"The semiconductor cycle is now in the bottom of the third inning... Regulations must be eased for proper support" [fn Insight]

[Financial News] Park Jea-Gun, distinguished professor in the Department of Convergence Electronics Engineering at Hanyang University, said in a meeting with Nodong-il, chief editorial writer at Financial News, on the 31st, "Semiconductors are not a business of scale, but a business of timing." He added, "A system in which the government simply responds to companies’ requests and then builds infrastructure is always too late. We need to completely shift to a 'first infrastructure' paradigm, where infrastructure is built in advance and companies enter according to market conditions." Having joined Samsung Electronics' Giheung plant in 1985 and spent 40 years in semiconductors, he described the current AI cycle as "the bottom of the third inning, or the top of the fourth inning" in baseball terms.
The interview can be viewed on Financial News' YouTube channel, fn Insight.— There are concerns that semiconductors are a cyclical industry, so excessive expansion could backfire.▲ Historically, semiconductors have gone through ups and downs every three to five years. But the average trend has always been upward, and this AI cycle is much longer than usual. I believe the memory shortage will continue at least until 2030. In the past, downturns came because markets such as the internet and smartphones grew and then became saturated. Smartphones were already approaching saturation before COVID-19. But AI keeps creating new markets. I would even say a downturn may not come at all.— If you compare it to a nine-inning baseball game, what inning are we in now?▲ Around the bottom of the third inning or the top of the fourth. So far, growth has been driven mainly by learning, and inference is only now starting to expand rapidly. On top of that come AI agents and on-device AI. By 2030, physical AI will arrive. Once robots and autonomous vehicles are fully deployed, every robot and car will need an AI accelerator, and inside that, the memory we specialize in will also be needed. By 2040, we will be in the era of artificial general intelligence (AGI), as AI experts call it. If new markets keep opening, the game will never really end.— Is the move by big tech companies to make their own chips with neural processing units (NPU) or application-specific integrated circuits (ASIC) a threat to us?▲ On the contrary, it is an opportunity. Google developed Tensor Processing Units (TPU) to move away from NVIDIA Corporation’s ecosystem, but HBM is used there too. Whether it is a GPU, TPU, or ASIC, nothing works without memory. In fact, as each customer’s needs become more specialized, memory becomes more diverse. That is why products such as Compute Express Link (CXL), which aims to maximize memory utilization, and High Bandwidth Flash (HBF), which stacks NAND flash vertically to improve both capacity and speed, are emerging. In the past, the structure was one in which the CPU or application processor pulled memory along behind it. Now, the era is shifting to one in which memory leads the way.— There are also concerns about the rapid catch-up by ChangXin Memory Technologies (CXMT) in China.▲ To be clear, the technology gap is at least three years, and in HBM it is even wider. If anything, the gap appears to be growing. CXMT’s wafer input volume is around 300,000 wafers per month, which is not small, but most of its products are legacy, general-purpose chips. We do not have the capacity to make all of those general-purpose products. It is actually better for them to handle the commodity side while we focus on high-performance, high-value-added products. That said, to maintain the gap, 3D DRAM is the key. The two-dimensional structure has reached its physical limit, so it must be built upward like an apartment building, and it needs to be applied starting in 2029. If CXMT succeeds in mass production first, the entire memory market could be reshaped. Investment in research and development for 3D DRAM must never be delayed, just as much as investment in capacity.— Do you think the special law for the mega-cluster promoted by the government will be effective?▲ I think it will have a very large effect. In Korea, when we plan a semiconductor complex, we conduct environmental assessments, designate the site as a complex, and acquire the land. Land acquisition alone takes four to five years. It is difficult because we are asking people who are living well to move out, and there are also issues with land speculation and blocking tactics. Only after that do we build infrastructure such as water and power. Because everything is done sequentially, it takes 10 years. In Kumamoto, they prepared the site while conducting the environmental assessment and laid the infrastructure at the same time, finishing in 1 year and 10 months. The core of the special law is to carry out all procedures in parallel and in a one-stop manner.— In the end, the biggest obstacles are water and power.▲ That is right. Yongin needs to draw water from the Han River basin, and power must come from Dangjin, the southwest region, and the East Coast of Korea. Since the route passes near the Seoul metropolitan area, underground cabling is necessary, and issues of local acceptance arise. Fab sites in the southwest region are relatively better. There are plans to raise dam levels in the Yeongsan River basin to secure 600,000 tons of water, and power can come from nearby nuclear plants or, if needed, from liquefied natural gas (LNG) plants built alongside them. Solar power is not suitable for semiconductors. Even a one-second outage can cause enormous losses, so stable, high-quality power is essential. The government’s plan is to bring forward four SK hynix fabs to 2032 and six Samsung Electronics fabs to 2040, but in the end, securing infrastructure is the key.
— You have long pointed out problems in the materials, parts, and equipment industry.▲ It is difficult for the public to fully understand the semiconductor ecosystem. When one fab is built, 2,000 to 3,000 engineers from Samsung Electronics or SK hynix are needed. But if equipment breaks down, it must be fixed immediately, so about 3,000 workers from the Materials, Parts, and Equipment Industry stay on site around the clock. Wherever a fab goes, the ecosystem must follow. The problem is the strength of domestic companies. ASML’s annual revenue is 53 trillion won, Applied Materials’ is 40 trillion won, and Lam Research’s is 26 trillion won. Yet all of Korea’s leading equipment makers combined reach only 7 trillion won. Operating margins are around 8% in Korea, while overseas they are above 25%. Because profits are small, R&D investment is small, and the gap widens.— Even so, why is domestic materials, parts, and equipment production necessary?▲ This is not a matter of slogans about localization. It is a matter of supply chain risk. Samsung Electronics and SK hynix are companies that buy materials, parts, and equipment and then make chips. During COVID-19, and again during Japan’s export restrictions, we saw how badly we were hit. If a country manufactures semiconductors, it must have a domestic ecosystem that minimizes supply chain risk. But if we cannot make world-class products, Samsung Electronics and SK hynix will have no reason to use them. That is why bold support is needed. I proposed creating a 300 billion won "Semiconductor Materials, Parts, and Equipment Ecosystem Innovation Task Force" with 100 billion won each from Samsung Electronics, SK hynix, and the government. Each year, it would identify 10 tasks aimed at world-class performance and develop them over four to five years. If demand companies invest and directly lead the effort, half the battle is already won. Over 10 years, that means 50 successes out of 100. That alone would make a huge contribution to supply chain stability.— Is there enough talent? Some say the southwest region can rely on local universities.▲ It is nowhere near enough. We should assume it is impossible. Each fab needs about 6,000 workers. If you add 10 fabs in Yongin and 4 in the southwest region, that makes 14 fabs. By simple calculation, that is 84,000 people, and in reality nearly 100,000 are needed. The southwest region alone needs more than 25,000. This is not something a few local universities can handle; top talent from across the country must be supplied. The government’s plan to train 100,000 semiconductor workers over 10 years was a well-calculated policy and got the numbers right. But because of AI demand, that timeline now needs to be moved up.— Do you mean numbers alone are not enough?▲ Exactly. What matters now is quality. Semiconductors require world-class science and technology, and they need skilled, high-level talent rather than simple labor. Theory alone is not enough; hands-on training must follow. But there are not many institutions that actually have R&D and manufacturing process equipment. We need to create new intensive practical training programs centered on places that have 12-inch wafer process fabs. Student interest is strong. At our university, the semiconductor engineering department under the SK hynix Department of Semiconductor Engineering had entrance scores this year higher than those for the natural sciences at Seoul National University or medical schools outside the capital region.— You also emphasized regional division of roles.▲ You have to understand the nature of the business. The first priority is timing, and the second is role division. The Seoul metropolitan area should be where R&D, pilot production, and manufacturing are all done together, while the southwest region should become a production base, just like overseas fabs. Some equipment and certain materials need to be located next to fabs for maintenance reasons, but inorganic materials such as ceramics do not need to be in the capital area. For lower costs, the provinces have the advantage. That is why Gumi was designated as a national advanced semiconductor industrial complex, and Busan as a specialized cluster for power semiconductors. Packaging will go to Cheongju and Cheonan-Onyang. Taiwan also has functions divided from the north to the south.— Finally, what should the government keep in mind?▲ Just one thing: semiconductors are a business of timing. Taiwan can build a factory in two and a half to three years, so why does it take us 10 years? And if we only start building after companies ask for it, it is already too late. What the government should do is build the infrastructure first. If the infrastructure is already in place, companies can adjust their investment pace to demand and supply, and that can actually help avoid a downturn. The southwest region fab plan is a shift in thinking in that sense. Completely changing the infrastructure-building paradigm of the past — that is the real content that should be included in the special law.
■ Profile of Professor Park Jea-Gun △ Head of the Semiconductor and Materials Technology Group at Samsung Electronics △ Ph.D. in Materials Engineering, North Carolina State University, USA △ Policy adviser in the economic innovation field for Busan Metropolitan City Government △ President of The Korean Society of Semiconductor & Display Technology

[email protected] Han Young-joon Reporter