GTC October 2025 Keynote with NVIDIA CEO Jensen Huang

America,

the land of innovation, where  invention shaped destiny

and technology helped dreams take flight.

At Bell Labs, the transistor was born,  

sparking the age of semiconductors  and giving rise to Silicon Valley.

Hedi Lamar reimagined communication,  paving the way for wireless connectivity.  

IBM's system 360 put a universal  computer at the heart of industry.

Intel's microprocessor drove the digital age  forward and Craig's supercomputers expanded  

the frontiers of science. So, we think we're at  the beginning of something with this technology,  

and we're going to go just as fast as we  can. Apple made computing personal. Hello,  

I imagine us. Microsoft opened the window to  a new world of software. Long before the web,  

you've got mail. US government researchers  built ARPANET connecting the first computers,  

the foundation for the internet. An iPod, a  phone. Are you getting it? Then Apple again  

put a thousand songs in your pocket and the  internet in your hand. Every era a leap. We  

choose to go to the moon in this decade and do  the other things not because they are easy but  

because they are hard. Every leap America leap for  mankind. Now the next era is here launched by a  

revolutionary new computing model. This is likely  going to be the most important contribution we've  

made to the computer industry. It will likely  be recognized as a revolution. Machine learning  

is a branch of artificial intelligence.  Computers that almost appear to think

amount of computational resource is ultimately  what's going to turbocharge this field.

Artificial intelligence, the  new industrial revolution.

At its heart, Nvidia GPUs invented here in  America. Like electricity and the internet,  

AI is essential infrastructure. Every  company will use it. Every nation will  

build it. Winning this competition will be a  test of our capacities unlike anything since  

the dawn of the space age. And today, AI  factories are rising. Built in America

for scientists, engineers, and dreamers across  universities, startups, and industry. I think  

we want to try to reach new heights as a  civilization, discovering the nature of  

the universe. And now, American innovators are  clearing the way for abundance, saving lives,

shaping vision into reality,

lending us a hand,

and delivering the future.

We will soon power it all  with unlimited clean energy.

We will extend humanity's reach to the stars.

This is America's next Apollo moment. Together,  

we take the next great leap to boldly  go where no one has gone before.

And here is where it all begins.

Welcome to the stage, Nvidia  founder and CEO, Jensen Wong.

Washington DC.

Washington DC. Welcome to GTC.

It's hard not to be sentimental and proud  of America. I got to tell you that. Was that  

video amazing? Thank you. Nvidia's creative team  does an amazing job. Welcome to GTC. We have a  

lot to cover with you today. Um, GTC is where  we talk about industry, science, computing,  

the present, and the future. So, I've got a lot  to cover with you today, but before I start,  

I want to thank all of our partners who helped  sponsor this great event. You'll see all of them  

around the show. They're here to meet with you  and uh uh really great. We couldn't do what we do  

without all of our ecosystem partners. This is the  Super Bowl of AI, people say. And therefore, every  

Super Bowl should have an amazing pregame show.  What do you guys think about the pregame show  

and our all all star allstar athletes and allstar  cast? Look at these guys. Somehow I turned out the  

buffest. What do you guys think? I don't know if  I had something to do with that. Nvidia invented  

a new computing model for the first time in 60  years. As you saw in the video, a new computing  

model rarely comes about. It takes an enormous  amount of time and set of conditions. We observed,  

we invented this computing model because we wanted  to solve problems that generalpurpose computers,  

normal computers could not. We also observed that  someday transistors will continue the number of  

transistors will grow but the performance and the  power of transistors will slow down that Moore's  

law will not continue beyond limited by the laws  of physics and that moment has now arrived dinard  

scaling has stopped it's called dinard scaling  dinard scaling has stopped nearly a decade ago  

and in fact the transistor performance and its  power associated ated has slowed tremendously and  

yet the number of transistor continued. We made  this observation a long time ago and for 30 years  

we've been advancing this form of computing we  call accelerated computing. We invented the GPU.  

We invented the the programming model called  CUDA. And we observed that if we could add a  

processor that takes advantage of more and more  and more transistors, apply parallel computing,  

add that to a sequential processing CPU that we  could extend the capabilities of computing well  

beyond well beyond. And that moment has really  come. We have now seen that inflection point.  

Accelerated computing its moment has now arrived.  However, accelerated computing is a fundamentally  

different programming model. You can't just take  a CPU software software written by hand executing  

sequentially and put it onto a GPU and have it  run properly. In fact, if you just did that, it  

actually runs slower. And so, you have to reinvent  new algorithms. You have to create new libraries.  

You have to in fact rewrite the application which  is the reason why it's taken so long. It's taken  

us nearly 30 years to get here. But we did it  one domain at a time. This is the treasure of  

our company. Most people talk about the GPU. The  GPU is important, but without a programming model  

that sits on top of it, and without dedication to  that programming model, keeping it compatible over  

generations, we're now CUDA 13 coming up with  CUDA 14, hundreds of millions of GPUs running  

in every single computer, perfectly compatible. If  we didn't do that, then developers wouldn't target  

this computing platform. If we didn't create these  libraries, then developers wouldn't know how to  

use the algorithm and use the architecture to its  fullest. One application after another. I mean,  

these this is really the this is really the  treasure of our company. CU litho computational  

lithography. It took us nearly seven years to  get here with KU Litho and now TSMC uses it,  

Samsung uses it, ASML uses it. This is  an incredible library for computational  

lithography. the first step of making a chip.  Sparse solvers for CAE applications. Co-op, a  

numerical optimization is broken just about every  single record. The traveling salesperson problem,  

how to connect millions of products with millions  of customers in the supply chain. Warp Python  

solver for CUDA for simulation. QDF a dataf frame  approach basically accelerating SQL dataf frame  

pro dataf frame databases. Um this library is  the one that started AI al together coupnn the  

the library on top of it called megatron core made  it possible for us to simulate and train extremely  

large language models. The list goes on. Uh,  Monai, really, really important, the number one  

medical imaging AI framework in the world. Uh,  by the way, we're not going to talk a lot about  

health care today, but be sure to see Kimberly's  keynote. She's going to talk a lot about the work  

that we do in healthcare. And the list goes on.  Uh, genomics processing, Ariel, pay attention.  

We're going to do something really important  here today. Um, coup quantum quantum computing.  

This is just a representative of 350 different  libraries in our company. And each one of these  

libraries redesigned the algorithm necessary for  accelerated computing. Each one of these libraries  

made it possible for all of the ecosystem partners  to take advantage of accelerated computing. And  

each one of these libraries opened new markets  for us. Let's take a look at what CUDA X can do.

Ready go.

Heat.

Heat.

Heat. Heat.

Heat.

Heat.

up

here.

Heat.

Heat.

Heat up here.

Is that amazing? Every everything you  saw was a simulation. There was no art,  

no animation. This is the beauty of  mathematics. This is deep computer science,  

deep math. And it's just incredible how beautiful  it is. Every industry was covered from healthcare  

and life sciences to manufacturing, robotics,  autonomous vehicles, computer graphics,  

even video games. That first shot that you saw was  the first application Nvidia ever ran. And that's  

where we started in 1993. And we kept believing  in what we were trying to do. And it took,  

it's hard to imagine that you could see that first  virtual fighter scene come alive and that same  

company believed that we would be here today. It's  just a really, really incredible journey. I want  

to thank all the NVIDIA employees for everything  that you've done. It's really incredible.

We have a lot of industries to cover today. I'm  going to cover AI, 6G, quantum, models, enterprise  

computing, robotics, and factories. Let's get  started. We have a lot to cover, a lot of big  

announcements to make, a lot of new partners that  would very much surprise you. Telecommunications  

is the backbone, the lifeblood of our economy, our  industries, our national security. And yet ever  

since the beginning of wireless where we defined  the technology, we defined the global standards,  

we exported American technology all around  the world so that the world can build on  

top of American technology and standards. It  has been a long time since that's happened.  

wireless technology around the world largely  today deployed on foreign technologies. Our  

fundamental communication fabric built on foreign  technologies. That has to stop and we have an  

opportunity to do that especially during this  fundamental platform shift. As you know computer  

technology is at the foundation of literally  every single industry. It is the single most  

important instrument of science. It's the single  most important instrument of industry. And I just  

said we're going through a platform shift. That  platform shift should be the once-in-a-lifetime  

opportunity for us to get back into the game for  us to start innovating with American technology.  

Today, today we're announcing that we're going  to do that. We have a big partnership with Nokia.  

Nokia is the second largest telecommunications  maker in the world. It's a three trillion dollar  

industry. Infrastructure is hundreds of billions  of dollars. There are millions of base stations  

around the world. If we could partner, we could  build on top of this incredible new technology  

fundamentally based on accelerated computing and  AI. and for United States, for America to be at  

the center of the next revolution in 6G. So today  we're announcing that Nvidia has a new product  

line. It's called the NVIDIA Arc. The Aerial  Radio Network Computer, Aerial RAM computer,  

ARC. A Arc is built from three fundamental new  technologies. the gray CPU, the Blackwell GPU,  

and our ConnectX Melanox Connect X networking  designed for this application. And all of that  

makes it possible for us to run this library,  this CUDA X library that I mentioned earlier  

called Aerial. Ariel is essentially a wireless  communication system running on top of CUDA X.  

We're going to we're going to create for the  first time a softwaredefined programmable computer  

that's able to communicate wirelessly and do AI  processing at the same time. This is completely  

revolutionary. We call it Nvidia Arc. And Nokia is  going to work with us to integrate our technology,  

rewrite their stack. This is a company with  7,000 fundamental essential 5G patents.

Hard to imagine any greater leader in  telecommunications. So, we're going to partner  

with Nokia. They're going to make Nvidia Arc their  future base station. Nvidia Arc is also compatible  

with Airscale, the current Nokia base stations.  So what that means is we're going to take this new  

technology and we'll be able to upgrade millions  of base stations around the world with 6G and AI.  

Now 6G and AI is really quite fundamental in  the sense that for the first time we'll be able  

to use AI technology AI for RAN to make radio  communications more spectral efficient doing  

using artificial intelligence reinforcement  learning adjusting the beam forming in real  

time in context depending on the surroundings  and the traffic and the mobility the weather  

all of that could be taken into account so that  we could improve spectral efficiency. Spectral  

efficiency consumes about 1 and a half to 2%  of the world's power. So improving spectral  

efficiency not only improves the amount of data  we can put through wireless networks without  

increasing the amount of energy necessary.  The other thing that we could do AI for RAN  

is AI on RAM. This is a brand new opportunity.  Remember the internet enabled communications  

but amazingly smart companies AWS built a cloud  computing system on top of the internet. We are  

now going to do the same thing on top of the  wireless telecommunications network. This new  

cloud will be an edge industrial robotics cloud.  This is where AI on RAN the first is AI for RAN  

to improve radio radio spectrum efficiency. The  second is AI on RAN essentially cloud computing  

for wireless telecommunications. Cloud computing  will be able to go right out to the edge where  

data centers are not are not because we have base  stations all over the world. This announcement is  

really exciting. Justin Hodar the CEO I think  he's somewhere in the room. Thank you for your  

partnership. Thank you for helping United  States bring telecommunication technology  

back to America. This is really a fantastic,  fantastic partnership. Thank you very much.

That's the best way to celebrate Nokia.

Let's talk about quantum computing. 1981 particle  physicist quantum physicist Richard Feman imagined  

a new type of computer that can simulate nature  directly to simulate nature directly because  

nature is quantum. He called it a quantum  computer. 40 years later the industry has  

made a fundamental breakthrough. 40 years later,  just last year, a fundamental breakthrough. It  

is now possible to make one logical cubit. One  logical cubit. One logical cubit that's coherent,  

stable, and error corrected in past. Now that one  logical cubit consists of could be sometimes tens,  

sometimes hundreds of physical cubits all  working together. As you know, cubits,  

these particles are incredibly fragile. They  could be unstable very easily. Any observation,  

any sampling of it, any environmental condition  causes it to become decoherent. And so,  

it takes an extraordinarily well-controlled  environments. And now also a lot of different  

physical cubits for them to work together and  for us to do error correction on these what are  

called auxiliary or syndrome cubits for us to  error correct them and infer what that logical  

cubit state is. There are all kinds of different  types of quantum computers. superconducting,  

photonic, trapped ion, stable atom, all kinds  of different ways to create a quantum computer.  

Well, we now realize that it's essential for  us to connect a quantum computer directly to  

a GPU supercomput so that we could do the  error correction so that we could do the  

artificial intelligence calibration and control  of the quantum computer and so that we could do  

simulations collectively working together.  the right algorithms running on the GPUs,  

the right algorithms running on the QPUs  and the two processors, the two computers  

working side by side. This is the future  of quantum computing. Let's take a look.

There are many ways to build a quantum computer.  Each uses cubits, quantum bits as its core  

building block. But no matter the method, all  cubits, whether superconducting cubits, trapped  

ions, neutral atoms, or photons, share the same  challenge. They're fragile and extremely sensitive  

to noise. Today's Qbits remain stable for only  a few hundred operations. But solving meaningful  

problems requires trillions of operations. The  answer is quantum error correction. Measuring  

disturbs a cubit which destroys the information  inside it. The trick is to add extra cubits in  

tangle so that measuring them gives us  enough information to calculate where  

errors occurred without damaging the cubits  we care about. It's brilliant but needs beyond  

state-of-the-art conventional compute. That's why  we built NVQLink, a new interconnect architecture  

that directly connects quantum processors with  NVIDIA GPUs. Quantum error correction requires  

reading out information from QIDS, calculating  where errors occur and sending data back  

to correct them. MVQLink is capable of moving  terabytes of data to and from quantum hardware,  

the thousands of times every second needed for  quantum error correction. At its heart is CUDAQ,  

our open platform for quantum GPU computing. Using  MVQL link and CUDAQ, researchers will be able to  

do more than just error correction. They will  also be able to orchestrate quantum devices and  

AI supercomputers to run quantum GPU applications.  Quantum computing won't replace classical systems.  

They will work together fused into one  accelerated quantum supercomputing platform.

Wow, this is a really long stage.

You know, CEOs, we don't just sit at our desk  typing. It's this is a physically job. Physical  

job. So, so today we're announcing the MV  MVQ link. MVQL link and it's made possible  

by two things. Of course, this interconnect that  does quantum computer control and calibration,  

quantum error correction as well as  connects two computers, the QPU and  

our GPU supercomputers to do hybrid simulations.

It is also completely scalable. It doesn't just  do error correction for today's number of few  

cubits. It does error correction for tomorrow  where we're going to essentially scale up these  

quantum computers from the hundreds of cubits  we have today to tens of thousands of cubits,  

hundreds of thousands of cubits in the future. So  we now have an architecture that can do control,  

co- simulation, quantum error correction and  scale into that future. The industry support has  

been incredible between the invention of CUDA Q.  Remember CUDA was designed for GPU CPU accelerated  

computing. Basically using both processors to do  use the right tool to do the right job. Now CUDA  

Q has been extended beyond CUDA so that we could  support QPU and have the two processors QPU and  

the GPU work and have computation move back and  forth within just a few microsconds. The essential  

latency to be able to cooperate with the quantum  computer. So now CUDAQ is such an incredible  

breakthrough adopted by so many different  developers. We are announcing today 17 different  

quantum computer industry companies supporting the  MVQ link and and I'm so excited about this eight  

different DOE labs. Berkeley, Brook Haven,  Fermy Labs in Chicago, Lincoln Laboratory,  

Los Alamos, Oakidge, Pacific Northwest, San  Diego Lancho Lab, just about every single DOE  

lab has engaged us working with our ecosystem  of quantum computer companies and these quantum  

controllers so that we could integrate quantum  computing in into the future of science. Well,  

I have one more additional announcement to make.  Today, we're announcing the Department of Energy  

is partnering with NVIDIA to build seven new AI  supercomputers to advance our nation's science.

I have to have a shout out for Secretary Chris  Wright. He has brought so much energy to the DOE,  

a surge of energy, a surge of passion to make sure  that America leads science. Again, as I mentioned,  

computing is the fundamental instrument of  science and we are going through several  

platform shifts. On the one hand, we're going to  accelerated computing. That's why every future  

supercomputer will be GPUbased supercomputer.  We're going to AI so that AI and principled  

solvers, principled simulation, principal  physics simulation is not going to go away.  

But it could be augmented, enhanced, scaled, use  surrogate models, AI models working together. We  

also know that principal solvers, classical  computing could be enhanced to understand the  

state of nature using quantum computing. We also  know that in the future, we have so much signal,  

so much data we have to sample from the world.  Remote sensing is more important than ever.  

And these laboratories are impossible to  experiment at the scale and speed we need to  

unless they're robotic factories, robotic  laboratories. So all of these different  

technologies are coming into science at exactly  the same time. Secretary Wright understands this  

and he wants the DOE to take this opportunity  to supercharge themselves and make sure the  

United States stay at the forefront of science.  I want to thank all of you for that. Thank you.

Let's talk about AI. What is AI? Most people  would say that AI is a chatbot and it it's  

rightfully so. There's no question that Chad  GPT is at the forefront of what people would  

consider AI. However, just as you see right  now, these scientific supercomputers are not  

going to run chatbots. They're going to do basic  science. Science, AI, the world of AI is much,  

much more than a chatbot. Of course, the chatbot  is extremely important and AGI is fundamentally  

critical. Deep computer science, incredible  computing, great breakthroughs are still  

essential for AGI. But beyond that, AI is a lot  more. AI is in fact, I'm going to describe AI in  

a couple different ways. This first way, the first  way you think about AI is that it has completely  

reinvented the computing stack. The way we used to  do software was hand coding. Hand coding software  

running on CPUs. Today AI is machine learning  training data inensive programming if you will  

trained and learned by AI that runs on a GPU. In  order to make that happen, the entire computing  

stack has changed. Notice you don't see Windows  up here. You don't see CPU up here. You see a  

whole different a whole fundamentally different  stack. Everything from the need for energy. And  

this is another area where our administration,  President Trump gets deserves enormous credit.  

His pro- energy initiative, his recognition that  this industry needs energy to grow. It needs  

energy to advance. and we need energy to win. His  recognition of that and putting the weight of the  

nation behind pro- energy growth completely  changed the game. If this didn't happen,  

we could have been in a bad situation and  I want to thank President Trump for that.

On top of energy are these GPUs and these GPUs are  connected into built into infrastructure that I'll  

show you later. On top of this infrastructure  which in consists of giant data centers like  

easily many times the size of this room enormous  amount of energy which then transfer transforms  

the energy through this new machine called  GPU supercomputers to generate numbers. These  

numbers are called tokens. the language, if you  will, the computational unit, the vocabulary  

of artificial intelligence. You can tokenize  almost anything. You can tokenize, of course,  

the English word. You can tokenize images. That's  the reason why you're able to recognize images  

or generate images, tokenize video, tokenize  3D structures. You could tokenize chemicals  

and proteins and genes. You could tokenize cells,  tokenize almost anything with structure, anything  

with information content. Once you could tokenize  it, AI can learn that language and the meaning of  

it. Once it learns the meaning of that language,  it can translate. It can respond just like you  

respond just like you interact with chat GPT. And  it could generate just as chat GPD can generate.  

So all of the fundamental things that you see  Chad GPD do, all you have to do is imagine what  

if it was a protein, what if it was a chemical,  what if it was a 3D structure like a factory,  

what if it was a robot and the token was  understanding behavior and tokenizing motion  

and action. All of those concepts are basically  the same, which is the reason why AI is making  

such extraordinary progress. And on top of these  models are applications transformers. Transformers  

is not a universal model. It's an incredibly  effective model. But there's no one universal  

model. It's just that AI has universal impact.  There are so many different types of models.  

There's in the last several years we enjoyed  the invention and went through the innovation  

breakthroughs of multimodality. There's so many  different types of models. There's CNN models,  

competition neuronet network models, their state  space models, their graph neuronet network models,  

multimodal models, of course, all the different  tokenizations and token methods that I just  

described. You could have models that are  spatial and it's understanding optimized for  

spatial awareness. You could have models that  are optimized for long sequence recognizing  

subtle information over a long period of time.  There are so many different types of models.  

On top of these models architectures, on top  of these model architectures are applications,  

the software of the past. And this is a a  profound understanding, a profound observation  

of artificial intelligence that the software  industry of the past was about creating tools.  

Excel is a tool. Word is a tool. A web browser is  a tool. The reason why I know these are tools is  

because you use them. The tools industry, just  as screwdrivers and hammers, the tools industry  

is only so large. In the case of IT tools, they  could be database tools. These IT tools is about  

a trillion dollars or so. But AI is not a tool.  AI is work. That is the profound difference.  

AI is in fact workers that can actually use tools.  One of the things I'm really excited about is the  

work that Irvin's doing at Perplexity. Perplexity  using web browsers to book vacations or do  

shopping basically an AI using tools. Cursor is an  AI anantic AI system that we use at NVIDIA. Every  

single software engineer at Nvidia uses cursor  has improved our productivity tremendously. It's  

basically a partner for every one of our software  engineers to generate code and it uses a tool and  

the tool it uses is called VS code. So cursor is  an AI agentic AI system that uses VS code. Well,  

all of these different industries, these different  industries, whether it's chat bots or digital  

biology where we have AI assistant researchers or  what is a robo taxi inside a robo taxi? Of course,  

it's invisible, but obviously there's a AI  chauffeur. That chauffeur is doing work and  

the tool that it uses to do that work is the car.  And so everything that we've made up until now,  

the whole world, everything that we've made up  until now are tools. Tools for us to use. For  

the very first time, technology is now able to do  work and help us be more productive. The list of  

opportunities go on and on, which is the reason  why AI addresses the segment of the economy that  

it has never addressed. It is a few trillion  dollars that sits underneath the tools of a  

hundred trillion dollar global economy. Now  for the first time AI is going to engage that  

hundred trillion dollar economy and make it more  productive, make it grow faster, make it larger.  

We have a severe shortage of labor. Having AI  that augments labor is going to help us grow. Now  

what's interesting about this from a technology  industry perspective also is that in addition to  

the fact that AI is new technology that addresses  new segments of the economy AI in itself is also  

a new industry this token as I was explaining  earlier these numbers after you tokenize all these  

different modalities of information there's a  factory that needs to produce these numbers unlike  

the computer industry indry and the chip industry  of the past. Notice if you look at the chip  

industry of the past, the chip industry represents  about 5 to 10% maybe less 5% or so of a multi-  

trillion dollar few trillion dollar IT industry.  And the reason for that is because it doesn't take  

that much computation to use Excel. It doesn't  take that much computation to use browsers. It  

doesn't take that much computation to use Word.  We do the computation. But in this new world,  

there needs to be a computer that understands  context all the time. It can't precomputee that  

because every time you use the computer for AI,  every time you ask the AI to do something, the  

context is different. So, it has to process all of  that information. Environmental, for example, in  

the case of a self-driving car, it has to process  the context of the car. context processing. What  

is the instruction you're asking the AI to do?  Then it's got to go and break down the problem  

step by step, reason about it, and come up with a  plan and execute it. Every single one of that step  

requires enormous number of tokens to be generated  which is the reason why we need a new type of  

system and I call it an AI factory. It's an AI  factory for short. It's unlike a data center of  

the past. It's an AI factory because this factory  produces one thing unlike the data centers of the  

past that does everything. Stores files for all  of us, runs all kinds of different applications.  

You could use that data center like you can use  your computer for all kinds of applications. You  

could use it to play game one day. You could use  it to browse the web. You could use it, you know,  

to do your accounting. And so that is a computer  of the past, a universal generalpurpose computer.  

The computer I'm talking about here is a factory.  It runs basically one thing. It runs AI and its  

purpose, its purpose is designed to produce  tokens that are as valuable as possible,  

meaning they have to be smart. And you want to  produce these tokens at incredible rates because  

when you ask an AI for something, you would like  it to respond. And notice during peak hours,  

these AIs are now responding slower and slower  because it's got a lot of work to do for a lot  

of people. And so you wanted to produce valuable  tokens at incredible rates and you wanted to  

produce it cost effectively. Every single word  that I used are consistent with an AI factory,  

with a car factory or any factory. It is  absolutely a factory. And these factories,  

these factories never existed before. And inside  these factories are mountains and mountains of  

chips. Which brings to today. What happened in the  last couple years? And in fact, what happened this  

last year? Something fairly profound happened this  year. Actually, if you look in the beginning of  

the year, everybody has some attitude about AI.  That attitude is generally this is going to be  

big. It's going to be the future. And somehow  a few months ago, it kicked into turbocharge.  

And the reason for that is several things. The  first is that we in the last couple years have  

figured out how to make AI much much smarter.  Rather than just pre-training, pre-training  

basically says let's take all of the all of the  information that humans have ever created. Let's  

give it to the AI to learn from. It's essentially  memorization and generalization. It's no it's not  

unlike going to school back when we were kids.  the first stage of learning. Pre-training was  

never meant just as preschool was never meant  to be the end of education. Pre-training,  

preschool was simply teaching you the basic skills  of intelligence so that you can understand how to  

learn everything else. Without vocabulary, without  understanding of language and how to communicate,  

how to think, it's impossible to learn everything  else. The next is post-training. Post-training  

after pre-training is teaching you skills. Skills  to solve problem. Break down problems. Reason  

about it. How to solve math problems. How to code.  How to think about these problems step by step.  

Use first principle reasoning. And then after that  is where computation really kicks in. As you know  

for many of us, you know, we went to school and  that's in my case decades ago. But ever since I've  

learned more, thought about more. And the reason  for that is because we're constantly grounding  

oursel in new knowledge. We're constantly doing  research and we're constantly thinking. Thinking  

is really what intelligence is all about. And so  now we have three fundamental technology skills.  

We have these three technologies. Pre-training,  which still requires enormous enormous amount of  

computation. We now have post training which  uses even more computation and now thinking  

puts incredible amounts of computation load on  the infrastructure because it's thinking on our  

behalf for every single human. So the amount of  computation necessary for AI to think inference  

is really quite extraordinary. Now I used  to hear people say that inference is easy.  

Nvidia should do training. Nvidia is going to  do you know they are really good at this. So  

they're going to do training. The inference was  easy. How could thinking be easy? Regurgitating  

memorized content is easy. Regurgitating the  multiplication tables easy. Thinking is hard.  

Which is the reason why these three scales,  these three new scaling laws which is all of  

it in in full steam has put so much pressure on  the amount of computation. Now another thing has  

happened from these three scaling laws. We get  smarter models and these smarter models need more  

compute. But when you get smarter models,  you get more intelligence. People use it

as if anything happens. I  want to be the first one out.

Jazz kick. I'm sure it's fine. Probably  just lunch. My stomach. Was that me? And so,  

so where was I? The smarter your models are, the  smarter your models are, the more people use it.  

It's now more grounded. It's able to reason.  It's able to solve problems it never learn how  

to solve before because it could do research.  Go learn about it. come back, break it down,  

reason about how to solve your how to answer  your question, how to solve your problem,  

and go off and solve it. The amount of thinking  is making the models more intelligent. The more  

intelligent it is, the more people use it. The  more intelligent it is, the more computation  

is necessary. But here's what happened. This  last year, the AI industry turned the corner,  

meaning that the AI models are now smart enough.  They're making they're worthy. They're worthy to  

pay for. Nvidia pays for every license of Cursor.  And we gladly do it. We gladly do it because  

Curser is helping a several hundred,000 employee  software engineer or AI researcher be many,  

many times more productive. So, of course, we'd be  more than happy to do that. These AI models have  

become good enough that they are worthy to be paid  for. Cursor, 11 Labs, Syntheasia, A Bridge, Open  

Evidence, the list goes on. Of course, Open AI, of  course, Claude. These models are now so good that  

people are paying for it. And because people are  paying for it and using more of it, and every time  

they use more of it, you need more compute. We  now have two exponentials. These two exponentials,  

one is the exponential compute requirement of the  three scaling law. And the second exponential,  

the more people, the smarter it is, the more  people use it, the more people use it, the more  

computing it needs. Two exponentials now putting  pressure on the world's computational resource

at exactly the time when I told you earlier that  Moore's law has largely ended. And so the question  

is what do we do? If we have these two exponential  demands growing and if we don't if we don't find a  

way to drive the cost down then this positive  feedback system this circular feedback system  

essentially called the virtuous cycle essential  for almost any industry essential for any platform  

industry. It was essential for Nvidia. We have  now reached the virtual cycle of CUDA. The more  

applications, the more the more applications  people create, the more valuable CUDA is. The  

more valuable CUDA is, the more CUDA computers are  purchased. The more CUDA computers are purchased,  

more developers want to create applications for  it. That virtual cycle for Nvidia has now been  

achieved after 30 years. We have achieved that  also. 15 years later, we've achieved that for AI.  

AI has now reached the virtual cycle and so  the more you use it because the AI is smart  

and we pay for it the more profit is generated  the more profit generated the more computes  

put to on the on the grid the more compute is  put into AI factories the more comput the AI  

becomes smarter the smarter more more people use  it more applications use it the more problems we  

can solve this virtual cycle is now spinning what  we need to do is drive the cost down tremendously  

So that one the user experience is better when you  prompt the AI it responds to you much faster and  

two so that we keep this virtual cycle going  by driving its cost down so that it could get  

smarter so that more people use it so that so on  so forth that virtual cycle is now spinning but  

how do we do that when Moore's law has really  reached this limit well the answer is called  

co-design you can't just design chips and hope  that things on on top of it is going to go faster.  

The best you could do in designing chips is add  I don't know 50% more transistors every couple  

of years and if you added more transistors just  you know we can add more transistors and TSMC's  

got a lot of transistor incredible company we  just keep adding more transistors however that's  

all in percentages not exponentials we need to  compound exponentials to keep this virtual cycle  

going extreme code design is the only company  in the world today that literally starts from  

a blank sheet of paper and can think about new  fundamental architecture, computer architecture,  

new chips, new systems, new software, new model  architecture and new applications all at the same  

time. So many of the people in this room are  here because you're different parts of that  

layer that different parts of that stack and  working with NVIDIA. We fundamentally rearchitect  

everything from the ground up and then because  AI is such a large problem, we scale it up. We  

created a whole computer, a computer for the  first time that has scaled up into an entire  

rack. That's one computer, one GPU. And then  we scale it out by inventing a new AI Ethernet  

technology we call Spectrum Ethernet. Everybody  will say Ethernet is Ethernet. Ethernet is hardly  

Ethernet. Ethernet spectrum X Ethernet is designed  for AI performance and it's the reason why it's so  

successful. And even that's not big enough. We'll  fill this entire room of AI supercomputers and  

GPUs. That's still not big enough because the  number of applications and the number of users  

for AI is continuing to grow exponentially.  And we connect multiple of these data centers  

together and we call that scale across spectrum  XGS gigascale X spectrum X gigascale XGS. By  

doing so, we do code design at such a such an  enormous level, such an extreme level that the  

performance benefits are shocking. Not 50% better  each generation, not 25% better each generation,  

but much much more. This is the most extreme  code-designed computer we've ever made and quite  

frankly made in modern times. Since the IBM system  360, I don't think a computer has been ground up,  

reinvented like this ever. This system was  incredibly hard to create. I'll show you the  

benefits in just a second. But essentially what  we've done, essentially what we've done, we've  

created otherwise Hey Janine, you can come out.  It's you have to have to meet me like halfway. All

right. So, this is kind of like Captain  America shield. So, MVLink 72, MVLink72,  

if we were to create one giant chip, one giant  GPU, this is what it would look like. This is  

the level of wafer scale processing we would  have to do. It's incredible. All of this,  

all of these chips are now put into one giant  rack. Did I do that or did somebody else do that?  

Into that one giant rack. You know, sometimes  I don't feel like I'm up here by myself. Just

this one giant rack makes all of these chips  work together as one. It's actually completely  

incredible. And I'll show you the benefits of  that. The way it looks is this. So, thanks,  

Janine. I I like this. All right,  ladies and gentlemen, Janine Paul.

I got it. In the future next,  I'm just going to go like Thor.

It's like when you're at home and and you can't  reach the remote and you just go like this and  

somebody brings it to you. That's Yeah. Same  idea. It never happens to me. I'm just dreaming  

about it. I'm just saying. Okay. So, so anyhow,  anyhow, um we basically this is what we created  

in the past. This is MVLink MVLink 8. Now, these  models are so gigantic. The way we solve it is we  

turn this model, this gigantic model, into a whole  bunch of experts. It's a little bit like a team.  

And so, these experts are good at certain types of  problems. And we collect a whole bunch of experts  

together. And so, this giant multi-trillion dollar  AI model has all these different experts and we  

put all these different experts on a GPU. Now,  this is NVLink 72. We could put all of the chips  

into one giant fabric and every single expert can  talk to each other. So the master the the primary  

expert could talk to all of the true work and all  of the necessary contexts and prompts and bunch of  

data that we have to bunch of tokens that we have  to send to all of the experts. The experts would  

whichever one of the experts are selected to solve  the answer would then go off and try to respond  

and then it would go off and do that layer after  layer after layer. Sometimes eight, sometimes  

16 and sometimes these experts, sometimes 64,  sometimes 256. But the point is there are more  

and more and more experts. Well, here MVLink 72,  we have 72 GPUs. And because of that, we could put  

four experts in one GPU. The most important thing  you need to do for each GPU is to generate tokens,  

which is the amount of bandwidth that you have  in HBM memory. We have one H one GPU generating  

thinking for four experts versus here because each  one of the computers can only put eight GPUs. We  

have to put 32 experts into one GPU. So this one  GPU has to think for 32 experts versus this system  

each GPU only has to think for four. And because  of that the speed difference is incredible. And  

this just came out. This is the benchmark done by  semi analysis. They do a really really thorough  

job and they benchmarked all of the GPUs that are  benchmarkable and it turns out it's not that many.  

If you look at the list of looks list of GPUs you  could actually benchmark is like 90% Nvidia. Okay.  

And but so we're comparing ourselves to ourselves  but the second best GPU in the world is the H200  

and runs all the workload. Grace Blackwell per GPU  is 10 times the performance. Now, how do you get  

10 times the performance when it's only twice the  number of transistors? Well, the answer is extreme  

code design. And by understanding the nature of  the future of AI models and we're thinking across  

that entire stack, we can create architectures  for the future. This is a big deal. It says we  

can now respond a lot faster. But this is even  bigger deal. This next one, look at this. This  

says that the lowest cost tokens in the world are  generated by Grace Blackwell MVLink72. The most  

expensive computer. On the one hand, GB200 is the  most expensive computer. On the other hand, its  

token generation capability is so great that it  produces it at the lowest cost because the tokens  

per second divided by the t by the total cost of  ownership of Grace Blackwell is so good. It is  

the lowest cost way to generate tokens. By doing  so, delivering incredible performance, 10 times  

the performance, incre delivering 10 times lower  cost, that virtual cycle can continue. Which then  

brings me to this one. I just saw this literally  yesterday. This is uh the CSP capex. People are  

asking me about capex these days and um this is a  good way to look at it. In fact, the capex of the  

top six CSPs and this one this one is Amazon,  Corewave, Google, Meta, Microsoft and Oracle.  

Okay, these CSPs together are going to invest this  much in capex and I would I would tell you the  

timing couldn't be better and the reason for that  is now we have the Grace Blackwell MVLink72 in all  

volume production supply chain everywhere in the  world is manufacturing it. So we can now deliver  

to all of them this new architecture so that  the capex invests in instruments computers that  

deliver the best TCO. Now underneath this there  are two things that are going on. So when you  

look at this it's actually fairly extraordinary  and it's fairly extraordinary anyhow. But what's  

happening under underneath is this there are  two platform shifts happening at the same time.  

One platform shift is going from general purpose  computing to accelerated computing. Remember  

accelerated computing as I mentioned to you before  it does data processing, it does image processing,  

computer graphics, it does com comput computation  of all kinds. It runs SQL, runs spark, it runs,  

you know, you you ask it, you tell us what you  need to have run, and I'm fairly certain we  

have an amazing library for you. You could be,  you know, a data center trying to make masks  

to manufacture semiconductors. we have a great  library for you. And so underneath irrespective  

of AI, the world is moving from general purpose  computing to accelerated computing irrespective  

of AI. And in fact, many of the CSPs already have  services that have been here long ago before AI.  

Remember, they were invented in the era of  machine learning. classical machine learning  

algorithms like XG Boost, algorithms like um uh  data frames that are used for recommener systems,  

collaborative filtering, content filtering,  all of those technologies were created in the  

old days of general purpose computing. Even  those algorithms, even those architectures  

are now better with accelerated computing. And  so even without AI, the world's CSPs are going  

to invest into acceleration. Nvidia's GPU is  the only GPU that can do all of that plus AI.  

And ASIC might be able to do AI, but it can't do  any of the others. Nvidia could do all of that,  

which explains why it is so safe to just  lean into Nvidia's architecture. We have now  

reached our virtual cycle, our inflection  point. And this is quite extraordinary.  

I have many partners in the room and all of you  are part of our supply chain and I know how hard  

you guys are working. I want to thank all of you  how hard you are working and thank you very much.

Now I'm going to show you why this is what's  going on in our company's business. We're seeing  

extraordinary growth for Grace Blackwell for all  the reasons that I just mentioned. It's driven by  

two exponentials. We now have visibility. I think  we're probably the first technology company in  

history to have visibility into half a trillion  dollars of cumulative blackwell and early ramps  

of Reubin through 2026. And as you know, 2025  is not over yet and 2026 hasn't started. This  

is how much business is on the books. Half a  trillion dollars worth so far. Now, this is  

out of that. We've already shipped 6 million of  the Blackwells in the first several quarters. I  

guess the first four quarters of production, three  and a half quarters of production. We still have  

one more quarter to go for 2025. And then we have  four quarters. So the next five quarters there's  

$500 million $500 billion half a trillion dollars.  That's five times the growth rate of Hopper. That  

kind of tells you something. This is Hopper's  entire life. This doesn't include China and and  

um and Asia. So this is just uh the West. Okay.  This is just uh we're excluding China. So Hopper  

in its entire life 4 million GPUs. Blackwell.  Each one of the Blackwells has two GPUs in it  

in one large package. 20 million GPUs of  Blackwells in the early parts of Reuben.  

Incredible growth. So, I want to thank all  of our supply chain partners. Everybody,  

I know how hard you guys are working. I made  a video to celebrate your work. Let's play it.

The age of AI has begun. Blackwell is its  engine, an engineering marvel. In Arizona,  

it starts as a blank silicon wafer. Hundreds of  chip processing and ultraviolet lithography steps  

build up each of the 200 billion transistors  layer by layer on a 12in wafer. In Indiana,  

HBM stacks will be assembled in parallel. HBM  memory dies with 1,024 IO's are fabricated using  

advanced EUV technology through silicon via is  used in the back end to connect 12 stacks of HBM  

memory and base dye to produce HBM. Meanwhile, the  wafer is scribed into individual Blackwell dye,  

tested and sorted, separating the good dyes to  move forward. The chip on wafer on substrate  

process attaches 32 Blackwell dyes and 128 HBM  stacks on a custom silicon interposer wafer.

Metal interconnect traces are etched directly into  it, connecting Blackwell GPUs and HBM stacks into  

each system and package unit, locking everything  into place. Then the assembly is baked, molded,  

and cured, creating the GB300 Blackwell Ultra  Super Chip. In Texas, robots will work around  

the clock to pick and place over 10,000  components onto the Grace Blackwell PCB.  

In California, Connect X8 Super Nix for scaleout  communications and Bluefield 3 DPUs for offloading  

and accelerating networking, storage, and security  are carefully assembled into GB300 compute trays.

MVLink is the breakthrough high-speed link that  Nvidia invented to connect multiple GPUs and scale  

up into a massive virtual GPU. The MVLink switch  tray is constructed with MVLink switch chips  

providing 14.4 terabytes per second of all toall  bandwidth. MVLink spines form a custom blindmated  

back plane with 5,000 copper cables connecting all  72 black wells or 144 GPU dies into one giant GPU  

delivering 130 tab per second of all to-all  bandwidth nearly the global internet's peak  

traffic. Skilled technicians assemble each of  these parts into a rack scale AI supercomput.  

In total, 1.2 million components, 2 m of copper  cable, 130 trillion transistors, weighing nearly  

2 tons. From silicon in Arizona and Indiana to  systems in Texas, Blackwell and future Nvidia AI  

factory generations will be built in America,  writing a new chapter in American history  

and industry. America's return to making and  reindustrialization, reignited by the age of AI.

The age of AI has begun. Made  in America. Made for the world.

We are manufacturing in America again.  It is incredible. The first thing that  

President Trump asked me for is bring  manufacturing back. Bring manufacturing  

back because it's it's necessary for national  security. bring manufacturing back because  

we want the jobs and we want that part  of the economy. And nine months later,  

nine months later, we are now manufacturing  in full production Blackwell in Arizona.

Extreme Blackwell GB 200 MV Grace Blackwell  Envy 72 extreme code design gives us 10x  

generationally. It's utterly incredible. Now,  the part that's really incredible is this. This  

is the first AI supercomput we made. This is  in 2016 when I delivered it to a startup in  

San Francisco which turned out to have been  open AI. This was the computer. And in order  

to do the create that computer, we designed  one chip. We designed one new chip in order  

for us to do code design. Now, look at all of  the chips we have to do. This is what it takes.  

You're not going to take one chip and make  a computer 10 times faster. That's not going  

to happen. The way to make computers 10  times faster that we can keep increasing  

the performance exponentially, we can keep  driving cost down exponentially is extreme  

code design and working on all these different  chips at the same time. We now have Ruben back  

home. This is Ruben. This is the Vera Rubin  and and Ruben. Ladies and gentlemen, Ruben

This is this is our third generation MVLink 72  rack scale computer. Third generation GB200 was  

the first one. All of our partners around the  world, I know how hard you guys worked. It was  

insanely hard. It was insanely hard to do. Second  generation, so much smoother. And this generation,  

look at this. Completely cableless. completely  cableless. And this is this is all back in the  

lab now. This is the next generation  Reuben. While we're shipping GB300's,  

uh we're preparing Reuben to be in  production. You know, this time next year,  

maybe slightly earlier. And so, every single  year, we are going to come up with the most  

extreme code design system so that we can keep  driving up performance and keep driving down  

the token generation cost. Look at this. This  is just an incredibly beautiful computer. Now,

so this is amazing. This is 100 pedlops. I  know this doesn't mean anything. 100 pedlops,  

but compared to the DGX1 I delivered to OpenAI  10 years ago, nine years ago, it's 100 times  

the performance right here versus 100 times of  that supercomput. A 100 times a 100 of those,  

let's see, a hundred of those would be like 25 of  these racks all replaced by this one thing. One  

Vera Rubin. Okay. So this is this is the compute  tray and this is so Vera Rubin super chip. Okay.  

And this is the compute tray. This Oh right  here. It's incredibly easy to install. Just  

flip these things open, shove it in. Even  I could do it. Okay. And this is the ver  

Vera Rubin compute tray. If you decide  you wanted to add a special processor,  

we've added another processor. It's called  its context processor because the amount of  

context that we give AIS are larger and larger.  We wanted to read a whole bunch of PDFs before  

it answered question. Wanted to read a whole bunch  of archive papers, watch a whole bunch of videos,  

go learn all this before you answer a question for  me. All of that context processing could be added.  

And so you see on the bottom eight connectx9 new  super nicks. You have CX you have uh CPXs eight  

of them. You have uh blue field 4 this new data  processor two Vera CPUs and four Reuben packages  

or eight Reuben GPUs. All of that in this one  node, completely cableless, 100% liquid cooled.  

And then this new processor, I won't talk too much  about it today. I don't have enough time, but this  

is completely revolutionary. And the reason for  that is because your AIs need to have more and  

more memory. You're interacting with it more.  You wanted to remember our last conversation.  

Everything that you've learned on my behalf,  please don't forget it when I come back next time.  

And so all of that memory is going to create  this thing called KV caching. And that KV caching  

retrieving it, you might have noticed every  time you go into your your your AIS these days,  

it takes longer and longer to refresh and retrieve  all of the previous conversations and and the  

reason for that is we need a revolutionary new  processor and that's called Blue Fuel 4. Next is  

the the ConnectX switch, excuse me, the MVLink  switch which is right here. Okay, this is the  

MVLink switch. This is what makes it possible for  us to con connect all of the computers together.  

And this switch is now several times the bandwidth  of the entire world's peak internet traffic. And  

so that spine is going to communicate and carry  all of that data simultaneously to all of the  

GPUs. On top of that, on top of that, this is the  this is the Spectrum X switch. And this Ethernet  

switch was designed so that all of the processors  could talk to each other at the same time and not  

gum up the network. Gum up the network. That's  very technical. Okay. So, um, so these are the  

these three combined. And then this is the quantum  switch. This is for Infiniband. This is Ethernet.  

We don't care what language you would like  to use, whatever standard you like to use. We  

have great scale out fabrics for you. Whether  it's Infiniban or Quant or Spectrum Ethernet,  

this one uses silicon photonics and is completely  co-acked options. Basically, the laser comes right  

up to the silicon and connects it to our chips.  Okay. So, this is the Spectrum X Ethernet. And so,  

now let's talk about Thank you. Oh, this is this  is what it looks like. This is a rack. This is  

two and a half. This is two uh 2000. This is two  tons. 1.5 million parts. And the spine, this spine  

right here carries the entire internet traffic in  one second. Same speed moves across all of these  

different processors. 100% liquid cooled. All for  the, you know, fastest token generation rate in  

the world. Okay, so that's what a rack looks like.  Now that's one rack. A gigawatt data center would  

have you know call it let's see 16 racks would be  a thousand um and then 500 of those. So whatever  

500 time 16 and so call it 9,000 of these 8,000 of  these would be a one gigawatt data center. Okay.  

And so that's a future AI factory. Now we used,  as you notice, Nvidia started out by designing  

chips and then we started to design systems and  we designed AI supercomputers. Now we're designing  

entire AI factories. Every single time we move  out and we integrate more of the problem to solve,  

we come up with better solutions. We now  build entire AI factories. This is going,  

this AI factory is what we're building for Vera  Rubin and we created a technology that makes it  

possible for all of our partners to integrate into  this factory digitally. Let me show it to you.

The next industrial revolution is here  and with it a new kind of factory. AI  

infrastructure is an ecosystem scale challenge  requiring hundreds of companies to collaborate.

NVIDIA Omniverse DSX is a blueprint for building  and operating gigascale AI factories. For the  

first time, the building, power, and cooling are  co-designed with NVIDIA's AI infrastructure stack.

It starts in the Omniverse digital twin. Jacob's  engineering optimizes compute density and  

layout to maximize token generation according  to power constraints. They aggregate SIM ready  

open USD assets from Seammen's Schneider Electric  Train and Vertive into PTC's product life cycle  

management. Then simulate thermals and electricals  with CUDA accelerated tools from EAB and Cadence.

Once designed, Nvidia partners like Bectal  and Vertive deliver pre-fabricated modules  

factory-built, tested, and ready to plug  in. This shrinks build time significantly,  

achieving faster time to revenues. When the  physical AI factory comes online, the digital  

twin acts as an operating system. Engineers  prompt AI agents from FIDRA and Emerald AI,  

previously trained in the digital twin to  optimize power consumption and reduce strain  

on both the AI factory and the grid. In total,  for a 1 gawatt AI factory, DSX optimizations can  

deliver billions of dollars in additional revenue  per year across Texas, Georgia, and Nevada.  

NVIDIA's partners are bringing DSX to life.  In Virginia, NVIDIA is building an AI factory  

research center using DSX to test and productize  Vera Rubin from infrastructure to software. With  

DSX, NVIDIA partners around the world can build  and bring up AI infrastructure faster than ever.

completely completely in digital long long  before Vera Rubin exists as a real computer  

we've been using it as a digital twin computer  now long before these AI factories exist we  

will use it we will design it we'll plan it  we'll optimize it and we'll operate it as a  

digital twin and so all of our partners that  are working with us I'm incredibly happy for  

all of you supporting us And Gio is here and G ver  Vernova is here. Schlider I I think um I think uh  

uh Olivia is here. Olivia Blum is here. Um uh uh  Seaman's incredible partners. Okay. Roland Bush,  

I think he's watching. Hi Roland. And so anyways,  uh really really great partners working with us.  

In the beginning we had CUDA and we have all  these different ecosystems of software partners.  

Now we have Omniverse DSX and we're building  AI factories and again we have these incredible  

ecosystem of partners working with us. Let's talk  about models. Open source models particularly  

in the last couple years. Several things have  happened. One, open source models have become  

quite capable because of reasoning capabilities.  It has become quite capable because they're  

multimodality and they're incredibly efficient  because of distillation. So all these different  

capabilities have become uh has made open source  models for the very first time incredibly useful  

for developers. They are now the lifeblood of  startups. Lifeblood of startups in different  

industries because obviously as I mentioned  before each one of the industries have its own  

use case it own use cases it own data it owned  data its own flywheels. All of that capability,  

that domain expertise needs to have the ability  to embed into a model. Open source makes that  

possible. Researchers need open-source. Developers  need open-source. Companies around the world,  

we need open source. Open- source models is  really, really important. The United States has  

to lead in open source as well. We have amazing  proprietary models. We have amazing proprietary  

models. We need also amazing open source models.  Our country depends on it. Our startups depend  

on it. And so NVIDIA is dedicating ourselves to  go do that. We are now the largest the largest  

we lead in open-source contribution. We have  23 models in leaderboards. We have all these  

different domains from language models the  physical AI models. I'm going to talk about  

robotics models to biolog biology models. Each one  of these models has enor enormous teams and that's  

one of the reasons why we built supercomputers  for ourselves to enable all these models to be  

created. We have number one speech model, number  one reasoning model, number one physical AI model.  

The number of downloads is really really terrific.  We are dedicated to this and the reason for that  

is because science needs it, researchers need  it, startups need it and companies need it.

I'm delighted that AI startups build on Nvidia.  They do so for several reasons. One, of course,  

our ecosystem is rich. Our tools work great. All  of our tools work on all of our GPUs. Our GPUs  

are everywhere. It's literally in every single  cloud. It's available on prem. You could build  

it yourself. You could you could you know build  up a an enthusiast gaming PC with multiple GPUs  

in it and you could download our software stack  and it it just works. We have the benefit of rich  

developers who are making this ecosystem richer  and richer and richer. So I'm really pleased with  

all of the startups that we're working with. I'm  I'm thankful for that. It is also the case that  

many of these startups are now starting to create  even more ways to enjoy our GPUs. the Cordwaves,  

Nscale, Nbius, Llama, Lambda, all of these  companies, Crusoe companies are building these  

new GPU clouds to serve the startups and I really  appreciate that this is all possible because  

Nvidia is everywhere. We integrate our libraries.  All of the CUDA X libraries I tal talked to you  

about. All the open-source AI models that I talked  about. All of the models that I talked about,  

we integrated into AWS, for example, really  love working with Matt. We integrated into  

Google Cloud, for example, really love working  with Thomas. Each one of these clouds integrate  

NVIDIA GPUs and our computing, our libraries as  well as our models. Love working with Satia over  

at Microsoft Azure. love working with uh Clay  at Oracle. Each one of these clouds integrate  

the NVIDIA stack. As a result, wherever you go,  whichever cloud you use, it works incredibly.  

We also integrate Nvidia libraries into the world  SAS so that each one of these SAS will eventually  

become agentic SAS. I love Bill McDerman's vision  for Service Now. There. Yeah, there you go.  

I think that might have been Bill. Hi, Bill. And  so, Service Now, what is it? 85% of the world's  

enterprise workloads, workflows. SAP, 80% of  the world's commerce. Christian Klein and I are  

working together to integrate NVIDIA libraries,  CUDA X and Nemo and Neotron, all of our AI systems  

into SAP, working with Ceine over at Synopsis,  accelerating the world CAE, CAD, EDA tools  

so that they could be faster and could scale,  helping them create AI agents. One of these days,  

I would love to hire a AI agent, ASIC designers  to work with our ASIC designers. essentially the  

cursor of synopsis if you will. We're working  with uh Annie Rude. Annie Rude here, I saw him  

earlier today. He was part of the pregame show.  Cadence doing incredible work accelerating their  

stack creating AI agents so that we can have  Cadence AI as designers and system designers  

working with us. Today we're announcing a new  one. AI will supercharge productivity. AI will  

transform just about every industry. But AI will  also supercharge cyber security challenges, the  

bad AIs. And so we need an incredible defender. I  can't imagine a better defender than CrowdStrike.  

George George is here. Uh he was Yeah, I saw him  earlier. We are partnering with CrowdStrike to  

make cyber security speed of light to create a  system that has cyber security AI agents in the  

cloud but also incredibly good AI agents on prem  or at the edge. This way you whenever there's a  

threat you are moments away from detecting it. We  need speed and we need a fast agentic AI super a  

super smart AIs. I have a second announcement.  This is the single fastest enterprise enterprise  

company in the world. Probably the single most  important enterprise stack in the world today.  

Palunteer ontology. Anybody from Palunteer  here? I was just talking to Alex earlier. This  

is Palunteer ontology. They take information, they  take data, they take human judgment and they turn  

it into business insight. We work with Palanteer  to accelerate everything Palanteer does so that  

we could do data processing data processing at  a much much larger scale and more speed whether  

it's structured data of the past and of course  we'll have structured data, human recorded data,  

unstructured data and process that data for  our government, for national security and for  

enterprises around the world. process that data at  speed of light and to find insight from it. This  

is what it's going to look like in the future.  Palunteer is going to integrate NVIDIA so that  

we could process at the speed of light and at  extraordinary scale. Okay, Nvidia and Palanteer.

Let's talk about physical AI. Physical AI requires  three computers just as it takes two computers to  

train a language model. One that's to train  it, evaluate it, and then inference it. Okay,  

so that's the large GB 200 that you see.  In order to do it for physical AI, you need  

three computers. You need the computer to train  it. This is GB the Grace Blackwell Invink 72.  

We need a computer that does all of the  simulations that I showed you earlier with  

Omniverse DSX. It basically is a digital twin  for the robot to learn how to be a good robot  

and for the factory to essentially be a digital  twin. That computer is the second computer,  

the omniverse computer. This computer has to be  incredibly good at generative AI and it has to  

be good at computer graphics, sensor simulation,  ray tracing, signal processing, this computer is  

called the omniverse computer. And once we train  the model, simulate that AI inside a digital twin  

and that digital twin could be a digital twin  of a factory as long as well as a whole bunch of  

digital twins of robots. Then you need to operate  that robot. And this is the robotic computer. This  

is this one goes into a self-driving car. Half  of it could go into a robot. Okay? Or you could  

actually have, you know, robots that are quite  agile and quite quite fast in operations. And  

it might take two of these computers. And so this  is the Thor Jetson Thor robotics computer. These  

three computers all run CUDA. And it makes  it possible for us to advance physical AI.  

AI that understand the physical world, understand  laws of physic, causality, permanence, you know,  

physical AI. We have incredible partners working  with us to create the physical AI of factories.  

We're using it ourselves to create our factory in  Texas. Now, once we create the robotic factory,  

we have a bunch of robots that are inside  it. And these robots also need the physical  

AI applies physical AI and works inside  the digital twin. Let's take a look at it.

America is re-industrializing, reshoring  manufacturing across every industry. In Houston,  

Texas, Foxcon is building a state-of-the-art  robotic facility for manufacturing NVIDIA AI  

infrastructure systems. With labor shortages  and skills gaps, digitalization, robotics,  

and physical AI are more important than ever,  the factory is born digital in Omniverse.

Foxcon engineers assemble their virtual  factory in a seaman's digital twin solution  

developed on Omniverse Technologies. Every  system, mechanical, electrical, plumbing,  

is validated before construction. Seaman's  plant simulation runs design space exploration  

optimizations to identify ideal layout. When  a bottleneck appears, engineers update the  

layout with changes managed by Seaman's team  center. In Isaac sim, the same digital twin  

is used to train and simulate robot AIS. In the  assembly area, Fanic manipulators build GB300  

tray modules by manual manipulators from FII and  Skilled AI install bus bars into the trays and  

AMRs shuttle the trays to the test pods. Then  Foxcon uses Omniverse for large-scale sensor  

simulation where robot AIs learn to work as  a fleet. In Omniverse, vision AI agents built  

on NVIDIA Metropolis and Cosmos. Watch the  fleets of robots and workers from above to  

monitor operations and alert Foxcon engineers of  anomalies and safety violations. or even quality  

issues. And to train new employees, agents power  interactive AI coaches for easy worker onboarding.

The age of US re-industrialization is here  with people and robots working together.

That's the the future of manufacturing,  the future of factories. I want to thank  

our partner Foxcon Younglu, the CEO, is  here, but all of these ecosystem partners  

make it possible for us to create the future of  robotic factories. The factory is essentially a  

robot that's orchestrating robots to build  things that are robotic. You know this is  

the amount of software necessary to do this is  so intense that unless you could do it inside  

a digital twin to dis to plan it to design it  to operate inside a digital twin the hopes of  

getting this to work is nearly impossible.  I'm so happy to see also that Caterpillar,  

my friend Joe Joe Creed and his hundredyear-old  company is also incorporating digital twins in  

the way they manufacture. Um these factories will  have future robotic systems and one of the most  

advanced is figure. Brett Atcock is here today. He  just he founded a company three and a half years  

ago. They're worth almost $40 billion. Today  we're working together in training the the AI,  

training the robot, simulating the robot and of  course the robotic computer that goes into figure  

really amazing. Uh I had the benefit of seeing  it. Uh it's really quite quite extraordinary.  

It is very likely that humano robots and uh my  friend Elon is also working on this that this  

is likely going to be one of the largest consumer  new consumer electronics markets and surely one of  

the largest industrial equipment market. Peggy  Johnson and the folks at Agility are working  

with us on robots for warehouse automation. the  folks at Johnson Johnson working with us again  

training the robot, simulating it in digital  twins and also operating the robot. These John  

Johnson and Johnson surgical robots are even  going to perform surgery that are completely  

noninv noninvasive surgery at a precision  the world's never seen before. And of course,  

the cutest robot ever, the cutest robot  ever, the Disney robot. And this is this is  

um something really close to our heart. We're  working with Disney research on a entirely new  

framework and sim simulation platform uh based  on revolutionary technology called Newton. And  

that Newton uh simulator makes it possible  for the the robot to learn how to be a good  

robot inside a physically aware physically  based environment. Let's take a look at it.

Excuse

me.

blue. Ladies and gentlemen, Disney Blue. Tell  me that's not adorable. He's not adorable. We  

all want one. We all want one. Now, remember  everything you were just seeing, that is not  

animation. It's not a movie. It's a simulation.  That simulation is an omniverse. Omniverse,  

the digital twin. So these digital twins  of factories, digital twins of warehouses,  

digital twins of surgical rooms, digital twins  where blue could learn how to manipulate and  

navigate and you know interact with the world.  All completely done in real time. This is going  

to be the largest consumer electronics product  line in the world. Some of them are just really  

working incredibly well now. This is a future  of human or robotics and of course blue. Okay.

Now, human robots is still in development. But  meanwhile, there's one robot that is clearly at  

an inflection point and it is basically here and  that is a robot on wheels. This is a robo taxi.  

A robo taxi is essentially an AI chauffeur. Now,  one of the things that we're doing today, we're  

announcing the NVIDIA drive Hyperion. This is a  big deal. We created this architecture so that  

every car company in the world could create cars,  vehicles could be commercial, could be passenger,  

could be dedicated to robo taxi. Create vehicles  that are robo taxi ready. The sensor suite with  

surround cameras and radars and LAR make it  possible for us to achieve the highest level of  

surround cocoon sensor perception and redundancy  necessary for the highest level of safety.  

Hyperion drive drive Hyperion is now designed into  Lucid Mercedes-Benz my friend Ola Ken Canel Kenius  

um the folks at Stalantis and there are many other  cars coming and once you have a basic standard  

platform then developers of AV systems and there's  so many talented ones wave wabby Aurora Momenta  

Neuro there's so many of them we ride there's  so many of them that can then take their AV  

system and run it on the standard chassis.  Basically, the standard chassis has now become  

a computing platform on wheels. And because it's  standard and the sensor suite is comprehensive,  

all of them could deploy their AI  to it. Let's take a quick look.

Okay, that's the be that's beautiful  San Francisco. And as you could see,  

as you could see, robo taxis inflection point  is about to get here. And in the future,  

a trillion miles a year that are driven, a  100 million cars made each year. There's some  

50 million taxis around the world. It's going to  be augmented by a whole bunch of robo taxis. So,  

it's going to be a very large market to connect  it and deploy it around the world. Today,  

we're announcing a partnership with Uber.  Uber, Derek, Dara K, Dara is going to we're  

working together to connect these Nvidia drive  Hyperion cars into a global network and now in  

the future you'll you know be able to hail up  one of these cars and the ecosystem is going  

to be incredibly rich and we'll have Hyperion  or Robo taxi cars all over the world. This is  

going to be a new computing platform for us and  I'm expecting it to be quite successful. Okay.

So this is what we talked about today. We talked  about a large large number of things we spoke  

about. Remember at the core of this is two or  two platform transitions from general purpose  

computing to accelerated computing. NVIDIA  CUDA and those suite of libraries called  

CUDA X has enabled us to address practically every  industry and we're at the inflection point. It is  

now growing as a virtual cycle would suggest.  The second inflection point is now upon us.  

The second platform transition AI from classical  handwritten software to artificial intelligence.  

two platform transitioning happening at the same  time which is the reason why we're feeling such  

incredible growth. quantum quantum computing.  We spoke about open models. We spoke about we  

spoke about enterprise with crowd strike and  uh palunteer accelerating their platforms.  

Uh we spoke about robotics a new potentially  one of the largest consumer electronics and  

industrial manufacturing sectors. And of course  we spoke about 6G. Nvidia has new platforms for  

6G. We call it ARC. We have a new platform for  robotics cars. We call that Hyperion. We have  

new platforms even for factories. Two types of  factories. The AI factory we call that DSX. And  

then factories with AI we call that mega. And  so now we're also manufacturing in America.  

Ladies and gentlemen, thank you for joining us  today and thank you for allowing me to bring

Thank Thank you for for allowing us to bring  GTC to Washington DC. We're going to do it  

hopefully every year. And thank you all for your  service and making America great again. Thank you.

We start with a handshake. Solid  and true. One step at a time,  

we're breaking through. Brick by brick,  we're stacking dreams high. Side by side,  

we'll touch the sky. Handshakes and high hopes  we're making our way. Shoulder to shoulder.  

Come what may. Shared vision brighter than the  sun. Friendship and business rolling as one.

Plans on paper but hearts in sink. Building  together faster than you think. Laughter's  

the glue in the grind we share. We've  got the spark. We're going somewhere.  

Handshakes and high hopes. We're making  our way. Shoulder to shoulder. Come what  

may. Shared vision brighter than the sun.  Friendship and business. Handshakes and  

high. We're making our way shoulder to come  one way. Share vision brighter than the sun.

Business rolling as one.