Lisa Su paid her second visit to South Korea within seven months, proposing collaborations with Samsung and SK Hynix for "three to five years, or even longer," and publicly disclosed for the first time that SK Hynix will supply HBM4 for AMD's AI accelerators. With the HBM usage per MI455X chip up 50% compared to the previous generation and supply remaining tight, AMD's urgency to secure the Korean supply chain has increased. Meanwhile, Samsung is leveraging HBM supply negotiations to advance foundry partnerships, but AMD remains cautious.
AMD CEO Lisa Su visited South Korea on October 7, holding intensive meetings that day with top executives from Samsung Electronics and SK Hynix. At a media briefing, she unusually put a concrete figure on the framework for cooperation: "Supply chain planning requires three to five years, or even longer, rather than short-term arrangements of one to three years." This was Lisa Su's second visit to South Korea within seven months.
This visit also brought an important 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 that Samsung was the sole source of AMD's HBM4. According to The Elec, Lisa Su stated, "The next-generation MI455 and Helios systems have both begun shipping, all using HBM4, with strong support from both Samsung and SK Hynix."
Lisa Su also emphasized, "HBM collaboration will not be limited to a single product generation, but will continue to advance across multiple generations."
HBM demand surges, supply remains tight.
AMD's latest AI accelerator, the MI455X, features 432 GB of HBM4 per chip, a 50% increase over the previous-generation MI355X's 288 GB of HBM3E. This means that even if shipment volumes remain unchanged, the total HBM capacity required will rise substantially.
Supply-side pressures should not be overlooked either. At the end of last month, market research firm TrendForce predicted that, as HBM shares advanced‑node capacity with conventional DRAM and as ramping up yields and production for new products will take time, memory supply constraints are expected to persist into next year.
Lisa Su stated, "Memory supply is currently very tight. On the one hand, we are working with customers to optimize memory utilization efficiency; on the other hand, we are urging our partners to ramp up production capacity as quickly as possible."
Samsung seeks to upgrade from a memory supplier to a foundry partner.
Samsung's demands in these talks extend beyond HBM supply.
According to reports, Samsung hopes to extend its memory supply relationship to the foundry business. During Lisa Su's first visit to South Korea this March, the two sides signed an MOU at Samsung's Pyeongtaek campus, confirming that Samsung will supply HBM4 for AMD's Instinct MI455X and high-performance DDR5 DRAM for the Helios platform and sixth-generation EPYC server CPUs, while also exploring the possibility of having AMD chips manufactured by Samsung.
At the time, Jun Young-hyun, head of Samsung's DS division, emphasized that Samsung not only offers memory products but also provides foundry and advanced packaging services, enabling the integration of multiple types of chips into a single package.
However, AMD adopts a cautious stance toward foundry partnerships. Currently, AMD's high-end CPUs and GPUs are both manufactured by Taiwan Semiconductor, and the MI455X likewise utilizes Taiwan Semiconductor's process technology. When asked about the progress of its foundry collaboration with Samsung, Lisa Su remained noncommittal: "AMD has grown rapidly in recent years, and having a broad supply chain is crucial. Samsung is an outstanding partner across many areas, and we will continue to explore opportunities for cooperation."
A senior executive in the semiconductor industry told the Chosun Ilbo, "AMD believes that entrusting its foundry business to Samsung carries significant risks. However, to secure a stable supply of HBM, it must accept Samsung's proposal. This divergence in positions is precisely what prompted these talks."
In its annual report last year, AMD also disclosed that shifting certain products to new manufacturers could result in delays in product launches or declines in yield and performance.
The software ecosystem and local deployment are advancing in tandem.
Beyond the hardware supply chain, Lisa Su also focused on advancing AMD's software ecosystem in South Korea during this trip.
That morning, Lisa Su met with Ryu Je-myeong, Vice Minister of South Korea's Ministry of Science and ICT, at the Four Seasons Hotel in Seoul. She then held a roundtable discussion with 13 Korean AI startups, including FuriosaAI, DeepX, Fadu, and Upstage, focusing on integrating local hardware and software developers into AMD's open-source ROCm platform—a tool widely regarded as a core contender to NVIDIA's CUDA ecosystem.
AMD also plans to establish an "AI Center of Excellence" in South Korea, with an anticipated recruitment of several hundred AI researchers. This initiative stems from a memorandum of understanding signed in July this year between AMD and South Korea's Ministry of Science and ICT.
Lisa Su was originally scheduled to stay in South Korea for two days, but due to a last-minute change in her Washington itinerary, her visit was shortened to one day. Jack Huynh, AMD's Senior Vice President of Computing and Graphics, will remain an extra day to conduct on-site inspections of domestic startup technologies, including the optimization techniques employed by the startup Nota to efficiently run AI models on AMD's Halo platform. Jack Huynh stated, "We work very closely with startups, and I'm very interested in seeing this firsthand."
Editor/KOKO