Around five in the evening on Tuesday, Lisa Su walked into Samsung Electronics’ offices in the Seocho district of Seoul. The chief executive of AMD was in South Korea to meet the company that supplies the memory inside her AI chips and, by her own account, to talk about something larger.
“Samsung is one of our most strategic partners across DDR4 and DDR5, as well as high-bandwidth memory and a wide range of other technologies,” Su told reporters. “We continue to explore opportunities to work with Samsung across memory and foundry.”
The last word is the one that matters. AMD has long relied on Taiwan Semiconductor Manufacturing Co. to fabricate its flagship processors, and that dependence has become a strategic question as AI demand strains the industry’s most advanced production lines. A foundry relationship with Samsung would give AMD a second source for the silicon it cannot make without a factory, and Su’s public mention of it turned a rumor into something closer to a plan.
The reliance on TSMC is the quiet risk inside AMD’s success. Taiwan’s advanced foundries make the overwhelming share of the world’s leading-edge chips, and every major chip designer, from AMD to Nvidia to Apple, books capacity there years ahead. A second source has been the industry’s unfulfilled wish for a decade, and Samsung’s foundry, with its advanced nodes, is the only credible alternative at that scale.
The relationship already runs deep in memory. Samsung supplies AMD with HBM3E and HBM4, the high-bandwidth memory that sits beside AI accelerators, and AMD has named Samsung its preferred supplier of HBM4 for the Instinct MI455X, its next-generation AI chip. In March the two companies signed an agreement to expand next-generation HBM supply and to explore foundry work for future products.
The memory side is no less strained than the manufacturing side. High-bandwidth memory is the component every AI accelerator now needs in larger quantities, and the two Korean makers, Samsung and SK hynix, control nearly all of it. Su’s visit was as much about locking in HBM supply as about foundry, and the fact that she met both suppliers in one trip shows how much of AMD’s future runs through Seoul.
Su’s trip was built around reinforcing that tie. She met SK hynix, the other Korean memory leader and AMD’s long-time HBM partner, and was scheduled to sit down with Samsung executives. The visit echoed one she made in March, when she toured Samsung’s production lines at its Pyeongtaek campus with Jun Young-hyun, the co-chief executive who runs Samsung’s chip business, and Han Jin-man, who heads its foundry division.
AMD has reason to claim a deep tie to Korean memory. The company worked with SK hynix on the early high-bandwidth-memory designs and brought the first HBM-equipped chips to market, and Su said this week that AMD made HBM happen starting in 2015. That history is part of why Seoul treats her visits as events rather than courtesy calls.
For AMD, the calculus is straightforward. The company’s AI business has grown quickly, and its bottleneck is no longer demand but supply of the two things every accelerator needs: advanced manufacturing and high-bandwidth memory. Spreading orders across more than one source is a way to keep that supply stable, and Su framed South Korea as a partner AMD cannot do without.
For Samsung, an AMD order would mean something it has been chasing. Samsung’s foundry has struggled to match TSMC’s yields on the most advanced nodes, and winning AMD as a customer would lift utilization at its fabrication plants and lend credibility to a business that has lagged the group’s memory division. The prospect of AMD, a leading fabless chip designer, routing production to Samsung is exactly the kind of validation the foundry has been seeking.
Neither side is announcing a contract. Su’s language was careful, and a comment that AMD will continue to explore opportunities is a long way from a purchase order. What she offered Seoul this week was not a deal but a signal that the door is open, and in an industry where capacity is booked years in advance, an open door is a real asset.
The trip also pointed at where AMD thinks the AI market is heading next. The company has talked up physical AI, the chips that run robots and machines outside the data center, and Korean memory and manufacturing sit on that path as squarely as they do on the data-center one. Su’s meetings this week were about the chips of the next two years and the factories that will make them.
AMD’s founder helped design the modern memory chip, and the company has spent decades tied to Korea’s memory makers. What changed this week is the suggestion that the relationship could extend from memory to manufacturing, and that the company that once bet everything on a single supplier is now quietly assembling a second one.
