Keynote by NVIDIA CEO Jensen Huang at 2024 SIEPR Economic Summit

welcome back everyone after the short

break I know that many of you are

looking forward to hearing from our next

speaker Jensen

Wong Jensen is at The Cutting Edge of

artificial

intelligence and all of the

innovation

technology and human capital that is

needed to support

it my good friend and Seer colleague

John Chauvin is going to introduce

Jensen and I hope he's here somewhere so

I'm just going to keep talking and then

the two of them will have a conversation

before taking some of your

questions John chovin certainly requires

very little introduction to many most In

This Crowd as my predecessor as the Tron

director of seer John is the one who

started the SE economic Summit 20 years

ago so I would just like right now for

all of us to give John chovin a huge

round of applause and appreciate the

community that he had the foresight to

build uh for those of you who haven't

been touched by John's research his

mentorship or his friendship here's what

here's just a snippet of what you might

like to know about him along with being

the former Seer director and a Seer

Senor senior fellow Meritus John is the

Charles R Schwab professor of Economics

he is also a senior fellow at the Hoover

institution and a research associate of

the National Bureau of economic research

he specializes in public finance and

corporate finance and has published many

articles over the years on social

security health economics corporate

personal taxation mutual funds pension

plans economic demography applied

General equilibrium economics and much

more uh John isn't one for long

introductions but I just will say that

if I can be on10th as helpful to my

successor as John uh has been to me I'll

feel like I've uh succeeded so I will

let you read more about his Publications

and accomplishments in the programs

you've received uh today and so please

join me in welcoming our good friend

John Chauvin and I'm really looking

forward to

this

thanks wow thank you so I have always

thought that the more famous the speaker

the shorter the appropriate

introduction and if I was to follow that

rule I would stop right now and say

Jensen Wong but I'm not going to do

that

um so the Oxford English

Dictionary defines the American

dream believe it or not it does that and

it says that it's a situation where

everybody has an equal opportunity for

Success Through hard work dedication and

initiative and I would like to say that

Jensen Wong is an example of the

American dream

Jensen uh was born in

Taiwan came to the US at age nine with

his brother not with his

parents went to a rough tough School in

Kentucky survived that his parents came

two years later he moved to Oregon

skipped two grades and graduated from

high school and went to Oregon state

electrical engineering major 150 men and

two

women he said he was 16 he looked like

he was 12 he had no chance with the

women

well he sort of liked one of them and

said why don't we work on homework

together did that over and over and over

again six months later he after out for

a date well he's still married to her so

another American

Dream now to skip to age 30 he co-founds

Nvidia he's the only CEO there's ever

been of

Nvidia it's had its ups and its down

more UPS than

Downs it's now the fourth largest

company in the world third largest

American uh company so that sounds to me

like the American

dream um I should add that he also got a

degree from Stanford master's degree I

think he did it mostly at night

uh and he was always good with homework

at worked with his wife at worked with

Stanford uh

too um now of course we were here last

week Nvidia announced its

earnings in the finance

crowd this got more attention than the

Super Bowl that occurred a couple weeks

earlier it was pretty uh amazing uh his

company is at the absolute center of the

most exciting develop vment I'd say of

the 21st century technology development

and uh so he's to be congratulated on

that let me just say uh he's received a

lot of

awards a lot of recognition Enid has

received a lot of awards a lot of

recognition but I should have a short

introduction so I'm about to quit I'm

just going to talk about one

award last month he was elected as a

member of the National Academy of

engineering this is a pretty prestigious

award there are only three that I know

of I actually asked chat GPT I didn't

get an absolute clear

answer how many CEOs of S&P 500

companies are members of the National

Academy of engineering but I think it's

three and two are in this room anaru

Devan of Cadence Design Systems was

awarded it last year so the two of them

have that in common but let me now just

conclude and

congratulate Jensen not only on this

award but on the amazing success of your

company and thank you for speaking to us

today at Seer

Jensen how

it thank you thank you you're here I'm

here I guess so

okay so why don't you start off with

maybe some opening remarks and then I'll

ask you a few questions and then then

you get the tough questions well I think

that after your opening remarks uh it is

smartest for me not to make any opening

remarks to to uh uh avoid risking uh

damaging all the good things you said

you know but but um let's see it's it's

always good to have a pickup line um and

mine was was uh do you want to see my

[Laughter]

homework and you're right we're married

still we have two beautiful kids I have

a perfect life uh two great puppies and

um I love my job and and uh she still

enjoys my

homework well if you want I can ask you

a few questions then yes please so if in

my lifetime I thought the biggest

technical development technology

breakthrough was the transistor now I'm

older than you yeah uh and it was pretty

fundamental deal but should I rethink is

AI now the biggest change in

technology that has occurred in the last

76 years to to hint at my age yeah um

well first first of all the the

transistor was obviously a great

invention but what

was the greatest capability that enabled

was

software the ability for humans to

express our ideas algorithms uh in a

repeatable way computationally

repeatable way uh was a was is the

Breakthrough um what have we done we

dedicated our company in the last 30

years 31 years uh to a new form of

computing called accelerated Computing

the idea is that general purpose

Computing is not ideal for every every

field of work and we said why don't we

in invent a new way of doing computation

such that we can solve problems that

general purpose Computing is ill

equipped at at

solving and and uh uh what we what we

have effectively done in in a particular

area of a domain of computation that is

that's that is algorithmic in nature

that can be paralyzed we've taken the

computational cost of computers to

approximately zero

so what happens when you when you uh are

able to take the marginal cost of

something to approximately zero some we

enabled a new way of doing software

where it used to be written by humans we

now can use computers to write the

software because the computational cost

is approximately zero and so you might

as well uh let the computer go off and

grind on just a massive amount of

experience we call data digital

experience human dig digital experience

called data and grind on it to find the

relationships and patterns that as a

result represents human knowledge and

that miracle happened about a decade and

a half ago we saw it coming and and we

took the whole company and we shaped our

computer which was already which was

already driving the marginal cost of

computing down to

zero and we pushed it into this whole

domain and as a result in the last 10

years we reduced the cost of computing

by 1 million times

the cost of deep learning by 1 million

times and a lot of people said said to

me but Jensen if you if you reduce the

cost of computing your your cost by a

million times then people buy less of it

and it's exactly the opposite we saw

that if we could reduce the marginal

cost of computing down to approximately

zero we might use it to do something

insanely amazing large language

models to literally extract all of

digital human knowledge from the

internet and put it into to a computer

and let it go figure out what the wisd

what the knowledge is that idea of

scraping the entire internet and putting

it in one computer let the computer

figure out what the program is is an

insane concept but you wouldn't ever

consider doing it unless the marginal

cost of computing was zero and so so we

made we made that breakthrough and now

we've enabled this new way of doing

software imagine you know for for all

the people that are still new to

artificial intelligence we figured out

how to use a computer to understand the

meaning not the pattern but the meaning

of almost all digital knowledge and

everything you can digit anything you

can digitize we can understand the

meaning so let me give you an example

Gene sequencing is digitizing genes but

now with large language models we can go

understand go un go learn the meaning of

that

Gene amino acids we

digitized you know through Mass Spec we

digitized

um Pro amino acids now we can understand

from the amino acid sequence without a

whole lot of work with cryms and things

like that we can go figure out what is

the structure of the protein and what it

does what is this meaning we can also do

that on a fairly large scale pretty soon

we can understand what's the meaning of

a cell a whole bunch of genes that are

connected together and this is from a

computer's perspective no

different than there's a a a whole page

of words and you asked it to what is the

meaning of it summarize what did it say

summarize it for me what's the meaning

this is no different than a hard you

know big huge long page of genes what's

the meaning of that big long page of

proteins what's the meaning of that and

so we're on the cusp of all this this is

just this is the miracle of of what

happened and so I would it's a

longwinded answer of saying John that

you're absolutely right that that that

that AI which was enabled by this form

this new form of computing we call

Accelerated Computing that took three

decades to do uh is probably the single

greatest invention of the computer of

the in of the technology industry this

will likely be the most important thing

of the 21st

century I agree with that 21st century

but maybe not the the 20th century which

was the transistor which it's got to be

close we'll let history decide that's

right we'll let history decide could you

look ahead you I I I take it that the

the GPU chip that is

behind uh artificial intelligence right

now is your h100 and I know you're

introducing an h200 and I think I read

that you plan to upgrade that each year

and so could you think ahead five years

March

2029 you're introducing the

H700 right what will it allow us to do

that we can't do

now um I'll go backwards but but let me

first say something about the chip that

John just described um as we say a chip

all of you in the audience probably

because you've seen a chip before you

you imagine there's a chip kind of like

you know like this um the chip that John

just described uh weighs 70

lbs it consists of 35,000

Parts eight of those parts came from

tsmc it that one

chip

replaces um a data center of old CPUs

like this into one

computer the savings because we compute

so fast the

savings of that one computer is

incredible and yet it's the most

expensive computer the world's ever seen

it's it's a quarter of a million dollar

per chip we sell the world's first quar

million dollar chip but the system that

it replaced the cables alone cost more

than the chip this

h100 the cables of connecting all those

old computers that's the that's the

incredible thing that we did we

reinvented Computing and as a result

Computing marginal cost of computing

went to zero that's what I just

explained we took this entire data

center We Shrunk it into this one chip

well this one

chip uh uh is really really great at

trying to figure out um uh uh this form

this form of computation that that

without without

without getting weird on you guys um

call Deep learning it's really good at

this thing called Ai and so so uh the

way that this chip

works it works not just at the chip

level but it works at the chip level and

the algorithm level and the data center

level it works

together it can't it doesn't do all of

its work by itself it works as a team

and so you connect a whole bunch of

these things together and it works at

you know networking as part of it and so

when you look at one of our computers it

it's a it's a magnificent thing you know

only only computer Engineers would think

it's magnificent but it's magnificent

okay um it weighs a lot miles and miles

of cables hundreds of miles of cables

and and the next one's soon coming is

liquid cooled and you know it's

beautiful in a lot of ways okay and and

um uh and it computes at data center

scales and together what's going to

happen in the next 10 years say John um

we'll increase the computational

capability for M for deep learning by

another million times and what happens

when you do that what happens when you

do that um today we we kind of learn and

then we apply it we go train inference

we learn and we apply it in the future

we'll have continuous

learning We could decide whether that

whatever that continuous learning um

result it will be uh uh deployed into

you know the world's applications or not

but the computer will will watch videos

and and new text and uh from all the

interactions that it's just continuously

improving itself the learning process

and the Train the the training process

and the inference process the training

process and the deployment process

application process will just become

one well that's exactly what we do you

know we don't have like between now and

o' in the morning I'm going to be doing

my learning and then after that I'll

just be doing inference you're learning

and inferencing all the time and that

reinforcement learning Loop will be

continuous and that reinforcement

learning will be grounded with real

world data that is been um uh through

interaction as well as synthetically

generated data that we're creating in

real time so this computer will be

imagining all the time does that make

sense just like just as when you're

learning you you take take pieces of

information and you go from first

principles it should work like this and

then we we do the the simulation the

imagination in our brain and that that

future imaginate imag imagin state in a

lot of ways manifests itself to us as

reality and so your AI computer in the

future will kind of do the same it'll do

synthetic data generation it'll do

reinforcement learning it'll continue to

be grounded by real world experiences um

it'll imagine some things it'll test it

with real world experience I'll be

grounded by that and that entire Loop is

just one giant

Loop that's what happens when you can

compute for a million times cheaper than

today and so as I as I'm saying this

notice what's what's at the core of it

when you can drive the marginal cost of

computing down to zero then there are

many new ways of doing something you're

willing to

do this is no different than I'm willing

to go further places because the

marginal cost of Transportation has gone

to zero I can fly from here to New York

relatively cheap cheaply if it would if

it would have taken a month you know

probably never go and so it's exactly

the same in transportation and all just

about everything that we do and so we're

we're going to take the marginal cost of

computing down to approximately zero as

a result we'll do a lot more

computation that causes

me as you probably know there have been

some recent stories that Nvidia will

face more competition in the inference

Market than it has in the training

Market but what you're saying is it's

actually going to be one market I think

can you comment about um you know is

there going to be a separate training

chip market and inference chip Market or

it sounds like you're going to be

continuously uh training and switching

to inference maybe within one chip I I

don't I don't know why don't you explain

more well today today whenever you uh

prompt uh an AI it could be chat GPT or

it could be co-pilot or it could be uh

if you're using a surface nail platform

you using mid Journey um using Firefly

from Adobe whenever you're prompting

it's doing inference you know inference

is right so it's it's generating

information for you whenever you do that

what's behind it 100% of them is NVIDIA

gpus and so Nvidia most of the time you

engage our our our platforms are when

you're inferencing and so we are 100% of

the world's inferencing today is NVIDIA

now is inferencing hard or Easy A lot of

people the the reason why people are

picking on inferences when you look at

training and you look at Nvidia system

doing training when you just look at it

you go that looks too hard I'm not going

to go do that I'm a chip company that

doesn't look like a

chip and so there's a natural and you

have to in order for you to even prove

that something works or not you're $2

billion doll into it

yeah and you turn it on to realize it's

not very effective you're $2 billion in

two years into it the risk the risk of

exploring something new is too high for

the for the customers and and so a lot

of a lot of competitors tend to say you

know we're not into we're not into

training we're into inference inference

is incredibly hard let's think about it

for a

second the the the the response time of

inference has to be really high but this

is the this is the easy part that's the

computer science part the the E the hard

part of inference is the goal of

somebody who's doing inference is to

engage a lot more users to to apply that

software to a large install

base inference is an install base

problem this is no different than

somebody who's writing a an application

on on on an iPhone um the reason why

they do so is because iPhone has such an

large install base almost everyone has

one and so if you wrote an application

for that phone it's going to have the

benefit of it it's going to be able to

benefit everybody well in the case of

Nvidia our accelerated Computing

platform is the only accelerated

Computing platform that's literally

everywhere and because we we've been

working on it for so long if you wrote

an application for inference and you

take that model and you Deploy on

invidious architecture it literally runs

everywhere and so you could touch

everybody you can enable have greater

impact and so the problem with inference

is is actually install base and that

takes enormous patience and years and

years of success and dedication to

architecture compatibility you know so

on so

forth you make completely State

of-the-art chips is it possible though

that you'll face

competition that is claims to be good

enough not as good as Nvidia but good

enough and and much cheaper is that a is

that a threat well first of all

competition um we we have more

competition than anyone on the planet

has competition

uh not only do we have competition from

competitors we have competition from our

customers and um and and I'm the only

competitor to a customer um fully

knowing they're about to design a chip

to replace ours and I show them not only

what my current chip is I show them what

my next chip is and I'll show them what

my chip after that is and so and the

reason for that is because because look

if you don't if you don't make an

attempt at uh uh explaining why you're

good at something

they'll never get a chance to to buy

your your products and so so we're we're

completely open book in working with

just about everybody in the industry um

and and the reason the reason for that

our our advantage is several our

advantage what we're about is several

things whereas you could build a chip to

to be good at one particular algorithm

remember Computing is more than even

Transformers there's this idea called a

Transformers there's a whole bunch of

species of Transformers and their new

Transformers being invented as we speak

and the number of different types of

software is really quite quite rich and

the reason for that is because software

Engineers love to create new things

Innovation and we want that what Nvidia

is good at is that our our architecture

not only does it accelerate algorithms

it's programmable meaning that that you

can use it for SE we're the only

accelerator for SQL SQL was came about

in the

1960s IBM 1970s in storage Computing I

mean sqls structured data is as

important as it gets uh 300 zettabytes

of data being created you know every

couple of years Mo most of it is in sqls

structured databases and so so we're we

can accelerate that we can Accel

accelerate quantum physics we can

accelerate shortes equations we can

accelerate just about you know every

fluids particles um you know lots and

lots of code and so what Nvidia is good

at is the General field of accelerated

Computing one of them is generative Ai

and so for a data center that wants to

have a lot of customers some of it in

financial services some of it you know

some of it in in manufacturing so on so

forth in the world of computing we're

you know we're we're a great standard

we're in every single Cloud we're in

every single computer company and so our

company's architecture has become a

standard if you will after some 30

somewhat years and and so that's that's

really our advantage if a customer can

can um do something specifically that's

more cost effective quite frankly I'm

even surprised by that and the reason

for that is

this remember artchip is only part think

of when you see a when you see computers

these days it's not a computer like a

laptop it's a computer it's a Data

Center and you have to operate it and so

people who buy and sell chips think

about the price of chips people who

operate data centers think about the

cost of

operations our time to deployment our

performance performance our utilization

our flexibility across all these

different applications in

total allows our operations cost they

call total cost of operations TCO our

TCO is so good that even when the

competitor's chips are free it's not

cheap

enough and that that is our goal to add

so much value that the alternative um is

not about cost and and so so we of

course of course that takes a lot of a

lot of hard work and we have to keep

innovating and things like that and we

don't take anything for granted but we

have a lot of

competitors as you know but maybe not

everybody in the audience knows there's

this term artificial general

intelligence which basically I was

hoping not to sound competitive but John