Jensen Huang Nvidia Morgan Stanley TMT Conference

Wow.

>> No music. No walk on music. No

no no no no roaring applause.

>> Um I'm just I'm just saying that I'm I'm not I'm not used to coming to work in this way. This total silence.

this way. This total silence.

>> I'm just kidding.

>> There there were a lot of Taylor Swift comments along the way. So the crowd is ready.

>> Needs humor, too. Is humor allowed here?

>> Humor is very allowed. I made investment banking jokes yesterday, Jensen. But uh

thank you uh for being here for the last I think 25 27 years. You've been such a great uh supporter uh of this

conference. I think we sometimes become

conference. I think we sometimes become uh numb to the scale of the numbers and the transformation we're experiencing. I

don't think I'm the only one in this audience. I'm getting billions and

audience. I'm getting billions and trillions confused uh constantly. My

partner Mark Edlestone and I 27 years ago, we sat on a stage much smaller than this one on the Morgan Stanley trading

floor and we announced uh and introduced Nvidia and um and you to the Morgan Stanley Salesforce and uh believe it or

not, $48 million IPO, 1998 revenue n trailing revenue, $30 million. Um,

Jensen and his team, Colette, were so generous. Two years ago, you hosted our

generous. Two years ago, you hosted our board meeting in your headquarters. I

think you had just announced a $30 billion quarter uh in terms of revenue and then last week uh a $46 billion net

income quarter. So we've moved from

income quarter. So we've moved from years to quarters, from millions uh to uh billions. It's really amazing and

uh billions. It's really amazing and unprecedented scale and growth. Um and

then you've changed our lives. You've

changed our lives. And so I guess my question after that is what had to come together strategically, culturally, technically to deliver that

type of hyperrowth at scale. And the

scale is really astounding. And again,

thank you.

>> That's going to take 37 minutes and 13 seconds and slightly more. Uh, you know, obviously Nvidia wasn't built overnight.

It's taken us 33 years. I I sort of remember somehow that when we went public, our price was $13 and I just read here is 12. I overstated them. I

remembered it to be much more optimistically than it was actually. The

the company's valuation at the time, I think, was like $300 million.

And and uh Mark did such a good job.

Mark Edlesen did such a good job preparing all of our investors that that they really only had one question. It

was literally a one question IPO road show and the the question was when are you going out of business?

I'm not kidding. It's and that that's that exact that question uh is about as hard to answer as the one you just gave me.

Um well the the the answer is the answer is um as it turns out we started the company with the idea of creating a new computing platform a new way of doing

computing and not not that the old way was wrong. It's just that the new way um

was wrong. It's just that the new way um a new way is is essential to solve some unique problems and the type of things that we were extremely good at are

algorithms. algorithms because the inner loop of the software tends to be about 5% of the code but 99% of the compute

time and and back then the algorithms in the world of computers was quite rare and one of the most important algorithms was computer graphics the simulation of

light and how light travels through through space and so so uh computer graphics uh was used for uh things like animation movies of course at the time

that were founded the cover of I forget which magazine was you know Jurassic Park was there and so it was really it was during that time where computer

graphics was was becoming uh more more capable and we could simulate you know virtual reality with it and we applied it to creating a a a new industry uh

which did not exist at the time uh called uh uh video games and so 3D graphics uh was was modernized in my

time uh consumerized in my time and and the whole video game industry was created in my time. And when I say in my time, meaning it was it was Nvidia that Paul pulled it all together.

We the reason why we're so beloved in the video game industry and we're so deep in it still is because in a lot of ways we created the modern video game industry from the algorithms associated

with it, the libraries, you know, the uh in in the computer graphics industry without RTX. there would be nothing

without RTX. there would be nothing today without our contribution of all the algorithms that goes into all of the game engines. Uh you wouldn't be able to

game engines. Uh you wouldn't be able to enjoy the type of video games you enjoy today. So Nvidia has been deep in the

today. So Nvidia has been deep in the world of algorithms since day one 33 years ago. Now accelerated computing

years ago. Now accelerated computing requires uh what is described as a full stack meaning the architecture the chip

design the libraries that sit on top of it how it's integrated forwardly you know I I'm using that apparently there's this new idea called for deployed engineers or something like that

Nvidia's had devte engineers 33 years ago we deployed them into the world's video game industries and video game companies in game engines and we

integrate our technology into their game engine. Today, if you look at Epics's

engine. Today, if you look at Epics's Unreal Engine, Nvidia's technology is all over it and and uh uh you go into every game developer, Nvidia's technology is all over it. That's the

reason why all the all the games run best on Nvidia for good reason. That's

the reason why Nvidia is the world's largest game platform. Uh you you probably don't don't know this, but there's several hundred million active GeForce gamers in the world. many of

them turned into AI researchers. It is

because of GeForce GTX 580 that that uh uh you know that Ilia Sus Cescover and uh um uh Alex Kasheeski uh and Jeff

Hinton, it was Jeff that told him to go buy it to discover CUDA. And so so I the the first idea about Nvidia is that we're full stack company. The second

idea about our company and this is this is you know really old history that long many people might might not have been born yet but during that time the PC the

PC architecture was incompatible with today's computer graphics capabilities and we created some new technology called direct NVIDIA it was a way for

applications to directly communicate with our our APIs and uh uh we exposed it to to uh some very important companies it became direct X. Uh if you

look at the way that we uh communicate between us and the application uh that was completely revolutionary to bypass uh uh a whole bunch of software that

makes it slow by to make accelerated computing possible. Uh we uh introduced

computing possible. Uh we uh introduced the idea of virtualized uh frame buffer memory into system memory. It was

initially called AGP which then became PCI Express. um many of the system

PCI Express. um many of the system architecture had to be reinvented so that we could accommodate video games uh and uh 3D graphics in a PC. Well, that

same sensibility of both innovating the full stack to be integrated into in algorithms as well as changing the architecture of systems so that we could

create new computer systems led to that same sensibility expertise led to DGX1 which was the world's first AI supercomputer that had delivered you

know by hand to San Francisco here very close by to a company that eventually became open AAI and so so the the fundamental attitude, if you will, expertise, um,

how we see the world propagated in this way. It's literally

33 years. The company's entire culture is designed to be full stack. The

organization is designed to be full stack. The entire system uh is designed

stack. The entire system uh is designed to create new new stacks and new system architectures uh that allow us to do

this. Well, we started with of course

this. Well, we started with of course the the if you look at Nvidia's graphics cards, GeForce, it's a technology marvel. How it's integrated into the

marvel. How it's integrated into the operating system, how it's integrated into the system architecture, completely reinvented how computers were before.

Well, we have no trouble with that with DGX1. I have no trouble with that with

DGX1. I have no trouble with that with the first supercomputing cluster, which then went to Satia um uh for their first supercomput. And

you you might you might uh uh you know people noticed that Microsoft first supercomputer and Nvidia's supercomputer had exactly the same benchmark like down

to the you measure that you measure the performance of the system across all of these GPUs that that was about 10,000 GPUs or so. It was exactly the same performance. The reason for that is

performance. The reason for that is because we designed it and we delivered to to Azure cloud. Uh it was all based on Infiniband. It was all based on based

on Infiniband. It was all based on based on uh uh based on um uh Ampear a this was the A100 which became the first computer that open AAI used and so um

we're quite comfortable with this full stack full system approach and without being able to do that it is impossible to stay at the bleeding edge. It is

literally impossible to keep up with a company that's building not just one chip each year, but we're building an entire infrastructure each year because we we own the CPU. We revolutionized the

new way of designing CPUs. And you'll

see you'll see more examples of that. Uh

we re revolutionized the way we do CPUs, revolutionized the way we obviously do GPUs, connect them together using this thing called MVLink, which revolutionized the way you built

computers all together. uh connected

together with a new type of AI Ethernet called Spectrum X. We connected

everything together. Now we own the entire stack. We know all the chips

entire stack. We know all the chips inside. When you own the entire stack

inside. When you own the entire stack and you own all the chips inside, you could change it every single year. If

you don't own the entire stack and you don't own all the chips, it's hard to innovate every year. And the reason for that is because you're com you're connecting too many cats and dogs and

there's too much innovation to pull together once a year if you can't control it because it's a full stack problem. So that's how we got here. It's

problem. So that's how we got here. It's

amazing in the la in the last two years since you were here last and our board meeting we've sort of gone from generative AI uh models to reasoning and

now agentic and Satcha just finished a panel on the enterprise and at the enterprise level you know we're working with Microsoft the open AI XAI Gemini we

had Daario here from Anthropic the capabilities are extraordinary uh what does it mean around the size of that enterprise market, how is it changing

and how is it going to be adopted and how do you sort of see that playing out uh over the years because it's a big big topic at the conf.

>> Yeah, really good. Um literally in the last two years we went through three inflection points in AI. Uh the f first inflection point first of all the

technology sat there in plain sight for months. GPT3 sat there in plain sight

months. GPT3 sat there in plain sight for months until somebody wrote essentially a wrapper around it and turned it into chat GPT turned it into an API u made it a available and ease

easy to use by everybody. But the first inflection point was generative as you mentioned the ability to translate convert information from one form to

another form and auto reggressively generate tokens and and the the second but of course the problem with generative AI is that that it's prone to

hallucinate and and the reason for that is because not because there's something fundamentally wrong with the technology not because it didn't it didn't learn all the right things but because it's not grounded on contextual information.

it's not grounded on re relevant information and so so uh the second thing that happened was 01 and uh reasoning came about but behind 01 is

also grounding on research grounding on on truth and um the ability to have to combine generative with semantic um uh

we call it retrieval augmented generation um but basically conditional generation okay conditional generation meaning that what you're about to

generate depends on context and ground truth or whatever research or whatever it is. And so so the second generation

it is. And so so the second generation uh uh introduced reasoning, self-reflection, um the ability to self-correct because sometimes what comes out you know your

mouth you kind of wish that you pulled back and you go ah you know and so in the case of AI it has the ability to do that in real time and so 01 became much more uh grounded and the information

that was generated was more reliable. So

what happened um what came out as a curiosity and incredible excitement um and the and the tech industry jumping onto it because we realized what's about to what can happen the next phase of it

the usefulness of chatpt just skyrocketed um but the amount of tokens that it generated was much much more than the first generation maybe you know a 100 times more tokens the model is

maybe 10 times larger so it's probably something like a thousand times more compute so from from 01 one over Chad GBT call it a thousand times and then

because it was so useful maybe a million times more usage.

Okay. So the combination of usage um uh uh uh and and its usefulness um and groundedness allows us to to uh we saw

that next phase of growth. But in the end, in the end, what 01 did was it provided information

essentially um a chatbot that was uh much more much more factual.

It was informational and of course for many of us we use it for research and we use it all the time uh instead of searching you know our goal isn't to search our goal is to get answers and so

so uh chat GPT gave us that that was kind of the second inflction the inflection that we're we're seeing here also sat in plain sight for quite a long

time and um uh and it's basically the ability for AI to use files access files and use tools and so now it could reason it think it could use tools. it can

solve problems and uh it could do search, it could do planning and so um probably the the biggest phenomenon that's happening and if you're paying

attention to it I'm sure you are open claw is probably the single most important um uh release of software you know probably ever and if you look at

look at open claw uh and the adoption of it uh you know Linux took right some 30 years to reach this level uh openclaw saw in what is it 3 weeks has now

surpassed Linux.

It is now the single most downloaded opensource software in history and it took 3 weeks. If you look at the line, you know, even in semi- log, this thing

is straight up. It's vertical. It looks

like the it looks like the the y- axis.

I've never seen anything like it. Okay,

it look really looks like the y- axis.

And um uh and so what's happening now that you could give a problem statement um create start where the the prompt goes create you know the last prompt the

way you kind of think about it the last prompt was what is when is who is right that's the last prompt this now prompt

goes create do build write does that make sense so what's happened the last prompt was queries.

This prompt are actions. They're tasks.

Do something for me. And you describe it as, you know, expressively as you like as with with a lot of intention, you know, and let it let it infer or very

specific and it goes off and it just churns. It just thinks it goes off and

churns. It just thinks it goes off and it does research and it reads it reads a manual. Um, if it has to use a tool,

manual. Um, if it has to use a tool, it's never used it before, it reads the manual of the tool.

it goes off and studies what's on the web and and uh um and it you know applies the tools and performs the task.

Now I just said we went from one you know one generative prompt uh one generative uh response to now one that

is a thousand times more tokens and uh agents you know we call them at the company claws. These claws are now consuming what a million times more

tokens. They're running continuously in

tokens. They're running continuously in the background. We have a whole bunch of

the background. We have a whole bunch of claws in the company and uh they're all continuously running doing things for us, writing, developing tools, developing software and so now the question is the implication the amount

of compute in our company uh that we need is just skyrocketed. The amount of compute every company needs is skyrocketing. So in that context I think

skyrocketing. So in that context I think over the last few days it's come out certainly at Morgan Stanley as a user maximum bullish on tokens maximum

bullish on doing and creating it does require the compute you just mentioned and the question is around the financing and the capex around that to support

that extraordinary large compute. How

does it all get financed as you see it from a a sort of top of the ecosystem and how do the factory AI factory economics play out and evolve?

>> Yeah.

>> And so there's a couple of thoughts of that's really important. Uh remember I I I appreciate you using the word factory.

You know several years ago I I described that these new these data centers what people call data centers is not for storing data as in a data center. they

are producing tokens and so a a a facility a plant with the fundamental purpose of producing tokens is a factory it's an AI

factory at a time people said ah Jensen that sounds so grungy um you know it's clean and and uh uh but it produces tokens and people nobody likes to build

data centers because that you know who knows what what kind of return you're going to get on a data center but everybody loves building factories. And

the reason for that is because factories make money. And we now know for we now

make money. And we now know for we now know for certain that these factories directly generate tokens. And these

tokens are monetizable.

And the more compute you have, the more tokens you could produce. The more

tokens you produce, the greater your top line. We now know for certain, we now

line. We now know for certain, we now know for certain that companies revenues are directly correlated to compute. And

we know that for a fact because if Anthropic had three times more compute, their revenues will be three times higher. We know that. We know that

higher. We know that. We know that Anthropic is capac compute limited, factory limited. It's no different than

factory limited. It's no different than Mercedes being factory limited or any company being factory limited. And so if they had more compute in their factories, they will have higher revenues. If open had right now had more

revenues. If open had right now had more compute, they will have higher revenues.

And so so so the first thought is that compute equals revenues. Now the big idea of course compute equals GDP. That

we also know. Compute equals a country's GDP. And so so that's one thought. The

GDP. And so so that's one thought. The

second thought, the reason why Nvidia is so successful is because we engineer these systems full stack end to end and they're architected from the ground up

to generate tokens at incredible effectiveness. Nvidia's tokens per watt

effectiveness. Nvidia's tokens per watt is an order of magnitude an order of magnitude

ahead of the competition alternative tokens per watt. Now, what does that mean? Remember your factory has one

mean? Remember your factory has one gigawatt and if your tokens per watt is 10 times the alternative your revenues are 10

times the alternative.

For the very first time in history the computer architecture chosen in a factory in a company's factory must go

through CEO review. No question about it.

CEO review. No question about it.

that company only has a gigawatt or 2.3 gawatts for next year. If they put the wrong system inside, it will affect their revenues the next year. I promise

you that and we see it.

And so our architecture being so advanced now and pulling further and further ahead, you know, there was a probably one of the most exhaustive benchmarking done is by by a firm called

Semi analysis and they declared they declared Nvidia inference king.

Inference King inference is tokens per second, tokens per watt. It's about

generating tokens and tokens per dollar.

When our performance per watt or per anything is so much ahead of the competition or the alternative, our tokens per dollar is also the best, which means we're the cheapest tokens

you can produce today. Not even close, an order of magnitude better. And so

that's the second thought. The second

big idea for AI is AI is a factory because factories are power limited always. Doesn't matter how many plants

always. Doesn't matter how many plants you have. Each plant is still 100

you have. Each plant is still 100 megawatts or gigawatt. And therefore

tokens per watt is the single most important thing for the top line of companies.

And they have to make those decisions very very carefully. You know, it's no longer just about PowerPoint slides.

You're not going to go put $50 billion dollars down on somebody's PowerPoint slides.

>> So, the token demand is extraordinary.

As you just mentioned, you're see we're seeing it in your numbers, right? I

think I mentioned 46 billion in net income, but $70 billion in revenue.

>> You were going to ask me something about how to fund it.

>> Um, can I can I just tell you how to fund it? First of all,

fund it? First of all, um I I just told you I just told you uh the reason why you have to build these factories in the future is because you

either you just believe that one software is important and and so I hope this audience believes software is important. Software runs the world. The

important. Software runs the world. The

first thought, the second idea is this.

There will be no software in the future that's not agentic. Do you guys agree with that?

How could you have software that's dumb?

And so it is absolutely true that every software company will become an agentic company. They're they're they're going

company. They're they're they're going to simultaneously use um uh open models.

Okay. Uh open models mean the ones that they download themselves and they fine-tune themselves. They're also going

fine-tune themselves. They're also going to use closed models. The combination of all that just like we in our all of our companies, we have employees that we hire. We have employees that we're

hire. We have employees that we're grooming. We have we have contractors

grooming. We have we have contractors that we bring bring in. We have

specialists like yourself that we bring into the company to to do our work. Our

job is not to do the job. Our job is to have the job be done.

That's what every company does. And so

therefore, every company will realize that these AI models, some of it you rent, some of it you build. That's not

illogical. Just like biological workers, you will do that with digital workers.

And so every single software company in the in the future will no longer just rent tools, but they'll rent also experts to use the tools.

They'll not just rent tools, but rent experts that use those tools because their agents are going to be extremely good at using their specialized tools.

And so, so every single software company, what did the IT industry is a couple trillion dollars today? They're

tool renters in the future. They will of course have they'll rent agents that use those tools which means that the software industry in the future will be

much larger than the software industry of today. You pick your favorite

of today. You pick your favorite software companies and I can imagine a much much larger future for them. You

cadence is going to be much larger.

Synopsis is going to be much larger.

Seammens is going to be much larger in the future. But their business profile

the future. But their business profile will change because today there's they're basically a software licensing company. In the future, they will also

company. In the future, they will also rent tokens, specialized tokens, which also means that that $2 trillion industry today with no token consumption

in the future will be extraordinary token consumers. That's where that money

token consumers. That's where that money is going to come from. They're all of those software industry of today, not the enterprise companies, the IT industry alone is going to shift an

enormous is going to consume enormous amounts of tokens in the clouds and they're either open models or >> so that extraordinary token economy is

facing some constraints. So we've got memory constraints, we've got power permitting constraints. Uh I was in

permitting constraints. Uh I was in Texas with builders, we have electrician constraints. How do you see that playing

constraints. How do you see that playing out? You're Satcha raised it in the last

out? You're Satcha raised it in the last session. You're closer to it. And also

session. You're closer to it. And also

if it takes a little longer, is it still okay or is it really negative if we just if the if the cycle on building this extraordinary >> I love constraints. I love constraints and the reason for that is because in a

world of constraint, you have no choice but to choose the best.

You can't squander your choice. If the

if if the data center if the land power and shell is constrained, you're not going to randomly put something in there um just to try it out.

You're going to put something that you know for certain is going to deliver the tokens per watt that you know for certain is going to allow you from the moment you you secure the capacity,

we're going to be able to stand up an entire factory for you. We're the only company in the world that can come into your company and help you stand up an entire AI factory, you know. So, anybody

here that needs an AI factory and you need, you know, I I'm happy to help.

You call one person and that one person comes in and next thing you know, you you're in the AI factory business. Okay?

And and and and so so we have the expertise. We know the architecture

expertise. We know the architecture works. We know there's enormous demand

works. We know there's enormous demand for the architecture, you know, after you're done standing it up so we can help you get into business. And so when you're constrained that way, you have no

choice but to make the best choice.

because your revenues next year is directly correlated to it. And this is one of those questions now for all the CEOs that are in the cloud, their cloud service providers or software um

providers. If they make poor choices,

providers. If they make poor choices, this is no different than me choosing the wrong foundry. This is no different than me choosing, you know,

the wrong memory than the wrong anything because I have so little everything is so constrained.

If I choose poorly, I my revenues are affected. Everything's affected. And so

affected. Everything's affected. And so

they can't choose they can't choose poorly. The second thing is, you know,

poorly. The second thing is, you know, Nvidia's, as you mentioned, working at such a large scale, our supply chain. One of the things that we do with our money, of course, is to

secure our supply chain. One of the things that we do with our our capital is to secure our supply chain. So that

when Satia asked me to help him stand up a few gigawatts, the answer is no problem.

And the reason for that is I got all the memories. I got all the wafers. I got

memories. I got all the wafers. I got

all the co-ass. I've got all the packaging. I've got all the system s

packaging. I've got all the system s systems. I've got all the connectors. I

got all the cables. You know, everything from copper to multi-layer ceramic pack ceramic capacitors. Everything secured.

ceramic capacitors. Everything secured.

That's one of the reasons why Nvidia's balance sheet being strong is so strategic.

A strong balance sheet today is not only helpful, it's strategic.

And so you look at the amount of revenues we're shipping into. Just look

backwards and look at the capac the amount of supply chain capacity we had to go secure or that they have to believe. You

know, if you if you set up a factory, a plant, a DRAM plant, and I come in and say, you know what, go ahead and set up the DRAM plant because I'm going to I'm going to use it. That goes a long ways.

You might as well take that to the bank, as many of them have. And so and so and and so I I think the the the fact that

everything is scarce is fantastic for us and I think it does create duration which I think is extraordinarily powerful for you. I think just another layer which is the ecosystem. You're one

of the great, you are the greatest cash flow generating company in history and then you've taken that capital and really created it feels like stability,

diversity in the ecosystem. And so how do you think about that in both a financial and a strategic context as you build I think both duration and

durability in the entire ecosystem?

Yeah, you know, when when Mark Mark took me public, um I think it was probably, you know, a little bit less energetic than I was develop

delivering it just now, but I am fairly certain I said all the same things.

Nvidia has been building. Remember,

accelerated computing requires that I build an ecosystem. You can't just take code and decompile it and it works.

There's no such thing as a universal accelerated computing system.

Accelerated computing is by definition proprietary.

There is nothing about our architecture that is compatible with somebody else's.

It's just not. The instruction set's different. The architecture is

different. The architecture is different. The micro architecture is

different. The micro architecture is different. Everything is different. And

different. Everything is different. And

so we hide it underneath these all you know these things in such a way that that makes it makes you feel like and because of NVIDIA we accelerate everything from data processing

molecular dynamics fluid dynamics particle systems you know biology chemicals you know all the way to deep learning right you know long sequence spatial 3D you

name it right >> sounds like a five layer >> Sounds like a five layer cake >> it's a five layer cake right exactly but because we've been working on it so long, it looks like everything's accelerated, but it's not true. It's

because I did it one at a time, one domain at a time, that all of the important domains in the world are now fully accelerated. And so, the thing

fully accelerated. And so, the thing that we do on the on the on the supply chain side, our balance sheet is incredibly valuable because it provides security for our customers.

On the upstream side, I'm cultivating new ecosystems for the future. All these

AI natives that I'm investing in, the companies that I'm partnering with, these are expanding, extending the CUDA ecosystem. 100% of

everything that we do is on top of CUDA.

Every investment that we've made is on top of CUDA. Um uh so recently there was a question about about um are we going to invest a hundred billion dollars in

Open AI? Uh we just just just for

Open AI? Uh we just just just for everybody's um update, we uh we uh finalized our agreement. We're going to

invest $30 billion in open AI. Um I

think the opportunity to invest $und00 billion in open AI is probably not in not in the not in the cards. And the

reason for that is because they're going to go public. And so I'm fairly sure that if we provide the capacity they need, which the compute capacity they need, which we're ramping up hard to go

do, um the revenues will more than uh follow. And uh um uh and they're going

follow. And uh um uh and they're going to go public towards the end of the year. And so this might be the last time

year. And so this might be the last time we'll have the opportunity to invest in a consequential, you know, company like this. And and our $10 billion investment

this. And and our $10 billion investment in Anthropic uh probably will be the last as well. And and speaking of that, the the one of the things that that I wanted to make sure I told you guys this

time and something new that you probably haven't internalized. You see all the

haven't internalized. You see all the news, uh you probably haven't internalized some of the the really great work that we did last year, the

last year and a half or so, last year or so. Um we expanded we expanded

so. Um we expanded we expanded uh OpenAI's capacity from Azure

to OCI to now AWS.

We expanded OpenAI's reach of capacity to AWS. We're ramping AWS like mad.

to AWS. We're ramping AWS like mad.

we're ramping them as hard as we can so that open AI has accessed even more capacity. That's one. The second thing

capacity. That's one. The second thing that we did and this was this was a a a really really great outcome is we're now also working with Anthropic

and in the case of Anthropic we're expanding their capacity as aggressively as we can at AWS

as well as Azure. And so notice what we're doing in both in both they used to be one and one now they're kind of cross productd um but the amount of capacity

that we're going to bring online for them um you know supporting supporting their revenues their quality of revenues are so good uh we just need a lot more capacity for them so I think that this

is something that I I that that is somewhat new and of course the third thing that happened is a brand new AI labed into the world isn't that right? I

don't nobody's mentioned them. A brand

new lab came into the world and they're going to need multi a few million GPUs and that's MSL. And so that MSL is a net new on top of Meta. So we've we've

worked with Meta a long time. MSL is a net new on top of Meta. And so these three things happened, three new growth vectors. um OpenAI at AWS,

vectors. um OpenAI at AWS, Anthropic totally at both AWS and Azure and MSL

and um uh so so the the the uh our demand profile uh went from uh being

incredibly high to higher than that.

Speaking of So, uh, speaking of more than that, uh, there's Whimos everywhere. I want to walk my new dog

everywhere. I want to walk my new dog with the my new robot. Physical AI could be the next place. How does that take TAM and tokens to a whole another level?

>> Yeah, that's really great. Nvidia,

>> that's really great. AI is all the stuff that we're doing inside the building, but obviously obviously ultimately the largest industries are outside the building and and um that AI needs to be

needs to have physical awareness, physical understanding, you know, causality. You push a bottle, it falls

causality. You push a bottle, it falls over and understands gravity, understands collision, you know, understands inertia, understand those things. Okay? And um understands, for

things. Okay? And um understands, for example, object permanence. um I take this and I put it behind my chair in your mind you can't see it but you realize it hasn't disappeared okay so

object permanence thing things like that that affects um uh physical behavior and physical intelligence uh you know fairly

importantly and so so um I you you probably also don't know this that that Nvidia is the frontier of physical AI cosmos is the most downloaded physical

AI model in the Nvidia is also the frontier of autonomous AI. Two versions, autonomous

autonomous AI. Two versions, autonomous vehicle called Alpamo. Look it up.

Number one downloaded. And then the next one, Groot, human or robotics physical AI. We are at the frontier on all three

AI. We are at the frontier on all three of those. We're also at the frontier of

of those. We're also at the frontier of digital biology AI. Look up La Proutina.

Incredibly successful. La Prertina for digital biology. There's a whole bunch

digital biology. There's a whole bunch of other models. um uh Groot N2 is now number one most downloaded uh uh human

robotics uh model in the world and so so we are at the frontier on physical AI uh uh physics laws of physics multifysics

earth 2 we're at the frontier of uh of uh physical uh physical AI that is physical AI and AI physics and so these whole area of physical AI Nvidia defines

the frontier It is completely open. Uh we open it because we want to enable every company

uh new or old industry to be able to take advantage of this capability. And

we've got the whole stack and the necessary computers for you to uh advance the AI for your own use. uh as

well as deploy it uh inside a robot, inside a plant, at the edge, at a radio tower. Um deploy it everywhere. This is

tower. Um deploy it everywhere. This is

the next frontier. Uh in two years time, we're going to be largely done talking about Agentic AI because we're all going to be using it. In two years time when and if you invite me back again um

>> every year, every year, every year >> we're going to be talking about all these new companies. Of course, we announced a very important one, a co-inovation lab with Lily. Um, there

will be others, but uh uh you know this in order to set up Lily's AI factory, unless you are unless you have the capabilities on Nvidia and this full software stack and the capabilities of

all the model and the expertise in that digital biology domain, how would you even do it? And so uh we're quite the things that we are building in the next couple years you'll see uh really come

to four and uh we're going to be talking about physical AI for you know starting next couple two three years and for a decade. So the speed of innovation and

decade. So the speed of innovation and the pace that you're operating in is truly extraordinary. So at the beginning

truly extraordinary. So at the beginning of the week my partner Joe Moore made uh Nvidia his number one pick. Is that

right?

>> It's his number one pick.

>> Thank you.

>> Uh uh thank you. Well, thank you.

>> Uh, good timing, Joe.

>> 30 33 years later.

>> Uh, how do you think about the stock? Do

you do you think about the stock? Do you

do you have perspectives on it? Uh,

you're so extraordinarily important and busy around driving all this innovation for in essence everything that's going on with 3,500 attendees and we had 40

trillion dollars of market cap here. How

do you how do you think about that?

>> Um, well, you know, of course I care about the stock. I I care about shareholders. I care about I care about

shareholders. I care about I care about our employees. I care about all of you.

our employees. I care about all of you.

Um, I the my and you might be referring to we just had the best earnings of earnings in the history of earnings. Is

that what you were saying? I I I think the I think somebody actually told me that that this might be the the single best print in the h in the history of

humanity and and I I I said it must be only only you know recorded humanity.

I'm sure somebody uh had better returns.

Uh but but anyways uh we did we had a very good quarter. Listen, you can't hold the stock back.

You can't hold it back. And the reason for that is very simple.

Compute equals revenues for companies.

In the in the future, every single company will need compute for revenues.

I'll just make that prediction for now.

Every single company will need compute for revenues. And the reason for that is

for revenues. And the reason for that is because compute translates to intelligence which translates to your digital workforce which translates to your revenues.

I am certain compute equals revenues.

I'm certain also that compute equals GDP. Therefore, every country will have

GDP. Therefore, every country will have it because not one country in the future will say, "Guess what? You know, we're going to opt out on intelligence.

We've got we got I don't know what we got, but we we don't need intelligence.

That's the one thing we don't need."

Okay? And so, if you need intelligence, you're going to need digital. You're

going to need AI. You're going to need compute. And so, compute equals GDP. I

compute. And so, compute equals GDP. I

know that for certain. I also know that we're at the beginning of this journey.

And I I see crystal clearly exactly how it's going to get funded. Um we know for a fact that all the CSPs took all of their capex and they converted to

generative agentic systems AI systems because it helps search because it helps shopping, because it helps ads, because it helps social because it helps

literally every single internet service in the world has been reinvented into generative AI. So they could take a 100%

generative AI. So they could take a 100% the entire internet industry could take a 100% of their capex and make it AI because it's better. We we've proven it to be better. Meta has proven to be

better. Google has proven to be better.

better. Google has proven to be better.

AWS has proven to be better. And so you could now take your capex and convert to this. Number two, I just said the entire

this. Number two, I just said the entire software industry will be token driven.

The entire software industry. You pick

your favorite software company and I can show you exactly how they're going to be token driven. And that token drip that

token driven. And that token drip that that token um you take your favorite you know software company their token uh

will be either produced by themselves which needs compute or they could be uh resold from anthropic or open AI and

that needs compute and so what that says for the first time is the entire IT industry will have to be fueled by compute that's exactly where all this is going to come from trillions of dollars of it

and we're at the beginning of that. So

that's my prediction.

>> Thank you Jensen for making history at this conference 27 years. Thank you.