TSMC: CoPoS Will Not Replace CoWoS, Improvement Margins on Wafers Are Still Wide

TSMC has commented on the rumors that emerged in recent hours regarding packaging solutions. Despite the race for artificial intelligence significantly boosting the development of new manufacturing processes, the company emphasized that CoPoS will represent a parallel path, but will not replace CoWoS even in the AI accelerator segment. In fact, as reported by Tom's Hardware USA, according to TSMC, the improvement margins for CoWoS are still wide.

According to Kevin Zhang, senior vice president of business development and global sales and deputy co-COO of TSMC, "the panel integration technologies have not yet reached the level of geometric complexity and interconnection density offered by wafer processes." The entire industry, and not just TSMC, actually possesses much more mature production tools in the realm of wafer-level packaging.

The roadmap for CoWoS therefore envisions further evolutions. TSMC claims it can extend this technology to configurations of 14X and even integrate 58 large dies within a single package. These numbers highlight that wafer-based packaging still represents a competitive solution for processors intended for AI workloads.

The main advantage of panel technologies, however, relates to size. The current substrates used by TSMC measure 120 x 150 mm, while future generations of CoWoS will reach 150 x 250 mm. The initial panels will start from 310 x 310 mm, with growth prospects up to 515 x 510 mm and even 750 x 620 mm, dimensions larger than those of a traditional wafer.

For TSMC, therefore, CoPoS represents a complementary path rather than an immediate substitute for CoWoS. The choice between the two technologies will depend on the characteristics of individual products and customer needs, who will be able to adopt the approach most suited to next-generation platforms.

The company expects to complete the first pilot line of CoPoS in June. Generally, two to three years pass between the experimental phase and large-scale production, but the use of new manufacturing tools makes it difficult to make accurate predictions. Ming-Chi Kuo's forecast thus does not seem so implausible, although mass production starting from 2029 appears more likely.