According to the latest research from TrendForce, the rapid growth in AI demand since 2023 has led to capacity bottlenecks for 3nm to 2nm wafers and 2.5D/3D packaging. The shortage of CoWoS has persisted, even causing tightness in related production equipment, downstream packaging substrates, and key peripheral raw materials. The front-end 3nm advanced process, temporarily supplied exclusively by Taiwan Semiconductor Manufacturing, is not only facing tighter capacity but has also become a scarce resource fiercely contested by global tech giants. TrendForce indicates that the resource competition sparked by the AI race has spread throughout the entire semiconductor supply chain. Leading firm NVIDIA, leveraging its long-term experience and strong control over the supply chain, was the first to observe signs of capacity constraints and pre-booked substantial advanced 4/3nm wafer foundry capacity, CoWoS capacity, and materials such as T-glass, substrates, PCBs, HBM, and SSDs. Other major tech companies like Google, despite having strong demand, failed to secure key component capacity in time, resulting in severe component shortages that hampered product growth momentum. As AI computing power demands increase packaging area, the consumption of wafer and packaging resources per chip has multiplied. Even with TSMC actively building new facilities to expand production, CoWoS has been in short supply since 2023, prompting customers to seek additional capacity resources. Besides OSAT providers like SPIL and Amkor benefiting from this trend, Intel's EMIB and SPIL's FOEB have attracted customer attention due to their similar technologies, with Intel also holding the advantage of local manufacturing in the U.S. TrendForce points out that due to significant order spillover effects and TSMC's plans to add over 60% more CoWoS capacity by 2027, the severe shortage in global 2.5D packaging capacity is expected to slightly ease by 2027. Regarding front-end advanced process demand, as mainstream AI computing chips shift en masse from 4nm to 3nm in the second half of 2025 to 2026, and high-end processors for smartphones and PCs have not yet widely transitioned to the next-generation 2nm node, all high-computing-power demands are concentrated on the 3nm process within a very short period. On the supply side, since Samsung and Intel lag behind TSMC in 3nm foundry progress, and many chip designs were finalized one to three years ago, TSMC currently holds a monopoly on supply. To alleviate the supply-demand imbalance, TSMC is actively expanding its 3nm facilities. After the new capacity comes online, global total 3nm capacity is expected to officially surpass that of 5/4nm by the end of 2026 and become the second most critical process node after 28nm by 2027.
Comments