Taiwan Semiconductor Manufacturing Co. is exploring plans to manufacture advanced 3D memory chips, including High Bandwidth Memory (HBM), at Nvidia's request, according to people familiar with the matter.
The move would pit TSMC against Samsung, SK Hynix and Micron in a segment TSMC has historically avoided. Those three companies currently dominate HBM production, which supplies memory for advanced AI GPUs.
HBM stacks memory vertically to increase bandwidth and cut power consumption—critical specifications for AI accelerators. The chips have become essential as Nvidia and other AI chip designers push toward denser, more efficient systems.
TSMC's push into memory represents a sharp departure from its pure-play foundry model, where it manufactures logic chips for clients like Apple and Nvidia. Memory production involves different fab processes, yield curves and capital intensity than logic—and would require substantial investment in new equipment and IP.
But Nvidia's request signals a strategic calculation: diversifying HBM suppliers reduces its dependence on Samsung and SK Hynix, both of which make competing GPUs and can use supply leverage as a competitive weapon. For Nvidia, TSMC's involvement adds optionality during a period of surging AI chip demand when memory has become a bottleneck.
For TSMC, the opportunity extends beyond contract manufacturing fees. Memory is a higher-margin, more consolidated business than logic foundry work. Breaking in requires overcoming entrenched players with existing relationships and proven yields—a costly undertaking with no guarantee of meaningful market share.
