老黄:未来 10 年算力提高 100 万倍,对手芯片免费也比不过英伟达!

Wallstreetcn
2024.03.11 09:27
portai
I'm PortAI, I can summarize articles.

黄仁勋斯坦福分享,本次访问信息量巨大,黄仁勋就加速计算的本质、模型训练的未来、推理芯片的竞争等话题发表了自己的看法,另外,他还对通用 AGI 何时实现,AI 增长需要多少额外芯片产能进行了预测。


Author: Shuqing Bu

Source: Hard AI

On March 4th local time, NVIDIA CEO Jensen Huang visited Stanford Graduate School of Business as an alumnus to share his challenging entrepreneurial journey and his views on the artificial intelligence revolution, causing quite a stir in the industry.

In less than a week, Jensen Huang once again appeared at Stanford, this time at the Stanford Institute for Economic Policy Research (SIEPR) Economic Summit, where he was interviewed. This visit was incredibly informative as Huang shared his views on topics such as the essence of accelerated computing, the future of model training, and the competition of inference chips. Additionally, he made predictions on when General Artificial Intelligence (AGI) will be achieved and how much additional chip capacity is needed for AI growth.

Huang believes that AI will pass the Turing Test within five years, and AGI will arrive soon. However, the answer largely depends on how we define this goal. He stated that if we define "computers that think like humans" as the ability to pass the Turing Test, then AGI will arrive soon.

He also mentioned 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, moving away from the current pattern of learning first and then applying.

Key Points by Jensen Huang:

  • If your definition of AGI is passing the Turing Test, then I can tell you, it can be achieved in 5 years. But if you slightly change the way you ask the question, that is, if AGI needs to have human intelligence, then I am not sure how to clearly define all 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, moving away from the current pattern of learning first and then applying.
  • In the past 10 years, we have reduced the cost of computing by 1 million times. Many people say that if you can reduce the cost of computing by 1 million times, naturally people will spend less money, but in fact, the demand has significantly increased... Therefore, we will do more computing, and we will reduce the marginal cost of computing to near zero.
  • One (H100) can replace a data center composed of traditional CPUs. Although it is priced at $25,000 per piece, the cost of just the cables in the corresponding old system exceeds the chip price. We have redefined the way of computing, condensing the entire data center into this chip.
  • Inference chips are very difficult. You might think that the response time for inference must be very fast, but that's relatively simple because it's considered an easy part of computer science. The challenging part is the goal of deploying inference, which is to attract more users and apply the software to a large installed user base.


  • Reasoning is a matter of installing the basics. This is similar to those who write applications for the iPhone, they do so because of the huge number of iPhone installations, almost everyone has one. Therefore, writing applications for the iPhone can bring huge profits that benefit all users.
  • People who buy and sell chips only consider the price of the chips, while those operating data centers consider the overall operating costs, deployment time, performance, utilization, and flexibility in all these different applications. In general, our total cost of ownership (TCO) is very low, even if competitors' chips are free, in the end, it is not cheap enough!
  • The models we are training are multimodal, which means we learn from sound, text, and vision, just like all of us watch TV and learn from it. This is important because we hope that AI is not only rooted in human values, which is the true innovation of ChatGPT, namely RLHF (Reinforcement Learning based on Human Feedback).
  • Every ten years, we will increase computing power by a million times, while demand will increase by a trillion times, these two must offset each other. Then there is technological diffusion and so on, it's just a matter of time, but it doesn't change the fact that one day, all the computers in the world will be 100% transformed, every data center, infrastructure worth trillions of dollars, will be completely changed, and new infrastructure will be built on top of it.
  • AI is likely to be the most important thing in the 21st century. As for the transistor, it is the greatest invention of the 20th century. However, both are approaching the greatest inventions, and history will judge.
  • The AI computers of the future will synthesize data generation, engage in reinforcement learning, and will continue to be based on real-world experiential data. It will imagine scenarios, test and correct them with real-world experience. The whole process is a huge loop.
  • One of my biggest advantages is that my expectations are very low... People with high expectations have low endurance. Unfortunately, endurance is crucial on the path to success. I don't know how to teach you unless I want you to suffer some hardships.
  • Coding is a reasoning process, which is a good thing. But can it guarantee you a job? Not at all. Programmers will certainly continue to exist in large numbers in the world, and NVIDIA also needs programmers. However, in the future, the way you interact with computers will not be C++.

Huang Renxun's Stanford Second Round Interview Full Translation

Opening

Huang Renxun:

I always feel that a good opening is important, my opening line is to ask: "Do you want to see my homework?"

You know, my wife and I have two beautiful children, I live a perfect life, two lovely dogs, and I love my work. She still likes to see my "homework" to this day.



Host:

Well, if you're willing, I can ask you a few questions.

Huang Renxun:

Sure, please go ahead.

The Essence of Accelerated Computing

Host:

In my lifetime, I believe the greatest technological development and breakthrough was the invention of the transistor. I am older than you, and indeed, it was a significant invention. But should I rethink this? Is AI now the biggest revolution in technology over the past 76 years? This implies my age.

Firstly, the transistor is obviously a great invention, but its greatest achievement is software, the ability for humans to express our ideas and algorithms in a repeatable way in computing. That is a true breakthrough.

Over the past 31 years, our company has been dedicated to a new computing method called accelerated computing. The core reason is that general-purpose computing is not suitable for every field of work. So we thought, why not develop a new way of computing to address the shortcomings of general-purpose computing.

In certain algorithmic and parallel computing fields, we have actually reduced the cost of computing to zero. What happens when you can reduce the marginal cost of something close to zero? We have created a new way of software development where software is no longer written by humans but by computers, because the cost of computation is close to zero.

We can have computers process a large amount of empirical data, which is the digitized experience of humans, to discover patterns and relationships, representing human knowledge. This miracle happened about 15 years ago. We foresaw this and adjusted the entire direction of the company to enter this brand-new field.

Over the past 10 years, we have reduced the cost of computing by a million times. Many people say, if you can reduce the cost of computing by a million times, naturally people will spend less money. But in fact, the opposite is true. We found that if we can reduce the marginal cost of computation close to zero, we might use it to do some very crazy things, and the demand would significantly increase.

For example, large language models can extract all human digital knowledge from the internet and input it into a computer for the computer to understand knowledge on its own. The idea of fetching all data from the internet and putting it into a computer for the computer to find answers is itself a crazy concept.

Imagine, for those who are just getting into AI, we can now use computers to understand the meaning of almost all digitized knowledge. Take gene sequences, for instance, we can now use large language models to learn the meanings of those genes. We can also digitize amino acids through techniques like mass spectrometry. Now we can understand the structure and function of proteins from amino acid sequences without the need for extensive work using technologies like cryo-electron microscopy (cryo-EMs). We can do this on a fairly large scale. Soon we will be able to understand the meaning of a cell, a long string of interconnected genes. Computers can also understand a whole page of text. You can ask it what the meaning of the text is, please summarize, what does the text want to convey.


This is the change brought about by the miracle of accelerated computing. So, what I want to say is that AI, supported by this new form of computing we call accelerated computing, which took thirty years to take shape, may be the greatest invention in the technology industry. AI is likely to be the most important thing in the 21st century. As for the transistor, it was the greatest invention of the 20th century. However, both are approaching the greatest inventions, and history will be the judge.

Plans to increase the computing power of deep learning by 1 million times in the next 10 years

Host:Can you look ahead to the next five years? I know that your H100 GPU is currently driving the development of AI, and you are also introducing the new H200. As far as I know, you plan to upgrade the chip every year. So, by March 2029 when you launch the H700, what will it enable us to do that we cannot do now?

Huang Renxun:First of all, as you just mentioned, the chip weighs 70 pounds and consists of over 35,000 components, with 8 components coming from TSMC. This chip can replace a data center composed of traditional CPUs. It is well known that our chip has a very fast computing speed, and the resources saved by this chip are unimaginable. However, at the same time, it is also the most expensive chip in the world, with a price tag of up to $25,000. But the corresponding old systems, just the cost of cables exceeds the chip price. We have redefined the way of computing, condensing the entire data center into this chip.

This chip is very good at a form of computation called deep learning, which is the core of AI. This chip not only works in the chip set but also at the algorithm and data center levels. They are an integrated whole and cannot run independently. They need to be connected together, and the network becomes a part of it. So when you see one of our computers, it is truly remarkable, although only computer engineers would find it remarkable. It is very heavy and requires hundreds of miles of cables. The next version will use liquid cooling technology, which is excellent in many ways.

It computes on a scale 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 mode of learning first and then applying. We will decide whether the results of continuous learning should be deployed in practical applications. Computers will watch videos and new texts, and through all these interactions, continuously improve themselves. The learning process, training process, reasoning process, and deployment process will merge into one. This is what we are doing.

In the future, computers will not focus on learning at a certain time and only do reasoning at another time, but will continuously learn and reason. This reinforcement learning will be continuous and based on real-world data interactions and real-time generated synthetic data.The capabilities of computers will continue to evolve. Just like when you learn, you start from first principles and think about how things should be. Then we simulate and imagine in our brains, and the future imagined state is, in many ways, reality for us. Therefore, future AI **computers will synthesize data, engage in reinforcement learning, and continue to be based on real-world experiential data. They will imagine scenarios, test and correct them with real-world experience. The whole process is a huge cycle. This will happen when your computing power is 1 million times cheaper than it is now.

So when I say this, pay attention to its core. When you can reduce the marginal cost of computing to zero, there will be many new feasible ways worth trying. This is similar to going further, as the marginal cost of transportation has been reduced to zero. I can fly relatively cheaply from here to New York. If it took a month, I might never go. The same goes for transportation, and everything else. Therefore, we will do more computing, and we will reduce the marginal cost of computing to near zero.

How to evaluate chip competition?

Host:

This reminds me, you may know that there have been reports recently that NVIDIA will face more competition in the inference market than in the training market. But as you mentioned, are these actually the same market? Can you comment on this? Will there be separate markets for training chips and inference chips? Or will you continue training and switch to inference on the same chip? Can you explain?

Huang Renxun:

Whenever you prompt an AI system, whether it's ChatGPT, Copilot, or a service platform you're using, if you use Midjourney, Adobe's Firefly, etc., once you make a prompt, it will engage in inference and generate the corresponding information. Whenever you do this, what's running in the background is NVIDIA GPU. So most of the time, when you interact with our platform, you are doing inference. Therefore, 100% of the world's inference is currently provided by NVIDIA.

So, is inference difficult or easy? Many companies challenge NVIDIA in the inference field because when they see NVIDIA's system training, they think it looks too difficult, and they won't do it. They are just a chip company, but this system doesn't look like a chip at all. Just to prove if something new is effective, you have to invest $2 billion first, and then you start it, only to find out that it may not work. You've invested $2 billion and two years just to prove that it doesn't work. By then, you've already invested a lot of money and time. The risk of exploring new things is too high for customers. Therefore, many competitors tend to say, we don't do training, we only do inference.

Let me tell you now, inference chips are very difficult. You might think that the response time for inference must be very fast, but that's relatively simple because it's considered an easy part of computer science. The difficult part is the goal of deploying inference, which is to attract more users and apply the software to a large installed user base.Therefore, reasoning is a matter of installing the foundation. This is similar to those who develop applications for the iPhone because of its massive user base, almost everyone has one. Therefore, developing applications for the iPhone can bring huge profits that benefit all users.

NVIDIA applies the same principle. Our CUDA accelerated computing platform is globally pervasive. With our extensive experience in this field, if you develop applications for reasoning and deploy models on our architecture, they can run anywhere.

So, you can reach every user and have a greater impact. Therefore, the core of the reasoning issue is actually the user base, which requires long-term patience, years of successful experience, and a focus on architecture compatibility.

Host:NVIDIA's chips are unparalleled. But is there a possibility of cheaper competitors claiming to be good enough, although not as good as NVIDIA? Is this a threat?

Huang Renxun:First of all, competition does exist. We face more competition than any other company. We not only face competition from competitors but also from customers. And in their eyes, I am the only competitor, showing them not only the current chips but also the next and future generations. This is because if you don't explain why you excel at something, they will never have the opportunity to buy your product. Therefore, our strategy is completely open and transparent, collaborating with almost everyone in the industry.

We have several reasons for doing this, our main advantages are as follows:Customers can build chips optimized for specific algorithms (ASIC), but remember, computing is not just about transformers, especially since we are constantly inventing new variants of transformers. In addition, there is a wide variety of software because software engineers like to create new things.

NVIDIA excels in accelerated computing, but our architecture can not only accelerate algorithms but is also programmable, meaning you can use it to process SQL. We can accelerate quantum physics, accelerate all fluid and particle codes, among many other areas, one of which is generative AI.

Therefore, NVIDIA excels in the vast field of accelerated computing, one of which is generative AI. For a data center that wants to have clients in various industries, we have become the de facto standard, present on every cloud platform and computing company.

After over 30 years of development, our company's architecture has become an industry standard, which is our main advantage.

I would even be surprised if customers could have a more cost-effective alternative. The reason is, when you look at today's computers, they are not like laptops, they are data centers that need to be operated. Therefore, those who buy and sell chips only consider the price of the chip, while those who operate data centers consider the entire operating cost, deployment time, performance, utilization, and flexibility in all these different applications.Overall, our total cost of ownership (TCO) is very low. Even if competitors' chips are free, in the end, it still doesn't come cheap! Our goal is to add so much value that the alternative is not just about cost.

Of course, this requires a lot of effort. We must continue to innovate, and we cannot take anything lightly. I originally hoped not to sound too competitive, but the host asked a competitive question, and I thought this was an academic forum... This triggered my competitive genes, I apologize, I could have handled this issue more artistically.

Next time, I will better control the situation, but I was surprised by the mention of competitors' questions. I thought this was an economic forum, but you guys are so straightforward.

Is AGI five years away?

Host:

So, to be more direct, when do you think we can achieve Artificial General Intelligence (AGI), reaching human intelligence levels? Is it 50 years from now? Or 5 years from now? What are your thoughts?

Huang Renxun:

I will give you a very specific answer. But first, let me tell you about some exciting things happening. First, the models we are training are multimodal, which means we will learn from sound, text, and vision, just like all of us watch TV and learn from it. This is important because we hope AI is not only rooted in human values, which is where ChatGPT truly innovates, also known as RLHF (Reinforcement Learning based on Human Feedback).

But until reinforcement learning, humans will anchor AI in what we consider good human values. Now, can you imagine, you have to generate images and videos, AI knows hands won't go through the podium, and if you step on water, you will fall in, so now AI is starting to anchor in physics.

Now, AI watches a lot of different examples, such as videos, to learn the rules that govern this world. It must create a so-called world model. So, we must understand multimodality, as well as other modalities, such as genes, amino acids, proteins, cells, and so on.

Secondly, AI will have more powerful reasoning abilities. Much of the reasoning we humans do is encoded in common sense. Common sense is the ability we all humans take for granted. There is a lot of reasoning and knowledge already encoded on the internet that models can learn from.

But there are higher levels of reasoning abilities, for example, now when you ask me a question, for most questions, I can quickly generate like a generative model without needing much time for reasoning. But for some questions, I need to think, that is planning. I might say, "Interesting, let me think about it," and then I might be mentally looping it, building multiple plans, traversing my knowledge system, and selectively processing, "This doesn't make sense, but this I can do," that is, I will simulate and run it in my mind, maybe do some calculations, and so on.My point is, the ability for long-term contemplation that AI currently cannot achieve. No matter what prompt you give to ChatGPT, it will respond immediately. We hope to input a question to ChatGPT, give it a goal, give it a mission, and let it contemplate for a while. Therefore, this kind of system is referred to as System 2 in computer science, or long contemplation, or planning system, used to solve planning and complex reasoning problems.

I believe that as we research these issues, you will see some breakthrough progress. Therefore, the way we interact with AI in the future will be very different. Sometimes you just need to ask a question and get a reply, sometimes you might say, "This is a question, take some time to think about it first," and then after a lot of computation, it will provide an answer. You can also say, "I give you this question, with a budget of $1000, but do not exceed this amount." It will provide the best reply within the budget. There are also other scenarios and applications, and so on.

So, back to the issue of AGI, what is the definition of AGI? In fact, this is the first question that needs to be answered now. If you ask me, if you say AGI is a series of tests, and please remember, only engineers know that we have achieved it, you know, anyone in that prestigious organization will definitely know, for engineers, there needs to be a standard, a definition of success, and a test. Now, if I give AI a bunch of math tests, reasoning tests, history tests, biology tests, medical exams, law exams, and any other tests you can think of, SAT and GMAT, etc., list all the exams you can think of, put them in front of the computer science industry, I guess within 5 years, computers will perform well in all these tests.

Therefore, if your definition of AGI is through human tests, then I will tell you, it can be achieved in 5 years. But if you slightly change the way the question is asked, that AGI is to have human intelligence, then I am not sure how to clearly define all your intelligence. In fact, no one is really sure. Therefore, as an engineer, it is difficult to achieve. Does that make sense? So the answer is we are not sure. But we are all working to make it better.

The Role of AI in Drug Discovery

Host:

I have two more questions I would like to ask, and then I will hand over the time to everyone, as I believe there are many good questions out there. The first question I want to ask is, could you delve deeper into your views on the role of AI in drug discovery?

The first role is to understand the meaning of the numerical information we have. As you know, we now have a large amount of amino acid sequences. With AlphaFold, we can now understand the structure of many 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 chat with ChatGPT, you know, someone can chat with a PDF. You can take a PDF file, whatever content it may have, my favorite is taking a research paper PDF file, load it into ChatGPT, and start chatting with it.

Just like chatting with a researcher, asking where the inspiration for this research came from, what problems it solves? What is the breakthrough point? What was the previous state? Any novel ideas? Just like talking to a person.

In the future, when we put a protein into ChatGPT, just like a PDF, it will ask what it is for? Which enzymes will activate it? For example, if there is a gene sequence representing a cell, you can put this cell in, and it will ask you what it is for? What is your role? What is your purpose? What are your hopes and dreams?

So, this is one of the most profound things we can do, understanding the significance of biology. If we can understand the significance of biology, as you know, once we understand the significance of almost any information in the world, outstanding engineers and scientists in the field of computer science and computation will know exactly how to leverage it.

But this is the breakthrough, multimodal understanding of biology. Therefore, if I can give you a straightforward answer, I think this may be the most profound single thing we can do.

Advice for Students: Lower Your Expectations

Host:

Oregon State University and Stanford University must be proud of you. If I may switch topics slightly, Stanford has many students with entrepreneurial dreams, they are entrepreneurs, perhaps students majoring in computer science or engineering. What advice can you give them to increase their chances of success?

Ren Xun Huang:

You know, I think one of my biggest advantages is that my expectations are very low, I'm serious. Most Stanford graduates have very high expectations. You should have high expectations because you come from a great school. You excel in school, you are among the best students. Obviously, you have the ability to pay tuition, and you graduate from one of the best institutions on Earth. You are surrounded by other incredible kids. Naturally, you will have high expectations.

People with high expectations have low endurance. Unfortunately, endurance is crucial on the path to success. I don't know how to teach you unless I hope you experience some hardships. I am fortunate to have grown up in an environment that allows for success but also has many setbacks and hardships. Even today, I feel immensely pleased when I use the phrase "pain and hardship" within the company.

The reason is, you want to cultivate the character of 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 those who have experienced hardships.Therefore, what I want to say is, if I may offer you some advice, even though I don't know how to do it, for all Stanford students, I hope you experience an ample amount of pain and hardships.

How to motivate employees?

Host:I plan to break the promise and ask you one last question. You seem full of drive and vitality, but how do you keep employees motivated and energetic, especially when they may be wealthier than expected?

Huang Renxun:Yes, I have 55 people around me, my management team. So, you know, there are 55 people in my management team who report directly to me. I have never written individual comments for any of them. But I will continue to give them feedback, and they will also give me feedback. The compensation I give them is just the number at the bottom right corner of an Excel sheet, which I directly copy and paste. The compensation for many of our executives is exactly the same. I know it sounds strange, but it works. I never meet with them individually unless they need me.

I never have one-on-one meetings with them, and they will never hear from me about something that only they know. I do not disclose information to any employee that others in the company do not know. Therefore, our company is designed for agility, to allow information to flow as quickly as possible, for people to gain power based on their abilities, not on what they know. That's the structure of our company. I don't remember your question, but I understand.

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 setbacks? Every day, I look for opportunities to instill company culture, what is important, what is not, what defines good, how to compare oneself to good, how to think about good, how to think about the journey, how to think about the results, and so on, all day long.

How much additional chip capacity is needed for AI growth?

Audience 1:I have two questions, the first one is about the story of your leather jacket. The second one is, based on your predictions and calculations, how much chip capacity needs to increase in the next 5 to 10 years to support AI growth?

Huang Renxun:Alright, I appreciate these two questions. The first question, this leather jacket was bought by my wife, and this is what I am wearing now. Because I don't shop at all. As long as she finds a piece of clothing that doesn't make me feel itchy, I will keep wearing it, because she has known me since I was 17, and she thinks all clothes will make me itchy. When I say I don't like a piece of clothing, I will say it makes me itchy. So as long as she finds something that doesn't make me itchy, look at my wardrobe, it's all the same shirt, because she doesn't want to shop for me anymore. So this is all she bought for me, this is all I wear. If I don't like it, I will shop for myself, otherwise, I will wear it, which is good enough for me.

The second question is about predictions. In terms of predictions, I am actually terrible, but I am 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 are doing today, the information is written by others, or created by someone, is basically pre-recorded.When I say everything, every word, audio, video, is retrievable. Someone has written and stored it somewhere, and then you retrieve it. Every mode you know in the past has been like this. As I said, every time you click on your phone, remember that someone has written and stored it somewhere, all pre-recorded.

In the future, because our AI can access all the latest news in the world, which means it can retrieve, it understands your context, meaning it understands what you are asking, and most of the calculations will be generative. Today, 100% of the content is pre-recorded.

If in the future, 100% of the content will be generative, the question is how this will change the form of computing. So I no longer dwell on details, this is how I deduce this issue, like do we need more networks? Do we need more memory? In simple terms, we need more semiconductor fabs.

However, we are also constantly improving algorithms and processing efficiency, which has greatly increased over time. It's not that the efficiency of computing is as it is today, so the demand is so high.

At the same time, every 10 years, we will increase computing power by a million times, while demand will increase by a trillion times, these two must offset each other. Then there is technological diffusion and so on, it's just a matter of time, but it doesn't change the fact that one day, all the computers in the world will change 100%, every data center, infrastructure worth trillions of dollars, will completely change, and new infrastructure will be built on top of it.

Learning to code does not guarantee finding a job

Audience 2:

Yes, welcome to Stanford. You recently said that you encourage students not to learn to code. If that's the case, it may imply a few things. But do you think, from the perspective of company formation and ownership, will the future trend towards the creation of more companies or towards the integration of a few large players?

First of all, I said it so badly that you repeated it badly too. I didn't mean that. If you want to code, my goodness, please code **. Coding is a reasoning process, it's a good thing. But can it guarantee you a job? Not at all. There will certainly be a lot of programmers in the world, and NVIDIA will need programmers too. However, in the future, your interaction with computers will not be in C++.

For some people, it may be, but for you, why bother programming in such a strange language like Python? In the future, you just need to tell the computer what you want. Then the computer will act according to your needs. You can say, "Hey, I want you to create a construction plan, list all the suppliers and materials needed, and complete it based on the sales forecast we give you."

Based on all the necessary components, it will give you a complete plan. If you don't like it, let it write a modifiable program for you in Python. So remember, the first time you interact with a computer, just talk in plain English. The second time, use a programming language. By the way, the best programming language in the future is human language.The essence of programming is changing, and future interactions with computers will be more natural and intuitive. Through AI, we are narrowing the technology gap. In the past, only a few who understood programming could work in related fields, but today, almost everyone can guide computers to perform tasks through simple verbal commands. There are many people on YouTube, including children, who interact with ChatGPT to accomplish amazing tasks without the need for traditional programming skills. This indicates 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 geopolitical risks, we are almost a typical representative of such risks because we produce tools crucial to AI, which is considered the decisive technology of our time. Therefore, the United States has every right to decide to restrict these tools to countries it deems appropriate.

This limitation restricts our opportunities in certain regions but also creates opportunities in other regions. In the past 6 to 9 months, every country and region has realized the need to master their own AI technology and not send their data abroad for processing before returning it domestically. This awakening to sovereign AI has created tremendous opportunities for us.

Lastly, concerning the issue of customizing solutions for customers, why is the current threshold relatively high? This is because each generation of our platform includes GPUs, CPUs, network processors, software, and two types of switches. I have built five chips for one generation of products, although people think there is only one GPU chip, in reality, there are five different chips, each with a research and development cost of hundreds of millions of dollars, just to meet our "release" standards. Then you have to integrate them into a system, along with network equipment, transceivers, fiber optic devices, and a large amount of software. Running a computer as big as this room requires a lot of software, so everything is quite complex.

If the customization requirements vary too much, you have to repeat the entire research and development process. However, if customization can leverage everything existing and add something on top of it, then it becomes very meaningful. It could be a proprietary security system, an encrypted computing system, a new numerical processing method, and more. We are very open to these possibilities.Our clients know that I am willing to do all these things, and they realize that if you change too much, you basically reset everything, wasting nearly a hundred billion dollars. So they hope to make the most of these in our ecosystem (reduce the cost of resetting as much as possible).