00:00:00
special episode this week. We've
00:00:02
preempted the weekly show and there's
00:00:05
only three people we preempt the show
00:00:06
for. President Trump, Jesus, and Jensen.
00:00:11
And uh I'll let you pick which order we
00:00:13
do that. Uh but what an amazing run
00:00:16
you've had and a great event. Uh
00:00:19
>> every industry is here. Every tech
00:00:21
company is here. Every AI company is
00:00:23
here. Incredible. Incredible. I'm going
00:00:26
all in.
00:00:27
>> If you were building a global financial
00:00:29
system from first principles today, you
00:00:31
wouldn't build it on 50-year-old legacy
00:00:34
rails, you'd build Airwall, one AI
00:00:37
native platform for global accounts,
00:00:39
cards, and payments. It's designed to
00:00:41
make the entire world feel like a local
00:00:44
market. Others are bolting AI onto
00:00:46
broken infrastructure, but AirWall was
00:00:48
built for the intelligent era from day
00:00:50
one. Stop paying the legacy tax and
00:00:52
start building the future at
00:00:54
awallix.comallin.
00:00:56
Arowallix. Build the future.
00:00:59
>> And um one of the great announcements of
00:01:02
the past year has been Grock. When you
00:01:04
made the purchase of Grock, did you
00:01:06
realize how insufferable Cha Chimath
00:01:08
would become?
00:01:10
>> I had I had an inkling that that that
00:01:13
>> we're his friends. We have to deal with
00:01:15
him every week.
00:01:16
>> I know it.
00:01:16
>> You had to deal with him for the six
00:01:18
week close.
00:01:19
>> I know. It's like two weeks. Two weeks.
00:01:21
>> It's all coming back to me now. It's
00:01:22
it's making me rather uncomfortable. The
00:01:25
the thing is uh many of our strategies
00:01:29
are are presented in in broad daylight
00:01:32
at GTC years in advance of when we do
00:01:36
it. Two and a half years ago, I
00:01:38
introduced the operating system of the
00:01:40
AI factory and it's called Dynamo.
00:01:43
Dynamo as you know is a piece of
00:01:46
instrument a machine that was created by
00:01:48
seammens to turn essentially water into
00:01:52
electricity and dynamo uh powered the
00:01:56
factory of the last industrial
00:01:58
revolution. So I thought it was the
00:01:59
perfect name for the operating system of
00:02:02
the next industrial revolution the
00:02:04
factory of that and so inside Dynamo the
00:02:07
fundamental technology is disagregated
00:02:10
inference.
00:02:11
Jason, I I know you're you're you're
00:02:14
super technical.
00:02:14
>> Absolutely. I know it.
00:02:15
>> I'll let you take this one. Go ahead and
00:02:17
define it for the audience. I don't want
00:02:18
to step on you.
00:02:19
>> Yeah. Thank you. I I I knew you wanted
00:02:21
to jump in there for a second, but it's
00:02:22
it's disagregating inference, which
00:02:25
means the the pipeline, the processing
00:02:27
pipeline of inference is extremely
00:02:29
complicated. In fact, it is the most
00:02:31
complicated computing problem today.
00:02:34
Incredible scale, lots of mathematics of
00:02:36
different shapes and sizes. And we came
00:02:39
up came up with the idea that you would
00:02:41
change you would you would disagregate
00:02:43
parts of the processing such that some
00:02:46
of it can run on some GPUs rest of it
00:02:49
can run on different GPUs and that led
00:02:52
to us realizing that maybe even
00:02:55
disagregated computing could make sense
00:02:58
that we could have different
00:02:59
heterogeneous nature of computing that
00:03:02
same sensibility led us to melanox
00:03:05
>> you know today Nvidia's computing is
00:03:08
spread across GPUs, CPUs, switches,
00:03:11
scale up switches, scale out switches,
00:03:13
networking processors, and now we're
00:03:15
going to add Grock to that. And we're
00:03:17
going to put the right workload on the
00:03:19
right chips. You know, we just really
00:03:21
evolved from a GPU company to an AI
00:03:23
factory company.
00:03:24
>> I mean, I think that was probably the
00:03:25
biggest takeaway that I had. You're
00:03:27
seeing this fundamental disagregation
00:03:29
where we've gone from a GPU and now you
00:03:31
have this complexion of all these
00:03:34
different options that will eventually
00:03:36
exist. The thing that you guys said on
00:03:38
stage or you said on stage was I I would
00:03:40
like the high value inference people to
00:03:43
take a listen to this and 25% of your
00:03:46
data center space you said should be
00:03:47
allocated to this gro lpu GPU
00:03:50
>> grock to about 25% of the ver rubins in
00:03:54
the g in the data center. So can you
00:03:56
tell us about how the industry looks at
00:03:58
this idea of now basically creating this
00:04:01
next generation form of disagregated
00:04:03
pre-filled decode DSAG and how people do
00:04:05
you think will react to it?
00:04:06
>> Yeah and take a step back and at the
00:04:09
time that we added this we went from
00:04:12
large language model processing
00:04:16
to agentic processing. Now when you're
00:04:19
running an agent you're accessing
00:04:22
working memory you're accessing
00:04:24
long-term memory. You're using tools.
00:04:26
You're really beating up on storage
00:04:28
really hard. You have agents working
00:04:30
with other agents. Some of the agents
00:04:32
are very large models. Some of them are
00:04:34
smaller models. Some of them are
00:04:36
diffusion models. Some of them are auto
00:04:39
reggressive models. And so there are all
00:04:41
kinds of different types of models
00:04:42
inside this data center. We created Vera
00:04:44
Rubin to be able to run this
00:04:46
extraordinarily diverse workload. My
00:04:49
sense is and so we added what used to be
00:04:52
a one rack company. we now added four
00:04:55
more racks,
00:04:55
>> right?
00:04:56
>> So, Nvidia's TAM, if you will, increased
00:04:59
from what whatever it was to probably
00:05:02
something, call it, you know, 33% 50%
00:05:05
higher. Now, part of that 33% or 50% a
00:05:09
lot of it's going to be storage
00:05:10
processors. It's called Blue Field. Some
00:05:13
of it will be a lot of it, I'm hoping,
00:05:15
will be Grock processors, and some of it
00:05:17
will be CPUs. And they're all and a lot
00:05:21
of it's going to be networking
00:05:22
processors. And so all of this is going
00:05:24
to be running basically the computer of
00:05:28
the AI revolution called agents, right?
00:05:30
>> The operating system of of um modern
00:05:33
modern industry.
00:05:34
>> What about embedded applications? So you
00:05:36
know my daughter's teddy bear at home
00:05:39
wants to talk to her. What goes in
00:05:41
there? Is it a custom ASIC or does there
00:05:43
end up becoming much more kind of a
00:05:45
broader set of TAM with developing tools
00:05:48
that are maybe different for different
00:05:49
use cases at the edge and an embedded
00:05:51
application? We think that there's three
00:05:53
computers in the problem at the at the
00:05:56
largest at the largest scale when you
00:05:58
take take a step back. There's one
00:05:59
computer that's really about training
00:06:01
the AI model, developing creating the
00:06:03
AI, another computer for evaluating it.
00:06:07
Depending on the type of problem you're
00:06:08
having, like for example, you look
00:06:10
around, there's all kinds of robots and
00:06:11
cars and things like that. You have to
00:06:13
evaluate these robots inside a virtual
00:06:18
gym that represents the physical world.
00:06:21
So it has to be software that obeys the
00:06:24
laws of physics.
00:06:25
>> And that's a second computer. We call
00:06:27
that omniverse. The third computer is
00:06:29
the computer at the edge, the robotics
00:06:31
computer.
00:06:32
>> That robotics computer, one of them
00:06:34
could be self-driving car. Another one's
00:06:35
a robot. Another one could be a teddy
00:06:37
bear.
00:06:37
>> Little tiny one for a teddy bear.
00:06:39
>> One of the most important ones is one
00:06:41
that we're working on that basically
00:06:43
turns the telecommunications base
00:06:45
stations into part of the AI
00:06:48
infrastructure. So now all of the it's a
00:06:51
$2 trillion industry all of that in time
00:06:54
will be transformed into an extension of
00:06:57
the AI infrastructure and so radios
00:07:00
radios will become a edge devices
00:07:02
factories warehouses you name it and so
00:07:05
so there are three these three basic
00:07:07
computers
00:07:08
all of them you know are going to be
00:07:10
necessary
00:07:11
>> Jensen last uh last year I think you
00:07:14
were ahead of the the rest of the world
00:07:15
in in in saying inference isn't going to
00:07:17
a thousand last year.
00:07:19
>> Yes. Is it going to hurt my feelings? Is
00:07:21
it is it going to 1 millionx? Is it
00:07:23
going to 1 billionx? Yeah.
00:07:24
>> Right. And I think people at the time
00:07:26
thought it was pretty hyperbolic because
00:07:28
the world was still focused on
00:07:29
pre-scaling on training. Here we are
00:07:32
now. Inference has exploded. We're
00:07:34
inference constrained. Um you announced
00:07:36
an inference factory that I think is
00:07:38
leading edge that's going to be 10x
00:07:40
better in terms of throughput to the
00:07:42
next factory. But yet if you if I listen
00:07:45
to what the chatter is out there, it's
00:07:47
that your inference factory is going to
00:07:49
cost 40 or 50 billion and the
00:07:51
alternatives the custom AS6 AMD others
00:07:54
are going to cost 25 to 30 billion and
00:07:56
you're going to lose share. So why don't
00:07:58
you talk to us? What are you seeing? How
00:07:59
do you think about share and does it
00:08:02
make sense for all these folks to pay
00:08:04
something that's a 2x premium to what
00:08:06
others are marketing? The big takeaway,
00:08:09
the big idea is that
00:08:13
you should not equate the price of the
00:08:17
factory and the price of the tokens, the
00:08:20
cost of the tokens. It is very likely
00:08:23
that the $50 billion factory, and in
00:08:26
fact, I can prove it that the $50
00:08:28
billion factory will generate for you
00:08:31
the lowest cost tokens. And the reason
00:08:34
for that is because we produce these
00:08:36
tokens at extraordinary efficiency
00:08:40
10 times you know the difference between
00:08:42
50 billion now it turns out 20 billion
00:08:45
is just land power and shell right
00:08:47
>> and then on top of that you have storage
00:08:50
anyways networking anyways you got CPUs
00:08:52
anyways you got servers anyways you got
00:08:54
cooling anyways the difference between
00:08:56
that GPU being 1x price or halfx price
00:09:00
>> is not between 50 billion and 30 billion
00:09:03
Pick your favorite number, but let's say
00:09:05
between 50 billion and 40 billion.
00:09:07
>> That is not a large percentage when the
00:09:10
$50 billion data center is actually 10
00:09:13
times the throughput.
00:09:14
>> Right, Jess?
00:09:16
>> That's the reason why I said that even
00:09:18
for most chips, if you can't keep up
00:09:21
with the state of the technology and the
00:09:22
pace that we're running, even when the
00:09:24
chips are free, it's not cheap enough.
00:09:26
>> Yeah.
00:09:27
>> Can I can I just ask a general strategy
00:09:29
question?
00:09:29
>> Yeah. I mean, you're running the most
00:09:32
valuable company in the world. This
00:09:33
thing is going to do 350 plus billion of
00:09:36
revenue next year. 200 billion of free
00:09:39
cash flow. It's compounding at these
00:09:40
crazy rates. How do you decide what to
00:09:43
do? Like, how do you actually get the
00:09:45
information? I mean, it's famous now
00:09:47
these sort of emails that are people are
00:09:49
meant to send you, but how do you really
00:09:51
decide to get an intuition of how to
00:09:54
shape the market, where to really double
00:09:55
down, where to maybe pull back, where to
00:09:57
actually go into a green field? How how
00:09:59
does that information get to you? How do
00:10:00
you decide these things?
00:10:01
>> In a final analysis, that's the job of
00:10:03
the CEO. Yeah.
00:10:04
>> And our job is to define the strategy,
00:10:06
define the vision, define the strategy.
00:10:08
We're informed, of course, by amazing
00:10:10
computer scientists, amazing
00:10:11
technologists, great people all over the
00:10:13
company, but we have to shape that
00:10:15
future. Well, part of it has to do with
00:10:17
is this something that's insanely hard
00:10:19
to do? If it's not hard to do, we should
00:10:22
back away from it. And the reason for
00:10:23
that if if it's easy to do obviously um
00:10:26
lots of competitors a lot of competitors
00:10:28
is this something that has never been
00:10:30
done before that's insanely hard to do
00:10:33
and that somehow taps into the special
00:10:36
superpowers of our company and so I have
00:10:38
to find this confluence of things to
00:10:40
that meets the standard and in the end
00:10:44
we also know that a lot of pain and
00:10:46
suffering is going to go into it. Yeah,
00:10:47
>> there no great things that are invented
00:10:50
because it was just easy to do and just
00:10:51
like first try here we are.
00:10:53
>> And so if it's super hard to do,
00:10:55
nobody's ever done it before, it's very
00:10:57
likely that you're going to have a lot
00:10:58
of pain and suffering and so you better
00:11:00
enjoy it.
00:11:00
>> So can you can you just look at maybe
00:11:02
three or four of the more longtail
00:11:04
things you announced
00:11:05
>> and just talk about the long-term
00:11:07
viability of whether it's the data
00:11:09
centers in space or whether it's what
00:11:10
you're trying to do with ADAS in autos
00:11:12
or you know what you're trying to do on
00:11:14
the biology side. just give us a sense
00:11:16
of like how you see some of these curves
00:11:18
inflecting upwards in some of these
00:11:19
longer tail businesses.
00:11:20
>> Excellent. U physical AI large category
00:11:24
we believe and I just mentioned we have
00:11:26
three computing systems all the software
00:11:28
platforms on top of it. physical AI as a
00:11:31
large category.
00:11:34
It's technology industry's first
00:11:36
opportunity
00:11:38
to address a $50 trillion industry that
00:11:41
has largely been, you know, void of
00:11:44
technology until now. And so, we need to
00:11:46
invent all of the technology necessary
00:11:48
to do that. I felt that that was a
00:11:49
10-year journey. We started 10 years
00:11:52
ago. We're seeing it inflecting now. It
00:11:54
is a multi-billion dollar business for
00:11:56
us. It's close to$10 billion a year now.
00:11:58
And so it's a big business and it's
00:12:00
growing exponentially. And so that's
00:12:02
number one. I think in the case of
00:12:04
digital biology, I think we are
00:12:05
literally near the chat GPT moment of
00:12:08
digital biology. We're about to
00:12:09
understand how to represent genes,
00:12:12
proteins, cells. We already know how to
00:12:15
understand chemicals. And so the ability
00:12:17
for us to represent and understand the
00:12:20
dynamics of the building blocks of
00:12:22
biology, that's a couple of two, three,
00:12:25
five years from now. In five years time,
00:12:28
I completely believe that the healthcare
00:12:29
industry where digital biology is going
00:12:31
to inflect. And so these are a couple of
00:12:33
the really great ones and you could see
00:12:35
they're all around us.
00:12:36
>> Agriculture,
00:12:37
>> agriculture
00:12:38
>> reflecting now.
00:12:39
>> No question. Yeah.
00:12:40
>> Hson, I want to take you from the data
00:12:42
center to the desktop. Uh the company
00:12:45
was built in large part on hobbyists,
00:12:48
video gamers, and and all those graphic
00:12:50
cards in the beginning. And you
00:12:52
mentioned in front of I think 10,000
00:12:54
people here just clawed open claw clawed
00:12:58
code and what a revolution agents have
00:13:00
become and specifically
00:13:03
the hobbyists who are really where a lot
00:13:05
of energy um we see you know a lot of
00:13:08
the innovation breaks want desktops you
00:13:10
announced one here uh I believe it's the
00:13:13
Dell 6800 uh this is a very powerful
00:13:16
workstation to run local models 750 gigs
00:13:18
of RAM obviously the the Mac uh studio
00:13:22
sold out everywhere in my company. We're
00:13:24
moving to openclaw everything. Freeberg
00:13:27
just got clawdelled. You got claw peld,
00:13:29
I understand. And you're obsessed with
00:13:30
these.
00:13:31
>> What is this from the streets movement
00:13:34
of creating open-source
00:13:37
agents and using open source on the
00:13:39
desktop mean to you? Great. Where is
00:13:40
that going?
00:13:41
>> Yeah. So great. First of all, let's take
00:13:42
a step back. Um in the last two years,
00:13:45
we saw basically three inflection
00:13:47
points. The first one was generative
00:13:50
chat GPT
00:13:53
brought AI to the common
00:13:56
everybody to our awareness. But the fact
00:13:58
of the matter is the technology sat in
00:14:00
plain sight months before GPT. It wasn't
00:14:03
until chat GPT put a user interface
00:14:07
around it made it easy for us to use
00:14:09
that generative AI took off. Now
00:14:11
generative AI as you know generates
00:14:14
tokens for internal consumption as well
00:14:16
as external consumption. internal
00:14:17
consumption is thinking which led to
00:14:20
reasoning. 01 and 03
00:14:23
continue that wave of chat GBT grounded
00:14:26
information made AI not only answer
00:14:28
questions but answer questions in a more
00:14:30
grounded way useful. We started seeing
00:14:33
the revenues and the e the economic
00:14:36
model of open AI start to inflect. Then
00:14:39
the third one was only inside the
00:14:42
industry that we saw clock code the
00:14:44
first agentic system that was very
00:14:46
useful really revolutionary stuff but
00:14:50
but cloud code was only available for
00:14:52
enterprises. Most people outside never
00:14:55
saw anything about cloud code until open
00:14:58
claw. Open claw basically put into the
00:15:02
po popular consciousness what an AI
00:15:05
agent can do. Mhm.
00:15:07
>> That's the reason why open claw is so
00:15:09
important from a cultural perspective.
00:15:11
Now the second second reason why it's so
00:15:14
important is that open claw is open but
00:15:17
it formulates
00:15:20
it structures a type of computing model
00:15:24
that is basically reinventing computing
00:15:26
all together. It has a memory system. It
00:15:30
scratch is a short-term memory file
00:15:31
system. It has it has it has scales.
00:15:34
Yeah. Did you say skills or scales?
00:15:36
>> Skills.
00:15:37
>> Oh, skills.
00:15:37
>> They do have scales theoretically. Yeah.
00:15:39
Yeah. Skills.
00:15:40
>> So, the first thing first thing it it,
00:15:42
you know, it has resources. It it
00:15:44
manages resources. It's it does
00:15:46
scheduling.
00:15:47
>> Yep.
00:15:47
>> Right. And it cron jobs. It could it
00:15:50
could spawn off agents. It could, you
00:15:52
know, it could decompose a task and and
00:15:54
cause and solve problems as does
00:15:56
scheduling. It has IO subsystems. It
00:15:59
could, you know, input. It has outputed
00:16:01
connect to WhatsApp. And also it has a
00:16:04
API that allows it to run multiple types
00:16:07
of applications called skills.
00:16:09
>> Yeah.
00:16:10
>> These four elements fundamentally define
00:16:13
a computer.
00:16:13
>> Yeah.
00:16:14
>> And therefore what do we have? We have a
00:16:17
personal artificial intelligence
00:16:21
>> computer for the very first time.
00:16:23
>> Open source.
00:16:24
>> It's open source. It runs literally
00:16:26
everywhere. And so this is now the this
00:16:28
is the op this is basically the
00:16:30
blueprint the operating system of modern
00:16:33
computing.
00:16:33
>> Yeah.
00:16:34
>> And it's going to run literally
00:16:35
everywhere. Now of course one of the
00:16:37
things that we had to help it do is
00:16:38
whenever you have agentic software you
00:16:41
have to make sure that and agentic
00:16:42
software has access to sensitive
00:16:44
information. It could execute code. It
00:16:46
could communicate externally. We have to
00:16:49
make sure that all of it has to be
00:16:50
governed. all of it has to be secure and
00:16:52
that we have policies that that gives
00:16:55
these agents two of the three things but
00:16:58
not all three things at the same time
00:16:59
and so the governance part of it we
00:17:02
contributed to Peter Peter Steinberger
00:17:04
was here and and so we've got a mountain
00:17:06
of great engineers working with him to
00:17:07
help secure and keep that thing so that
00:17:10
it could protect our privacy protect our
00:17:12
security
00:17:12
>> Jensen that paradigm shift makes some of
00:17:15
the AI legislation that has passed
00:17:17
around the country to regulate AI and a
00:17:20
lot of the proposed legislation
00:17:21
effectively moot, doesn't it? Can you
00:17:23
just comment for a second on how quickly
00:17:25
the paradigm shift kind of obiates a lot
00:17:27
of the models for regulatory oversight
00:17:30
of AI, which is becoming a very hot
00:17:32
topic in politics right now.
00:17:34
>> Well, this is this is the part that that
00:17:36
we just with policy makers, we need to
00:17:39
we need to always get in front of them
00:17:41
and Brad, you do a great job doing this.
00:17:42
We had to get in front of them and
00:17:43
inform them about the state of the
00:17:45
technology, what it is, what it is not.
00:17:48
It is not a biological being. It is not
00:17:53
alien. It is not conscious.
00:17:57
Um it is computer software.
00:18:00
>> Yeah. Exactly.
00:18:01
>> And and it is not something that um we
00:18:04
say things like we don't understand it
00:18:06
at all.
00:18:07
>> It is not true. We don't understand at
00:18:09
all. We understand a lot of things about
00:18:10
this technology. and and so so I think
00:18:13
one we have to make sure that we
00:18:14
continue to inform the policy makers and
00:18:17
not affect not allow doomerism and
00:18:20
extremism to affect how policy makers
00:18:24
think and understand about this
00:18:25
technology. However, however, we still
00:18:27
have to recognize the technology is
00:18:28
moving really fast and don't get policy
00:18:30
ahead of the technology too quickly and
00:18:33
the risk that we we run as a nation. Our
00:18:37
greatest source of national security
00:18:39
concern with respect to AI is that other
00:18:41
countries adopt this technology while we
00:18:44
are so angry at it or afraid of it or
00:18:48
somehow paranoid of it that our
00:18:51
industries our society don't take
00:18:53
advantage of AI and so I'm just mostly
00:18:55
worried about the diffusion of AI here
00:18:57
in the United States.
00:18:58
>> Can you just double click if you were in
00:18:59
the seat in the boardroom of Anthropic
00:19:02
over that whole scuttlebutt with the
00:19:04
department of war? It sort of builds on
00:19:06
this idea of people didn't know what to
00:19:09
think. It's sort of added to this layer
00:19:11
of either resentment or fear or just
00:19:14
general mistrust that people have
00:19:16
sometimes at the software levels of AI.
00:19:19
What would do you think you would have
00:19:20
told Daario and that team to do maybe
00:19:22
differently to try to change some of
00:19:24
this outcome and some of this
00:19:25
perception?
00:19:25
>> The first thing that I I would I would
00:19:27
say about anthropic is first of all the
00:19:29
technology is incredible. We are a large
00:19:31
consumer of anthropic technology. really
00:19:33
admire their focus on security, really
00:19:36
admires their focus on safety. Um the
00:19:38
the the the culture by which we they
00:19:42
went about it, the the technology
00:19:43
excellence by which they went about it
00:19:45
really fantastic. Um I I would say that
00:19:48
that the the desire to warn people about
00:19:53
the capability, the technology is is
00:19:55
also uh really terrific. We just have to
00:19:57
make sure that we understand that the
00:19:59
world has a spectrum and that that
00:20:02
warning is good, scaring is less good,
00:20:05
>> right?
00:20:06
>> Um and because this technology is too
00:20:08
important to us,
00:20:09
>> right?
00:20:10
>> And and I think that it is fine to uh
00:20:15
predict the future, but we need to be a
00:20:17
little bit more circumspect. We need to
00:20:19
have a little bit more humility that in
00:20:21
fact we can't completely predict the
00:20:24
future and the abil and to say things
00:20:26
that that are quite extreme, quite
00:20:29
catastrophic that there's no evidence of
00:20:32
it happening um could be more damaging
00:20:35
than people think. And and of course we
00:20:37
are technology leaders.
00:20:40
>> There were there was a time when nobody
00:20:41
listened to us. Yeah. Um but now because
00:20:44
technology is so important in the social
00:20:47
fabric, such an important industry, so
00:20:50
important to national security, our
00:20:52
words do matter and I think we have to
00:20:54
be much more circumspect. We have to be
00:20:55
more moderate. We have to be more
00:20:57
balanced. We have to be more for more
00:20:58
thoughtful.
00:20:59
>> Well, I you know, I would nominate you.
00:21:02
I think the industry's got to get
00:21:03
together. 17% popularity of AI in the
00:21:07
United States. I mean, we see what
00:21:08
happened to nuclear, right? We basically
00:21:11
shut down the entire nuclear industry
00:21:12
and now we have a 100 fision reactors
00:21:14
being built in China and zero in the
00:21:16
United States. Um we hear about
00:21:18
moratoriums on data centers. So I think
00:21:20
we have to be a lot more proactive about
00:21:21
that. But but I want to go back to this
00:21:23
agentic explosion that you're seeing
00:21:25
inside your company, the efficiencies,
00:21:27
the productivity gains inside your
00:21:29
company. There's a lot of debate whether
00:21:31
or not we're seeing ROI, right? And you
00:21:33
and I entering into this year, the big
00:21:36
question was are the revenues going to
00:21:38
show up? are the revenues going to scale
00:21:40
like intelligence and then we had this
00:21:42
kind of Oenheimer moment a five6 billion
00:21:45
month by anthropic in February. Um do
00:21:49
you think as you look ahead you
00:21:51
announced a trillion dollar you know
00:21:53
visibility into a trillion dollars of
00:21:54
just Blackwell and Vera Rubin over the
00:21:56
course of the next couple years when you
00:21:58
see this happening at Anthropic and
00:22:00
OpenAI do you think we're on that curve
00:22:02
now where we're going to see revenues
00:22:04
scale in the way that intelligence is
00:22:06
scaling
00:22:07
>> when you look around when you I'll
00:22:08
answer this a couple different ways when
00:22:09
you look around this audience you will
00:22:11
see that anthropic and open AAI is
00:22:13
represented here but in fact everybody
00:22:16
99% of Everything that is here is all AI
00:22:19
and it's not anthropic and open AI.
00:22:21
>> Right. Right.
00:22:21
>> And the reason for that is because AI is
00:22:24
very diverse.
00:22:26
>> I would say that the second most popular
00:22:29
model as a category is open models.
00:22:32
>> Number one is yeah open open source open
00:22:35
ways open source.
00:22:36
>> Open AI is number one. Open source is
00:22:39
number two. Very distant third is
00:22:40
anthropic. And that tells you something
00:22:42
about the scale of all of the AI
00:22:45
companies that are here. And so so it's
00:22:48
important to recognize recognize that um
00:22:51
let me let me come back and say a couple
00:22:52
things. One when we went from generative
00:22:55
to reasoning the amount of computation
00:22:58
we needed was about a hundred times.
00:23:00
>> When we went from reasoning to agentic
00:23:03
the computation is probably another 100
00:23:06
times. Now we're looking at in just two
00:23:10
years computation went up by a fact
00:23:13
10,000x.
00:23:15
Meanwhile,
00:23:17
people pay for information, but people
00:23:21
mostly pay for work.
00:23:23
>> Yes,
00:23:25
>> talking to a chatbot and getting an
00:23:26
answer is super great,
00:23:28
>> right?
00:23:29
>> Helping me do some research,
00:23:30
unbelievable, but getting work done,
00:23:32
I'll pay fordeed.
00:23:34
>> And so that's where we are.
00:23:36
>> Agentic systems get work done. They're
00:23:38
helping our software engineers get work
00:23:39
done. And and so then you take that you
00:23:43
got 10,000x more compute you get
00:23:46
probably at this point 100x more
00:23:48
consumption now.
00:23:49
>> Yes.
00:23:49
>> Yeah.
00:23:50
>> And we haven't even started scaling yet.
00:23:52
>> We are absolutely at a millionx
00:23:54
>> which is I think a great place to talk
00:23:56
about the number of people have 20
00:23:59
30,000 at the company something.
00:24:01
>> We have 43,000 employees. You know I
00:24:03
would say 38,000
00:24:05
are engineers. The conversation we've
00:24:08
had on the pod a number of times is, "Oh
00:24:10
my god, look at the token usage in our
00:24:12
companies. It is growing massively." And
00:24:15
some people are asking, "Hey, when I
00:24:17
join a company, how many tokens do I get
00:24:19
cuz I want to be an effective employee?"
00:24:21
And you postulated, I believe, during
00:24:23
your 2 and 1/2 hour keynote, pretty long
00:24:26
keynote, well done, that you were
00:24:30
spending,
00:24:30
>> if it was well done, it would be
00:24:31
shorter. Yeah, he didn't have time to do
00:24:33
a time to write an hour.
00:24:36
>> So you guys So you guys know So you guys
00:24:38
know there is no practice
00:24:40
>> and so it's a gripping and ripping
00:24:41
>> rip and rip. Yeah.
00:24:43
>> So So I just want to let you know I was
00:24:44
writing the speech while I was giving
00:24:46
the speech. Okay. So
00:24:48
>> you never know.
00:24:49
>> But does that mean if we do back
00:24:51
>> I apologize.
00:24:52
>> Back of the envelope math 75,000 in
00:24:55
tokens for each engineer or something
00:24:56
like that. So, are you spending in
00:24:58
Nvidia a billion2 billion dollars on
00:25:00
tokens for your engineering team right
00:25:02
now?
00:25:02
>> We're trying to Let me give you a
00:25:03
thought experiment. Let's say you have a
00:25:05
software engineer or AI researcher and
00:25:07
you pay them $500,000 a year. We do that
00:25:10
all the time.
00:25:11
>> Okay, this is happening all over the
00:25:13
time. Um, that $500,000 engineer at the
00:25:16
end of the year, I'm going to ask them
00:25:18
how many token how much did you spend in
00:25:20
tokens? If that person said $5,000, I
00:25:23
will go ape something else.
00:25:24
>> Yes.
00:25:25
>> Right. If that if that $500,000 engineer
00:25:28
did not consume at least $250,000 worth
00:25:31
of tokens, I am going to be deeply
00:25:34
alarmed. Okay? And this is no different
00:25:37
than one of our chip designers who says,
00:25:40
"Guess what? I'm just going to use paper
00:25:42
and pencil. I don't think I'm going to
00:25:44
need any CAD tools."
00:25:45
>> This is a real paradigm shift to start
00:25:47
thinking about these all-star employees.
00:25:50
It almost reminds me of of what we
00:25:51
learned in the NBA when LeBron James
00:25:53
started spending a million dollars a
00:25:55
year just on his health of his body like
00:25:57
and maintaining it. That's right.
00:25:58
>> Here he is at age 41 still playing. It
00:26:01
really is, hey, if these are incredible
00:26:03
knowledge workers, why wouldn't we give
00:26:05
them
00:26:06
>> superhuman abilities?
00:26:08
>> That's exactly right.
00:26:08
>> Where does that go? If we if we
00:26:10
extrapolate out two or three years from
00:26:12
now, what is the efficiency of that
00:26:15
allstar at an Nvidia and what they're
00:26:16
able to accomplish? What do they look
00:26:18
like? Well, first of all, things that
00:26:20
that that um wow, this is too hard. That
00:26:24
thought is gone. Uh this is going to
00:26:26
take a long time. That thought is gone.
00:26:28
Uh we're going to need a lot of people.
00:26:29
That thought is gone. This is no
00:26:32
different than in this in the last
00:26:33
industrial re revolution. Somebody goes,
00:26:36
"Boy, that building really looks heavy."
00:26:38
Nobody says that. Nobody, wow, that
00:26:40
mountain looks too big. Nobody says
00:26:42
that. Right.
00:26:43
>> Everything that's too big, too heavy,
00:26:45
takes too long,
00:26:46
>> those thought, those ideas are all gone.
00:26:48
>> You're reduced to creativity.
00:26:50
>> That's right.
00:26:50
>> What can you come up with?
00:26:51
>> Exactly. Which means now the question is
00:26:53
how do you how do you work with these
00:26:55
agents? Well, it's just a new way of
00:26:57
doing computer programming. In the f in
00:26:59
the past, we code. In the future, we're
00:27:01
going we're going to write ideas,
00:27:03
architectures, specifications.
00:27:06
We're going to organize teams. We're
00:27:08
going to give him, we're going to help
00:27:09
them define how to evaluate the
00:27:11
definition of good versus bad. What's
00:27:14
the what does it look like when
00:27:15
something is a great outcome, how to
00:27:17
iterate with you, how to brainstorm.
00:27:19
That's really what you're looking for.
00:27:21
And I'm I think that every engineer is
00:27:23
going to have hundred hundred agents.
00:27:25
>> Back to the PR problem the industry has
00:27:28
right now. You have executives uh like
00:27:31
David Freeberg with Oho who's looking at
00:27:35
literally taking through the use of
00:27:36
technology your technology and AI the
00:27:39
number of calories produced and making
00:27:42
high quality cal calories what is the
00:27:44
factor you think you can bring the cost
00:27:47
down Freeberg and what impact does this
00:27:49
vision have for what you're doing
00:27:51
>> zero shot genomic modeling and it works
00:27:54
>> and then you have that moment and you're
00:27:56
like holy
00:27:58
>> honestly like and and that's after
00:28:01
people are replacing entire enterprise
00:28:03
software stacks in a night. I did
00:28:05
something in 90 minutes. I was telling
00:28:06
the guys about replaced the whole
00:28:08
software stack and like a whole bunch of
00:28:10
workload. 90 minutes on claude, ran this
00:28:12
agentic system, built the whole thing,
00:28:13
deployed it, and we got we were on a
00:28:15
Sunday night.
00:28:15
>> On a Sunday night, 10 p.m. I was done at
00:28:17
11:30. I went to bed.
00:28:18
>> As the CEO, you replaced
00:28:20
>> Yeah. And everyone on my management team
00:28:22
had to do a similar exercise over the
00:28:24
weekend. What we saw on Monday, I was
00:28:26
like, it's over. But the technical
00:28:30
stuff, the science stuff, we did
00:28:32
something in 30 minutes using auto
00:28:33
research. And I'd love your view on auto
00:28:34
research and what that tells us about
00:28:36
how far we still have to go in terms of
00:28:38
efficiency. But using auto research and
00:28:40
a chunk of data, something was published
00:28:43
internally that we said, "Oh my god."
00:28:46
And that would normally be a PhD thesis
00:28:47
that would take seven years. It would be
00:28:49
one of the most celebrated PhD thesis
00:28:50
we've ever seen in this field and it
00:28:52
would be in the journal science. And it
00:28:54
was done in 30 minutes on a desktop
00:28:55
computer running on auto research with
00:28:58
all the data we just ingested. We got it
00:28:59
on Friday. We're like, "Hey, let's try
00:29:00
it." try booted up, went to GitHub,
00:29:02
downloaded auto research and ran it and
00:29:05
you see everyone's face just go like and
00:29:08
then the potential of what this is
00:29:09
unlocking for us is like the kind of
00:29:11
thing that would take seven years and it
00:29:13
happened in 30 minutes and we're
00:29:15
experiencing it in genomics and we're
00:29:17
like this is unbelievable. So I I think
00:29:19
like the acceleration is widening the
00:29:22
aperture for everyone in a way that like
00:29:25
you didn't imagine a few years ago. But
00:29:27
just going back to the auto research
00:29:28
point, can you just comment on what you
00:29:30
think about the fact that this thing got
00:29:32
published with 600 lines of code in a
00:29:34
weekend and the capacity that it has to
00:29:36
run locally and achieve what it can
00:29:38
achieve with all of these diverse data
00:29:40
sets and what that tells us about the
00:29:42
early stages we are in terms of
00:29:43
optimization on algorithms and hardware.
00:29:46
The fundamental reason why open claw is
00:29:48
so incredible number one is its com its
00:29:52
confluence its timing with the
00:29:55
breakthroughs in large language model.
00:29:57
>> Yeah.
00:29:57
>> Its timing was perfect. It was
00:29:59
impeccable. Now in a lot of ways Peter
00:30:01
wouldn't have come up with it probably
00:30:03
if not for the fact that claude and GPT
00:30:06
and chat GPT have reached a level that
00:30:08
is really very good.
00:30:10
>> Right. It is also a new capability that
00:30:13
allows these models to tool use the
00:30:16
tools that we've created over time web
00:30:19
browsers and Excel spreadsheets and you
00:30:22
know in the case of chip design synopsis
00:30:25
and cadence and uh omniverse and blender
00:30:29
and autodesk and all of these tools are
00:30:31
going to continue to be used. There's
00:30:33
some some people say that that the
00:30:35
enterprise IT software industry is going
00:30:38
to get destroyed. There's it's there's a
00:30:41
let me give you the alternative view.
00:30:43
The enterprise software industry is
00:30:45
limited by butts and seats. It's about
00:30:48
to get a hundred times more agents
00:30:50
banging on those tools. They're going to
00:30:52
be agents banging on SQL. They're going
00:30:53
to be agents bang on vector databases,
00:30:55
agents bang on Blender, agents banging
00:30:57
on Photoshop. And the reason for that is
00:30:59
because those tools are first of all do
00:31:02
a very good job. Second, those tools are
00:31:05
the conduit between us in the final
00:31:08
analysis. When the work is done, it has
00:31:10
to be represented back to me in a way
00:31:12
that I can control,
00:31:13
>> right?
00:31:14
>> And I know how to control those tools.
00:31:17
And so I need everything to be put back
00:31:19
in the synopsis. I want everything to
00:31:20
put back in the cadence because that's
00:31:22
how I control it. That's how I've ground
00:31:24
truth.
00:31:24
>> Let me ask you a question about open
00:31:26
source. So we have these closed source
00:31:28
models. They're excellent.
00:31:29
>> We have these openweight models. Many of
00:31:31
the Chinese models are incredible.
00:31:34
Absolutely incredible. Two days ago, you
00:31:36
may not have seen this because you were
00:31:37
busy on stage, but there was a training
00:31:40
run that happened in this crypto project
00:31:42
called Bit Tensor Subnet 3. They managed
00:31:45
to train a 4 billion parameter llama
00:31:47
model, totally distributed with a bunch
00:31:50
of people contributing
00:31:52
excess compute, but they were able to do
00:31:54
it statefully and manage a training run,
00:31:56
which I thought was like a pretty crazy
00:31:59
technical accomplishment. Yeah,
00:32:00
>> because it's like random people and each
00:32:02
person gets a little share.
00:32:04
>> Our our modern version of folding at
00:32:05
home.
00:32:06
>> Exactly. So what what do you think about
00:32:08
the end state of open source? Do you see
00:32:11
this decentralization of architecture as
00:32:13
well and decentralization of compute to
00:32:16
support open weights and a totally open-
00:32:19
source approach to making sure AI is
00:32:21
broadly available to everyone? I believe
00:32:23
we fundamentally need
00:32:26
models as a first class product,
00:32:28
proprietary product as well as models as
00:32:32
open source. These two things are not A
00:32:35
or B. It's A and B. There's no question
00:32:37
about it. And the reason for that is
00:32:39
because models is a technology, not a
00:32:42
product. Model is a technology, not a
00:32:44
service. For the vast majority of
00:32:46
consumers, the horizontal layer, the
00:32:48
general intelligence, I would really
00:32:51
really love not to go fine-tune my own,
00:32:54
right? I would really love to keep using
00:32:55
chat GPT. I'd love to use cloud. I love
00:32:58
to use Gemini. I love to you use X. And
00:33:00
they all have their own personalities as
00:33:02
you know, which is kind of depends on my
00:33:04
mood and depends on what problem I'm
00:33:05
trying to solve. You know, I might, you
00:33:07
know, do it on X or I might do it on on
00:33:08
chat GBT. And so that that segment of
00:33:11
the of the industry is thriving. is
00:33:13
going to be great. However, there all
00:33:17
these industries their domain expertise
00:33:20
their specialization has to be channeled
00:33:22
has to be captured in a way that they
00:33:25
can control and that it can only come
00:33:27
from open models. The open model
00:33:29
industry we're contributing tremendously
00:33:31
to it is near the frontier
00:33:35
and quite frankly even if it reaches the
00:33:38
frontier
00:33:40
I think that products as a service
00:33:42
worldass products as a models as a
00:33:45
product is going to continue to thrive.
00:33:46
>> Every startup we're investing in now is
00:33:50
open source first and then going to the
00:33:52
proprietary models.
00:33:53
>> Yeah. The beautiful thing is because you
00:33:55
have a great router you connect it to by
00:33:58
on on first day every single day you're
00:34:01
going to have access to the world's best
00:34:02
model and and then it gives you time to
00:34:06
cost reduce and fine-tune and specialize
00:34:08
and so you're going to have worldass
00:34:09
capabilities out to shoot every single
00:34:11
time. Let
00:34:12
>> J can I
00:34:13
>> ask a question?
00:34:14
>> Nobody wants the US to win the global AI
00:34:16
race more than you, right? But a year
00:34:20
ago, the Biden era diffusion rule really
00:34:23
was an anti- American diffusion of AI
00:34:26
around the world. So here we are a year
00:34:28
into the new administration.
00:34:30
Give us a grade. Where is where are we
00:34:33
in terms of global diffusion and the
00:34:35
rate at which we're spreading US AI
00:34:38
technology around the world? Are we an
00:34:40
A? Are we a B? or we see what what's
00:34:42
working, what's not working.
00:34:44
>> Well, first of all, President Trump
00:34:46
wants American industry to lead. He
00:34:50
wants American technology industry to
00:34:51
lead. He wants American technology
00:34:53
industry to win. He wants us to spread
00:34:56
American technology around the world. He
00:34:58
wants United States to be the wealthiest
00:35:00
country in the world. He wants all of
00:35:02
that. At the current moment, as we
00:35:05
speak,
00:35:07
Nvidia gave up a 95% market share in the
00:35:10
second largest market in the world, and
00:35:13
we're at 0%.
00:35:15
>> President Trump, That's right. President
00:35:16
Trump wants us to get back in there. And
00:35:19
and uh the first thing is uh to get
00:35:23
licens licensed for the companies that
00:35:25
we're going to be able to sell to. We've
00:35:27
got many companies who have requested
00:35:30
for licenses. We've applied for licenses
00:35:32
for them and we've got approved licenses
00:35:34
from sec secretary lutnik. Uh now uh
00:35:38
we've we informed the Chinese companies
00:35:40
and many of them have given us purchase
00:35:41
orders and so we're going to we're going
00:35:43
to we're in the process of cranking up
00:35:45
our supply chain again to go ship. I
00:35:48
think at the highest level Brad um I
00:35:50
think one of the things that we should
00:35:51
acknowledge is this. Our national
00:35:54
security
00:35:55
is diminished when we don't have access
00:36:00
to miniature motors, rare earth
00:36:03
minerals. It's diminished when we don't
00:36:05
control our telecommunications networks.
00:36:07
It's diminished when we can't provide
00:36:10
for sustainable energy for our country.
00:36:12
It is fundamentally diminished. Every
00:36:14
single one of these industries is an
00:36:16
example of what I don't want the AI
00:36:19
industry to be.
00:36:20
>> Right? When we look forward in time and
00:36:23
we say what do we want? What is the what
00:36:25
does it look like when American
00:36:28
technology industry American AI industry
00:36:30
leads the world? We can all acknowledge
00:36:33
that there is no way that AI models is
00:36:37
one universally. It is we can all
00:36:40
acknowledge that that is an outcome that
00:36:42
makes no sense. However, we can all
00:36:45
imagine that the American tech stack
00:36:48
from chips to computing systems to the
00:36:52
platforms are used broadly by the world
00:36:56
where they build their own AI, they use
00:36:58
public AI, they use private AI whatever
00:37:01
and they can build their applications in
00:37:02
their society. I would love that the
00:37:05
American tech stack is 90% of the world.
00:37:07
Yes, I would love that. The alternative
00:37:10
if it looks like solar, rare earth,
00:37:14
magnets, motors, telecommunications, I
00:37:17
consider that a very bad outcome for
00:37:20
national security.
00:37:21
>> Agreed.
00:37:21
>> Yeah.
00:37:22
>> How much are you monitoring the
00:37:23
situation with the conflicts around the
00:37:26
world right now? And how much does it
00:37:27
worry you Jensen? So, China and Taiwan
00:37:30
and then helium availability coming out
00:37:32
of the Middle East, I understand, can be
00:37:34
a supply chain risk to semiconductor
00:37:35
manufacturing. How much do these
00:37:37
situations worry you? How much are you
00:37:39
spending on them?
00:37:40
>> Well, first of all, I think the in
00:37:42
Middle East, I have we have 6,000
00:37:44
families there.
00:37:45
>> Yeah,
00:37:45
>> we have a lot of Iranians uh at NVIDIA
00:37:48
and their families are still in Iran.
00:37:50
And so so we have we have a lot of
00:37:52
families there. The first thing is is
00:37:53
they're quite anxious. They're quite
00:37:55
concerned, quite scared. Um we're
00:37:57
thinking about them all the time. Uh
00:37:58
we're monitor and keeping an eye on them
00:38:00
all the time. They have 100% of our
00:38:02
support. Uh I've been asked several
00:38:04
times, are we still considering uh being
00:38:06
in Israel? We are 100% in Israel. We are
00:38:10
100% behind the families there. We are
00:38:12
100% in the Middle East. I was also
00:38:15
asked, you know, given what's happening
00:38:17
in the Middle East, uh is that an area
00:38:20
where we believe that we can expand
00:38:22
artificial intelligence to? Um I believe
00:38:25
that there's a reason we went to war and
00:38:27
I believe at the end of the war, Middle
00:38:29
East will be more stable than before.
00:38:32
And so if we were there, if we're
00:38:34
considering it before, we should
00:38:35
absolutely be considering it after. And
00:38:37
so I'm 100% in on that. With respect to
00:38:40
with with with respect to to Taiwan,
00:38:43
>> we have to do three things. One, we have
00:38:46
to make sure that we re-industrialize
00:38:48
the United States as fast as we can.
00:38:50
Yeah.
00:38:50
>> And whether it's the chip manufacturing
00:38:52
plants, the the computer manufacturing
00:38:54
plants, or the AI factors.
00:38:55
>> How are we doing on that? We're doing
00:38:57
excellent with by by gaining the
00:39:01
strategic support by gaining the
00:39:03
friendship of the supply chain of
00:39:05
Taiwan.
00:39:07
By gaining their friendship, by gaining
00:39:09
their support, we were able to build
00:39:13
Arizona and Texas, California at
00:39:15
incredible rates. They're they are
00:39:18
genuinely a strategic partner. Um we we
00:39:21
we really they deserve our support. They
00:39:25
deserve our friendship. They deserve our
00:39:27
uh generosity and they're doing
00:39:29
everything they can to accelerate the
00:39:31
manufacturing process for us. And so, so
00:39:34
I think that's number one. Number two,
00:39:36
we ought to diversify the manufacturing
00:39:38
supply chain. And whether it's South
00:39:40
Korea, whether it's it's Japan, it's
00:39:43
Europe, we got to we got to diversify
00:39:45
the supply chain, make it more
00:39:46
resilient. And number three, let's be
00:39:48
let's let's demonstrate restraint. And
00:39:53
while we're reducing uh increasing our
00:39:56
diversity and resilience, let's not
00:40:00
press push um you know
00:40:03
>> unnecessary we need to be patient.
00:40:05
>> Is helium a problem?
00:40:07
>> A lot of reports,
00:40:08
>> you know, I think helium could be a
00:40:09
problem, but it's also the case that the
00:40:11
supply chain probably has a lot of
00:40:13
buffer in it.
00:40:14
>> These kind of things tend to have a lot
00:40:15
of buffer. Uh but but um you know
00:40:19
>> Yeah,
00:40:19
>> you've um made massive progress in
00:40:22
self-driving. You made a big
00:40:24
announcement. You've added many more
00:40:26
partners including BYD. There was just a
00:40:28
video of you driving around in a
00:40:30
Mercedes and uh huge announcement uh
00:40:34
with Uber that you're going to have a
00:40:36
number of cars on the road from many
00:40:38
different manufacturers. your bet is I
00:40:41
believe that there's going to be an
00:40:43
Android
00:40:44
type open-source platform that you're
00:40:47
going to play a major part in with
00:40:49
dozens of uh car providers and then
00:40:52
maybe on the other side there could be
00:40:53
an iOS with Tesla or Whimo. What's your
00:40:56
strategy thinking there and how that
00:40:59
chessboard emerges because it feels like
00:41:02
you have a a pretty deep stack and in
00:41:05
some ways you're competing and in other
00:41:07
places you're collaborative. Yeah. Um,
00:41:10
it's taking a step back. We believe that
00:41:13
everything that moves will be autonomous
00:41:16
completely or partly
00:41:18
someday. Number one. Number two, we
00:41:20
don't want to build self-driving cars,
00:41:22
but we want to enable every car company
00:41:24
in the world to build self-driving cars.
00:41:26
And so, we built all three computers,
00:41:28
the training computer, the simulation
00:41:29
computer, the valuation evaluation
00:41:31
computer, as well as the car computer.
00:41:33
We develop the world's safest driving
00:41:36
operating system. Uh we also created the
00:41:40
world's first reasoning autonomous
00:41:43
vehicle so that it could decompose
00:41:45
complicated scenarios into simpler
00:41:47
scenarios that it knows how to navigate
00:41:49
through just like us reasoning systems.
00:41:52
And so that reasoning system called Al
00:41:54
Pomayo has enabled us to achieve
00:41:56
incredible results.
00:41:59
We
00:42:00
open this we ver we vertical
00:42:03
optimization. We horizontally innovate
00:42:06
and we let everybody decide. Do you want
00:42:08
to buy one computer from us? In the case
00:42:09
of Elon and Tesla, they buy our training
00:42:12
computers. Um, do they want to buy our
00:42:14
training computer and our simulation
00:42:15
computers or do you want to let us uh
00:42:18
work with us to do all three and even
00:42:19
put the car computer in your car. So,
00:42:21
we, you know, our attitude is we want to
00:42:24
solve the problem. We're not the
00:42:27
solution provider.
00:42:29
And we're delighted however you work
00:42:31
with us. Let me build on this question
00:42:33
because I think it's like it's so
00:42:34
fascinating. You actually do create this
00:42:36
platform. A thousand flowers are
00:42:38
blooming.
00:42:40
>> But it's also true that some of those
00:42:41
flowers want to now go back down in the
00:42:43
stack and try to compete with you a
00:42:45
little bit. Google has TPU, Amazon has
00:42:48
inferentia and tranium. You know,
00:42:50
everybody's sort of spinning up their
00:42:52
own version of I think I can out Nvidia
00:42:54
Nvidia
00:42:55
>> even though they also tend to be huge
00:42:58
customers.
00:42:58
>> How do you navigate that? And what do
00:43:01
you think happens over time and
00:43:03
>> where do those things play in the
00:43:05
complexion of this kind of vision?
00:43:06
>> Yeah, really great. You know, first of
00:43:08
all, um, we're the only AI company,
00:43:11
we're an AI company. We build foundation
00:43:13
models. We're at the frontier in many
00:43:15
different domains. We build every single
00:43:17
every single layer, every single stack.
00:43:19
Um, we're the only AI company in the
00:43:21
world that works with every AI company
00:43:22
in the world. They never show me what
00:43:25
they're building and I always show them
00:43:26
exactly what I'm building.
00:43:28
>> Right.
00:43:28
>> Yeah. And so so the confidence comes
00:43:31
from this one. Uh we are delighted to
00:43:35
compete on what is the best technology
00:43:38
and to the extent that to the extent
00:43:40
that we can continue to run fast I
00:43:42
believe that buying from Nvidia still is
00:43:45
one of the most economic things they
00:43:46
could do and that's just incredible
00:43:48
confidence there. Number one. Number
00:43:49
two, we're the only architecture that
00:43:51
could be in every cloud and that gives
00:43:53
us some fundamental advantages. We're
00:43:55
the only architecture you could take
00:43:56
from a cloud and put into onrem in the
00:43:59
car in any region
00:44:00
>> in space.
00:44:01
>> That's right. In space. And so there's a
00:44:03
whole whole part of our market about 40%
00:44:06
of our of our business most people don't
00:44:08
realize this 40% of our business unless
00:44:11
you have the CUDA stack unless you can
00:44:12
build an entire AI factory you have the
00:44:14
customers don't know what to do with
00:44:16
you. They're not trying to build chips.
00:44:18
They're not trying to buy chips. They're
00:44:20
trying to build AI infrastructure. And
00:44:22
so they want you to come in with the
00:44:24
full stack. And we've got the whole
00:44:25
stack. And so surprisingly, Nvidia is
00:44:28
gaining market share. If you look at
00:44:30
where we are today, we're gaining share.
00:44:32
>> Do you think what happens is these guys
00:44:33
try and they realize, oh my god, it's
00:44:35
too much. And then they come back. Is
00:44:36
that why the share grows?
00:44:38
>> Well, we're gaining share for several
00:44:39
reasons. One, um, our velocity has gone,
00:44:44
we help people realize it's not about
00:44:45
building the chip, it's about building
00:44:47
the system.
00:44:48
>> And that system is really hard to build.
00:44:50
uh and and so their their their business
00:44:53
with us is increasing. In the case of
00:44:55
AWS, I think they just announced, I
00:44:56
think it was yesterday, that they're
00:44:58
going to buy a a million chips uh in the
00:45:01
next couple years. I mean, that's a lot
00:45:03
of chips from from AWS. And that's on
00:45:05
top of all the chips they've already
00:45:06
bought. And so, we're delighted to do
00:45:08
that. But number one, we're gaining
00:45:10
share this last couple years because we
00:45:13
now have Anthropic coming to Nvidia.
00:45:16
Meta SL is coming to Nvidia. And the
00:45:20
growth of open models is incredible. And
00:45:23
that's all on Nvidia. And so we're
00:45:25
growing in share because of the number
00:45:27
of models. We're also growing in share
00:45:29
because out all of these companies are
00:45:33
outside of the cloud and they're growing
00:45:35
regionally in enterprise in industries
00:45:37
at the edge and that entire segment of
00:45:40
growth is you know really hard to do if
00:45:42
it's just building an as
00:45:44
>> Brad
00:45:44
>> related to that um and not to get in the
00:45:47
weeds on the numbers but analysts don't
00:45:49
seem to believe right so if you look at
00:45:52
the consensus forecast you said compute
00:45:54
could 1 millionx right and Yet they have
00:45:58
you growing next year at 30%, the year
00:46:00
after that at 20%. And in 2029, which is
00:46:03
supposed to be a monster year at 7%.
00:46:06
Right? So if you just if you take your
00:46:08
TAM and you apply their growth numbers,
00:46:11
it suggests that your share will
00:46:12
plummet. Do you see anything in your
00:46:15
future order book that would make that
00:46:18
correct?
00:46:18
>> Yeah. First of all, they just don't
00:46:20
understand the scale and the breadth of
00:46:23
AI.
00:46:23
>> Yes.
00:46:24
>> Yeah.
00:46:24
>> I think that's true. Most people think
00:46:26
that AI is in the top five hyperscalers,
00:46:29
>> right? That's right. There's also an
00:46:31
orthodoxy around these law of large
00:46:33
numbers where,
00:46:34
>> you know, they have to go back to their
00:46:36
investment banking risk committee and
00:46:37
show some model.
00:46:39
>> They're not going to believe in their
00:46:41
minds that 5 trillion goes to 15
00:46:43
trillion. They're like go to it can go
00:46:45
to seven or they can have a 10 trillion
00:46:48
company.
00:46:48
>> It's all just CIA stuff that I think
00:46:50
>> it's never happened before. So you can't
00:46:51
say it will
00:46:52
>> and and because because you have to
00:46:54
redefine what it is that you do. There
00:46:56
was somebody who made an observation
00:46:57
recently that Nvidia
00:47:00
Jensen how can you be larger than Intel
00:47:04
in servers and the reason for that is
00:47:06
because the CPU market of the entire
00:47:09
data center was about $25 billion a
00:47:11
year,
00:47:12
>> right?
00:47:12
>> We do $25 billion a year as you guys
00:47:14
know in a very in the time that we were
00:47:16
sitting here.
00:47:17
>> And so obviously obviously
00:47:21
That was a joke.
00:47:22
>> No, it's but it's
00:47:23
>> all in podcast.
00:47:25
Don't worry. Everything on this show is
00:47:27
roughly. Don't worry about it. It's all
00:47:28
in here. Anyway, that was not guidance.
00:47:32
But anyhow, anyhow, it the the point is
00:47:35
how big you can be
00:47:37
>> depends on what is it that you make,
00:47:39
>> right?
00:47:40
>> Nvidia is not making chips. Number one,
00:47:42
making chips does not help you solve the
00:47:44
AI infrastructure problem anymore. It's
00:47:46
too complicated. Number three, most
00:47:49
people think that AI is narrowly in the
00:47:51
things that they talk about and hear and
00:47:53
see.
00:47:54
>> It's AI is much open AI is incredible.
00:47:57
They're going to be enormous. Anthropic
00:47:59
is incredible. They're going to be
00:48:00
enormous. But AI is going to be much
00:48:03
much bigger than that.
00:48:05
>> And we addressed that segment.
00:48:06
>> Tell us about data centers in space for
00:48:08
a second.
00:48:09
>> Yeah.
00:48:09
>> Um
00:48:10
>> we're already in space. How should the
00:48:12
layman think about what that business is
00:48:16
versus when you hear about these big
00:48:17
data center buildouts that's happening
00:48:19
in in on the ground?
00:48:21
>> Well, we should definitely work on the
00:48:23
ground first because we're already here
00:48:25
and number one. Number two, we should
00:48:27
prepare to be out in space and obviously
00:48:28
there's a lot of energy in space. Um the
00:48:31
challenge of course is that cooling
00:48:34
you can't take advantage of conduction
00:48:36
and convection and so you can only use
00:48:38
radiation and radiation requires very
00:48:41
large surfaces and so now that's not an
00:48:43
impossible thing to solve and there's a
00:48:45
lot of lot of space in space. Um but
00:48:48
nonetheless
00:48:49
the expense is still quite there is is
00:48:51
there uh we're going to go explore it.
00:48:53
We're already there. We're already
00:48:55
radiation hardened. Uh we have we have
00:48:57
uh uh uh CUDA in satellites around the
00:49:00
world. Um they're doing imaging, image
00:49:02
processing, AI imaging and um and that
00:49:05
kind of stuff ought to be done in space
00:49:07
instead of sending all the data back
00:49:08
here and do imaging down here. We ought
00:49:10
to just do imaging out in space. And so
00:49:12
there's a lot of things that we ought to
00:49:13
done do do in space. And in the
00:49:15
meantime, uh we're going to explore what
00:49:17
is the architecture of data centers look
00:49:18
like uh in space. And it'll take it'll
00:49:21
take years. It's okay. We got I got
00:49:23
plenty of time. I wanted to um double
00:49:25
click on healthcare. I know you've got a
00:49:26
big effort there. We're all of a certain
00:49:28
age where we're thinking about lifespan,
00:49:30
health span. I mean, we all look great.
00:49:32
I think
00:49:34
>> some better than others.
00:49:34
>> I think some better than others. I don't
00:49:36
know what your secret is, Jensen.
00:49:37
>> Pretty good these these
00:49:38
>> I mean, what's what are you taking?
00:49:40
What's off the menu? You got to talk to
00:49:42
me when we're backstage. I want to know
00:49:43
in the green room what you got going on.
00:49:45
>> Squats and push-ups and sit-ups.
00:49:47
>> Perfect. Okay. Um but
00:49:49
>> that works. what you know in terms of
00:49:52
the buildout in healthcare
00:49:56
where is that going and what kind of
00:49:58
progress are we making? I was just using
00:50:00
Claude to do some analysis and saying
00:50:02
like where are all these billing codes?
00:50:04
We spend twice as much money in the US.
00:50:06
We get seem to get half as much. It
00:50:08
seemed like uh 15 to 25% of the dollar
00:50:11
spent were on these first GP visits. And
00:50:14
I think we all know like chat GBT and a
00:50:17
large language model does a better job
00:50:19
more consistently today at a first
00:50:22
visit. So what has to happen there to
00:50:25
kind of break through all that
00:50:26
regulation and have AI have a true
00:50:28
impact on the health care system?
00:50:29
>> There's several way several areas that
00:50:31
we're involved in in um in healthcare.
00:50:34
One is uh AI
00:50:38
uh physics uh and and that's or AI
00:50:41
biology using AI to understand represent
00:50:45
predict biology behavior biological
00:50:47
behavior and so that's one that's very
00:50:49
important in drug discovery. There's
00:50:51
second which is AI agents and that's
00:50:54
where the assistance and helping
00:50:56
diagnosis and things like that. Open
00:50:58
evidence is a really good example.
00:50:59
Hypocratic is a really good example.
00:51:01
Love working with those companies. Um I
00:51:03
really think that this is an area uh
00:51:04
where agentic technology is going to
00:51:06
revolutionize how we interact with
00:51:09
doctors and how do we interact for
00:51:10
healthcare. The third part that we're in
00:51:12
involved in is physical AI. The first
00:51:14
one is AI physics using AI to predict
00:51:16
physics. The second one is physical AI.
00:51:19
AI that understand the properties of the
00:51:21
laws of physics and that's used for a uh
00:51:24
robotic surgery huge amounts of
00:51:27
activities there. Every single
00:51:29
instrument whether it's ultrasound or
00:51:31
you know CT or whatever instrument we
00:51:34
interact with in a hospital in the
00:51:35
future will be agentic.
00:51:36
>> Yeah.
00:51:37
>> You know open claw in a safe version
00:51:39
will be inside every single instrument.
00:51:42
And so in a lot of ways that instrument
00:51:44
is going to be interacting with patients
00:51:45
and nurses and doctors in a very unique
00:51:48
way. so much investment in AI weapons.
00:51:50
It would be wonderful to see some
00:51:52
investment in AI EMTs and paramedics and
00:51:55
saving lives, not just taking them,
00:51:57
which I think is a great segue into
00:51:59
robotics. You've got dozens of partners.
00:52:02
We have this very weird
00:52:04
>> I I don't know I want to call a lost
00:52:05
decade or 20 years of Boston Dynamics.
00:52:09
Google bought a bunch of companies. They
00:52:10
then wound up selling them and spinning
00:52:12
them out where people just thought
00:52:14
robotics is just not ready for prime
00:52:16
time. And now here we have the world's
00:52:18
greatest entrepreneur at this time. Uh
00:52:20
tied with you, uh Elon Musk doing well,
00:52:23
that was a good save, I hope. Optimus,
00:52:25
uh pretty impressive. And then other
00:52:27
companies in China. How how close is
00:52:30
that to actually being in our lives
00:52:34
where we might see a chef, a robotic
00:52:37
chef, a robotic nurse, a robotic
00:52:39
housekeeper, you know, this humanoid
00:52:41
factor actually working in the real
00:52:43
world, knowing what you know with those
00:52:46
partners and the fidelity, especially in
00:52:48
China where they seem to be doing as
00:52:49
good a job as we're doing here or maybe
00:52:51
better.
00:52:52
>> Um,
00:52:54
we invented the industry largely.
00:52:57
America invented. We c you could argue
00:52:59
we got into it too soon.
00:53:00
>> Yeah.
00:53:01
>> And and we got exhausted. We got tired
00:53:04
um about five years before the enabling
00:53:08
technology appeared.
00:53:09
>> The brain.
00:53:09
>> Yeah. Yeah. And we we just got tired of
00:53:11
it just a little too soon. Okay. That's
00:53:14
number one. But it's here now. Now the
00:53:16
question is how much longer? From the
00:53:18
point of high functioning existence
00:53:21
proof, high functioning exist existence
00:53:24
proof to reasonable products
00:53:28
technology never takes more than a
00:53:30
couple two three cycles. And so a couple
00:53:33
two three cycles basically be somewhere
00:53:34
around 3 years to 5 years. That's it. 3
00:53:37
years to 5 years we're going to have
00:53:38
robots all over the place. Uh I think I
00:53:40
think um uh China is is uh formidable
00:53:43
and the reason for that is because their
00:53:46
micro electronics, their uh motors,
00:53:49
their rare earth, their magnets, which
00:53:50
is foundational to robotics,
00:53:53
>> they are the world's best. And so in a
00:53:55
lot of ways, our robotics industry
00:53:57
relies deeply on their ecosystem and
00:53:59
their supply chain. Um and uh and and
00:54:02
they're, you know, obviously moving very
00:54:04
quickly. Uh we're going to, you know,
00:54:06
our robotics industry will have to rely
00:54:08
a lot on it. the world's robotics
00:54:09
industry will have to rely on a lot on
00:54:11
it. And so so I think um you're gonna
00:54:14
see some fast fast movements here.
00:54:16
>> Ultimately, one for one. Elon seems to
00:54:19
think we're going to have one robot for
00:54:20
every human. 7 billion for 7 billion, 8
00:54:22
billion for 8 billion.
00:54:23
>> Well, I'm hoping more. Yeah, I'm hoping
00:54:25
more. Yeah. Uh well, first of all,
00:54:27
there's a whole bunch of robots that are
00:54:29
going to be in factories working around
00:54:30
the clock. There's going to be a whole
00:54:32
bunch of fac
00:54:34
that that don't move. They move a little
00:54:36
bit. Uh almost everything will be
00:54:37
robotic. What does the world look like?
00:54:39
>> Sorry, let me I think like this is one
00:54:41
of the robotics for me is one of the
00:54:43
pieces that I think unlocks uh economic
00:54:45
mobility opportunities for every
00:54:47
individual. Everyone now like when
00:54:49
everyone got a car, they could now go
00:54:51
and do a lot of different jobs. When
00:54:53
everyone gets a robot, their robot can
00:54:55
do a lot of work for them. They can
00:54:57
stand up an Etsy store, a Shopify store.
00:54:59
They can create anything they want with
00:55:01
their robot. They could do things that
00:55:03
they independently cannot do. I think
00:55:05
the robot is going to end up being the
00:55:07
greatest unlock for prosperity for more
00:55:09
people on Earth than we've ever seen
00:55:11
with any technology before.
00:55:13
>> Yeah, no doubt. I mean, just a simp the
00:55:15
simple math at the moment is we're
00:55:17
millions of people short in labor today.
00:55:19
Right. Yeah.
00:55:19
>> Right. We're we're we're actually really
00:55:22
desperate in need of robotics and so
00:55:24
that all of these companies could grow
00:55:27
more if they had more labor. I mean,
00:55:29
we're we're number one. Some of the
00:55:32
things that you mentioned are super fun.
00:55:33
I mean, because of robots, we'll have
00:55:35
virtual presence. Uh, you know, I'll be
00:55:38
able to go into the robot of my house
00:55:42
and virtually operate it. I'm on a
00:55:44
business trip,
00:55:45
>> right?
00:55:46
>> Walk around the house and walk the dog.
00:55:48
>> Yeah. Walk the dog.
00:55:49
>> Break the leaves.
00:55:50
>> Yeah. Exactly. Freak out the dog.
00:55:51
>> Maybe not quite that, but just, you
00:55:53
know, just, you know, wander around and
00:55:55
just see what's going on in the house.
00:55:56
You know, chat with the dog, chat with
00:55:58
the kids.
00:55:58
>> Yeah.
00:55:59
>> Yeah. And that's time travel is also
00:56:01
we're going to be able to travel at the
00:56:02
speed of light, you know, and so, you
00:56:04
know, clearly we're going to send our
00:56:06
robots ahead of us.
00:56:07
>> Yeah.
00:56:08
>> Not going to send myself. I'm going to
00:56:09
send a robot, you know.
00:56:10
>> Check it out.
00:56:11
>> Yeah. Yeah. And then I'm going to upload
00:56:12
my AI. Yeah.
00:56:13
>> Well, it's inevitable. It unlocks the
00:56:14
moon and it unlocks Mars as um targets
00:56:17
for for colonization, which gives us
00:56:19
>> infinite resources. Getting back from
00:56:21
the moon is effectively zero energy cost
00:56:23
to move material back because you can
00:56:25
use solar and accelerate. So you could
00:56:27
have factories that make everything the
00:56:29
world needs on the moon and the robots
00:56:31
are going to be the unlock for enabling
00:56:32
that.
00:56:32
>> That's right. Distance no longer
00:56:33
matters.
00:56:34
>> Distance doesn't matter. Yeah.
00:56:35
>> The more the more revenue we get out of
00:56:38
models and agents, the more we can
00:56:40
invest in building the infrastructure
00:56:41
which then unlocks more capabilities on
00:56:43
models and agents. Daario on Dwaresh's
00:56:46
podcast recently said by 2728 we'll have
00:56:49
hundreds of billions of dollars of
00:56:51
revenue out of the model companies and
00:56:53
the agent companies. and he forecasts a
00:56:55
trillion dollars by 2030. Right? This is
00:56:58
non-infrastructure AI revenue. Um
00:57:01
>> I think he I think he's he's being very
00:57:04
conservative. I believe Dario and
00:57:05
Anthropic is going to do way better than
00:57:07
that.
00:57:08
>> Wow.
00:57:08
>> Way better than that.
00:57:09
>> Wow. So from 30 billion to a trillion.
00:57:11
>> Yeah. and not and and the reason for
00:57:13
that is the one part that he hasn't
00:57:15
considered is that I believe every
00:57:17
single enterprise software company will
00:57:20
also be a reseller
00:57:23
value added reseller of anthropic code
00:57:25
anthropics tokens value added reseller
00:57:28
open AAI that's right and they're going
00:57:31
to that that that part of their
00:57:34
>> get this logarithmic expansion
00:57:35
>> yes
00:57:36
>> their go to market is going to expand
00:57:39
tremendously this year
00:57:41
>> what do you think in that world is the
00:57:42
moat what's left over. I mean you have
00:57:44
some modes that are frankly I think as
00:57:47
this scales almost insurmountable. The
00:57:49
best one that nobody talks about is
00:57:51
probably CUDA which is just like an
00:57:54
incredible strategic advantage. But in
00:57:56
the future if a model can be used to
00:57:59
create something incredible then the
00:58:01
next spin of a model can be used to
00:58:03
maybe disrupt it. Sort of in your mind
00:58:05
what do you think for these companies
00:58:06
that are building at that application
00:58:08
layer? What's their moat? like how do
00:58:10
they differentiate themselves?
00:58:12
>> Deep specialization.
00:58:14
Deep specialization. I believe that um
00:58:17
these models they're going to have
00:58:19
general general models that are
00:58:20
connected into the software company's
00:58:24
agentic system,
00:58:25
>> right?
00:58:26
>> Many of those models are cloud models
00:58:29
and proprietary models, but many of
00:58:31
those models are specialized sub aents
00:58:34
that they've trained on their own.
00:58:37
>> Right. So the call to arms for you for
00:58:39
entrepreneurs is look
00:58:40
>> know your vertical.
00:58:42
>> That's right.
00:58:42
>> Know it as deep and as better than
00:58:44
everybody else.
00:58:45
>> That's right.
00:58:45
>> And then wait for these tools because
00:58:47
they're catching up to you and now you
00:58:48
can imbue it with your knowledge.
00:58:49
>> That's right. The sooner you connect
00:58:51
your agent,
00:58:52
>> the sooner you connect your agent with
00:58:54
customers,
00:58:55
>> that flywheel is going to cause your
00:58:57
agent to get
00:58:58
>> it very much is an inversion of what we
00:59:00
do today because today we build a piece
00:59:02
of software and we say what generalizes
00:59:04
>> and then let's try to sell it as broadly
00:59:06
as possible and then sell the
00:59:07
customization around it
00:59:08
>> and we in fact in fact exactly right we
00:59:11
we create a horizontal but notice there
00:59:15
are all these gsis and all of these
00:59:17
consultants who are specialists who then
00:59:20
take your horizontal platform and
00:59:22
specializes it into
00:59:24
>> and that's arguably a five or six time
00:59:26
bigger industry is the customization.
00:59:28
>> It is absolutely the whole very much is
00:59:30
>> that's right. So I think that these
00:59:32
platform companies have an opportunity
00:59:34
to become that specialist to become that
00:59:37
vertical.
00:59:38
>> Yeah. Domain expert.
00:59:38
>> You know, I just want to give you your
00:59:39
flowers. I think it was 3 years ago you
00:59:41
said you're not going to lose your job
00:59:43
to AI. You're going to lose your job to
00:59:44
somebody using AI. And here we are. The
00:59:47
entire conversation has revolved around
00:59:49
this concept of agents making people
00:59:52
superhuman and the business opportunity
00:59:55
expanding and entrepreneurship
00:59:56
expanding. You actually saw it pretty
00:59:58
clearly. Yeah.
00:59:58
>> You changed your view.
01:00:01
This is Doom Dmer. No, I'm not Doomer. I
01:00:04
do I do have Dmer. No, I you can hold
01:00:07
space for I think two ideas. One is
01:00:08
there are going to be a lot
01:00:09
>> that's viral Jake.
01:00:11
>> Oh, no. There you can.
01:00:12
>> But that's just because he doesn't hang
01:00:13
out with me enough.
01:00:15
>> Well, we I mean we a little bit. We
01:00:17
don't talk about it. He will show THE
01:00:19
TABLE. HE'LL FOLLOW YOU AROUND.
01:00:21
>> I'm not asking for it. I'm just
01:00:22
>> follow you around. I'm not asking for
01:00:24
it.
01:00:24
>> You can come with me and Tucker. We ski
01:00:26
in Japan every January. Love it. and
01:00:29
Tucker go road trip.
01:00:31
There is going to be job displacement
01:00:33
and then the question becomes,
01:00:35
>> you know, do those people have the
01:00:37
fortitude, the resolve to then go
01:00:39
embrace these,
01:00:40
>> you know, technologies. We're we're
01:00:42
going to see 100% of driving go away by
01:00:44
humans. That's just it's that's a
01:00:47
beautiful thing and the lives saved, but
01:00:49
we have to recognize that's 15 million
01:00:50
people in the United States, 10 to 15
01:00:52
million who are employed in that way.
01:00:54
And and so that is going to happen. Yes,
01:00:57
>> I I think I think that jobs will change.
01:00:59
For example, um there are many
01:01:01
chauffeers today uh who drives the car.
01:01:04
I believe that though many of those
01:01:05
chauffeers will actually be in the car
01:01:08
sitting behind the drive the steering
01:01:10
wheel while the car is driving by
01:01:12
itself. And the reason for that is
01:01:14
because remember what a chauffeur does
01:01:16
in the end. These chauffeers, they're
01:01:18
helping you they're your assistants.
01:01:20
They're helping you with your luggage.
01:01:21
They're helping you. I mean, they're
01:01:22
helping you with a lot of things and and
01:01:24
so I wouldn't be surprised actually if
01:01:27
the chauffeers of the future becomes
01:01:29
your mobility assistant and they are
01:01:32
helping you do on a whole bunch of other
01:01:33
stuff
01:01:33
>> to the hotel.
01:01:34
>> Yeah. And the car is driving by itself.
01:01:36
>> The autopilot in planes created a lot
01:01:39
more pilots and didn't take any of the
01:01:41
pilots out of the cockpit even though
01:01:43
the autopilot is flying the plane 90% of
01:01:45
the time. By the way, while that car is
01:01:47
driving itself, that chauffeur is going
01:01:49
to be doing a bunch of other work on his
01:01:50
phone and he's going to be
01:01:52
>> arranging, for example, coordinating a
01:01:54
bunch of things for you, getting, you
01:01:55
know, it's all the pie just grows in a
01:01:57
way that
01:01:58
>> one of the things that that that
01:02:01
yes, every job will be will be
01:02:02
transformed. Um, some jobs will be
01:02:04
eliminated. However, we also know that
01:02:07
many many jobs will be recre will be
01:02:08
created. The one thing that I will say
01:02:10
to young people who are coming out of
01:02:12
school who are concerned who are anxious
01:02:14
about AI be the expert of using AI
01:02:18
>> how much look we all want our employees
01:02:20
to be expert at using AI and it's not
01:02:23
not
01:02:25
>> not trivial not trivial and so knowing
01:02:28
how to specify not to overprescribe
01:02:32
leaving enough room for the AI to
01:02:34
innovate and create while we guide it to
01:02:37
the outcome we want. it. All of that
01:02:40
requires artistry.
01:02:41
>> You had you had this great advice to
01:02:43
when you were at Stanford, I think it
01:02:45
was, which is I wish to you pain and
01:02:46
suffering. Do you remember that?
01:02:47
>> Yeah.
01:02:48
>> Fantastic.
01:02:49
>> What's your advice to young people
01:02:50
around what they should be studying? So,
01:02:52
if they're sort of about to leave high
01:02:54
school because now those are the kids
01:02:56
that are at this like really native,
01:02:58
they haven't made a decision about
01:02:59
college, what to study, if at all go to
01:03:02
college. How do you guide those kids?
01:03:04
What would you tell them? I I still
01:03:07
believe that deep science, deep math, um
01:03:10
language skills, you know, as you know,
01:03:13
language is the programming language of
01:03:16
AI,
01:03:17
>> the ultimate programming language.
01:03:18
>> And so, as it turns out, it it could be
01:03:20
that the English major could be the most
01:03:22
successful. Yeah.
01:03:23
>> And and so so I think I think um I I
01:03:26
would just advise whatever whatever
01:03:28
education you get, just make sure that
01:03:30
you're deeply deeply expert in using
01:03:33
AIS. One of the things that I wanted to
01:03:35
say with respect to jobs and I want
01:03:36
everybody to hear it that in fact at the
01:03:39
beginning of the deep learning
01:03:40
revolution, one of the the finest
01:03:43
computer scientists in the world deeply
01:03:45
deeply I deeply uh deeply uh um respect
01:03:49
uh predicted that computer vision will
01:03:51
completely eliminate radiologists
01:03:54
and and that the one the one field he
01:03:57
advises everybody to not go into is
01:04:00
radiology. 10 years later, his
01:04:03
prediction was at 100% right. Computer
01:04:06
vision has been integrated into all of
01:04:09
the radiology technologies and radiology
01:04:11
platforms in the world 100%. The
01:04:14
surprising outcome is the number of
01:04:16
radiologists actually went up and the
01:04:18
demand for radiologists is skyrocketed.
01:04:21
The reason for that is because
01:04:23
everybody's job
01:04:25
has a purpose and its task. The task
01:04:29
that you do is studying the scans,
01:04:33
>> but your purpose is to diag helping the
01:04:36
doctors, helping the patient diagnose
01:04:38
disease.
01:04:38
>> And so what's surprising is because the
01:04:41
scans are now being done so quickly,
01:04:44
>> they could do more scans, improving
01:04:46
healthcare.
01:04:47
>> Yes.
01:04:47
>> But doing more scans more quickly allows
01:04:50
patients to
01:04:51
>> be
01:04:53
onboarded a lot more quick, treated a
01:04:55
lot more quickly. And as it turns out,
01:04:57
because hospitals enjoy making money,
01:05:00
too.
01:05:00
>> Yeah.
01:05:01
>> Right.
01:05:02
>> They're doing more scans,
01:05:04
>> they're treating more customers and
01:05:05
patients, the revenues go up. Guess
01:05:07
what?
01:05:09
>> And and a country that grows faster,
01:05:11
productivity increases. A wealthier
01:05:14
country can put more teachers in the
01:05:16
classroom, not less teachers in the
01:05:17
classroom. That's right. You just give
01:05:19
every one of those teachers a
01:05:20
personalized curriculum for every
01:05:22
student in the room. It makes them all
01:05:23
bionic and leads to a lot more. Every
01:05:26
single student will be assisted by AI,
01:05:28
but every single student will need great
01:05:30
teachers.
01:05:31
>> Yeah. Amazing. Uh Jensen,
01:05:33
congratulations. I know your success and
01:05:35
really this is an incredibly positive,
01:05:37
uplifting discussion. We really
01:05:39
appreciate you taking the time for us.
01:05:40
He is the steward we need.
01:05:42
>> You are you are the more vocal. I'm
01:05:45
being very vocal about the positive side
01:05:47
of it. I think there's too much dumerism
01:05:48
is
01:05:49
>> but I also think it takes the humility
01:05:50
to have this level of success and be
01:05:52
humble about we're making software guys.
01:05:55
Yeah.
01:05:56
>> And I think that that's actually really
01:05:58
healthy for people to hear. We have done
01:06:00
this before. We have invented categories
01:06:02
and industries before.
01:06:04
>> We don't need to go to this
01:06:06
>> scaremongering place. It does nothing.
01:06:08
>> And we get to choose, right? We have
01:06:10
autonomy and and agency. We get to pick
01:06:12
how to
01:06:13
>> we sure do
01:06:13
>> employ this. Okay, everybody. We'll see
01:06:15
you next time on the All-In interview.
01:06:18
Okay.
01:06:19
>> Well done, brother.
01:06:20
>> Thanks, man.
01:06:20
>> Good job.
01:06:21
>> Thank you, sir. That was awesome.
01:06:23
>> Good. Good. Appreciate you.
01:06:24
>> You guys are awesome.
01:06:25
>> Look at this. Look at this big crowd
01:06:27
behind you guys,
01:06:27
>> man. I think they're here for you.
01:06:36
>> I'm going all in.
