AMD CEO Lisa Su traveled to South Korea on October 7, holding a packed schedule of meetings with top executives from Samsung Electronics Co., Ltd. (SSNLF) and SK Hynix, and at a media briefing she unusually quantified the partnership framework in explicit terms — "supply chain planning needs to span three to five years or even longer, rather than a short-term arrangement of one to three years." It marked Su's second visit to South Korea within seven months.

The trip also delivered a significant disclosure: AMD publicly confirmed for the first time that SK Hynix will supply HBM4 for its AI accelerators and server platforms. Previously, the market widely believed Samsung Electronics Co., Ltd. (SSNLF) was AMD's sole source for HBM4. According to The Elec, Su said, "The next-generation MI455 and Helios systems have both begun shipping, all using HBM4, and both Samsung Electronics Co., Ltd. (SSNLF) and SK Hynix have provided strong support."

Su also stressed that "HBM cooperation will not stop at a single product generation, but will continue across multiple generations."

HBM demand surges and supply tightens

AMD's latest AI accelerator, the MI455X, carries 432GB of HBM4 per unit, a 50% increase over the previous-generation MI355X's 288GB of HBM3E. That means even if shipment volumes remain unchanged, the total HBM capacity required rises substantially.

Pressure on the supply side cannot be overlooked either. Market research firm TrendForce predicted late last month that HBM and ordinary DRAM share advanced process capacity, and combined with the time needed for new products to ramp yields and output, memory supply tightness is expected to persist into next year.

Su commented on this: "Memory supply is currently very tight. On one hand, we are working with customers to optimize memory usage efficiency; on the other, we are asking partners to expand capacity as quickly as possible."

Samsung seeks to upgrade from memory supplier to foundry partner

Samsung Electronics Co., Ltd. (SSNLF)'s ambitions in these talks go beyond HBM supply.

According to reports, Samsung Electronics Co., Ltd. (SSNLF) hopes to extend the memory supply relationship into its foundry business. During Su's first visit to South Korea in March this year, the two sides signed an MOU at Samsung's Pyeongtaek campus, confirming that Samsung Electronics Co., Ltd. (SSNLF) would supply HBM4 for AMD's Instinct MI455X, provide high-performance DDR5 DRAM for the Helios platform and sixth-generation EPYC server CPUs, and explore the possibility of Samsung manufacturing AMD chips on a foundry basis.

Jun Young-hyun, head of Samsung Electronics Co., Ltd. (SSNLF)'s DS division, emphasized at the time that Samsung can offer not only memory but also foundry and advanced packaging services, integrating multiple chips into one.

However, AMD remains cautious on foundry cooperation. AMD currently has its high-end CPUs and GPUs produced by TSMC, and the MI455X also uses TSMC's process. When asked about progress on foundry cooperation with Samsung Electronics Co., Ltd. (SSNLF), Su chose measured words: "AMD has grown quickly in recent years, and having a broad supply chain is very important. Samsung Electronics Co., Ltd. (SSNLF) is an excellent partner in many areas, and we will continue to seek cooperation opportunities."

A senior semiconductor industry figure told Chosun Ilbo: "AMD believes handing foundry business to Samsung Electronics Co., Ltd. (SSNLF) carries significant risk. But to secure HBM stably, it must respond to Samsung's proposal. There is a temperature gap between the two sides, and that is exactly why this meeting was arranged."

AMD also disclosed in last year's annual report that transferring some products to a new manufacturer could lead to product launch delays or declines in yield and performance.

Software ecosystem and local footprint advance in tandem

Beyond the hardware supply chain, Su's trip also focused on advancing AMD's software ecosystem buildout in South Korea.

On the morning of that day, Su met with Ryu Je-myeong, second vice minister of South Korea's Ministry of Science and ICT, at the Four Seasons Hotel Seoul, and then held a roundtable with 13 South Korean AI startups including FuriosaAI, DeepX, Fadu and Upstage, focusing on connecting local hardware and software developers to AMD's open-source ROCm platform — a platform seen as a core tool for challenging Nvidia's CUDA ecosystem.

AMD also plans to establish an "AI Center of Excellence" in South Korea, expected to recruit hundreds of AI researchers. The plan stems from a memorandum of understanding AMD signed with South Korea's Ministry of Science and ICT in July this year.

Su had originally planned to stay in South Korea for two days, but due to a last-minute scheduling conflict in Washington, the visit was shortened to one day. Jack Huynh, AMD's senior vice president of computing and graphics, stayed an extra day to examine domestic startup technologies on the ground, including startup Nota's optimization technology for running AI models efficiently on AMD's Halo platform. Jack Huynh said: "We will work very closely with startups. I am very interested and want to see it with my own eyes."