TSMC Packaging Bottleneck Keeps Nvidia AI GPUs in Short Supply
TSMC's chairman reveals that a lack of advanced packaging capacity, rather than chip production limits, drives the ongoing shortage of Nvidia's AI GPUs. The company expects this supply constraint to last for approximately 18 months.
TSMC's chairman admits that the ongoing shortage of Nvidia's AI GPUs stems from limited chip-on-wafer-on-substrate (CoWoS) packaging capacity rather than a lack of actual silicon production. Because high-performance AI processors require this specific advanced packaging to integrate HBM memory, TSMC's facilities experience severe strain as demand for generative AI surges.
Currently, TSMC fulfills only about 80 percent of its customers' needs for these crucial AI components, which power popular tools like ChatGPT. The company reports that demand for CoWoS packaging triples year-over-year, creating a significant bottleneck for Nvidia, AMD, and other tech giants developing artificial intelligence hardware.
To resolve the supply crunch, TSMC accelerates the expansion of its advanced packaging capabilities by installing additional tools at existing facilities. However, because this equipment setup takes considerable time, the company anticipates that its CoWoS capacity only doubles by the end of 2024, meaning the GPU shortage persists for roughly 18 months.