NVIDIA’s Blackwell GPUs Push CoWoS Limits, FOPLP Adoption Looms

Due to the surging demand for artificial intelligence (AI) chips, NVIDIA’s data center GPU sales have been exceptionally robust, leading to a persistent strain on TSMC’s CoWoS packaging capacity over the past year. In response, both NVIDIA and TSMC have been striving to expand capacity and optimize the supply chain to alleviate the tight supply situation.

NVIDIA’s Blackwell architecture-based GPUs are among the most powerful AI chips on the market. These products are set to be mass-produced soon, posing another significant challenge to CoWoS packaging capacity. According to Wccftech, NVIDIA is interested in introducing FOPLP (panel-level fan-out packaging) technology to the Blackwell architecture chips, specifically for the GB200, likely starting in 2025.

Market research firms indicate that NVIDIA’s GB200 supply chain is already in motion, currently in the design tweaking and testing phase. It is projected that shipments will reach approximately 420,000 units this year and between 1.5 million to 2 million units next year. Amid the ongoing CoWoS capacity shortage, the introduction of FOPLP packaging technology offers NVIDIA more options in packaging and somewhat mitigates the pressure from insufficient packaging capacity. Reports suggest that the chosen FOPLP packaging employs glass substrates, which can withstand higher temperatures for longer periods while maintaining optimal performance.

It is understood that NVIDIA initially planned to introduce FOPLP packaging technology in 2026. However, with the shifting market dynamics, the timeline has been advanced.