Micron has officially confirmed that TSMC will manufacture the base dies for its next-generation high-bandwidth memory (HBM). With this development, all three of the world’s major memory makers — SK Hynix, Samsung Electronics, and Micron — have now brought TSMC into their HBM base die supply chains. This signals a rapid restructuring of HBM component production away from memory makers’ in-house processes toward specialized foundry partnerships.

Taiwan’s DIGITIMES reported Micron’s decision on September 16. SK Hynix had already committed to outsourcing its base dies to TSMC’s 12nm-class process starting with the HBM4 generation, while Samsung Electronics announced a so-called “dual track” strategy on September 15 that uses both its own foundry and TSMC in parallel. With Micron now joining, HBM base die procurement has effectively shifted from a single-source model to a multi-source competitive landscape.

The report noted that this shift does not simply mean cost reduction. A second source does not automatically guarantee lower prices, and assumptions that were once taken for granted are now being challenged. In fact, TSMC’s HBM4 base dies are reportedly three to four times more expensive than the core chips SK Hynix produces on its in-house 1b process. Despite this, customers are willingly paying a premium for higher performance and customized features.

CompanyBase Die ProcurementTimelineSK HynixOutsourced to TSMC, 12nm-class processStarting with HBM4 generationSamsung ElectronicsDual track: in-house foundry + TSMCCustom HBM, announced September 15MicronOutsourced to TSMCStarting with HBM4E, mass production targeted for 2027

The remaining components — the DRAM core dies — are still manufactured in-house by each memory maker. Micron held a groundbreaking ceremony for a new cleanroom at its Hiroshima plant in Japan on July 4, 2026. The first phase covers approximately 28,000 square meters. According to Micron, Hiroshima has been the hub leading 1β DRAM production and the introduction of EUV for 1γ DRAM, and the local team is participating in the development of next-generation AI DRAM and HBM. With base dies outsourced to TSMC, HBM4E-generation stacks will combine Taiwan-made logic dies with Japan-made DRAM dies.

The spread of foundry outsourcing for base dies is driven by the direction of HBM evolution. As HBM moves from standardized products to customer-specific solutions like Nvidia’s NVHBM, the logic functions that must be integrated into the base die have expanded dramatically. The HBM4 standard (JESD270-4) finalized by JEDEC in April 2025 doubles the interface width from HBM3’s 1,024 bits to 2,048 bits, expands the channel count to 32, and enables bandwidth of up to 2TB/s per stack. This means the control and interface logic the base die must handle has grown accordingly. NVHBM, unveiled on August 26, 2026, places the memory controller inside the base die rather than on the XPU die. Nvidia says NVHBM delivers up to 30% higher bandwidth, 15% lower HBM power consumption, and frees up 25% of die area on the XPU compute die compared to standard HBM4E. Amazon’s Annapurna Labs is the first partner, with Trainium4 supporting NVLink Fusion. This structure favors foundries with logic-process strengths over traditional DRAM processes. TSMC is riding this wave to extend its influence beyond conventional logic chip contract manufacturing into memory-based components. HBM is becoming a new growth curve for TSMC.

This supply chain shift coincides with expectations that AI memory demand will remain robust for the foreseeable future. Citi projects global HBM bit demand will grow 62% next year from this year’s level to reach 75.2 billion Gb. By 2028, demand is expected to climb to 127 billion Gb, or 2.8 times this year’s figure. Citi believes that as “continual learning” — where AI continuously learns from new data — gains traction, memory supply shortages could persist through 2031.

Continual learning is an approach where AI learns new information during service operations and integrates it with existing knowledge. This drives simultaneous demand growth not only for HBM but also for server DRAM and enterprise SSDs. Since not all data can be stored in expensive HBM, capacity expansion across the entire memory hierarchy becomes necessary. Citi explained, “Despite memory spec downgrades and recent concerns about AI safety, continual learning in AI will reshape long-term memory demand.”

KB Securities also diagnosed that the center of gravity in AI infrastructure is shifting from GPUs to memory. Kim Dong-won, head of research at KB Securities, identified the key message from the “AI Infrastructure Summit 2026” held in Santa Clara, California from September 15 to 17: “AI data center construction bottlenecks are rapidly spreading beyond simple GPU procurement to memory, networking, storage, and power across the entire AI infrastructure stack.”

He relayed that Intel’s CEO emphasized that the current memory supply shortage is not a temporary phenomenon and could intensify further next year. The analysis is that HBM capacity expansion for AI is cannibalizing general-purpose DRAM production capacity, meaning AI demand growth constrains overall DRAM supply capacity beyond just HBM.

Samsung Electronics and SK Hynix have also unveiled next-generation technologies to respond to these structural changes. Samsung Electronics presented “zHBM,” a three-dimensional stacking architecture that places HBM directly on top of AI accelerators. The goal is to break away from the existing 2.5D structure and reduce data travel distances to boost AI response speeds by 10 times. SK Hynix introduced PIM (Processing-In-Memory), which performs computation inside the memory, and SALT-KV technology, which optimally places data across HBM, DRAM, and SSD tiers based on data characteristics.

KB Securities expects Samsung Electronics to see expanded strategic opportunities in the memory-centric AI infrastructure restructuring, given its one-stop solution capabilities spanning base dies, interposers, advanced packaging, and HBM. Samsung Electronics held a 33% share of the HBM market in the second quarter of this year, and KB Securities projects that share could approach 40% in the fourth quarter, driven by expanded HBM4 shipments.

The spread of foundry outsourcing for HBM base dies and the structural growth in AI memory demand are mutually reinforcing trends. As memory makers leverage TSMC’s logic processes to build more complex customized HBM, AI systems become increasingly memory-dependent. In this process, TSMC is positioning itself not merely as a chip manufacturer but as a core pillar of the AI memory supply chain.