Nvidia Reportedly Seeks TSMC 3nm Base Dies for Custom AI Memory

Key Takeaways

• Nvidia has approached Samsung, SK Hynix, and Micron to utilize TSMC’s 3nm process for the base dies of its custom NVHBM memory.
• The NVHBM architecture aims to enhance bandwidth, lower power consumption, and streamline integration across different suppliers.
• This move could broaden TSMC's involvement in AI memory production while imposing new qualification standards for memory manufacturers.

Nvidia's Reported Request Puts AI Memory Supply in Focus

Nvidia is reportedly pursuing a unified manufacturing strategy for its next-generation custom high-bandwidth memory (NVHBM). The company has requested that major memory suppliers—Samsung Electronics, SK Hynix, and Micron Technology—adopt TSMC’s 3nm process for the base dies of NVHBM. This strategy aims to facilitate easier integration of memory from various manufacturers into Nvidia's technology ecosystem.

While Nvidia has publicly introduced NVHBM and its architecture, the specific requirement for all three suppliers to use TSMC’s 3nm process remains unverified through official announcements.

What NVHBM Changes Inside AI Chips

The NVHBM architecture integrates Nvidia’s custom memory controller into the base die of the memory stack, which is positioned beneath the memory layers. This design allows for closer proximity between the memory and the AI processor, potentially leading to a 30% increase in memory bandwidth and a 15% reduction in power consumption compared to standard HBM4E.

This architecture is part of Nvidia’s NVLink Fusion offering, which is crucial for customers developing custom AI infrastructure, thus extending its significance beyond just memory procurement.

Why Standardisation Matters for Nvidia

A standardized manufacturing approach could enhance Nvidia's efficiency in managing multiple memory suppliers. Variations among suppliers often necessitate additional integration and testing, so reducing these differences could simplify the qualification of alternative sources and future product planning.

Commercially, this could bolster Nvidia's negotiating power regarding capacity and delivery schedules. However, a shared specification does not imply that all suppliers will be interchangeable, as each must still meet performance and reliability standards.

Samsung, SK Hynix, and Micron Face Different Implications

For Samsung, this arrangement could be strategically significant as it operates both memory and foundry businesses. Utilizing TSMC for NVHBM base dies could enhance Samsung's competitiveness in high-end memory while outsourcing part of the manufacturing process.

SK Hynix already has a partnership with TSMC for HBM4 development, which could facilitate collaboration on NVHBM base dies. Micron's involvement could provide access to the growing demand for custom AI memory, contingent on successful qualification and scalable production.

TSMC Could Gain a Broader Role in AI Memory

The proposal underscores the increasing importance of advanced logic manufacturing in memory products. If multiple memory suppliers adopt TSMC’s foundry process, it could expand TSMC's role in AI infrastructure beyond just processor fabrication.

However, the report lacks details on wafer volumes, contract values, and customer allocations, which are essential for assessing TSMC's potential revenue impact.

What Investors Should Watch Next

Key developments to monitor include official confirmations from suppliers, disclosures regarding the manufacturing process, and evidence of product qualification. Progress through various production stages—sampling, qualification, and volume manufacturing—will be critical in determining readiness for market introduction.

Investors should also keep an eye on any announcements regarding committed capacity, shipment schedules, or changes in capital spending related to custom memory.