The Unsung Heroes of AI Chips: MLCCs and Substrates

MLCCs and substrates are almost entirely monopolized by Japanese companies. However, a unique ceramic formulation technology known only to Samsung Electro-Mechanics and Murata determines competitiveness in this market. MLCCs must have a high dielectric constant while being made thin, and many layers must be stackable on top. The ceramic formulation that simultaneously meets all these conditions is the core technology that Samsung Electro-Mechanics and Murata have protected.

Samsung Electro-Mechanics already has factories in multiple regions worldwide and is expanding them, anticipating a significant increase in demand. It has a factory in the Philippines, a substrate factory in Sejong where expansion work is underway, and factories in Mexico.

In the past, MLCCs were primarily used in mobile phones, but recently with AI, they are being sold in enormous quantities to data centers where AI products are deployed. Nvidia's AI servers require 2-3 times more MLCC units and capacity each year, continuously driving up demand. MLCCs will also be incorporated into products like autonomous vehicles and humanoid robots. MLCCs are used more as power consumption increases, which is exactly why demand is surging in the AI era.

Copper posts and glass substrates are part of FC-BGA. As AI advances, more and more semiconductors are mounted on substrates. Recently, AI products have 4-5 SOCs (System-on-Chips) and 8-10 HBMs (high-bandwidth memories) mounted on substrates.

In the past, substrates the size of a thumbnail were common. Recently, substrates the size of a palm have become the norm. In area, this represents about 100 times increase. Wiring layers have also increased significantly to accommodate more semiconductors.

FC-BGA size is approximately 100mm × 100mm with 20-30 layers. In contrast, FC-CSP (flip-chip chip scale package) is about 10mm × 10mm or 15mm × 15mm with only 2-3 wiring layers. Therefore, FC-BGA has 100 times the area and 10 times the layers, requiring 100 × 10 = 1,000 times more processes, and prices are proportionally expensive.

Japan's Ibiden and others are ahead, with Samsung Electro-Mechanics representing Korea among leading companies.

FC-BGA has a glass core in the center, with white cross-sections visible above and below—these are thin films that assist stacking. Copper wiring is arranged on top, totaling about 20 layers. Value-added is much higher than process count.

FC-CSP mounts 1 semiconductor chip. However, recent FC-BGA used in AI semiconductors typically mount about 15 chips. These many chips must be connected through complex wiring, requiring high-speed performance.

Globally, only 2-3 companies possess equivalent mass production or development technology to Samsung Electro-Mechanics.

For a long time, MLCC and substrate markets were dominated by Japan globally. However, a new competitive arena emerged in the AI era, and Samsung Electro-Mechanics has risen prominently.

When growth potential emerges, it enhances technology development and execution while investing boldly. This is the same strategy that made Samsung Electronics grow.

No company globally can support all products the market demands. Japanese companies and Samsung Electro-Mechanics must focus on high-end. This leaves low-end products that Samsung Electro-Mechanics doesn't touch, and while unprofitable, China sees this as opportunity and aggressively tries to capture it. In slowly evolving product areas, China has monopolized markets in numerous cases. Samsung Electro-Mechanics seeks strategies that avoid low-cost investment while securing profitability.

China also secures significant profits from these budget products. There's growing strategic need to defend the low-cost market.

From his position selling products and talking with customers, Park Chul-min assessed that AI market activity remains truly vigorous. Among those actually executing, concerns about business failure from not investing far outweigh fears of inventory buildup or investment failure. Microsoft and Google emphasized in investor relations that the risk of not investing exceeds the risk of investing, reflecting field evaluation of questions about Nvidia's funding sources.

AI chip competition is evolving from GPUs through packaging to ecosystem competition. For Korea to win, strategy must be clear. Specifically, Korea must define what it should do, how to advance the ecosystem, and what our advantages are. This must be recognized as a uniquely Korean domain and strategy must be built accordingly.

Data centers appear to operate as an unrelenting force. The data center concept emerged over 10 years ago, initially thought to serve only simple server functions. But currently, data center scale and innovations they generate operate as a ceaseless active volcano. Data centers are expected to remain the greatest driver of the electronic components market ecosystem in the near term.

While Park Chul-min doesn't know what new applications physical AI will bring, he hopes it will create differences in our lifestyle and expects physical AI to deliver that.