00:00:02
[Music]
00:00:04
We're sitting here at the Google Plex
00:00:06
with the CEO of Alphabet, Sundar. Thanks
00:00:08
for being here. Great to have you here,
00:00:10
David. Look forward to it. Is Google at
00:00:12
risk of being truly disrupted from AI?
00:00:14
Recently, we're testing it in labs. This
00:00:16
whole new dedicated AI experience called
00:00:18
AI mode coming to search. Open AI has
00:00:21
SAM, XAI has Elon, Meta has Zuck,
00:00:25
Microsoft has Satcha. Are you willing to
00:00:27
kind of share your perspectives on those
00:00:29
four competitors? I think maybe only one
00:00:31
of them has invited me to a dance, not
00:00:33
the others. Biggest regret. Look, there
00:00:35
are acquisitions. We debated hard came
00:00:39
close. Just give me one
00:00:41
name or get in trouble. Maybe Netflix.
00:00:43
We just leaned into the user experience
00:00:46
and over time we figured out
00:00:48
monetization to follow. It's like one of
00:00:50
the original principles of Google.
00:00:52
Follow the user. All will follow. There
00:00:55
you go. I'm going all in. All right,
00:00:57
besties. I think that was another epic
00:01:00
discussion. People love the interviews.
00:01:02
I could hear him talk for hours.
00:01:03
Absolutely. Oh, he crushed your
00:01:05
questions in a minute. We are giving
00:01:06
people ground truth data to underwrite
00:01:08
your own opinion. What do you guys
00:01:10
think? That was fun. That was great.
00:01:11
What's going on? I'm really excited for
00:01:14
this conversation. You and I started
00:01:16
working at Google on the same day in
00:01:20
2004. I didn't quite realize that. Same
00:01:23
nougler class. We had the the hats on
00:01:25
that same week on the Friday all hands.
00:01:27
I'm now a podcaster. You've done a
00:01:29
little bit differently. You're more than
00:01:31
you're more than a podcaster, but you're
00:01:34
very good at podcasting. Well, I
00:01:36
appreciate it. I think I respect the
00:01:37
other stuff you've done as well. So, no,
00:01:39
I appreciate it. But in your tenure at
00:01:41
Google, you ran Chrome, Chrome OS Drive,
00:01:45
Google Maps, and it's been 10 years now
00:01:48
since you've been the CEO here at Google
00:01:51
Now, Alphabet. Amazing. And
00:01:53
congratulations. Under your tenure as
00:01:54
CEO, the stock has gone up by 4 and a.5x
00:01:57
to a$2 trillion market cap today. You've
00:02:00
grown revenue from 20 billion a quarter
00:02:01
to nearly hundred billion a quarter.
00:02:03
It's been a really like incredible run
00:02:05
to see someone that kind of started as a
00:02:09
a PM and you know grew your way into
00:02:11
this incredible role. So congrats. How
00:02:13
have you liked the job?
00:02:15
No, look, I mean, uh, I love building
00:02:18
products and in some ways, you know,
00:02:21
Google was really set up, I think the
00:02:23
founders set up this kind of a deep
00:02:26
computer science approach and like you
00:02:29
take that and apply it to build things
00:02:31
which can impact people on a day-to-day
00:02:33
basis. And so, you know, it's that kind
00:02:35
of a product and uh technical culture
00:02:38
which, you know, is is the essence of
00:02:41
the company. So, I love doing that. And
00:02:44
you know there's not a single week which
00:02:46
goes by where I I feel like I don't get
00:02:48
to do that. So those are the parts I
00:02:50
really enjoy. But obviously you know
00:02:52
running a company of this scale where
00:02:54
you impact uh so many people I think
00:02:56
it's a privilege. So enjoyed every part
00:02:58
of it. You're at a pivotal moment in the
00:03:00
company's history today. Have you read
00:03:02
the innovators dilemma? Uh you know my I
00:03:05
I I'm obviously very very familiar with
00:03:07
the concept. I don't think I've read the
00:03:08
book actually. But but you know, it's
00:03:10
one of those things which is so much in
00:03:11
the ether. You think you know it, you
00:03:13
know. I say it in justest because that's
00:03:16
the talk of the town, the talk on Wall
00:03:18
Street, the talk in Silicon Valley. Is
00:03:20
Google getting disrupted in this moment.
00:03:23
AI seems to create a fundamentally
00:03:25
different paradigm for human computer
00:03:26
interaction. Consumers are asking AI
00:03:29
questions through chat interfaces.
00:03:31
They're getting complete answers.
00:03:33
They're engaging with AI systems in a
00:03:35
way that they traditionally didn't do
00:03:37
with the classical search interface. Is
00:03:39
Google at risk of being truly disrupted
00:03:41
from AI. Is the core search business,
00:03:44
which the ad revenue and search is about
00:03:46
a $200 billion run rate out of 360
00:03:48
billion of your total revenue, most of
00:03:51
your profits. And it seems like Google's
00:03:54
in a really challenging quandry where if
00:03:56
you disrupt yourselves too quickly, all
00:03:58
of that revenue can go away. It can be
00:04:00
really impactful. So is Google being
00:04:02
disrupted by AI at this moment or is
00:04:04
Google leading? It's a good framework,
00:04:06
good question to talk about. Uh you know
00:04:08
I've definitely uh you know for almost a
00:04:11
decade uh you know one of the first
00:04:13
things I did was to think of the company
00:04:14
as AI first. It was very clear to us. Uh
00:04:18
we had Google brain underway in 2012. We
00:04:20
acquired deep mind in 2014. 2015 when I
00:04:23
became the CEO I said look the
00:04:25
technology is really evolving. The
00:04:27
reason we were excited to be approach
00:04:29
our work as AI first uh is because we
00:04:32
really felt that AI is what will drive
00:04:34
the biggest progress in
00:04:35
search. And so you know I I think even
00:04:40
the last couple of years I viewed this
00:04:42
as an extraordinary opportunity for
00:04:44
search. I think if you look at how much
00:04:47
information means to people I think
00:04:50
they're going to each person is going to
00:04:52
have access to information in a way
00:04:55
they've never had before. So it feels
00:04:57
very far from a zero sum construct to me
00:05:00
and we are seeing it empirically when
00:05:02
people are using search. Obviously there
00:05:04
are a couple of major things uh we've
00:05:06
done with search. Uh you know
00:05:09
transformers drove some of the biggest
00:05:10
innovations in search with BERT and mom
00:05:13
dramatically improved search quality. Um
00:05:16
we we launched AI overviews about a year
00:05:19
ago. It's now being used by over one and
00:05:22
a half billion users uh in over 150
00:05:25
countries.
00:05:26
It's expanding the types of queries
00:05:29
people can type in and we see it
00:05:31
empirically the nature of queries is
00:05:33
expanded. So there are whole new use
00:05:35
cases coming into search. We find for
00:05:38
queries where we trigger AI overviews.
00:05:40
Uh you know we see query growth and the
00:05:43
growth continues over time. You know
00:05:45
getting the feedback from AI overviews.
00:05:47
We are you know we we've recently we're
00:05:49
testing it in labs. There's a whole new
00:05:51
dedicated AI experience called AI mode
00:05:53
coming to search. We'll speak about it
00:05:56
more at Google IO. And in AI mode, you
00:05:59
can have a full-on AI experience in in
00:06:02
search including follow on
00:06:04
conversational queries and we're
00:06:07
bringing our cutting edge models there.
00:06:10
Uh where the models are actually working
00:06:12
to answer your questions using search as
00:06:14
a real native tool, right? And and there
00:06:18
the queries people are typing in queries
00:06:20
like literally long paragraphs, right?
00:06:22
the average query length is somewhere
00:06:24
two to three times is what we see in uh
00:06:27
search as it existed uh two years ago.
00:06:30
So we are seeing people respond. Uh you
00:06:32
know search is always from the outside
00:06:34
people look at it and say search all
00:06:36
kind of looks easy to do. The craft of
00:06:38
search is very hard. Over two decades I
00:06:40
think we've had a real northstar of
00:06:42
understanding what users want in search.
00:06:45
And you know we you know you've been
00:06:46
here we kind of a very metrics driven
00:06:49
company. We kind of know what works. Uh
00:06:52
users are are our northstar and and
00:06:55
empirically we see that people are
00:06:57
engaging more and using the product
00:06:58
more. Right. So uh so all that uh to
00:07:02
your question about innovators dilemma I
00:07:04
think the dilemma only exists if you
00:07:07
treat it as a dilemma right like you
00:07:09
know so for me all along in technology
00:07:12
you have these massive uh periods of
00:07:14
innovation and you lean into it as hard
00:07:16
as you can it's the only way to do it
00:07:19
you know when mobile came everyone was
00:07:21
like well you know it's like you're not
00:07:22
going to have the real estate like how
00:07:24
will ads work all that stuff uh you know
00:07:27
mobile was a transition which ended
00:07:29
ended up working great. I can give great
00:07:31
examples, right? Like Tik Tok has come
00:07:33
in. YouTube has thrived since the moment
00:07:37
Tik Tok has come in, right? And uh it
00:07:39
was a whole new format. We we uh did
00:07:43
shorts when we launched shorts. Shorts
00:07:45
absolutely didn't monetize anywhere near
00:07:47
long form, but we just leaned into the
00:07:50
user experience and over time and we
00:07:53
figured out monetization to follow. So
00:07:56
you know to me you know you don't think
00:07:58
about it as a dilemma like you know
00:08:00
because users you have to innovate to
00:08:02
stay ahead and you kind of lean in that
00:08:04
direction. It's like one of the original
00:08:06
principles of Google follow the user all
00:08:09
else will follow. There you go. And I
00:08:11
think the the Google is dead disruptor
00:08:15
narrative has as you point out been kind
00:08:17
of repeated a number of times. Today
00:08:19
people are pointing specifically and I
00:08:21
appreciate your points about there's new
00:08:22
search experiences coming. The search
00:08:24
experience, it sounds like, is going to
00:08:26
evolve. As people look at standalone
00:08:29
apps, they compare Gemini as a
00:08:31
standalone app to chat GPT to the meta
00:08:33
experience. The stats that came out in
00:08:35
the recent court testimony that had some
00:08:37
data revealed from March. I don't know
00:08:39
where the data came from, but it said
00:08:41
the uh Gemini AI app had 350 million
00:08:44
monthly users compared to Chat GPT at
00:08:46
600 and Meta AI at 500. Is that the
00:08:50
wrong way to think about it? that the
00:08:51
Gemini standalone app isn't the future
00:08:53
or the AI bet that Google's making, but
00:08:55
it sounds like there's going to be much
00:08:57
more of a kind of timed out integration
00:08:59
into how the search experience evolves
00:09:01
and what happens to Gemini, you know, in
00:09:03
search, you know, maybe the most widely
00:09:06
used GI product today might be search
00:09:08
with AI overviews, right? You know, uh
00:09:10
people people are using it intensely.
00:09:12
Obviously, uh we have a standalone
00:09:15
Gemini app. Um I think I think uh we are
00:09:19
making progress there. Particularly with
00:09:20
the introduction of Gemini 2.5 Pro,
00:09:23
we've seen a real uh uptick in
00:09:26
engagement uh and usage growth uh in the
00:09:30
product. We have a lot more to come.
00:09:32
Just in the last few weeks, we've
00:09:34
shipped deep research and updated canvas
00:09:37
audio overviews. You can now go and
00:09:40
generate v do video generation with V2
00:09:43
straight in the Gemini app uh on Android
00:09:46
phones with Gemini live you know you can
00:09:48
screen share it can talk to what's on
00:09:50
your screen so the the you know there's
00:09:54
a lot coming that way and you know users
00:09:56
are responding
00:09:58
um look chat GPD is obviously had
00:10:00
phenomenal success um you know but but I
00:10:04
think it's still early days and you know
00:10:05
we are definitely seeing traction seeing
00:10:09
growth to me what matters is if you
00:10:11
innovate are users responding and using
00:10:13
it more and that seems to be the case so
00:10:16
it's in our hands to continue innovating
00:10:18
right I think it's a fiercely
00:10:19
competitive moment but I would say
00:10:22
across our products people are coming
00:10:24
and using and consuming information
00:10:27
across search using the Gemini model
00:10:30
increasingly in YouTube in the Gemini
00:10:33
app and so on so I think I think it's a
00:10:35
much broader view we have if I were to
00:10:37
think about the unit economics of
00:10:39
Google's business. There's a cost to
00:10:42
serve a search query and there's revenue
00:10:45
per search query, ad revenue per search
00:10:47
query. How is that number changing or
00:10:50
how will that change in this kind of
00:10:53
evolution in search towards more of an
00:10:55
AI interface? Because I've got to assume
00:10:58
that to serve an AI driven query is much
00:11:00
more expensive than to serve a search
00:11:03
query. Look, this is something I think
00:11:05
you know people are really worried about
00:11:06
uh two years ago, but you I've always
00:11:09
felt to the extent that something is
00:11:12
about the cost of serving it, Google
00:11:15
with its infrastructure, I' I'd wager on
00:11:17
that, right? And you know, on our
00:11:19
chances to do that better than pretty
00:11:20
much anyone else. And you know, we have
00:11:24
actually seen like, you know, for a
00:11:27
given query, the cost to serve that
00:11:29
query is fallen dramatically in a
00:11:31
18-month time frame. What is probably
00:11:33
more of a constraint is latency I would
00:11:35
say. So it's less the cost per query. I
00:11:38
think our ability to serve the
00:11:40
experience at the right latency. You
00:11:42
know search has been near instant. So
00:11:45
how do you uh think about that frontier
00:11:48
has been more of a question. Uh the cost
00:11:50
per query is not what I think will end
00:11:52
up uh you know I think I think we'll be
00:11:54
able to we've done the transition well.
00:11:56
That's that's not a primary driver of
00:11:59
how it'll impact things. Yeah. And do
00:12:01
you have a point of view on ad revenue
00:12:03
per AI query? You know, we already with
00:12:05
AI overviews um you know, we we are at
00:12:08
the baseline of uh uh you know, it's the
00:12:10
same as without AI overviews. And so we
00:12:13
we've reached that stage in uh so but
00:12:15
from there we can improve, right? And I
00:12:17
think uh you I've always felt uh you
00:12:22
know the reason ads have worked well in
00:12:24
search is because commercial information
00:12:26
is also information people in when they
00:12:29
have that intent are looking for that
00:12:30
most relevant information. So I don't
00:12:33
see any reason why AI, you know, just
00:12:36
from a first principal standpoint, why
00:12:38
won't AI do a better job there as well,
00:12:40
right? And and and and so I think I
00:12:44
think, you know, I think we're
00:12:45
comfortable that we can work the
00:12:46
transition through. Some of it may take
00:12:48
time, but all indicators are that we'll
00:12:50
be able to do it well over time. Over
00:12:53
time, but you know, it's already, you
00:12:54
know, already AI overviews when we show
00:12:56
ads. We've kind of reached the baseline.
00:12:58
Do you feel that pressure on Wall Street
00:13:00
and the board and like what's the
00:13:03
tension that you feel as a leader in
00:13:06
trying to manage this transition on the
00:13:08
product on the revenue model for an
00:13:11
organization of this scale? I don't know
00:13:13
how many leaders have done it
00:13:14
successfully in the history of business.
00:13:17
Where do you feel the tension? Where do
00:13:19
you feel the pressure? And how much
00:13:20
leeway are you being given by the
00:13:22
founders and the board to do what's
00:13:24
needed here? Two things. I mean the main
00:13:26
it's a moment of acceleration right so
00:13:28
if anything um the good thing about
00:13:30
these moments is you don't even have
00:13:32
time a lot of times to think about you
00:13:34
know some of those questions you are uh
00:13:36
I think I think a lot about making sure
00:13:38
we have the best models we are we
00:13:41
pushing the frontier as a company uh and
00:13:44
I think the last few months have shown
00:13:46
the breadth and range of what we are
00:13:48
doing uh you know uh we are there and we
00:13:50
have to continue to stay there so for me
00:13:54
you know you you think and you worry a
00:13:55
lot more about execution from within
00:13:57
that's all you know are we are we
00:13:59
executing are we moving fast are we
00:14:01
innovating and I think you know over the
00:14:03
past 12 months I think we've really
00:14:05
picked up pace as a company and uh you
00:14:07
know to to meet the moment so that's
00:14:09
where I do spend a lot of time look as a
00:14:12
as a CEO one of the first things I did
00:14:13
in 2015 in addition to being AI first
00:14:15
was to really bet big on you know we had
00:14:19
great products like YouTube we had
00:14:21
workspace and cloud but really turning
00:14:24
them into robust businesses, right? As
00:14:27
well as great products. Last year, we
00:14:29
exited a combination of YouTube and
00:14:31
cloud at $110 billion. I think, you
00:14:35
know, people don't internalize that
00:14:37
Google is one of the largest enterprise
00:14:39
software companies in the world now. Uh
00:14:41
and and so look, I think and the largest
00:14:43
media company, you know, in in some
00:14:45
ways, right? And you know, definitely
00:14:47
we're doing a podcast. I think we're the
00:14:49
largest podcasting service in the world.
00:14:51
And so you know so I I I feel like you
00:14:54
know as a company we are set up well for
00:14:56
the first time you have this
00:14:57
crosscutting technology you know to to
00:15:00
our earlier point thinking of us as a
00:15:02
deep computer science company what
00:15:04
better technology than AI which
00:15:06
horizontally can impact all aspects of
00:15:08
our business search YouTube cloud and
00:15:12
the other new things we are doing so it
00:15:14
feels like an exciting time so not a lot
00:15:16
of what you know we've continued to do
00:15:18
well in search we are doing well in
00:15:21
these other businesses. Um, and so to me
00:15:24
it feels like, you know, one of the
00:15:26
biggest opportunities ahead as a company
00:15:28
too. I think the next decade ahead looks
00:15:30
to me as exciting as the past decade. So
00:15:33
as I think about my time at Google right
00:15:37
below us in the
00:15:39
garage and his team were building these
00:15:41
super secret shipping container data
00:15:44
centers. They had these like data center
00:15:45
in a box that you could ship anywhere as
00:15:48
long as you had access to water and
00:15:49
power. could connect to the internet and
00:15:51
you could scale data center capacity all
00:15:53
over the world. That was 20 years ago.
00:15:56
Um it's always seemed to me that one of
00:15:59
Google's core and not well understood
00:16:02
advantages was its infrastructure
00:16:03
advantage. Something that Google's
00:16:05
invested in to its core from the
00:16:07
beginning. Can you tell me a little bit
00:16:09
about where you view Google's
00:16:12
infrastructure advantage playing out in
00:16:15
the AI competitive landscape today? How
00:16:19
does it translate
00:16:20
into cost, speed, product quality and
00:16:26
where do you guys think about investing
00:16:28
the 70 billion of capex this year in the
00:16:31
chip layer in the networking the data
00:16:34
center? We can unpack both right like
00:16:36
where our capex is going but on your
00:16:38
first part right like one of the ways
00:16:40
you know we look at the parto frontier
00:16:42
of performance and cost Google literally
00:16:45
is on the parto frontier so we deliver
00:16:48
the best models at the most cost
00:16:51
effective price point right like you
00:16:53
know and our flash series of models are
00:16:56
a real workhorse uh in the industry
00:16:58
right and and part of why we are able to
00:17:02
do that uh is because you know we train
00:17:05
and serve our models on our
00:17:06
infrastructure including TPUs right and
00:17:10
we are in our seventh generation of TPUs
00:17:12
and uh we built our first version in
00:17:15
2017 um I remember talking about it at
00:17:18
Google IO probably people didn't pay
00:17:20
attention to it because like you know
00:17:21
why are you building a specific machine
00:17:23
learning accelerated chip uh look it
00:17:26
plays out everywhere to your earlier
00:17:27
question on cost per query in
00:17:29
search the reason we feel comfortable we
00:17:32
can serve it at that scale scale uh
00:17:34
because we are constantly innovating
00:17:37
through each generation including chips
00:17:39
which are really really good at
00:17:40
inference right and ironwood which is
00:17:43
our latest in our TPU series a single
00:17:46
part of ironwood is over 40 exoflops
00:17:49
right and and so the scale of these
00:17:52
things are incredible and and we have
00:17:55
thought about our in all the way from
00:17:57
subc cables to the scale at which we do
00:17:59
infrastructure is unparalleled and I
00:18:02
I've always viewed that full stack
00:18:05
approach, you know, deep
00:18:08
infrastructure foundational fundamental
00:18:10
R&D on top of it and then you build and
00:18:13
innovate on top of that and I think that
00:18:15
approach will serve us well over time.
00:18:18
But it really, you know, empirically
00:18:20
plays out in, you know, the the cost at
00:18:22
which we are able to provide our models.
00:18:24
Part of the reason we've had a lot of
00:18:26
traction with Gemini 2.5 series is not
00:18:29
only are they great models, but we are
00:18:31
offering it at a very attractive value.
00:18:33
Uh, and we can do that because, you
00:18:35
know, we we are driving our
00:18:37
infrastructure cost down on the $75
00:18:40
billion in capex for uh
00:18:44
2025. You know, obviously majority of
00:18:46
that goes
00:18:48
into servers, data centers uh and so on.
00:18:52
uh servers being the vast portion of it.
00:18:55
Uh I would I would say on on looking at
00:18:58
2025
00:19:00
uh and looking at the compute part of
00:19:01
the uh spend
00:19:04
uh half of that is going towards our
00:19:06
cloud business in 2025 and obviously
00:19:09
that is a very different uh u uh it's a
00:19:13
very different business to search and so
00:19:14
on. So uh a lot of it is to power the
00:19:17
innovations in uh from Google deep mind
00:19:20
pushing the frontier and we're doing it
00:19:22
across many dimensions right not just
00:19:24
large language models but you know even
00:19:26
there doing it across not just text
00:19:29
images video etc building uh world
00:19:32
models right so there's just a lot of uh
00:19:35
innovation which we are pushing on the
00:19:37
frontier obviously to support our core
00:19:39
products like search YouTube gemini etc
00:19:42
but 50% of the compute goes towards
00:19:44
Google cloud Let's just talk about chips
00:19:46
for a second. This is a big part of the
00:19:47
conversation is Nvidia's got the real
00:19:49
market monopoly in AI is what everyone
00:19:51
says. Um, do TPUs provide a wholesale
00:19:55
replacement for your need for NVIDIA in
00:19:58
the supply chain or is NVIDIA still a
00:20:00
core part of the mix in the data center
00:20:02
for training versus inference uh in LLM
00:20:05
versus other models? Maybe just share
00:20:07
your your understanding of where like
00:20:09
the mix evolves to for you guys. Look,
00:20:12
first of all, at a high level, Nvidia is
00:20:13
a phenomenal company. Uh, you know,
00:20:15
Jensen is awesome. We we have been
00:20:18
working with Nvidia now for a very very
00:20:21
long time and we continue to do so,
00:20:22
right? And we serve a lot of the Gemini
00:20:26
traffic on GPUs as well, right? And so
00:20:29
we give customers choice, etc.
00:20:31
Internally, we train our Gemini models
00:20:33
on TPUs, right? And and and we serve it
00:20:35
that way across our products. But uh we
00:20:39
use both and uh I do think look I I do
00:20:42
think everyone in the industry is going
00:20:43
to try and do something like that. But
00:20:45
uh but you know it's it's you know
00:20:47
Nvidia's R&D uh and and their ability to
00:20:51
drive that innovation uh their software
00:20:54
stack is world class. So you know they
00:20:56
have a lot of uh advantages as a company
00:20:58
and I have extraordinary respect for
00:21:00
them. Uh you know but we've always had
00:21:03
you know we are committed you know we
00:21:04
are actually deploying GPUs internally
00:21:06
as well. I think I like that flexibility
00:21:09
and and uh but I I I we are also
00:21:12
long-term committed to the TPU direction
00:21:14
as well. So I think it's a good
00:21:16
combination to have both and and I think
00:21:20
we push each other um you know and drive
00:21:22
the frontier forward. Just going back so
00:21:25
there there's an infrastructure
00:21:26
advantage inherent in all of the
00:21:29
investment that's been made for 20 plus
00:21:30
years and the continued investment. A
00:21:32
lot of folks have said that some of the
00:21:34
performance in foundational LLMs is kind
00:21:38
of starting to plateau and as a result
00:21:41
we're seeing a less kind of
00:21:42
differentiated landscape amongst the
00:21:44
competitors and that's should be a
00:21:47
consideration for Google. That's the
00:21:48
outside kind of narrative. Can you share
00:21:50
a little bit about and then I want to
00:21:51
come back to nonLLM models where there's
00:21:54
other advantages for Google in a minute
00:21:55
but maybe just on this point how much
00:21:58
more um of uh an opportunity to continue
00:22:02
to evolve LLM is there where's Google's
00:22:05
advantage lie in maintaining better
00:22:07
performance in the models over time I
00:22:10
think maybe it was Andre Karpati who
00:22:12
used the term AJI which is like he
00:22:14
called it artificial jag jag
00:22:17
intelligence right so I think the
00:22:19
progress is not going to be always
00:22:21
smooth, right? Like you you go through
00:22:22
these periods, it looks like something
00:22:24
slow and then you see a paradigm
00:22:26
breakthrough, etc. And it's been going
00:22:27
like that for a while. uh I think
00:22:30
obviously over the last couple of years
00:22:32
uh you know all of us scaled up on
00:22:33
pre-training and then there was a lot of
00:22:35
momentum with post-training and then
00:22:37
with inference compute uh and and and
00:22:41
now you know there's progress with how
00:22:44
do you take all that and stitch together
00:22:45
in agentic workflows and you know and
00:22:48
and and so on. So I do think there's a
00:22:51
lot of progress and it feels pretty
00:22:54
continuous to me, right? I think it's
00:22:57
both true progress gets
00:22:59
harder which I think will distinguish
00:23:02
the elite teams at least on the
00:23:05
foundational side. Uh you know I think I
00:23:07
think I think that that might be a
00:23:09
factor. Uh I felt the the harder the
00:23:12
problem is I think you know we are well
00:23:14
set up for that. Uh I think I think we
00:23:16
are well set up for that. I do think we
00:23:18
are uh pushing the research frontier in
00:23:22
a much broader way than most other
00:23:25
people beyond just LLMs transformer
00:23:28
based models I mean diffusion you know
00:23:30
do you do uh diffusion based models all
00:23:34
those areas we are exploring in a deep
00:23:35
deep way right so
00:23:38
um and you know there's always the
00:23:41
chance that we may reach a point where
00:23:43
you know you quite don't get that
00:23:45
returns to the additional compute you're
00:23:47
going put in, but I quite haven't seen
00:23:50
it yet. Right. It it the progress looks
00:23:53
maybe harder because you're now dealing
00:23:55
with a lots more compute. So, you're
00:23:57
really running into the loss of like can
00:23:59
I actually get as many electricians as I
00:24:01
can to build the data centers at the
00:24:03
speed like I you know all that
00:24:05
stuff. But I haven't seen or at least
00:24:08
talking to our researchers haven't seen
00:24:10
anything fundamentally hey like we are
00:24:12
not going to be able to move past this
00:24:15
point or something like that. Does
00:24:16
Google have a data advantage with
00:24:19
YouTube or other products or services?
00:24:21
Are you able to train on that data in a
00:24:23
way that others can't? I think we have
00:24:24
the opportunity to create much better
00:24:26
experiences for people. I think people
00:24:28
use products like Gmail, Calendar, Docs,
00:24:31
YouTube, search
00:24:32
etc. So with their permission taking
00:24:35
that personal context into account I
00:24:38
think we can deliver much better
00:24:40
experiences. We are working on that but
00:24:42
it's it's something on which we have to
00:24:43
deliver but I view that as one of the uh
00:24:46
uh differentiated innovation
00:24:48
opportunities we have ahead as a company
00:24:51
but it's something we are thoughtfully
00:24:52
working on uh we'll make progress there
00:24:55
that makes a lot of sense if search
00:24:57
evolves and I've been using a lot of
00:24:59
voice AI tools I find them incredible I
00:25:03
can have a conversation access the news
00:25:06
dive deep on a topic it's just it's so
00:25:08
incredible
00:25:10
what do Do you view the future of human
00:25:12
computer interaction being 5 to 10 years
00:25:15
from now as AI evolves? As computing
00:25:19
evolves, am I looking at a screen? Am I
00:25:21
typing in a chat? Am I using an AirPod
00:25:24
and just getting audio? Am I doing audio
00:25:26
plus a screen? Is it just a personalized
00:25:29
interface? And there's no even concept
00:25:31
of the web. What does the future look
00:25:33
like for accessing
00:25:35
information and pursuing my interests in
00:25:37
life as a human using compute? It's a
00:25:40
great question. Uh I do think you know
00:25:42
the answer has got to be you know we've
00:25:44
always you humans have adapted to
00:25:47
computing and it's always been that way
00:25:50
but over time the answer will be that
00:25:52
you need to do less of the hard work
00:25:54
less of the adaptation and computing
00:25:56
kind of works for you right and that's
00:25:58
the holy grail I think and and we are
00:26:02
making progress right be touch be voice
00:26:05
everything inches us towards this future
00:26:09
um for example When I wear AR glasses, I
00:26:13
already wear glasses, so it's not that,
00:26:15
you know, but the AR glasses aren't
00:26:17
quite as comfortable as my normal
00:26:18
glasses, but they're getting there. It's
00:26:22
obvious to me that that'll push it to
00:26:24
the next level of seamlessness where it
00:26:26
kind of is ambiently there and doing
00:26:28
stuff for you. So, I think that's the
00:26:31
arrow there, you know, the air of uh uh
00:26:34
how it'll, you know, it has to be more
00:26:36
seamless and just be there for you, you
00:26:38
know. Will it be like neural link down
00:26:39
the line right you know like you know
00:26:41
when I when I want to understand
00:26:42
something you know is it is it that
00:26:44
seamless right you know I I think all of
00:26:47
that is a possibility but I think in the
00:26:49
immediate
00:26:52
world given you're going to have really
00:26:56
natively multimodal models which can
00:26:58
take you know audio vision language uh
00:27:03
all of that and be there in your uh line
00:27:06
of view so I think when AR really works.
00:27:09
I think that'll wow people. Um, I'm not
00:27:12
talking about immersive displays. I'm
00:27:15
talking more about AR glasses, right?
00:27:18
And I think I think that paradigm looks
00:27:20
very interesting to me having used it.
00:27:23
You can kind of feel that next leap,
00:27:25
right? Where uh I think we'll all enjoy
00:27:28
using it in a way, but you still have a
00:27:31
little bit of system integration
00:27:32
challenges to work through. So, we have
00:27:33
maybe couple cycles away to get to that
00:27:36
sweet spot.
00:27:38
uh what smartphones were in around 2006
00:27:41
2007. So, but maybe that's the next
00:27:43
leap, right? And and so probably that's
00:27:45
what's exciting for me. Are you spending
00:27:47
a lot of time on hardware? Yes. Right. I
00:27:50
think um we are definitely excited about
00:27:54
be AR glasses, the next for form
00:27:56
factors,
00:27:58
um you know, uh robotics is another
00:28:01
area, all that. uh and we obviously
00:28:04
build pixel phones uh you know we build
00:28:07
vast data centers so we are definitely
00:28:09
in the physical world you can think of
00:28:10
Whimo as a big robot we are driving
00:28:12
around everywhere so we're making uh
00:28:15
with our partners cars that way so
00:28:17
definitely yes I just want to zoom out
00:28:19
and look at there's this competitive
00:28:21
landscape that's emerged for Google that
00:28:24
maybe maybe it's always been challenging
00:28:26
maybe there's always been competitors
00:28:27
but they're getting a lot of money and
00:28:30
they're investing a lot of money more
00:28:32
than to compete with Google. How have
00:28:36
the founders of Google, I've seen both
00:28:38
of them recently. Sounds like Sergey's
00:28:41
spending time here. Uh they both
00:28:43
independently shared with me that this
00:28:45
is the most exciting thing they've ever
00:28:47
seen in computer science and it's
00:28:48
transforming everything. How engaged are
00:28:51
they? How much time do you spend with
00:28:52
them? And what's your relationship like
00:28:54
there? They are, you know, obviously
00:28:56
fortunate to have both of them uh
00:28:58
involved in their own unique ways uh
00:29:00
deeply. I talk to them all the time. Um,
00:29:04
look, I think both Larry, Sergey, I, you
00:29:06
know, you know, credit to them. They
00:29:09
always envision like where AI would be.
00:29:11
I think, uh, you know, I think I think
00:29:14
their ability to understand trends and
00:29:16
and you know, I I swear I've had
00:29:19
conversations maybe as early as like 15,
00:29:22
20 years ago about moments like this
00:29:24
with them. I think they are both would
00:29:27
argue that this is the most exciting uh
00:29:30
time in the field uh you know and and I
00:29:32
they both engage in their own ways. I
00:29:35
think Sergey is definitely uh spending
00:29:37
time with the with the Gemini team at a
00:29:41
uh you know in a pretty hardcore way
00:29:42
like you know sitting and coding and uh
00:29:45
spending time with the engineers uh and
00:29:47
that gives the energy to the team which
00:29:49
I think it's unparalleled right like to
00:29:51
have a founder sitting there looking at
00:29:54
loss curves giving feedback on model
00:29:56
architectures
00:29:58
uh how can we improve post training etc
00:30:00
I think I think you know it's a it's a
00:30:02
rare rare place to be But you know my
00:30:05
favorite conversations are sometimes
00:30:07
when the three of us sit and talk. Uh
00:30:09
the combination of I mean they are very
00:30:12
nonlinear thinkers. So I feel like it
00:30:15
expands the conversation into ways which
00:30:17
you always don't expect and out of it
00:30:20
which comes interesting ideas. So I
00:30:22
think I always have access to that but I
00:30:24
think I've worked with them uh for such
00:30:26
a long time you know you know there is
00:30:29
friendship, respect, mutual dialogue. uh
00:30:32
we love doing that and uh and I think
00:30:34
it's uh we'll I'll always have that.
00:30:36
Your competitors out there have active
00:30:39
founders. OpenAI has SAM, XAI has Elon,
00:30:43
Meta has Zuck and um Microsoft has
00:30:47
Satcha. Are you willing to kind of share
00:30:49
your perspectives on those four
00:30:51
competitors, both the companies and the
00:30:53
leaders? Look, it's a obviously by by
00:30:56
definition it's a very impressive group,
00:30:58
right? And um I think I think you're
00:31:00
talking about some of the best
00:31:02
companies, some of the best
00:31:03
entrepreneurs uh uh all that look I
00:31:07
I you know it shows how uh both how much
00:31:10
progress we are going to see because
00:31:13
you're basically talking about many
00:31:15
people uh who are working hard to drive
00:31:19
that progress. Right? So to the earlier
00:31:22
question when you were talking about are
00:31:23
we going to see progress the answer has
00:31:25
got to be yes because of the uh you know
00:31:28
the the unique types of people here
00:31:30
pushing progress right so look each of
00:31:34
them they're they're different people I
00:31:35
I fortunate to know all of them and I
00:31:38
think maybe only one of them has invited
00:31:39
me to a dance not the others uh right
00:31:43
but I just look I spent time with Elon
00:31:45
maybe two weeks ago um when I talked to
00:31:48
him and his ability to will future
00:31:51
technologies into existence. I think
00:31:54
it's just unparalleled. So like look,
00:31:55
these are phenomenal people. I respect
00:31:57
all of them. U you know there's
00:31:59
partnerships involved, there's
00:32:00
competition involved. But if I were to
00:32:03
step back and say, you know, at the end
00:32:07
of the day, I
00:32:10
love, you know, driving technology
00:32:12
progress in a way that impacts people
00:32:14
positively. um when you think about
00:32:17
areas like healthcare and and other
00:32:20
important areas education and you know
00:32:22
like we are now talking about this is
00:32:23
why AI is so profound so the opportunity
00:32:26
is what excites me I think all of us are
00:32:29
going to do well in this scenario that's
00:32:31
how I think about it right I think
00:32:32
that's what a lot of people don't gro
00:32:34
and I think this is an important point
00:32:36
everyone out there says there's
00:32:37
competitors there's a winner and
00:32:39
everyone else is a loser but this is an
00:32:41
entirely new world that's going to be a
00:32:43
lot bigger than the world we had last
00:32:45
here and uh everyone's building down
00:32:48
their own path, but there's going to be
00:32:50
a lot of success. It's not just that
00:32:52
who's going to beat whom in the in the
00:32:55
marketplace. When the internet happened,
00:32:57
right? Google wasn't even around. Right.
00:32:59
Right. So, we obviously So, the the
00:33:02
other thing you can say is there are
00:33:04
companies we don't even know, haven't
00:33:06
been started yet, their names aren't
00:33:09
known, might be extraordinarily big
00:33:12
winners in the AI thing, right? So it's
00:33:15
it's going to be AI is a much bigger uh
00:33:19
landscape opportunity landscape than all
00:33:21
the previous technologies we have known
00:33:23
combined combined and so you know so
00:33:26
which is why I think it's all about uh
00:33:28
you know the companies which will end up
00:33:30
doing well or you will do well because
00:33:32
you're able to innovate and execute with
00:33:34
the best talent you know that's that
00:33:36
ends up being the driver well let's talk
00:33:38
about that and let's talk about the
00:33:40
unknown competitor Deepseek popped up
00:33:42
tell me about your impression of the
00:33:44
model, the performance, the rumors about
00:33:46
the next model, and what does that tell
00:33:48
you about what's going on in China and
00:33:50
what's going on that we're not seeing?
00:33:52
Look, I think the main moment from
00:33:54
Deepseek was uh look always, you know,
00:33:59
if you if you kind of follow the AI
00:34:01
research and scan through papers and
00:34:03
read them, no, nobody who does that
00:34:05
would underestimate China, right? Like,
00:34:07
you know, so when you look at the amount
00:34:09
of research output from China, right? Um
00:34:13
they have extraordinary talent.
00:34:16
Um and and and so but I do think all of
00:34:20
us had to adjust our priors a little bit
00:34:23
after the deepseek moment which was like
00:34:25
wow they are even closer to the frontier
00:34:29
than most people maybe assume you know
00:34:31
and so I think I think that was a
00:34:32
moment. I think internally for us uh I
00:34:35
think externally people are very
00:34:36
impressed and rightfully so with how
00:34:38
efficient their models were.
00:34:40
Interestingly for us internally we
00:34:42
benchmarked it to flash and you know
00:34:44
flash was uh as efficient or you know
00:34:47
you could argue in some ways better. So
00:34:50
you know I think I think to our earlier
00:34:52
conversations I I do think this is more
00:34:55
maybe internal baseball for us you know
00:34:57
we were benchmarking and saying look
00:34:59
it's good to see you know because they
00:35:00
had to work in a hardware constrained
00:35:02
way I think which is what drove a lot of
00:35:03
their uh innovations and efficiency
00:35:05
improvements and you know uh and so I
00:35:08
was pleased with that but you know it
00:35:10
tells you that the frontier is is uh
00:35:15
evolving rapidly there are more players
00:35:17
closer to it than people fully realize
00:35:19
And it's going to be a very dynamic
00:35:21
moment uh in the industry. I think China
00:35:24
will will will be uh very very
00:35:27
competitive on the AI frontier is just
00:35:28
what I always assumed and much of the
00:35:33
narrative and I think probably the fact
00:35:36
around the ability to deploy AI at scale
00:35:40
is one that is predicated on
00:35:43
availability of electricity. Mhm. Even
00:35:46
Elon, and I've been talking about this
00:35:47
for a while on my podcast, but Elon this
00:35:49
week is saying, "Hey, I need a terowatt
00:35:51
of compute." Uh, terowatt is roughly the
00:35:53
power production or the electricity
00:35:55
production capacity of the entire United
00:35:56
States. US is going from 1 to two
00:35:59
between now and 2040. China's going from
00:36:01
3 to 8 and there's probably upside given
00:36:04
all the new electricity production
00:36:05
technologies that they're rolling out
00:36:06
now, which will be additive to that. How
00:36:09
much is electricity generation going to
00:36:11
play a role in who is going to
00:36:13
economically benefit from AI over the
00:36:16
next 10 to 15 years and where is the US
00:36:19
compared to China and maybe where is
00:36:20
Google. Well, look, you are uh
00:36:22
definitely uh hitting on you know what
00:36:25
is
00:36:26
uh you know when you when you look at
00:36:28
any system you want to find where the
00:36:30
constraint is because that's what like
00:36:31
gates the whole system and you are
00:36:34
rightfully identifying uh the most
00:36:36
likely constraint for uh AI progress and
00:36:40
and hence by definition GDP growth and
00:36:43
all that stuff right so I do worry about
00:36:45
it a lot um but you know the answers or
00:36:51
you know sometimes you run into
00:36:52
challenges which are you know you have
00:36:54
to solve you know you're running into
00:36:56
physics barriers or something like that.
00:36:58
This is not a problem like that, right?
00:37:00
Like we already know the technologies
00:37:04
that can work to supply the demand we
00:37:07
need. So it's more to me an an execution
00:37:12
challenge, right? I would I would phrase
00:37:14
the energy problem as uh it's obviously
00:37:17
multifaceted. Uh but I think I think be
00:37:20
it really embracing we shouldn't have
00:37:24
innovators dilemma in the energy sector,
00:37:26
right? So we should lean into all the
00:37:29
possible innovations ahead and there are
00:37:31
many of them. Obviously first of all
00:37:33
people perpetually I think will
00:37:35
underestimate solar right you know solar
00:37:37
plus batteries will end up being huge.
00:37:40
uh you know obviously uh the amount of
00:37:42
innovation that's going into nuclear
00:37:45
geothermal all of that are uh
00:37:48
opportunities to uh uh embrace and more
00:37:51
I'm not mentioning but I think you know
00:37:54
upgrading the grid
00:37:56
uh you know solving for
00:37:58
transmission uh you know permitting to
00:38:02
make all of that progress faster and
00:38:04
then actually I think we maybe workforce
00:38:06
constraint like you know to to my
00:38:08
earlier point Right? You know, I think
00:38:11
we are all if you look at the number of
00:38:14
electricians leaving the
00:38:16
workforce versus
00:38:18
suddenly all of us and if you you
00:38:20
project out this demand, there's a huge
00:38:22
mismatch, right? So, literally how you
00:38:25
know how do you make sure there is
00:38:28
incentives and workforce development to
00:38:30
address shortages like that over the
00:38:32
next decade will end up being important
00:38:35
policies. Uh I think we are fortunate
00:38:37
you know people like Secretary Wright
00:38:39
and Secretary Bergam I mean they are
00:38:42
very I think deeply aware uh of the
00:38:44
issue and I think they are uh hitting
00:38:47
hitting the problem hard but I
00:38:49
definitely think it's solvable but I
00:38:51
think we all have to put our mind
00:38:52
towards it but for your business today
00:38:54
you don't see electricity constraining
00:38:56
growth in the business in this moment or
00:38:58
in the projectable future. No, I won't
00:39:01
say that right like just for example we
00:39:03
are supply constraint this year in our
00:39:04
cloud business right and when we are all
00:39:09
of us are simultaneously looking to
00:39:11
scale up data centers right so we are
00:39:15
running into real constraints the way
00:39:17
the constraints play out today is delays
00:39:19
in projects because of permitting or you
00:39:23
know not having access to electricians
00:39:25
all of that is realities all of us are
00:39:27
dealing with right So if this trend line
00:39:31
continues the pace at which we are all
00:39:33
ramping up and obviously for it to
00:39:36
continue we all have to generate the
00:39:37
returns on it and you know and and so it
00:39:39
has to really impact the economy in a
00:39:41
more substantive way. So they go hand in
00:39:43
hand uh if the trend continues these
00:39:46
constraints will be much more visible I
00:39:48
think today we are all working through
00:39:49
these constraints uh so I think there
00:39:52
are real constraints today uh but I
00:39:54
expect it to for us to be competitive
00:39:56
with China etc. I think we have to solve
00:39:58
these constraints in the near future.
00:40:01
What does that look like then? Fast
00:40:03
forward 15 years. The US has 25% of the
00:40:07
electricity of China. Mhm. Is China just
00:40:10
bigger GDP in that moment? Is the pie
00:40:13
going to grow for everyone? You know,
00:40:15
how do we kind of think about how, you
00:40:17
know, the way I've assumed is the US is
00:40:19
always there's never been a time where
00:40:21
US just doesn't meet these moments,
00:40:23
right? So to me I look at it and say it
00:40:26
just means that you know the capitalist
00:40:28
solutions will innovate through this
00:40:30
moment right that's why people are
00:40:32
working hard to build SMRs and uh
00:40:35
nuclear fusion etc. So I've kind of
00:40:37
assumed we will meet that moment and if
00:40:41
we don't or if if the if the lines don't
00:40:43
match I think the conversations will get
00:40:45
louder and louder till we meet the
00:40:47
moment. Uh that's that's the way I
00:40:48
internalize it. There's a history of
00:40:50
Google investing in innovative
00:40:52
technologies and being ignored or being
00:40:56
told that they don't make much sense.
00:40:58
Good luck. The TPU is a great example.
00:41:01
The acquisition of Deep Mind is a great
00:41:03
example. The investment in
00:41:04
infrastructure is a great example. The
00:41:07
insane continued investment forever in
00:41:10
Whimo. Yeah. Is a great example. And
00:41:12
suddenly it looks like Whimo's on track
00:41:14
to be a hundred billion dollar business
00:41:16
and this is actually going to work.
00:41:18
mindblowing persistence and patience. By
00:41:22
the way, we have we are doing the same
00:41:24
patient approach in many other areas.
00:41:26
That's my next question. Quantum is one.
00:41:28
So tell me about quantum but you know
00:41:30
but because everyone ignores quantum.
00:41:32
You've had this investment for some
00:41:34
time. Why is quantum so important?
00:41:38
Because again I want to use the
00:41:39
historical data that it does it seems
00:41:41
like a small bet. Good luck. But what
00:41:42
does quantum evolve to from a compute
00:41:45
perspective for humanity and when does
00:41:47
that happen do you think? Obviously
00:41:48
quantum has gotten a lot more attention
00:41:50
in the last 12 months or so. Uh but
00:41:52
we've been work just like Whimo we work
00:41:55
through these things whether there's
00:41:56
attention from the outside or not
00:41:58
because we are working on these things
00:41:59
out of conviction on the long-term
00:42:03
trends. Right? So it's it comes from
00:42:04
those first principles. Um obviously the
00:42:08
universe is fundamentally quantum. uh
00:42:10
you know to to do any kind of large
00:42:12
scale uh simulations in a way that truly
00:42:14
represent nature you know you would need
00:42:17
uh some versions of quantum computing. I
00:42:19
think to me quantum feels like where AI
00:42:23
was
00:42:25
around you know
00:42:27
2015. So I would say in a 5-year time
00:42:32
frame you would have that moment where
00:42:34
some a really useful practical
00:42:38
computation you know is done in a
00:42:41
quantum way far superior to classical
00:42:43
computers and that'll be that aha moment
00:42:46
uh you know I think which will really
00:42:48
show the promise of the industry.
00:42:50
uh I'm absolutely confident that we will
00:42:53
get there when I see the progress and I
00:42:56
can pattern match to progress in the
00:42:58
other fundamental areas we have worked
00:43:00
on. So it really doesn't feel like
00:43:03
obviously look these are very
00:43:04
challenging areas. You may hit a
00:43:05
constraint. I do think a lot of people
00:43:08
are making announcements in quantum. So
00:43:09
in some ways it's tough to distinguish
00:43:11
them. We had the same scenario in
00:43:13
self-driving maybe 3 years ago. There
00:43:15
were so many people doing
00:43:17
self-driving. It looked like everyone
00:43:19
was roughly the same but they weren't.
00:43:22
Uh I could internally tell the
00:43:24
difference that how far ahead Whimo was.
00:43:26
I feel that way about our quantum effort
00:43:28
too. I think there are a lot of
00:43:29
announcements, a lot of noise in the
00:43:30
industry. There are few good people, but
00:43:33
like you know, you know, but I I do
00:43:34
think we are at the at the frontier
00:43:36
there and so you know I'm I'm pretty
00:43:40
excited about it in a 3 to 5 year time
00:43:43
frame, but we'll be patient and get
00:43:44
there. Yeah. Do you want to speculate on
00:43:47
a business in quantum? Look, I I we are
00:43:49
committed to, you know, in almost all
00:43:52
these cases, our goal would be to, you
00:43:56
know, demonstrate more and more useful
00:43:59
practical algorithms and show progress
00:44:01
on that and and give access to it
00:44:04
through cloud, right? And
00:44:07
and I think, you know, I always say it's
00:44:11
tough to project innovation on top of a
00:44:13
platform, right? Nobody could say just
00:44:15
because you had smartphones and GPS and
00:44:17
payments something like Uber would get
00:44:19
invented. You couldn't linearly sit and
00:44:22
project Uber from the underlying
00:44:24
innovation. That's how the world works.
00:44:26
And so uh for me quantum is that
00:44:28
foundational again just like AI there's
00:44:31
going to be extraordinary innovations on
00:44:32
top of it. We don't know the algorithms
00:44:34
yet. It's almost like trying to predict
00:44:35
how people would use personal computers
00:44:37
in 1977 or something. That's right.
00:44:39
We're very early and you know some of
00:44:41
the the constraints in quantum are that
00:44:44
there aren't quantum computers to test
00:44:46
them out new algorithms to test them
00:44:48
out. There's a lot of theory in quantum
00:44:49
algorithm development but not a lot of
00:44:51
testability experimentation at this
00:44:53
point. We are working on all of that
00:44:55
too. I think we'll have more exciting
00:44:56
moments to share this year. So look
00:44:58
forward to making I think that that's
00:45:00
what's interesting. It will expand
00:45:01
people's minds of the potential of what
00:45:02
you can actually do. Right now, no one
00:45:04
really knows how to think about quantum,
00:45:06
where it's going to take us. But those
00:45:07
announcements, I think, are going to be
00:45:09
really preient. And then I'm assuming
00:45:11
all your friends will show up and say,
00:45:13
"We've got a quantum effort." Now, too.
00:45:16
Tell me about robotics. I think this was
00:45:17
going to be the year of the robot. We
00:45:19
see so many models being trained on
00:45:22
simulation data or real world kind of
00:45:24
observational data that are then being
00:45:26
used to control physical systems. Call
00:45:28
it physical AI, call it robotics. Lots
00:45:31
of startups, lots of big companies.
00:45:32
Google bought Boston Dynamics and a
00:45:35
bunch of other robotic companies. I
00:45:36
think Andy Rubin was overseeing these
00:45:38
for a while and then you sold them off
00:45:40
and decided it was too early. What's
00:45:42
your point of view on the opportunity in
00:45:44
robotics today? How does Google play
00:45:46
here? We are definitely for robotics,
00:45:49
you
00:45:50
know, we again have probably, you know,
00:45:53
one of the most advanced frontier R&D
00:45:56
teams in the world now. uh you know and
00:45:59
and the Gemini robotics efforts around
00:46:01
vision language, action models etc are
00:46:03
world class. I do think um you know
00:46:07
robotics uh you know so we are now
00:46:10
thinking through how we either partner
00:46:13
or where we actually bring products out.
00:46:15
you are right we we tried the
00:46:18
application layer too early uh where I
00:46:22
think robotics wasn't really being
00:46:23
influenced by AI as much but now it's
00:46:26
it's really the combination of AI plus
00:46:28
robotics that gives that next sweet spot
00:46:30
right and and so we are uh making plans
00:46:34
there uh nothing to share today but you
00:46:36
will see us make more announcements in
00:46:38
the space but we are definitely uh
00:46:40
foundationally driving uh the underlying
00:46:44
uh models and We are building
00:46:46
state-of-the-art models there. We are
00:46:47
working with partners and testing it.
00:46:51
You know that when I look at the
00:46:52
progress of humanoid robots etc. I mean
00:46:55
they are uh you know in the past I would
00:46:58
say oh these this is obviously uh you
00:47:02
know you can see how janky they are. Now
00:47:05
I have to take 5 seconds to look at it
00:47:07
and say closely and say is this fake or
00:47:09
is this an actual robot doing it? Right.
00:47:11
Right. like already I'm in that moment
00:47:13
and so and so you can see the progress
00:47:16
uh in the field underway. So I think you
00:47:18
know we are probably two to three years
00:47:20
away from that magical moment in
00:47:23
robotics too and so so that's the next
00:47:25
exciting phase is a good way to think
00:47:27
about it that Google could potentially
00:47:29
develop the Android for robotics and
00:47:31
ultimately have a broad play here. Yeah,
00:47:33
we have intrinsic so one of our bets is
00:47:35
effectively doing that. So uh supporting
00:47:38
uh robotics manufacturers
00:47:40
um you know we we are committed to
00:47:43
having the Gemini as a model you know
00:47:47
will will take all modalities into
00:47:49
account work very very well for robotics
00:47:51
it's definitely something we are
00:47:53
committed to being on how we actually
00:47:55
bring products out first party versus
00:47:56
third party etc is where we are thinking
00:47:58
I want to talk a little bit about
00:48:00
culture which seems to be a key
00:48:03
differentiator on the kind of
00:48:05
competitive landscape I go back to
00:48:07
thinking about Google offering free
00:48:09
food, massages at work, 20% time as a
00:48:11
way to attract and win in the early days
00:48:14
of the talent wars in Silicon Valley,
00:48:16
early 2000s and and and that persisted.
00:48:19
But what happened is it grew and it
00:48:20
became more amenities and the narrative
00:48:22
is that Google ended up creating a
00:48:25
culture that kind of moved away from
00:48:28
more accountability and performance and
00:48:30
was much more about coddling employees.
00:48:33
Can you just comment on kind of your
00:48:35
observations on the evolution of Google
00:48:36
over the 20 years that you've been here
00:48:38
and what you've tried to do lately as a
00:48:39
leader? How you think about the culture
00:48:42
you want to foster and what you're doing
00:48:43
about it? Look, I think it's important
00:48:45
to step back and say, you know, the
00:48:48
underpinnings of a culture in which you
00:48:50
really invest in employees and and you
00:48:52
empower them and and even some of the
00:48:55
perks was to create a a a culture where
00:49:01
it's positive, optimistic, you're in an
00:49:04
innovation mindset, people are talking
00:49:06
to each other. Maybe by giving lunch
00:49:08
here, people are all sitting and talking
00:49:10
ideas through lunch, you're
00:49:12
cross-pollinating. Imagine. So you know
00:49:14
that is the thesis of it. Not that we
00:49:15
are trying to give lunch to people right
00:49:17
and and so I till today feel you know we
00:49:23
still get a lot of innovation in the
00:49:25
company at all levels of the company and
00:49:27
I think people wake up um you know and
00:49:31
say well I can go do this notebook LM
00:49:33
etc are great examples right and so
00:49:35
people do that all the time. So I think
00:49:37
empowering
00:49:38
employees has been and is and will be a
00:49:41
source of strength for Google right I
00:49:43
think we can attract higher caliber
00:49:45
people who feel like they have agency to
00:49:47
do that right and but that doesn't mean
00:49:51
like you know I think I think people
00:49:52
shouldn't confuse that with like today
00:49:54
for example you can take something like
00:49:55
Google deep mind I think there is all
00:49:58
the way from Demis and others you know
00:50:00
extraordinary leadership team be korai
00:50:03
Jeff Oral no me etc all these leaders
00:50:05
have strong opinions depends on how to
00:50:07
drive that frontier forward and and
00:50:10
that's happening too right so I think
00:50:12
it's important to strike a balance
00:50:14
balance uh between the two I think when
00:50:17
you empower employees a lot in some ways
00:50:19
like you know we have allowed for more
00:50:20
free speech than other companies that's
00:50:22
one way you can think about it so you're
00:50:24
going to hear voices sometimes you can
00:50:27
hear like what is effectively 500 people
00:50:30
in the company but that doesn't
00:50:32
represent the company as a whole so in
00:50:34
some ways we are different from other
00:50:35
companies and can confuse use it on the
00:50:36
outside I think but I think overall look
00:50:40
we have a clear sense of where we are
00:50:41
going I think we want to empower people
00:50:45
all in the service of our mission so if
00:50:47
anything you know over the past few
00:50:49
years and you are right there are
00:50:50
moments not just us but as an industry I
00:50:54
think uh I think some of the other
00:50:57
things became more of the focus than the
00:50:59
mission of the company and why we are
00:51:01
all here right like we are we are not
00:51:03
all here in the company to resolve all
00:51:06
our personal differences or something.
00:51:08
We are here because you're excited
00:51:10
about, you know, innovating in the
00:51:13
service of the mission of the company
00:51:14
and and the impact you can have. And so
00:51:16
bringing that focus back, that's
00:51:18
something I've been very deliberate
00:51:19
about for for the past few years. And I
00:51:22
think it needs reinforcing. I think one
00:51:24
of the lessons for me was we all grew so
00:51:27
much that you assumed everyone always
00:51:30
understood those underpinnings but then
00:51:33
when you added so many people you
00:51:35
realize you have to go back and repeat
00:51:37
that a lot uh to to help people
00:51:39
internalize that. We've done that and we
00:51:41
do that all the time. Uh I think moments
00:51:44
like this help a lot too. The current
00:51:47
moment is just genuinely both so
00:51:49
exciting and so intense. It actually
00:51:52
reminds me a lot of early Google, right?
00:51:54
You know, when I walk into the GDM
00:51:57
building, you know, some of our earliest
00:51:59
engineers are all sitting there working
00:52:03
together. People come in 5 days a week
00:52:06
at a minimum, right? And so you have
00:52:08
that intensity and you have that
00:52:10
excitement and I feel that same sense of
00:52:13
optimism. So that's what I'm focused on,
00:52:15
right? You know, to me that's the
00:52:17
hardcordness which matters, right? like
00:52:19
are people smart
00:52:20
people really working with a passion and
00:52:24
that's where that intensity comes from
00:52:26
and and you have to work hard to create
00:52:28
that and you know there are pockets of
00:52:30
the company if that doesn't happen you
00:52:31
figure out what are the changes you need
00:52:33
to make uh to to do that right and
00:52:36
sometimes for example I recreated the
00:52:39
notion of
00:52:39
labs right and and and because I said
00:52:43
well there are things that are possible
00:52:45
with 10erson teams and so we need go and
00:52:49
do that again and and there are quite a
00:52:52
few projects both we have shipping and
00:52:54
are underway to come which will be an
00:52:56
outcome of those efforts as well. So you
00:52:58
know your culture your values are
00:53:00
enduring culture is something you're
00:53:02
constantly tweaking to make sure you're
00:53:04
true to your values and so by definition
00:53:07
there's going to be drift and you know
00:53:10
you work hard to uh uh snap it back. Was
00:53:13
there a moment in the last 10 years
00:53:16
where you said I've got to spend more
00:53:17
time on this? Oh, you know, uh, for
00:53:19
sure. I look, I think CO was such a big
00:53:22
distortion to to the to our way of
00:53:24
working, right? So, you know,
00:53:26
fundamentally Google was designed to be
00:53:29
uh a culture in which people were seeing
00:53:32
each other, engaging with each other.
00:53:35
So, losing that continuity, right, I I
00:53:37
think definitely impacted our culture.
00:53:39
So when we have we've gotten people back
00:53:42
uh in a 3-2 model and some teams are you
00:53:45
know work uh beyond that I think it's
00:53:48
been important I've I've spent time to
00:53:50
get those connections back like you know
00:53:52
for example GDM we were intentional in
00:53:54
creating a physical space where we can
00:53:57
get all of them back in the same
00:53:59
building both in London both in Mountain
00:54:02
View and and and taking our newest
00:54:05
building um you know with that kind of a
00:54:08
tentlike roof structure and putting all
00:54:10
the people in and being intentional
00:54:12
about it has made a massive difference.
00:54:15
Have you found a shift in your ability
00:54:18
to recruit top talent? A lot of great
00:54:20
talent has started other great
00:54:22
companies, other great companies in
00:54:24
Silicon Valley have recruited folks. I
00:54:26
know there's always a talent war going
00:54:28
on, but has there been a tenor shift for
00:54:30
Google in the last period of time
00:54:33
because of some of the underlying
00:54:34
advantages in AI or some of the cultural
00:54:36
changes that are underway? the talent
00:54:38
market you know we go through these
00:54:41
fierce moments for talent AI is one of
00:54:43
them and whenever there are these Google
00:54:47
you know obviously we we are fortunate
00:54:49
to have some of the most talented
00:54:50
employees so we are a source I'm equally
00:54:52
proud of the fact that I think Googleers
00:54:54
have left to start over 2,000 companies
00:54:56
right and and so you know there is a
00:54:59
virtuous cycle I think people come back
00:55:01
we acquire companies I think all of that
00:55:03
keeps the company uh fresh but in the
00:55:06
current AI moment
00:55:08
Look, I think we are both holding on to
00:55:10
critical talent. We are recruiting. You
00:55:12
know, I always look at the tip of the
00:55:13
tree of are we able to attract the best
00:55:16
PhD researchers coming out of the top
00:55:19
programs and the answer is yes. Um you
00:55:22
know and and there are people who have
00:55:23
left who have come back and so I feel
00:55:26
good about the position we are but you
00:55:28
work at it hard every week, every month
00:55:30
and so on. Do you think this is going to
00:55:31
change in the future with how we do
00:55:35
education and how AI plays a role in
00:55:37
education? Are you going to be able to
00:55:39
identify, recruit, and then teach and
00:55:42
train talent out of high schools and at
00:55:44
an earlier age? And the traditional kind
00:55:46
of college education system is going to
00:55:48
change because of AI on the job
00:55:51
training. There's a lot of potential to
00:55:52
change. I just there's a part of me
00:55:55
which feels maybe we've all
00:55:56
misunderstood what colleges are about
00:55:58
and maybe colleges are about that
00:56:00
community and people getting together
00:56:03
and exchanging. So it you know there may
00:56:06
be intangibles which which would still
00:56:08
uh maybe make the uh uh uh you know it
00:56:12
more valuable than like we all perceive
00:56:14
it to be but um but the way I think
00:56:17
about it is you're going to get
00:56:19
extraordinary talent at more places
00:56:21
around the world. So that's the way I
00:56:23
think about it because people have
00:56:25
access to with AI. So you don't need to
00:56:28
be in a few certain places to be that uh
00:56:32
that great uh great talent. So I think
00:56:34
the nature of that changes. By the way,
00:56:36
I think it's an important thing to
00:56:37
internalize. We often talk about talent.
00:56:39
We've always been able to recruit the
00:56:41
best talent uh in the country. But now
00:56:44
there's extraordinary talent emerging in
00:56:46
other parts of the world too. So I think
00:56:49
it's something not to lose uh line of
00:56:51
sight and you know maybe that's the way
00:56:53
I would think about it. So just taking a
00:56:54
step back zooming back I had a
00:56:56
conversation 10 years ago with Larry
00:56:58
Paige where he talked about the
00:56:59
transition from Google to Alphabet.
00:57:01
Alphabet is going to be this holding
00:57:02
company. It's going to discover or
00:57:05
develop the next hundred billion dollar
00:57:07
revenue business. At the time, I think
00:57:09
Google wasn't quite at hundred billion.
00:57:10
There have been a lot of these
00:57:11
investments and other bets since that
00:57:13
time. Do you still think about Alphabet
00:57:16
as a holding company? Are there still
00:57:19
multiple businesses that you want to
00:57:21
kind of stand up and foster and have
00:57:23
this kind of holding company model? Is
00:57:26
that still hold or is Google really the
00:57:29
core engine that's going to continue to
00:57:31
evolve and continue to have ancillary
00:57:34
businesses that are you know somewhat
00:57:35
adjacent to Google? I'll answer it two
00:57:37
ways, right? So I think the way we are
00:57:41
not a holding company in the sense that
00:57:43
we are we are not just like looking to
00:57:45
invest capital in other attractive
00:57:48
businesses that's that's not who we are
00:57:50
right uh we are you know from a
00:57:54
foundational technology basis if you can
00:57:56
take that
00:57:57
technology and that R&D we do and
00:58:01
identify problems in which we can
00:58:03
innovate and bring a differentiated
00:58:06
value proposition we'll do do that,
00:58:08
right? So that's the way we approach and
00:58:09
so the the structure is an outcome of
00:58:11
that, right?
00:58:13
So and which means you will have
00:58:16
businesses uh on on on on paper they may
00:58:20
look like very desperate but there's a
00:58:22
common strand underneath them right so
00:58:25
like mayo is going to keep getting
00:58:28
better because of the same work we do in
00:58:30
Gemini and AI over time as Google cloud
00:58:34
to search to YouTube to isomorphic to
00:58:37
robotics etc. So that is the unifying
00:58:40
layer right and and then it's a
00:58:43
continuum is Google cloud a Google
00:58:45
business or an alphabet business right
00:58:47
you know we segment it out right and so
00:58:51
uh so you know so the the the branding
00:58:54
matters less I think right you know
00:58:55
we'll have a range of companies some of
00:58:57
them will leave and IPO out because
00:59:00
maybe that's the best way they can make
00:59:01
progress so all of that is a possibility
00:59:04
but what I think I founders think about
00:59:08
is like the underlying innovation by
00:59:11
which so we think at the units of
00:59:14
quantum right you know alpha fold and
00:59:17
hence isomorphic right you know
00:59:19
self-driving and and building the way
00:59:21
more driver and hence all the businesses
00:59:23
on top of it so it's more maybe maybe
00:59:26
that's how we think about it right does
00:59:28
X still play a big role in driving
00:59:30
innovation and you continue to invest
00:59:31
there yeah look I think if anything um X
00:59:34
over time look a lot of lot of these
00:59:37
innovations did that come out of X,
00:59:38
right? And so, um, including Whimo, uh,
00:59:42
you know, the early incarnations of
00:59:44
Google Brain, right? Yeah. So, so I
00:59:47
think, uh, X as an incubator, um, allows
00:59:51
us to, uh, you know, push the
00:59:53
boundaries. They're thinking about uh,
00:59:56
tapestries, thinking about the grid
00:59:58
problem. um that are uh you know uh
01:00:02
extraordinary but but it's all rooted in
01:00:05
computer science, physics, kind of a
01:00:07
deep uh technology R&D and I think
01:00:10
that's the foundation across everything
01:00:11
we do. As we wrap up um I want to ask
01:00:14
you one last question to hopefully frame
01:00:17
your experience of the last 10 years as
01:00:19
a CEO. Biggest regret, biggest mistake
01:00:23
and what you're most proud of. Proud is
01:00:25
obvious. Look, I I think I think we have
01:00:27
as a company I think there aren't that
01:00:30
many companies which can push the
01:00:31
technology frontier like you don't hear
01:00:34
of companies winning Nobel prizes often.
01:00:36
That level of foundational R&D we do and
01:00:39
then apply it to create businesses and
01:00:41
value. I think I think we've done an
01:00:44
extraordinary job at that and we aspire
01:00:46
to do that. You know, I'm I'm really
01:00:48
proud of that. I think we're pretty
01:00:50
unique as a company that way. You know,
01:00:52
there a lot of small regrets, you know,
01:00:54
by nature. I tend to look forward and I
01:00:56
learn from mistakes we make. But look,
01:00:59
there are acquisitions we debated hard
01:01:02
came close and you know some of them are
01:01:06
just give me one name
01:01:10
or get in trouble. Maybe Netflix, right?
01:01:11
Like we debated Netflix at some point
01:01:13
super intensely inside. So you go
01:01:16
through these moments, right? and and uh
01:01:18
and uh you know and so uh I wouldn't
01:01:22
call it regrets but you always look back
01:01:23
and like you know like you know in in a
01:01:25
world of butterfly effects there were
01:01:27
alternate paths but maybe they are in a
01:01:29
different part of the multiverse. Yes.
01:01:31
Yes. I always um tell people I I think
01:01:34
they underappreciate the role that Bell
01:01:38
Labs played in driving innovation and
01:01:41
ultimately human prosperity in the early
01:01:42
20th century. And I do think a lot of
01:01:44
people underappreciate the role that
01:01:46
Alphabet is playing in driving
01:01:48
innovation across so many different
01:01:50
lanes which drives prosperity,
01:01:53
businesses, competition, all that stuff
01:01:55
aside, the innovation that's being
01:01:57
driven out of Alphabet continues to
01:01:59
impress and benefit us all. And so I
01:02:01
want to thank you for your leadership
01:02:02
and the time. Thanks, David. Real
01:02:04
pleasure.
01:02:10
[Music]
01:02:14
I'm going all in.
