
through a single MCP end point.

01:27:31.410 --> 01:27:35.647
And because tools essentially,
that means governance does as

01:27:35.730 --> 01:27:38.567
well. I've applied a guard rile that

01:27:38.650 --> 01:27:42.207
blocks PII from leaking, and all

01:27:42.290 --> 01:27:44.147
I have to do is apply this once,

01:27:44.230 --> 01:27:47.347
and all my agents are protected.

01:27:47.430 --> 01:27:51.097
Now let's make this agent
enterprise-ready.

01:27:51.180 --> 01:27:53.937
To deploy this agent to Foundry,
all I have to do is add this one

01:27:54.020 --> 01:27:58.557
block of code, and then I push
my changes and GitHub actions

01:27:58.640 --> 01:28:02.297
takes it from there.
Once deployed, I can actually

01:28:02.380 --> 01:28:06.727
use this agent in the same
Foundry extension I showed you

01:28:06.810 --> 01:28:09.487
earlier. And let's now test it out.

01:28:09.570 --> 01:28:14.107
I'm asking it to track a few open items,
from a standup I had earlier today.

01:28:14.190 --> 01:28:18.377
And what's actually happening
right now, is Foundry is

01:28:18.460 --> 01:28:21.847
spinning up a dedicated
microVM just for this session.

01:28:21.930 --> 01:28:26.367
And the session even gets its
own persistent file system.

01:28:26.450 --> 01:28:29.657
What you're going to see
in a minute here, if I go to the

01:28:29.740 --> 01:28:34.927
Files tab to track the open
items, the agent's actually

01:28:35.010 --> 01:28:36.027
writing to a file.

01:28:36.110 --> 01:28:42.787
Pretty cool, right?

01:28:42.870 --> 01:28:46.767
Off the top of my head has
server side traces and evals to

01:28:46.850 --> 01:28:48.437
show me exactly what
happened on every run.

01:28:48.520 --> 01:28:50.207
But how do I know if my agent's

01:28:50.290 --> 01:28:51.577
doing a good job?

01:28:51.660 --> 01:28:56.517
That's what Foundry's brand-new

01:28:56.600 --> 01:28:59.897
rubric evaluators are for.

01:28:59.980 --> 01:29:04.557
One easy command, Foundry
reads my agent and generates the

01:29:04.640 --> 01:29:08.697
evaluation criteria for me.
It creates a rubric personalized

01:29:08.780 --> 01:29:12.067
just to this agent.
Now let's check out this rubric

01:29:12.150 --> 01:29:15.417
in Foundry portal.

01:29:15.500 --> 01:29:20.067
Look at these dimensions.
Governance outcome correctness,

01:29:20.150 --> 01:29:22.937
prescribed source usage.
I didn't write any of these.

01:29:23.020 --> 01:29:25.677
Foundry generated them
from production traces.

01:29:25.760 --> 01:29:28.417
And with this rubric, I can

01:29:28.500 --> 01:29:33.137
score my agent, and run evals,
but we can do so much more than that.

01:29:33.220 --> 01:29:37.787
That's where Foundry's brand-new
agent optimizer takes over.

01:29:37.870 --> 01:29:43.897
Here's how it works. It tunes four
things: The model, instructions, tool

01:29:43.980 --> 01:29:46.667
descriptions, and scales.
And then it generates improved

01:29:46.750 --> 01:29:51.037
candidates and scores each one
using the rubric I just showed you.

01:29:51.120 --> 01:29:52.037
Here, I can view the candidates.

01:29:52.120 --> 01:29:55.707
I can see the strategy used.
I can see their scores.

01:29:55.790 --> 01:29:59.677
And I can actually view
exactly what changed.

01:29:59.760 --> 01:30:02.347
This candidate, for example,
improved its score by updating

01:30:02.430 --> 01:30:05.397
the model and the system prompt.

01:30:05.480 --> 01:30:08.787
Foundry then makes it super easy
to deploy the best candidate as

01:30:08.870 --> 01:30:14.657
a brand-new agent version.
But this is not a one time

01:30:14.740 --> 01:30:16.657
thing. Every run feeds the next eval.

01:30:16.740 --> 01:30:20.917
And every eval tells the optimizer
where to improve next.

01:30:21.000 --> 01:30:24.467
So your agents now get
better the more they're used.

01:30:24.550 --> 01:30:27.677
But now, let's put this agent to work.

01:30:27.760 --> 01:30:30.847
I've published my agent to Teams

01:30:30.930 --> 01:30:34.917
and M365 Copilot.
Right now I'm going to ask it to

01:30:35.000 --> 01:30:39.597
just catch me up on what I've
missed, because I've been of

01:30:39.680 --> 01:30:41.617
offline all morning, but my agent hasn't.

01:30:41.700 --> 01:30:46.927
While it works on that, though,
let me explain what makes this

01:30:47.010 --> 01:30:50.497
agent different. This is an autopilot
agent, which means it has its own

01:30:50.580 --> 01:30:52.397
identity and productivity license.

01:30:52.480 --> 01:30:56.347
So it can work across
M365 on its own behalf.

01:30:56.430 --> 01:31:00.107
Earlier, you saw me using
this agent by myself.

01:31:00.190 --> 01:31:04.007
But shipping of features and new
release takes a whole team.

01:31:04.090 --> 01:31:07.677
And that why this agent
lives in our Teams group chat.

01:31:07.760 --> 01:31:10.497
Let's check out how it did.

01:31:10.580 --> 01:31:13.617
Serums as you didded --

01:31:13.700 --> 01:31:15.757
summariesed all updates and call

01:31:15.840 --> 01:31:20.897
out features I need to track.
But putting this to work in your

01:31:20.980 --> 01:31:22.857
enterprise, governance
needs to come first.

01:31:22.940 --> 01:31:27.317
That's why every autopilot agent

01:31:27.400 --> 01:31:32.537
requires admin approval to she
can review it and choose who has

01:31:32.620 --> 01:31:33.867
the ability to talk to it.

01:31:33.950 --> 01:31:37.477
But it doesn'tent there.

01:31:37.560 --> 01:31:41.547
An over approval, add minutes
anterior and block -- monitor it

01:31:41.630 --> 01:31:45.837
and block it at any time. But governance
can't apply to just one agent.

01:31:45.920 --> 01:31:47.717
Every agent in your organization

01:31:47.800 --> 01:31:53.227
needs to be managed with the same
rigor as users, apps, and devices.

01:31:53.310 --> 01:31:57.327
And that's exactly what
Microsoft Agent 365 provides.

01:31:57.410 --> 01:32:01.447
So let's zoom out.
Today we saw Foundry accelerate

01:32:01.530 --> 01:32:04.667
development, take your agent
from local to enterprise-ready,

01:32:04.750 --> 01:32:09.207
and put it to work in M365.
Foundry makes it simple.

01:32:09.290 --> 01:32:11.777
You build the agent; we handle the rest.

01:32:11.860 --> 01:32:15.200
Now back to you, Satya!

01:32:15.530 --> 01:32:19.667
>> Thank you, Amanda.

01:32:19.750 --> 01:32:26.057
You know, what you just saw
was how we're building security for

01:32:26.140 --> 01:32:29.527
AI. But there's also one other
critical aspect, especially, you

01:32:29.610 --> 01:32:31.997
know, the news today is all

01:32:32.080 --> 01:32:37.637
about, you know, how do you
defend yourself using AI against

01:32:37.720 --> 01:32:40.467
attacks that may in fact be using AI?

01:32:40.550 --> 01:32:40.677
Right?

01:32:40.760 --> 01:32:46.677
So last month, we announced our

01:32:46.760 --> 01:32:51.327
multimodel agentic security
system in dash, an agent harness

01:32:51.410 --> 01:32:53.837
for security, essentially, that we build.

01:32:53.920 --> 01:32:56.557
We're bringing together a

01:32:56.640 --> 01:33:01.097
hundred agents across the frontier,
and custom models, to

01:33:01.180 --> 01:33:05.227
really find these exploitable
bugs better than any single

01:33:05.310 --> 01:33:05.667
model does.

01:33:05.750 --> 01:33:09.987
In fact, when we Dubdebuted thi

01:33:10.070 --> 01:33:15.537
harness, it was on top of cyber
gym [dog barking] benchmark --

01:33:15.620 --> 01:33:20.507
Jim benchmark. So look what you
can do to defend against AI attacks.

01:33:20.590 --> 01:33:27.590
Over to you, Sarah.

01:33:29.500 --> 01:33:33.757
>> Thank you. We know that security
scans can take a while, so I'm not going

01:33:33.840 --> 01:33:38.597
to do one life. Let me show you the
results of an ender scan I already ran on

01:33:38.680 --> 01:33:41.847
my code base.
The system runs as a stand-alone

01:33:41.930 --> 01:33:47.907
CLI, but today, I'm using it
in my GitHub Copilot app on my

01:33:47.990 --> 01:33:52.907
local dev machine. The scan is broken
down but vulnerability, demands, and

01:33:52.990 --> 01:33:57.667
severity. And it's also an additional

01:33:57.750 --> 01:34:01.617
to -- finding traditional
issues like coding errors and

01:34:01.700 --> 01:34:06.787
hard-coated secrets. It's identifying
AI-specific vulnerabilities in the code

01:34:06.870 --> 01:34:08.697
base. What happened under the hood,

01:34:08.780 --> 01:34:10.627
over 100 specialized agents

01:34:10.710 --> 01:34:13.917
working together to discover,

01:34:14.000 --> 01:34:17.837
debate, and improve exploitable
vulnerabilities end to end.

01:34:17.920 --> 01:34:20.607
When the scan finishes, it

01:34:20.690 --> 01:34:25.807
generates both a log and an HTML
report that I can give to my

01:34:25.890 --> 01:34:31.467
management. The defender details
command allows me to dig into these

01:34:31.550 --> 01:34:35.057
vulnerabilities, and you can
see what it is, where it is in my

01:34:35.140 --> 01:34:38.287
code, and the severity
to help me prioritize.

01:34:38.370 --> 01:34:40.357
And from here, of course, we are

01:34:40.440 --> 01:34:42.957
going to fix it.

01:34:43.040 --> 01:34:47.617
Now, using the Defender
fix command, the system will

01:34:47.700 --> 01:34:51.967
remediate suggest the fixing
directly in my local dev environment.

01:34:52.050 --> 01:34:55.077
When that's done,
I can check out the diff.

01:34:55.160 --> 01:34:57.847
So I have full transparency
about what the harness has done,

01:34:57.930 --> 01:35:02.177
and I still have a human
in the loop check.

01:35:02.260 --> 01:35:05.597
But of course, everything
I've shown you so far has run

01:35:05.680 --> 01:35:06.267
locally.

01:35:06.350 --> 01:35:09.387
But I can also create a PR to

01:35:09.470 --> 01:35:13.227
plug into my existing workflows

01:35:13.310 --> 01:35:17.227
and push up to my repo,
and I can take the output from the

01:35:17.310 --> 01:35:21.477
scan and upload it to tools,
like GitHub advanced security,

01:35:21.560 --> 01:35:24.537
and manage everything alongside

01:35:24.620 --> 01:35:28.677
my other indication security findings.

01:35:28.760 --> 01:35:32.977
I've shown it working on my code, but let
me show you a vulnerability our security

01:35:33.060 --> 01:35:38.167
research teams identified
using MDASH that you can look up

01:35:38.250 --> 01:35:40.737
yourself. So the TL;DR, because this is a

01:35:40.820 --> 01:35:46.687
lot to read of this bug,
is the time reads an out of date map of

01:35:46.770 --> 01:35:50.967
an object. Runs off the end.
And then crashes the host.

01:35:51.050 --> 01:35:55.917
So this is exactly the kind of bug
that the harness was build for.

01:35:56.000 --> 01:36:00.037
Because the floor spread across three
different parts of the code base.

01:36:00.120 --> 01:36:03.277
So no single file looks wrong on its own.

01:36:03.360 --> 01:36:06.077
They look absolutely fine.

01:36:06.160 --> 01:36:10.447
And we can even see here be

01:36:10.530 --> 01:36:14.117
probably my favorite part,
a statement from the developers

01:36:14.200 --> 01:36:17.817
claiming everything is fine.
This is exactly the sort of

01:36:17.900 --> 01:36:22.157
reassurance that fools normal
scanners and single AI models.

01:36:22.240 --> 01:36:27.417
But MDASH wasn't fooled. One team
of agents spotted the suspicious gap.

01:36:27.500 --> 01:36:29.187
Another team argue they it in part.

01:36:29.270 --> 01:36:34.337
And a third team build a working example
that actually triggered the crash.

01:36:34.420 --> 01:36:36.407
And it did all of this
in open source code base.

01:36:36.490 --> 01:36:41.847
And this is a kind of joined-up
reasoning that previously

01:36:41.930 --> 01:36:44.327
required significant manual
security research effort.

01:36:44.410 --> 01:36:48.557
This is MDASH helping developers
create secure code from the

01:36:48.640 --> 01:36:52.187
start, coming soon to your CLI
and the Microsoft Defender portal.

01:36:52.270 --> 01:36:56.180
Back to you, Satya!

01:36:56.540 --> 01:37:00.627
>> Thank you, Sarah.
So that was the stack.

01:37:00.710 --> 01:37:05.337
Before, though, we move to,
you know, unpacking more of the

01:37:05.420 --> 01:37:07.407
opportunity, I want
to do something different.

01:37:07.490 --> 01:37:10.437
I want to introduce two people

01:37:10.520 --> 01:37:12.707
whose LinkedIn profiles were

01:37:12.790 --> 01:37:14.717
both super impressive and

01:37:14.800 --> 01:37:18.667
slightly perplexing.

01:37:18.750 --> 01:37:22.217
Under role had says they're
general partners on Mantis.

01:37:22.300 --> 01:37:25.627
On the previous role, sold out

01:37:25.710 --> 01:37:29.127
Madison Square Garden.

01:37:29.210 --> 01:37:36.210
Welcome Alex and Drew from the

01:37:38.650 --> 01:37:40.437
Chainsmokers. Thank you so much.

01:37:40.520 --> 01:37:43.737
One of the things when my team
came to me and said, hey, we're

01:37:43.820 --> 01:37:47.377
going to have Chainsmokers
at Build, I thought maybe that's

01:37:47.460 --> 01:37:51.567
what we're calling
our new GitHub Copilot app.

01:37:51.650 --> 01:37:55.557
But maybe they're your fans.
But tell me about how you got into this.

01:37:55.640 --> 01:38:00.957
I mean, you're being added
for what -- at it for what, 12-plus

01:38:01.040 --> 01:38:03.997
years as investors first,
and you had your firm now for

01:38:04.080 --> 01:38:06.667
seven-plus years, and you even
picked what I would have not

01:38:06.750 --> 01:38:09.987
thought is the natural place,

01:38:10.070 --> 01:38:15.607
which is B2B SaaS. Give us the back story.

01:38:15.690 --> 01:38:18.677
>> I'm sure you guys are
wondering what time line you're

01:38:18.760 --> 01:38:20.447
on, where we're at Microsoft

01:38:20.530 --> 01:38:24.497
Build, but hey, how are you.

01:38:24.580 --> 01:38:28.057
We've been in the Chainsmokers
for 14 years, and you know, when

01:38:28.140 --> 01:38:30.487
our music started to take off in

01:38:30.570 --> 01:38:35.457
the mid two thousand teens,
we got to play a few events like this.

01:38:35.540 --> 01:38:40.167
And we met a lot of the founders from
the consumer mobile cloud era of startups.

01:38:40.250 --> 01:38:41.617
And they taught us a lot and

01:38:41.700 --> 01:38:48.207
give us a front row seat into
what early stage investing was

01:38:48.290 --> 01:38:51.437
like. We fell love with it,
had a lot in common with the founders and

01:38:51.520 --> 01:38:54.447
the way they build their
business was the same way that

01:38:54.530 --> 01:38:59.637
we thought about starting the Chainsmokers
and breaking through the noise there.

01:38:59.720 --> 01:39:03.757
And we got to participate
in a few of their deals.

01:39:03.840 --> 01:39:08.927
And we decided to institutionalized
in 2020 and start our own fund

01:39:09.010 --> 01:39:12.327
>> That's cool. When you hear about
all the AI stuff and what's happening,

01:39:12.410 --> 01:39:16.247
there's a lot going on, even in your
portfolio we were talking back staining.

01:39:16.330 --> 01:39:19.237
Things are changing. How do you
see the opportunity going forward?

01:39:19.320 --> 01:39:22.877
>> I mean, there's so many
vectors you could talk about.

01:39:22.960 --> 01:39:26.347
There's the creative
output side -- >> Yep.

01:39:26.430 --> 01:39:28.017
>> -- we've been
experimenting with in music.

01:39:28.100 --> 01:39:31.697
But always important to have your
authenticity when it comes to creativity.

01:39:31.780 --> 01:39:36.857
But on the invest side, we're moving from
producing outputs to producing actions.

01:39:36.940 --> 01:39:42.757
Which I think presents a very interesting
opportunity to reimagine the entire

01:39:42.840 --> 01:39:45.427
architecture of the way
software enterprise is build.

01:39:45.510 --> 01:39:49.187
Instead of humus producing
outputs -- humans producing

01:39:49.270 --> 01:39:52.937
outputs, it's machines, and rethinking
the entire space in that context.

01:39:53.020 --> 01:39:56.107
>> It's fantastic to see that.

01:39:56.190 --> 01:40:00.877
And maybe to close out,
any sort of -- as artists who have had

01:40:00.960 --> 01:40:04.437
great success, when you look at

01:40:04.520 --> 01:40:08.217
a founder, what sort of is the
advice you give them when you're

01:40:08.300 --> 01:40:12.087
creating, you know, you're
bringing something new, and it's

01:40:12.170 --> 01:40:16.057
a creative process. So what is --
what are you looking for in founders?

01:40:16.140 --> 01:40:18.497
What is it that you're
giving them as advice?

01:40:18.580 --> 01:40:21.077
>> Totally. There really
are so many parallels.

01:40:21.160 --> 01:40:22.947
Maybe it seems like that;
maybe it doesn't.

01:40:23.030 --> 01:40:26.737
But it's really difficult
as artists to really find, I guess,

01:40:26.820 --> 01:40:29.637
what your sound s
what's authentically you.

01:40:29.720 --> 01:40:33.677
What you can continue to do over
and over, because with this much

01:40:33.760 --> 01:40:35.847
competition, you have
to be moving authentically.

01:40:35.930 --> 01:40:39.597
You have to be connected
to the product that you're creating.

01:40:39.680 --> 01:40:44.927
And -- 'cause you're going
to have to iteratedate and keep it

01:40:45.010 --> 01:40:47.957
going for a long time to lock
into your fan base and make

01:40:48.040 --> 01:40:50.957
something special and nearby.
And we try -- and unique.

01:40:51.040 --> 01:40:54.717
And we try to advise the same thing.

01:40:54.800 --> 01:40:58.037
>> Are you game for playing for us
this evening for our closest friends here?

01:40:58.120 --> 01:41:01.607
>> We'll be there.
>> 6:00! 6 this evening.

01:41:01.690 --> 01:41:02.607
Thank you very much.

01:41:02.690 --> 01:41:03.107
[Cheers and applause.]

01:41:03.190 --> 01:41:08.817
Thank you, thank you.

01:41:08.900 --> 01:41:11.697
We will be right here at 6 p.m.

01:41:11.780 --> 01:41:15.887
tonight, and really looking forward to it.

01:41:15.970 --> 01:41:19.227
So that effectively sort
of talks a little bit about the

01:41:19.310 --> 01:41:23.757
developer stacks.
So now let's talk about what is

01:41:23.840 --> 01:41:26.427
the opportunity for every company.

01:41:26.510 --> 01:41:30.517
Right? At the end of the day,
we are an institution, an organizational

01:41:30.600 --> 01:41:35.777
builders, whether it's AI nave
company, whether it's a SaaS

01:41:35.860 --> 01:41:38.607
company, or whether enterprise,
the first thing we want to do is

01:41:38.690 --> 01:41:40.747
make sure that as you build

01:41:40.830 --> 01:41:43.817
things black --

01:41:43.900 --> 01:41:45.937
Build things like plugins or

01:41:46.020 --> 01:41:51.757
agents or AI apping, we want
them to be discovered through

01:41:51.840 --> 01:41:54.087
the Microsoft ecosystem.
That's job one for us.

01:41:54.170 --> 01:41:58.227
That's why we're doing things
in Windows or Microsoft 365 Copilot

01:41:58.310 --> 01:41:59.897
or in Teams or GitHub.

01:41:59.980 --> 01:42:03.847
We want to structure them that

01:42:03.930 --> 01:42:07.237
your applications, agents,
plugins are discovered.

01:42:07.320 --> 01:42:10.307
Customers are also building lots

01:42:10.390 --> 01:42:14.607
of line of business applications
and agents using Copilot Studio,

01:42:14.690 --> 01:42:17.577
and we want to make that all
discoverable again, as part of

01:42:17.660 --> 01:42:21.987
the Copilot experience.
And Teams in some sense has

01:42:22.070 --> 01:42:26.657
become this destination
for multiplayer, human to agent

01:42:26.740 --> 01:42:32.027
interaction. We want you to be able
to find agents, interact with agents,

01:42:32.110 --> 01:42:37.667
right in Teams.
And we are super-charging all of

01:42:37.750 --> 01:42:41.207
this. Copilot continues
to evolve very quickly.

01:42:41.290 --> 01:42:43.037
It started first with chat, with

01:42:43.120 --> 01:42:45.277
some of the best models, with

01:42:45.360 --> 01:42:49.977
great access to Work IQ.

01:42:50.060 --> 01:42:53.737
We didn't have the name,
but that's where it started.

01:42:53.820 --> 01:42:57.517
Then came Cowork, a new way
of working, and generating the

01:42:57.600 --> 01:43:01.227
stunning artifacts,
and solve these multistep problems.

01:43:01.310 --> 01:43:05.697
Right? So you assign
a multistep task to Cowork.

01:43:05.780 --> 01:43:07.527
You saw GitHub, which has

01:43:07.610 --> 01:43:09.447
continued to evolve as well.

01:43:09.530 --> 01:43:13.937
And now, as of, you know, come

01:43:14.020 --> 01:43:18.477
summer, you will be bringing
coding to all knowledge work

01:43:18.560 --> 01:43:21.847
within one Copilot super app. Right?

01:43:21.930 --> 01:43:27.450
That's going to be really
exciting, to see -- yep!

01:43:27.550 --> 01:43:31.557
You will have chat, cowork,
and code all in Copilot.

01:43:31.640 --> 01:43:34.157
But today, we're introducing

01:43:34.240 --> 01:43:40.297
something completely new.

01:43:40.380 --> 01:43:43.087
Autopilots. When you
think of autopilots as

01:43:43.170 --> 01:43:45.267
enterprise-grade claws.

01:43:45.350 --> 01:43:50.477
These are autonomous,
long-running agents with full

01:43:50.560 --> 01:43:52.547
enterprise compliance that run

01:43:52.630 --> 01:43:53.877
in your tenant.

01:43:53.960 --> 01:43:57.967
Autopilots can have a name,

01:43:58.050 --> 01:44:01.137
personality, custom connecters,
context, and memory.

01:44:01.220 --> 01:44:04.987
And they're totally new way
to reduce toil and get you back to

01:44:05.070 --> 01:44:07.657
what you love, to kick things

01:44:07.740 --> 01:44:11.157
off, the the first autopilot

01:44:11.240 --> 01:44:14.547
we're introducing is Scout.

01:44:14.630 --> 01:44:21.630
Let's take a look.

01:45:11.740 --> 01:45:17.057
>> All right. As you can see, Scout works

01:45:17.140 --> 01:45:21.247
where you work, joining group
chats and teams, handling

01:45:21.330 --> 01:45:25.367
threats in Outlook.
Starting today for those of you

01:45:25.450 --> 01:45:27.807
on Copilot Frontier,
you can try out Scout.

01:45:27.890 --> 01:45:32.647
And in the coming months, we will build
this out to a complete digital team of

01:45:32.730 --> 01:45:34.447
autopilots, right inside of Copilot.

01:45:34.530 --> 01:45:36.647
Right? So you can go to the Copilot app.

01:45:36.730 --> 01:45:38.137
Scout is the one that comes by default.

01:45:38.220 --> 01:45:42.317
But you can build more
of these autopilots.

01:45:42.400 --> 01:45:46.357
And so that's the future of what
we think of as the Copilot

01:45:46.440 --> 01:45:50.397
ecosystem itself.
And so far, we have talked a lot

01:45:50.480 --> 01:45:53.397
about what you can do to build
your own agents and how these

01:45:53.480 --> 01:45:58.067
agents are discovered and things
like Copilot and Teams and Windows.

01:45:58.150 --> 01:45:59.907
But when you think about what

01:45:59.990 --> 01:46:01.937
makes any organization, any

01:46:02.020 --> 01:46:05.647
enterprise, any company unique,

01:46:05.730 --> 01:46:11.467
it is its Tatacit knowledge
tha is continuously compounding

01:46:11.550 --> 01:46:14.267
through its operations.
So the key consideration at some

01:46:14.350 --> 01:46:16.987
level in an AI age is to ask the

01:46:17.070 --> 01:46:21.397
question, "What's the future of the firm?"

01:46:21.480 --> 01:46:24.427
How do you continue to preserve
and compound that tacit

01:46:24.510 --> 01:46:27.867
knowledge in the age of AI,
where models can learn anything?

01:46:27.950 --> 01:46:29.517
From the data and the

01:46:29.600 --> 01:46:30.887
trajectories they see.

01:46:30.970 --> 01:46:35.067
To do that, we believe that

01:46:35.150 --> 01:46:37.977
every organization, whether it's

01:46:38.060 --> 01:46:42.447
an AI native company, a SaaS
company, or an enterprise, will

01:46:42.530 --> 01:46:43.747
need to build their own

01:46:43.830 --> 01:46:45.197
hill-climbing machine.

01:46:45.280 --> 01:46:49.857
It's a system that continuously

01:46:49.940 --> 01:46:54.757
improves against your objectives,
your private evals,

01:46:54.840 --> 01:46:56.757
compounding your advantage over time.

01:46:56.840 --> 01:46:59.327
Not someone else's.

01:46:59.410 --> 01:47:04.687
To deliver on this, we're taking
the next step with Frontier

01:47:04.770 --> 01:47:08.477
tuning. We've been working with many
customers already to help, you

01:47:08.560 --> 01:47:11.677
know, really build
their own learning loop.

01:47:11.760 --> 01:47:17.217
The environment, the context,
the tools, the ribosomes -- the

01:47:17.300 --> 01:47:18.967
rubrics, and even train their own models.

01:47:19.050 --> 01:47:23.357
And today, we're really thrilled
and excited about super charging

01:47:23.440 --> 01:47:28.257
that frontier tuning capability with the
innovation that's coming out of our super

01:47:28.340 --> 01:47:35.270
intelligence lab, to tell you all about
this, let me hand it over to Mustafa.

01:47:35.360 --> 01:47:40.137
>> Thank you. [Music].
Thank you, thank you.

01:47:40.220 --> 01:47:43.337
Thank you, good morning, everybody.

01:47:43.420 --> 01:47:45.277
You know, we really are living

01:47:45.360 --> 01:47:47.907
in the most remarkable times.

01:47:47.990 --> 01:47:50.647
Since I started working in AI,

01:47:50.730 --> 01:47:53.237
the computer that we used to

01:47:53.320 --> 01:47:55.417
train frontier models has

01:47:55.500 --> 01:47:58.517
increased by one trillion-fold.

01:47:58.600 --> 01:48:02.157
That is 12 orders of magnitude

01:48:02.240 --> 01:48:03.697
of computation in just 15 years.

01:48:03.780 --> 01:48:07.097
It's clear that a consistent,

01:48:07.180 --> 01:48:10.847
exponential increase
in computation leads to predictable

01:48:10.930 --> 01:48:15.507
advances in AI capabilities.
And in the next few years, we're

01:48:15.590 --> 01:48:17.337
going to see three more orders

01:48:17.420 --> 01:48:24.420
of magnitude of compute apply to

01:48:25.380 --> 01:48:29.287
train Frontier models.
Long linear hill climbing is the norm.

01:48:29.370 --> 01:48:32.117
The scaling laws are clearly
holding, and it is a remarkable

01:48:32.200 --> 01:48:35.427
time in our industry.

01:48:35.510 --> 01:48:39.567
And so in this context,
we at MAI are building towards what we

01:48:39.650 --> 01:48:42.277
call humanist superintelligence.

01:48:42.360 --> 01:48:47.437
State of the art AI capabilities
that are explicitly designed to

01:48:47.520 --> 01:48:52.147
serve people and organizations
and not replace them.

01:48:52.230 --> 01:48:54.407
Because the type of AI that we

01:48:54.490 --> 01:48:56.977
create really does matter.

01:48:57.060 --> 01:48:59.537
We need an AI that places

01:48:59.620 --> 01:49:01.537
humanity first.

01:49:01.620 --> 01:49:05.457
But always prioritizes human

01:49:05.540 --> 01:49:07.287
well-being and human progress.

01:49:07.370 --> 01:49:12.867
This is the core philosophy and

01:49:12.950 --> 01:49:15.847
modification behind our efforts
at Microsoft, and it shapes

01:49:15.930 --> 01:49:18.287
everything that we do.
And as a platform company, our

01:49:18.370 --> 01:49:23.877
job and our commitment is to keep
you developers building the

01:49:23.960 --> 01:49:27.177
absolute frontier.
So today, we're very excited to

01:49:27.260 --> 01:49:32.467
announce a family of seven
new models across image, voice,

01:49:32.550 --> 01:49:33.567
transcription, and coding.

01:49:33.650 --> 01:49:38.157
[Cheers and applause.]

01:49:38.240 --> 01:49:42.657
These are all built with real
attention to detail and a

01:49:42.740 --> 01:49:45.697
commitment to making very
practical and efficient tools

01:49:45.780 --> 01:49:49.817
that are typed to -- tuned
to just how you work in the real

01:49:49.900 --> 01:49:53.407
world. First up: MAI image 2.5 and its

01:49:53.490 --> 01:49:55.007
flash variant, two super strong

01:49:55.090 --> 01:49:57.607
models that deliver a step

01:49:57.690 --> 01:50:03.117
change in quality, not -- now at

01:50:03.200 --> 01:50:09.200
#2, surpassing the core of Nanabanana2.

01:50:09.300 --> 01:50:12.787
They gave you precise editing
with control and consistency,

01:50:12.870 --> 01:50:16.997
flashes here for super efficient
production work loads, while 2.5

01:50:17.080 --> 01:50:21.947
gives you that maximum fidelity
and professional grade performance.

01:50:22.030 --> 01:50:25.737
Alive in PowerPoint today,
they're rolling out to One

01:50:25.820 --> 01:50:29.907
Drive, and you can access them
on Foundry at a market leading

01:50:29.990 --> 01:50:34.847
quality per dollar.
Next, MAI transcribe 1.5.

01:50:34.930 --> 01:50:37.547
This is the best transcription
model in the world.

01:50:37.630 --> 01:50:41.887
State of the art accuracy across

01:50:41.970 --> 01:50:47.587
43 languages, beating out Gemini
and open AI's flagship models.

01:50:47.670 --> 01:50:48.827
We've optimized it for real

01:50:48.910 --> 01:50:54.397
world use so you can produce

01:50:54.480 --> 01:50:57.247
highly accurate triceps,
five times -- transcripts, five times

01:50:57.330 --> 01:51:02.177
faster than all rival models.
It's integrated inside of

01:51:02.260 --> 01:51:05.877
GitHub, Copilot, dynamic 365
contact center, and available in

01:51:05.960 --> 01:51:11.117
Foundry where I'm excited
to say it is the fastest, most

01:51:11.200 --> 01:51:13.667
official, and most -- efficient,
and most cost-effective

01:51:13.750 --> 01:51:19.970
transcription model of any
of the hyperscalers out there.

01:51:20.070 --> 01:51:22.457
So paired with that, we've got

01:51:22.540 --> 01:51:23.397
MAI Voice 2.

01:51:23.480 --> 01:51:28.127
The latest speech generation.

01:51:28.210 --> 01:51:29.687
Beautiful prosody, natural

01:51:29.770 --> 01:51:34.437
delivery, fine grained control,

01:51:34.520 --> 01:51:37.577
in 15 languages with more to come.

01:51:37.660 --> 01:51:39.277
And voice 2 flash, providing the

01:51:39.360 --> 01:51:45.397
best value and speed for ultra

01:51:45.480 --> 01:51:48.137
lateliency voice, which
is the giggest thing in 2026.

01:51:48.220 --> 01:51:50.857
Next up, text foundation model.

01:51:50.940 --> 01:51:55.427
MAI thinking one. This is the first
reasoning model, and it's exceptionally

01:51:55.510 --> 01:51:58.257
strong, and our target use cases

01:51:58.340 --> 01:52:00.427
of reasoning and sweet tasks.

01:52:00.510 --> 01:52:03.517
It's a 35 billion active

01:52:03.600 --> 01:52:10.407
parameter E, meaning it competes
in the medium sized weight

01:52:10.490 --> 01:52:12.037
class. And independent human raters on

01:52:12.120 --> 01:52:13.877
surge prefer it in overall

01:52:13.960 --> 01:52:19.737
quality side by side versus

01:52:19.820 --> 01:52:20.837
sonnet 4.6.

01:52:20.920 --> 01:52:21.767
[Cheering]

01:52:21.850 --> 01:52:27.527
It's achieved 97% on AMI2025,

01:52:27.610 --> 01:52:30.297
the key measure of its
capabilities, but most

01:52:30.380 --> 01:52:34.467
importantly, it's now at 53% on

01:52:34.550 --> 01:52:35.847
bench pro that places it

01:52:35.930 --> 01:52:39.137
alongside opus 4-6, at least on

01:52:39.220 --> 01:52:42.507
the toughest coding benchmark

01:52:42.590 --> 01:52:46.790
out there. We're very happy with that.

01:52:46.890 --> 01:52:51.147
There's plenty more for us as we
get this into production and

01:52:51.230 --> 01:52:55.017
hill climb against real-world
tasks and traffic, but what's

01:52:55.100 --> 01:52:58.787
most remarkable about this model,
we think, it has climbed

01:52:58.870 --> 01:53:00.627
entirely from the bottom.

01:53:00.710 --> 01:53:02.327
And that means that it hasn't

01:53:02.410 --> 01:53:07.267
targeted any of the benchmarks

01:53:07.350 --> 01:53:09.987
specifically, with zero distillation.

01:53:10.070 --> 01:53:12.067
And to us, this is critical.
It means that the model is

01:53:12.150 --> 01:53:17.307
created with an enterprise-grade
clean and commercially licensed

01:53:17.390 --> 01:53:21.077
data lineage that means that you
can put it into production in a

01:53:21.160 --> 01:53:24.047
very trustworthy way
with complete confidence.

01:53:24.130 --> 01:53:29.487
Now, finally, I'm incredibly
excited to announce MAI code 1

01:53:29.570 --> 01:53:32.587
flash. This is our new inference

01:53:32.670 --> 01:53:34.427
sufficient coding model tuned

01:53:34.510 --> 01:53:41.510
for VS code and GitHub CLI.

01:53:42.400 --> 01:53:45.287
5 51%, despite five billion parameters.

01:53:45.370 --> 01:53:49.577
Close to haiku in terms of size
but cheapner cost, delivering

01:53:49.660 --> 01:53:51.737
strong coding performance at

01:53:51.820 --> 01:53:56.817
great sufficiencily.
And it's rolling out inside of

01:53:56.900 --> 01:54:00.750
VS Code.

01:54:00.850 --> 01:54:04.787
Now, alongside distribution
on Foundry and optimization for our

01:54:04.870 --> 01:54:07.657
1P products, we're excited
to make our models available on

01:54:07.740 --> 01:54:11.357
open router as well
as Fireworks and base ten.

01:54:11.440 --> 01:54:13.997
This means for the first time
you will be able to tune the

01:54:14.080 --> 01:54:17.487
weights directly yourself
in an ecosystem of your choice.

01:54:17.570 --> 01:54:21.707
Now, across this entire family,
safety and security has been

01:54:21.790 --> 01:54:26.707
build in from the start.
Voice models come with

01:54:26.790 --> 01:54:29.147
protectiontion against
unauthorized cloning.

01:54:29.230 --> 01:54:32.497
Everything is water marked from scratch.

01:54:32.580 --> 01:54:36.657
Reduced the reFOOUFLS,
improvedish -- refuseals, and

01:54:36.740 --> 01:54:40.257
accomplishing a very detailed
technical report today to give

01:54:40.340 --> 01:54:46.127
you a full and transparent understanding
of how we put all of this together.

01:54:46.210 --> 01:54:50.647
One of the things I'm particularly excited
about is that we have been carefully

01:54:50.730 --> 01:54:53.037
codesigning our models
with our own silicon.

01:54:53.120 --> 01:54:55.767
That means that we've optimized

01:54:55.850 --> 01:54:59.107
MAI thinking one on our very own

01:54:59.190 --> 01:55:04.047
myo200 Chip and benchmarked

01:55:04.130 --> 01:55:07.597
against the GB2300.

01:55:07.680 --> 01:55:09.917
So top of the improvement, we're

01:55:10.000 --> 01:55:13.727
seeing a further 1.4X

01:55:13.810 --> 01:55:19.027
performance per Watt gain when
we run them on the MIO model end

01:55:19.110 --> 01:55:23.217
to end, and that's huge.
As everybody knows, this scale,

01:55:23.300 --> 01:55:27.007
every Watt counts.
And silicon and model codesign

01:55:27.090 --> 01:55:31.007
is a really key advantage that
we think is going to help keep

01:55:31.090 --> 01:55:36.047
everybody here right on the
Frontier with the most efficient

01:55:36.130 --> 01:55:38.667
and powerful thinking
and coding agents out there.

01:55:38.750 --> 01:55:41.367
And we're excited that these
faster and more efficient MAI

01:55:41.450 --> 01:55:45.057
models are coming to the N1X
that Satya mentioned a few

01:55:45.140 --> 01:55:48.087
moments ago, and we think that's
going to be able to deliver the

01:55:48.170 --> 01:55:52.697
very best performance on mounds

01:55:52.780 --> 01:55:55.847
in a new -- on windows
in a few months time.

01:55:55.930 --> 01:56:00.637
This is what owning the full
end to end stack looks like.

01:56:00.720 --> 01:56:04.577
It lets you customize the MAI
models using our full hill

01:56:04.660 --> 01:56:05.777
climbing machine right where you want it.

01:56:05.860 --> 01:56:10.017
And it means that the discipline
and very relentless engineering

01:56:10.100 --> 01:56:13.767
that has gone into building our models
is now available to all of you.

01:56:13.850 --> 01:56:17.987
On a platform that you can trust,
working on your behalf,

01:56:18.070 --> 01:56:22.257
to create custom agents
that you will control.

01:56:22.340 --> 01:56:26.927
So the really big thing, of course, that's
happened in the last year is these RLEs.

01:56:27.010 --> 01:56:28.467
Reinforcement learning environments.

01:56:28.550 --> 01:56:31.447
These unique training gyms for your AIs.

01:56:31.530 --> 01:56:35.767
They create company
and task-specific agents, adapted

01:56:35.850 --> 01:56:38.777
only to you, built on MAI models.

01:56:38.860 --> 01:56:40.447
For example, within Microsoft we

01:56:40.530 --> 01:56:42.947
use our RLEs combined with MAI

01:56:43.030 --> 01:56:47.297
models to climb towards the best

01:56:47.380 --> 01:56:48.637
agentic ace uses on Excel.

01:56:48.720 --> 01:56:53.787
The model is on par with GPT5.4

01:56:53.870 --> 01:56:59.767
on public and private benchmarks, and
being ten times more efficient on cost.

01:56:59.850 --> 01:57:03.187
[Cheering] Thank -- [chuckling]
And many other earlier adopters

01:57:03.270 --> 01:57:04.947
are seeing similar results.

01:57:05.030 --> 01:57:07.867
When we've tuned our models on

01:57:07.950 --> 01:57:12.177
McKensie's tasks,
it delivered the highest win rate,

01:57:12.260 --> 01:57:16.747
outperforming GPT5.5, and delivering
10X grater efficiency on cost.

01:57:16.830 --> 01:57:21.197
To us, this is the advantage
of very carefully calibrated

01:57:21.280 --> 01:57:25.057
frontier tuning.
And importantly, unlike with

01:57:25.140 --> 01:57:26.187
some of the other companies,

01:57:26.270 --> 01:57:33.270
with shared model

01:57:35.080 --> 01:57:39.307
only you keep benefits from the
hard earned workflows, know how,

01:57:39.390 --> 01:57:41.707
knowledge and your own institutional data.

01:57:41.790 --> 01:57:45.037
Only you get to control
the resulting model.

01:57:45.120 --> 01:57:47.707
So with us, the RLEs and the

01:57:47.790 --> 01:57:50.377
models that you build inside of

01:57:50.460 --> 01:57:55.237
them they become your mote.

01:57:55.320 --> 01:58:00.027
It marks a new era in AI that we
are all really very excited about.

01:58:00.110 --> 01:58:04.157
Okay. So now just one final announcement
that I am very excited about.

01:58:04.240 --> 01:58:05.897
We are taking customization and

01:58:05.980 --> 01:58:08.197
co creation of our models to the

01:58:08.280 --> 01:58:12.537
highest level possible on what I

01:58:12.620 --> 01:58:16.707
think offers the most important
application of AI, healthcare.

01:58:16.790 --> 01:58:18.077
Today we are very proud to be

01:58:18.160 --> 01:58:19.777
announcing we are partnering

01:58:19.860 --> 01:58:24.077
with Mayo Clinic for a new

01:58:24.160 --> 01:58:27.137
frontier for health and deploy
it in their hospitals and beyond.

01:58:27.220 --> 01:58:30.787
Help me in welcoming to the stage
physician ground breaking

01:58:30.870 --> 01:58:32.957
researcher, president and CEO of

01:58:33.040 --> 01:58:40.040
the Mayo Clinic Dr. Gianrico

01:58:42.830 --> 01:58:47.390
Farrugia.

01:58:47.490 --> 01:58:50.807
Thank you for being here.
Everyone recognizes Mayo as the

01:58:50.890 --> 01:58:53.977
leading hospital in the world
with incredible track record of

01:58:54.060 --> 01:58:55.277
research, innovation and

01:58:55.360 --> 01:58:56.217
clinical practice.

01:58:56.300 --> 01:59:01.637
Tell us more about what hope to

01:59:01.720 --> 01:59:07.087
get collaboration. >> Mayo Clinic is known
for being able to live up to primary

01:59:07.170 --> 01:59:09.687
value the needs
of the patient comes first.

01:59:09.770 --> 01:59:11.597
We deliver outstanding healthcare.

01:59:11.680 --> 01:59:14.127
We are ranked number one
healthcare organization in the

01:59:14.210 --> 01:59:17.047
world, yet we know most people
in the world will not have

01:59:17.130 --> 01:59:21.407
access to Mayo Clinic.
So seven years ago we decided to

01:59:21.490 --> 01:59:23.807
create a platform, Mayo Clinic

01:59:23.890 --> 01:59:29.577
platform, moving all
of healthcare from a pipeline to a

01:59:29.660 --> 01:59:33.567
platform. With our partners that
platform is in four continents and

01:59:33.650 --> 01:59:35.787
reaches about 100 million people.

01:59:35.870 --> 01:59:38.817
It has created the largest, to

01:59:38.900 --> 01:59:42.987
our knowledge, longitudal

01:59:43.070 --> 01:59:47.497
healthcare center in the world,
multimodal genomics.

01:59:47.580 --> 01:59:50.747
We have the opportunity to do
what we do best together which

01:59:50.830 --> 01:59:52.967
is create a frontier model for healthcare.

01:59:53.050 --> 01:59:55.977
What it means if you are
a patient, if you are somebody

01:59:56.060 --> 02:00:01.347
interested in healthcare,
you can get clinical and logistical

02:00:01.430 --> 02:00:04.347
answers to your healthcare.
If your healthcare provider,

02:00:04.430 --> 02:00:10.217
your physician it can give
you insight, it can act as your

02:00:10.300 --> 02:00:13.957
real-time team member it can
tell you what is likely to

02:00:14.040 --> 02:00:17.397
happen next, but it can also
prevent harm and therefore

02:00:17.480 --> 02:00:20.227
increase patient safety and give
valuable insights that make the

02:00:20.310 --> 02:00:24.787
team better at giving you what you
need most which is better healthcare.

02:00:24.870 --> 02:00:28.407
>> One of the things we are most
excited about is the models are

02:00:28.490 --> 02:00:32.407
already pretty incredible
about textbook knowledge.

02:00:32.490 --> 02:00:34.377
Read all of the journals and papers.

02:00:34.460 --> 02:00:39.667
What they are looking is clinical practice
and clinical expertise of your team and

02:00:39.750 --> 02:00:42.417
clinicians you developed
over the last many decades.

02:00:42.500 --> 02:00:47.087
How do you think you might go
about the clinical practice for

02:00:47.170 --> 02:00:50.727
the model and production?
>> The exciting part here is we

02:00:50.810 --> 02:00:54.567
each do what we do best.
We can tackle something that

02:00:54.650 --> 02:00:58.767
Hazel lewded healthcare for a long
time trusted scalable solutions.

02:00:58.850 --> 02:01:02.007
To do that you need to have
the right data, certainly need to

02:01:02.090 --> 02:01:07.147
have the right people, but you need
to have a very patient focused lens.

02:01:07.230 --> 02:01:10.617
Between the two of us we have
all of this together so we can

02:01:10.700 --> 02:01:13.967
build this frontier model and

02:01:14.050 --> 02:01:19.087
offer safe, secure,
trustworthy and effective healthcare

02:01:19.170 --> 02:01:23.797
solutions for all. >> Number one objective
is put the patient first, deliver high

02:01:23.880 --> 02:01:28.397
quality that we can as trusted
as we can and share the world over.

02:01:28.480 --> 02:01:30.087
Excited to share this partnership.

02:01:30.170 --> 02:01:31.887
Can't wait to share more in the future.

02:01:31.970 --> 02:01:37.190
>> Us too. Thank you.

02:01:37.420 --> 02:01:40.777
>> So today marks some very,
very exciting steps that we are

02:01:40.860 --> 02:01:45.597
taking on our journey to create
humanist super intelligence at

02:01:45.680 --> 02:01:49.647
Microsoft. We have an incredible roster
of seven new world class models to

02:01:49.730 --> 02:01:51.617
keep everybody working
at the absolute frontier.

02:01:51.700 --> 02:01:55.257
We are really looking forward
to everybody being able to co

02:01:55.340 --> 02:01:57.957
create your own unique agents

02:01:58.040 --> 02:02:01.497
adapted to you at your control.

02:02:01.580 --> 02:02:04.187
I feel this is a new era in AI.

02:02:04.270 --> 02:02:06.607
An era of AI that you
control on your terms.

02:02:06.690 --> 02:02:07.967
Let's build it together.

02:02:08.050 --> 02:02:15.050
Thank you very much everyone.

02:02:17.760 --> 02:02:20.397
Now over to Tanaya to show us

02:02:20.480 --> 02:02:23.387
how frontier learning works in

02:02:23.470 --> 02:02:28.610
practice.

02:02:28.870 --> 02:02:33.267
>> Thanks Mustafa.
With frontier tuning we are

02:02:33.350 --> 02:02:37.087
making it possible for you
to create your own enterprise AI,

02:02:37.170 --> 02:02:42.637
with the models and the harness
tuned on your data and your workflows.

02:02:42.720 --> 02:02:43.707
Let me show you how you can

02:02:43.790 --> 02:02:49.147
build your own health planning

02:02:49.230 --> 02:02:54.037
machine. MAI thinking one is available
in private preview in the foundry

02:02:54.120 --> 02:02:54.437
model catalog.

02:02:54.520 --> 02:02:59.157
You can go ahead and deploy the

02:02:59.240 --> 02:03:03.357
mo model as is or if you want to go

02:03:03.440 --> 02:03:05.357
on your health journey click the

02:03:05.440 --> 02:03:07.767
fine tune button.

02:03:07.850 --> 02:03:13.167
The first thing I will set.

02:03:13.250 --> 02:03:15.167
Next I am going to add a greater

02:03:15.250 --> 02:03:20.477
and that's it.
I go ahead andee ate a -- and

02:03:20.560 --> 02:03:22.077
create a job.

02:03:22.160 --> 02:03:24.947
It has through logs, scores them

02:03:25.030 --> 02:03:28.497
and we start the hill climb.

02:03:28.580 --> 02:03:32.817
If you want full control over
the training loop let me give

02:03:32.900 --> 02:03:35.827
you sneak peak into
low level training API.

02:03:35.910 --> 02:03:39.697
You can see I have the MEI sinking model.

02:03:39.780 --> 02:03:42.417
I can also configure my role out

02:03:42.500 --> 02:03:44.687
strategy as well as my hyper

02:03:44.770 --> 02:03:50.207
parameters to define exactly how
I want my training algorithm to

02:03:50.290 --> 02:03:50.677
work.

02:03:50.760 --> 02:03:54.547
In fact I can also incorporate

02:03:54.630 --> 02:03:57.847
my own RLM by defining the tools

02:03:57.930 --> 02:03:59.297
this model interacts with.

02:03:59.380 --> 02:04:02.837
You just saw all of the code,

02:04:02.920 --> 02:04:05.687
but as M365 customer you never

02:04:05.770 --> 02:04:10.327
start from scratch.
Let's hop on to Copilot.

02:04:10.410 --> 02:04:11.927
As a part of frontier tuning we

02:04:12.010 --> 02:04:15.167
are introducing a new way to

02:04:15.250 --> 02:04:18.087
build environments based on your

02:04:18.170 --> 02:04:20.437
data and your workflows.

02:04:20.520 --> 02:04:25.637
One of our customers land
O lakes one of the largest

02:04:25.720 --> 02:04:27.777
businesses in the United States

02:04:27.860 --> 02:04:33.867
is using this to perfect
that butter on your moerning --

02:04:33.950 --> 02:04:36.917
morning toast. Let me walk
you through their environment.

02:04:37.000 --> 02:04:39.427
The environment has soft skills,

02:04:39.510 --> 02:04:41.327
knowledge and tools.

02:04:41.410 --> 02:04:46.797
But in the back end we create
an entire origin for your agents to

02:04:46.880 --> 02:04:47.767
continuously learn the way that

02:04:47.850 --> 02:04:54.207
you work. Let's look at the butter report

02:04:54.290 --> 02:04:56.837
generation here.
These tasks are very complex.

02:04:56.920 --> 02:05:01.207
They require many manual steps

02:05:01.290 --> 02:05:02.717
and a high degree of precision.

02:05:02.800 --> 02:05:06.467
In these tasks almost 80%
accuracy isn't good enough.

02:05:06.550 --> 02:05:10.217
To hill climb to higher accuracy,
we are extending the

02:05:10.300 --> 02:05:12.787
industry definition of skills to

02:05:12.870 --> 02:05:18.227
include Rubriks on what good looks like.

02:05:18.310 --> 02:05:22.577
This is one task. How do I
scale this to qualify all o

02:05:22.660 --> 02:05:24.247
all of the tasks in enterprise.

02:05:24.330 --> 02:05:28.437
You spend a lot of time in M365

02:05:28.520 --> 02:05:31.807
in teams, outlook, word,
excel and PowerPoint.

02:05:31.890 --> 02:05:35.707
We use those signals to suggest
skills and rubrics that define

02:05:35.790 --> 02:05:40.137
the way you work. Next you can add your

02:05:40.220 --> 02:05:43.687
organizational knowledge for branding
like from OneDrive and SharePoint.

02:05:43.770 --> 02:05:47.927
The environment comes built
in with Microsoft tools and you can

02:05:48.010 --> 02:05:50.527
add custom tools to it.

02:05:50.610 --> 02:05:52.597
Because these tools tap into

02:05:52.680 --> 02:05:57.817
real workflows we simulate

02:05:57.900 --> 02:06:01.167
execution so the model can learn
without actually impacting the

02:06:01.250 --> 02:06:07.177
live state of your business.
Now my most favorite part, the

02:06:07.260 --> 02:06:09.317
science. By generalizing all of these

02:06:09.400 --> 02:06:14.567
learnings into the main model
as well as in the embedding model,

02:06:14.650 --> 02:06:20.857
we are able to hill climb
for tasks that require high

02:06:20.940 --> 02:06:21.227
accuracy.

02:06:21.310 --> 02:06:26.157
And not only this, we are able

02:06:26.240 --> 02:06:30.747
to hill climb to more than 90%

02:06:30.830 --> 02:06:32.417
accuracy for Land-O-Lakes class

02:06:32.500 --> 02:06:35.737
for using the model.

02:06:35.820 --> 02:06:38.837
We estimate the model to be 10X

02:06:38.920 --> 02:06:45.677
efficient than the baseline

02:06:45.760 --> 02:06:47.367
models. Let's run the task.

02:06:47.450 --> 02:06:50.657
This time using this environment

02:06:50.740 --> 02:06:54.887
as an entrancing harness.
Usually this task takes a couple

02:06:54.970 --> 02:06:57.497
of minutes to run.

02:06:57.580 --> 02:07:04.580
In the meantime I will show you

02:07:07.720 --> 02:07:12.377
a carshed response. They have research
with multiple models including a fine tune

02:07:12.460 --> 02:07:15.477
model. What you see here is a summary

02:07:15.560 --> 02:07:20.967
that doesn't feel generic,
it feels undoubtedly Land-O-Lakes.

02:07:21.050 --> 02:07:25.387
That's not it.
The task holds itself to high

02:07:25.470 --> 02:07:31.297
standards. It continuously retrospect and

02:07:31.380 --> 02:07:35.767
evaluate success.
With frontier tuning your agent

02:07:35.850 --> 02:07:37.847
continuously improves with

02:07:37.930 --> 02:07:41.267
environment on your data

02:07:41.350 --> 02:07:45.137
encoding the way you work.

02:07:45.220 --> 02:07:47.947
That is what I call frontier

02:07:48.030 --> 02:07:53.397
tuning as smooth as butter.

02:07:53.480 --> 02:07:57.957
We can't wait to see
the environment you build.

02:07:58.040 --> 02:08:00.510
Back to you Satya.

02:08:00.870 --> 02:08:01.287
>> Thank you very much.

02:08:01.370 --> 02:08:08.370
Thank you for the partnership

02:08:08.980 --> 02:08:11.387
here. What you saw is a pretty

02:08:11.470 --> 02:08:14.707
significant shift.

02:08:14.790 --> 02:08:19.847
Times come from just consuming
frontier model to fully

02:08:19.930 --> 02:08:20.847
participating at the frontier

02:08:20.930 --> 02:08:25.217
and frontier ecosystem.

02:08:25.300 --> 02:08:28.367
That is the transition.

02:08:28.450 --> 02:08:30.857
You can have private evals and

02:08:30.940 --> 02:08:33.117
outcomes private RLEs and

02:08:33.200 --> 02:08:36.187
traces, your enterprise

02:08:36.270 --> 02:08:38.827
knowledge create scaffolding

02:08:38.910 --> 02:08:42.127
from models to hill climb.

02:08:42.210 --> 02:08:44.987
That will allow you to

02:08:45.070 --> 02:08:50.067
differentiate on-line what you control.

02:08:50.150 --> 02:08:51.737
There is a new operating point

02:08:51.820 --> 02:08:53.147
at the frontier.

02:08:53.230 --> 02:08:59.217
Where you can use a very effic

02:08:59.300 --> 02:09:04.087
efficient reasoning model
and coding model and achieve

02:09:04.170 --> 02:09:07.687
frontier because you have done
the hard work to create the RLE

02:09:07.770 --> 02:09:10.727
the hill climbing machine that

02:09:10.810 --> 02:09:13.767
create the model your hill

02:09:13.850 --> 02:09:16.847
traces can make to the frontier.

02:09:16.930 --> 02:09:19.767
We think the combination
of the two is a game-changer and how

02:09:19.850 --> 02:09:24.037
people think about what does it
mean to operate in the frontier.

02:09:24.120 --> 02:09:27.747
What does the token look like
how are you in control what is

02:09:27.830 --> 02:09:31.717
the future of the firm those
are the questions and really the

02:09:31.800 --> 02:09:33.897
ecosystem that gets built around

02:09:33.980 --> 02:09:36.757
as opposed to a few models

02:09:36.840 --> 02:09:41.487
hungry for all data.

02:09:41.570 --> 02:09:44.897
To close out I want to talk

02:09:44.980 --> 02:09:48.367
about the frontiers beyond to

02:09:48.450 --> 02:09:50.687
push the frontier what's the

02:09:50.770 --> 02:09:57.337
next best thing.

02:09:57.420 --> 02:09:59.077
Building the scientific

02:09:59.160 --> 02:10:03.747
discovery loop can have perhaps

02:10:03.830 --> 02:10:10.257
the biggest societal impact.
Science today is a little too leni

02:10:10.340 --> 02:10:14.157
lenient. You form a hypothesis, you run

02:10:14.240 --> 02:10:18.997
an experiment you wait for lab
results and then you begin

02:10:19.080 --> 02:10:24.247
again. What if the scientific
method itself can become more

02:10:24.330 --> 02:10:25.847
continuous, more program able.

02:10:25.930 --> 02:10:29.367
That's what we are doing
with Microsoft discovery.

02:10:29.450 --> 02:10:30.907
Discovery brings together the

02:10:30.990 --> 02:10:36.907
models the ACP compute,
the knowledge graphs of all of the

02:10:36.990 --> 02:10:37.777
scientific knowledge automated

02:10:37.860 --> 02:10:41.137
lab and simulation into one

02:10:41.220 --> 02:10:45.357
agentic discovery loop.

02:10:45.440 --> 02:10:50.327
I am thrilled to GA and Microsoft
Discovery, to show you

02:10:50.410 --> 02:10:54.110
this in action let me
invite David up on the stage.

02:10:54.480 --> 02:10:56.947
>> Hello. I work in the Microsoft

02:10:57.030 --> 02:10:59.737
Discovery and quantum team.

02:10:59.820 --> 02:11:02.067
Today to recycle this plastic

02:11:02.150 --> 02:11:05.807
like this bottle you have to

02:11:05.890 --> 02:11:09.717
sled it and melt it.
You cannot use it to make this

02:11:09.800 --> 02:11:13.227
bottle again. It is down cycling.

02:11:13.310 --> 02:11:17.057
We are using Microsoft discovery
to advance that research.

02:11:17.140 --> 02:11:22.187
Let me show it to you.
This is the Microsoft discovery

02:11:22.270 --> 02:11:24.757
app. It is VS code.

02:11:24.840 --> 02:11:29.027
Agentic Discovery has many with

02:11:29.110 --> 02:11:34.067
software engineers.
I want to do three things as a

02:11:34.150 --> 02:11:36.307
scientist. I want to write a scientific

02:11:36.390 --> 02:11:40.707
paper about this topic.

02:11:40.790 --> 02:11:43.317
Instead of melting the plastic I

02:11:43.400 --> 02:11:45.577
can use proteins so you can

02:11:45.660 --> 02:11:48.697
recycle it again and again.

02:11:48.780 --> 02:11:50.857
Second I want to perform the

02:11:50.940 --> 02:11:56.127
actual discovery to define this.

02:11:56.210 --> 02:12:01.297
Third I want to create
a protocol to test the results in

02:12:01.380 --> 02:12:02.227
a real lab.

02:12:02.310 --> 02:12:04.587
If you are a developer the steps

02:12:04.670 --> 02:12:11.670
will be very familiar to you.

02:12:15.560 --> 02:12:16.717
Let's launch it.

02:12:16.800 --> 02:12:21.097
This will is discovery engine.

02:12:21.180 --> 02:12:24.617
These are specialized agents
always running in the background

02:12:24.700 --> 02:12:26.387
following the scientific methods.

02:12:26.470 --> 02:12:28.687
You can see the agents you can add more.

02:12:28.770 --> 02:12:30.797
Microsoft discovery includes a

02:12:30.880 --> 02:12:34.027
community of agents, models and

02:12:34.110 --> 02:12:37.077
tools across many domains.

02:12:37.160 --> 02:12:41.137
You can have officers fair
parties or create your own.

02:12:41.220 --> 02:12:42.507
This is still running.

02:12:42.590 --> 02:12:44.407
It will be running for a while.

02:12:44.490 --> 02:12:51.490
Even hours or days.

02:12:55.820 --> 02:12:59.017
It is running simulations

02:12:59.100 --> 02:13:03.357
dynamically. You can see
the files created here.

02:13:03.440 --> 02:13:05.497
This one is the research paper

02:13:05.580 --> 02:13:08.567
that I was asking for.

02:13:08.650 --> 02:13:14.837
Discovery using no less graph

02:13:14.920 --> 02:13:16.367
internally. It is internal knowledge.

02:13:16.450 --> 02:13:19.207
This is a critical asset because

02:13:19.290 --> 02:13:23.247
it provide compete visibility.

02:13:23.330 --> 02:13:30.330
Scientists can be in full

02:13:34.310 --> 02:13:35.357
control.

02:13:35.440 --> 02:13:42.440
How do you come up with the

02:13:44.700 --> 02:13:45.887
plastic.

02:13:45.970 --> 02:13:47.807
You need a way to generate the

02:13:47.890 --> 02:13:53.237
proteins and identify the most

02:13:53.320 --> 02:14:00.320
promising ones.

02:14:00.580 --> 02:14:02.397
There is no out of the box agent for that.

02:14:02.480 --> 02:14:07.227
That's okay. With Microsoft discovery
there is not an agent for the task.

02:14:07.310 --> 02:14:10.227
You can create one on the
fly which is pretty cool.

02:14:10.310 --> 02:14:12.397
Here is all of the files that it

02:14:12.480 --> 02:14:19.187
created. This one is the YAMMER for the

02:14:19.270 --> 02:14:21.737
agent. Now we need to generate

02:14:21.820 --> 02:14:22.967
candidates for the proteins.

02:14:23.050 --> 02:14:28.277
That task requires a lot of

02:14:28.360 --> 02:14:31.517
compute. It is integrated with HPC.

02:14:31.600 --> 02:14:36.137
You can use it. You can see the process.

02:14:36.220 --> 02:14:40.587
You start with C protein
and created variations by replacing

02:14:40.670 --> 02:14:41.527
small segments.

02:14:41.610 --> 02:14:47.227
You can see the segments there.

02:14:47.310 --> 02:14:51.847
You apply to see if it helps or

02:14:51.930 --> 02:14:52.517
not.

02:14:52.600 --> 02:14:56.937
This is done millions of times

02:14:57.020 --> 02:15:01.347
exploring a huge tree of proteins.

02:15:01.430 --> 02:15:03.607
The result is 80 proteins sent

02:15:03.690 --> 02:15:06.467
to the lab for testing.

02:15:06.550 --> 02:15:08.537
It is like the software.

02:15:08.620 --> 02:15:11.757
Creating a protein is a little

02:15:11.840 --> 02:15:17.897
more complicated.
The bacteria create the protein

02:15:17.980 --> 02:15:20.827
for me. So discovery created another

02:15:20.910 --> 02:15:24.847
file.

02:15:24.930 --> 02:15:26.017
This one here.

02:15:26.100 --> 02:15:30.067
This contains the DNA sequences.

02:15:30.150 --> 02:15:37.150
I can send them to the lab if I

02:15:37.460 --> 02:15:42.437
have a lab I can integrate it

02:15:42.520 --> 02:15:48.527
directly with the agent.
Go to proteins and submit to

02:15:48.610 --> 02:15:52.697
lab. This will use a custom agent
that sends instructions for the

02:15:52.780 --> 02:15:54.097
lab appointment.

02:15:54.180 --> 02:16:00.267
This is very real.
That's an maut ma automated lab

02:16:00.350 --> 02:16:03.237
I have it right here.
This is application controlling the lab.

02:16:03.320 --> 02:16:05.137
It has a Copilot interface so

02:16:05.220 --> 02:16:08.147
the scientist can interact with

02:16:08.230 --> 02:16:11.917
the lab to the science

02:16:12.000 --> 02:16:15.567
experiment. In this case Discovery has

02:16:15.650 --> 02:16:19.587
submitted the job.
You can see it right here.

02:16:19.670 --> 02:16:22.887
This will take some time.
Let me open a previous run.

02:16:22.970 --> 02:16:26.767
There you go. These are all
of the steps that the agents are doing.

02:16:26.850 --> 02:16:30.987
Most of them completely
automated with human supervision.

02:16:31.070 --> 02:16:38.070
How cool is that?

02:16:39.290 --> 02:16:41.707
This is bringing the together

02:16:41.790 --> 02:16:45.817
the physical and agents in

02:16:45.900 --> 02:16:50.217
unified Discovery loop.
This is one example of what

02:16:50.300 --> 02:16:54.017
Microsoft Discovery can do.
Industries are using it today to

02:16:54.100 --> 02:16:57.657
embrace a new era of scientific discovery.

02:16:57.740 --> 02:17:02.667
Thank you. Back to you Satya.
>> Thank you so much David.

02:17:02.750 --> 02:17:04.587
Talking about scientific

02:17:04.670 --> 02:17:05.587
discovery, we are also

02:17:05.670 --> 02:17:12.337
continuing to make rapid

02:17:12.420 --> 02:17:16.247
progress on our long-term goal
of building a scalable computer.

02:17:16.330 --> 02:17:19.447
We announced last year our first QPU.

02:17:19.530 --> 02:17:21.067
We created a new state of matter

02:17:21.150 --> 02:17:27.787
that was only theorized 100

02:17:27.870 --> 02:17:29.927
years ago. Proved it out that it exists.

02:17:30.010 --> 02:17:33.097
Our vision was to take
a radically different approach to

02:17:33.180 --> 02:17:37.217
addressing the fundamental
barriers to building a scalable

02:17:37.300 --> 02:17:41.607
quantum machine which is all about
reliability, speed as well as size.

02:17:41.690 --> 02:17:43.767
Since then, we have continued to

02:17:43.850 --> 02:17:47.037
make progress across the full

02:17:47.120 --> 02:17:49.177
quantum stack with both our

02:17:49.260 --> 02:17:52.547
academic and industry partners.

02:17:52.630 --> 02:17:57.717
In Q north we will have
a quantum computer powered by atom

02:17:57.800 --> 02:18:00.487
computers natural atom computers
with our stack in there.

02:18:00.570 --> 02:18:02.127
We are working with algorithmic

02:18:02.210 --> 02:18:05.497
and Columbia and Zurich.

02:18:05.580 --> 02:18:08.627
We ourselves use this discovery

02:18:08.710 --> 02:18:11.387
agentic loop to accelerate the

02:18:11.470 --> 02:18:14.337
work in quantum compressing

02:18:14.420 --> 02:18:18.277
years of research into this last

02:18:18.360 --> 02:18:18.537
year.

02:18:18.620 --> 02:18:23.447
Today I am really thrilled to

02:18:23.530 --> 02:18:30.530
announce Majorana2.

02:18:39.240 --> 02:18:42.727
The resulting cube bits are

02:18:42.810 --> 02:18:46.427
reliable and maintaining
their state much longer while other

02:18:46.510 --> 02:18:48.337
common approaches deliver a

02:18:48.420 --> 02:18:52.677
lifetime of micro seconds.

02:18:52.760 --> 02:18:57.607
Majorana 2 provide cubic
lifetime of 20 seconds or up to

02:18:57.690 --> 02:19:00.777
even a minute.
Essentially a thousand times

02:19:00.860 --> 02:19:02.177
higher than what we were able to

02:19:02.260 --> 02:19:09.260
achieve with high -- my or --

02:19:11.640 --> 02:19:14.827
Majorana 1. They are enabling complex

02:19:14.910 --> 02:19:18.727
quantum computation in that lifetime.

02:19:18.810 --> 02:19:21.407
All of this in the same cubic

02:19:21.490 --> 02:19:27.737
form factor of Majorana 1 in one
100th of a millimeter control of

02:19:27.820 --> 02:19:30.537
digital. Which again
is a super important aspect.

02:19:30.620 --> 02:19:35.337
Making it all possible to fit
a million of these cube bits in a

02:19:35.420 --> 02:19:37.187
chip smaller than a credit card.

02:19:37.270 --> 02:19:39.457
It is this combination of the

02:19:39.540 --> 02:19:41.127
reliability, the speed, the size

02:19:41.210 --> 02:19:44.357
that makes the topological

02:19:44.440 --> 02:19:46.857
approach so unique.

02:19:46.940 --> 02:19:48.357
With Majorana 1 we had proven

02:19:48.440 --> 02:19:52.057
out the foundational physics and

02:19:52.140 --> 02:19:54.377
with Majorana 2 we begin the

02:19:54.460 --> 02:19:59.760
engineering scale.

02:19:59.860 --> 02:20:02.907
Yup.

02:20:02.990 --> 02:20:04.807
Ultimately it is never about

02:20:04.890 --> 02:20:07.877
tech for tech sake.

02:20:07.960 --> 02:20:12.027
It is about tackling those
pressing challenges of people

02:20:12.110 --> 02:20:16.787
and planet. It is also the fundamental
point of this conference.

02:20:16.870 --> 02:20:20.857
The question is not about
whether you can build the next

02:20:20.940 --> 02:20:24.957
great model or platform
even this quantum machine.

02:20:25.040 --> 02:20:29.477
It is the question of how do
we build the frontier ecosystem

02:20:29.560 --> 02:20:32.677
together. Because there are really two

02:20:32.760 --> 02:20:36.677
stories people can tell about this moment.

02:20:36.760 --> 02:20:39.607
One is that technology
concentrates power, reduces

02:20:39.690 --> 02:20:43.927
human agency and leaves
the society to absorb the

02:20:44.010 --> 02:20:46.887
consequences.
The other is that we use this

02:20:46.970 --> 02:20:49.587
next wave to unlock opportunity

02:20:49.670 --> 02:20:52.487
for developers, scientists,

02:20:52.570 --> 02:20:54.027
enterprises in every community.

02:20:54.110 --> 02:20:57.587
Our job is to make the second

02:20:57.670 --> 02:21:01.977
story true. That's our North
Star for the frontier ecosystem.

02:21:02.060 --> 02:21:03.737
Let's all go build together.

02:21:03.820 --> 02:21:06.577
Thank you all very, very much.

02:21:06.660 --> 02:21:13.660
>> Thank you.

02:21:59.840 --> 02:22:04.187
>> Within reach.

02:22:04.270 --> 02:22:07.777
Quantum is our most fundamental

02:22:07.860 --> 02:22:12.607
physical theorem.

02:22:12.690 --> 02:22:14.337
For quantum to deliver real

02:22:14.420 --> 02:22:16.637
world impact it has to be both

02:22:16.720 --> 02:22:21.907
reliable and scalable.
>> Reaching quantum at scale

02:22:21.990 --> 02:22:25.157
will allow us to make a more
direct approach to solving

02:22:25.240 --> 02:22:27.687
problems in chemistry.

02:22:27.770 --> 02:22:30.457
>> Discovery in life sciences.

02:22:30.540 --> 02:22:35.957
>> We bring drugs to market it
opens up millions of possibilities.

02:22:36.040 --> 02:22:37.027
Quantum will change the world.

02:22:37.110 --> 02:22:44.110
This change is just beginning to

02:22:46.090 --> 02:22:48.290
happen.