Core Viewpoint - TSMC is facing capacity constraints in its CoWoS advanced packaging, leading to order overflow to other packaging companies, including Intel's EMIB technology, raising concerns about TSMC's market share loss. However, TSMC is actively investing in the CoPoS technology for future growth [1] Group 1: TSMC's Capacity and Market Dynamics - TSMC's CoWoS advanced packaging capacity is expected to reach 70,000 to 80,000 units this year, doubling from last year's 30,000 units, with projections of 110,000 to 120,000 units by 2026 [1] - Despite TSMC's capacity expansion, there is a growing gap in demand, leading to orders being diverted to OSAT manufacturers and even Intel's EMIB process, which is attracting clients like Google and Meta due to its lower costs and higher yields [1] - Concerns about TSMC's CoWoS capacity expansion are overshadowed by a more significant shortage in N3 advanced process capacity, which is critical for upcoming chips from NVIDIA, AMD, and Google [1] Group 2: Capital Expenditure and Technology Transition - Discussions are ongoing regarding TSMC potentially increasing its capital expenditure for next year, primarily focused on N3 capacity expansion, which is expected to rise from around 120,000 units to 150,000 to 160,000 units [2] - TSMC's cautious approach to expanding CoWoS capacity is attributed to the increasing size of Blackwell chips, which leads to more waste when using round wafers. The company is developing CoPoS technology to optimize future production [2] - Intel's EMIB packaging business is positioned advantageously due to its smaller embedded bridging chips, which reduce material costs and offer a price advantage of 50% compared to TSMC's CoWoS [2] Group 3: Revenue Growth and Market Trends - TSMC's AI revenue has surpassed that of Apple this year, with AI chips for clients like Microsoft and NVIDIA commanding higher average selling prices (ASP), leading to a projected revenue growth of 26% by 2026, exceeding market consensus of 15-20% [3] - The market for ASICs (custom chips) is expected to surpass that of GPUs by 2027, driven by the increasing usage of AI models, as ASICs offer lower operational costs compared to general-purpose GPUs [3] - Companies like Google and AWS are promoting their self-developed chips (TPUs) alongside cloud services, enhancing customer willingness to rent computing power, with Google offering competitive pricing at $0.5 per MW of computing power [3]