Taiwan and South Korea Make AI Chips Together. Who Designs the Next One?

Imagine a shipping order and a blank drawing sheet on the same table. The order lists quantities, deadlines and the parts each company will deliver. The blank sheet is harder. When a customer asks for a faster machine that uses no more power, who gets to pick up the pencil and propose a design?

What came back was a team that could build

In 1974, Sun Yun-suan, then Taiwan's minister of economic affairs, discussed developing integrated circuits with a group of officials, engineers and researchers. In 1976, Taiwan's Industrial Technology Research Institute signed a technology transfer agreement with RCA and sent engineers to learn design, manufacturing, testing and equipment. Back in Taiwan, they put those skills to work together in a demonstration factory. The question I take from that episode is what happens after someone hands over the technology. Has the team built anything with it? Can they fix what goes wrong? Will they have somewhere to run the next experiment? ITRI's history of Taiwan's semiconductor industry

Taiwan can now help others solve manufacturing problems. That changes the question. When customers are still deciding what problem they want solved, who gets a seat at the table?

One chip stack, two teams changing the drawings

SK hynix and TSMC announced plans in 2024 to collaborate on HBM4 and next-generation packaging. High-bandwidth memory stacks several layers of memory to move data quickly to the processor. A logic base die at the bottom routes signals. SK hynix said it would use a TSMC advanced logic process for that base die starting with HBM4 and work with TSMC to integrate memory and logic through advanced packaging. That is a division of work between two companies. SK hynix still makes the memory, and the design of the full AI accelerator also involves the customer. SK hynix partnership announcement · SK hynix at the 2026 TSMC symposium

Suppose a customer needs data to move faster but cannot make the package any larger. The Korean memory team may need to change how signals travel, and the Taiwanese logic and packaging teams may need to redo the layout. Taiwan has a chance to suggest a design that will be easier to make at scale while the drawings are still open. That takes engineers and testing resources up front, some of which may have to be spent again if the specifications change. South Korea has a chance to turn memory into something customers plan around early. Once an outside partner handles the base process or packaging, that partner also has a say in the schedule and how a redesign happens.

There are already shipments to discuss. TSMC reported strong demand for leading-edge processes in the third quarter of 2026, and SK hynix said its HBM4 shipments began in earnest that quarter. Those announcements describe demand and shipment progress. They leave open which approach particular customers chose and who could see which test results. When the next redesign comes, the lasting value will depend on whether engineers can diagnose failures and carry what they learned about manufacturing into another project while protecting customer secrets. TSMC third-quarter results · SK hynix third-quarter results

Partners on one generation, rivals on the next

South Korea also has another route. In February 2026, Samsung announced commercial HBM4 shipments using its own memory and a four-nanometre logic base die. It has since promoted a package of services that combines memory, foundry and packaging. Some customers may save coordination time by buying several services from one company. Others may want to choose the strongest partner for each part. Both approaches still depend on qualification and delivery. Taiwanese and Korean companies can complement each other on one project and compete for the same customer's foundry or packaging work on another. Samsung's HBM4 shipment announcement · Samsung's integrated manufacturing strategy

In 2027, I would watch who gets invited while a new project's specifications can still change. I would also watch what happens when a design fails a test. Who can see enough of the data to explain what went wrong, propose a different approach and persuade the customer to try again? If the next order is half as big, some of the boom will have lasted if teams can still use what they learned for a different customer or product.

If you were the supplier, would you take one more guaranteed order this year, or give up some volume for a role in designing the next generation?

Listen to the conversation

Listen to Who Designs the Next Generation?. In this Mandarin episode, M.K., Dorami and a dramatized Sun Yun-suan ask what Taiwan and South Korea learn from working on AI chips together.