AI Revolution: Semiconductor Industry and Future Developments

The AI revolution is causing stock fluctuations stemming from policy uncertainty and liquidity reduction, impacting semiconductor companies.

### The Current and Future of the AI Revolution: Challenges and Opportunities
* 📉 Recent stock market fluctuations in the semiconductor industry were triggered by **policy uncertainty, excessive tech investments, and liquidity reduction**, affecting AI semiconductor companies like NVIDIA.
* 🎮 The AI revolution, which started suddenly with OpenAI's ChatGPT in 2022, created a **shortage of AI semiconductors**, benefiting NVIDIA due to its **strong software ecosystem**.
* 💻 NVIDIA's dominance in AI semiconductors may be challenged as big tech companies like Microsoft and Google start designing their own chips to **reduce costs for inference services**.
* 📈 NVIDIA's CEO expects the AI semiconductor market to **grow 15-20 times** in the next 10 years, aiming to maintain a **60% market share**.
* 🔋 The memory semiconductor market is shifting towards specialized, made-to-order solutions like **HBM**, with **SK Hynix leading** and **Samsung playing catch-up**.
* 🇰🇷 South Korea's semiconductor industry, traditionally strong in memory, is facing challenges in catching up with TSMC in foundry services due to **organizational culture and IP limitations**.
* 🧠 The AI value chain includes foundation AI models (L&M models), specialized models, and application AI models, with global big tech companies investing heavily to **catch up with OpenAI's ChatGPT**.
* 🌐 Sovereign AI, the concept of developing AI based on a country's own language, is seen as **crucial for national security and economic competitiveness**.

Today, we will meet the present and future of the Plus AI revolution. We have Lee Young-soo, the CEO of HSL Partners, with us. Welcome! 

Hello! It's nice to meet you. I occasionally watch your videos on YouTube, but this is the first time I'm seeing you in person. 

Oh, I'm really a big fan of yours! 

Oh, why are you saying that? Thank you! If someone saw this, they would think we are just acting. 

For a while, the stock prices in the semiconductor industry have only been rising, but recently they seem to be fluctuating significantly. What’s going on? 

It seems the trigger came from Trump. After that shooting incident, he started making all sorts of statements, claiming, I became president, and then he began to shoot off his mouth. He touched on TSMC, saying that TSMC took our semiconductors and that we need to protect them even though they aren't paying us properly. There’s a bit of nuance regarding the CHIPS Act, which is one of the two key policies of the Biden administration, along with the Inflation Reduction Act.

There has been a change in the law. I thought that Trump viewed the IRA law negatively, as he has always favored traditional fossil fuels. The market had already priced that in to some extent. However, suddenly bringing up the Chips Act and discussing it became a negative signal. Additionally, during the earnings announcements of big tech companies, there was too much investment; the capital expenditures were excessive. People started questioning how they could make money from AI. As a result, Sundar Pichai from Google Alphabet struggled to respond adequately. The market is currently crashing, and this is due to the carry trade funds ultimately reducing liquidity. Up until now, the liquidity had been concentrated in...

It's about AI semiconductors. It seems that we have now reached a point where we are facing these issues. I see, there are concerns about policy uncertainty and whether there will be results in relation to investments. Additionally, a lot of liquidity is being withdrawn, and all these factors are overlapping, causing stock prices to fluctuate. 

But one more thing I'm curious about is that the AI boom has mostly been concentrated in semiconductor manufacturing or companies related to manufacturing. What is the reason for that? They say we are in an AI revolution, and this revolution is said to have started after OpenAI released ChatGPT in November 2022, which has almost become a widely accepted fact. However, revolutions are never prepared; they start suddenly, right? When they start suddenly, the urgent matters tend to receive the most attention. If we look back at the mobile revolution over the past decade, we can see that when mobile phones transitioned from feature phones to smartphones...

In Korea, the most urgent change was the shift in the user interface. The user interface transitioned from pressing keypad buttons to suddenly using a touchscreen. Because there were no touchscreen panels available, the stocks of touchscreen panel manufacturers surged. After that, there were developments in semiconductors, multi-core processors, display resolution, and cameras, and later on, metal cases, all of which were quite predictable. However, with the recent AI revolution, it started suddenly, leading to a shortage of AI semiconductors needed for AI servers. In fact, it seems that the OpenAI initiative triggered this AI revolution, which was not anticipated. If I had known that the AI revolution was about to begin, I would have focused on the core components of NVIDIA chips. Consequently, the foundries faced supply shortages, which affected the entire value chain.

NVIDIA and its associated stocks have seen significant price movements over the past year and a half. Your explanation about how this relates to the smartphone app boom makes it very clear. Back then, we experienced a sudden shift in interface methods to touch screens, which caused related stock prices to soar. I understand that this phenomenon is quite similar. In this process, it seems that NVIDIA has been the clear beneficiary. Despite the presence of various semiconductor companies, what do you think has allowed NVIDIA to dominate the market to such an extent? It seems that when a sudden change occurs, people tend to look for the most effective products available. NVIDIA has consistently produced high-performance chips, but I believe that the software ecosystem they have developed has played a crucial role in their success.

Since 2006, NVIDIA has started to provide CUDA software for free. This means that if you buy an NVIDIA chip, you can use CUDA for free. As a result, a lot of source code accumulates, similar to how past exam papers are important when we take university exams. The programs that have the most accumulated source code are often used to create AI models, as they tend to utilize existing source code extensively. They also make significant use of open source. Then, they upload their own developed core algorithms to create models. The source code that has accumulated the most is from NVIDIA. Consequently, there has been a greater focus on that area. The AI revolution is ultimately a race for speed, so securing computing power quickly and creating models is crucial, and in that regard, NVIDIA is seen as the most attractive option.

AI ecosystem is managed directly by NVIDIA, and the biggest reason for providing services to people is that it helps develop the AI ecosystem based on their hardware. Yes, that's correct. In fact, Jensen Huang once confessed that he didn't know whether AI chips would be used. The technology behind AI is very similar to the history of computer technology. Originally, computers were first conceived by the genius John von Neumann, and that concept has continued to evolve. In fact, there was a contemporary of von Neumann named Alan Turing, who can be considered the father of AI. He became very famous for decrypting the German encryption system, Enigma, during World War II, but at that time, he was not recognized for it and passed away too soon. So, AI technology is essentially a part of computer history.

The speed of technological advancement has actually slowed down, and we can say that it has stagnated for almost 40 years. However, after that, the speed of computer development increased, leading to higher computing power and a significant accumulation of data, which enabled the advancement of AI. In the past, everything was programmed individually by engineers based on rule-based systems, which was highly inefficient. This led to edge cases, where unexpected scenarios would arise. Since then, we have thrown massive computing power and big data at computers, allowing them to learn on their own. This resulted in improved efficiency. The technology of machine learning became the trend in the 2000s. However, there is another important technology that emerged in the 2010s, which is what we are discussing now: deep learning.

There is a technology, and there is a genius named Jeffrey Hinton who appears. Originally, deep learning is a technology that mimics the human neural network within machine learning. It was a technology that didn't work well at first, but Dr. Hinton completely transformed it, making it a globally renowned technology. The important point here is that until then, AI research was based on CPUs, but after deep learning, it shifted to GPUs. At that time, Jensen Huang noticed that the PhD students at Stanford were taking our chips and using them for strange purposes, so he secretly looked into it. He found out that they were programming AI with those chips. Since then, Jensen Huang has put a lot of effort into AI, and that ultimately led to the Google Transformer that came out in 2017.

There is an undeniable trend emerging, and I don't think other companies will just sit idly by during this process. What is the current situation? What are other companies doing? 

First of all, when we talk about AI semiconductors, NVIDIA still holds about 90% of the market share in AI semiconductors and AI acceleration. However, AI semiconductors have two main purposes: one is for data training, and the other is for providing services based on the trained data, which is referred to as inference. While there are two uses, it is still challenging to replace NVIDIA chips for data training. Researchers in AI still need to use NVIDIA chips for data training because they rely on CUDA, which is a significant factor. However, as the focus shifts more towards inference, the service costs from the perspective of platform companies are becoming increasingly important.

So, there is a need to reduce costs in this area. Data training, in a way, doesn't quite align with cost reduction in investment transactions. However, inference is ultimately being conducted based on service pricing. Notably, big tech companies are now starting to design their own chips. For instance, Microsoft has recently introduced the Maya 100 chip, and Google has developed the TPU, or Tensor Processing Unit. They are increasingly using custom-designed chips to cut costs, especially since NVIDIA chips are quite expensive. Listening to your comments, it seems that if companies like Microsoft and Google, which could be NVIDIA's biggest customers, are starting to take matters into their own hands, it raises questions about whether NVIDIA's dominance can continue in the future.

I wonder if we are now entering a competitive landscape. What is your outlook on this, our representative? Well, rather than my outlook, Jensen Huang recently made some comments. He said that having a 90% market share, as we do now, is abnormal. Ultimately, it is evident that our company's market share will fall below 70%. It's not easy for a monopoly to say such things. However, the important point is that within the next ten years, the AI semiconductor market is expected to grow 15 to 20 times, and we aim to maintain a market share in the 60% range, which is exactly the position of TSMC. TSMC maintains a 60% market share in the foundry market. From their perspective, it is clear that there is currently a significant concentration of market share, but since the AI market is expected to grow nearly 20 times,

Even if our share is a bit better, the overall pie we have is growing. That's what you meant, right? Yes, that's correct. So, if we summarize what the CEO has said, it seems that the AI market is still in its early stages, and it is a time when we need to invest diligently. It appears that some big tech companies have just started investing. So, how long do you think this aggressive investment will continue? Because we can't keep investing indefinitely; after the investment, it needs to turn into profit. As you mentioned earlier, there is a lot of money being invested, which raises the question of when profits will actually materialize. Until when should we continue investing? I believe this is something the professor has experienced much more than we have. In the early 2000s, during the late 90s and into the 2000s...

In the early days of the World Wide Web, during the internet revolution, and then again in the 2010s when the iPhone was released, the mobile revolution began. At that time, people in the market were saying the same things. In the early 2000s, there were discussions about why so much money was being poured into dominating just a single search bar. When the mobile era began, Apple was internalizing everything, while Google started offering Android for free to smartphone manufacturers. However, Google was pondering how to make money from this. Ultimately, when a monopoly is established, it can generate significant revenue. A platform that attracts a massive number of users can earn a lot of money. However, in the case of AI, a monopoly has not yet been fully established.

I believe that in order to dominate the service sector, we inevitably have to engage in a continuous arms race. Many people find it hard to visualize this because AI is still in progress. For example, if we say that AI monopolization has been completed, does that mean that all portal sites are transforming into AI-based portal sites? What kind of landscape are we talking about? Or are the big tech companies working on the software behind the scenes, whether it's an L&M model or something else? Using that engine, each app ecosystem is not something we search for individually; rather, it should automatically provide us with services. Only when every single service unit is transformed by AI can we say that AI has truly changed everything. At what stage do you think we are? How far do you believe investment needs to go? Ultimately, I...

I believe that the ultimate goal of big tech is AGI, or Artificial General Intelligence. It is often translated as general artificial intelligence, but I think it is more accurately described as superintelligence. This means AI that surpasses human capabilities in all areas. In George Orwell's 1984, we see the character Big Brother, which is quite similar to this concept. If this AI is a good entity, it could become something like JARVIS from the movie Iron Man. However, if it has malicious intentions, it could turn into Skynet from the Terminator series. Ultimately, I believe that until AGI is secured, big tech will continue to invest heavily in this area. I see AI as the most crucial infrastructure in modern civilization. In the realm of technological civilization, I think there are two key components: the first is electricity. Electricity is essential for everything.

That technology is the foundation. And considering the two revolutions we've had, the World Wide Web and the mobile revolution, communication has now become a basic infrastructure for all modern civilization. I believe that AI will become the infrastructure of the third technological civilization. So, understanding AI is not just about a specific industry's position; rather, it will be integrated across all industries, functioning like an operating system. 

I see. Thank you for your insights. If I may ask, for an industry to mature after it has been conceived, it needs to be interconnected both upstream and downstream, right? As you mentioned, one of the most important elements in the current industry is electricity, followed by communication. The same goes for AI.

We all know that this consumes a lot of electricity, but there seems to be a doubt about whether the electricity supply is sufficient for this industry to develop. There aren't many countries with a network environment as good as ours. In this context, can AI-based smart cities, smart homes, and autonomous driving really be born? It seems that legal adjustments also need to be made. If all these things are not fully prepared, should investments still be made? There might be concerns that AI has already reached its peak. What do you think? I believe the keyword that the market loves the most is supply shortage. When there is a supply shortage, people in the industry get anxious, but the market really loves this keyword. Yes, as long as the keyword supply shortage is alive,

The stock price continues to fluctuate. In fact, the stock price of NVIDIA started to decline somewhat as the supply shortage began to ease. Previously, if you ordered an NVIDIA chip, you would have to wait about 40 weeks, but now, with increased supply, you can receive it in about 7 to 8 weeks. It was around that time that the stock price began to hesitate a bit. Currently, we are facing shortages in electricity, communication infrastructure, and other areas, which indicates that the AI cycle still has a long way to go. Therefore, it cannot be ignored. This is because AI is directly linked to defense technology. We can see how much AI robots are contributing to military power in the ongoing Russia-Ukraine war. Additionally, many of the big tech companies in the U.S. are directly or indirectly involved, and the impact of AI on economic growth is becoming increasingly significant.

Now, in terms of economic growth and development, the level of AI adoption in a country affects its economic growth. From the perspective of nations, companies, or individuals, this can foster a sense of selfishness, as it allows one to feel they can get ahead of others. As long as this exists, capitalism will never falter. Consequently, we are still facing a supply shortage in AI, which has garnered significant attention from many people. Speaking of this, let's briefly discuss our semiconductor companies. Given that our semiconductor companies are still in the midst of this supply shortage cycle, can we expect this situation to continue for a while? This is from the perspective of our industry. However, the viewpoint of the stock investment market is somewhat different. From the industrial perspective, as you mentioned, it is currently maintaining a strong performance. The semiconductor exports are astonishing. Recently, our country has been exporting to Taiwan.

The export data for semiconductors is very good. This is because of the HBM that goes into TSMC, which stands for High Bandwidth Memory. It connects RAM vertically, and it is indeed gaining a lot of traction. Additionally, the solid-state drives (SSDs) that are entering the U.S. market, particularly for server storage, are also showing very positive data. However, the market has already largely priced this in, so to some extent, this issue may have already been reflected in the stock prices.

Now, let me ask about our country. Traditionally, the memory semiconductor market has been a stronghold, right? What position do you think the memory market will establish in the future, especially in the context of artificial intelligence? I believe the landscape of memory will change. If we think about the PC era, our system semiconductors were previously categorized as non-memory.

We had a system semiconductor subcontractor, and when you plug it into the board, it works. So, we went to Yongsan to assemble our PCs. But now, with the start of the mobile revolution, what has changed is that our mobile phones have become computers. The processors that used to be general-purpose, like CPUs and GPUs, are now application processors, meaning that companies like Apple and Google have started to design their own. Companies like Qualcomm also design these, and then they are produced by foundries. As a result, the processor brain chips have become more specialized. They have shifted from being general-purpose to being customized. However, now with the beginning of the AI revolution, the memory that used to be general-purpose, like RAM or NAND flash, which could work anywhere, is now...

It has changed to a made-to-order system. This is the HBM3 item that goes into NVIDIA, and it needs to be designed and manufactured according to NVIDIA's specific protocol. In the case of existing memory semiconductors, they are subject to commodity regulations, which means there is significant price volatility. Orders can be cut suddenly after being placed. However, with HBM, we are contracting to purchase a fixed quantity at a fixed price for a year. Therefore, I believe that the demand from companies will increasingly lead to memory that is optimized for their specific systems. The memory semiconductor market is undergoing significant changes; it is no longer a market where semiconductors are made generically, stored, and then sold in small quantities. Instead, it has shifted to a world where they must be tailored to fit specific systems.

주셨는데, 그 과정에서 가장 뜨거워지고 있는 개념이 HBM인 것 같아요. 지금은 하이닉스가 굉장히 앞서가는 것처럼 보이는데, 삼성전자와 다른 회사들, 예를 들어 마이크론 같은 회사들은 HBM에 대해 현재 어떤 상황인지 궁금합니다. 사실 SK 하이닉스는 디램이라는 회사로, 굉장히 오랫동안 개발을 해왔습니다. 이 기술은 2.5D 패키징을 사용하는데요, 요즘에 좀 어렵게 이야기하면 이종 집적 반도체라고 할 수 있습니다. 즉, 프로세서와 메모리가 최대한 가까이 붙어 있는 구조로, 약간 주상복합 건물과 비슷하다고 생각하시면 될 것 같아요. 한 건물에 상가도 있고 주택도 있는 구조라고 보시면 되는데, 이를 구현하기 위해서는 서로 간에 굉장히 많은 연구 개발을 오랫동안 해오면서 데이터를 확보해야 하고, 이런 것들이 필요하잖아요. 그런데 사실 디램도 이렇게 칩이 대박이 날지 몰랐고, SK...

Hynix probably didn't expect that HBM would become such a big hit. They were caught off guard. So, when the ARI revolution suddenly started, they benefited significantly. In the case of Samsung, they were originally researching and developing HBM, but I heard that they disbanded the HBM team in 2019. Recently, Samsung Electronics clarified that they didn't actually disband the team; they just reduced its size. This meant that they couldn't properly develop in the UK. As a result, several years have passed, and a gap has formed. Now, as they try to catch up, they find that they need more time than they initially thought. HBM didn't attract much interest initially, contrary to what we expected. Hynix didn't anticipate it would grow this much, and if Samsung Electronics is at the point of saying they reduced their team, it indicates a significant shift.

It seems that the technology related to HBM was not considered a promising technology in the past. Back then, AMD decided to create something to use in their gaming graphics processors instead of GDDR, and they just provided the concept. However, they left the manufacturing entirely to memory companies. In 2013, NVIDIA decided to use HBM in their high-performance graphics processors and began mass production of the first generation of HBM. The quantity produced was not as large as expected. Nevertheless, Samsung Electronics and SK Hynix continued to invest in research and development. Samsung actually developed the HBM 2 and 3 generations first. The fourth generation product, HBM...

Since 2013, Samsung has somewhat stepped back, and during that time, SK Hynix has been diligently conducting research and development. The HBM was originally designed for gaming. Now, you mentioned that Samsung had taken a break for a while. What is the current situation? Is the gap still significant, making it difficult to catch up? 

In fact, last year, Samsung Electronics confidently stated that they could catch up quickly, but they ultimately failed to mass-produce. This was announced in September of last year, and despite their promise to start mass production, almost a year has passed, and they still haven't been able to deliver substantial quantities. However, a significant change occurred recently: the head of the DS division was replaced.

So, Vice Chairman Jeon Young-hyun has started reorganizing the company since May, and it seems that the HBM3, which is the fifth-generation product, is currently undergoing successful cold testing. Therefore, it is likely that some supplies will be delivered to NVIDIA around autumn. Additionally, there are separate 8-layer and 12-layer products for HBM3. The 8-layer product is expected to enter mass production this year, while the 12-layer product will likely begin next year. SK Hynix is currently producing and supplying well, and the 12-layer product should be available soon. It is encouraging that our domestic companies are competing for the top positions in the HBM field. However, I also have some thoughts. The market capitalization of companies like Samsung and SK Hynix, which excel in HBM production, compared to that of GPUs and CUDA...

When comparing the market capitalization of companies with ecosystems like NVIDIA, there seems to be a significant difference. While sales were larger than Samsung's, what this means is that in the new industry of AI semiconductors, NVIDIA plays a crucial anchor role. I wonder if it can be expressed as a supplier-buyer relationship, but to secure a position as a leader, South Korean semiconductor companies should not become overly focused on HBM. I think about what strategies and tactics are necessary for the semiconductor sector to lead in the AI era. In fact, Samsung needs to perform well. Samsung must take the lead, although I feel like SK might be disappointed to hear that, as they are doing well too. Professor Chris Miller, the author of Chip War, mentioned that Samsung's crisis is not with HBM but rather with its foundry business. So now, in this context of AI...

The demand for semiconductors is exploding, and right now, TSMC is unable to meet the supply. As a result, people are pulling waiting lists and waiting for a long time, and prices keep rising. Naturally, there are many complaints. So, people are definitely looking for alternatives. If you go to a popular restaurant and the line is too long, you might check out another restaurant nearby, right? But there aren't any alternatives available. 

In fact, when looking at the global landscape, the only companies that can provide advanced foundry processes below 5 nanometers are TSMC and Samsung Foundry. However, rather than narrowing the technology gap between the two companies, it continues to widen. Some companies that initially switched to Samsung after waiting on TSMC's list found that it wasn't working out and returned to TSMC. This situation is unfolding. 

So, what are the reasons Samsung is unable to catch up with TSMC? Well, there are many reasons.

In the end, the most important thing is like a kingpin. I think that’s it. Ultimately, it’s about organizational culture. Samsung, as a company, is very systematic and efficient. In a way, it operates vertically like a military organization, and that worked well during the semiconductor era; it was the most efficient. However, the foundry business is essentially a service industry. We need to cater to what the customers want. We have to try to meet their needs as much as possible. But since Samsung is fundamentally a memory company, it’s not set up for that. So, it’s like a restaurant that serves great soup, but when you ask the grumpy grandmother to sprinkle some chili powder and adjust the serving, she just gives you what she wants.

You need to eat properly. This feeling indicates that the service mindset is lacking, and in fact, there hasn't been training in that area. We need to be able to adjust to those things. In the end, semiconductors are all about IP, which stands for intellectual property. This is what Intel refers to. We create and develop something that aligns with what the customers want, and then IP is generated. When the next customer comes, we can provide that solution as it is. However, in comparison to TSMC, we are only at about level one. TSMC has an IP cell library with over 30,000 entries, while Samsung has only about 7,000. Until now, it seems we have only asked about the hardware side, so I would like to ask a bit about software as well. In the AI value chain, it seems that the software field is also very important, not just hardware. This is not just AI.

소프트웨어라고 그냥 통칭해서 부르는 것이 아니라 여러 층위가 있는 것 같습니다. 대표적으로 어떤 것들이 있나요? 글로벌 빅테크 기업들이 벌이고 있는 경쟁이 있습니다. 그게 바로 파운데이션 AI 모델이라고 합니다. 요즘 우리가 이야기하는 LM, 즉 라지 랭귀지 모델이라고 하는데요. 이 모델은 사실 컴퓨팅 파워를 최대한 많이 키우고, 빅데이터를 최대한 많이 확보하고, 인력을 정말 엄청나게 투입해야 합니다. 이제 AGI 확보를 위한 경쟁이 벌어지고 있는데, 여기에는 현재 오픈AI와 마이크로소프트의 연합군이 있는 ChatGPT가 있습니다. 그리고 여기에 대항하는 것이 구글의 Gemini라는 모델입니다. 최근에는 아마존의 지원을 받아 성장하고 있는 회사가 엔트로픽이라는 회사인데, 그곳은 Claude라는 모델을 가지고 거의 2강 구도를 형성하고 있습니다.

We can consider it to be at a medium level. So, what should companies that cannot enter the AGI competition do? They are developing specialized models, which we refer to as small language models. Additionally, there are application AI models that utilize APIs from language model companies to create these specialized models. Last year, there was a significant event in the industry involving a company called MetaLite, which was originally developing an LML. That model is called the Llama model, and they released it as open source. They encouraged anyone who needed it to take it, modify it, and use it freely. This has ultimately played a significant role in the development of small language models and application AI models, leading to the creation of many specialized models in areas such as healthcare and biotechnology.

There is an L&M model, which can be understood as a software that serves as an engine for artificial intelligence to process and control operations based on the natural language we use in our daily lives, providing services. It could also be a platform or even content itself. In the past, if we wanted to make a machine do something, we had to program it manually, at least using languages like Java or Python. But now, to make a machine work, we can simply use natural language, just like how Iron Man's JARVIS operates. This means that there is no longer a need to learn programming. This leads me to think that if a large company with many smart employees creates this L&M model...

If I want to use it, I can just start a business, but why are all the big tech companies trying to create their own L&M models? In the end, they will become Big Brother, right? Then, if they become monopolies, they will have to recover the massive amounts of money they have invested using that monopoly power. But if they share it, they have to divide it, right? Platform companies, especially those we refer to as big tech, seem to inherently aim for some level of monopoly. So, is it impossible for us to develop an L&M model based on the Korean language independently of big tech companies? Or are all the big tech companies already covering our Korean-based L&M model? They are covering it. In fact, when ChatGPT first came out based on the 3.5 model, experts also mentioned that this is an L&M model based on the Korean language.

It will be very useful. So, I had high expectations for Hyper Clover, a company by Naver. Since then, it has continuously updated from GPT 3.5 to version 4, and even Google is consistently upgrading its versions, performing very well in Korea. This made me realize that it's not easy. So, should we use AI from global big tech companies? This is a complex issue. In the past, during the internet and mobile revolutions, the government's role was more of a supporter, helping companies to thrive and connecting them with overseas opportunities. However, in the case of AI, the government is a player. If we don't engage, it could lead to serious consequences. For instance, if we fall behind in military technology due to a lack of AI, it could be detrimental, and later on, if AI adoption decreases, it could hinder economic development as well.

The growth rate is declining. Yes, that's why the government needs to step in as a player, rather than just providing support. This is where the concept of Sovereign AI comes in. What is Sovereign AI? It means that we must have AI based on our own language. If we don't, we will have to operate our military weapon systems on top of the AI from America's big tech companies. Ah, I see. At that time, Elon Musk became a game changer by providing internet through Starlink during the Russia-Ukraine war. However, the Ukrainian military executed a military operation that Elon opposed, and during that time, he cut off the internet. As a result, the operation failed. Yes, yes. This shows how frightening it is from a national perspective. In other words, we are relying on American companies for our military systems.

Relying on something poses a significant risk. Even if American non-tech companies are leading in developing Korean language-based AI models, we must continue to create our own language-based models in collaboration with the government and businesses. I understand that. However, despite this, it seems that ChatGPT is becoming a common noun. We are in a situation where platforms and solutions for communicating with artificial intelligence based on natural language are being referred to simply as ChatGPT. Are companies like Google and Meta investing heavily to catch up? Yes, that's correct. They are investing a lot. Although ChatGPT initially launched with version 3.5, it provided a lot of back data for development. For this, 15 billion parameters were needed.

Variables, or parameters, were provided, and to implement that, we used 10,000 Nvidia Ampere chips and trained with 5 trillion data sheets. It was mentioned that this required a total investment of about 4 trillion won. Currently, AI's performance continues to improve linearly as computing power and big data are increased. The transformer architecture now has around 500 million to 1 trillion parameters, which means that computing power has significantly increased. Last year, Kyung-kyu, who was the head of Samsung's DS division, mentioned that to implement AGI, he believes a minimum of around 1 trillion parameters of computing power is necessary. This means that even the best-performing companies, like Google, would need to increase their computing power by a thousand times. So, with 1 trillion parameters, what about humans?

It seems like the level of AI is almost unmanageable, and if that's the case, there still appears to be quite a long way to go. If the AI power struggle is completed, what kind of world will we find ourselves in? I'm really curious about that. It will be very efficient and much more convenient. However, the important thing is the gap between those who embrace this technology and those who turn away from it. The disparity will be enormous between countries, between companies, and even more so between individuals. 

So, let me ask a more specific question. Currently, major companies and anchor firms from various countries are engaged in this AI power struggle. If this struggle comes to an end, what will happen next?

I think that, including the L&M model, the hardware aspect is similar. It seems like only a specific company or a couple of companies will remain, while the rest will fade away. Do you think that will happen, or will there be various needs that lead us to insist on using our own technologies for security reasons? It seems like we might end up in a situation where we have to force ourselves to use our own capabilities. 

In my opinion, the United States and its allied countries will mostly use AI provided by American big tech companies. Most countries cannot develop AI on their own. How many countries do you think can develop models like LLM globally? I would say around 20 countries, right? Actually, there are only four: the United States, China, Israel, and... Japan can't do it either, can it? No, Japan can't either. The reason is that they don't have a platform. So recently, we have been focusing on our own line of development.

We brought it to do that, and now, there are only three countries in the world that can design and manufacture chips for this service, including our country. It’s obvious. Therefore, I don’t think it’s so bad for our country to transition to AI quickly. My judgment may be wrong. Our country is not in a bad position in the AI revolution, but the first place is too far ahead. The first place is doing overwhelmingly well, and the second place, China, is also overwhelmingly ahead. Our country is still within the top 12, which is a very good position, but the gap is incredibly large. The outlook is not very good either. Recently, I had high expectations for this; we have quite good big data in the English-speaking world.

There was a line, right? Yes, that's correct. Now that has become uncertain, and that's quite painful. In fact, the computing power, even now, in the case of Naver's Hyper Cloverb X, has about 200 trillion parameters. There, they are trying to increase computing power by securing NVIDIA chips and investing in data centers. If there is a social consensus, the government could also provide support, and that could resolve the issue. However, the fact that the data pipeline has disappeared is very uncertain. In our country, besides Naver, there is also Kakao, which is a leading company in this field. Why has Kakao shown less interest in increasing cloud server computing power? As I mentioned, to work on foundation models, the cost based on GPT 3.5 is around 4 trillion to 5 trillion won.

They have entered the market, but now they are entering even more. In the past, big data was mostly gathered by crawling the web and using it roughly. However, these days, the intellectual property related to data has significantly increased. Recently, Apple has prioritized a payment of 50 million dollars to develop this LM, specifically to media companies and publishers, to address copyright issues. That's right; we can no longer just take what was on the web and use it carelessly like before. This means that development costs are becoming a greater hurdle, and much more money is needed to implement computing power. Kakao seems to have a strategy that focuses on expanding applications rather than just developing foundational AI models, aiming to dominate the service market.

It was different. The reason I asked about Kakao is that, back when it transitioned from web to app, it emerged like a comet and grew into a representative company of our country, right? So, just as it changed from web to app, when the ecosystem shifts from apps to AI, are there opportunities for promising small companies or startups like Kakao to emerge? That’s the most regrettable part. In the past, during the World Wide Web or mobile revolution, large companies took most of the market, but through the trickle-down effect, many promising startups emerged. However, the AI era has a slightly different characteristic: the leaders take it all. The existing representative case is this: Open...

AI is starting to enhance its multi-modal services. Initially, it only provided text, but now it has begun generating images through a tool called DALL-E. By the end of the year, it will also offer video generation services. This has had a significant impact, especially on companies like Adobe, which were previously considered to have very robust services. Now, Adobe is becoming less necessary. Companies that were focused on VFX and advertising engines are also finding themselves in a similar situation. The foundation model doesn't just provide a platform; it absorbs all the services as well. It's true that for small companies to challenge and make an impact in this industry, it is becoming increasingly difficult.

The structure is in a very different context than before. However, as you mentioned earlier, companies like Adora and Gilhwa-seong in the graphics sector are also struggling these days. This indicates that when a new service is launched, the impact is significant. If you, as the representative, were to develop a killer app to prepare for the AI era, what would be the most impactful in the market? Please give me a hint today. Ultimately, I believe that you, Professor, are already doing well. You have an IP that is merely a substitute. In fact, you are very knowledgeable; I recall that you were once referred to as a polymath. I can think of one or two people like that, and one of them is Professor Park Jeong-hoon. You have a unique IP, and with that unique asset, I believe you can survive in the AI era as well.

I think I can survive. AI can change my face to that of a handsome person. With that, I can now utilize it. Having something unique that others cannot possess is an important source of competitiveness. That's what you're saying, right? Now, regarding the current AI boom, not only in semiconductors but also in power supply companies, interest is rising in various sectors, from telecommunications to many other fields. Recently, investment performance has been good in relation to AI. If there are areas we should pay attention to in the future, where do you think they might be? Well, the scenario I envision is this: initially, the biggest shortage is in AI semiconductors. Once the shortage of AI semiconductors is somewhat alleviated, there will be a shortage of data centers. So, data centers...

We have invested tremendously, and once the data center is set up, there will be a shortage of electricity. Therefore, we need to invest in electricity as well. However, a recent variable has emerged. We thought that if we just increased our computing power, we could easily access big data. But the recent news is that the speed of computing power development is so rapid that there is not enough data to learn from. Starting next year, we could face a depletion of data. Consequently, companies involved in big data and data management may gain attention starting next year. If the data centers are fully operational, then the issues will be resolved. However, we also need to invest in telecommunications immediately. Since the release of ChatGPT, global data traffic has increased by 20%. In our country, we don't really feel this because our telecommunications infrastructure is already very well developed.

However, when you go to Europe or places like that, you really can't use your teeth. So, we need to invest in that area. And now, the services that are behind it will inevitably continue to evolve. I see the cycle as being about ten years. It was the same with the hair revolution; it lasted about ten years. So, I think the cycle will continue for around ten years. But the important thing is that if you want to invest, you need to do it before the peak. I think about five years is the best time frame. If you look at smartphones, the stock market peaked around 2014 and 2015, and after that, the smartphone value chain wasn't very interesting. So, if AI has been sustained for about a year and a half, I think we need to work hard for at least three years. These days, the AI field is the hottest, but it's also quite complex. Thanks to Lee Young-soo, the CEO of HSL Partners, we can diagnose this very clearly today.

I think I was able to do it today. Thank you.

거품 꺼진다? AI 혁명은 이제 시작입니다 - 이형수 대표(HSL파트너스)

손경제유니버스 Official 채널｜이진우의 '손에 잡히는 경제' ｜박정호의 '손에 잡히는 경제 플러스'｜MBC 표준FM 95.9MHz

라디오만 25년 했습니다. 유튜브는 몇년이나 하겠습니까?
어차피 수십년은 함께 할 채널, 누구보다 먼저 구독해보세요!


손에잡히는경제

- The fluctuation in semiconductor stock prices is influenced by Trump's statements, excessive investment in big tech, and reduced liquidity from carry trade funds, causing policy uncertainty and fluctuating stock prices.

- The AI revolution's concentration in semiconductor manufacturing is due to the sudden and urgent need for AI semiconductors triggered by OpenAI, leading to supply shortages and significant price movements, with NVIDIA dominating the market due to its high-performance chips and software ecosystem.

- NVIDIA's dominance is attributed to its free CUDA software, accumulation of source code, open source utilization, and its crucial role in providing computing power for quickly creating AI models, ultimately shaping the AI ecosystem based on their hardware.

The Evolution of AI Technology and the Changing Landscape of AI Semiconductors

The speed of technological advancement stagnated for 40 years before the development of AI, which was driven by increased computing power and data accumulation, leading to the rise of machine learning and deep learning.

NVIDIA currently dominates the AI semiconductor market with its chips for data training, but there is a growing focus on cost reduction for inference, leading big tech companies to design their own custom chips.

Jensen Huang, the CEO of NVIDIA, acknowledges that the company's market share is expected to decrease, as other big tech companies like Microsoft and Google are shifting towards using their own custom-designed AI chips.

However, with the projected significant growth of the AI semiconductor market, there is still room for expansion and profitability.

The evolving landscape of AI technology and the increasing investment from big tech companies indicate that the AI market is still in its early stages, with the potential for future profitability similar to the establishment of monopolies in the internet and mobile

Big tech's heavy investment in AI is driven by the goal of achieving Artificial General Intelligence (AGI), which is seen as the next crucial infrastructure for modern civilization.

The ongoing supply shortage in AI technology is still a significant concern, impacting stock prices and global economic growth.

The memory semiconductor market is expected to undergo a transformation in the context of artificial intelligence, moving away from its traditional stronghold position.

The Evolution of Memory Semiconductors and the HBM Market

Memory semiconductors have shifted from being general-purpose to being customized, driven by increasing demand for specialized systems.

This has led to significant changes in the market, with HBM (High Bandwidth Memory) becoming a hot concept in the industry.

SK Hynix has taken the lead in HBM development, while Samsung Electronics initially struggled to catch up, experiencing delays in mass production.

However, with a recent reorganization and successful testing of HBM3, they are expected to start delivering supplies to NVIDIA soon, with SK Hynix also performing well in production.

The Role of NVIDIA and Samsung in the AI Semiconductor Industry

NVIDIA plays a crucial role in the AI semiconductor industry as a key player in the market with larger sales than Samsung.

Samsung's struggle to catch up with TSMC is due to a widening technology gap, organizational culture, and a lack of focus on the service mindset in the foundry business.

The global competition in the AI industry is focused on developing advanced AI models, with smaller companies contributing to the development of specialized models for various applications.

The Significance of L&M Models in Natural Language Processing

L&M models serve as the engine for AI to process natural language, eliminating the need for manual programming.

Big tech companies are striving to create their own L&M models, raising concerns about potential monopolies.

The development of an L&M model based on the Korean language is challenging, with big tech companies like Google and Naver leading the way.

These companies are making significant investments in computing power and big data to improve AI performance, highlighting the growing importance of L&M models in natural language processing.

The Global Impact of the AI Power Struggle

The AI power struggle will result in a significant gap between the adopters and non-adopters of the technology, creating disparities on a global scale.

As the competition resolves, only a few countries will have the capability to develop AI models and manufacture chips, leading to a concentration of power.

The rising costs of AI development, particularly in data gathering and computing power, are creating barriers for smaller companies to emerge, unlike previous technology revolutions.

AI's expansion into image and video generation is disrupting companies like Adobe and those in the VFX and advertising industries.

The AI boom is making it challenging for small companies to compete in the market.

The increasing demand for AI semiconductors, data centers, and telecommunications infrastructure is creating investment opportunities in these areas for the next few years.