THE CRUNCH
PieceMakers, a Nanya-backed DRAM designer, has listed on Taiwan’s Emerging Stock Board with a strategy to supply custom-designed memory that fuses DRAM stacks directly to processors using hybrid bonding. The company argues that edge AI inference memory diverges from the high-bandwidth memory used in data centres, positioning its products between Nvidia’s SRAM-only Groq LPU and traditional HBM solutions. Revenue is to
come from design fees rather than chip sales, with AI custom design work accounting for around 40% of revenue in the first half of 2026. The company’s first volume customer program is not expected to contribute to financials until 2027 at the earliest. PieceMakers is developing HiBaLL, a 3D-stacked memory rated at more than 1 TB/s, and HBLL, a 2D die rated at 144 GB/s. The strategy aligns with Nanya’s broader plan to focus on custom, edge-oriented memory rather than competing in HBM3 or HBM3E markets.
Samsung has outlined a roadmap for zHBM, which stacks DRAM directly on top of the processor using wafer-on-wafer hybrid copper bonding, projecting roughly 2.3x the bandwidth of a four-stack HBM4E system with about 70% less I/O power. SK hynix has indicated that hybrid bonding will not be ready for HBM4E, leaving HBM5 as the earliest point for such technology. This suggests that direct stacking may become a key differentiator for future AI inference accelerators.