AMD CEO Visits Samsung for HBM, Samsung Wants Foundry Orders in Return - TechPowerUp

The strategic weight of this move lies not in the HBM purchase itself, but in what it unlocks for Samsung's qualification pathway. SK Hynix has effectively cornered the AI memory market by being first to pass NVIDIA's HBM3E validation. Samsung's HBM3E has repeatedly stumbled on yield and signal integrity, leaving it locked out of the dominant accelerator supply chain. AMD's engagement provides an alternative certification route—once MI-series products qualify Samsung's HBM, the company gains independent shipping credentials that bypass NVIDIA's ecosystem entirely, structurally eroding SK Hynix's near-monopoly. The foundry component is the more consequential piece. AMD's compute dies remain tethered to TSMC's CoWoS packaging in the near term. What Samsung realistically captures is I/O die, chiplet interconnect, or analog module fabrication—a second packaging node within a chiplet architecture. The memory-for-foundry barter creates mutual dependency with artificially elevated exit costs, far more binding than a standard procurement contract. Competitive responses are predictable: SK Hynix will accelerate 16-high HBM4 stacking and deploy aggressive pricing before AMD's MI400 design freeze. TSMC will tighten CoWoS-L allocation, treating AMD's packaging share as strategic leverage. The 12-24 month inflection point is the HBM4 node transition. If AMD's 2026 production run incorporates Samsung HBM4, the global HBM market fractures from a two-player oligopoly into a three-pole structure, permanently diluting SK Hynix's pricing power.