Chairman, Sharp Corporation
Foxconn Chief Young Liu Keynote at Computex 2025 | Tech Vision | AT World News LIVE
Foxconn Chairman Young Liu delivers a powerful keynote at Computex 2025 in Taipei, Taiwan. He shares groundbreaking insights into AI, semiconductors, and the global tech supply chain. Watch the full speech to explore Foxconn’s vision for the future of innovation and sustainability.
#Foxconn #YoungLiu #Computex2025 #TaiwanTech #AI #Semiconductors #KeynoteSpeech #TechInnovation #SupplyChain #Electronics
Subscribe to the Amaravati Today, AT World News channel and tap the bell icon to receive notifications whenever we go live
/ @atnewsroom
#atworldnews #atworld #AmaravatiToday
Transcript (38 segments)
U
Unknown3:55
Let's rock.
Let's rock.
Ladies and gentlemen, please welcome chairman of Taitra, James Huang.
J
James Huang26:26
Good morning, everyone. Welcome to Computex. I'm James Huang, the host of Computex. As the organizer of Computex, Taitra is honored to host global technology leaders. Today is particularly special as we welcome Foxconn to the Computex keynote stage for the very first time. Under the leadership of Chairman Liu, Foxconn has been undergoing a remarkable transformation, placing AI at the core of its growth strategy. From AI-optimized manufacturing to smart electric vehicles, AI data centers and edge computing, Chairman Liu has redefined Foxconn's role not just as a manufacturer, but as a global enabler of intelligent manufacture. With deep engineering expertise and a clear vision for the future, he is driving Foxconn to the forefront of the AI economy. But before Chairman Liu joins us, let's watch a quick video on Foxconn's AI vision.
U
Unknown28:49
Ladies and gentlemen, please join me in warmly welcoming the chairman of Foxconn, Mr. Young Liu.
Y
Young-way Liu29:06
Thank you. Thank you, James. Good morning, everyone. It is a pleasure to be here with you today. You have seen so many factories with so many robots running in our current factories. And these factories, later I'm going to show it to you that it has changed. It's changed because of AI. So, where's James? Is James gone? Okay. Anyway, I'd like to thank James for inviting me to be here for this very important event, Computex, to say something about Foxconn's endeavor in AI. And after taking this mission, I've been thinking about what should I deliver. So what can our customer, our audience, feel beneficial after hearing this speech? So I've been thinking about this. After so many days of thinking, finally I thought I'm not going to talk to you about technologies, because we have very technical experts who are going to talk about technologies. I'm not going to share with you our business in AI, because some of it is confidential. But what I decided to share with you is the train of thoughts of how Foxconn got to this point with our AI factories. So next, it actually all starts from here.
About a year and a half ago when we had our HHTD '23, we sat in the meeting room in the hotel with Jensen. And Jensen hand-drawn this picture. I looked at this picture and said, 'Well, what does it mean?' Ever since, this picture's spirit is there but it's changed, it evolved. So let's look at what we have done, with a video that we prepared together with BCG, to show you what we have gone through in the last year and a half. Let's play the video.
Okay, so this is Genesis. And this is still a work in progress. So there are still things evolving. Through this journey we came up with many new ideas. And the very first new idea is about the future factories. You will have one physical factory together with two other factories which are very, very important. One you've probably heard about — the digital twin, the Omniverse digital twin factories. So before you create your physical factory, you can build your factory on Omniverse. And practice it and optimize it without the actual factory being built. In the past we thought with Omniverse that was already good enough. But with Omniverse it can already create data. And with that data, if we use it properly with the AI factory, then we may be able to create a base set of models that can be used when the actual factory is built. So we think in the future the factory will be like this. The physical factory will appear last. The first one you will have is the Omniverse digital twin factory. And then the AI factory will be the second one. And once you have the model built and the digital twin optimized to a certain level, then you can have your actual factory. And that's the vision we see for the future factories.
And what you have seen is that with the BCG effort, we have gone through so many use cases. These use cases, after accumulating many, many hours of work with our technicians, we have a very important finding which I would like to share with you. Before this experiment or this study, we thought with GenAI we could maybe replace humans. But very quickly we realized no, it will not. If you look at this chart, the white curve is a human technician. How many times it will take the technician to try to reach a certain level. That's typical human. And with the blue and green lines, this is done with GenAI's help. And you can tell with GenAI, very quickly, with two to three times of trying, it can reach 80%. But after 80%, look at the curve. It flattened. The human can do much better than GenAI after 80%. So we came up with this temporary conclusion: with GenAI, it can help for the 80% of the work. The rest of the 20% will still have to be done by skillful workers or technicians. That's a very important finding I would like to share with you.
And because of the scale of Foxconn's factories, we have central divisions. And we also have many BGs. And the BGs have factories across the world. How can we come up with an architecture that will be able to support this large-scale and very complicated implementation? And at the same time allows the model to evolve by itself? It requires a very, very complicated architecture that becomes very, very important. It's still evolving. But we think with this initial architecture we should be able to get to what we want, so that the base model created by the central department can then be utilized by the BGs' factories all over the world. And then they can continue to train with their factory actual data. And those trained models can then feed back to the central model. And the model in the central can then share the results to all the factories in the company. So this is the architecture that we came up with. I just want to remind the audience that the architecture is very, very critical. It's not just the LLM itself. How you're going to deploy the result of the LLM or GenAI is becoming very, very critical.
While we're doing smart manufacturing, we are at the same time working on EVs. You probably heard of our HHT Tech Day. This is the annual event where we show our new EVs that we created in the last year. The very first product that we delivered, that we shipped to the market, is the e-bus. And with that e-bus running on the street in Kaohsiung, with a lot of sensors on the e-bus, we collected a lot of data of the street. And very quickly we realized: how are we going to utilize this data? We started looking for ways to share the data with the government and with enterprises. And then we realized that a lot of the so-called smart city applications all work in silo. We'll talk about the smart city platform on the next page to address that problem. But with smart EV, we are going to provide the smart EV platforms to our potential customers. These platforms will include onboard and offboard applications. We will open these platforms to our customers, and also we plan to make it a reference design for our MIH members.
And with smart EV, the most recent news is that we made the announcement together with Mitsubishi. Mitsubishi will be the very first car OEM to utilize our reference models. With that reference model, our customers can use our models — maybe 80% of their work can be eliminated. They only have to put 20% of their effort into making their new EVs. That's very, very common in the PC industry. A good reference design gives a good foundation including very thorough tests required for the systems. That saves our customers a lot of time. That significantly improves the time to market and time to cost for our customers. Please stay tuned for this new EV announcement.
We talked about smart city while we were doing the e-bus. We realized that most of the smart city applications — be it smart transportation, smart hospitals, smart schools, or smart buildings — they all work in silo. They don't work together. And we realized that we need a platform for all these smart city applications to build on top of. So that with this platform, it will connect the government, the enterprises, and the citizens through this platform. And this platform will support not only data sharing but also knowledge sharing.
With this smart city platform, we are currently implementing it in Kaohsiung. We also have a number of other cities in Taiwan currently working with us on the smart city project. We also have cities outside of Taiwan working with us. Once we have the smart city platform done — sometime in about a year — we can show you how easy and useful it is to share data and knowledge between all these applications.
Therefore, with smart manufacturing, smart EV, and smart city, all these ideas came about because of that little drawing. We've been thinking, maybe this is it — we need AI factories. AI factories to power these three smart platforms. That's how we put all this together. And we made these three platforms examples of what an AI factory can do. We demonstrated this idea at GTC '24 and '25. You've probably heard about the 3+3 initiatives in the past from Foxconn. Now we add another three. Now the 3+3+3 smart platforms become Foxconn's new initiatives.
With these three smart platforms, we need to find a model to support them. We've been looking for models adequate to support these three smart platforms. We kept looking and realized that most foundation models are very generalized. They don't understand manufacturing. Especially the data — they're all numbers. But the numbers from different machines mean different things. You have to have some understanding of this data in order to understand what it means. With that finding, together with EVs, a general-purpose LLM can't handle it. A general-purpose LLM can probably answer where you want to go, but will it know if something is going to break? The signal sent from a certain module means something will happen in the near future. This knowledge and these tokens are all very, very different. If you don't have special training or a special way of processing these tokens, it will give you not just hallucination but also wrong answers. Therefore, we better come up with our own model to support these three applications. And therefore the Foxconn Brain idea came about.
It is built on top of, of course, open-source models. Currently we're using Llama 3. And we've created high-quality, large-scale pre-training corpus. Foxconn's Kaohsiung data center will be used, and currently it's being used to build this model. This first-generation model is specialized in reasoning, and that reasoning is a little different from general reasoning because of the applications — the agentic workflow for very domain-specific applications. This Foxconn Brain is going to solve the problem I just described. And we plan to make it open-source to the communities. Hopefully, through your joining the open-source communities, we can make this model more powerful, more useful.
So with the Foxconn Brain, we think: now that we have the applications — the three smart platforms — we know we're going to build a Foxconn model to support them. Then, what do we build those on? We looked around and realized, Nvidia is not just a chip company. It has a full stack of software to support it. We realized that with all the platforms we talked about and the Brain we're going to create, Nvidia has a full stack to support us. So we're going to build our platforms on top of the Nvidia AI software stack. Then, what hardware should we use? Right, but definitely Nvidia hardware. We need hardware to run it. We need tremendous compute power to support it. And that's why you heard the announcement yesterday.
The AI factory will be built in this very first NCP AI data center. This AI data center is targeted to have 100 megawatts of power. Power is a very, very critical resource. So it takes us a couple of steps to build it up to 100 MW. We will build about 20 MW to start with, then add another 40 in the next phase, then another 40. Some of it will be in Kaohsiung. For the others, we'll look around in Taiwan for where we can get the power the quickest, and we will build the data center there.
So with all this AI-related — the applications, the Brain, the software stack, the hardware, and computing power — all comes within the last 18 months. We've gone through this entire journey and exercise to come up with the current status. Together, we work with BCG and Nvidia and other partners to create the whole ecosystem. Hopefully, with all these ecosystems, we're able to become the leader in AI manufacturing at the least. And if we're lucky, we can become the leader in the smart city platforms. That's very likely to happen. If that happens, we estimate the compute power needed for supporting the smart cities will be huge. About a year ago, when people asked whether I see a slowdown of compute power requirement, I told them no — this is just the beginning. We see the applications coming, and we also see the evolution of the models. It will only become bigger and bigger. Don't worry. It will not slow down, at least in the foreseeable future.
This is another discovery we had while working on the smart city. We looked at demographic data and the issues in countries. We realized that when a country becomes developed, what happens? If you look at the GDP pyramid, the triangle — when a country becomes more and more prosperous, the pyramid or triangle will shift up, leaving the bottom GDP empty. After your country becomes more and more developed, the GDP will become higher and higher. Then the low-GDP work will have fewer and fewer countrymen to do it. How did countries solve that problem in the past? Two ways. One is to outsource the low-GDP work to low-GDP countries. We've seen that — the US outsourced work to Taiwan, and then Taiwan outsourced work to China. When we outsourced from Taiwan to China about 40 years ago, the GDP there was less than 1,000. With their hard work, after 40 years, the GDP has grown over 10,000. What happened? What was the impact to society? Most people want high-paid jobs, and the society is able to pay. But what about the low-GDP, low-paid jobs? Nobody will be interested. Eventually they're going to outsource low-GDP jobs to other low-GDP countries. Eventually you will run out of low-GDP countries. That's the limit.
Another way to solve that problem is to import people from low-GDP countries — the so-called immigrant workers. This has been done in Europe, in the US, in Japan, and now in Taiwan. But that creates social problems. In the past, there was no other way to solve that problem. Now, with the advent of GenAI, we see great potential: GenAI plus robotics will fill the void, the gap. And that's the opportunity I see. It's a tremendous opportunity — GenAI plus robotics to fill the void or the gap when a country becomes more prosperous. The low-GDP work will be done by GenAI plus robotics. It's not just service. It's not entertainment.
I think that's the real challenge for all developed countries. So, that's a great opportunity for all of us. And I urge you and the leaders of the developed countries to watch the development of this very, very carefully.
Okay. So, before I introduce, this talk is broken into two parts. The first part, I am almost done with it. The second part, we'll have our special guest joining us. But before I introduce that man, I would like to make a brief announcement that Foxconn will work with Tech Orange to create Taiwan's leading robots community to drive innovation and industry growth because of that GDP paradigm shift syndrome. Okay?
Is our guest here? Okay. Good. Wait a second. Okay. Let's play. Okay, ladies and gentlemen, let's welcome the superstar of the AI industry and the leader of Team Taiwan. Go Team Taiwan.
J
James Huang1:06:33
I still love being introduced.
Y
Young-way Liu1:08:42
The very first question. What is Team Taiwan from your point of view?
J
James Huang1:08:47
Well, Taiwan is where our most important partners are. We've been coming here for 30 years. Somebody did a count. We have 350 partners here. Of course, I don't know them all, but apparently, either directly or indirectly, we work with some 350 companies. And you know that the technology supply chain is incredibly right. And it's much more complicated than people think. As a computing company, this is the epicenter of the world's computing industry. This is where it starts. Everything from the chips, of course, to the systems and now increasingly the pioneering work that is being done in the automation of robotic factories and building robots. This is one of the largest electronic manufacturing regions in the world. Not only is it the largest, it's the most advanced. The things that we build here are incredibly complicated. For example, GB200 servers have 1.2 million components, weigh almost 2 tons, are completely liquid cooled, and each rack costs a couple of million dollars. The technology is incredibly complicated. To build one of those Grace Blackwell racks, many companies are involved and much robotics technology is used. Since we started talking about AI and robotics, Foxconn is using Nvidia Metropolis for smart cities and factories, Nvidia Isaac Groot for robotics, Nvidia Drive for autonomous vehicles, and Nvidia Omniverse as the digital twin operating system. Yesterday, we announced building an AI factory for Taiwan's ecosystem. Foxconn is becoming a world-class regional AI cloud provider. The breadth of work covers everything from edge agents to robots, including manufacturing inspection agents with computer vision. This is robots building robots. Foxconn is one of the most advanced technology companies, applying technology to manufacturing.
Y
Young-way Liu1:13:45
My next question is, have you heard about Tema? T-E-E-M-A, the Taiwan Electronic and Electrical Manufacturers Association. I was just made the chairman of Tema. It's the largest association in Taiwan with more than 3,000 members, accounting for over 50% of Taiwan's export value. With this rapidly changing AI industry, what advice do you have for your suppliers in Tema?
J
James Huang1:15:30
I do. Let me make an observation. Vast majority of industrial companies have been left behind during the software revolution because they understand mechatronics, mechanics, chemicals, and the deep physics of industry. Taiwan is excellent in these industries. But when software came, it was a new skill, unusual and unnatural for these companies. The second observation is that when AI started 10 years ago, it was interesting but not very useful for industrial companies because AI understands English but not physics, which is needed for physical understanding. Now we're developing both. AI is great because you teach it, you don't have to program it. Software programming skills are no longer as vital. You have an opportunity to leap the software chasm and go directly to AI. Everybody is a software programmer now. If you don't know how to tell ChatGPT to do something, ask it to teach you. It will write a prompt for you. This computer is now intelligent, and intelligent systems are easier to use. We combine all of that into Omniverse, which you use as a digital twin. Within the digital twin, you can design, operate, and simulate your plants. Every company in Tema is now a technology company and has to be.
Y
Young-way Liu1:21:52
My next question for you: Do you see any limit in terms of hardware speed for AI?
J
James Huang1:21:58
Remember, Moore's law ended because we're at the limits of physics. The cost of transistors is going up, and energy efficiency is not improving fast. We have to try different skills and techniques, like multi-chip packages, 3D packages. NVLink connects all chips together into one giant chip. We use advanced mechanical and liquid cooling systems to compress systems into one giant rack. Software and algorithms allow us to reformulate AI models and distribute work across supercomputers. Networking is a big challenge and opportunity. Instead of passing information back and forth, we can do in-network, in-fabric reduction, reducing traffic practically in half. These clever ideas combined have allowed us to improve AI performance tremendously, probably faster than 2x every 6 months, more like 2x every 3 months. It's full stack innovation: chips, systems, data center, operating system, and models all at the same time. From where I see, only the sky is the limit.
Y
Young-way Liu1:27:01
I want to ask you something personal. There's an old Chinese saying that every successful man stands a terrific woman. I'm not going to ask you who that woman is, but do you want to say something to her openly?
J
James Huang1:27:20
I say the same thing to her every day: Love you, honey. I met her when I was 17 and she was 19. We were lab partners in electrical engineering in Oregon. There were 250 students in electrical engineering fundamentals, only three girls. I was 128 lbs going to college. I noticed this beautiful girl and used my superpowers by asking, 'Would you like to see my homework?' I told her if she studied with me every Sunday, she'd get straight A's. So, I had a date every Sunday.
Y
Young-way Liu1:30:36
Thank you very much, Jensen. We are very happy to have you here. Through this conversation, we know the other side of Jensen, and hopefully, it will help people to know Jensen more besides his chips and success in business.
J
James Huang1:30:58
My chips are pretty good. I am delighted by our partnership. We're manufacturing the most advanced computers the world's ever seen. During our meeting yesterday, one of your leaders said no one has ever manufactured systems this complicated or supercomputers at this rate. We're cranking it out every hour. The idea that we're doing this at scale is extraordinary. The thing I'm really excited is that you use this incredible AI technology to revolutionize Foxconn and make it available for all of Taiwan. You're revolutionizing your own company with technology. You were already an amazing technology company; now you're extraordinary. Everything you do for Taiwan, I'm very grateful.
Y
Young-way Liu1:32:15
Thank you. Thank you very much, Jensen, and thank you everybody. This is it. Bye-bye.
U
Unknown1:34:13
Yeah.