Taiwan Semiconductor Manufacturing Co. is weighing entry into 3D DRAM production at Nvidia's request, according to sources familiar with the matter, signaling how acute memory bottlenecks have become for AI chip makers.
Nvidia's data center GPUs require high-bandwidth memory (HBM) in volumes that memory suppliers Samsung, SK Hynix and Micron cannot reliably meet. The shortfall has directly constrained Nvidia's ability to ramp production of its most profitable accelerators.
For TSMC, the move would mark a fundamental departure from its 27-year foundry model. The company has built its $70 billion revenue base on pure-play contract manufacturing of logic chips—not on competing in the capital-punishing memory business, where Samsung and SK Hynix spend $10 billion-plus annually on fab construction and yield optimization.
Memory fabrication requires entirely different process flows, equipment roadmaps and operational complexity than logic. A dedicated 3D DRAM fab would demand billions in upfront capex and tie up manufacturing capacity that TSMC can currently monetize at higher margins through advanced logic work for Apple, AMD and others.
The economic calculus hinges on whether TSMC can charge a memory manufacturing premium sufficient to justify the capital intensity and execution risk. If Nvidia locks in a long-term offtake agreement at favorable pricing, the venture becomes more attractive. If TSMC must compete against entrenched memory players on cost, the business case deteriorates quickly.
No final decision has been made. TSMC declined to comment on the discussions.
