TSMC will double its CoWoS packaging capacity this year

In recent months, the rise of Artificial Intelligence (AI) tools, spearheaded by ChatGPT, has dramatically increased the demand for data center GPUs like Nvidia’s A100 and H100. This surge in demand has placed considerable pressure on TSMC, responsible for manufacturing and packaging these GPUs. To meet the escalating needs, TSMC has urgently expanded its 2.5D packaging capacity.

As reported by DigiTimes, TSMC is fully committed to addressing the high demand for CoWoS (Chip-on-Wafer-on-Substrate) packaging, with plans to double its capacity this year. TSMC is optimistic about the robust chip demand, including AI and high-performance computing, stating that “almost all AI entities are collaborating with TSMC.” Industry insiders note that the influx of ‘urgent orders’ from clients like Nvidia and AMD is incessantly increasing. Nvidia accounts for nearly half of these orders, and despite a reduced lead time for the H100, it still stands at a lengthy 10 months, indicative of sustained high demand. The shipment rate of AI servers this year has notably accelerated, fueling revenue growth for related companies.

Projections suggest that by the end of 2024, TSMC’s CoWoS packaging capacity will reach at least 32,000 wafers per month, with a further increase to 44,000 wafers per month by the end of 2025. Reports from the end of last year indicated that CoWoS packaging capacity had risen to around 15,000 wafers per month, with Nvidia occupying 40% of this capacity and AMD 8%. By the first half of 2024, TSMC is expected to boost its CoWoS packaging capacity to 20,000 wafers per month.

Currently, foundries are increasingly focusing on packaging investments. Beyond TSMC, both Samsung and Intel have intensified their investments in advanced packaging technologies and capacities, recognizing advanced packaging as a critical factor in shaping the future competitive landscape of the semiconductor industry.