台积电，颠覆封装？

Core Viewpoint - The article discusses the significant advancements and challenges in TSMC's CoWoS (Chip-on-Wafer-on-Substrate) packaging technology, particularly in relation to NVIDIA's evolving needs in the AI sector, highlighting the shift towards CoWoS-L and the emergence of CoPoS (Chip-on-Panel-on-Substrate) as a potential alternative [1][3][11]. Group 1: TSMC and NVIDIA Collaboration - TSMC has become a crucial partner for NVIDIA, especially in the CoWoS technology, with NVIDIA's CEO stating that they have no alternative to TSMC in this area [1]. - NVIDIA is transitioning to use more CoWoS-L packaging for its upcoming Blackwell series products, which require high bandwidth interconnects [3][6]. Group 2: CoWoS Technology Developments - TSMC has been expanding its CoWoS capacity significantly over the past two years and is reportedly surpassing ASE Group to become the largest packaging player globally [1]. - The CoWoS technology is evolving, with TSMC planning to introduce CoWoS-L with a mask size of 5.5 times by 2026 and a record 9.5 times by 2027 [9]. Group 3: Challenges in CoWoS Technology - The increasing chip sizes pose challenges for CoWoS packaging, as larger chips reduce the number of chips that can fit on a 12-inch wafer [6]. - TSMC is facing difficulties with flux usage in CoWoS, which is essential for chip bonding, and is exploring flux-free bonding technologies [7][9]. Group 4: Emergence of CoPoS Technology - CoPoS technology is being developed as a next-generation alternative to CoWoS, allowing for higher chip density and efficiency by using a panel instead of a wafer [11][14]. - TSMC's AP7 factory is set to become a key hub for advanced packaging, focusing on CoPoS production [12]. Group 5: Comparison of FOPLP and CoPoS - FOPLP (Fan-out Panel-Level Packaging) and CoPoS both utilize large panel substrates but differ in architecture and application, with CoPoS offering better signal integrity due to its use of an interposer [12][13]. - CoPoS is positioned for high-end applications in AI and HPC systems, while FOPLP is more suited for mid-range applications [13][14].