TSMC's denial and the sources of the rumor
TSMC has denied the report that it would be working on three-dimensional DRAM memory at the request of customer NVIDIA. Chinese technology newspaper IT之家 reports that the company responded to Tom's Hardware by asserting its identity as a provider of foundry services for specialized integrated circuits, with the role of being a reliable partner to the global logic industry in terms of technology and production capacity.
The news had been reported by The Information, which cited two people familiar with the project. According to this version, TSMC would have started plans to produce 3D DRAM after a request from its largest customer, NVIDIA, with supervision attributed to Vice President Hsu Kuo-jin, head of advanced packaging and formerly head of Micron's Taiwan operations. These are claims reported by a source that TSMC has not confirmed: the denial concerns the project as a whole, not individual details.
Why memory weighs on AI and why 3D DRAM is not a quick solution
The context described by IT之家 is a global memory shortage that, with growing demand for artificial intelligence hardware, is pushing prices upward. Memory, the article says, is shifting from a standard product to a strategic resource, capable of affecting the timeliness of deliveries to customers. This is the reason why a TSMC move on the DRAM front would carry considerable weight in the AI supply chain.
However, the material reiterates that TSMC does not have the expertise and production capacity needed to manufacture DRAM, much less a three-dimensional version, although it collaborates with memory manufacturers and integrates memory chips into system-in-package solutions for customers. 3D DRAM, which stacks cells vertically to increase density per wafer, is still at an early stage: according to Micron's roadmap, cited in the reconstruction, at least four generations of planar DRAM will be needed before reaching it, and even the world's three major manufacturers have projects underway but no mass production. Hence an explicitly stated consequence: even if TSMC proceeded, the move would not alleviate the current shortage for at least five years.
The same material lists the difficulties of a possible entry: without a competitive base in planar DRAM, it is almost impossible to develop a mature 3D process, and the alternative of acquiring a small manufacturer with its own fab would still require billions in investment and years of research. Added to this are the risks of the memory cycle, with possible overcapacity and price declines, and complications in relations with the major manufacturers, particularly SK hynix, which in 2024 entered into a collaboration with TSMC to produce base logic chips for future HBM. The text also mentions possible antitrust reviews for vertical integration among logic, packaging, and memory.
An alternative path indicated by the source does not require DRAM manufacturing: producing base logic chips for HBM, integrating memory and processor with SoIC hybrid bonding technology, and collaborating with GUC and memory manufacturers on customized solutions. The existence and status of the project reported by The Information, which TSMC denies, remain to be verified.
