Nvidia CEO Jensen Huang went to Stanford Business School and shared his difficult entrepreneurial journey as a Stanford alumnus and his views on the artificial intelligence revolution, which caused quite a stir in the industry.
Less than a week later, Jen-Hsun Huang made another appearance at Stanford and was interviewed at the Stanford Institute for Economic Policy Research (SIEPR) Economic Summit. There was a huge amount of information in this interview. Huang Renxun expressed his views on topics such as the nature of accelerated computing, the future of model training, and the competition of inference chips. In addition, he also expressed his views on when general artificial intelligence (AGI) will be realized and how much AI growth will be needed. Additional chip production capacity is forecast.
Huang Renxun believes that AI will pass human testing within five years and AGI will arrive soon. However, the answer depends largely on how we define this goal. He said that if our definition of "a computer that thinks like a human" is to test the capabilities of humans, then AGI will soon arrive.
He also said that in the next 10 years, NVIDIA plans to increase the computing power of deep learning by another 1 million times. By then, NVIDIA will achieve continuous learning, no longer the current model of learning first and then applying.
Huang Renxun’s core views:
If your definition of AGI is to pass human testing, then I will tell you that it can be achieved in 5 years. But if you change the question a little bit and say that AGI is about human intelligence, then I'm not quite sure how to clearly define all of your intelligence. In fact, no one is really sure.
In the next 10 years, we plan to increase the computing power of deep learning by another 1 million times. By then, we will achieve continuous learning, no longer the current model of learning first and applying later.
In the past 10 years, we have reduced the cost of computing a million times. A lot of people say, but if you can reduce the cost of computing a million times, people will spend less money, but the opposite is true... the demand has grown significantly... so we will do more calculations, We will reduce the marginal cost of computation to close to zero.
One piece (H100) can replace a data center composed of traditional CPUs. Although it sells for up to $25,000 a piece, the cost of the cable alone for the corresponding old system exceeds the price of the chip. We have redefined computing and condensed the entire data center into one chip.
Reasoning about chips is very difficult. You would think that the response time for inference would have to be very fast, but that's easy because it's the easy part of computer science. The rare part is that the goal of those deploying inference is to attract more users and apply the software to a large installed base.
Reasoning is a question of installed base. It's the same people who write apps for the iPhone, and they do it because there are so many iPhones installed, almost everyone has one. Therefore, if you write applications for iPhone, you will gain huge benefits and benefit all users.
People who buy and sell chips only think about the price of the chip, while people who run data centers think about the entire operating cost, deployment time, performance, utilization and flexibility in all these different applications. Overall, our total cost of operations (TCO) is very low, and even if a competitor's chip is free, it's not cheap enough in the end!
These models that we're training are multimodal, which means we're going to learn from sounds, we'll learn from text, we'll learn from sights, just like we all do, watching TV and learning from it. This is important because we want AI to not only be rooted in human values, but of course this is where ChatGPT’s real innovation lies, which is RLHF (reinforcement learning based on human feedback).
We increase computing power by a million times every 10 years while demand increases by a trillion times, and the two must cancel each other out. And then there's technology proliferation and so on, it's just a matter of time, but that doesn't change the fact that one day all computers in the world will be 100% changed, every data center, trillions of dollars worth of infrastructure The facilities will be completely changed and then new infrastructure will be built on top of that.
AI is quite possibly the most important thing of the 21st century. As for the transistor, it is the greatest invention of the 20th century. However, both are close to the greatest inventions, and we can let history be the judge.
Future AI computers will generate synthetic data, perform reinforcement learning, and will continue to be based on real-world experience data. It imagines situations, tests and corrects them using real-world experience. The whole process is a huge cycle.
I think one of my biggest strengths is that I have low expectations...people with high expectations have low stamina. Unfortunately, endurance is important on the road to success. I don't know how to teach it to you unless I want you to suffer a little.
Coding is a reasoning process, and that's a good thing. But does it guarantee you a job? Not at all. Programmers will certainly continue to exist in large numbers in the world, and NVIDIA needs programmers. However, the way you interact with computers in the future won't be C++.
Jensen Huang’s second interview at Stanford
opening remarks
Jen-Hsun Huang:
I always feel it's important to have a good opening line, and my opening line is to ask: "Do you want to see my homework?"
You know, we have two beautiful children, I have a perfect life, two adorable puppies, and I love my job. She still likes to look at my "homework".
host:
Well, I can ask you a few questions if you want.
Jen-Hsun Huang:
OK, please ask.
The nature of accelerated computing
host:
In my lifetime, I think the biggest technological development and technological breakthrough is the invention of the transistor. I'm older than you and this is indeed a major invention, but should I rethink it? Is AI now the biggest change in technological development in the past 76 years? This hints at my age.
First, the transistor is obviously a great invention, but the greatest capability it enables is software, the human ability to computationally express our ideas and algorithms in a repeatable way. This is a real breakthrough.
For the past 31 years, our company has been working on a new form of computing called accelerated computing. The core reason is that general-purpose computing is not suitable for every field of work, so we said, why not develop a new computing method that solves the problems that general-purpose computing is not good at.
In some areas of algorithmic, parallelizable computing, we have actually reduced the cost of computing to zero. What happens when you can reduce the marginal cost of something to close to zero? We have created a new way of developing software, where software is no longer written by humans, but by computers because the computational cost is close to zero.
We can let computers process large amounts of empirical data, that is, human digital experience, and discover patterns and relationships, which represents human knowledge. This miracle happened about 15 years ago. We anticipated this and reoriented our entire company to enter this new space.
In the past 10 years, we have reduced the cost of computing a million times. A lot of people say, but if you can reduce the cost of computing a million times, people will naturally spend less money, but the opposite is true. We found that if we can reduce the marginal cost of computing to close to zero, we may use It does some pretty crazy things, and demand has grown significantly.
For example, large language models can extract all human digital knowledge from the Internet and put it into a computer, allowing the computer to understand the knowledge on its own. The idea of grabbing all the data from the Internet and putting it into a computer and letting the computer figure it out is a crazy concept in itself.
Imagine, for those new to AI, that we can now use computers to understand the meaning of almost any digital knowledge. Taking gene sequences as an example, we can now use large language models to learn the meaning of those genes. We can also digitize amino acids through mass spectrometry and other techniques. We can now understand the structure and function of proteins from their amino acid sequences without having to work extensively using techniques such as cryo-EMs. We can also do this on a fairly large scale. Soon we will be able to understand what a cell is, a long chain of interconnected genes. Computers can also understand a full page of text. You can ask it what the meaning of the text is. Please summarize what the text wants to express.
This is what the miracle of accelerated computing brings. So, I would say that AI, enabled by this new form of computing that we call accelerated computing, which took thirty years to come into being, is probably the greatest invention in the technology industry. AI is quite possibly the most important thing of the 21st century. As for the transistor, it is the greatest invention of the 20th century. However, both are close to the greatest inventions, and we can let history be the judge.
It is planned to increase the computing power of deep learning by 1 million times in the next 10 years.
host:
Can you look ahead five years? I know that what's driving AI right now is your H100 GPU, and you're also launching a new H200, and I understand that you plan to upgrade the chip every year. So when you launch the H700 in March 2029, what will it allow us to do that we can't do now?
Jen-Hsun Huang:
First of all, the chip you just mentioned weighs 70 pounds and is composed of more than 35,000 parts, 8 of which are from TSMC. This chip can replace a data center composed of traditional CPUs. As we all know, the computing speed of our chips is very fast. The resources saved by this chip are unimaginable. But at the same time, this is also the most expensive chip in the world, priced at as high as 25,000 US dollars a piece. But for the corresponding old system, the cost of the cable alone exceeded the price of the chip. We have redefined computing and condensed the entire data center into one chip.
The chip is very good at a form of computing called deep learning, which is at the core of AI. This chip not only works in the chipset, but also at the algorithm level and data center level. They are a whole and cannot operate alone. They need to be connected together, and the network becomes part of it. So when you look at one of our computers, it's really an amazing thing, although only a computer engineer would find it amazing. It is very heavy and requires hundreds of miles of cable. The next version will feature liquid cooling, which is great in many ways.
Its computing scale is equivalent to a data center. In the next 10 years, we plan to increase the computing power of deep learning by another 1 million times. By then, we will achieve continuous learning, no longer the current model of learning first and applying later. We will decide whether the results of continuous learning are deployed into real-world applications. The computer will watch videos and new text and continue to improve itself through all these interactions. The learning process, training process, reasoning process, and deployment process will all become one. That's exactly what we do.
In the future, computers will not focus on learning in a certain period of time and only reasoning in another period of time, but will continue to learn and reason. This reinforcement learning will be continuous and based on real-world data interactions as well as synthetic data generated in real time.
The capabilities of computers will continue to be imagined. Just like when you study, you start from first principles and think this is how things should be. We then simulate and imagine in our brains, and the imagined state of the future becomes reality for us in many ways. Therefore, future AI computers will generate synthetic data, perform reinforcement learning, and will continue to be based on real-world experience data. It imagines situations, tests and corrects them using real-world experience. The whole process is a huge cycle. That's what happens when your computing power is a million times cheaper than it is now.
So when I say this, please pay attention to what is at its core. When you can reduce the marginal cost of computing to zero, there are many new possibilities worth trying. This is the same as going further because the marginal cost of transportation has dropped to zero. I can fly from here to New York relatively cheaply. If it had taken a month, I probably would have never gone. The same goes for shipping, it goes for anything. So we're going to do more computation, and we're going to reduce the marginal cost of computation to close to zero.
How to evaluate chip competition?
host:
Which reminds me, you may be aware of some recent reports that Nvidia will face more competition in the inference market than in the training market. But according to you, these two are actually the same market? Can you comment? Will there be independent training chip market and inference chip market? Or will you keep training and switch to inference within the same chip? Can you explain it?
Jen-Hsun Huang:
Whenever the AI system is prompted, whether it is ChatGPT, Copilot, or a service platform you are using, if you use Midjourney, Adobe's Firefly, etc., once a prompt is made, it will perform inference and generate corresponding information. Whenever you do this, it's the NVIDIA GPU that's working behind the scenes. So most of the time when you interact with our platform, you're making inferences. As a result, 100% of the world's inference is now provided by NVIDIA.
So is reasoning difficult or easy? The reason why many companies challenge NVIDIA in the field of inference is because when they see NVIDIA's system for training, you will think that it looks too difficult, I will not do it, I am just a chip company, but this system can see It doesn't look like a chip at all. Just to prove that something new works, you have to invest $2 billion, and then you launch it and find out that it may not work. You invest $2 billion and two years just to prove that it doesn't work. But at this point, you have already invested a lot of money and time. The risk of exploring something new is too high for customers. Therefore, many competitors tend to say that we do not do training, only inference.
Let me tell you now, reasoning about chips is very difficult. You would think that the response time for inference would have to be very fast, but that's easy because it's the easy part of computer science. The rare part is that the goal of those deploying inference is to attract more users and apply the software to a large installed base.
Reasoning is therefore a question about the installed base. It's the same people who write apps for the iPhone, and they do it because there are so many iPhones installed, almost everyone has one. Therefore, if you write applications for iPhone, you will gain huge benefits and benefit all users.
The same principle applies to NVIDIA. Our accelerated computing platform CUDA is the only one available around the world. Because we've been in this space for so long, if you write an application for inference and deploy a model on our architecture, it can run anywhere.
So you can reach every user and have a greater impact. Therefore, the core of the reasoning problem is actually the installed capacity, which requires long-term patience, years of successful experience, and a focus on architectural compatibility.
host:
Nvidia makes chips like no other. But is it possible that some cheaper competitor will emerge that claims to be good enough, but not as good as Nvidia? Is this a threat?
Jen-Hsun Huang:
First, competition does exist. We face more competition than any other company. We have competition not only from our competitors, but also from our customers. And I'm the only competitor in their eyes, and I'm showing them not just the current chips, but the next generation and future generations. This is because if you don’t put in the effort to explain why you’re good at something, they’ll never get the chance to buy your product. Therefore, our strategy is to be completely open and transparent, working with almost everyone in the industry.
We do this for several reasons, and our main advantages are as follows:
Customers can build chips (ASICs) that are optimized for a specific algorithm, but remember, computing is not just about transformers, not to mention we are constantly inventing new transformer variants, and beyond that, the variety of software is very rich , because software engineers love to create new things.
What NVIDIA is good at is accelerating computing, but our architecture not only accelerates algorithms, but is programmable, which means you can use it to process SQL, we can accelerate quantum physics, accelerate all fluids and particle codes, etc. in a wide range of fields , one of which is generative AI.
Therefore, what NVIDIA is good at is the broad area of accelerated computing, one of which is generative AI. For a data center looking to have customers in every industry, we have become the de facto standard, present in every cloud platform and computer company.
Therefore, after more than 30 years of development, our company's structure has become an industry standard, which is our main advantage.
I would even be surprised if the customer had a more cost effective alternative. The reason is, when you look at a computer today, it's not like a laptop, it's a data center and you need to operate it. So, people who are buying and selling chips are just thinking about the price of the chip, whereas people who are running data centers are thinking about the entire cost of operations, deployment time, performance, utilization and flexibility in all these different applications.
Overall, our total cost of operations (TCO) is very low, and even if a competitor's chip is free, it's not cheap enough in the end! The goal is to add so much value that alternatives aren't just about cost.
Of course, it takes a lot of effort, we have to keep innovating, and we can’t take anything lightly. I was hoping not to sound too competitive but the moderator asked a competition question and I thought this was an academic forum....it triggered my competitive gene and I apologize, I could have handled this more artistically .
I'll be more measured next time, but he surprised me with a reference to a competitor. I thought this was an economic forum, but you guys are so straightforward.
Five years until AGI?
host:
Let me be more straightforward. When do you think we can achieve artificial general intelligence (AGI), that is, reach the level of human intelligence? Is it 50 years from now? Or in 5 years? What do you think?
Jen-Hsun Huang:
I'll give you a very specific answer. But first let me tell you some exciting things going on. First of all, these models that we're training are multimodal, which means we're going to learn from sounds, we'll learn from words, we'll learn from visions, just like we all do, watching TV and learning from it. This is important because we want AI to not only be rooted in human values, but of course this is where ChatGPT’s real innovation lies, which is RLHF (reinforcement learning based on human feedback).
But until reinforcement learning, humans anchored AI in what we thought were good human values. Now, you can imagine, you have to generate images and videos, and the AI knows that hands don't penetrate the podium and you fall in when you step on water, so now the AI starts to be anchored in physics.
Now, AI watches a lot of different examples, such as videos, to learn the rules that are followed in the world. It must create what is called a world model. So, we have to understand multimodality, and there are other modalities, like genes, amino acids, proteins, cells, etc.
The second point is that AI will have more powerful reasoning capabilities. Many of the reasoning we humans do are encoded in common sense. Common sense is an ability that all of us humans take for granted. There is a lot of reasoning and knowledge on the Internet that we have encoded and models can learn from.
But there is also a higher level of reasoning ability. For example, if you ask me a question now, I do generate most of the questions as quickly as a generative model. I don’t need to spend too much time reasoning, but there are some questions that I need to think about. It's planning, I'll say, "That's interesting, let me think about it," and then I might cycle through it in my head, construct multiple plans, go through my knowledge system, and selectively process, "This doesn't make sense. , but I can do this” meaning I’ll simulate it in my head and run it, maybe I’ll do some calculations and so on.
I mean, today's ability to think for a long time is something that AI cannot yet do. No matter what prompt you give ChatGPT, it responds immediately. We hope that by inputting a question into ChatGPT, giving it a goal and a mission, it can think about it for a period of time. Therefore, this system is called System 2 in computer science, or long thinking, or planning system, which is used to solve problems such as planning and complex reasoning.
I think we're looking at these issues and you're going to see some breakthroughs. As a result, the way you interact with AI will be very different in the future. Sometimes it's just a matter of asking a question and getting an answer, and sometimes you say "here's a question and you think about it for a while" and then it goes through a lot of calculations and gives the answer. You could also say "I'll give you this problem with a budget of $1,000, but no more than that." It will give the best answer within budget. There are other scene applications and so on.
So, back to the question of AGI, what is the definition of AGI? In fact, this is now the first question that needs to be answered. If you ask me, if you say Renxun, AGI is a series of tests, and remember, only engineers know that we've done it, you know, anyone in that reputable organization must know that, for engineers Say, there needs to be a specification, you need to know what the definition of success is, and you need to have a test. Now, if I give the AI a bunch of math tests, reasoning tests, history tests, biology tests, medical tests, bar exams, and any other tests you can think of, SAT and GMAT, etc., list all the tests you can think of. Come out, put it in front of the computer science industry, and I guess within 5 years, computers will be performing well on all of these tests.
Therefore, if your definition of AGI is to pass human testing, then I will tell you that it can be achieved in 5 years. But if you change the question a little bit and say that AGI is about human intelligence, then I'm not quite sure how to clearly define all of your intelligence. In fact, no one is really sure. Therefore, it is difficult to achieve as an engineer. Does this make sense? So the answer is we're not sure. But we're all trying to make it better.
The role of AI in drug discovery
host:
I want to ask two more questions, and then I'll turn my time over to you guys because I think there's a lot of good questions out there. The first question I want to ask is, can you dig a little deeper into your thoughts on the role of AI in drug discovery?
The first role is to understand the meaning of the digital information we have. As you know, we now have a large number of amino acid sequences. Thanks to AlphaFold, we can now understand the structure of many of these proteins. But the question is, what is the significance of that protein? What does it mean? What is the function of this protein? That would be great.
Just like you can talk to ChatGPT, you all know, there are people who can chat with PDFs. You can take a PDF file, no matter what the content is, my favorite is to take a PDF file of a research paper, load it into ChatGPT, and just start talking to it.
Like talking to researchers, like what was the inspiration for this study, what problem did it solve? What's the breaking point? What was the previous status? Any novel ideas? Just like talking to people.
In the future, when we put a protein into ChatGPT, like a PDF, it will ask what is this for? Which enzymes are activated? For example, there is a string of gene sequences that represent a cell. You can put this cell in, and it will ask you what it is used for. What is your role? What do you use it for? What are your hopes and dreams?
So that's one of the most profound things we can do, understand what biology means. If we can understand the meaning of biology, as you know, once we understand the meaning of almost any information in the world, in computer science and computing, brilliant engineers and scientists will know exactly how to exploit it.
But this is where the breakthrough comes, understanding biology multimodally. So if I could give you a deep answer, I think that's probably the single most profound thing we can do.
Advice for students: Lower your expectations
host:
OSU and Stanford must be proud of you. If I can switch gears a little bit, there are a lot of students at Stanford who have entrepreneurial aspirations, who are entrepreneurs, maybe computer or engineering majors. What advice can you give them to improve their chances of success?
Jen-Hsun Huang:
You know, I think one of my biggest strengths is that I have low expectations, and I mean that. Expectations are high for most Stanford graduates. You should have high expectations because you come from a great school. You are outstanding in school and are among the best students. Obviously, you can afford your tuition, and you graduated from one of the best institutions on earth. You are surrounded by other incredible kids. You will naturally have high expectations.
People with high expectations have low endurance. Unfortunately, endurance is important on the road to success. I don't know how to teach it to you unless I want you to suffer a little. I was lucky enough to grow up in an environment where I could succeed but also had a lot of setbacks and hardships. To this day, I still take great joy when the phrase “pain and suffering” is used within the company.
The reason is, you want to build character in the company, you want it to become great. Greatness is not intelligence, as you know. Greatness comes from character, and character is not shaped by smart people, but by people who have gone through hardships.
So that's what I would say, if I could give you any advice, although I don't know how, but for all the Stanford students out there, I hope you experience enough pain and suffering.
How to motivate employees?
host:
I'm going to break my promise and ask you one last question. You seem motivated and energized, but how do you keep your employees motivated and energized, especially when they might be richer than they expected?
Jen-Hsun Huang:
Yeah, I have 55 people around me, my management team. So you know, the management team that reports directly to me is 55 people. I have not written individual reviews for any of them. But I will continue to comment on them, and they will comment on me. The salary I gave them was just the number in the lower right corner of the Excel sheet, which I would just drag and copy. Many of our executives are paid exactly the same amount. I know it's weird, but it works. I never meet with people alone unless they need me.
I never have a private meeting with them, and they never hear from me what only they know. I will not disclose information to any employee that is not known to others in the company. So our company is designed to be agile, to allow information to flow as quickly as possible and to empower people based on what they can do, not what they know. This is the structure of our company. I don't remember your question, but I get it.
The answer lies in my actions. How do I celebrate success? How do I celebrate failure? How do I talk about success? How do I talk about frustration? I'm looking for opportunities every day to instill company culture and what's important, what's not important, what's the definition of good, how to compare yourself to good, how to think about good, how to think about the journey, how to think about the outcome, etc., all day long.
How much additional chip capacity will be needed for AI growth?
Audience 1:
I have two questions, the first is about your leather jacket story. The second is, based on your predictions and calculations, how much will chip production capacity need to increase in the next 5 to 10 years to support the growth of AI?
Jen-Hsun Huang:
OK, I appreciate both questions. The first question is, my wife bought me this leather jacket and that's what I'm wearing now. Because I don’t shop at all. As long as she finds something that doesn't make me feel itchy, I'll wear it all the time because she's known me since she was 17 and she thinks all clothes are itchy to me. When I say I don't like a piece of clothing, I say it makes me itchy. So as soon as she finds something that doesn't make me itch, you look in my closet and it's all the same shirt because she doesn't want to shop for me anymore. So that's all she bought me and that's all I wore. If I don't like it I just go shopping myself, otherwise I just wear it and it's good enough for me.
The second question is about prediction. When it comes to forecasting, I'm actually terrible, but I'm very good at deducing the size of opportunities based on first principles. I don't know how many fabs there are, but what I do know is that the calculations we do today, the information was written by someone else, or created by someone, it was basically pre-recorded.
What I'm talking about is that everything, every word, voice, video, is retrieval. Someone has written it and stored it somewhere, and then you retrieve it. Every modality you know used to be That's right. Like I said, every time you click on your phone, remember that someone wrote it and it's stored somewhere, it's all pre-recorded.
In the future, because our AI can access all the latest news in the world, which means it can be retrieved, it understands your context, which means it understands what you ask. The key is that most of the calculations will be generative. 100% of today's content is pre-recorded.
If in the future, 100% of content is generative, the question is how this will change the shape of computing. So instead of getting hung up on the details, this is how I frame the question, like do we need more networks? Do we need more memory? Simply put, we need more fabs.
However, we are also constantly improving the algorithm and processing process, and the efficiency has been greatly improved in time. It's not that computing is as efficient as it is today and therefore the demand is as high as it is.
At the same time, we are increasing computing power by a million times every 10 years while demand is increasing by a trillion times, which must offset each other. And then there's technology proliferation and so on, it's just a matter of time, but that doesn't change the fact that one day all computers in the world will be 100% changed, every data center, trillions of dollars worth of infrastructure The facilities will be completely changed and then new infrastructure will be built on top of that.
Learning to code doesn’t guarantee a job
Audience 2:
Yes, welcome to Stanford. Recently you said you were encouraging students not to learn to code, and if that's the case, that could mean a few things. But do you think the future, from a company formation and business ownership perspective, will move towards the creation of more companies or towards the consolidation of a few large players?
First, I said it so badly that you repeated it badly. I didn't mean that. If you want to code, oh my god, please code. Coding is a reasoning process, and that's a good thing. But does it guarantee you a job? Not at all. Programmers will certainly continue to exist in large numbers in the world, and NVIDIA needs programmers. However, the way you interact with computers in the future won't be C++.
For some people it is, but for you, why program in such a strange language as Python? In the future, you'll just tell the computer what you want. The computer will then do what you want it to do. You could say, "Hey, I want you to create a build plan that lists all the suppliers and materials needed, and do it based on the sales forecast we give you."
It will give you a complete plan based on all necessary components. If you don't like it, just let it write a modifiable program for you in Python. So remember, the first time I interacted with the computer, I was just talking in normal English. Only the second time I used a programming language. By the way, the best programming language of the future is human language. How do you talk to a computer? How to prompt it? How to carry out prompt project? How to communicate with people? How to communicate with computers? How to make a computer do what you want it to do? How do you fine-tune the instructions to your computer? This is called the prompt project. This requires an artistic talent.
For example, this would be surprising to most people, but not to me. For example, you ask Midjourney to generate an image of a puppy on a surfboard at sunset in Hawaii. Then it generates a picture, and you say "be cuter" and it becomes even cuter. You say "No, it's cuter than that". Here comes another cuter one. Why does the software do this? This illustrates the software's ability to fine-tune to your requirements, reflecting how computers in the future will understand and respond to human instructions.
As a result, the nature of programming is changing, and future interactions with computers will be more natural and intuitive. Through AI, we are closing the technology gap. In the past, only a few people who knew how to program could perform related tasks. Today, almost everyone can instruct a computer to perform tasks by simply communicating commands. There are many people on YouTube, including children, who communicate with ChatGPT and let it do amazing things, and these people do not need to have traditional programming skills. This shows that in the future, interacting with computers will be as natural as interacting with people, which is a great contribution of the computer science industry to the world.
Regarding the issue of geopolitical risks, we can almost say that it is a typical representative of geopolitical risks. Because we produce tools that are crucial for AI, which is considered the defining technology of our generation. The United States therefore has every right to decide to restrict these tools to countries it deems appropriate.
On the one hand, this limits our opportunities in some areas, but on the other hand, it also creates opportunities for us in other areas. In the past 6 to 9 months, every country and region has realized that it needs to master its own AI technology and cannot let its data be sent abroad for processing and then flow back to the country. This awakening to sovereign AI has created a new world for us. A huge opportunity.
Finally, regarding the issue of customizing solutions for customers, why is the threshold relatively high now? Because each generation of our platform products first has a GPU, a CPU, a network processor, software, and two types of switches. I built five chips for one product generation, and people thought there was just one chip, the GPU, but it was actually five different chips, each costing hundreds of millions of dollars in R&D just to get to what we call "launch." standards, and then you have to integrate them into a system, and then you also need network equipment, transceivers, fiber optic equipment, and a lot of software. Running a computer as big as this room requires a lot of software, so it's all complicated.
If the customization needs differ too much, then you have to repeat the entire development process. However, if customization can take everything that's already there and add something on top of it, then it makes a lot of sense. Maybe a proprietary security system, maybe a cryptographic computing system, maybe a new way of processing numbers, and more, we're very open to those.
Our clients know that I'm willing to do all of these things and realize that if you change too much, you basically reset it all and waste nearly a hundred billion dollars. So they want to leverage these as much as possible in our ecosystem (reduce replacement costs).