The battle for Taiwan Semiconductor Manufacturing's (TSMC) most advanced manufacturing capacity has officially commenced. Global tech giants are rapidly converging on the 2nm process node, with its entire capacity already fully booked, while advanced packaging supply is simultaneously tightening, highlighting the sustained pressure on the semiconductor supply chain from the combined demand for AI and mobile chips. Nvidia CEO Jensen Huang, while hosting a dinner for key supply chain executives on January 31st, stated that TSMC must operate at full capacity this year, directly pointing to the current tightness in advanced process capacity. This remark confirms the industry's assessment of the critical shortage in TSMC's 2nm production capabilities. According to industry sources, Apple has secured more than half of the initial 2nm capacity, with Qualcomm also being a major client for 2026. AMD plans to commence production of CPUs based on the 2nm process in 2026, while Google and Amazon Web Services (AWS) are targeting adoption of this technology in the third and fourth quarters of 2027, respectively. Nvidia is looking even further ahead to 2028, with its anticipated "Feynman" AI GPU expected to utilize TSMC's A16 process, which integrates backside power delivery technology. The capacity crunch is projected to persist until 2027. AI accelerators and mobile processors are simultaneously competing for the limited available capacity, while yield challenges in advanced packaging are further exacerbating the supply-demand imbalance. Institutional investors anticipate that TSMC's monthly CoWoS capacity will increase by over 70% year-over-year by 2026, yet this is still expected to fall short of market demand. Mobile chips are dominating the initial capacity allocation, with AI clients scaling up significantly from 2027. TSMC's 2nm and 3nm process nodes are both facing capacity constraints, with high-performance computing and mobile chips vying for the limited supply. Apple and Qualcomm are the primary 2nm customers for 2026. According to Wccftech citing sources, Apple has secured over half of the first batch of 2nm capacity. Starting in 2027, general-purpose GPUs and custom ASICs will see much broader adoption. Analysis indicates this includes AMD's MI-series GPUs, Google's eighth-generation TPU, and AWS's Trainium 4. Industry sources predict that TSMC's 2nm family will become a long-lifecycle node, with its initial production ramp potentially exceeding that of the 3nm generation. The N2 process is scheduled to enter mass production in 2026, with the N2P and A16 processes following in the second half of the year. The A16 process is specifically designed for certain high-performance computing products requiring complex wiring and high-density power delivery. Nvidia is skipping 2nm, heading straight for 1.6nm, betting on backside power delivery technology. Nvidia's process roadmap reveals a different strategy. Reports indicate the company plans to launch its "Feynman" AI GPU in 2028, expected to use TSMC's A16 process, noted for its backside power delivery feature. The A16 process represents TSMC's 1.6nm node, tailored for high-performance computing products. Backside power delivery technology relocates the power delivery network to the chip's backside, which can improve signal integrity and enhance power transmission efficiency—a critical factor for large AI accelerators. This timeline suggests Nvidia may bypass or only minimally adopt the 2nm process, moving directly to a more advanced node, reflecting the aggressive pursuit of process technology by AI chip manufacturers. Advanced packaging emerges as a new bottleneck, with CoWoS capacity growth struggling to keep pace with demand. The capacity tightness is not limited to the wafer fabrication stage. Reports indicate TSMC is upgrading its advanced packaging ecosystem. As AI chips fully enter the era of chiplet architectures and超大封装尺寸, single-die designs can no longer meet computational demands, making technologies like CoWoS-L, SoIC, and hybrid bonding effectively standard. Citing institutional investor information, reports state that TSMC aims for its monthly CoWoS capacity to grow by over 70% year-over-year by 2026, while concurrently validating next-generation technologies such as CoWoP (Chip-on-Wafer-on-PCB) and CPO (co-packaged optics). However, the supply-demand imbalance remains a critical bottleneck. Beyond the tight 2nm foundry capacity, improving yields for large-scale system-level packaging presents another significant challenge. As the physical size of AI chip packages continues to increase, the difficulty of maintaining high yields rises substantially, potentially further constraining the supply of advanced chips.
Comments