Samsung is reportedly allocating more of its 4nm manufacturing capacity to logic base dies used in HBM4 memory. The shift reflects how the AI accelerator boom is changing the demand for advanced manufacturing capacity. Samsung has already confirmed that its HBM4 uses a 4nm logic base die. The company says its HBM4 entered mass production with 11.7Gbps transfer speed and can reach 13Gbps. It has also described 12-layer products up to 36GB and a path to 16-layer, 48GB versions. The base die is important because it manages communication between the stacked memory layers and the host processor. As accelerator bandwidth rises, the logic underneath the memory becomes a larger part of the performance and power equation.
HBM is consuming more of the semiconductor stack
AI systems are driving demand for memory that can feed processors at high speed. That makes HBM different from ordinary DRAM in both value and manufacturing requirements. Samsung can coordinate memory and foundry operations internally, but it still has to decide how much advanced capacity to devote to HBM versus other products. The reported allocation is So a sign of demand, but it also exposes a possible bottleneck. Advanced logic capacity is scarce, and HBM base dies compete for the same kind of manufacturing resources needed by other high-performance chips. Samsung expects HBM sales to rise sharply in 2026 and is preparing HBM4E and custom products. The company is effectively betting that AI data-centre investment will continue to absorb advanced memory. The next question is whether that demand grows fast enough to justify the capacity shift without creating shortages elsewhere.The HBM4 shift matters because the memory package is becoming a critical component of an AI accelerator rather than a separate commodity. Faster memory can feed compute units with more data, while the logic base die coordinates the interface between the stacked memory and the host system.
Samsung’s position is unusual because it operates both a memory business and a foundry business. That gives it an opportunity to coordinate the two parts of the HBM4 supply chain, although it also creates a capacity-allocation decision. Advanced manufacturing resources can be used for many products, and demand from AI is competing with other high-value applications.
If the reported expansion continues, it would be another sign that AI demand is pulling advanced semiconductor manufacturing toward memory-related components. The outcome will depend on customer qualification and volume, not simply on how much capacity Samsung allocates on paper.
HBM4 also changes the balance between memory suppliers and logic manufacturers. A base die produced on an advanced process is closer to a high-performance processor component than to conventional memory manufacturing.
For AI accelerator customers, that integration can affect power consumption and bandwidth and capacity. It is one reason why HBM qualification takes time and why a technically capable product still needs to pass customer testing before it can become a major revenue source.
Samsung’s expansion plans So point to a broader industry trend: memory is becoming an increasingly strategic part of AI computing. The companies that can provide capacity, bandwidth and reliable yields together will be competing for a growing portion of the AI hardware bill.