Jensen Huang Special Address from NVIDIA AI Summit Japan

Build me car on an empty road. Hello, I'm James, a digital human. Feel free to ask me anything about NVIDIA and its products. Ladies and gentlemen, please take your seats and welcome to the stage, NVIDIA Founder and CEO, Jensen Huang. Hello, Japan! Hello, Japan. Welcome to NVIDIA AI Summit. Everything you saw just now was simulated. Everything was simulation. No animation. At the core, NVIDIA is a simulation technology company. We simulate physics. We simulate virtual worlds, we simulate intelligence. Through our simulations, we help you predict the future. So in many ways, NVIDIA builds time machines. Today

we're going to talk to you about some of our latest breakthroughs. But most importantly, this is an event about the Japan ecosystem. We have so many partners here. 350 startups, 250,000 developers, hundreds of companies. We've been coming here for a very long time. Japan has been very dear to us since the very early foundings of our company. It is here in Japan we did many firsts, many firsts for NVIDIA. It was our first game developer that worked with us. SEGA, Yu Suzuki, the famous developer, 3D game developer, who first worked with us to port SEGA's

incredible 3D games to NVIDIA's GPUs. It was the first time that someone used NVIDIA CUDA to build a supercomputer. The Tsubame 1.2 at Tokyo Tech, which enabled NVIDIA to utilize our GPUs to advance scientific computing. Japan has been first in so many ways. It was also the first time that we were able to create mobile processors, which led to one of our very, very dear projects, the Nintendo Switch. So many firsts. Now we're in the beginning of a new age. The AI revolution: a new industry, extraordinary technology changes. This is a time that is very

exciting, but it's also very critical. So we're here to partner with the Japan ecosystem, the amazing companies here, to bring AI to Japan. So that we can all together take advantage of this extraordinary opportunity ahead of us. We have many partners here today that I would like to thank GMO Internet Group, Hewlett Packard Enterprise, Microsoft Azure, MITSUI and Company are the platinum sponsors. I want to thank all of you for that. Thank you. And 56 other sponsors. Thank you all for coming and thank you for supporting us. NVIDIA invented accelerated computing. Accelerated computing does not

replace the CPU. In fact, we were just about the only company in computing who did not want to replace the CPU, but to augment the CPU, so that we can take the workloads that are very computationally intensive and offload them onto the GPU. These are GPUs, that's the CPU. Working together, we can take advantage of the best capabilities of both processors, one that is extremely good at sequential processing, the CPU, and a processor that's incredibly good at parallel processing, called the GPU. This I'll talk about in just a second. But this is accelerated computing. Not

just parallel computing, but accelerated computing. CPU and GPU working together. This computing model is brand new to the world. In fact, just using the CPU has been around since 1964, the year after my birth. 60 years in the making. The vast majority of everything we see in the world today running on computers are running on CPUs. But there's a new change, a fundamental change in the computing model. But in order to do that, you cannot just take the CPU software that's running sequentially and put it on a GPU to run parallel. We have to create

a whole bunch of new algorithms. Just as OpenGL made it possible for computer graphics applications to connect to acceleration through the graphics processor, we had to create many domain-specific libraries like OpenGL, but for many different applications. These are some of the 350 different libraries we have in our company, some very, very important libraries. CuLitho was for computational lithography, takes computational lithography. Making the mask of chips may require many, many weeks because there are so many layers. We take many weeks of computation and reduce it down to hours. We could, of course, speed up the cycle

time of building chips, but very importantly, we also make it possible to advance the algorithms of lithography much, much more sophisticated so that we can advance semiconductor physics well beyond two nanometers, one nanometer and beyond. So computational lithography is going to be accelerated by cuLitho. CuDSS for Sparse Solver, AI Aerial— I'll talk about that today. An incredible new library that makes it possible for this computer to run the 5G radio stack. Basically a radio in real time running on the CUDA accelerator. CUDA-Q for quantum simulation, quantum circuit simulations, Parabricks for gene sequencing, cuVS for vector

storage, or vector databases for indexing and querying vector databases that is used for AI. CuPyNumeric is a numerical processing library. NumPy is the most popular numerical processing library in the world. Five million different developers use it. It has been downloaded 30 million times just last month. Incredible number of downloads. It is now fully accelerated for multi-GPU multi-node. CuPyNumeric, please go look it up. It's incredible. CuDF for data frame processing, structured data processing like SQL, Pandas, Polars. CuOPT, the classic incredibly complicated traveling salesperson problem. This problem has now been accelerated for CuOPT, hundreds of times faster.

Modulus for AI physics, and then one of the most important libraries we ever created, called cuDNN. CuDNN for deep neural networks, which processes CuDNN for deep neural networks, which processes the different layers, the various layers of the deep learning stack. It enabled us to do something extraordinary. By creating cuDNN, and accelerating and democratizing deep learning, We made it possible in the last decade to scale artificial intelligence and scale machine learning by 1 million times. Increasing the scale of machine learning by 1 million times enabled, of course, the incredible breakthrough that we now know—ChatGPT. The advent

of artificial intelligence. CuDNN did something very special. It enabled a change in the way that software is done. This is the way that software was done before. Software 1.0. Software programmers writing code to describe an algorithm— that's a function. That function is software. You then apply input, it predicts an output. That code written by human runs best on CPUs. Software 1.0 was coding, writing software that runs on CPUs. Software 2.0, because the computer is now so incredibly fast, you can give it an enormous amount of examples—observed data— so that it can learn, it can predict

what the function is by itself. We call that software 2.0. So instead of coding, it is now machine learning. Instead of code running on CPU, it is now neural networks running on GPUs. And neural networks, these neural networks that are running on GPUs Is now the formation of a new operating system, a new way of using computers. The operating system of the modern computer, Large Language Models. This machine learning approach has proven to be incredibly scalable. You can use it for all kinds of things. Of course, digitized text, languages, of course, digitized sound, speech, images,

video. It could be multimodal. You could teach it amino acid sequences. You could teach it to understand just about anything, anything where you have large observed data. Well, that was the first step, Was to understand the meaning of the data. Just by studying an enormous amount of text on the internet, we were able to understand the words, the vocabulary, syntax, grammar, and even the meaning of the words. by finding patterns and relationships. Using the same approach, we're now able to not only understand the meaning of the different data types, connect it to different modalities— for

example words and images. Connect the image of the word 'cat' and the image of a cat are now connected together, learning a multi-modality. We can now even translate and generate, so we can understand data of all kinds, and we can generate intelligent information, intelligence of all kinds. Well, if you look at all the amazing startups that are being created and the amazing applications that are being created, you can find it in this slide in one of the two combinations. One side to the other side, Text to text, would be summarization, question and answering, text generation,

storytelling. Video to text would be captioning. Image to text— image recognition. Text to images— image generation like Midjourney. Text to video—video creation like Runway ML. All of these different combinations are truly breakthrough. You can even have proteins to text, explain what a protein does. Text to chemicals, describe properties of a chemical that might be a successful drug for drug discovery. You can even have video and text to machine articulation, robotics. Each one of these combinations is a new industry, new company, new application use case, incredible Cambrian explosion of the number of applications have now been

created. And we're just at the beginning. One of the properties of machine learning, of course, is that the larger the brain, the more data we can teach it, the smarter it becomes. We call it the scaling law. There's every evidence that as we scale up the size of the models, the amount of training data, the effectiveness, the quality, the performance of the intelligence improves. Every single year, the industry is scaling up the size of the models by 2X or so, which correspondingly need 2X the data, and therefore we need 4X the amount of compute. The

amount of computing resource necessary to drive to the next level of artificial intelligence is extraordinary. We call that the scaling law, the training scaling law. Pre-training is part of it, post-training is part of it. Post-training with reinforcement learning, human feedback, reinforcement learning, AI feedback. So many different ways now of using synthetic data generation in the post-training stage. So training, pre-training, post-training is enjoying very significant scaling and we're continuing to see excellent results. Well, when Strawberry or OpenAI’s o1 was announced, it exposed the world to a new type of inference. Inference is when you interact with

the AI, just like ChatGPT. But ChatGPT is one shot. You ask it a question, you ask it to do something for you. Whatever question you have, whatever prompts you provide, through one shot, it delivers an answer. However, we know that thinking is oftentimes more than just one shot, and thinking requires us to maybe do multi-plans, multiple potential answers that we choose the best one from. Just like when we're thinking. We might reflect on the answer before we deliver the answer. Reflection. We might take a problem and break it down into step by step by step,

chain of thought. There are many different technologies that we've invented that makes it possible for inference to perform better and better as we apply more and more compute. Now we have the second scaling law: inference scaling law. Not just generation of the next word, but thinking, reflecting, planning. These two simultaneous scale laws are going to require us to drive computing at extraordinary speeds. Every single time when we deliver a new generation, a new architecture, we increase the performance by X factors, But we also decrease the power by the same X factor. We decrease the cost

by the same X factor. So driving up performance is exactly the same as reducing cost. Driving up the performance is exactly the same as reducing energy. And so therefore as the world continues to absorb and embrace artificial intelligence, it is our mission, it is our duty, it is our duty, to continue to drive the performance up as fast as we can. In the process, expanding the reach of artificial intelligence, driving up its effectiveness, driving down its cost, driving down its power consumption. That's the reason why we went to a one year cycle. However, AI is

not a chip problem. These AI systems are enormous. This is the Blackwell system. Blackwell is the name of a GPU, but it's also the name of this entire system. The GPU is extraordinary in itself. There are two Blackwell dies. Each Blackwell die is the largest chip the world's ever made. 104 billion transistors made by TSMC and their most advanced four nanometer node. Two of these Blackwell dies are connected together across 10TB/s, low energy link. Right in the middle, right there, that line, that seam. Thousands of interconnections between the two dies. Ten terabytes per second. It's

connected by eight HBM3E memories from SK Hynix and from Micron. And these memories together run at 8 TB/s. And 8 TB/s, these two GPUs are connected to the CPU with another low, very low energy, very energy efficient series, 1 TB/s. Each one of the GPUs are connected through NVLink at 1.8 TB/s. That's a lot of terabytes per second. And the reason for that is because this system cannot work alone. Even the most advanced computer the world has ever made cannot work alone for artificial intelligence. Sometimes it has to work with thousands of other computers like

this, nodes like this, together as one computer, and sometimes they have to work separately because they are responding to a different customer, different query. So sometimes separately, sometimes as one. In order to enable the GPUs to work as one, we of course have networking, two ConnectX-7s that connect this GPU with thousands of other GPUs, but we still need this NVLink. This NVLink allows us to connect the few GPUs in one rack that's standing behind me. That one rack behind me, that one rack is connected to, with this NVLink 5.0, 1.8TB per second, 35 times higher

bandwidth than the highest bandwidth networking in the world, which allows us to connect all of these GPUs together to this NVLink switch. There are nine NVLink switches in one rack. Each rack has 72 computers like this connected through this spine. This is the NVLink spine. This is cables, copper. 50 pounds of copper, driven directly by this incredible series, incredible IO we call NVLink. They connect into, connect into the computer, into NVLink this way, and this switch connects all of these computers together as one. And so what results, Are 72 of these computers connected as one

large GPU. One incredibly large GPU. From the software perspective, it is just one giant chip, and these racks, these NVLinked 72 systems. This one rack. This one rack is 3,000 pounds. It is impossible to put on this stage otherwise I'd show it to you. That is 3,000 pounds and 120kW. That is, I have my my friends here. That's many, many Nintendo Switch power. It is not portable but it's very powerful. And so this is the Blackwell system. We designed this so that it can be configured as one superpod like this, or one entire gigantic data

center with thousands and thousands of them, hopefully hundreds of thousands of them, And they're connected to them by the switches. Some of them are Quantum InfiniBand switches. If you want to have a dedicated AI factory or a Spectrum-X, NVIDIA Spectrum-X revolutionary ethernet system that you can integrate into your existing ethernet environments. We can build AI supercomputers with these, we can integrate them into enterprise data centers, into hyperscalers, or Configure them for the edge. The Blackwell system is not only incredibly powerful, it is also incredibly adaptable, so that it can fit into every corner of the

world's computing infrastructure. So this is this is Blackwell. On top of Blackwell, of course, is the computer. But most importantly, without all of the software that runs on top of it, this computer is just simply impossible to operate. When you see these computers with all of the liquid cooling, all of the wiring, the mind will blow up. How do you program such an incredible computer? This is where NVIDIA's software stack, this is where all of our effort in CUDA, NCCL, all of our Megatron, Megatron-Core, all of the software that we created, TensorRT LLM, Triton, all

of the software that we created over the years integrated into the system, makes it possible for everyone, anyone, to deploy AI supercomputers around the world. And then, of course, on top we have AI software that makes it easy for people to build AI. And so what is AI? What is AI? We talk about AI in a lot of different ways, but I think there are two types of AIs that will be extremely popular. And there are two mental models that I think are very helpful. It's very helpful to me. The first AI is basically a

digital AI worker. These AI workers can understand, they can plan and they can take action. Sometimes the digital AI workers are being asked to execute a marketing campaign, support a customer, come up with a manufacturing supply chain plan, optimize a chip, help us write software. Maybe be a research assistant, a lab assistant in the drug discovery industry. Maybe this agent you know is a tutor to the CEO. Maybe there's a tutor for all of our employees. These AI, these digital AI workers, we call them AI agents are essentially like digital employees. And just like digital

employees, you have to train them. You have to create data to welcome them to your company, teach them about your company. You train them for their particular skills, depending on what function you would like them to have, you evaluate them after you're done training them to make sure that they learned what they're supposed to learn, you guardrail them to make sure that they perform the job they're asked to do, and not the jobs they're not asked to do. And of course, you operate them, you deploy them. Provide them energy from Blackwell, the AI tokens from

Blackwell. And they interact with other agents to work as a team to solve problems. Well, you're going to see all kinds of different agents. And we created several things to make it easier for the ecosystem to be able to build AI agents for companies. NVIDIA is not in the service business. And we don't create, we don't deliver final product, we don't deliver solutions, but we do deliver the enabling technology to make it possible for the ecosystem to create AI, to deliver AI, to continuously improve AI. The AI agent lifecycle libraries, the lifecycle platform is called

NeMo, And NeMo has libraries for each one of the stages that I mentioned, from data curation to training, to fine tuning to synthetic data generation to evaluation, to guard railing. And these libraries are integrated into workflows and frameworks all over the world. We're working with AI startups, service providers like Accenture and Deloitte, companies all over the world to bring this to all of the large companies. We also work with ISVs like ServiceNow so that they can create agents that use ServiceNow. Today, you use ServiceNow by licensing the platform and your employees interact with the ServiceNow

platform to get assistance. In the future, ServiceNow will also provide a large number of AI agents that you can rent. Essentially digital employees that you can rent to help you solve problems. We're working with ServiceNow, we're working with SAP, Cadence, Ansys, companies all over the world, Snowflake, companies all over the world so that we can all build agents that can be helpful to you in driving productivity in your company. Now, these agents are able to understand, reason, plan, take action. And these agents are collections or systems of AI models. It's not just one AI model,

but a system of AI models. And NeMo helps us build those. We also create pre-trained AI models that we package up in what is called a NIM. And so these NIMs are microservices. They're basically AI packaged. In the old days, software was packaged in a box and they'd come delivered with CD-ROMs. Today, AI is packaged in a microservice and inside the software is smart. You could talk to the software and you could talk to the software because it understands what you mean, and you can connect the software with other software. You can connect this AI

with other AI, and together you could create essentially an agent, an AI agent. So this is the first thing. Let me give you an example of some of these agents. Agenetic AI is transforming every enterprise using sophisticated reasoning and iterative planning to solve complex, multi-step problems. AI agents help marketing campaigns go live faster with instant insights, helping optimize supply chain operations, saving hundreds of millions in costs, and reduce software security processes from days to seconds by helping analysts triage vulnerabilities. What makes agentic AI so powerful is its ability to turn data into knowledge, and knowledge

into action. A digital agent in this example can educate individuals with insights from a set of informationally dense research papers. It was built using NVIDIA AI Blueprints. These are reference workflows featuring NVIDIA acceleration libraries, SDKs, and NIM microservices that help you rapidly build and deploy AI apps. The multi-modal PDF Data Extraction Blueprint helps build a data ingest pipeline, while the Digital Human Blueprint provides smooth, human-like interactions. “Hi, I’m James.” A digital agent ingests PDF research papers, including complex data like images, charts and tables, and generates a high level summary delivered through an interactive digital human

interface. “What an exciting breakthrough in weather forecasting. The development of CorrDiff, a new generative model, Is a significant step forward in accurately predicting weather patterns by combining a U-Net regression model with a diffusion model.” James can also answer questions or generate new content based on the papers. NVIDIA AI gives enterprises the power to automate processes, tap into real time insights, and enhance workflow efficiency. AI agents. Three parts: NeMo, NIMs, and Blueprints. These are all references. They're available to you in source form so that you could use it as you like and build your AI agent

workforce. None of these agents can do 100% of anyone's task. Anybody's job. None of the agents can do 100%. However, all of the agents will be able to do work for 50% of your work. This is the great achievement. Instead of thinking about AI as replacing the work of 50% of the people, you should think that AI will do 50% of the work for 100% of the people. By thinking that way, you realize that AI will help boost your company's productivity, boost your productivity. You know, people have asked me, you know, is AI going to

take your job? And I always say, because it's true, AI will not take your job. AI used by somebody else, will take your job. And so be sure to activate using AI as soon as you can. So the first is digital AI agents. Digital. These are digital AIs. The second application is physical AI. The same basic technology is now embodied, sits inside a mechanical system. Of course, robotics is going to be one of the most important industries in the world. Until now, robotics has been limited and the reason is very, very clear. In fact, here

in Japan, 50% of the world's manufacturing robots are built. Kawasaki, FANUC, Yaskawa, Mitsubishi. Four of the leaders that build half of the world's robotic systems. As much as robots have driven the productivity of manufacturing, it has been very difficult to expand. The robotics industry has been largely flat for a long time. And the reason for that is because it is too specific. It is not flexible enough. It is not flexible enough to be able to apply to different scenarios and different conditions and different work. We need AI that is much, much more flexible, that it

can adapt and learn by itself. Notice the technology that we described up until now, agentic AI. Irrespective of who you are, you should be able to interact with the agents. It can give you response. Of course, sometimes the response is not as good as the response that you would produce, but many of them are in fact even better than what we can produce. And so we can now apply this general AI technology into the world of embodied AI or physical AI, or otherwise known as robotics. In order to enable robotics to happen, we need to

build three computers. The first computer is training the AI, just like we did with all of the examples I've given you so far. Second is to simulate the AI, you need to give the AI a place to practice, A place to learn, a place to retreat, to receive synthetic data that it can learn from. We call that Omniverse. Omniverse is our virtual world digital twin suite of libraries that could be used for creating physical AIs, robotics. Omniverse, then, after validation, after training, after evaluation, then you can take the model and put it into a physical

robot. Inside that we have processors that are designed for robotics. We call it Jetson Thor. Thor is a robotics processor designed for humanoid robots. This loop goes on and on and on. Just as there is a NeMo AI agent life cycle platform, There is this Omniverse platform that enables you to create AI. Well, ultimately, what you want is the AI— notice on the left—it sees a world, tt sees the video, it sees the surrounding, the circumstance. You tell it what you want. And this AI will generate articulation motion. Just as we take text, we can

generate video, we can take text and generate chemicals for drugs. We can take text and generate articulation motion. So this concept is very similar to generative AI. And this is the reason why we think That now we have the necessary technology between Omniverse and all the computers that we built, these three computers and the latest generative AI technology, that the time has come for humanoid robotics. Now, why is humanoid robotics so hard? Well, obviously the software developed for humanoid robots is extremely hard. However, the benefit is incredible. There are only two robotic systems that can

be easily deployed into the world. The first robot is a self-driving car, and the reason for that is because we created the world to adapt to cars. The second is humanoid robots. These two robotic systems could be deployed In brownfields anywhere in the world. Because we created the world for us. This is both extraordinarily hard technology that the time has come, but the impact can be enormous. Well, this last week at the Conference for Robotics Learning, we announced a new framework that is super important. It's called Isaac Lab. Isaac Lab is a reinforcement learning virtual

simulation system that allows you to teach humanoid robots how to be humanoid robots. And on top of it, we have several workflows that we have created. The first workflow is Groot-Mimic. Groot-Mimic is a framework for demonstrating to the robot How to perform a task. You use human demonstration and then mimic that environment, using domain randomization generates hundreds of other examples like your demonstration. So that the robot can learn to generalize. Otherwise it can only perform that very specific task. Using Mimic, we can generalize its learning. The second is Groot-Gen. Groot-Gen, using generative AI technology in

Omniverse, we can create an enormous number of random, domain randomized examples of environments and the actions that we would like the robot to perform. And so we're generating a whole bunch of tests, evaluation systems, evaluation scenarios that the robot can try to perform and improve itself, learn how to be a good robot. And the third is Groot-Control. Groot-Control is a model distillation framework that allows us to take all of the skills that we've learned and distill it into one unified model that allows the robot to perform kinematic skills. Not only will robots be autonomous, but

remember that the future factories will also be robotic. And so these factories are going to be robotic factories that are orchestrating robots, building mechanical systems that are robotic. Let me show it to you. Physical AI embodies robots like self-driving cars that safely navigate the real world, manipulators that perform complex industrial tasks, and humanoid robots who work collaboratively alongside us. Plants and factories will be embodied by physical AI capable of monitoring and adjusting its operations or speaking to us. NVIDIA builds three computers to enable developers to create physical AI. The models are first trained on DGX,

then the AI is fine-tuned and tested using reinforcement learning physics feedback in Omniverse. And the trained AI runs on NVIDIA Jetson AGX robotics computers. NVIDIA Omniverse is a physics-based operating system for physical AI simulation. Robots learn and fine tune their skills in Isaac Lab, a robot gym built on Omniverse. And with GR00T workflows like GR00T-Gen to generate diverse learning environments and layouts, GR00T-Mimic to generate large scale synthetic motion datasets based on a small number of real-world captures, and GR00T-Control for neural whole body control. This is just one robot. Future factories will orchestrate teams of robots

and monitor entire operations through thousands of sensors. For factory digital twins, they use an Omniverse blueprint called Mega. With Mega, the factory digital twin is populated with virtual robots and their AI models—the robots' brains. The robots execute a task by perceiving their environment, reasoning, planning their next motion, and finally converting it to actions. These actions are simulated in the environment by the World Simulator in Omniverse, and the results are perceived by the robot brains through Omniverse sensor simulation. Based on the sensor simulations, the robot brains decide the next action and the loop continues while Mega

precisely tracks the state and position of everything in the factory digital twin. This software in the loop testing brings software defined processes to physical spaces and embodiments, letting industrial enterprises simulate and validate changes in an Omniverse digital twin before deploying to the physical world, saving massive risk and cost. The era of physical AI is here, transforming the world's heavy industries and robotics. Incredible times. So we have two robotic systems. One that's digital. We call AI agents, and you use them in your office, collaborating with your employees. And then the second is a physical AI, physical

AI system, robotics. And these physical AIs will be products that the company's built. And so companies will use AI to increase the productivity of our employees. And we will use AI to drive and power the products that we sell. Car companies will have two factories in the future one factory to build cars, one factory to produce the AI that runs in the car. Well, this is the robotics revolution. So much activity is going around the world and I can't imagine a better country to lead the robotics AI revolution than Japan. And the reason for that

is, as you know, this country loves robots. You love robots. You have created some of the world's best robots. These are the robots we grew up with. These are the robots we've loved our whole lives. I didn't even show some of my favorites, Mazinger Z, Gundam. I hope that Japan will take advantage of the latest breakthroughs in artificial intelligence and combine that with your expertise in mechatronics. No country in the world has greater skills in mechatronics than Japan. This is an extraordinary opportunity that you must seize. And so I hope that we can work together

to make that dream possible. NVIDIA AI in Japan is doing incredibly well. We have so many partners here. We have partners that are building large language models. Institute of Science Tokyo, Rakuten, SoftBank Intuitions, NTT, Fujitsu, NEC, Nagoya University, Kotoba Technologies. We also have if you go to the upper right, the AI clouds. AIST, SoftBank, Sakura Internet, GMO Internet Group, HIGHRESO, KDDI, Rutilea, building AI clouds here for the ecosystem to flourish here in Japan. So many robotics companies are starting to understand the capabilities that AI is now providing to take advantage of the opportunity, Yaskawa, Toyota,

Kawasaki, Rapyuta. Medical imaging systems— Canon, Fuji Films, Olympus— all taking advantage of AI because in the future, these medical instruments will be much more autonomous. It's almost like a nurse AI agent inside the medical instrument, helping the nurse guide the diagnosis. So many different ways, the drug discovery industry, so many different ways that AI is being used. And so I'm delighted. I'm delighted by the advancements here. And we want to go even faster to take advantage of the revolution of AI. Well, this industry is changing. As I said earlier, The computer industry has fundamentally changed.

From coding that runs on CPU to now machine learning running on GPUs. From an industry that produced software, we have now become an industry that is manufacturing artificial intelligence, and the artificial intelligence is produced in factories. They're running 24/7. When you license software, you install it onto your computer. The manufacturing, the distribution of that software is complete. However, intelligence is never complete. And you're interacting with all of the AIs, Whether they're AI agents or AI robots. And the tokens, the intelligence is expressed in tokens, which is the unit of intelligence. It's a number. And these

numbers are constituted. These tokens are constituted in ways that becomes intelligence and words, intelligence in steering wheels, intelligence for self-driving cars, intelligence in motors to articulate human robots, intelligence in proteins and chemicals and drug discovery. All of these tokens are being produced in these factories, these infrastructure, these factories never existed before. It's a brand new thing, Which is the reason why we're seeing so much development around the world. For the very first time, we have a new industry, a new factory producing something brand new, that we call artificial intelligence. These factories will be built by

companies. They'll be built. Every company will be an AI manufacturer. Of course, no company can afford not to manufacture, produce, artificial intelligence. How can any company afford not to produce intelligence? How can any country afford not to produce intelligence? You don't have to produce chips. You don't have to produce software, but you have to produce intelligence. It is vital. It is core to who you are. It is core to who we are. And so we have the new industry, AI factories. And that's the reason why I call it. It's a new industrial revolution. The last

time this happened was 300 years ago when electricity was discovered, the generation of electricity and the distribution of electricity and a new type of factory was created. And that new factory was a power plant. And then a new industry was created called energy. Several hundred years ago, there was no such thing as an energy industry. It happened during the Industrial Revolution. Now we have a new industry, never existed before. Artificial intelligence sits on top of the computer industry, but it's utilized, it is created by every industry. You have to create your own AI. The drug

industry—creates your own AI. The automotive industry—creates your own AI. The robotics industry—creates your own AI. Every industry, every company, every country must produce your own AI. A new industrial revolution. I have a very big announcement to make. Today we're announcing that we're partnering with SoftBank to bring and to build an AI infrastructure for Japan. Together, we're going to build Japan's largest AI factory. It's going to be built out of NVIDIA DGX. When it's built, it'll have 25 AI exaflops. Just remember, the largest supercomputer in the world just recently was 1 exaflop. This is 25 exaflops

for an AI factory to produce the AI. But in order to distribute the AI, SoftBank is going to integrate NVIDIA's Aerial, which was the engine that I mentioned earlier, running the 5G radio on CUDA. By doing that, we can unify and combine the radio computers, the basebands and the AI computers from 5G-RAN. We can now evolve and reinvent the telecommunications network into AI-RAN. It will be able to carry voice, data, video, but in the future we will also carry AI, a new type, a new type of information: intelligence. This will be distributed across SoftBank's 200,000

sites here in Japan, serve 55 million customers. AI factory and to produce the AI, and AI distribution networks to distribute AI-RAN to distribute the AI. We're also going to put on top of it a new type of store, an AI store, so that the AIs that were created by SoftBank and the AIs created by third parties could be provided to the 55 million customers. So we will build those applications on top of NVIDIA AI enterprise, as I mentioned, showed you before and there will be a new store that makes AI available to everybody. This is

just a magnificent development. What will result is an AI grid that runs across Japan. Now this AI grid will be part of infrastructure and one of the most important infrastructures. Remember, you need factories and roads, part of the infrastructure So that you could just make and distribute goods. You need energy and communications, it's part of the infrastructure. Every time you create something fundamentally new for the infrastructure, new industries and new companies are created, new economic opportunities, new prosperity. If not for roads and factories, how would we have the Industrial Revolution? If not for energy and

communications, how would we have the IT revolution? Each one of these new infrastructures open up new opportunities, and so it's incredibly exciting for me to partner with SoftBank to make this possible in Japan. Miyakawa-san's team, they should be in the audience. Partnering with you is incredible. And, really, really happy that we're doing this. This is completely revolutionary. This is the first of its kind to transform the telecommunications network, the communications network, into an AI network. Well, let me show you what you can do. There's some amazing things you could do. For example, I'm standing underneath

one of the base stations, one of the radio towers, and the car has video, and the car’s video is streamed to the radio tower and the radio tower has AI. This radio tower has video intelligence. It has vision intelligence. So it could see what the car sees and it can understand what the car is seeing. That AI model may be too heavy to put inside the car, but it's not too heavy to put in the base station. And using the video that's streamed to the base station, it can understand anything that's happening to the car

and surrounding. Okay, so this is just one example of using AI at the edge to keep people safe. Or maybe this is the air traffic control, essentially, for self-driving cars. The applications are endless. We can also use this basic concept To turn a whole factory into an AI. This is a factory on this side. Lots of cameras. The cameras are streamed to the base station. And what's amazing is this: because of all the cameras and the AI model in the AI-RAN, now this factory became an AI. You can talk to the factory. Ask the factory,

what is happening? Ask the factory, were there any accidents? Is there anything abnormal happening? Was anybody injured today? Could it give you a daily report? You just have to ask the factory, because the factory now has turned into an AI. That AI model doesn't have to run in the factory. That AI model can run in the SoftBank radio. Okay, so that's another example. But the countless examples, you can basically turn every single physical object into an AI, a stadium, a road, a factory, a warehouse, an office, a building. They can all become AIs and you

can just talk to it, just like you can talk to ChatGPT. Okay. What's the condition of the aisles? Any obstruction or spills? You're just talking to the factory. And the factory has observed everything. It understands what it sees. It can reason and it can plan an action or just talk to you. And here it says "No, the aisles in the warehouse did not have any obstruction, spills or hazards. The conditions of the aisles in the video appears to be well-organized, clean, and free of obstacles or hazards." Okay, so you're talking to the factory. It's incredible.

You're talking to the warehouse. You're talking to the car. Because all of these have now become intelligent. Well, I have a very special guest to talk about our announcement today and also to talk about bringing AI to Japan. And this friend of mine, this friend of mine. You might know him. That's the two of us. You might know him. This friend of mine is really, genuinely unique. Where is Masa? Where is Son-san? Is he here? - He’s here. Hey! Son-san. Ladies and gentlemen. The great Masa. Great Jensen. Let me let me tell you something. I

don't know if you know this. So we've been in, I've been in the technology industry a long time. Started in the PC revolution. The computer industry went from PC to Internet, to cloud, to mobile cloud, to AI. Long journey. Masa is the only entrepreneur only entrepreneur, only innovator in the world that has selected the winner and partnered with the winner, in every single generation. Remember, it was Masa that brought Bill Gates to Japan. It was Masa that invented or brought, Jerry Yang to Japan. It was Masa that made it possible for China's cloud industry to

happen— Alibaba. It was Masa that brought Steve Jobs to Japan and the iPhone. And maybe many of you probably don't know this, but at one point Masa was the largest shareholder of NVIDIA. Yes. Oh. It’s okay. We can we can cry. We can cry together. Yeah. Sit down. Sit. How did you do it? How did you pick the innovators, the creators Of each one of the technology revolutions in the history of computing? 100% record. Well, I think I was just lucky. I was born in the right time, and I met with great entrepreneurs like yourself,

right? It's a passion. It's a dream. And it's an instinct that you smell, is the real pioneer who is a real, you know, innovator. And, I really think I was lucky, but it's the same vision that we can smell, right? It's like, wolf smell wolf. I think we smell each other. I have two puppies. I don't like that mental image. No. No. Masa, each, as you think back on the history. - Yes. - How is this transition, this platform shift, this revolution, how is this one different than the previous ones? How does it feel different

to you? Well, I say this is the most exciting, most dynamic, front end of the future. This is 100 times, a thousand times bigger. This is the biggest wave. I thought that's how it feels. - Yeah, yeah. I think mathematically or from an industrial perspective, the important thing to realize is that although AI is software, it is a very different type of software. Yeah. The software industry you and I created, and are part of is an industry of tools. It's tools used by humans. Yeah. For the very first time, this new type of software, neural

network, Large Language Models, agents, robots are not tools, but they are skills. They're tasks. They do work. - Yeah. They can perform work. And the industry, the market, the industry of work is not $1 trillion, it is $100 trillion. And that's the reason why we realized that this industry is, in fact, not a transition of the IT industry. But this is a transition of every industry. - Yes. Which is the reason why it’s such a big deal. Yeah. Mankind is the only animal that has the super brain compared to any other species. Because of brain

power, mankind is so powerful. If you compare just the muscles, lions and the elephants, they have a bigger muscle. But mankind has the smartest brain. And every activity of GDPs today is based on mankind's type of brain activity. So I think every industry will be affected by this revolution. And one of the one of the amazing things, of course, is that, in the industry that is governed by atoms, the size of the industry is limited. Because there's only so much atoms you can move around. It's heavy. - That's right. But the industry of AI is

industry of electrons. - Exactly. It's governed by quantum mechanics. - Right, right, right. It can be infinitely large. - Yeah. And intelligence. Intelligence is so much more valuable compared to just moving things. You know, you think, chain of thought, reasoning. It's amazing. It's amazing. - Yeah, Masa one of the things That we announced today together is building the AI grid of Japan. - Yeah, yeah. And the AI grid will have AI factories for developing the AI models. It will have AI-RAN distribute the AI models all over Japan. The architecture of the AI factory in the

AI-RAN that we worked on together is revolutionary. - Yeah. The world has nothing like this. Japan will be the world's first. - Yeah, every other telco will have to follow this new wave. - Yeah. And. And so several things that that I want to ask you. So one, how would SoftBank use this yourself for SoftBank and your subsidiaries? And how do you imagine this AI grid will revolutionize AI here in Japan? - Yeah, as you just mentioned our cell towers used to just carry bits for the telecommunication and internet surfing and so on. However, now

with this intelligence network that we densely connect each other, becomes one big neural brain, right? For the infrastructure of intelligence for Japan. That will be amazing. - Yeah. You know, and of course, you could use it for your subsidiaries like LINE, Yahoo Japan. - And PayPay. - That's right. - Yeah. And so it will create AIs to make all of your services more enjoyable and more useful to your consumers. But one of the things that I'm very excited about is making this resource available so that, so that, researchers and students, startup companies can flourish here

in Japan. - Yes, definitely. And with your support, we are creating the largest AI data center here in Japan, which, I'm talking with Miyakawa that we should provide this platform to many of those researchers, the students, the startups, so that we, we can encourage with, we are trying to subsidize so that they have a better access, much more compute. - Yeah. - Well, building infrastructure is very capital intensive. And you're making a very big bet on Japan. - Yes. - You know that you and I have spoken about this before. In a lot of ways,

Japan technology led doing the mechatronics era. - Yes. - During that industrial revolution when when mechanical technology electronics came together. In fact, even consumer electronics, during that age, Japan really led the world, right? And then when, when the IT industry came and software came, there was a, there was a miss that I think the last three decades, as the software industries flourished in the West and in China, really, Japan could have been more aggressive. - Totally agree. Totally. Yeah. And back in those days, even to some extent today, big enterprise, the medias, the big grown

up guys, they used to say, 'monozukuri'— meaning, ‘making physical stuff.’ Physical stuff has the real value and the real meaning. And software is something virtual that they don't trust, the value of software. That has been the mentality for many, many years in Japan that led young, startup guys, the young, you know, generations discouraged, especially after the net bubble crash. Everybody criticized. I was criticized a lot. And that really, you know, was sort of like a punishment to the young stars. Made, you know, the, depressed feeling. I think we have to revive this passion which the

robotics is bringing with AI. As you say, you know, bring AI intelligence into robotics. Astro Boy, in Japan, we call ‘Tetsuwan Atomu.’ My favorite cartoon. You cannot just have the muscle. It has to have, intelligence so that the robot can talk. Robot can have a passion, as a friend, you know. I think that kind of, front end challenge is, is very, very needed here. - Yeah. Yeah, I think the software era was, is now, The good news is that this is the beginning of a new era. - Yeah. Reset. Once more, reset. Reset button. -

That’s right. The industry is in reset. And you could see that the entire stack is being reset because the companies of the last generation are not doing so well in this new generation. - That's right. - And so there's an emergence of a whole new stack, a whole new opportunity. - Right. - Japan must take advantage. Must take advantage of this time. And artificial intelligence is very different than software. Artificial intelligence requires that you have data, that you have domain expertise. Yes. If you are an artist, you have domain expertise, if you develop video games,

you have domain expertise. If you look for drugs, you discover drugs, you have domain expertise. If you have domain expertise, you can describe your expertise in data. That data could be used to train an AI model. That AI model becomes your artificial intelligence. - That's right. At least, Japanese government is not trying to, you know, depress the, this AI revolution. Some of the other countries, they try to be overprotective of the, intelligence, artificial intelligence, so the regulation goes a little crazy. Here in Japan, at least we are lucky that this time Japanese government seems not,

you know, trying to, depress, at least no handicap. So I think, but they should encourage more. They should encourage more. As you said, this is the reset. This is the catch up moment for this new revolution. We can't miss this time. - Cannot miss this time. And of course, in order to be part of the AI revolution, you need, this time, you need infrastructure. - Right. - This type of software, because it's machine learning. - We’re gonna buy a lot of your chips. - Well, thank you, thank you. - Yes. - And and, you need

infrastructure. You need AI factories. Without the infrastructure, it's not possible to create AI. - Totally. - That's the reason why SoftBank is building the AI grid in Japan. And you will catalyze, you will activate, you’ll turbocharge all of the activity that's here already. - We're going to show by our example. - Yeah. - And hopefully, today we have 350 startups in Japan. 350 startups that we work with out of 22,000 in the world. Yeah. It makes no sense. - That's right. - And so we we must encourage, the young entrepreneurs, the innovators. - Yeah, yeah.

- To jump, to leap, to engage AI. The infrastructure is coming. - Yeah. So as I said, Miyakawa and I are discussing, we're going to create the largest AI data center here in Japan. So we are going to provide a lot of encouraging program. Subsidize the compute power, so that they can they can use almost free. Okay, almost free to try out the new models, try out the, you know, whatever, the applications of AI. You should also help with some donation, right? - Okay. - Hey, Jensen! For some sort of startup, and researchers— - This

is the last time I'm inviting myself to— Every time I see Masa, it costs me money. - Yeah. That's good for everybody. - Yes, yes. Yeah. Very happy about that. Masa, what are you most excited about for the future of AI in Japan? What do you hope? What is your dream? - Well, look, you know, as you say, I am passionate about, AI robotics. I think the medical solutions, I think medical agents is definitely coming. I think a lot of, new agents. You know, so we have LINE, we have Yahoo and other services, Paypay. We

can make many specific AI agents for, for helping Japanese lifestyle. That cannot just come from US, you know, we know about Japanese behaviors, the cutures, the local intelligence and APIs to many, many, sites in Japan. So that I think, agents, I think every—you mentioned about, enterprise AI agents, definitely I support that. I'm excited about that. But also, I think personal agents will really come to every one of us. Bill gates said, the PC on every desktop. Steve said, a smartphone in every hand. I think now we should say, AI agents to everybody. So each of

us should have our own personal agents. - That's right. - That will help make our plans for trips, vacations. Right? Education. - And it follows you your whole life. - Yeah. Yeah. - Could you imagine an AI agent that knows you your whole life? - Exactly. - Our grandkids, they grow up with, iPhones from, you know, age one, right? They talk, and every time they see some picture, they do this even with a still picture. Right? Because they are born the, you know, with two fingers, they can they think, you know, every picture can be,

blown up. - In the future, they'll see a picture, they'll talk to it. - Right. They talk. - They'll hope it talks back. You can ask it questions. - They will have agent, their personal agents since the age of one and their personal agent, like your, you know, second buddy to see you grow together with you knows everything. You are sick. You know, your health. - Your tutor. - Yeah. Yeah. - Since you were a child. - Exactly. - Yeah. Remember everything that you read? Remember everything that it taught you, right? That. Yeah. Your personal

Aristotle. - Totally, totally digital twin. You know? I think that is really coming and having, you know, Japanese domestic knowledge, culture and so on, homegrown agents will have a huge, amazing future. - You know, one of the things Masa, that countries are awakening to is that the data of their country, their, their citizens' data, encodes the country's knowledge, its culture, its intelligence. And that data belongs to the country like its national resources. - National security, national security. - So that country, every country, should process that data, process and turn it into its AI for its

own people. - Totally, totally. - Makes no sense to outsource that to somebody else. - Oh, this is this is very, very important. It's a sovereign data center. Their each sovereign, each country, each government have to migrate, their national security data into their own data center, AI data center. That becomes the, something that you have to have in your own country. Each country have to have national security, data security. I think that will become, regulated to each protection of each country. - Yeah. Every country, every country will produce its own intelligence. Of course, every company

will produce their own intelligence. Their own AI. How is it possible that a company, does not create its own AI? - Yeah. - A company doesn't— - It’s like giving your brain away to somebody else. - That's right, that's right. And so I think, I think, the world is awakening to this idea. And the most important part is the first part, is that a national grid, an AI grid, has to exist. - Right. - You cannot have an automotive industry if there are no roads. - Totally, totally. - You know, and so so now you

have built the AI roads for Japan. On top of it, you know, all kinds of new services, new companies can flourish. - Totally. - So really, really excited about that. Well, Masa-san. Could you imagine if today, you were the largest shareholder. - Oh my God. - [laughs] - Yeah. We tried this. - We tried, we tried to. - It was three times. - Was it three times? I thought it was two times. - No, no. The first time, we became your shareholder by ourselves buying from the market. We talked about the, you know, the... We

even talked about. - Some things we can share. - Right? You and I had a private talk. - I don't want, don't say it. - You know, ten years ago— - Oh, I know. I have regretted it. - —in my garden. - All right. There's nothing. Okay, let me tell you what Masa said. Are you guys ready? So Masa said, “Jensen, the market does not understand the value of NVIDIA. Your future is incredible. But the market doesn't understand it.” - That was ten years ago. - “And your journey of suffering will continue for some time.

Because you are inventing this future.” “ So, let me give you the money to buy NVIDIA.” He wanted to lend me money to buy NVIDIA. All of it. Now I regret not taking you up on it. That was a great idea. - Oh, that was a great idea right? - Now, now we're both sad. - That was one month after I acquired Arm. - Yes, yes. And then we talked about combining the two companies. That was another. Another. That was a good dream too. - That was the third one. That was the third attempt. -

Yeah, right. - The first we talked about this privatizing and then second was we I just bought from the market. And then the third time was merging. Three attempts! Oh my god. - Well now we are going to We're going to create incredible value together. So NVIDIA and SoftBank are going to partner together. - Yeah. The market is so huge! - It's incredible. - So huge, so huge. - And so I am so happy. I'm so happy that that we're doing something of this great significance. - Yeah. - Here in Japan, I, I am very

hopeful. - Well, this is just the beginning. We're going to do many, many things together. - Thank you. - And the industry is so big and, you know, Arm has lots of mobile and IoTs and autos. - And you have a great data centers and games and so forth. Many, many things that we can collaborate. - I'm looking forward to it. Yeah. Masa-san! Ladies and gentlemen, Son-san. - Yeah. - Unquestionably one of the great entrepreneurs the world ever seen. - No, no. You are the greatest. - Thank you so much, thank you. Thank you. Masa-san.

You could see you could see his passion for AI. The work, the partnership that we're developing is going to bring an AI grid throughout Japan from the factory to the distributed AI-RANs. Before I before I leave, I want to welcome all of you to AI Summit. We have so many sessions here, so many partners here. Our purpose. Our purpose is to partner with you, to bring AI, to activate AI here in Japan, and to use this reset, this opportunity, where technology has reset for us to transform the companies and to create the next great companies

here in Japan. Japan has always been dear to me. Most of you do not know that if not for Japan and I mean this very seriously, If not for Japan, NVIDIA would not be here today. And there's a reason for that. If not for Japan, we would have not seen the first AI supercomputer, CUDA supercomputer built. So Japan has been very dear to me for a very long time. I'm very grateful for our partnership. I'm very grateful for your friendship. And I made a small video to share with you. Please run it. “If there is

a heart in the fish, there is a heart in the water.” Our friendship with Japan began 30 years ago. <font color=#E7D500FF>“Ready? Go!”</font> We had a dream to reinvent 3D graphics. SEGA, with its groundbreaking 3D games, chose NVIDIA to build its next game console. But it became clear our unconventional approach would not live up to our vision. With humility, we asked SEGA’s Irimajiri-San to release us from our obligations. He generously agreed, giving us the time needed to invent the GPU, a chance for success — saving our company. In the years since, we've worked with Japan to

redefine gaming on the PlayStation 3, SEGA Arcades, and the Nintendo Switch. And created a medium for telling beautiful stories. With the invention of CUDA, we transitioned from a gaming company to a computing company. As we expanded into the world of Supercomputing and AI, it was Japan's then Tokyo Tech that first saw the promise of GPU accelerated computing. <font color=#E7D500FF>“It’s just the beginning of GPU acceleration becoming mainstream.”</font> Less than a decade after Tsubame, we introduced DGX-1. The AI computer that changed the world. <font color=#E7D500FF>“A GPU is not just a GPU anymore.”</font> NVIDIA and AI are now

at the center of a new industrial revolution. From the beginning, Japan has been a friend on the NVIDIA journey. And our story together is just getting started. Thank you, Japan. Have a great AI Summit. Thank you.