Keynote by NVIDIA CEO Jensen Huang at 2024 SIEPR Economic Summit

Stanford Institute for Economic Policy Research (SIEPR)
7 Mar 202455:21

Summary

TLDRIn a thought-provoking conversation, Jensen Huang, co-founder and CEO of Nvidia, discusses the transformative impact of artificial intelligence and accelerated computing. Huang shares insights on the future of AI, emphasizing the importance of multimodality and reasoning capabilities. He also addresses the challenges and opportunities presented by geopolitical risks in the tech industry, highlighting the rise of sovereign AI and the need for countries to control their digital intelligence. Furthermore, Huang touches on the role of Nvidia in custom solutions and the evolving landscape of computer programming, where prompt engineering becomes crucial.

Takeaways

  • πŸš€ Jensen Huang, co-founder of Nvidia, emphasizes the significance of artificial intelligence and accelerated computing as the defining technologies of the 21st century.
  • 🌟 Nvidia has been dedicated to a new form of computing called accelerated computing, aiming to solve problems that general-purpose computing is not equipped to handle efficiently.
  • πŸ’‘ The cost of computing has been reduced by 1 million times in the last decade, enabling large-scale AI applications like deep learning and data analysis.
  • 🧠 AI has reached a point where it can understand the meaning of digital knowledge, not just patterns, allowing for advancements in fields like gene sequencing and protein structure analysis.
  • πŸ”„ The future of AI involves continuous learning and real-time adaptation, with reinforcement learning loops becoming integrated into everyday operations and decision-making processes.
  • 🌐 Nvidia's GPU technology is foundational for the world's inferencing capabilities, with the company's platforms being the backbone for AI interactions globally.
  • πŸ› οΈ The importance of character and resilience over intelligence is highlighted, as these traits are seen as crucial for achieving greatness and success in both individuals and companies.
  • 🌍 Geopolitical risks impact Nvidia as a company, but also create opportunities for the development of sovereign AI capabilities in various countries.
  • πŸ’Ό The future may see Nvidia and other tech companies offering more customized solutions to clients, leveraging their existing ecosystems and technological advancements.
  • πŸŽ“ Jensen Huang advises aspiring entrepreneurs to embrace challenges and suffering as a means to build resilience and character, which are essential for long-term success.

Q & A

  • What is Jensen Wong's background and how did he become involved in the field of artificial intelligence?

    -Jensen Wong was born in Taiwan and moved to the US at age nine. He skipped two grades and graduated from high school early, then went on to study electrical engineering at Oregon State. He co-founded Nvidia at age 30 and has been instrumental in the development of accelerated computing, which has played a significant role in the advancement of AI technologies.

  • What is the significance of the transistor in the history of technology?

    -The transistor is considered one of the greatest inventions in the history of technology. It revolutionized the electronics industry by enabling the miniaturization of electronic components, which in turn led to the development of modern computing devices.

  • How has Nvidia's approach to computing contributed to the development of AI?

    -Nvidia has focused on a new form of computing called accelerated computing. This approach involves creating specialized hardware and software to solve problems that general-purpose computing is not well-equipped to handle, such as large-scale data processing and machine learning tasks that are fundamental to AI.

  • What is the role of AI in the future of technology according to Jensen Wong?

    -Jensen Wong believes that AI, enabled by accelerated computing, is the single greatest invention of the computer industry and will likely be the most important development of the 21st century. He sees AI as a transformative technology that will continue to evolve and improve, leading to continuous learning and real-time problem-solving capabilities.

  • What is the significance of Nvidia's H100 chip and its impact on data centers?

    -The H100 chip is a significant advancement in computing as it can replace an entire data center of old CPUs in a single unit. This has led to a dramatic reduction in the marginal cost of computing, enabling new possibilities in AI and data processing at scales previously unattainable.

  • How does Jensen Wong view the future of AI in relation to human intelligence?

    -Jensen Wong suggests that while we are making significant strides in AI, defining and achieving artificial general intelligence (AGI) that mirrors human intelligence is still a challenge. He emphasizes the need for clear specifications and tests to measure success in achieving AGI.

  • What is the role of AI in drug discovery and biology according to Jensen Wong?

    -Jensen Wong sees AI playing a crucial role in understanding the meaning of biological data. He envisions a future where AI can be used to understand the function of proteins, the meaning of genes, and the complex interactions within cells, which could revolutionize drug discovery and medical research.

  • How does Jensen Wong maintain motivation and inspire his employees at Nvidia?

    -Jensen Wong maintains motivation through a culture of transparency, constant feedback, and a focus on company values. He believes in empowering employees with information and opportunities, and fostering an environment where everyone is working towards a common goal.

  • What is Jensen Wong's perspective on the future of coding and programming?

    -Jensen Wong believes that while coding will remain important, the future of interacting with computers will be less about traditional programming languages and more about 'prompt engineering,' where users interact with AI through natural language, making technology more accessible to everyone.

  • How does Jensen Wong respond to the challenges of geopolitical risks in the tech industry?

    -Jensen Wong acknowledges the impact of geopolitical risks on the tech industry, particularly for companies like Nvidia that create essential AI tools. He emphasizes the need for companies to understand and comply with policies, and to be agile enough to adapt to changing geopolitical landscapes.

  • What is the potential impact of AI on the job market and economy?

    -Jensen Wong suggests that AI has the potential to close the technology divide by enabling more people to interact with and program computers through natural language. However, he also acknowledges that this shift could lead to changes in the job market, with new opportunities arising in areas like prompt engineering and AI interaction design.

Outlines

00:00

🎀 Introduction and Recognition

The speaker welcomes the audience and introduces Jensen Wong, highlighting his achievements in artificial intelligence and technology. The introduction includes a brief overview of John Chauvin's accomplishments and contributions to the community, as well as his role in starting the SE economic Summit 20 years ago. The speaker expresses gratitude towards John and looks forward to the conversation between Jensen and John.

05:00

πŸš€ Nvidia's Impact and Jensen's Background

The speaker discusses Nvidia's recent earnings announcement and its significance in the technology sector. Jensen Wong's background is detailed, from his early life and education to his co-founding of Nvidia. The speaker emphasizes the transformative impact of Nvidia and accelerated computing, and congratulates Jensen on his recent election to the National Academy of Engineering.

10:02

🌟 AI and the Future of Computing

Jensen Wong shares his perspective on the significance of AI and accelerated computing, comparing it to the invention of the transistor. He discusses the evolution of computing, particularly the reduction in the cost of computing and deep learning. Jensen explains how Nvidia's focus on accelerated computing has led to breakthroughs in AI, enabling the creation of large language models and the extraction of digital human knowledge.

15:02

πŸ’‘ Innovation in Chip Technology

The speaker and Jensen discuss the advancements in chip technology, specifically Nvidia's h100 and the upcoming h200. Jensen explains how these chips represent a reinvention of computing, reducing the marginal cost of computing and enabling new possibilities in AI and data processing. He also touches on the future of chip development and the potential for continuous learning and real-world application.

20:02

🌐 The Role of AI in Various Industries

Jensen elaborates on the application of AI in various fields, such as gene sequencing and protein structure analysis. He emphasizes the importance of understanding the meaning behind digital information and how AI can contribute to significant advancements in these areas. Jensen also discusses the potential for AI to improve itself through continuous learning and interaction with real-world data.

25:03

πŸ€– The Future of AI and Its Integration

The conversation turns to the future of AI, with Jensen discussing the potential for AI to develop human-like intelligence. He addresses the challenges in defining and achieving artificial general intelligence (AGI) and the importance of multimodality and reasoning capabilities. Jensen also shares his thoughts on the evolving role of AI in drug discovery and the broader implications for the understanding of biology.

30:04

πŸŽ“ Advice for Aspiring Entrepreneurs

Jensen offers advice to aspiring entrepreneurs, emphasizing the importance of resilience over high expectations. He shares his belief that success is rooted in character, which is built through suffering and setbacks. Jensen encourages students to embrace challenges and to approach their entrepreneurial journeys with a mindset of continuous learning and adaptation.

35:06

πŸ’Ό Employee Motivation and Company Culture

Jensen discusses his approach to maintaining a motivated and agile workforce, highlighting the importance of transparency, constant feedback, and a shared vision. He explains that his management style involves open communication and a focus on collective goals rather than individual performance reviews. Jensen believes that this culture of openness and empowerment is key to the company's agility and success.

40:10

🌍 Geopolitical Impact on the AI Industry

Jensen addresses the geopolitical risks associated with the AI industry, particularly in relation to Nvidia's products. He acknowledges the challenges posed by international regulations and the need for companies to adapt. However, he also sees these challenges as opportunities for growth, as countries around the world invest in their own sovereign AI capabilities. Jensen views the current geopolitical landscape as a catalyst for a broader global commitment to AI development.

45:10

πŸ›  Customization and Future Directions

Jensen discusses the potential for customizing solutions for customers, particularly those with unique needs or large scale. He explains that while significant customization can be costly, leveraging existing platforms and adding value is beneficial. Jensen is open to such collaborations, as long as they do not compromise the extensive investment made in developing the company's current technology ecosystem.

50:10

πŸ™ Closing Remarks

The session concludes with a final word of thanks from the speaker to John and Jensen for their insightful discussion. The audience is left with a deeper understanding of the AI industry's current state and its promising future.

Mindmap

Keywords

πŸ’‘Artificial Intelligence (AI)

AI refers to the simulation of human intelligence in machines that are programmed to think and learn like humans. In the video, Jensen Wong discusses AI as the biggest technological breakthrough in the last 76 years, highlighting its role in accelerated computing and its potential to understand and generate digital human knowledge.

πŸ’‘Accelerated Computing

Accelerated Computing is a form of computing where the processing power is increased to solve complex problems more efficiently. In the context of the video, Jensen Wong's company, Nvidia, has been dedicated to this form of computing, which has led to significant advancements in AI and deep learning.

πŸ’‘Innovation

Innovation refers to the process of introducing new ideas, methods, or products. In the video, Jensen Wong is portrayed as a pioneer in the field of AI and technology, having co-founded Nvidia and contributed significantly to technological innovation.

πŸ’‘Transistor

A transistor is a semiconductor device used to amplify or switch electronic signals and electrical power. It is a fundamental component in modern electronic devices. In the video, the transistor is mentioned as a great invention, setting the stage for the discussion on AI as the next biggest technological change.

πŸ’‘Deep Learning

Deep Learning is a subset of machine learning that uses artificial neural networks to model and solve complex problems. It is inspired by the structure and function of the human brain. In the video, deep learning is discussed as a domain where accelerated computing has significantly reduced costs, enabling new applications and breakthroughs.

πŸ’‘National Academy of Engineering

The National Academy of Engineering is a prestigious U.S.-based organization that recognizes outstanding engineering achievements and provides advice on engineering and technology issues. In the video, Jensen Wong's election as a member of the National Academy of Engineering is highlighted as a significant honor and recognition of his contributions to the field.

πŸ’‘Sovereign AI

Sovereign AI refers to the concept of a country or region developing and controlling its own artificial intelligence capabilities, rather than relying on foreign technology. In the video, Jensen Wong discusses the geopolitical implications of AI, emphasizing the importance of countries investing in their own sovereign AI to protect their interests and culture.

πŸ’‘Prompt Engineering

Prompt Engineering is the art of crafting inputs or prompts for AI systems to elicit desired responses or actions. It involves understanding how to interact with AI effectively to achieve specific outcomes. In the video, Jensen Wong suggests that in the future, prompt engineering will be a critical skill, allowing users to harness AI's capabilities without traditional programming knowledge.

πŸ’‘Geopolytics

Geopolytics refers to the political dynamics and strategies among nations and how they influence international relations and policy decisions. In the video, the impact of geopolitics on the AI industry is discussed, with Jensen Wong noting the challenges and opportunities that arise from countries seeking to control and develop their own AI technologies.

πŸ’‘Entrepreneurship

Entrepreneurship is the process of designing, launching, and running a new business, often involving innovation, risk-taking, and strategic management. In the video, Jensen Wong shares advice for aspiring entrepreneurs, emphasizing the importance of resilience and the ability to endure setbacks as key to success.

Highlights

The video is abnormal, and we are working hard to fix it.
Please replace the link and try again.

Transcripts

00:00

welcome back everyone after the short

00:03

break I know that many of you are

00:07

looking forward to hearing from our next

00:10

speaker Jensen

00:12

Wong Jensen is at The Cutting Edge of

00:16

artificial

00:17

intelligence and all of the

00:20

innovation

00:21

technology and human capital that is

00:24

needed to support

00:27

it my good friend and Seer colleague

00:30

John Chauvin is going to introduce

00:33

Jensen and I hope he's here somewhere so

00:36

I'm just going to keep talking and then

00:38

the two of them will have a conversation

00:41

before taking some of your

00:43

questions John chovin certainly requires

00:46

very little introduction to many most In

00:50

This Crowd as my predecessor as the Tron

00:53

director of seer John is the one who

00:55

started the SE economic Summit 20 years

00:58

ago so I would just like right now for

01:02

all of us to give John chovin a huge

01:04

round of applause and appreciate the

01:13

community that he had the foresight to

01:16

build uh for those of you who haven't

01:18

been touched by John's research his

01:20

mentorship or his friendship here's what

01:22

here's just a snippet of what you might

01:25

like to know about him along with being

01:27

the former Seer director and a Seer

01:28

Senor senior fellow Meritus John is the

01:31

Charles R Schwab professor of Economics

01:33

he is also a senior fellow at the Hoover

01:35

institution and a research associate of

01:37

the National Bureau of economic research

01:39

he specializes in public finance and

01:41

corporate finance and has published many

01:43

articles over the years on social

01:45

security health economics corporate

01:47

personal taxation mutual funds pension

01:49

plans economic demography applied

01:52

General equilibrium economics and much

01:54

more uh John isn't one for long

01:57

introductions but I just will say that

01:59

if I can be on10th as helpful to my

02:02

successor as John uh has been to me I'll

02:05

feel like I've uh succeeded so I will

02:07

let you read more about his Publications

02:09

and accomplishments in the programs

02:10

you've received uh today and so please

02:13

join me in welcoming our good friend

02:15

John Chauvin and I'm really looking

02:16

forward to

02:17

this

02:21

thanks wow thank you so I have always

02:26

thought that the more famous the speaker

02:29

the shorter the appropriate

02:32

introduction and if I was to follow that

02:36

rule I would stop right now and say

02:40

Jensen Wong but I'm not going to do

02:44

that

02:46

um so the Oxford English

02:50

Dictionary defines the American

02:53

dream believe it or not it does that and

02:57

it says that it's a situation where

03:00

everybody has an equal opportunity for

03:03

Success Through hard work dedication and

03:08

initiative and I would like to say that

03:11

Jensen Wong is an example of the

03:15

American dream

03:19

Jensen uh was born in

03:22

Taiwan came to the US at age nine with

03:26

his brother not with his

03:28

parents went to a rough tough School in

03:34

Kentucky survived that his parents came

03:37

two years later he moved to Oregon

03:40

skipped two grades and graduated from

03:42

high school and went to Oregon state

03:46

electrical engineering major 150 men and

03:50

two

03:52

women he said he was 16 he looked like

03:54

he was 12 he had no chance with the

03:58

women

04:00

well he sort of liked one of them and

04:04

said why don't we work on homework

04:07

together did that over and over and over

04:09

again six months later he after out for

04:12

a date well he's still married to her so

04:15

another American

04:20

Dream now to skip to age 30 he co-founds

04:26

Nvidia he's the only CEO there's ever

04:30

been of

04:32

Nvidia it's had its ups and its down

04:35

more UPS than

04:37

Downs it's now the fourth largest

04:39

company in the world third largest

04:43

American uh company so that sounds to me

04:48

like the American

04:50

dream um I should add that he also got a

04:54

degree from Stanford master's degree I

04:57

think he did it mostly at night

05:00

uh and he was always good with homework

05:01

at worked with his wife at worked with

05:03

Stanford uh

05:05

too um now of course we were here last

05:10

week Nvidia announced its

05:12

earnings in the finance

05:15

crowd this got more attention than the

05:17

Super Bowl that occurred a couple weeks

05:20

earlier it was pretty uh amazing uh his

05:24

company is at the absolute center of the

05:27

most exciting develop vment I'd say of

05:31

the 21st century technology development

05:34

and uh so he's to be congratulated on

05:38

that let me just say uh he's received a

05:43

lot of

05:44

awards a lot of recognition Enid has

05:47

received a lot of awards a lot of

05:50

recognition but I should have a short

05:52

introduction so I'm about to quit I'm

05:54

just going to talk about one

05:56

award last month he was elected as a

06:00

member of the National Academy of

06:03

engineering this is a pretty prestigious

06:07

award there are only three that I know

06:10

of I actually asked chat GPT I didn't

06:13

get an absolute clear

06:15

answer how many CEOs of S&P 500

06:19

companies are members of the National

06:21

Academy of engineering but I think it's

06:23

three and two are in this room anaru

06:28

Devan of Cadence Design Systems was

06:31

awarded it last year so the two of them

06:34

have that in common but let me now just

06:38

conclude and

06:40

congratulate Jensen not only on this

06:43

award but on the amazing success of your

06:46

company and thank you for speaking to us

06:49

today at Seer

06:55

Jensen how

06:57

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

07:00

here I guess so

07:03

okay so why don't you start off with

07:06

maybe some opening remarks and then I'll

07:07

ask you a few questions and then then

07:10

you get the tough questions well I think

07:11

that after your opening remarks uh it is

07:14

smartest for me not to make any opening

07:17

remarks to to uh uh avoid risking uh

07:23

damaging all the good things you said

07:26

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

07:29

always good to have a pickup line um and

07:32

mine was was uh do you want to see my

07:36

[Laughter]

07:40

homework and you're right we're married

07:42

still we have two beautiful kids I have

07:44

a perfect life uh two great puppies and

07:47

um I love my job and and uh she still

07:51

enjoys my

07:53

homework well if you want I can ask you

07:55

a few questions then yes please so if in

07:59

my lifetime I thought the biggest

08:02

technical development technology

08:04

breakthrough was the transistor now I'm

08:07

older than you yeah uh and it was pretty

08:10

fundamental deal but should I rethink is

08:13

AI now the biggest change in

08:17

technology that has occurred in the last

08:20

76 years to to hint at my age yeah um

08:24

well first first of all the the

08:27

transistor was obviously a great

08:29

invention but what

08:31

was the greatest capability that enabled

08:35

was

08:36

software the ability for humans to

08:39

express our ideas algorithms uh in a

08:42

repeatable way computationally

08:44

repeatable way uh was a was is the

08:47

Breakthrough um what have we done we

08:50

dedicated our company in the last 30

08:53

years 31 years uh to a new form of

08:55

computing called accelerated Computing

08:57

the idea is that general purpose

08:59

Computing is not ideal for every every

09:01

field of work and we said why don't we

09:04

in invent a new way of doing computation

09:07

such that we can solve problems that

09:09

general purpose Computing is ill

09:11

equipped at at

09:12

solving and and uh uh what we what we

09:15

have effectively done in in a particular

09:17

area of a domain of computation that is

09:20

that's that is algorithmic in nature

09:22

that can be paralyzed we've taken the

09:24

computational cost of computers to

09:28

approximately zero

09:30

so what happens when you when you uh are

09:33

able to take the marginal cost of

09:35

something to approximately zero some we

09:38

enabled a new way of doing software

09:40

where it used to be written by humans we

09:43

now can use computers to write the

09:45

software because the computational cost

09:48

is approximately zero and so you might

09:50

as well uh let the computer go off and

09:53

grind on just a massive amount of

09:55

experience we call data digital

09:58

experience human dig digital experience

09:59

called data and grind on it to find the

10:01

relationships and patterns that as a

10:05

result represents human knowledge and

10:09

that miracle happened about a decade and

10:11

a half ago we saw it coming and and we

10:13

took the whole company and we shaped our

10:14

computer which was already which was

10:16

already driving the marginal cost of

10:18

computing down to

10:19

zero and we pushed it into this whole

10:22

domain and as a result in the last 10

10:25

years we reduced the cost of computing

10:27

by 1 million times

10:31

the cost of deep learning by 1 million

10:32

times and a lot of people said said to

10:35

me but Jensen if you if you reduce the

10:36

cost of computing your your cost by a

10:40

million times then people buy less of it

10:42

and it's exactly the opposite we saw

10:44

that if we could reduce the marginal

10:45

cost of computing down to approximately

10:47

zero we might use it to do something

10:49

insanely amazing large language

10:52

models to literally extract all of

10:56

digital human knowledge from the

10:57

internet and put it into to a computer

10:59

and let it go figure out what the wisd

11:01

what the knowledge is that idea of

11:04

scraping the entire internet and putting

11:06

it in one computer let the computer

11:07

figure out what the program is is an

11:10

insane concept but you wouldn't ever

11:13

consider doing it unless the marginal

11:15

cost of computing was zero and so so we

11:18

made we made that breakthrough and now

11:20

we've enabled this new way of doing

11:22

software imagine you know for for all

11:24

the people that are still new to

11:26

artificial intelligence we figured out

11:27

how to use a computer to understand the

11:32

meaning not the pattern but the meaning

11:35

of almost all digital knowledge and

11:37

everything you can digit anything you

11:38

can digitize we can understand the

11:40

meaning so let me give you an example

11:41

Gene sequencing is digitizing genes but

11:46

now with large language models we can go

11:48

understand go un go learn the meaning of

11:51

that

11:52

Gene amino acids we

11:55

digitized you know through Mass Spec we

11:57

digitized

11:59

um Pro amino acids now we can understand

12:02

from the amino acid sequence without a

12:05

whole lot of work with cryms and things

12:06

like that we can go figure out what is

12:08

the structure of the protein and what it

12:09

does what is this meaning we can also do

12:12

that on a fairly large scale pretty soon

12:15

we can understand what's the meaning of

12:16

a cell a whole bunch of genes that are

12:19

connected together and this is from a

12:22

computer's perspective no

12:24

different than there's a a a whole page

12:28

of words and you asked it to what is the

12:31

meaning of it summarize what did it say

12:33

summarize it for me what's the meaning

12:35

this is no different than a hard you

12:37

know big huge long page of genes what's

12:39

the meaning of that big long page of

12:42

proteins what's the meaning of that and

12:44

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

12:47

just this is the miracle of of what

12:48

happened and so I would it's a

12:50

longwinded answer of saying John that

12:52

you're absolutely right that that that

12:55

that AI which was enabled by this form

12:58

this new form of computing we call

13:00

Accelerated Computing that took three

13:01

decades to do uh is probably the single

13:04

greatest invention of the computer of

13:07

the in of the technology industry this

13:09

will likely be the most important thing

13:11

of the 21st

13:13

century I agree with that 21st century

13:16

but maybe not the the 20th century which

13:18

was the transistor which it's got to be

13:20

close we'll let history decide that's

13:21

right we'll let history decide could you

13:23

look ahead you I I I take it that the

13:28

the GPU chip that is

13:31

behind uh artificial intelligence right

13:33

now is your h100 and I know you're

13:35

introducing an h200 and I think I read

13:38

that you plan to upgrade that each year

13:42

and so could you think ahead five years

13:45

March

13:46

2029 you're introducing the

13:49

H700 right what will it allow us to do

13:52

that we can't do

13:54

now um I'll go backwards but but let me

13:57

first say something about the chip that

13:59

John just described um as we say a chip

14:02

all of you in the audience probably

14:04

because you've seen a chip before you

14:05

you imagine there's a chip kind of like

14:07

you know like this um the chip that John

14:10

just described uh weighs 70

14:15

lbs it consists of 35,000

14:20

Parts eight of those parts came from

14:26

tsmc it that one

14:29

chip

14:31

replaces um a data center of old CPUs

14:35

like this into one

14:38

computer the savings because we compute

14:41

so fast the

14:44

savings of that one computer is

14:47

incredible and yet it's the most

14:49

expensive computer the world's ever seen

14:51

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

14:53

per chip we sell the world's first quar

14:56

million dollar chip but the system that

14:59

it replaced the cables alone cost more

15:01

than the chip this

15:03

h100 the cables of connecting all those

15:06

old computers that's the that's the

15:08

incredible thing that we did we

15:10

reinvented Computing and as a result

15:12

Computing marginal cost of computing

15:14

went to zero that's what I just

15:16

explained we took this entire data

15:18

center We Shrunk it into this one chip

15:20

well this one

15:21

chip uh uh is really really great at

15:25

trying to figure out um uh uh this form

15:29

this form of computation that that

15:31

without without

15:32

without getting weird on you guys um

15:35

call Deep learning it's really good at

15:37

this thing called Ai and so so uh the

15:40

way that this chip

15:42

works it works not just at the chip

15:44

level but it works at the chip level and

15:47

the algorithm level and the data center

15:49

level it works

15:51

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

15:54

its work by itself it works as a team

15:56

and so you connect a whole bunch of

15:58

these things together and it works at

16:00

you know networking as part of it and so

16:02

when you look at one of our computers it

16:04

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

16:07

only only computer Engineers would think

16:09

it's magnificent but it's magnificent

16:11

okay um it weighs a lot miles and miles

16:13

of cables hundreds of miles of cables

16:16

and and the next one's soon coming is

16:18

liquid cooled and you know it's

16:20

beautiful in a lot of ways okay and and

16:22

um uh and it computes at data center

16:26

scales and together what's going to

16:28

happen in the next 10 years say John um

16:31

we'll increase the computational

16:33

capability for M for deep learning by

16:36

another million times and what happens

16:38

when you do that what happens when you

16:40

do that um today we we kind of learn and

16:44

then we apply it we go train inference

16:47

we learn and we apply it in the future

16:50

we'll have continuous

16:52

learning We could decide whether that

16:55

whatever that continuous learning um

16:57

result it will be uh uh deployed into

17:01

you know the world's applications or not

17:03

but the computer will will watch videos

17:06

and and new text and uh from all the

17:09

interactions that it's just continuously

17:10

improving itself the learning process

17:14

and the Train the the training process

17:15

and the inference process the training

17:17

process and the deployment process

17:18

application process will just become

17:21

one well that's exactly what we do you

17:25

know we don't have like between now and

17:28

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

17:29

my learning and then after that I'll

17:31

just be doing inference you're learning

17:33

and inferencing all the time and that

17:35

reinforcement learning Loop will be

17:37

continuous and that reinforcement

17:39

learning will be grounded with real

17:41

world data that is been um uh through

17:44

interaction as well as synthetically

17:47

generated data that we're creating in

17:50

real time so this computer will be

17:53

imagining all the time does that make

17:55

sense just like just as when you're

17:57

learning you you take take pieces of

17:59

information and you go from first

18:00

principles it should work like this and

18:02

then we we do the the simulation the

18:04

imagination in our brain and that that

18:06

future imaginate imag imagin state in a

18:10

lot of ways manifests itself to us as

18:14

reality and so your AI computer in the

18:17

future will kind of do the same it'll do

18:18

synthetic data generation it'll do

18:20

reinforcement learning it'll continue to

18:22

be grounded by real world experiences um

18:25

it'll imagine some things it'll test it

18:27

with real world experience I'll be

18:28

grounded by that and that entire Loop is

18:30

just one giant

18:32

Loop that's what happens when you can

18:34

compute for a million times cheaper than

18:37

today and so as I as I'm saying this

18:40

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

18:42

when you can drive the marginal cost of

18:44

computing down to zero then there are

18:46

many new ways of doing something you're

18:48

willing to

18:49

do this is no different than I'm willing

18:52

to go further places because the

18:54

marginal cost of Transportation has gone

18:55

to zero I can fly from here to New York

18:57

relatively cheap cheaply if it would if

18:59

it would have taken a month you know

19:00

probably never go and so it's exactly

19:03

the same in transportation and all just

19:05

about everything that we do and so we're

19:07

we're going to take the marginal cost of

19:09

computing down to approximately zero as

19:11

a result we'll do a lot more

19:14

computation that causes

19:16

me as you probably know there have been

19:19

some recent stories that Nvidia will

19:24

face more competition in the inference

19:28

Market than it has in the training

19:30

Market but what you're saying is it's

19:33

actually going to be one market I think

19:35

can you comment about um you know is

19:39

there going to be a separate training

19:41

chip market and inference chip Market or

19:45

it sounds like you're going to be

19:47

continuously uh training and switching

19:50

to inference maybe within one chip I I

19:54

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

19:56

more well today today whenever you uh

19:58

prompt uh an AI it could be chat GPT or

20:02

it could be co-pilot or it could be uh

20:04

if you're using a surface nail platform

20:06

you using mid Journey um using Firefly

20:09

from Adobe whenever you're prompting

20:11

it's doing inference you know inference

20:14

is right so it's it's generating

20:16

information for you whenever you do that

20:18

what's behind it 100% of them is NVIDIA

20:20

gpus and so Nvidia most of the time you

20:23

engage our our our platforms are when

20:26

you're inferencing and so we are 100% of

20:29

the world's inferencing today is NVIDIA

20:31

now is inferencing hard or Easy A lot of

20:33

people the the reason why people are

20:35

picking on inferences when you look at

20:38

training and you look at Nvidia system

20:41

doing training when you just look at it

20:43

you go that looks too hard I'm not going

20:45

to go do that I'm a chip company that

20:48

doesn't look like a

20:49

chip and so there's a natural and you

20:52

have to in order for you to even prove

20:54

that something works or not you're $2

20:56

billion doll into it

20:58

yeah and you turn it on to realize it's

21:02

not very effective you're $2 billion in

21:04

two years into it the risk the risk of

21:07

exploring something new is too high for

21:09

the for the customers and and so a lot

21:11

of a lot of competitors tend to say you

21:13

know we're not into we're not into

21:14

training we're into inference inference

21:16

is incredibly hard let's think about it

21:18

for a

21:20

second the the the the response time of

21:23

inference has to be really high but this

21:24

is the this is the easy part that's the

21:26

computer science part the the E the hard

21:28

part of inference is the goal of

21:31

somebody who's doing inference is to

21:34

engage a lot more users to to apply that

21:37

software to a large install

21:40

base inference is an install base

21:42

problem this is no different than

21:44

somebody who's writing a an application

21:46

on on on an iPhone um the reason why

21:49

they do so is because iPhone has such an

21:51

large install base almost everyone has

21:52

one and so if you wrote an application

21:55

for that phone it's going to have the

21:57

benefit of it it's going to be able to

21:59

benefit everybody well in the case of

22:01

Nvidia our accelerated Computing

22:03

platform is the only accelerated

22:05

Computing platform that's literally

22:06

everywhere and because we we've been

22:08

working on it for so long if you wrote

22:09

an application for inference and you

22:12

take that model and you Deploy on

22:13

invidious architecture it literally runs

22:15

everywhere and so you could touch

22:16

everybody you can enable have greater

22:18

impact and so the problem with inference

22:20

is is actually install base and that

22:23

takes enormous patience and years and

22:27

years of success and dedication to

22:29

architecture compatibility you know so

22:31

on so

22:32

forth you make completely State

22:36

of-the-art chips is it possible though

22:39

that you'll face

22:42

competition that is claims to be good

22:45

enough not as good as Nvidia but good

22:48

enough and and much cheaper is that a is

22:50

that a threat well first of all

22:53

competition um we we have more

22:55

competition than anyone on the planet

22:57

has competition

22:58

uh not only do we have competition from

23:01

competitors we have competition from our

23:04

customers and um and and I'm the only

23:06

competitor to a customer um fully

23:09

knowing they're about to design a chip

23:11

to replace ours and I show them not only

23:14

what my current chip is I show them what

23:15

my next chip is and I'll show them what

23:17

my chip after that is and so and the

23:20

reason for that is because because look

23:22

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

23:24

attempt at uh uh explaining why you're

23:27

good at something

23:28

they'll never get a chance to to buy

23:30

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

23:32

completely open book in working with

23:34

just about everybody in the industry um

23:37

and and the reason the reason for that

23:39

our our advantage is several our

23:42

advantage what we're about is several

23:43

things whereas you could build a chip to

23:46

to be good at one particular algorithm

23:49

remember Computing is more than even

23:51

Transformers there's this idea called a

23:53

Transformers there's a whole bunch of

23:55

species of Transformers and their new

23:57

Transformers being invented as we speak

24:00

and the number of different types of

24:02

software is really quite quite rich and

24:05

the reason for that is because software

24:07

Engineers love to create new things

24:09

Innovation and we want that what Nvidia

24:11

is good at is that our our architecture

24:14

not only does it accelerate algorithms

24:16

it's programmable meaning that that you

24:18

can use it for SE we're the only

24:21

accelerator for SQL SQL was came about

24:24

in the

24:26

1960s IBM 1970s in storage Computing I

24:30

mean sqls structured data is as

24:32

important as it gets uh 300 zettabytes

24:35

of data being created you know every

24:36

couple of years Mo most of it is in sqls

24:39

structured databases and so so we're we

24:42

can accelerate that we can Accel

24:44

accelerate quantum physics we can

24:46

accelerate shortes equations we can

24:48

accelerate just about you know every

24:50

fluids particles um you know lots and

24:53

lots of code and so what Nvidia is good

24:56

at is the General field of accelerated

24:59

Computing one of them is generative Ai

25:02

and so for a data center that wants to

25:04

have a lot of customers some of it in

25:06

financial services some of it you know

25:09

some of it in in manufacturing so on so

25:11

forth in the world of computing we're

25:13

you know we're we're a great standard

25:15

we're in every single Cloud we're in

25:17

every single computer company and so our

25:19

company's architecture has become a

25:21

standard if you will after some 30

25:23

somewhat years and and so that's that's

25:25

really our advantage if a customer can

25:27

can um do something specifically that's

25:30

more cost effective quite frankly I'm

25:32

even surprised by that and the reason

25:33

for that is

25:35

this remember artchip is only part think

25:38

of when you see a when you see computers

25:40

these days it's not a computer like a

25:42

laptop it's a computer it's a Data

25:43

Center and you have to operate it and so

25:46

people who buy and sell chips think

25:48

about the price of chips people who

25:49

operate data centers think about the

25:51

cost of

25:52

operations our time to deployment our

25:57

performance performance our utilization

26:00

our flexibility across all these

26:02

different applications in

26:04

total allows our operations cost they

26:08

call total cost of operations TCO our

26:11

TCO is so good that even when the

26:14

competitor's chips are free it's not

26:17

cheap

26:18

enough and that that is our goal to add

26:22

so much value that the alternative um is

26:25

not about cost and and so so we of

26:28

course of course that takes a lot of a

26:30

lot of hard work and we have to keep

26:31

innovating and things like that and we

26:32

don't take anything for granted but we

26:34

have a lot of

26:36

competitors as you know but maybe not

26:38

everybody in the audience knows there's

26:40

this term artificial general

26:43

intelligence which basically I was

26:45

hoping not to sound competitive but John

26:47

asked a question that kind of triggered

26:50

a competitive Gene and I came AC I I

26:53

want to say I want to apologize I came

26:56

across you know if if you will a little

26:59

[Laughter]

27:02

competitive I apologize for that I could

27:05

have probably done that more

27:07

artfully I will next time but he

27:10

surprised me with a competitive I I I I

27:13

thought I was on an economic

27:17

Forum you know just walking in here I

27:19

asked him I'd sent some questions to his

27:22

team and I said did you look at the

27:24

questions he says no I didn't look at

27:26

the questions cuz I wanted to be

27:28

spontaneous besides I might start

27:30

thinking about it and then uh that that

27:32

would be bad so we're just kind of

27:34

winging it here um both of us um so I

27:38

was asking when when do you think and of

27:42

course it when do you think we will

27:44

achieve artificial general intelligence

27:47

the sort of human level intelligence is

27:51

that is that 50 years away is it five

27:54

years away what's your

27:56

opinion um I'll give you a very specific

27:59

answer but but first let me let me just

28:01

tell you a couple things about what's

28:02

happening that's super exciting first uh

28:05

of course of course um uh we're training

28:07

these models to be multimodality meaning

28:11

uh that we will learn from sounds we

28:12

will learn from uh words we'll learn

28:15

from uh vision and we'll just watch TV

28:17

and learn uh so on so forth okay just

28:19

like all of us and the reason why that's

28:21

so important is because we want AI to be

28:24

grounded grounded not just by human

28:27

value use which is what chat GPT um

28:30

really innovated I remember we had large

28:32

language models before but if it wasn't

28:34

until reinforcement learning human

28:36

feedback that human feedback that

28:38

grounds the AI to something that that we

28:42

feel good about human values okay um and

28:46

now could you imagine now you have to

28:48

generate images and videos and things

28:50

like that how does it the AI know that

28:53

hands don't penetrate through you know

28:55

podiums uh that feet stand above the

28:58

ground that when you step on water you

28:59

all fall into it so you have to ground

29:02

it on physics and so so now ai has to

29:05

learn um by watching a lot of different

29:07

examples and ideally mostly video uh

29:11

that certain be certain properties um uh

29:14

are are obeyed in in in the world okay

29:16

it has to create what is called a world

29:18

model and so so one we have to we have

29:21

to understand multimodality there's a

29:22

whole bunch of other modalities like as

29:24

I mentioned before genes and amino acids

29:26

and proteins and cells which leads to

29:28

organs and you know so on so forth and

29:30

so we would like to uh multim modality

29:33

second is um uh greater and greater

29:35

reasoning capabilities a lot of a lot of

29:38

the things that we already do uh

29:40

reasoning skills are encoded in common

29:42

sense you know Common Sense is reasoning

29:44

that we all kind of take for granted and

29:46

so there are a lot of things in our

29:48

knowledge in the internet that already

29:50

encodes reasoning and and and models can

29:52

learn that um but there's higher level

29:54

reasoning uh capabilities for example

29:57

example there's some questions that you

29:59

ask me right now when we're talking I'm

30:02

mostly doing generative

30:04

AI I'm not spending a whole lot of time

30:07

reasoning about the question however

30:10

there are certain problems like for

30:11

example planning problems where I'm

30:13

going to that's interesting let me think

30:15

about that and I'm cycling it in the

30:17

back and I'm coming up with the multiple

30:19

plans I've got I'm traversing a tree

30:22

maybe I'm going through my graph and you

30:23

know I'm I'm I'm pruning my tree and

30:25

saying this doesn't make sense but this

30:27

I'm going to put and I simulate it in my

30:29

head and maybe I do some calculations

30:31

and so on so forth that long thinking

30:34

that long thinking AI is not good at

30:37

today everything that you prompt into

30:39

chat gbt it responds instantaneously we

30:41

would like to prompt something into chat

30:43

gbt give it a mission statement give it

30:45

a problem and for it to think a while

30:48

isn't that right and so so that kind of

30:50

system you know what computer science

30:52

call system 2 thinking or long thinking

30:54

or planning those kind of things

30:57

reasoning reasoning and planning those

30:59

kind of problems I think we're going to

31:00

we're working on those things and I

31:02

think that you're going to see some

31:03

breakthroughs and so in the future the

31:04

way you're interact with AI will be very

31:06

different some of it will be just just

31:08

give me a question question and answer

31:10

some of it say here's a problem go work

31:12

on it for a while okay tell me tomorrow

31:15

and it it it does the the largest amount

31:17

of computation it can do U by tomorrow

31:20

you you could also say I'm going to give

31:22

you this problem U you know spend $1,000

31:24

on it but don't spend more than more

31:26

than that and it comes back with the

31:27

best answer within the Thousand or you

31:29

you know so on so forth okay so so

31:31

that's now

31:33

AGI the question on AGI is what's the

31:37

definition yeah in fact that's kind of

31:40

the Supreme question now if you ask me

31:43

uh if you say Jensen uh AGI is a list of

31:47

a list of tests and remember an engineer

31:50

can only know an engineer knows that

31:52

we've you know anybody in the in in that

31:56

you know prestigious organization that

31:58

I'm now powered of it knows for sure

32:00

about engineers is that you need to have

32:02

a specification and you need to know

32:04

what the definition of successes you

32:06

need to have a test now if I if I gave

32:09

uh an AI a lot of math tests and

32:11

reasoning tests and a history test and

32:14

biology tests and medical exams and bar

32:16

exams and you name it SATs and mcats and

32:19

every single test that you can possibly

32:21

imagine you make that list of tests and

32:24

you put it in front of put it in front

32:25

of the computer science Industry

32:27

I'm guessing in 5 years time we'll do

32:29

well on every single one of

32:32

them and so if your definition of AG is

32:36

that it passes human

32:38

tests yep then I will tell you five

32:42

years if you tell me but is it if you

32:45

asked it to me a little bit differently

32:47

the way you asked it that AGI is going

32:50

to be have human intelligence well I'm

32:53

not exactly sure how to specify all of

32:54

your intelligence yet and nobody does

32:57

really and therefore it's hard to

32:58

achieve as an engineer does that make

33:00

sense okay and so so the answer is we're

33:02

not sure and and um uh but we're we're

33:06

all endeavoring to make it you know

33:08

better and better so I'm going to ask

33:10

two more questions and I'm going to turn

33:12

it over because I think there's lots of

33:14

uh good questions out there the first

33:17

one I was going to ask about is could

33:19

you just dive a little deeper into what

33:23

you see as ai's role in drug discovery

33:29

the first role is to understand

33:32

understand the meaning of the digital

33:34

information that we

33:36

have right now we have we have all as

33:38

you know we have U uh we have a whole

33:41

lot of amino acids we can now uh because

33:43

of alpha fold um understand the protein

33:46

structure in many of them but the

33:47

question is now what is the meaning of

33:49

that

33:50

protein what is the meaning of this

33:52

protein what is this function uh it

33:54

would be great just as you can chat with

33:56

GPT

33:57

uh as you guys know uh there's you can

33:59

chat with a PDF you take a PDF file

34:02

doesn't matter what it is my favorites

34:04

are you take a PDF file of a of a

34:07

research paper and you load it into chat

34:09

G and you start at just talking to it

34:12

it's like talking to the

34:14

researchers is you know just ask what

34:17

what inspired this this research what

34:19

problem does it solve you know what was

34:21

the Breakthrough what what was the what

34:23

was the state- of art before then what

34:25

were the what were the novel ideas

34:27

just talk to it like a human okay in the

34:29

future want to take a protein put it

34:32

into chat GPT just like

34:34

PDF what are you

34:37

for what what enzymes activate you you

34:40

know what makes you

34:43

happy for

34:45

example there'll be a whole whole

34:48

sequence of genes and you're going to

34:49

take the and represents a cell you you

34:51

going to put that cell in what are you

34:52

for what do you do what are you good for

34:56

you know what do you hopes and dreams

34:59

and so so that that's that's one of the

35:01

most profound things we can do is to

35:04

understand the meaning of biology does

35:06

that make sense if we can understand the

35:07

meaning of biology as you guys know once

35:09

we understand the meaning of almost any

35:11

information that it's in the world the

35:12

computer science in the world of

35:14

computing amazing engineers and amazing

35:16

scientists know exactly what to do with

35:17

it but that's the Breakthrough the

35:20

multiomic multi multi-omic um

35:24

understanding of

35:25

biology and so that's if I could you

35:29

know deep and shallow answer to your I

35:32

think that's probably the single most

35:34

profound thing that we can do boy Oregon

35:37

State and Stanford are really proud of

35:40

you so if I could switch gears just a

35:42

little bit and just say Stanford has a

35:46

lot of

35:47

aspiring entrepreneurs students that are

35:51

entrepreneurs and maybe they're computer

35:53

science Majors or or engineering majors

35:56

of some

35:58

sort please don't build

36:01

gpus what what advice would you give

36:05

them uh to improve their chances of

36:09

success um you

36:12

know one one of my one of I think one of

36:16

my my great advantages is that I have

36:18

very low

36:20

expectations um and

36:24

and and I mean that um most of most of

36:29

the Stanford graduates have very high

36:32

expectations you you and you deserve to

36:34

have have expectations because you came

36:36

from a great school um uh you were very

36:40

successful you're on top of your top of

36:42

your class uh obviously you were able to

36:44

pay for tuition um and and uh uh and

36:49

then you're graduating from one of the

36:50

finest institutions on the planet you're

36:52

surrounded by other kids that are just

36:54

incredible you should have very you you

36:57

naturally have very high

36:58

expectations um people with very high

37:01

expectations have very low

37:07

resilience and unfortunately resilience

37:10

matters in

37:12

success I don't know how to teach it to

37:14

you except for I hope suffering happens

37:16

to

37:17

you and and uh I I was fortunate that I

37:21

grew up with a with a with you know with

37:24

my parents um

37:27

uh uh providing a condition for us to be

37:30

successful on the one hand um but there

37:33

were plenty of plenty of opportunities

37:35

for setbacks and suffering and um you

37:38

know and and to to this day I use the

37:40

word the phrase pain and suffering

37:42

inside our company with great Glee and

37:44

the reason and I mean that you know boy

37:47

this is going to cause a lot of pain and

37:48

suffering and I mean that in a happy way

37:51

um because because you want to train you

37:53

want to refine the character of your

37:55

company you want want that you want

37:57

greatness out of them and greatness is

37:59

not intelligence as you know greatness

38:01

comes from character and character isn't

38:03

isn't formed out of smart people it's

38:05

formed out of people who

38:07

suffered and and so so that's that's

38:10

kind of and so if I could if I could

38:12

wish upon you I don't know how to do it

38:15

but you know for all of you Stanford

38:17

students I I wish upon you you know

38:20

ample doses of pain and

38:25

suffering

38:30

I'm going to back out of my promise and

38:32

ask you one more

38:34

question how do you you seem incredibly

38:38

motivated and energetic but how do you

38:41

keep your employees motivated and

38:44

energetic when they probably become

38:46

richer than they ever expected

38:49

to I'm surrounded I'm surrounded by 55

38:52

people my management team so you know my

38:54

I I have a man my management team my

38:56

director reports is 55

38:59

people um uh I write no reviews for any

39:02

of them I give them constant

39:06

reviews uh and they provide the same to

39:08

me uh my compensation for them uh is the

39:13

the bottom right corner of excel I just

39:16

drag it down

39:18

literally many of our executives are

39:21

paid the same exactly to the

39:24

dollar I know it's weird

39:27

it works and and uh I don't do one-on

39:30

ones with any of

39:32

them unless they need me then I'll drop

39:35

everything for

39:36

them uh I never have meetings with them

39:39

just alone and they never hear me say

39:42

something to them uh that is only for

39:44

them to

39:45

know there's not one piece of

39:47

information that I that I somehow

39:50

secretly tell eaff that I don't tell the

39:52

rest of the company um uh and so in in

39:56

that in that way our company was

39:58

designed for agility for information to

40:01

be to flow as quickly as possible uh for

40:04

people to be empowered by what they are

40:05

able to do not what they know um and uh

40:09

I and so that that's the architecture of

40:11

our

40:13

company um I don't remember your

40:15

question but but oh oh oh oh oh oh oh I

40:19

got it I got it I got it I got it uh and

40:22

the the answer the answer for that is my

40:24

behavior yeah

40:27

the it's uh how do I celebrate success

40:29

how do I celebrate failure how do I talk

40:31

about success how do I talk about

40:33

setbacks um every single thing that I'm

40:36

looking for opportunities to instill

40:38

every single day I'm looking for

40:39

opportunities to to keep on uh

40:42

instilling the culture of the company

40:43

and what is important what's not

40:45

important what's the definition of good

40:47

how do you compare yourself to good how

40:49

do you think about good um uh how do you

40:52

think about a journey how do you think

40:53

about results uh all of that all day

40:56

long

40:57

Mark dougen can you help us okay good so

41:01

let's open it up uh for some questions

41:03

let me start with Winston and I'll come

41:05

to

41:05

you oh we need a microphone can you just

41:08

Ben you got this

41:13

yeah board member Winston I have a

41:15

couple question what's a story about

41:17

your leather

41:20

jacket and the second the second is

41:23

according to your projection and

41:25

calculation

41:27

in 5 to 10 years how much more

41:30

semiconductor manufacturing

41:33

capacity is

41:34

needed to support the growth of

41:39

AI okay uh I appreciate two questions um

41:43

uh the the uh the first question is this

41:45

is what my wife bought for me and this

41:47

is what I'm

41:48

[Laughter]

41:51

wearing and and because I do I do 0% of

41:55

my own shopping

41:56

uh as soon as something doesn't as soon

41:58

as she finds something that doesn't make

42:00

me

42:01

itch because she knows she's known me

42:04

since I was 17 years old and she thinks

42:06

that everything makes me itch and the

42:09

way I say I don't like something is it

42:10

makes me

42:11

itch and so as soon as she finds me

42:14

something that doesn't make me itch if

42:16

you look at my closet the whole closet

42:18

is a

42:19

shirt because she doesn't want to shot

42:21

for me

42:23

again and so so that's why uh this is

42:27

all she bought me and this is all I'm

42:29

wearing and if I if I don't like the

42:31

answer I can go shopping otherwise I

42:33

could wear it and it's good enough for

42:36

me we second question on this the

42:39

forecast is actually very this is very

42:42

I'm horrible at

42:43

forecasting but I'm very good at first

42:46

principled reasoning of the size of the

42:51

opportunity and so let me first reason

42:53

for you um uh I have no idea how many f

42:56

ABS but here's here's the thing that I

42:58

do know the way that we do Computing

43:00

today the the the information was was

43:05

written by someone created by someone

43:07

it's basically

43:09

pre-recorded all the words all the

43:11

videos all the sound everything that we

43:13

do is retrieval based it was

43:15

pre-recorded does that make sense as I

43:18

say that every time you touch on a phone

43:20

remember somebody wrote that and stored

43:21

it somewhere it was

43:23

pre-recorded okay every modality that

43:25

you know

43:26

in the

43:27

future because we're going to have

43:30

AIS it understands the current

43:33

circumstance and because it can it's

43:35

tapped into all of the world's you know

43:37

latest news and things like it's called

43:38

retrieval based okay and it understand

43:41

your context meaning it understood why

43:43

you asked what you're asking about when

43:46

you and I ask about the economy we

43:48

probably are meeting very different

43:50

things and for very different context

43:54

and based on that it can generate at

43:56

exactly the right information for you so

43:58

in the future it already understands

44:01

context and most of computing will be

44:05

generative in the today 100% of content

44:09

is

44:10

pre-recorded if in the future 100% of

44:13

content will be generative the question

44:15

is how many how does that change the

44:17

shape of computing and so without

44:20

torturing you anymore um I'll that's how

44:23

I reason through things how much more

44:26

networking do we need more less of that

44:27

do we need memory of this and and the

44:29

answer is we're going to need more

44:32

Fabs however uh remember that we're also

44:35

improving the algorithms and the

44:37

processing of it um tremendously over

44:40

time it's not as if the efficiency of

44:43

computing is what it is today and

44:45

therefore the demand is this much in the

44:47

meantime I'm improving Computing by a

44:48

million times every 10 years while

44:50

demand is going up by a trillion

44:53

times and that has to offset each other

44:55

does that make sense and then there's

44:57

technology diffusion and so on so forth

45:00

that's just a matter of time but it

45:02

doesn't change the fact that one day all

45:05

of the computers in the world will be

45:06

changed 100% every single data center

45:10

will be all of those general purpose

45:12

Computing data centers 100% of the

45:14

trillion dollars worth of infrastructure

45:15

will be completely changed and then

45:17

there'll be new infrastructure built on

45:19

even on top of that okay next question

45:22

right here

45:23

Ben and then over here to Rand so yeah

45:27

thanks for coming today so recently you

45:29

said that you encourage students not to

45:31

learn how to code yeah um and that's the

45:34

case it means one of maybe a few things

45:36

but do you think the world starts to

45:38

look like from a company formation an

45:40

entrepreneurship perspective that it

45:42

goes towards many many more companies

45:45

that are created or do you think it's

45:48

consolidation to just a number of the

45:50

big big players so so first of all um I

45:53

I I said it so poorly that you repeat it

45:55

back

45:56

poorly I I didn't if you would like to

45:59

code for God's sakes code okay if if you

46:03

want to make omelets make omelets I'm

46:04

not not you coding has coding is a

46:08

reasoning process it's

46:11

good does is it going to guarantee you a

46:14

job no not even a little

46:16

bit uh the the number of coders in the

46:18

world uh surely uh will continue to to

46:22

uh uh be important and we Nvidia needs

46:25

coders

46:26

however in the

46:27

future the way you interact with the

46:29

computer is not going to be C++ mostly

46:32

for some of us that's true for some of

46:33

us that's but for you you know why why

46:36

programming python so weird in the

46:40

future you'll tell the computer what you

46:42

want and the computer will will you you

46:45

say hi I would like you to come up with

46:47

a uh a build plan with all of the

46:49

suppliers and build a material for a

46:51

forecast that we have for you and based

46:54

on all of the equip all the necessary

46:56

components necessary coming up with a

46:57

bill plan okay and then if you if you

47:00

don't like that you write me a Python

47:03

program that I can modify of that bill

47:06

plan and so remember the first time I

47:09

talk to the computer I'm just speaking

47:11

in plain English the second time so

47:14

English by the way human is the best

47:16

programming language of the

47:18

future how you talk to a computer how do

47:21

you prompt it how do you prompt it it's

47:23

called prompt engineering how you

47:25

interact with people how do you interact

47:27

with computers how do you make a

47:28

computer do what you want it to do um

47:30

how do you fine-tune uh the instructions

47:33

with that computer that's called prompt

47:34

engineering there's an there's an

47:35

Artistry to that okay so for example

47:39

most people are surprised by this but

47:40

it's it's not surprising to me but but

47:42

it's surprising for example you ask mour

47:44

to generate a pcture an image of a puppy

47:47

on a on a surfboard um uh uh in Hawaii

47:51

uh at Sunset okay and then and then and

47:55

it generates one and go and you say oh

47:57

more

47:58

cute make it more cute and it comes back

48:02

it's more cute and you go no no cuter

48:04

than that and it comes back why is it

48:08

that software would do that there's a

48:09

there's a structural reason why it does

48:11

that but for example you need to know

48:13

that that that capability exists in a

48:15

computer in the future isn't that right

48:17

that you if you don't like the answer

48:18

first time you could you can find tuna

48:20

and get it to within the context that

48:22

you you know you can make it give you

48:24

better and better results and once you

48:26

you can even ask it to write the program

48:28

Al together to generate that result in

48:29

the future and so my point is that

48:33

programming has has changed in a way

48:36

that is probably less valuable on the

48:38

other hand let me I will tell you this

48:41

that because of artificial intelligence

48:43

we have closed the technology divide of

48:45

humanity today about

48:49

about 10 million

48:51

people are gainfully employed because we

48:54

know how to program

48:58

computers which leaves the other 8

49:00

billion

49:02

behind that's not true in the future we

49:05

all can program computers does that make

49:07

sense you all know how to prompt a

49:09

computer to make it do things and look

49:11

at all you to do is look at YouTube and

49:13

look at all the people who are using

49:15

prompt engineering all the kids and you

49:17

know who are making a do amazing things

49:19

they don't know how to program they're

49:20

just talking to chat

49:22

GPT they just know that if I tell it to

49:24

do this if do that you know and so it's

49:27

no different than interacting with

49:29

people in the future that's that's the

49:31

great contribution we've the computer

49:33

science Industry has made to the world

49:34

we've closed the technology divide so

49:38

that's that's inspiring okay over here

49:41

we've got that sounds very we've got

49:43

Randy with a question right over here oh

49:45

um thank you very much I'm just

49:46

wondering um about do you think very

49:49

much about geopolitical risk and um how

49:53

do you see it impacting your industry if

49:56

you

49:57

do uh geopolitical risk you know we we

50:00

are almost a poster child of

50:02

geopolitical

50:03

risk and the reason for that is because

50:06

uh we make a very important instrument

50:08

for artificial intelligence and

50:10

artificial intelligence as John and I

50:11

were talking about earlier is the

50:13

defining technology of this of this of

50:15

this

50:16

time and and um and so the United States

50:21

uh has every right to determine that

50:24

this instrument should be limited to uh

50:27

to uh countries that that it determines

50:30

that uh it should be limited limited uh

50:33

with and so so the United States have

50:35

has that right and they they exercise

50:37

that right um and your question has to

50:41

do with what is the implication to us I

50:45

uh we first of all we we just have to

50:46

understand these policies and we have to

50:48

stay agile so that we can comply with

50:50

the policies uh number one on the one

50:53

hand it limits our opportunity and in

50:55

some places and it it opens up

50:57

opportunities in others one of the

50:58

things that has happened in the last I

51:01

would say maybe even 6 to n months is

51:04

the Awakening of every single country

51:07

every single Society The Awakening that

51:10

they have to control their own digital

51:15

intelligence that India can't Outsource

51:17

its data so that some country transforms

51:21

that Digital Data into India's

51:24

intelligence and imports that

51:26

intelligence back to India that

51:29

Awakening that Sovereign AI that you

51:31

have to you have to dedicate yourself to

51:33

control your Sovereign AI your Sovereign

51:36

intelligence protect your language

51:38

protect your culture for your own

51:40

industries that Awakening I think

51:43

happened in the last 6 nine months the

51:45

first part was we have to be we have to

51:47

be mindful about safety then the second

51:49

part was hold on a second we we all have

51:52

to do this and so every single country

51:54

from from India um uh Canada's doing

51:57

this uh the UK France um Japan uh

52:01

Singapore Malaysia uh the list goes on

52:05

uh just about every single country now

52:06

realize that they have to invest in

52:09

their own Sovereign AI so geopolitics in

52:12

the one hand limited opportunities but

52:14

it created just enormous opportunities

52:16

elsewhere and so hard hard to say okay

52:20

so I think we I have multiple hands but

52:22

I have time for one more question I am

52:24

going to go

52:25

right here you had to you were further

52:27

on the now remember the last question

52:29

has all big pressure you guys agree with

52:32

that do you can we all agree right here

52:34

the the person who La asked the last

52:36

question don't don't leave us all

52:41

depressed I'm going to don't trigger me

52:43

please I'm I'm that's all I'm saying I'm

52:46

just kidding I'm going to invoke your

52:49

commandment to have low expectations at

52:51

this

52:52

juncture um you you mentioned your

52:55

competing with your customers and I'm

52:57

wondering you know given the advantages

52:59

that you have why they're doing that and

53:01

I'm wondering if in the future you see

53:03

yourself building more customized

53:06

solutions for customers of a certain

53:09

scale um as opposed to you know uh the

53:13

solutions that you have now which are

53:14

more

53:15

horizontal uh the the so so are we

53:19

willing to customize the answerers yes

53:21

now why is it that the bar is relatively

53:23

High the the reason why the bar is high

53:25

is because each generation of our our

53:27

platform first of all there's a GPU

53:29

there's a CPU there's a networking

53:32

processor there's a SW there two types

53:34

of

53:35

switches I just build five chips for one

53:38

generation people thinks it's one chip

53:39

but it's five different chips each one

53:41

of those chips are hundreds and hundreds

53:43

of millions of dollars to do just

53:45

hitting launch which is tape out for us

53:48

launching a rocket is several hundred

53:49

million dollar each time okay I I got

53:52

five of them per generation then you've

53:54

got to put them into into a system and

53:56

then you got to put you know you got

53:57

networking stuff you got C transceiver

53:59

stuff you got optic stuff you got a

54:01

mountain of software to do it takes a

54:03

lot of software to run a computer as big

54:05

as this room and so so all of that is

54:09

complicated if I if if the customization

54:13

is so

54:14

different then then you have to repeat

54:17

the entire R&D however if the

54:19

customization leverages everything and

54:22

adds something to it then it makes it's

54:25

makes a great deal of sense maybe it's a

54:28

it's a proprietary security system maybe

54:30

it's a confidential Computing system

54:33

maybe it's a a a new way of doing uh

54:36

numerical processing um that that could

54:40

be extended we're very open-minded to

54:42

that and the custo our our customers

54:44

know that I'm willing to do all that and

54:46

recognizes the the the if you change it

54:49

too far you've basically reset and

54:52

you've squandered you know the the

54:55

nearly hundred billion dollars that's

54:57

taken us to get here um uh to to redo it

55:00

from from scratch and so they want to

55:02

leverage our ecosystem to the extent

55:04

that that that that will be done I'm

55:06

very open to it yeah and they know and

55:08

they know that

55:10

yeah okay so with that I think we need

55:12

to wrap up thank you so much to John and

55:19

Jensen