THE CRUNCH

Lisa Su has visited Samsung's Seoul headquarters to press for additional HBM memory supply for AMD's AI accelerators, less than seven months after the two companies signed a memorandum of understanding covering next-generation memory, joint development and possible foundry work. Samsung has reportedly become a principal supplier of HBM4 for AMD's Instinct MI400-series accelerators, and Su confirmed that AMD has started shipping its MI455 accelerators with 432GB of HBM4 onboard, along with rack-scale Helios AI systems.

The two sides reportedly want different things from the relationship. According to a Korean media report, Samsung hopes to convert its memory supply role into foundry orders for logic chips, while an industry official quoted in the report said AMD sees entrusting foundry work to Samsung as too risky. Neither company has confirmed those claims. Su's own public comments were more diplomatic: she told reporters AMD would "talk about our broader partnership" and continues to explore opportunities with Samsung in both memory and foundry, without confirming any plans to build chips there.

The caution is not hard to explain. AMD already uses Samsung Foundry for niche programmable logic devices, but moving more advanced products to a new contract chipmaker carries real hazards. In its most recent annual report, AMD warned that transitioning production to new manufacturers could mean delayed product introductions, lower yields or poorer performance. As Tom's Hardware notes, Samsung badly needs large customers, and AMD's appetite for HBM gives it leverage in the talks.

Memory urgency is genuine on AMD's side. Su called Samsung Electronics and SK hynix key supply-chain partners, said AI demand is extremely high, and urged suppliers to "build faster, as fast as possible" while AMD works with customers to trim their memory footprint. She added that AMD plans its supply chain three to five years ahead with partners.