YT YouTube2Text

YouTube2Text → Transcripts → Satya Nadella on the AI Doomer Slowdown, Microsoft’s Master Plan & Who Wins AI

Satya Nadella on the AI Doomer Slowdown, Microsoft’s Master Plan & Who Wins AI — Transcript

by All-In Podcast · 6,783 words · 954 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:01has generated $250 billion with a B in
0:04market value for Microsoft. Scott
0:05Nadella, chairman and CEO of Microsoft.
0:07>> Since you've been the CEO, three and a
0:09half years, the stock is up about uh I
0:12guess it's about 120%. I'm good for my
0:1580 billion. I am going to spend $80
0:17billion building out Azure. Maybe after
0:19the industrial revolution, this is the
0:21biggest thing. That's our goal with our
0:23frontier model. Our model should be the
0:26best model that they can use as a base.
0:28We create technology so that others can
0:31create more technology. That's who we
0:33are. We're tool maker.
0:35>> Please welcome Satia Nadella.
0:39>> All right.
0:43>> Hi guy. Good to see you coming out.
0:48>> Good to see you.
0:51>> Good morning guys.
0:52>> How are you?
0:53>> Good.
0:54>> Thanks for joining us.
0:55>> Crazy weekend, but here we are. Do we
0:57need to paste the frontier?
1:03>> So, let's start with the common sense
1:06part first, which is we should do what
1:10it takes to build stuff that serves
1:13humanity first and is in human control.
1:17You know, it's kind of crazy that we
1:18have to start with that level of common
1:20sense, but I think it's a good place.
1:23Then when I think about pacing whatever
1:26the first thing that at least I believe
1:29is the broad diffusion of this
1:32technology is the most critical thing
1:34because the benefits of this tech
1:37showing up everywhere is really what's
1:41all about right so at the end of the day
1:42if you sort of say serving humanity let
1:45it actually reach humanity in ways that
1:47it serves humanity and that means you
1:50got to have choice you have to have
1:52compet competition. You have to have all
1:54kinds of business models whether they're
1:55open weights, close weights, what have
1:57you. Then the other aspect I think that
2:01is not talked about when we talk about
2:03control is actually the control that for
2:06example customers have, enterprises or
2:10businesses have around this technology
2:12because sometimes this is so opaque,
2:14right? I want my privacy. I want to be
2:17able to embed my knowledge in a set of
2:19weights I control. I want to see all of
2:22the coot uh that's being generated. I
2:25want to use it to do fine-tuning of my
2:28own models. My IP shouldn't leak. So,
2:30there's an entire body of things that
2:32nobody's talking about as much, which is
2:35my I really want to make sure that this
2:37tech is in my control. Then we get to uh
2:41what is I think a real issue of safety
2:45and we should take it seriously which is
2:49we should take all the time we want uh
2:51to test things. In fact I love this idea
2:54of having third party testers. Oh wow.
2:57You know I you know I grew up in a
2:58company that's always done testing. Uh
3:01so it's novel that we should say wow
3:03they're having embedded third party
3:05testers. Why not? It's a great idea. In
3:08fact, the only thing I would say is we
3:09should avoid like these, you know, cozy
3:12arrangements of who's testing what, who
3:15has access to what, and it should be
3:16broad.
3:17>> Were you were you surprised though when
3:20both the essay landed and then it seemed
3:23like there was a circling of the wagons
3:25amongst the frontier companies? I
3:27>> I I think that it comes my my suspicion
3:30is it comes genuinely from this place
3:33where when you start seeing in fact it's
3:35fascinating, right? We are when you
3:37start seeing reward hacking um and
3:40what's happening in these environments
3:42right with these agent swarms there is
3:45the mundane there is some DevOps error
3:48where somebody misconfigured
3:50a container
3:51>> right right or these API keys
3:53>> or an API keys or yeah exactly there's
3:55no monitoring uh there's internet access
3:57there's sort of classic I would call it
3:59basic devops and then there is real
4:02novel new stuff right which is what is
4:04this uh reward board hacking uh that you
4:07know with these persistent agents and so
4:09on and that's a place where I'll admit
4:11that the science is not there it's I
4:13thought Yakob's post which is a good one
4:15which he said he called it we're growing
4:17intelligence not building intelligence
4:19so it's an experimental science and so
4:22the more experimental sciences uh then
4:26you really need to make sure you're
4:27doing those experiments in controlled
4:29environments if anything the place where
4:32I would love is taking even the hugging
4:34face incident in other places more
4:36transparency on what would it take in
4:39fact one of the fascinating things right
4:41now is the insider risk I mean think
4:43about it right if you're sitting in an
4:44enterprise this is all test time compute
4:46by the way right so it's not like oh
4:47it's going to only happen when in some
4:49training run it can happen for a very
4:52mundane task uh that I give one of these
4:55frontier models inside an enterprise uh
4:58where I say you know I don't I was you
5:00know telling David this suppose I say
5:01hey go optimize my working capital it
5:04may fake my books uh right because this
5:07is like a new type of insider risk
5:10>> and so what is the way to do that I
5:12would say oh go build a maybe a causal
5:15model like a semantic model that
5:17actually checks and verifies so I think
5:18there's a lot of product building um I I
5:22would say making things more robust
5:24which is classic engineering that we
5:26should be talking a lot more about
5:28transparently versus saying hey this is
5:30so mystical that you know we can't
5:32figure this out. Do you do you buy this
5:33argument that it's mystical?
5:35>> I I mean I I buy the argument that we do
5:38not understand the latent space. Uh
5:42right other than I thought you know as
5:43you said like do we understand the
5:45brain? We don't. We do functional MRIs
5:47and do neuroscience and we're trying to
5:49figure this out continuously getting a
5:52little better understanding. So I do
5:54think that in that sense we don't
5:56exactly uh have a complete un that's why
5:59by the way I also I don't believe in new
6:02release right so that's why I think
6:03making sure that the coots are in
6:06language that we can all understand in
6:08fact they're transparent so that when
6:11when I go back to an enterprise that's
6:13using all these models and if you have
6:15the full coot uh then you can
6:18>> chain of thought
6:18>> chain of thought and so then you can
6:20really go look at it deeply in fact you
6:22can have multiple models uh and you can
6:24look at the coot across those I think
6:26these are all things that I think will
6:28become very important
6:28>> satia you've worked with you've worked
6:30with technologists for decades
6:33uh and when you see as a leader of one
6:36company Microsoft which has very crisp
6:39communications with the public uh and
6:42you see what's happening with Daario and
6:44his team people coming out saying 10%
6:47chance we all die uh what do you think
6:50is going through those technologists
6:52minds. Do you believe they actually
6:55believe that this is going to kill
6:57humanity or are they going through some
6:58psychosis or are they seeing something
7:01working on those frontier models that is
7:03terrorizing them? You're not a
7:04psychologist, but you have worked with
7:06technologists for a long time. Handicap
7:08what's going on in these organizations
7:10that's all the making people feel the
7:12need to resign and say we're all going
7:15to die.
7:17>> Yeah. you know, it's it's hard for me to
7:20speak to what's happening in any of
7:22these places, but let let's just say uh
7:25how we I grew up even inside of
7:27Microsoft, you know, for example, you
7:29know, one of the biggest things you
7:30learn as an early sort of engineering
7:33lead is how to deal with a showstopper
7:35bug.
7:36>> Yeah.
7:36>> Right. I mean, that's kind of like 101,
7:38right? Which is why you're faced, you're
7:40like, you know, you have a bug. Um what
7:43do you do? do you stop uh and fix or you
7:47defer or you go in and say hey this is
7:50such an edge case that's kind of the
7:52judgment so I do think and as the stakes
7:56go up you want to like transaction
7:58processing I remember working on
7:59databases right you know wow like you
8:01know you got to take very seriously any
8:04bug uh where if the transaction is going
8:07to get lost right data loss is a thing
8:09that you stop the thing for so I feel a
8:13little bit culture culturally in the AI
8:15industry rediscovering maybe because
8:17when you see and it's possible that they
8:20see stuff which are showstoppers before
8:23the rest and if you see a showstopper
8:25stop the show um right to fix the bugs
8:29yeah when you saw the the hugging face
8:33run and it was super performative
8:36Dwaresh did his whole post civilizations
8:39what do you think what what's your take
8:42on that testing they ran because they
8:44could have run a test where they had
8:463,000 agents defend a bunch of websites.
8:48Instead, they instructed them to hack
8:50websites and you know the hiding of
8:53information all this
8:54anthropomorphicizing
8:56whatever of the agents. I mean the way
8:58at least I understand it was it was
9:00actually you know basically trying to uh
9:03do an eval uh for cyber gym and um as I
9:08understand it given that eval it sort of
9:12figured out a way to say let's just say
9:14reward hack uh and that's what led it to
9:17hugging phase in fact it speaks to I
9:20think what's the pre you know clear
9:21issue right now which is you can have
9:24these things if they're are longunning
9:26persistent agents
9:28become essentially like new insider
9:30risks. Uh and so that I would start from
9:33the very basics of saying okay what is
9:35containment look like. So for example
9:37like one of the things that I think is
9:38going to be really an issue and a thing
9:41that needs great solutions is true
9:44aggressive monitoring of agent activity.
9:48Uh that's behavioral
9:49>> evidence
9:50>> evidence and so everything has got to be
9:52auditable. uh and then every object it
9:54access, right? If it goes and gets a
9:56secret, oh, it's going to go chain a
9:58couple of things, you should be able to
9:59see it when it's starting to chain a
10:01couple of uh vulnerabilities uh to go
10:04hack. And so I think that these are the
10:06ways um that you really have to sort of
10:09deal with these situations versus saying
10:12um in in fact I think the core of my
10:15take is we will have to get the
10:19engineering process around building out
10:23this experimental science to be more
10:26robust.
10:27>> Yeah.
10:28>> Thanks.
10:29>> So I think I think that's a great point.
10:31I love how you uh differentiated in the
10:34HuggingFace uh episode between the
10:36mundane things they got wrong like the
10:38misconfigured sandbox and HuggingFace
10:40had credentials just sitting in a public
10:41repository and there was no monitoring
10:43and then you have the genuinely novel
10:45behavior, the swarms of agents, the
10:47reward hacking. That's the stuff that
10:49has everyone freaked out. I agree that,
10:51you know, we have to now figure out how
10:53to fix the bugs or, you know, fix the
10:55deeper problem that's coming from that
10:57reward hacking. What what do you think
10:59that means for and and and and I think
11:01to their credit I think what the
11:03Frontier Labs are saying is we are now
11:05going to slow down the pace of let's say
11:08raw power and shift towards reliability
11:11and predictability and you know what
11:13they call alignment which I think is
11:14good business practice I guess what do
11:17you think that means for what we see in
11:19terms of new products for the next year
11:21or two does it mean we just kind of
11:23improve what we already have or do we
11:25see new capabilities what do you think
11:27this going to mean A great question,
11:28David. I I do think there's already a
11:30massive model overhang, right? I mean,
11:33um capability overhang in the sense of
11:35the models are very good except the
11:39broad diffusion uh requires a lot of
11:42things, right? even requires uh
11:44essentially if you're compressing
11:46workflows and changing workflows to
11:48happen differently u the amount of
11:51change management that needs to happen
11:53in order to even incorporate these
11:55systems is sort of what's taking time so
11:58to some degree I would say the and also
12:01uh the the ability to create these new
12:03form factors right I mean if you think
12:05about coding agents and coding agents
12:06became really usable when you discovered
12:09that you could have an agent loop with a
12:11file system uh and that was the
12:14breakthrough that just made coding
12:15agents work. Um and I think now maybe
12:18with KUA right so which is with Astra
12:20with KUA uh could be a way for us to
12:23even do computer use or we just use long
12:26trajectory tasks that can get completely
12:29automated. So I think these type of
12:31product innovations where the model plus
12:34the harness allow us to do things that
12:38then lead to broad adoption. Right? I
12:40even go back to the chat GPT moment for
12:42me, right? Which was it was that RHF at
12:45the very end that made a chat
12:48conversation possible. Mhm.
12:50>> Uh and so I think that yes, so there's
12:52some science, there is some form factor
12:55that then leads to broad diffusion and
12:58we now need to find the next level of
13:00these things that are doing real work in
13:02the real enterprise. Um and in that
13:05context by the way the other thing is
13:07it's going to be a multimodel world
13:09right so at this point just out of
13:10resilience right I mean think about
13:12right every enterprise now comes to me
13:13and says hey this model does refusals
13:16here this model I want weights here I
13:18don't and so the people are going to
13:20want multiple models so one of the other
13:23things that we have to get right is some
13:25standards of interop right like even KV
13:28cache like why the heck can't I use
13:30multiple model families and have KV
13:32cache reuse
13:34uh right we've had document standards
13:36you and I lived through it right but
13:37we've sort of you know you kind of have
13:40things that are interoperable in the
13:42real world everywhere else so I think
13:43this industry also has to wake up and
13:45say hey in fact if I were talking about
13:48the most important pressing things is
13:50how do I have more standards on uh
13:53interoperability how do I have a harness
13:55that is external to a model so that my
13:57memory is not tied to one model I mean
14:00this is the first time you're going to
14:01have a technology where your use of it
14:04and the exhaust in the data could not be
14:07yours. Uh I mean that you know like it's
14:09like if I g sold you a database and said
14:11hey the data you put into your database
14:13is not yours and it's mine. It goes away
14:15if I took away the license. How would
14:17you feel about it? So therefore I think
14:19we have some serious issues like that to
14:21deal with.
14:21>> I think that's a good segue.
14:22>> Sorry. Let me just ask one question to
14:24connect the um economic incentive
14:27argument on what's going on. The
14:29argument is the Frontier Labs are facing
14:33token compression. 50 bucks for OpenAI's
14:37kind of million token output versus I
14:41think someone estimated Deep Seeks new
14:42is like can go as low as 15 cents for a
14:45million tokens of output. Let's call it
14:4660 cents. 99% cost reduction.
14:50If that is the the big kind of economic
14:53crux of what the frontier labs are
14:55facing, why would most tokens be paying
14:5850 bucks? Most enterprises pay 50 bucks
15:00when they could pay 60 cents for most of
15:02their tasks. Doesn't that also beg the
15:04question, are they in the wrong business
15:06model? And I I asked this for you as the
15:08CEO of Microsoft, what's the right
15:10business model? Do you want to be making
15:11the frontier model? Do you want to be
15:14running the compute and charging for
15:16rent on your compute? Or do you want to
15:18be in the application layer? I know you
15:19talk about this a lot, but I just love
15:20your perspective from where we sit today
15:22and how this all kind of
15:24>> um
15:25>> Yeah, I think the the fundamental thing
15:26that I think we're observing is good
15:28old-fashioned competition, right? I
15:30mean, for me, if I look back at it, we
15:32were we had like some real great closed
15:34source assets, Windows. What was the
15:36check against it? It was of course the
15:38Mac, but Linux
15:41>> uh we had a great closed source product
15:44called SQL Server. What was the check
15:45against it? there was always a
15:47substitute called Postgress or MySQL. So
15:50I think that's what's happening a little
15:51bit of it is there's real competition
15:53between closed source and the open-
15:56source check is real. Um, and that's
15:59good quite frankly uh because without it
16:01I don't think we're going to have a
16:02broad frontier ecosystem or broad
16:04diffusion because otherwise we'll just
16:06we'll be back to some uh you know
16:08mainframe uh locket that's just not uh a
16:11thing to your point about if anything
16:15given that we will now hopefully
16:17continue to have a much richer choice in
16:21every layer. Right. So to me hopefully
16:24we can start building these AI because
16:26today the royalty of an AI product all
16:29going to just the model layer doesn't
16:32make sense if you really want to build a
16:34product company right it just cannot be
16:36in fact if anything like that's the same
16:38thing right which is if you take the
16:39database if there was no open-source
16:41check on closed source uh the prices
16:44wouldn't have been at a place where
16:46people could have built the app tier
16:47successfully and the with a margin and
16:50so I think the apps are going to become
16:52you know much more viable economically
16:54which is great for the ecosystem. uh
16:57there are going to be all these other
16:59layers of middleware call it right which
17:01is hey what's my memory system what's my
17:03harness and orchestration layer so
17:06there's going to be a very rich tools
17:08ecosystem there the model companies will
17:10do fine uh in fact you know the paro
17:12they can manage the token pricing based
17:15on their model family if anything I want
17:17them to work on even the KV you know
17:19these these standards
17:21>> such that we can use multiple model f in
17:24fact it's better for them in fact I
17:25worked on Windows interrupt with Unix
17:28first.
17:29>> In fact, it was counterintuitive, right?
17:31We used to think, oh my god, this
17:32interrupt means we'll be less used
17:35except we were more used.
17:37>> In fact, we became weirdly enough
17:39because there were so many variants of
17:41Unix at that time that Windows interrupt
17:44made Unix better and Windows better. And
17:46in fact, we were able to penetrate the
17:48enterprise primarily because we did that
17:51interrupt work. And so that's at least
17:54how I think about it. Satya one of these
17:56we're in this interesting moment where
17:58on the one hand you have these experts
18:01asking for regulation asking for
18:04oversight governance it typically always
18:07leads to some restriction of freedom
18:11and general society
18:14are put in a position where now we have
18:16to opine on whether this is right or
18:18wrong but then on the other side most
18:21people's lived experience
18:24is not this magical productivity boost
18:26of AI. At best, it's integrating our
18:29Apple Eyewatch data to tell us why we're
18:31sleeping less. That's like functionally
18:33the bar for most people. Or why is my
18:36kid an into chat GPT? Uh so can
18:41you just help us bridge this? I mean,
18:43you see so many enterprise applications.
18:45Where's the magic? Like where is the
18:47where are the gains in profits? Where
18:49are the huge upside breakthroughs that
18:52AI is creating that will somehow make
18:55all of this tension understandable for
18:58everybody?
18:58>> Yeah, it's a great it's a great point. I
19:00mean, I think this is the real question
19:03which is how do we truly see this in the
19:06productivity stats? How do we really see
19:08it in the GDP growth? That's broadbased.
19:10It's not just supplier
19:12>> or supply side. Um I mean the the one
19:15example that I I love and I get back to
19:18in fact healthcare is a good one right
19:20if you think about um health care and
19:23even the simple doctor patient
19:26interaction in our case we have this
19:28thing called DAX copilot um that's the
19:31place which is the most tangible example
19:33I can always point to when a doctor can
19:36spend more time with the patient caring
19:38for them versus just the entry into an
19:40EMR system that's a good productivity
19:43gain If it can triage uh the inbox for
19:46the doctor so that they can be more
19:48responsive uh that's helpful for uh uh
19:52for the patient and the care system the
19:55administrator in fact keying like the
19:57insure like because it's the
19:58triangulation of the pay patient and the
20:02health system. Yeah. Uh that's of all in
20:05fact most of healthcare is sort of all
20:07workflow cost. Uh so taming of that
20:10workflow complexity that's a helpful
20:12thing. But do you see that in Microsoft
20:14with the people that you're helping?
20:15>> Yeah, absolutely. We see that and and by
20:17the way even in in simple co-pilot
20:19cases, right, which is if you look at
20:21the amount most people think about jobs
20:24which I think there is going to be
20:25displacement there is but the bottom
20:27line is what are the new jobs that get
20:29created uh is going to be one of the key
20:32aspects of it. But also a lot of
20:35knowledge work unfortunately is drudgery
20:38right who you know I get up in the
20:40morning and I think about like man all I
20:42do is email triage right you know
20:44>> uh what if uh even just these workflows
20:48that are taking away time from things
20:50that you could be spending time on
20:52>> okay well you're bring you're bringing
20:53up this great point if you go all the
20:54way back to like the turn of the century
20:55the industrial revolution when we had a
20:577-day work week you know a lot of people
20:59forget why did we introduce the weekends
21:01it was to sort of manage the tension
21:03between different uh religious groups
21:05that had to work in the same factory.
21:06And then when you look at long run GDP
21:09outside of some exogenous events, it
21:11sort of is, you know, between two and
21:13400 basis points.
21:15>> And so what happens is as productivity
21:17boosts come in,
21:18>> human work steps back and you kind of
21:21accomplish the same amount of work.
21:23>> Do you think that that happens here? Is
21:24that is there a risk that we have a
21:27three-day work week and we're just still
21:29growing at two and a half%. Yeah, that's
21:31a great qu or will we find new things
21:34and this is where the excitement at
21:36least I have for what the real impact of
21:38AI would be is instead of just thinking
21:41about hey it has helped me augment some
21:44workflow or simplify something that's
21:47happening today is it inventing new
21:49things uh is it speeding up drug
21:52discovery um is it taking the u I don't
21:56know let's again go back to my example
21:58of okay the working capital management
22:00of a small business has become so much
22:02more efficient uh that suddenly it's no
22:06longer just oh I have an ERP or a
22:07QuickBooks like thing but I truly am
22:10making decisions based on the ability to
22:12introspect my invoices my emails and
22:15what have you and some somehow optimize
22:17my working capital that's productivity
22:20that didn't exist and so I do hope that
22:24we will start seeing GDP growth which we
22:26did see in the industrial era um during
22:29the first phase of it.
22:30>> Yeah.
22:31>> Right. So, so that I think is what is
22:34needed, right? Which is in order for all
22:35of this to play out quite frankly, we do
22:38need to see at least 7 8% GDP growth
22:42that is real and that's broad-based.
22:45>> What's the what business is Microsoft in
22:48in relation to AI? Obviously, Azure has
22:51been crushing it. you're turning away
22:53customers uh and you're doing $175
22:56billion in capex buildout, but your
22:59capex is far below what Meta is doing,
23:02far below what Google's doing. They're
23:04doing secondary raises and raising debt,
23:06350 billion. The Frontier Labs are
23:08spending 500 billion. You were so early
23:10to the party with the precient open AI
23:13investment, but then co-pilot didn't
23:16exactly land. I don't think it didn't
23:18get great reviews. You don't have a
23:20frontier model. What's the business?
23:21Please come back.
23:23>> No, but what's the business here? What's
23:24the
23:26>> Do you need to have a frontier model?
23:29>> Did Did we tell you there was one
23:30journalist on the panel?
23:31>> No, no, no. It's I mean I mean it
23:33sincerely because I'm just curious.
23:35You're a great strategist. We know that
23:37about you. Microsoft missed the mobile
23:39revolution.
23:40>> Is Microsoft going to miss the AI
23:42revolution? You don't have a frontier
23:44model? Because I always found it
23:45perplexing that you didn't. And what's
23:46the strategy there in all seriousness?
23:48Like do you think open source is going
23:49to win? you should have that play.
23:51>> Yeah. So, let me walk you uh through the
23:53sort of where we are and what we're up
23:55to on each of these. By the way, on the
23:56capex side and the buildout side, we
23:59started early. So we if you sort of
24:01cumulatively look um it's a good I'm not
24:05sort of saying you know right right now
24:07speaking about a lot of capex is not a
24:09feature it's a bug but that said but if
24:11you really go actually add up the math
24:14uh given when we started because we
24:15started multiple years before people
24:17woke up to even actually needing to
24:19build and so that's kind of one aspect
24:20of it. The other aspect of it is we are
24:23calibrating our capex in such a way that
24:24we don't we don't want to build for one
24:27or two customers right so we want to
24:29build for the long tail right because
24:31that's I think most important and that's
24:32I mean that if you're a hyperscaler
24:34you're not a supplier to two model
24:36companies that's not a business uh you
24:38have to sort of basically build a system
24:40that is great for lots of third parties
24:43uh and our own one in that context we're
24:46pretty thrilled with the progress we're
24:48making uh with even copilot if you sort
24:50of look at the subscriber numbers we
24:52gave which is this is goes back in fact
24:54to Chamat's fundamental point which is
24:55these are real enterprises using it for
24:58real workflows u and the fact that we
25:00now have 30 plus million not over forum
25:02remember the total knowledge worker base
25:05right where most people talk about 3
25:07billion people 4 billion people on the
25:09internet the entire office 365 or
25:11Microsoft 365 is the the the sort of the
25:14standard when it comes to knowledge work
25:16there's 450 million that's including all
25:18students in the world
25:19>> oh wow
25:20>> right So when we talk like the market
25:22quote unquote as defined is maybe 300 uh
25:26250 even of real enterprise users and of
25:29that we've got the penetration of close
25:31to 30 million on that and it's growing
25:33and so on. The aspect on the model side
25:37is we're thrilled about obviously our
25:39investment in open AAI the access we
25:41have to their IP which we have for a
25:43long time we're going to use that but we
25:45are well on our way building our MAI
25:46models right if you look at it we have a
25:49flash cyber model that you know with our
25:52harness orchestrating other models
25:54outperforms
25:56um on cyber gym even a mythos uh same
25:59thing we're seeing in coding same thing
26:01we're seeing in uh knowledge work right
26:04So our goal is to basically hill climb
26:06from the bottom by the way uh not
26:08distilling anything. So from the very
26:10bottom using our RLES our data uh and
26:14then also have a differentiated position
26:16with enterprises going back to
26:18addressing some of the things that they
26:19want which is hey can I have the weights
26:21can I have the weights that I can then
26:24add to my knowledge uh these are the
26:27things that we will do with our
26:28foundation. Your best advice I think to
26:30enterprises is AI sovereignty is
26:32important. Putting your data into a
26:35frontier model probably not a good idea
26:37and then you're going to be that harness
26:38for them to to help them. So my
26:40implement my advice is more like use all
26:43but be independent of all. So for
26:46example my asset test is you should
26:49always eval
26:51that matter to you right. So what's the
26:53outcome you want? you should go run that
26:57outcome through all the models. Then
27:00here's the test I would do. I would pull
27:01out a model and see whether I can retain
27:03the eval. If I can't, that means you
27:06really are dependent on something that
27:09may or may not be yours.
27:11>> Right.
27:11>> Right. That's so so my fundamental
27:14enterprise architecture would say you
27:16should have a model system that
27:18fundamentally allows you to be able to
27:21continuously hill climb on your own on
27:23eval
27:26uh while using all models closed open u
27:29if you want you can even fine-tune any
27:32of these models but you can even
27:34substitute models
27:34>> s just to build on Jason's question you
27:36had this um incredible moment I think we
27:39put it here where you said you know
27:41we're good for our 80 billion. But just
27:42to expand the question, um there's
27:45effectively this sort of bank of AI that
27:48has emerged and there's this financing
27:50mechanism that just is so important to
27:53the entire ecosystem and now broadly to
27:54the entire economy. But you've been very
27:57disciplined. You have an enormous
27:58balance sheet. You're also an investment
28:00grade issuer. So you could do what
28:03Jensen did, but you've taken a very
28:05different capital allocation approach,
28:06much larger bets, very concentrated, and
28:08you've kind of stayed into your own
28:10ecosystem. just talk us through your
28:12mindset as a capital allocator at
28:14Microsoft and that balance sheet. Yeah.
28:16So the way I'm sort of looking at our
28:20book of business whether it's the hypers
28:22scale our model or our app tier and the
28:26shape of the demand um and then what's
28:29the way to build out for it. And so if
28:31you think about these assets right there
28:32are two classes of it. There are the
28:34long lead um long duration assets like
28:37the the land power cold shell let's call
28:40it. Then there is the kit. the kit is
28:43the short-term uh asset uh that you can
28:47much more you know uh be demand driven
28:49in other words right I have to forecast
28:51let's say two years three year out
28:53demand and then and then also
28:54>> the kit means the racks the chips
28:56>> the racks the chips and what have you
28:57and that's 60% of the cost or what have
28:59you right so therefore so what we do is
29:01we go build as much um we lease we even
29:05rent now right now we're even renting
29:07quite a bit because we kind of were
29:09short on supply uh But the overall goal
29:13is to build more lease some and then if
29:17really need to surge we will even rent
29:19that's kind of on the on the on the uh
29:22assets and then the chips themselves we
29:26will try to be first of all make sure
29:28that we're matching demand and as I said
29:31my goal is not to have just two
29:33customers three customers uh it's great
29:35to have openi being one of our largest
29:37customers it's great that they're
29:39growing uh but we need more uh is the
29:41kit over earning right now and do do we
29:45need is the is the industry pushing for
29:48diversification more silicon more memory
29:51more vendors
29:52>> yeah what's happening is the workloads
29:56that are now at scale uh they obviously
29:59grew up from what GPUs were there but
30:03now the the shape is so well understood
30:06uh that you're able to optimize for a
30:09very different world right So you can
30:11sort of start building
30:13>> um and saying well you know there are
30:15these multiple phases in um an inference
30:19or a training phase so why not build
30:21silicon that's optimized for these uh
30:23and that's just going to lead to a
30:25systems architecture that I think is
30:27going to by definition have a lot more
30:29uh diversity uh I mean I know you have
30:31Jensen coming he himself if you look at
30:33his own architecture is changing quite
30:36drastically
30:36>> quite drastically
30:37>> um and so I think that there is going to
30:39be a lot more choice even there in that
30:41layer. So ours we have Jensen stuff
30:44which is I think our primary thing. We
30:45have our own uh OpenAI is building their
30:48chip so that's also going to be there.
30:50AMD is in there. So we I I my thing is
30:53to run whether it's the OpenAI models,
30:55the anthropic models or our own models
30:57on a heterogeneous kit.
30:58>> Sax I want to let you get in here before
30:59we run out of time.
31:00>> Yeah. So you know we've heard now from
31:02the the various frontier lab leaders Sam
31:05Dario Elon Demis that we need to
31:08prioritize alignment like we're talking
31:10predictability reliability robustness uh
31:13as opposed to maybe just say raw raw
31:16power. Do you think the Chinese labs
31:18will follow suit?
31:20I think that that's the dialogue um that
31:23is I think should be prioritized right
31:25so because at some level my own premise
31:28would be that
31:31that China should also deeply care uh
31:36about the same safety concerns if the
31:39United States uh cares about them right
31:42why should it be different for them it's
31:44not like they won't have the same
31:45hacking problem
31:47>> uh it's not as if uh they don't want to
31:50make sure that their citizens um are
31:53benefiting from AI just like we will
31:55want our citizens to benefit from AI. So
31:57I think that there's a possibility of
32:00international norms around it. If we
32:03really are concrete about what's the
32:04risk, why is this risk so idiosyncratic
32:07that the only people who are worried
32:09about it is the Americans. Uh it doesn't
32:12make sense, right? It's not like a thing
32:14that is sort of said, "Oh, I'm going to
32:16only show up in the United States. I'm
32:17going to be something. If it is going to
32:20go wrong, it's going to go wrong
32:21everywhere at the same time." So I think
32:22the Chinese should care. I mean they're
32:25they are a superpower.
32:27>> Well that's you use the word
32:28idiosyncratic and I think that is the
32:30right word is I don't think we know yet
32:32is this um you know conversation we're
32:35having in the US over the past week. Is
32:37it idiosyncratic to us because we have
32:39you know the strong I guess you could
32:41say doomer type uh school of thought or
32:45is it something that the rest of the
32:46world will basically feel as well?
32:48>> It's a great question
32:48>> and if they do then presumably they'd
32:50want to act on it as well. Yeah, I I
32:52just feel my my take there is that we
32:54are ahead
32:56>> and we are who we are which is we argue
33:00we sort of we compete uh we are more
33:03transparent which is all by the way
33:05virtues as far as I'm concerned so
33:07therefore the fact that this debate is
33:09happening here the world will be better
33:11off for it right so to some degree us
33:13setting if anything I would love a US
33:16set us to lead in the norms that allow
33:20us to defuse use this technology broadly
33:22and create safety standards uh that work
33:26for the world including China. But
33:27>> what do you think we should be doing
33:29that we're not doing and what are you
33:31doing at Microsoft
33:33to change the narrative the populist
33:35sentiment that we have to shut down
33:37super intelligence stop building data
33:39centers
33:40>> etc. So, so to me I think this is I am
33:44squarely focused on one of the to
33:47answering Chamat's question from earlier
33:50which is whom is it benefiting and give
33:53me concrete stories right uh we talked
33:55about the productivity benefits a bit uh
33:58whether it's in healthcare or in general
34:00knowledge work coding but I'll give you
34:03another example right I was looking at
34:05data centers because after all we didn't
34:07talk much uh today on that but there's a
34:10challenge on how does one earn
34:12permission uh to open a data center in a
34:15region. In fact, we just have some of
34:18the best longitudinal data now for a
34:21data center we built out in Quinsey,
34:23Washington, uh for 20 years, close to,
34:26you know, 2008 is when we started it.
34:28And when I look at that data and what it
34:31has meant for that community, right,
34:33where uh the tax revenues have gone up
34:3512 times, uh the paidin taxes have gone
34:40down by a third. Um the growth is higher
34:44than Seattle in Quinsey. This is a rural
34:47town. Uh they have a new school, a new
34:50hospital, a new town center, a new
34:53aquatic center. Wow.
34:54>> Uh we have two and most people say, "Oh,
34:56there not that many jobs." In fact,
34:58there have been 1,200 construction jobs
35:00in that region all through that 20-year
35:03period, right? Because it's not like you
35:05just build it and leave. You
35:06continuously refurbishing, building,
35:08expanding.
35:09>> And how big, how big is that data
35:10center?
35:10>> Uh I think it's now going to be at least
35:124 or 500 megawatt
35:14>> and it sort of will keep expanding.
35:16>> Um and so so these are uh so that's a
35:21real like that community. So earning it
35:24like just not saying hey these are all
35:25the benefits but seeing it
35:26>> but how do you get people to tell that
35:28story because that's what's missing
35:29today is those stories aren't being
35:31organically told and if a Microsoft
35:33executive gets on stage and says don't
35:34worry it's good for the community.
35:36>> Yeah. No I don't think Yeah. So I think
35:38storytelling is one thing. The other one
35:40is I think we just need more people
35:43outside of the tech industry to say yeah
35:45because if you go to Quinsey Washington
35:47they will tell you thank god for this
35:49data center. It's part of like you know.
35:51So to me that's like when it's tangible
35:55uh like that uh because that's the only
35:57way to earn permission because at some
35:58level the skepticism of any of us in the
36:01tech industry just saying things uh is
36:04so high that I think we have to now do
36:06the hard yards of actually doing things
36:09in the world uh which allow people to
36:12say okay I now believe you.
36:13>> It's a new muscle.
36:14>> It's a new muscle. It's a new muscle.
36:16>> So I think you're a good spokesperson to
36:18flex that muscle. I hope you do it more.
36:20Thank you for being with us.
36:21>> Thank you so much.
36:22>> We appreciate you.
36:29>> Thank you, sir. Appreciate your time.

About this transcript

This page contains the full transcript of Satya Nadella on the AI Doomer Slowdown, Microsoft’s Master Plan & Who Wins AI by All-In Podcast, generated from the public captions YouTube serves with the video. The transcript has 6,783 words across 954 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.