Samsung co-CEO Jun Young-hyun described the company’s foundry business as “primed for a great leap forward” in his New Year address, citing recent supply deals with major global customers. The statement attracted less attention than the HBM4 headlines, but the foundry dimension of Samsung’s semiconductor strategy may prove equally consequential for the company’s competitive position and for the broader AI hardware supply chain. Samsung’s ability to fabricate the 4-nanometer logic dies that sit at the base of its HBM4 stacks in-house, rather than sourcing them from TSMC, creates a vertically integrated production model that provides structural advantages in cost, scheduling, and time-to-market.
The logic die in an HBM stack is a sophisticated piece of silicon that manages data flow between the DRAM layers and the external interface. It handles error correction, power management, thermal monitoring, and the signaling protocol that connects the memory stack to the AI processor. In HBM4, the transition to a 2,048-bit interface doubles the complexity of the logic die relative to HBM3E, requiring a more advanced process node to meet the performance and power specifications. Samsung fabricates this die on its own 4-nanometer GAA process, a technology that the company developed for its foundry customers and adapted for internal use.
SK Hynix, by contrast, depends on TSMC for its HBM logic dies. This dependency creates a supply chain linkage that adds lead time, introduces allocation risk, and requires coordination between two organizations with different production schedules and priorities. TSMC’s logic die production for SK Hynix competes for wafer starts with TSMC’s foundry customers, including Nvidia, AMD, Apple, and Qualcomm, meaning that SK Hynix’s HBM production schedule is indirectly affected by the demand patterns of companies with no relationship to the memory market. Samsung’s self-sufficiency eliminates this dependency, allowing the company to align logic die and DRAM production schedules without external coordination.
The cost advantage is significant. When Samsung fabricates its own logic die, the transfer price is an internal accounting entry rather than a market transaction. TSMC charges market rates for its foundry services, which include margin that reflects the company’s pricing power and the opportunity cost of allocating capacity to HBM logic dies rather than to higher-revenue chip designs. Samsung’s internal fabrication cost for the same die is lower by the amount of TSMC’s margin, a difference that flows directly to Samsung’s HBM profitability. At HBM4 prices near $700 per unit, the logic die represents a meaningful fraction of total production cost, and the internal sourcing advantage translates into a margin benefit of several percentage points.
The foundry technology development also creates spillover benefits. The 4-nanometer GAA process that Samsung uses for HBM logic dies is the same technology it offers to external foundry customers. Improvements in yield, reliability, and performance that Samsung achieves through the high-volume production of HBM logic dies feed back into the foundry’s commercial offerings, making the process more attractive to external customers. The HBM program, in this sense, functions as a captive customer for Samsung’s foundry that provides the production volume needed to mature a process technology that can then be marketed globally.
For investors, Samsung’s foundry-HBM integration represents a competitive advantage that the market has not fully valued. The analysis of Samsung’s HBM business typically focuses on the memory side: DRAM pricing, HBM market share, and qualification progress with Nvidia. The foundry dimension, which affects production cost, schedule control, and technology development, is a less visible but equally important component of Samsung’s competitive position. As HBM stacks become more complex and the logic die becomes a larger share of the total stack’s value, the advantage of in-house fabrication will grow proportionately. Investors evaluating Samsung should model the foundry contribution to HBM margins explicitly rather than treating it as an undifferentiated component of the memory business.
