[Chip Talk] "A Single Chip Tailored to Each Customer" ... AI Semiconductor Supremacy Hinges on This

Customer-Tailored Memory Chips, "Custom HBM"

At the recently held "Semicon Korea 2026" at COEX in Gangnam, Seoul, a key term emerged that could determine the future of South Korea's semiconductor industry, the pillar of the national economy: "custom high-bandwidth memory" (cHBM).

HBM is, simply put, a "data-only express superhighway." For artificial intelligence (AI) to process an enormous volume of data, the path that data travels must be wide and fast, and HBM is what plays that role. It is also the flagship product of Samsung Electronics and SK Hynix, the "twin engines of semiconductors" that rank first and second in South Korea's market capitalization.

The fact that the letter "c" (custom) is now placed in front of the word HBM means that, just as the era of ready-made clothing has given way to the era of tailored suits, in semiconductors as well, the ability to custom-build exactly to the specifications of clients (such as Nvidia and Google) has become the decisive battleground. Up to now, it was like mass-producing and selling roads of the same design; from now on, however, the industry is entering an era in which it must respond to each client's different demands, such as "we would like more tunnels" or "we need a road with fewer curves."

Samsung Electronics and SK Hynix have offered different responses to this shift in trend. SK Hynix has declared that it will become the "BTS" of the HBM world. This does not refer to the pop group BTS, but is an acronym for B (Bandwidth), T (Thermal Dissipation), and S (Space Efficiency). Its strategy is to provide customized chips that are either extremely fast, excel at dissipating heat, or have high density, depending on customer preferences.

Samsung Electronics, on the other hand, has signaled a breakthrough beyond the limits of conventional semiconductor design. It has expressed its ambition to develop a "multiplayer" semiconductor that not only supplies the necessary inputs to the graphics processing unit (GPU), which serves as the brain of AI, but also helps the GPU with part of its workload. Samsung Electronics Chief Technology Officer (CTO) Song Jae-hyuk has stated, "We are preparing custom HBM that can secure more bandwidth by taking the lead in adopting die-to-die interface IP," adding, "We are also considering 'Samsung Custom HBM,' in which the base die takes on a certain portion of the tasks currently handled by the GPU."

However, there are not only positive aspects to custom semiconductors. A tailored suit requires a great deal of effort for a single customer, so if that customer cancels the order or goes elsewhere, it becomes difficult to find another buyer for the remaining garment. The same is true for semiconductors. There is a risk of becoming overly dependent on specific customers, and shifting factories from mass production to small-batch, custom production requires considerable time and cost.

This is precisely why many experts unanimously say that "the government must become a strong supporter." In China, the government takes the lead in cultivating talent and providing subsidies to companies in order to foster high-tech industries such as semiconductors. In the United States, big tech companies such as Meta Platforms do not hesitate to offer annual salaries in the tens of billions of won to attract highly capable AI developers.

Yoo Hoejoon, Dean of the KAIST Graduate School of AI Semiconductors, said, "China is the best example of what role a government should play in high-tech industries such as semiconductors," and added, "South Korea, too, must have the government systematically foster professionals and ventures and provide subsidies to markets and companies." He argued that the government must step up with full-scale research and development (R&D) support so that South Korea's talented researchers do not flow overseas.