00:00:00
All right, everybody. We're thrilled to
00:00:02
have the one, the only Tata Nadella
00:00:04
here, the third CEO of Microsoft for a
00:00:08
uh impromptu fireside chat with David
00:00:11
Saxs, Arzar of AI and crypto. Satia
00:00:17
third, CEO of Microsoft, born in India.
00:00:21
What an incredible story. came here
00:00:24
right after college and uh you had a
00:00:27
little round trip to pick up your wife
00:00:30
in your book to to bring her here. Tell
00:00:31
everybody briefly uh how that occurred.
00:00:35
>> Well um uh you know so that's that's a
00:00:38
great story of uh the uh the labyrinth
00:00:42
that is the immigration policies of the
00:00:45
United States I think. Um I my wife and
00:00:49
I went to college together in India. I
00:00:51
came here for grad school. We then got
00:00:53
married. Uh I got my green card. Um and
00:00:58
she couldn't come join because we got
00:01:00
married. So the story goes basically I
00:01:04
had to give up my green card. So the
00:01:06
funny thing is I went to the American
00:01:07
embassy in Delhi
00:01:09
>> and I said where's the line to give up
00:01:11
my green card? And they said there is no
00:01:12
such line. Um
00:01:14
>> that would be a crazy thing to do in the
00:01:16
'9s. So it was a strange thing to give
00:01:18
up your green card, get an H1 so that
00:01:21
she could join, but it all worked out.
00:01:22
So um you know it's a long lost memory
00:01:25
but it was you know a way to work around
00:01:27
it. Um I wanted to ask you uh having
00:01:34
launched a co-pilot first with GitHub
00:01:37
then having a co-pilot on the desktop.
00:01:39
You made a very bold move for Microsoft
00:01:41
to put that in the Windows product,
00:01:43
which I use every day, on the desktop,
00:01:45
but you did that before it really could
00:01:47
recognize the file system and interact
00:01:49
with applications. Got a little bit of a
00:01:52
lukewarm reception, but now you've been
00:01:54
doubling down, doubling down, and there
00:01:56
seems to be, in my estimation, three
00:01:58
modalities for knowledge workers.
00:02:01
Elon's building at XAI,
00:02:04
what they're calling a human emulator,
00:02:06
if you saw that leak this week. Yeah.
00:02:08
uh where they're just building employees
00:02:10
and just putting them into their chat
00:02:12
rooms and email. Then you have Claude
00:02:14
came out with co-work this week.
00:02:16
Incredibly powerful. People are kind of
00:02:18
losing their minds over it. I've been
00:02:20
playing with it for the last 40 hours.
00:02:21
Truly impressive.
00:02:23
What's your vision for Microsoft and how
00:02:26
knowledge workers will actually put this
00:02:28
to use because there seems to be a gap
00:02:30
between you know playing around with
00:02:32
chatbt and getting some interesting
00:02:34
results and getting business results.
00:02:36
>> Yeah. So I think it one of the most um
00:02:39
perhaps illustrative examples um of
00:02:43
trying to understand these various form
00:02:44
factors is looking at coding which is
00:02:46
obviously a form of knowledge work or uh
00:02:49
probably the best example of knowledge
00:02:51
work and if you think about the journey
00:02:53
coding has been it started with u
00:02:56
essentially uh uh the next edit suggest
00:03:00
right that was the first time in fact my
00:03:02
own belief in this entire uh generation
00:03:05
of tech really sort of got formulated
00:03:08
where I started seeing I think it was G
00:03:10
you know there's a codeex model back in
00:03:12
the day it was preGPT35
00:03:14
uh that's when next edits started
00:03:17
working with some real accuracy then we
00:03:19
went to chat then we went to actions and
00:03:23
now to full autonomous agents and then
00:03:25
the autonomous agents can be both
00:03:28
foreground background in the cloud or
00:03:30
local right so that's all the form
00:03:32
factors that exist today when you're
00:03:34
coding and interestingly If you look at
00:03:36
it, you use all of them, right? It's not
00:03:38
like there's only one form factor. So
00:03:40
that's I think probably one of the other
00:03:41
lessons. So for example, when I'm in a
00:03:43
CLI, I can you go a foreground agent,
00:03:47
background agent, and then just
00:03:49
literally go edit in VS Code, right?
00:03:52
There all happening in parallel, right?
00:03:53
So that sort of shows how these form
00:03:56
factors even composed. So then you bring
00:03:59
that to knowledge work to your point. We
00:04:02
started with chat. chat with reasoning
00:04:06
sort of goes beyond just request
00:04:08
response because you now have that chain
00:04:10
of thought uh where you can see it work
00:04:13
now they're actions right essentially
00:04:15
either through computer use or through
00:04:18
uh AP you know basically skills uh and
00:04:21
agent calls so you can do actions so
00:04:24
that's kind of the state of the copilot
00:04:26
today now there is a way to think about
00:04:31
you know the the theory of the mind
00:04:33
evolution Right? Because you need like
00:04:34
if you remember you know Jobs had the
00:04:36
best line I would say for PCs or
00:04:38
computers was to say if you it's a
00:04:41
bicycle for the mind. Bill had a line
00:04:44
which I liked as well which was it's
00:04:45
information at your fingertips. We kind
00:04:48
of need now a new concept metaphor for
00:04:52
how we use computers in the AI age. And
00:04:54
you have one
00:04:55
>> and the one I like actually came from
00:04:56
the CEO of notion which I you know that
00:04:58
manager of
00:04:59
>> incredible product. Yeah.
00:05:00
>> You haven't bought it yet? I've not got
00:05:02
that. Uh but the it's both management of
00:05:06
you know basically
00:05:08
a manager of infinite minds. That's a
00:05:10
nice way to think about it right when
00:05:12
you sort of really look at all the
00:05:14
agents that you are working with. You
00:05:17
kind of need to understand what I in
00:05:19
fact the other term I like is we macro
00:05:21
delegate and micro steer. In fact you
00:05:25
kind of need that in in in coding you
00:05:27
kind of have it right. So you do a macro
00:05:29
delegation and then I can in parallel
00:05:31
give it instructions while it is doing
00:05:33
work. So that's sort of the state even
00:05:36
today of co-pilot or what have you. you
00:05:38
bring up a little bit of what one of one
00:05:40
of the form factors I'm very excited
00:05:42
about and you'll see us even in the next
00:05:44
week even uh do things is while I'm
00:05:47
sitting in GitHub copilot
00:05:50
what it's not as if software developers
00:05:52
sit in isolation right it's not like the
00:05:54
only thing I work on is my repo I attend
00:05:57
meetings um I write specs or others have
00:06:01
written specs that I'm implementing uh I
00:06:04
need to have my repo be consistent with
00:06:06
that so that means using either a
00:06:08
straightforward MCP server or a skill I
00:06:11
want to be able to call into my work IQ
00:06:14
which is the co-pilot bring that in
00:06:16
that's the type of composition uh of
00:06:20
knowledge work that'll happen same thing
00:06:21
with security say you're a security
00:06:23
professional you have lots of logs uh
00:06:25
how do you sort of really analyze them
00:06:27
you drop them into a file system then
00:06:29
write code on top of it create a
00:06:32
dashboard what have you those are the
00:06:33
types of knowledge work that we can
00:06:35
enable there I think you bring up one
00:06:38
more thing which is can you create quote
00:06:40
unquote digital employees digital
00:06:41
co-workers or what have you and it's all
00:06:43
about credentials right so the I today
00:06:46
you could like you can literally assign
00:06:47
>> are you working on that as well
00:06:48
>> yeah so in fact we introduced something
00:06:50
called agent 365 as a way to give
00:06:53
identities in fact extending the
00:06:55
identities we have for humans today uh
00:06:58
and the endpoint protection we have for
00:07:00
their compute devices to agents so
00:07:02
>> so you might clone me working in the HR
00:07:04
department or working in the marketing
00:07:06
department and have a virtual correct
00:07:07
>> version of me inside of office.
00:07:09
>> That's correct. So, so there are two
00:07:11
sort of modalities there. One is you
00:07:12
give every knowledge worker infinite
00:07:15
minds. That's kind of one and then you
00:07:18
create even infinite minds independent
00:07:21
of the your identity because the
00:07:22
identity is one of the key things you
00:07:24
got to get right even for it to work
00:07:26
right. Right. So
00:07:27
>> permissions
00:07:28
>> and decision making
00:07:29
>> permissions decision- making and like
00:07:31
one of the key in things is who did what
00:07:34
to whom is sort of the most important
00:07:36
query in an organization right at the
00:07:38
end of the day the organization needs to
00:07:40
understand what work got done h and
00:07:43
what's the provenence of that work uh
00:07:45
and how do you trace it back right so
00:07:47
therefore you kind of want either what
00:07:49
if it's a human with a lot of agents
00:07:52
then it's really macro delegation micro
00:07:55
steering by the human whose identity was
00:07:57
passed on. So it's delegation versus a
00:08:00
separate identity.
00:08:01
>> And that was done by a level of
00:08:03
management, product management that
00:08:05
you've eliminated, that Alphabets
00:08:07
eliminated. Meta has started to
00:08:09
eliminate in their organization four
00:08:11
years ago. You had the same number of
00:08:13
employees you have at Microsoft now, but
00:08:15
you put a $90 billion onto the top line
00:08:18
of the revenue in that time and you
00:08:20
doubled your income during that time. So
00:08:24
how did that happen? Is that automation
00:08:26
of those jobs? Is it you were a little
00:08:29
bit overstaffed?
00:08:31
>> Unpack.
00:08:31
>> I think it's it's actually you're
00:08:33
pulling on a very interesting thread
00:08:34
which is at some level what's the big
00:08:37
structural change that needs to happen.
00:08:38
In fact, I would say this is probably
00:08:41
the biggest change in knowledge work
00:08:44
since PCs. I mean I always you know
00:08:46
think about like how did work happen
00:08:48
precise right? I mean think about a
00:08:50
multinational company like ours trying
00:08:52
to do a forecast. uh right faxes went
00:08:55
around, inter office memos got sent and
00:08:58
then you kind of created a you know uh a
00:09:00
forecast then suddenly you know PCs
00:09:02
became standard issue uh you put an
00:09:04
Excel spreadsheet put some numbers sent
00:09:06
it in email everybody entered numbers
00:09:07
and you had a forecast so the work the
00:09:10
work artifact and the workflow all
00:09:12
changed that's what's happening so for
00:09:15
example I'll give you uh at LinkedIn
00:09:18
uh we used to have product managers we
00:09:20
had designers uh we had front-end
00:09:23
engineers and then we had CIS backend
00:09:25
engineers and so on. So what we did is
00:09:27
we sort of took those first four roles
00:09:30
and combined them in fact increased
00:09:32
scope and said let's they're all full
00:09:34
stack builders. So I like that because
00:09:37
that's a structural change that allows
00:09:39
for us to increase the change both the
00:09:43
work and the workflow between these
00:09:45
functions and I would assume the
00:09:46
velocity because you don't have four
00:09:48
people communicating and that throughput
00:09:50
of ideas it's just one person and vibe
00:09:52
coding. Exactly. And there's a new
00:09:53
workflow. So what at the same time as
00:09:55
you can imagine if to build an AI
00:09:57
product today there's a complete new
00:09:59
workflow right it starts with eval right
00:10:01
so basically there's this eval to
00:10:04
science to infrastructure and so eval
00:10:08
are done by these full stack builders
00:10:10
and what have you and product managers
00:10:12
in the new form the infrastructure is
00:10:14
built by the systems engineers at the
00:10:16
back end because they support the
00:10:17
science that supports the product. So in
00:10:19
some sense there's a new loop uh and you
00:10:22
have to structurally change and so a lot
00:10:24
of what is happening inside a tech is
00:10:27
that change uh which is I think going to
00:10:30
be pretty massive uh and at the same
00:10:32
time a company like ours I have to do
00:10:34
everything. It's not like I can just so
00:10:36
go live in the future. I have to make
00:10:37
sure we're doing a fantastic job of
00:10:39
doing hot patching on Windows is done
00:10:42
with quality uh while at the same time
00:10:44
building the evals that are improving
00:10:46
co-pilot quality. Right? And so both of
00:10:48
those have to be first class.
00:10:50
>> I assume this is the most challenging
00:10:51
moment of your career because Microsoft
00:10:54
was so dominant duopoly in some spaces.
00:10:58
Um but you really weren't up against the
00:11:01
competition level you're up against now.
00:11:03
I was talking to Elon, you know, and he
00:11:05
was sort of saying, well, building cars
00:11:07
was pretty easy. Uh because I was up
00:11:09
against the legacy car makers and now
00:11:11
I'm up against
00:11:13
just look at the set you're up against.
00:11:15
Yeah, it's it's a pretty intense time. I
00:11:18
mean, so the way I I I always think is
00:11:20
it's always helpful uh when you have a
00:11:23
complete new set of competitors every
00:11:25
decade because that keeps you fit. Uh if
00:11:28
you think about it, I joined uh
00:11:29
Microsoft in '92 when I had Noel as the
00:11:32
big existential competitor we had. Um
00:11:36
and here we are in 2026. Uh and it
00:11:39
you're absolutely right, it's a pretty
00:11:41
intense time. I'm glad there's the
00:11:42
competition. uh it's it's quite honestly
00:11:45
at the end of the day when I look at it
00:11:46
right as a percentage of GDP 5 years
00:11:49
from now where will tech be right uh it
00:11:52
will be higher so we're blessed to be in
00:11:54
this industry it's lot of intense
00:11:57
competition but it's not so zero sum as
00:12:00
some people make it out
00:12:02
>> it's getting much bigger
00:12:03
>> much the TAM and the you know just the
00:12:05
impact of this tech is going to be so
00:12:07
massive um the question then of course
00:12:09
is what is like I I always go back to
00:12:12
what's the brand identity Microsoft has
00:12:14
brand permission. We have what do
00:12:16
customers expect from us. It's sometimes
00:12:19
we kind of overthink somehow that every
00:12:22
customer wants the same thing from all
00:12:24
of the competitors and finding that out,
00:12:26
right? It's kind of a different take on
00:12:27
the Peter Theal thing which is you got
00:12:29
to avoid competition by really
00:12:32
understanding what customers really want
00:12:35
from you uh versus thinking everybody's
00:12:37
a competitor.
00:12:39
>> David.
00:12:40
>> Yeah. So there are a lot of heads of
00:12:41
state here obviously at Davos as well as
00:12:43
CEOs of Fortune 500 companies and I
00:12:46
think you got asked a question last
00:12:47
night at the dinner about how they
00:12:49
should think about AI and how to be
00:12:51
successful and I recall they used the
00:12:54
word diffusion and I was wondering if
00:12:56
you could expand on those remarks
00:12:58
because that really resonated with some
00:13:00
of the policy work I've been doing.
00:13:01
>> No, absolutely. In fact, uh what you all
00:13:03
have been doing to make sure in in this
00:13:05
context of the American tech stack um is
00:13:09
broadly used around the world and is
00:13:12
trusted around the world because I think
00:13:14
uh when I look back David to me um uh at
00:13:18
the end of the day you create the
00:13:21
technology but really the benefits
00:13:27
come only by intense use. In fact, one
00:13:30
of my favorite studies has always been
00:13:32
this work that uh an economist I think
00:13:34
out of Dartmouth did uh his name is
00:13:36
Diego Comman where he studied uh
00:13:39
basically what happened during the
00:13:40
industrial revolution um how did
00:13:42
countries get ahead? Um and the simple
00:13:45
sort of takeaway from that was any
00:13:48
country uh that brought the latest
00:13:50
technology into their uh country and
00:13:54
then did value add technology on top of
00:13:56
it. Right? So it's like don't reinvent
00:13:58
the wheel. Bring the latest and then
00:14:00
build on top of it. That's to me uh what
00:14:03
happens you know when you have
00:14:04
diffusion. So especially with general
00:14:06
purpose technologies like AI it needs to
00:14:08
spread like right in our you know in our
00:14:10
own country in the United States we now
00:14:12
need we have the tech. The question is
00:14:15
is it being used in healthcare? Is it
00:14:16
being used um in financial services? Is
00:14:19
it being used in every sector of the
00:14:21
economy by large businesses, small
00:14:22
business, public sector? Um so to me un
00:14:26
unless and until we see that diffusion
00:14:29
and intense use uh we're not going to
00:14:32
have the success. Um uh and so that's
00:14:34
the phase we are in it's f you know it's
00:14:36
diffusing faster um and so some of the
00:14:39
work policy work you have done um and in
00:14:42
general all you know the good news here
00:14:44
is the technologies there the rails
00:14:46
around cloud and mobile that were laid
00:14:49
out make it possible for this thing to
00:14:51
spread right it's not you know
00:14:52
impossible to get the tokens the
00:14:54
question is what are the use cases how
00:14:56
do and how do you manage the change in
00:14:58
all of that um you know like one of the
00:15:00
questions at least in Davos is it's one
00:15:02
for the west and the developed nations.
00:15:04
What about the global south? Um I think
00:15:07
global south has a huge opportunity too
00:15:09
quite frankly because to me like let's
00:15:11
say you know 40% 50% of the GDP of most
00:15:15
global south countries is public sector.
00:15:17
So just imagine this tech making a
00:15:21
difference in how the governments uh
00:15:23
really parlay their taxpayer money into
00:15:26
services for citizens and there's if
00:15:28
there's efficiency gains that's probably
00:15:30
couple of points of GDP growth right
00:15:32
there and so I'm very optimistic that
00:15:35
there's going to be a pull uh and that
00:15:37
we should as the United States given the
00:15:39
technology stack we have uh in Europe in
00:15:43
Asia in you know in South America in
00:15:45
Africa and everywhere
00:15:47
get it to be broadly deployed.
00:15:50
>> You one of the questions I get asked a
00:15:51
lot about the AI race is how do you know
00:15:55
if you're winning or how do you know if
00:15:57
the United States is ahead of its global
00:15:59
competitors? And the answer I give is
00:16:01
market share. You know, if we look
00:16:03
around the world in 5 years and we see
00:16:06
that American companies, American
00:16:09
technology has say 80% market share, it
00:16:11
means we did a good job. If we look
00:16:12
around the world in five years and see
00:16:14
that it's say Chinese chips and Chinese
00:16:16
models that are being used all over the
00:16:18
world, well means we probably lost. So
00:16:20
you know ultimately usage is the proof
00:16:23
of the pudding is in the eating of it. I
00:16:25
mean the in this case the way that you
00:16:27
know that you're succeeding is through
00:16:31
market shares through usage.
00:16:32
>> I and I I would agree with that. But
00:16:35
David, since you even worked at
00:16:36
Microsoft for a few years, um you you
00:16:39
know, one of the things that I'm very
00:16:41
grounded on is always uh that Bill Gates
00:16:43
line of a platform, right? So, one of
00:16:45
the things that I always think about is
00:16:47
it's market share, but it's also
00:16:49
ecosystem effects, right? See, what the
00:16:51
United States always has done is not
00:16:54
just about our market share or even um
00:16:57
the revenues to US companies. In fact,
00:16:59
one of the things I learned at Microsoft
00:17:01
is whenever I did a country visit, the
00:17:03
data I would first study is in let's say
00:17:06
in the UK or in Switzerland or what have
00:17:09
you is what is the total employment
00:17:11
created in Switzerland uh in our channel
00:17:15
that used to be like the number one
00:17:16
thing uh in our country reports right
00:17:19
and the total number
00:17:20
>> that be like the number of IT workers
00:17:22
the number
00:17:24
>> office workers channel so channel
00:17:26
partners we so number of ISVs
00:17:29
uh who were there. So we used to have a
00:17:31
complete marker of how did the ecosystem
00:17:34
around the platform get built one
00:17:36
country at a time and that is what the
00:17:39
United States has always done. In fact
00:17:41
the US tech stack
00:17:42
>> including in China got built because
00:17:46
others built around our tech stack the
00:17:48
same thing is going to happen. So that's
00:17:50
why I think the work you're doing around
00:17:51
diffusion,
00:17:52
>> right,
00:17:52
>> is about really increasing the size of
00:17:55
the pie, the trust in the platform so
00:17:59
that there is true economic opportunity
00:18:01
quite frankly.
00:18:02
>> Well, you're right and and I remember
00:18:03
actually you you brought back some
00:18:05
memories from this is about a decade ago
00:18:07
when my company Yammer was acquired by
00:18:09
Microsoft. we were part of uh the
00:18:11
SharePoint group and I remember that the
00:18:14
um the product managers there were very
00:18:15
proud of the fact that the revenue from
00:18:19
the SharePoint ecosystem meaning
00:18:21
non-Microsoft
00:18:22
the um the consulting community the
00:18:25
implementers who would go into companies
00:18:27
implement SharePoint I think their
00:18:29
revenue is something like seven times
00:18:30
greater than Microsoft's own software
00:18:33
revenue and I think
00:18:34
>> in aggregate
00:18:35
>> in aggregate and I think and I think
00:18:37
Bill had a line about you're not an
00:18:39
ecosystem or platform until the revenue
00:18:41
on top of your platform is some, you
00:18:44
know, factor of your own revenue. And
00:18:46
and I and I think that I think what's
00:18:48
really important about this is when we
00:18:50
talk about diffusion and obviously want
00:18:52
the United States to have this leading
00:18:55
position, it doesn't mean it's bad for
00:18:57
the rest of the world because they're
00:18:58
able to build on top of those platforms
00:19:00
and create even more value.
00:19:02
>> 100%. In fact, that's sort of the most
00:19:04
important point, right? So this is not
00:19:06
that this is not about uh American tech
00:19:09
um and America revenues to the United
00:19:11
States. It's actually creating
00:19:13
opportunity using a new platform
00:19:15
everywhere. And in fact you know the you
00:19:17
know like I remember I worked on our
00:19:19
database products uh in the '90s u you
00:19:22
know with SAP in fact the combination of
00:19:25
uh SQL Server and R3 were successful on
00:19:29
both sides. There's a lot talked about
00:19:30
Intel and Microsoft, but one of the
00:19:32
other things that I grew up in which has
00:19:34
sort of been foundational in how I look
00:19:36
at the world is what we did with a
00:19:38
European software company that is still
00:19:40
uh you know a giant. And so that you
00:19:42
know who knows what the next big AI app
00:19:45
will be and where and what will happen.
00:19:47
But uh I I sort of go in with the
00:19:49
attitude that there will be tech
00:19:51
companies uh maybe even top five tech
00:19:54
companies that could emerge everywhere
00:19:56
with even the American tech stack. you
00:19:59
have um done some amazing acquisitions
00:20:02
and you're quite a dealmaker on top of
00:20:03
being a technologist. It's probably the
00:20:05
least reported aspect of your
00:20:08
spectacular tenure and the massive
00:20:10
growth you've had.
00:20:12
But you did a deal with OpenAI and
00:20:15
probably one of the most savvy
00:20:18
slashcontroversial dealmakers of all
00:20:20
time, Sam Alman. That deal was looked at
00:20:23
as
00:20:25
you you you're you're set up to get a
00:20:27
windfall in cash which you don't need as
00:20:29
Microsoft. always nice I'm guessing if
00:20:31
they IPO but did you create potentially
00:20:34
and this was the criticism of it
00:20:37
an ultimate competitor to Microsoft and
00:20:40
how do you think about that and how can
00:20:42
Microsoft which missed Steve Bombber's
00:20:44
biggest regret missing the mobile
00:20:46
revolution how can you not have a Gemini
00:20:50
an XAI a claude that is your own or in
00:20:53
your mind do you have that because you
00:20:56
have the source code of open AI
00:20:57
>> yeah I think that that's right so when
00:20:58
when people say uh where is your
00:21:00
foundation model? I mean at the end of
00:21:02
the day we do have the IP but that said
00:21:04
I think you bring up a couple different
00:21:06
things right one is to us the most
00:21:08
important thing when I look at what is
00:21:10
Microsoft's uh strategy today one is we
00:21:14
want to build token factories right so
00:21:16
our biggest business today is Azure
00:21:18
business and the Azure business the TAM
00:21:20
given what's going to happen is is so
00:21:22
huge that we now need to be fantastic uh
00:21:25
at building these token factories and um
00:21:27
that's means a heterogeneous fleet of
00:21:30
infrastructure
00:21:31
and that every hyperscaler has always
00:21:33
done which is use software to make
00:21:35
maximum use of it and for TCO and
00:21:37
utilization. So that's one side of it.
00:21:40
Then there's the app server business
00:21:41
right which is everybody we you talked
00:21:43
about like if everyone's going to be
00:21:45
building agents have infinite minds have
00:21:47
these RL gyms have eval what have you
00:21:50
there's an entire just like every
00:21:51
platform has had an app server this one
00:21:54
has an app server that's what we're
00:21:55
doing with foundry and what have you
00:21:57
right so there's an app server business
00:21:59
in that app server one of the things
00:22:01
that structurally now is pretty clear is
00:22:04
anyone building any application or any
00:22:06
company is going to use not one model
00:22:08
but all the models Right? Why would I
00:22:10
not? Right? Which is in fact I will
00:22:12
orchestrate for any given task even
00:22:15
multiple models. Right? There's this one
00:22:17
nice thing that we came out in our
00:22:19
healthcare practice called the decision
00:22:21
orchestrator. What it proves is that by
00:22:24
assigning roles, right? So investigator,
00:22:27
data analyst, domain expert, just giving
00:22:30
even prompted roles to models and then
00:22:33
orchestrating them gets better results
00:22:36
than any one single frontier model. Am I
00:22:38
right to read into that then that you're
00:22:41
bullish on the open- source models and
00:22:43
think large language models will largely
00:22:45
be commoditized and that's not where the
00:22:47
value will occur. In fact, the way I
00:22:48
think about it is that just like what
00:22:50
happened
00:22:51
>> and Apple thinks that too by the way.
00:22:52
>> By the way, what the way you think about
00:22:54
what happened in the database market,
00:22:56
right? You know, I used to be like
00:22:57
everything is just a SQL database until
00:22:59
it was not, right? There was I mean,
00:23:01
think about it. There dock databases,
00:23:03
there is no SQL databases. The
00:23:05
proliferation of databases, right? Who
00:23:08
would have thought that the database
00:23:09
market would have such a richness to it
00:23:12
>> or that it could ever be open source?
00:23:13
That was
00:23:14
>> that's true. I mean talk about Postgress
00:23:16
or what has happened even with
00:23:17
which is open but there are even
00:23:18
companies that have backed it and so so
00:23:20
to me that's what's going to happen and
00:23:22
to me a model is like the database
00:23:24
market you know it's it's got it's going
00:23:26
to differences but I sort of somehow
00:23:28
think that uh it's not there are
00:23:30
definitely going to be frontier models
00:23:31
that are closed source you know there
00:23:33
going to be open source models that are
00:23:35
going to be uh uh frontier class in fact
00:23:38
if anything I think in this next year
00:23:40
what'll be probably a big part of the
00:23:43
discussion is what's the future of a
00:23:46
firm? A firm should be able to take the
00:23:50
tacet knowledge it has and embed it
00:23:54
inside a weights in a model that they
00:23:57
control. Right? So when somebody asks me
00:23:59
how many models should be there, I'll
00:24:01
say as many models as firms in the
00:24:03
world. Right? That's sort of the an
00:24:04
extreme way. Uh because because to me
00:24:07
that's how I think this you know this
00:24:10
knowledge economy becomes an AI economy.
00:24:12
>> Are you
00:24:13
secretly and you can say it here since
00:24:15
we're on allin working on an LLM to
00:24:18
exist on the Windows desktop because
00:24:21
that you are you have it like today
00:24:23
there's a five silica model which is
00:24:25
completely resident using NPUs and of
00:24:28
course using uh GPUs in fact the largest
00:24:31
installation
00:24:33
um of high power in fact it's one of the
00:24:34
fascinating the workstation is back I'm
00:24:36
one of the most if you went to see
00:24:38
>> which is great for Microsoft because you
00:24:41
you have a nice desktop business.
00:24:43
>> Absolutely. And so we and in fact we
00:24:45
think that that form factor especially I
00:24:47
mean I I always say this which is u you
00:24:50
know I started my career on a command
00:24:52
line. Who knows I may just end it in a
00:24:54
command line.
00:24:55
>> Well you started at Sun which was the
00:24:56
original 5 $10,000 workstation. Do you
00:24:59
see a time where you'll be meeting with
00:25:01
your customers here and advocating a 10
00:25:03
$20,000 desktop machine that has an LLM
00:25:07
and the hardware? You can you can put a
00:25:09
DGX card and you can have like just a
00:25:12
fantastic machine and the models I and
00:25:14
by the way you know we are one
00:25:16
architecture tweak away from even having
00:25:18
some kind of a distributed model
00:25:20
architecture right even ane architecture
00:25:22
that shows knows how to really
00:25:25
distribute itself right that's the type
00:25:27
of breakthrough that can completely
00:25:29
change uh what hybrid AI may look like
00:25:32
but we're absolutely committed and
00:25:34
focused on making the PC a great place
00:25:37
for local models uh and local models
00:25:40
that then do even a lot of the prompt
00:25:42
processing and call into the cloud,
00:25:44
right? So there's a whole lot of work
00:25:45
that can happen and that's sort of
00:25:46
definitely something that's underway.
00:25:48
>> Yeah, I think that the cloud co-work has
00:25:50
kind of shown the power of tapping into
00:25:52
the local file drive and be able to use
00:25:54
that. That that brings up another point.
00:25:56
you you got me thinking about Yammer and
00:25:57
for people who don't know um you know
00:25:59
Yammer's claim to fame this is about 15
00:26:01
years ago was that it pioneered a lot of
00:26:04
um well it used a lot of consumer growth
00:26:05
tactics to attack enterprise software
00:26:08
I'm wondering as you think about
00:26:09
enterprise adoption of AI how do you
00:26:13
think it's going to spread over the next
00:26:15
year it feels like we're at sort of a a
00:26:17
a critical point do you think it's going
00:26:19
to be top down is it going to come from
00:26:21
the CEO directing a team giving them a
00:26:24
strategic transformation
00:26:26
project and they're going to do an RFP
00:26:27
or do you think it's going to spread
00:26:29
bottom up in the enterprise through AI
00:26:32
native employees who are adaptable who
00:26:35
are using the tools in their own lives
00:26:37
and they start to bring these things to
00:26:39
work and start accomplishing amazing
00:26:41
things.
00:26:41
>> Yeah. No, I think you know like all
00:26:43
things David I think it's both the top
00:26:45
down bottom up right. uh the that the
00:26:47
reason I say that top down is if I look
00:26:49
at the ROI uh of uh applying AI in
00:26:53
customer service uh or in supply chain
00:26:56
or in HR self-service those are the easy
00:26:59
projects where uh IT and CXOs can make
00:27:04
calls and that's where you'll see the
00:27:05
first drop of uh real AI adoption but
00:27:09
the bottom up is what ultimately will
00:27:12
happen right I mean with even with the
00:27:13
PCs in fact if you think back at the
00:27:16
lawyers brought word in and then finance
00:27:18
bought Excel in and then email came and
00:27:21
then it became standard issue. That's
00:27:23
what's happening right now. So for
00:27:24
example, these agents when I sort of
00:27:27
talk about everybody's building agents,
00:27:29
they are figuring out a way to go create
00:27:33
these things that are changing workflow
00:27:35
and removing drudgery in their work.
00:27:37
Right? That's sort of the beginning of
00:27:40
what is a bottomup transformation. Um I
00:27:43
you I was in fact the thing that I'm
00:27:46
most excited about is this bottomup
00:27:47
change even at Microsoft for example we
00:27:49
manage something like 500 odd fiber
00:27:52
operators around the world in in Azure
00:27:55
today and by the way I not myself
00:27:57
realized it a lot of it you know it's
00:27:58
called DevOps but it's a it's a physical
00:28:00
asset things get cut and when you sort
00:28:03
of say DevOps that means you literally
00:28:05
are emailing people and saying hey what
00:28:07
happened to that fiber cut how do we
00:28:08
repair it so there's a lot of back and
00:28:10
forth so this network the the person who
00:28:13
runs our global network basically has
00:28:15
built to your point about these person
00:28:17
they're just digital employees
00:28:19
essentially that are doing all of that
00:28:22
devops uh and so that's and there's a
00:28:25
completely bottoms up uh where you see
00:28:27
the tools it's kind of like hey I have
00:28:29
the new way to build agents it's there
00:28:32
I'm going to use it to create levels of
00:28:34
automation uh that remove drudgery
00:28:36
improve efficiency improve quality and
00:28:39
that ultimately is a skilling thing
00:28:41
which is sort of the big issue which is
00:28:44
um and skilling is not mystical it's
00:28:46
just by doing right so it's not like I
00:28:48
go to a class per se it's like the
00:28:50
diffusion of the tools uh and using the
00:28:52
tools and that I think is what's really
00:28:55
going to be happening
00:28:56
>> and and we're in a very interesting
00:28:57
moment empowering an existing employee
00:29:00
with these tools is so much easier than
00:29:04
hiring and mentoring and bringing up the
00:29:06
next generation so it feels like we're
00:29:08
in a little bit of an indigestion moment
00:29:10
at Microsoft, do you think who's going
00:29:13
to have my job
00:29:14
>> in 30 or 40 years, if the company stays
00:29:17
the same size? Because given your
00:29:21
technology first approach, there's
00:29:23
really no reason to ever add another
00:29:24
Microsoft employee at the pace this is
00:29:26
going and you haven't for four years.
00:29:29
So, how you may have swapped some in and
00:29:32
out and changed the texture of it. So,
00:29:34
how do you think about maybe this next
00:29:37
generation? What advice would you have
00:29:38
for these college graduates who maybe
00:29:40
don't have an offer for Microsoft right
00:29:42
now? And you used to spend a lot of time
00:29:43
on that building that group, but maybe
00:29:46
you don't have that luxury now. Do you
00:29:48
think about it ever?
00:29:49
>> No, I I mean it's a great question. I
00:29:51
you know there's a little bit of a
00:29:52
debate what happens to early in career
00:29:54
and how is college recruiting. I still
00:29:56
am a big believer in uh college
00:29:58
recruiting because at the end of the day
00:30:00
um this is going to change the curve by
00:30:05
which anyone can pick up proficiency in
00:30:08
a codebase. Let's just it takes sort of
00:30:10
just regular CS hiring. uh what has
00:30:13
changed is perhaps for someone who comes
00:30:16
in new into a team and to be able to
00:30:19
ramp up thanks to all of uh the
00:30:23
markdowns, the skills, uh the fact that
00:30:26
I can go ask the agent. I mean, think
00:30:28
about it, right? It's like having an
00:30:30
unbelievable mentor who is getting you
00:30:33
onboarded onto a codebase faster. So in
00:30:36
some sense the productivity curve uh of
00:30:39
a college hire is going to be much
00:30:41
steeper than it ever before. So I think
00:30:44
there might be a difference. In fact,
00:30:45
one of the things we're experimenting
00:30:46
with is a different type of
00:30:48
apprenticeship, right? Which is you take
00:30:50
somebody who's an IC senior dev have
00:30:53
like a cohort of college uh hires
00:30:56
working with them because it's a new way
00:30:58
of working. It's like I remember like
00:31:00
all you know everybody who joined
00:31:01
Microsoft would say go how how did you
00:31:03
know whatever um Cutler implement Malik
00:31:06
or what have you right he would go try
00:31:08
to read uh his code to understand uh
00:31:11
what great craftsmanship looks like
00:31:14
nowadays I think that great
00:31:15
craftsmanship uh comes by looking at
00:31:18
even how the 10x 100x engineers use AI
00:31:22
to build great quality products uh and
00:31:25
that is what these new college grads
00:31:28
will learn and learn faster and so
00:31:30
that's a beneficial thing for a company
00:31:31
like us because at the end of the day
00:31:33
you know until we saw longevity or
00:31:35
something we need people to come into
00:31:38
the workforce be successful at Microsoft
00:31:40
so we are very committed but we are also
00:31:42
making sure that the scopes of the jobs
00:31:45
make sense for what the aspirations of
00:31:47
people are going to be both who are
00:31:49
currently in the workforce and people
00:31:50
who are entering the workforce.
00:31:52
Okay, on that note, Sache Nadella,
00:31:55
>> thank you so much.